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 <trace/events/qdisc.h>
135 #include <linux/inetdevice.h>
136 #include <linux/cpu_rmap.h>
137 #include <linux/static_key.h>
138 #include <linux/hashtable.h>
139 #include <linux/vmalloc.h>
140 #include <linux/if_macvlan.h>
141 #include <linux/errqueue.h>
142 #include <linux/hrtimer.h>
143 #include <linux/netfilter_ingress.h>
144 #include <linux/crash_dump.h>
145 #include <linux/sctp.h>
146 #include <net/udp_tunnel.h>
147 #include <linux/net_namespace.h>
148 #include <linux/indirect_call_wrapper.h>
149 #include <net/devlink.h>
150 #include <linux/pm_runtime.h>
151 #include <linux/prandom.h>
152 #include <linux/once_lite.h>
154 #include "net-sysfs.h"
156 #define MAX_GRO_SKBS 8
158 /* This should be increased if a protocol with a bigger head is added. */
159 #define GRO_MAX_HEAD (MAX_HEADER + 128)
161 static DEFINE_SPINLOCK(ptype_lock);
162 static DEFINE_SPINLOCK(offload_lock);
163 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
164 struct list_head ptype_all __read_mostly; /* Taps */
165 static struct list_head offload_base __read_mostly;
167 static int netif_rx_internal(struct sk_buff *skb);
168 static int call_netdevice_notifiers_info(unsigned long val,
169 struct netdev_notifier_info *info);
170 static int call_netdevice_notifiers_extack(unsigned long val,
171 struct net_device *dev,
172 struct netlink_ext_ack *extack);
173 static struct napi_struct *napi_by_id(unsigned int napi_id);
176 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
179 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
181 * Writers must hold the rtnl semaphore while they loop through the
182 * dev_base_head list, and hold dev_base_lock for writing when they do the
183 * actual updates. This allows pure readers to access the list even
184 * while a writer is preparing to update it.
186 * To put it another way, dev_base_lock is held for writing only to
187 * protect against pure readers; the rtnl semaphore provides the
188 * protection against other writers.
190 * See, for example usages, register_netdevice() and
191 * unregister_netdevice(), which must be called with the rtnl
194 DEFINE_RWLOCK(dev_base_lock);
195 EXPORT_SYMBOL(dev_base_lock);
197 static DEFINE_MUTEX(ifalias_mutex);
199 /* protects napi_hash addition/deletion and napi_gen_id */
200 static DEFINE_SPINLOCK(napi_hash_lock);
202 static unsigned int napi_gen_id = NR_CPUS;
203 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
205 static DECLARE_RWSEM(devnet_rename_sem);
207 static inline void dev_base_seq_inc(struct net *net)
209 while (++net->dev_base_seq == 0)
213 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
215 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
217 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
220 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
222 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
225 static inline void rps_lock(struct softnet_data *sd)
228 spin_lock(&sd->input_pkt_queue.lock);
232 static inline void rps_unlock(struct softnet_data *sd)
235 spin_unlock(&sd->input_pkt_queue.lock);
239 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev,
242 struct netdev_name_node *name_node;
244 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL);
247 INIT_HLIST_NODE(&name_node->hlist);
248 name_node->dev = dev;
249 name_node->name = name;
253 static struct netdev_name_node *
254 netdev_name_node_head_alloc(struct net_device *dev)
256 struct netdev_name_node *name_node;
258 name_node = netdev_name_node_alloc(dev, dev->name);
261 INIT_LIST_HEAD(&name_node->list);
265 static void netdev_name_node_free(struct netdev_name_node *name_node)
270 static void netdev_name_node_add(struct net *net,
271 struct netdev_name_node *name_node)
273 hlist_add_head_rcu(&name_node->hlist,
274 dev_name_hash(net, name_node->name));
277 static void netdev_name_node_del(struct netdev_name_node *name_node)
279 hlist_del_rcu(&name_node->hlist);
282 static struct netdev_name_node *netdev_name_node_lookup(struct net *net,
285 struct hlist_head *head = dev_name_hash(net, name);
286 struct netdev_name_node *name_node;
288 hlist_for_each_entry(name_node, head, hlist)
289 if (!strcmp(name_node->name, name))
294 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net,
297 struct hlist_head *head = dev_name_hash(net, name);
298 struct netdev_name_node *name_node;
300 hlist_for_each_entry_rcu(name_node, head, hlist)
301 if (!strcmp(name_node->name, name))
306 int netdev_name_node_alt_create(struct net_device *dev, const char *name)
308 struct netdev_name_node *name_node;
309 struct net *net = dev_net(dev);
311 name_node = netdev_name_node_lookup(net, name);
314 name_node = netdev_name_node_alloc(dev, name);
317 netdev_name_node_add(net, name_node);
318 /* The node that holds dev->name acts as a head of per-device list. */
319 list_add_tail(&name_node->list, &dev->name_node->list);
323 EXPORT_SYMBOL(netdev_name_node_alt_create);
325 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node)
327 list_del(&name_node->list);
328 netdev_name_node_del(name_node);
329 kfree(name_node->name);
330 netdev_name_node_free(name_node);
333 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name)
335 struct netdev_name_node *name_node;
336 struct net *net = dev_net(dev);
338 name_node = netdev_name_node_lookup(net, name);
341 /* lookup might have found our primary name or a name belonging
344 if (name_node == dev->name_node || name_node->dev != dev)
347 __netdev_name_node_alt_destroy(name_node);
351 EXPORT_SYMBOL(netdev_name_node_alt_destroy);
353 static void netdev_name_node_alt_flush(struct net_device *dev)
355 struct netdev_name_node *name_node, *tmp;
357 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list)
358 __netdev_name_node_alt_destroy(name_node);
361 /* Device list insertion */
362 static void list_netdevice(struct net_device *dev)
364 struct net *net = dev_net(dev);
368 write_lock_bh(&dev_base_lock);
369 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
370 netdev_name_node_add(net, dev->name_node);
371 hlist_add_head_rcu(&dev->index_hlist,
372 dev_index_hash(net, dev->ifindex));
373 write_unlock_bh(&dev_base_lock);
375 dev_base_seq_inc(net);
378 /* Device list removal
379 * caller must respect a RCU grace period before freeing/reusing dev
381 static void unlist_netdevice(struct net_device *dev)
385 /* Unlink dev from the device chain */
386 write_lock_bh(&dev_base_lock);
387 list_del_rcu(&dev->dev_list);
388 netdev_name_node_del(dev->name_node);
389 hlist_del_rcu(&dev->index_hlist);
390 write_unlock_bh(&dev_base_lock);
392 dev_base_seq_inc(dev_net(dev));
399 static RAW_NOTIFIER_HEAD(netdev_chain);
402 * Device drivers call our routines to queue packets here. We empty the
403 * queue in the local softnet handler.
406 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
407 EXPORT_PER_CPU_SYMBOL(softnet_data);
409 #ifdef CONFIG_LOCKDEP
411 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
412 * according to dev->type
414 static const unsigned short netdev_lock_type[] = {
415 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
416 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
417 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
418 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
419 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
420 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
421 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
422 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
423 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
424 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
425 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
426 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
427 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
428 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
429 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
431 static const char *const netdev_lock_name[] = {
432 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
433 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
434 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
435 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
436 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
437 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
438 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
439 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
440 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
441 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
442 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
443 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
444 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
445 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
446 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
448 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
449 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
451 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
455 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
456 if (netdev_lock_type[i] == dev_type)
458 /* the last key is used by default */
459 return ARRAY_SIZE(netdev_lock_type) - 1;
462 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
463 unsigned short dev_type)
467 i = netdev_lock_pos(dev_type);
468 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
469 netdev_lock_name[i]);
472 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
476 i = netdev_lock_pos(dev->type);
477 lockdep_set_class_and_name(&dev->addr_list_lock,
478 &netdev_addr_lock_key[i],
479 netdev_lock_name[i]);
482 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
483 unsigned short dev_type)
487 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
492 /*******************************************************************************
494 * Protocol management and registration routines
496 *******************************************************************************/
500 * Add a protocol ID to the list. Now that the input handler is
501 * smarter we can dispense with all the messy stuff that used to be
504 * BEWARE!!! Protocol handlers, mangling input packets,
505 * MUST BE last in hash buckets and checking protocol handlers
506 * MUST start from promiscuous ptype_all chain in net_bh.
507 * It is true now, do not change it.
508 * Explanation follows: if protocol handler, mangling packet, will
509 * be the first on list, it is not able to sense, that packet
510 * is cloned and should be copied-on-write, so that it will
511 * change it and subsequent readers will get broken packet.
515 static inline struct list_head *ptype_head(const struct packet_type *pt)
517 if (pt->type == htons(ETH_P_ALL))
518 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
520 return pt->dev ? &pt->dev->ptype_specific :
521 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
525 * dev_add_pack - add packet handler
526 * @pt: packet type declaration
528 * Add a protocol handler to the networking stack. The passed &packet_type
529 * is linked into kernel lists and may not be freed until it has been
530 * removed from the kernel lists.
532 * This call does not sleep therefore it can not
533 * guarantee all CPU's that are in middle of receiving packets
534 * will see the new packet type (until the next received packet).
537 void dev_add_pack(struct packet_type *pt)
539 struct list_head *head = ptype_head(pt);
541 spin_lock(&ptype_lock);
542 list_add_rcu(&pt->list, head);
543 spin_unlock(&ptype_lock);
545 EXPORT_SYMBOL(dev_add_pack);
548 * __dev_remove_pack - remove packet handler
549 * @pt: packet type declaration
551 * Remove a protocol handler that was previously added to the kernel
552 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
553 * from the kernel lists and can be freed or reused once this function
556 * The packet type might still be in use by receivers
557 * and must not be freed until after all the CPU's have gone
558 * through a quiescent state.
560 void __dev_remove_pack(struct packet_type *pt)
562 struct list_head *head = ptype_head(pt);
563 struct packet_type *pt1;
565 spin_lock(&ptype_lock);
567 list_for_each_entry(pt1, head, list) {
569 list_del_rcu(&pt->list);
574 pr_warn("dev_remove_pack: %p not found\n", pt);
576 spin_unlock(&ptype_lock);
578 EXPORT_SYMBOL(__dev_remove_pack);
581 * dev_remove_pack - remove packet handler
582 * @pt: packet type declaration
584 * Remove a protocol handler that was previously added to the kernel
585 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
586 * from the kernel lists and can be freed or reused once this function
589 * This call sleeps to guarantee that no CPU is looking at the packet
592 void dev_remove_pack(struct packet_type *pt)
594 __dev_remove_pack(pt);
598 EXPORT_SYMBOL(dev_remove_pack);
602 * dev_add_offload - register offload handlers
603 * @po: protocol offload declaration
605 * Add protocol offload handlers to the networking stack. The passed
606 * &proto_offload is linked into kernel lists and may not be freed until
607 * it has been removed from the kernel lists.
609 * This call does not sleep therefore it can not
610 * guarantee all CPU's that are in middle of receiving packets
611 * will see the new offload handlers (until the next received packet).
613 void dev_add_offload(struct packet_offload *po)
615 struct packet_offload *elem;
617 spin_lock(&offload_lock);
618 list_for_each_entry(elem, &offload_base, list) {
619 if (po->priority < elem->priority)
622 list_add_rcu(&po->list, elem->list.prev);
623 spin_unlock(&offload_lock);
625 EXPORT_SYMBOL(dev_add_offload);
628 * __dev_remove_offload - remove offload handler
629 * @po: packet offload declaration
631 * Remove a protocol offload handler that was previously added to the
632 * kernel offload handlers by dev_add_offload(). The passed &offload_type
633 * is removed from the kernel lists and can be freed or reused once this
636 * The packet type might still be in use by receivers
637 * and must not be freed until after all the CPU's have gone
638 * through a quiescent state.
640 static void __dev_remove_offload(struct packet_offload *po)
642 struct list_head *head = &offload_base;
643 struct packet_offload *po1;
645 spin_lock(&offload_lock);
647 list_for_each_entry(po1, head, list) {
649 list_del_rcu(&po->list);
654 pr_warn("dev_remove_offload: %p not found\n", po);
656 spin_unlock(&offload_lock);
660 * dev_remove_offload - remove packet offload handler
661 * @po: packet offload declaration
663 * Remove a packet offload handler that was previously added to the kernel
664 * offload handlers by dev_add_offload(). The passed &offload_type is
665 * removed from the kernel lists and can be freed or reused once this
668 * This call sleeps to guarantee that no CPU is looking at the packet
671 void dev_remove_offload(struct packet_offload *po)
673 __dev_remove_offload(po);
677 EXPORT_SYMBOL(dev_remove_offload);
679 /*******************************************************************************
681 * Device Interface Subroutines
683 *******************************************************************************/
686 * dev_get_iflink - get 'iflink' value of a interface
687 * @dev: targeted interface
689 * Indicates the ifindex the interface is linked to.
690 * Physical interfaces have the same 'ifindex' and 'iflink' values.
693 int dev_get_iflink(const struct net_device *dev)
695 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
696 return dev->netdev_ops->ndo_get_iflink(dev);
700 EXPORT_SYMBOL(dev_get_iflink);
703 * dev_fill_metadata_dst - Retrieve tunnel egress information.
704 * @dev: targeted interface
707 * For better visibility of tunnel traffic OVS needs to retrieve
708 * egress tunnel information for a packet. Following API allows
709 * user to get this info.
711 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
713 struct ip_tunnel_info *info;
715 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
718 info = skb_tunnel_info_unclone(skb);
721 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
724 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
726 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
728 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack)
730 int k = stack->num_paths++;
732 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX))
735 return &stack->path[k];
738 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr,
739 struct net_device_path_stack *stack)
741 const struct net_device *last_dev;
742 struct net_device_path_ctx ctx = {
746 struct net_device_path *path;
749 stack->num_paths = 0;
750 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) {
752 path = dev_fwd_path(stack);
756 memset(path, 0, sizeof(struct net_device_path));
757 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path);
761 if (WARN_ON_ONCE(last_dev == ctx.dev))
764 path = dev_fwd_path(stack);
767 path->type = DEV_PATH_ETHERNET;
772 EXPORT_SYMBOL_GPL(dev_fill_forward_path);
775 * __dev_get_by_name - find a device by its name
776 * @net: the applicable net namespace
777 * @name: name to find
779 * Find an interface by name. Must be called under RTNL semaphore
780 * or @dev_base_lock. If the name is found a pointer to the device
781 * is returned. If the name is not found then %NULL is returned. The
782 * reference counters are not incremented so the caller must be
783 * careful with locks.
786 struct net_device *__dev_get_by_name(struct net *net, const char *name)
788 struct netdev_name_node *node_name;
790 node_name = netdev_name_node_lookup(net, name);
791 return node_name ? node_name->dev : NULL;
793 EXPORT_SYMBOL(__dev_get_by_name);
796 * dev_get_by_name_rcu - find a device by its name
797 * @net: the applicable net namespace
798 * @name: name to find
800 * Find an interface by name.
801 * If the name is found a pointer to the device is returned.
802 * If the name is not found then %NULL is returned.
803 * The reference counters are not incremented so the caller must be
804 * careful with locks. The caller must hold RCU lock.
807 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
809 struct netdev_name_node *node_name;
811 node_name = netdev_name_node_lookup_rcu(net, name);
812 return node_name ? node_name->dev : NULL;
814 EXPORT_SYMBOL(dev_get_by_name_rcu);
817 * dev_get_by_name - find a device by its name
818 * @net: the applicable net namespace
819 * @name: name to find
821 * Find an interface by name. This can be called from any
822 * context and does its own locking. The returned handle has
823 * the usage count incremented and the caller must use dev_put() to
824 * release it when it is no longer needed. %NULL is returned if no
825 * matching device is found.
828 struct net_device *dev_get_by_name(struct net *net, const char *name)
830 struct net_device *dev;
833 dev = dev_get_by_name_rcu(net, name);
838 EXPORT_SYMBOL(dev_get_by_name);
841 * __dev_get_by_index - find a device by its ifindex
842 * @net: the applicable net namespace
843 * @ifindex: index of device
845 * Search for an interface by index. Returns %NULL if the device
846 * is not found or a pointer to the device. The device has not
847 * had its reference counter increased so the caller must be careful
848 * about locking. The caller must hold either the RTNL semaphore
852 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
854 struct net_device *dev;
855 struct hlist_head *head = dev_index_hash(net, ifindex);
857 hlist_for_each_entry(dev, head, index_hlist)
858 if (dev->ifindex == ifindex)
863 EXPORT_SYMBOL(__dev_get_by_index);
866 * dev_get_by_index_rcu - find a device by its ifindex
867 * @net: the applicable net namespace
868 * @ifindex: index of device
870 * Search for an interface by index. Returns %NULL if the device
871 * is not found or a pointer to the device. The device has not
872 * had its reference counter increased so the caller must be careful
873 * about locking. The caller must hold RCU lock.
876 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
878 struct net_device *dev;
879 struct hlist_head *head = dev_index_hash(net, ifindex);
881 hlist_for_each_entry_rcu(dev, head, index_hlist)
882 if (dev->ifindex == ifindex)
887 EXPORT_SYMBOL(dev_get_by_index_rcu);
891 * dev_get_by_index - find a device by its ifindex
892 * @net: the applicable net namespace
893 * @ifindex: index of device
895 * Search for an interface by index. Returns NULL if the device
896 * is not found or a pointer to the device. The device returned has
897 * had a reference added and the pointer is safe until the user calls
898 * dev_put to indicate they have finished with it.
901 struct net_device *dev_get_by_index(struct net *net, int ifindex)
903 struct net_device *dev;
906 dev = dev_get_by_index_rcu(net, ifindex);
911 EXPORT_SYMBOL(dev_get_by_index);
914 * dev_get_by_napi_id - find a device by napi_id
915 * @napi_id: ID of the NAPI struct
917 * Search for an interface by NAPI ID. Returns %NULL if the device
918 * is not found or a pointer to the device. The device has not had
919 * its reference counter increased so the caller must be careful
920 * about locking. The caller must hold RCU lock.
923 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
925 struct napi_struct *napi;
927 WARN_ON_ONCE(!rcu_read_lock_held());
929 if (napi_id < MIN_NAPI_ID)
932 napi = napi_by_id(napi_id);
934 return napi ? napi->dev : NULL;
936 EXPORT_SYMBOL(dev_get_by_napi_id);
939 * netdev_get_name - get a netdevice name, knowing its ifindex.
940 * @net: network namespace
941 * @name: a pointer to the buffer where the name will be stored.
942 * @ifindex: the ifindex of the interface to get the name from.
944 int netdev_get_name(struct net *net, char *name, int ifindex)
946 struct net_device *dev;
949 down_read(&devnet_rename_sem);
952 dev = dev_get_by_index_rcu(net, ifindex);
958 strcpy(name, dev->name);
963 up_read(&devnet_rename_sem);
968 * dev_getbyhwaddr_rcu - find a device by its hardware address
969 * @net: the applicable net namespace
970 * @type: media type of device
971 * @ha: hardware address
973 * Search for an interface by MAC address. Returns NULL if the device
974 * is not found or a pointer to the device.
975 * The caller must hold RCU or RTNL.
976 * The returned device has not had its ref count increased
977 * and the caller must therefore be careful about locking
981 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
984 struct net_device *dev;
986 for_each_netdev_rcu(net, dev)
987 if (dev->type == type &&
988 !memcmp(dev->dev_addr, ha, dev->addr_len))
993 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
995 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
997 struct net_device *dev, *ret = NULL;
1000 for_each_netdev_rcu(net, dev)
1001 if (dev->type == type) {
1009 EXPORT_SYMBOL(dev_getfirstbyhwtype);
1012 * __dev_get_by_flags - find any device with given flags
1013 * @net: the applicable net namespace
1014 * @if_flags: IFF_* values
1015 * @mask: bitmask of bits in if_flags to check
1017 * Search for any interface with the given flags. Returns NULL if a device
1018 * is not found or a pointer to the device. Must be called inside
1019 * rtnl_lock(), and result refcount is unchanged.
1022 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1023 unsigned short mask)
1025 struct net_device *dev, *ret;
1030 for_each_netdev(net, dev) {
1031 if (((dev->flags ^ if_flags) & mask) == 0) {
1038 EXPORT_SYMBOL(__dev_get_by_flags);
1041 * dev_valid_name - check if name is okay for network device
1042 * @name: name string
1044 * Network device names need to be valid file names to
1045 * allow sysfs to work. We also disallow any kind of
1048 bool dev_valid_name(const char *name)
1052 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1054 if (!strcmp(name, ".") || !strcmp(name, ".."))
1058 if (*name == '/' || *name == ':' || isspace(*name))
1064 EXPORT_SYMBOL(dev_valid_name);
1067 * __dev_alloc_name - allocate a name for a device
1068 * @net: network namespace to allocate the device name in
1069 * @name: name format string
1070 * @buf: scratch buffer and result name string
1072 * Passed a format string - eg "lt%d" it will try and find a suitable
1073 * id. It scans list of devices to build up a free map, then chooses
1074 * the first empty slot. The caller must hold the dev_base or rtnl lock
1075 * while allocating the name and adding the device in order to avoid
1077 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1078 * Returns the number of the unit assigned or a negative errno code.
1081 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1085 const int max_netdevices = 8*PAGE_SIZE;
1086 unsigned long *inuse;
1087 struct net_device *d;
1089 if (!dev_valid_name(name))
1092 p = strchr(name, '%');
1095 * Verify the string as this thing may have come from
1096 * the user. There must be either one "%d" and no other "%"
1099 if (p[1] != 'd' || strchr(p + 2, '%'))
1102 /* Use one page as a bit array of possible slots */
1103 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1107 for_each_netdev(net, d) {
1108 struct netdev_name_node *name_node;
1109 list_for_each_entry(name_node, &d->name_node->list, list) {
1110 if (!sscanf(name_node->name, name, &i))
1112 if (i < 0 || i >= max_netdevices)
1115 /* avoid cases where sscanf is not exact inverse of printf */
1116 snprintf(buf, IFNAMSIZ, name, i);
1117 if (!strncmp(buf, name_node->name, IFNAMSIZ))
1120 if (!sscanf(d->name, name, &i))
1122 if (i < 0 || i >= max_netdevices)
1125 /* avoid cases where sscanf is not exact inverse of printf */
1126 snprintf(buf, IFNAMSIZ, name, i);
1127 if (!strncmp(buf, d->name, IFNAMSIZ))
1131 i = find_first_zero_bit(inuse, max_netdevices);
1132 free_page((unsigned long) inuse);
1135 snprintf(buf, IFNAMSIZ, name, i);
1136 if (!__dev_get_by_name(net, buf))
1139 /* It is possible to run out of possible slots
1140 * when the name is long and there isn't enough space left
1141 * for the digits, or if all bits are used.
1146 static int dev_alloc_name_ns(struct net *net,
1147 struct net_device *dev,
1154 ret = __dev_alloc_name(net, name, buf);
1156 strlcpy(dev->name, buf, IFNAMSIZ);
1161 * dev_alloc_name - allocate a name for a device
1163 * @name: name format string
1165 * Passed a format string - eg "lt%d" it will try and find a suitable
1166 * id. It scans list of devices to build up a free map, then chooses
1167 * the first empty slot. The caller must hold the dev_base or rtnl lock
1168 * while allocating the name and adding the device in order to avoid
1170 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1171 * Returns the number of the unit assigned or a negative errno code.
1174 int dev_alloc_name(struct net_device *dev, const char *name)
1176 return dev_alloc_name_ns(dev_net(dev), dev, name);
1178 EXPORT_SYMBOL(dev_alloc_name);
1180 static int dev_get_valid_name(struct net *net, struct net_device *dev,
1185 if (!dev_valid_name(name))
1188 if (strchr(name, '%'))
1189 return dev_alloc_name_ns(net, dev, name);
1190 else if (__dev_get_by_name(net, name))
1192 else if (dev->name != name)
1193 strlcpy(dev->name, name, IFNAMSIZ);
1199 * dev_change_name - change name of a device
1201 * @newname: name (or format string) must be at least IFNAMSIZ
1203 * Change name of a device, can pass format strings "eth%d".
1206 int dev_change_name(struct net_device *dev, const char *newname)
1208 unsigned char old_assign_type;
1209 char oldname[IFNAMSIZ];
1215 BUG_ON(!dev_net(dev));
1219 /* Some auto-enslaved devices e.g. failover slaves are
1220 * special, as userspace might rename the device after
1221 * the interface had been brought up and running since
1222 * the point kernel initiated auto-enslavement. Allow
1223 * live name change even when these slave devices are
1226 * Typically, users of these auto-enslaving devices
1227 * don't actually care about slave name change, as
1228 * they are supposed to operate on master interface
1231 if (dev->flags & IFF_UP &&
1232 likely(!(dev->priv_flags & IFF_LIVE_RENAME_OK)))
1235 down_write(&devnet_rename_sem);
1237 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1238 up_write(&devnet_rename_sem);
1242 memcpy(oldname, dev->name, IFNAMSIZ);
1244 err = dev_get_valid_name(net, dev, newname);
1246 up_write(&devnet_rename_sem);
1250 if (oldname[0] && !strchr(oldname, '%'))
1251 netdev_info(dev, "renamed from %s\n", oldname);
1253 old_assign_type = dev->name_assign_type;
1254 dev->name_assign_type = NET_NAME_RENAMED;
1257 ret = device_rename(&dev->dev, dev->name);
1259 memcpy(dev->name, oldname, IFNAMSIZ);
1260 dev->name_assign_type = old_assign_type;
1261 up_write(&devnet_rename_sem);
1265 up_write(&devnet_rename_sem);
1267 netdev_adjacent_rename_links(dev, oldname);
1269 write_lock_bh(&dev_base_lock);
1270 netdev_name_node_del(dev->name_node);
1271 write_unlock_bh(&dev_base_lock);
1275 write_lock_bh(&dev_base_lock);
1276 netdev_name_node_add(net, dev->name_node);
1277 write_unlock_bh(&dev_base_lock);
1279 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1280 ret = notifier_to_errno(ret);
1283 /* err >= 0 after dev_alloc_name() or stores the first errno */
1286 down_write(&devnet_rename_sem);
1287 memcpy(dev->name, oldname, IFNAMSIZ);
1288 memcpy(oldname, newname, IFNAMSIZ);
1289 dev->name_assign_type = old_assign_type;
1290 old_assign_type = NET_NAME_RENAMED;
1293 pr_err("%s: name change rollback failed: %d\n",
1302 * dev_set_alias - change ifalias of a device
1304 * @alias: name up to IFALIASZ
1305 * @len: limit of bytes to copy from info
1307 * Set ifalias for a device,
1309 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1311 struct dev_ifalias *new_alias = NULL;
1313 if (len >= IFALIASZ)
1317 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1321 memcpy(new_alias->ifalias, alias, len);
1322 new_alias->ifalias[len] = 0;
1325 mutex_lock(&ifalias_mutex);
1326 new_alias = rcu_replace_pointer(dev->ifalias, new_alias,
1327 mutex_is_locked(&ifalias_mutex));
1328 mutex_unlock(&ifalias_mutex);
1331 kfree_rcu(new_alias, rcuhead);
1335 EXPORT_SYMBOL(dev_set_alias);
1338 * dev_get_alias - get ifalias of a device
1340 * @name: buffer to store name of ifalias
1341 * @len: size of buffer
1343 * get ifalias for a device. Caller must make sure dev cannot go
1344 * away, e.g. rcu read lock or own a reference count to device.
1346 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1348 const struct dev_ifalias *alias;
1352 alias = rcu_dereference(dev->ifalias);
1354 ret = snprintf(name, len, "%s", alias->ifalias);
1361 * netdev_features_change - device changes features
1362 * @dev: device to cause notification
1364 * Called to indicate a device has changed features.
1366 void netdev_features_change(struct net_device *dev)
1368 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1370 EXPORT_SYMBOL(netdev_features_change);
1373 * netdev_state_change - device changes state
1374 * @dev: device to cause notification
1376 * Called to indicate a device has changed state. This function calls
1377 * the notifier chains for netdev_chain and sends a NEWLINK message
1378 * to the routing socket.
1380 void netdev_state_change(struct net_device *dev)
1382 if (dev->flags & IFF_UP) {
1383 struct netdev_notifier_change_info change_info = {
1387 call_netdevice_notifiers_info(NETDEV_CHANGE,
1389 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1392 EXPORT_SYMBOL(netdev_state_change);
1395 * __netdev_notify_peers - notify network peers about existence of @dev,
1396 * to be called when rtnl lock is already held.
1397 * @dev: network device
1399 * Generate traffic such that interested network peers are aware of
1400 * @dev, such as by generating a gratuitous ARP. This may be used when
1401 * a device wants to inform the rest of the network about some sort of
1402 * reconfiguration such as a failover event or virtual machine
1405 void __netdev_notify_peers(struct net_device *dev)
1408 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1409 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1411 EXPORT_SYMBOL(__netdev_notify_peers);
1414 * netdev_notify_peers - notify network peers about existence of @dev
1415 * @dev: network device
1417 * Generate traffic such that interested network peers are aware of
1418 * @dev, such as by generating a gratuitous ARP. This may be used when
1419 * a device wants to inform the rest of the network about some sort of
1420 * reconfiguration such as a failover event or virtual machine
1423 void netdev_notify_peers(struct net_device *dev)
1426 __netdev_notify_peers(dev);
1429 EXPORT_SYMBOL(netdev_notify_peers);
1431 static int napi_threaded_poll(void *data);
1433 static int napi_kthread_create(struct napi_struct *n)
1437 /* Create and wake up the kthread once to put it in
1438 * TASK_INTERRUPTIBLE mode to avoid the blocked task
1439 * warning and work with loadavg.
1441 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d",
1442 n->dev->name, n->napi_id);
1443 if (IS_ERR(n->thread)) {
1444 err = PTR_ERR(n->thread);
1445 pr_err("kthread_run failed with err %d\n", err);
1452 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1454 const struct net_device_ops *ops = dev->netdev_ops;
1459 if (!netif_device_present(dev)) {
1460 /* may be detached because parent is runtime-suspended */
1461 if (dev->dev.parent)
1462 pm_runtime_resume(dev->dev.parent);
1463 if (!netif_device_present(dev))
1467 /* Block netpoll from trying to do any rx path servicing.
1468 * If we don't do this there is a chance ndo_poll_controller
1469 * or ndo_poll may be running while we open the device
1471 netpoll_poll_disable(dev);
1473 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack);
1474 ret = notifier_to_errno(ret);
1478 set_bit(__LINK_STATE_START, &dev->state);
1480 if (ops->ndo_validate_addr)
1481 ret = ops->ndo_validate_addr(dev);
1483 if (!ret && ops->ndo_open)
1484 ret = ops->ndo_open(dev);
1486 netpoll_poll_enable(dev);
1489 clear_bit(__LINK_STATE_START, &dev->state);
1491 dev->flags |= IFF_UP;
1492 dev_set_rx_mode(dev);
1494 add_device_randomness(dev->dev_addr, dev->addr_len);
1501 * dev_open - prepare an interface for use.
1502 * @dev: device to open
1503 * @extack: netlink extended ack
1505 * Takes a device from down to up state. The device's private open
1506 * function is invoked and then the multicast lists are loaded. Finally
1507 * the device is moved into the up state and a %NETDEV_UP message is
1508 * sent to the netdev notifier chain.
1510 * Calling this function on an active interface is a nop. On a failure
1511 * a negative errno code is returned.
1513 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1517 if (dev->flags & IFF_UP)
1520 ret = __dev_open(dev, extack);
1524 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1525 call_netdevice_notifiers(NETDEV_UP, dev);
1529 EXPORT_SYMBOL(dev_open);
1531 static void __dev_close_many(struct list_head *head)
1533 struct net_device *dev;
1538 list_for_each_entry(dev, head, close_list) {
1539 /* Temporarily disable netpoll until the interface is down */
1540 netpoll_poll_disable(dev);
1542 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1544 clear_bit(__LINK_STATE_START, &dev->state);
1546 /* Synchronize to scheduled poll. We cannot touch poll list, it
1547 * can be even on different cpu. So just clear netif_running().
1549 * dev->stop() will invoke napi_disable() on all of it's
1550 * napi_struct instances on this device.
1552 smp_mb__after_atomic(); /* Commit netif_running(). */
1555 dev_deactivate_many(head);
1557 list_for_each_entry(dev, head, close_list) {
1558 const struct net_device_ops *ops = dev->netdev_ops;
1561 * Call the device specific close. This cannot fail.
1562 * Only if device is UP
1564 * We allow it to be called even after a DETACH hot-plug
1570 dev->flags &= ~IFF_UP;
1571 netpoll_poll_enable(dev);
1575 static void __dev_close(struct net_device *dev)
1579 list_add(&dev->close_list, &single);
1580 __dev_close_many(&single);
1584 void dev_close_many(struct list_head *head, bool unlink)
1586 struct net_device *dev, *tmp;
1588 /* Remove the devices that don't need to be closed */
1589 list_for_each_entry_safe(dev, tmp, head, close_list)
1590 if (!(dev->flags & IFF_UP))
1591 list_del_init(&dev->close_list);
1593 __dev_close_many(head);
1595 list_for_each_entry_safe(dev, tmp, head, close_list) {
1596 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1597 call_netdevice_notifiers(NETDEV_DOWN, dev);
1599 list_del_init(&dev->close_list);
1602 EXPORT_SYMBOL(dev_close_many);
1605 * dev_close - shutdown an interface.
1606 * @dev: device to shutdown
1608 * This function moves an active device into down state. A
1609 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1610 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1613 void dev_close(struct net_device *dev)
1615 if (dev->flags & IFF_UP) {
1618 list_add(&dev->close_list, &single);
1619 dev_close_many(&single, true);
1623 EXPORT_SYMBOL(dev_close);
1627 * dev_disable_lro - disable Large Receive Offload on a device
1630 * Disable Large Receive Offload (LRO) on a net device. Must be
1631 * called under RTNL. This is needed if received packets may be
1632 * forwarded to another interface.
1634 void dev_disable_lro(struct net_device *dev)
1636 struct net_device *lower_dev;
1637 struct list_head *iter;
1639 dev->wanted_features &= ~NETIF_F_LRO;
1640 netdev_update_features(dev);
1642 if (unlikely(dev->features & NETIF_F_LRO))
1643 netdev_WARN(dev, "failed to disable LRO!\n");
1645 netdev_for_each_lower_dev(dev, lower_dev, iter)
1646 dev_disable_lro(lower_dev);
1648 EXPORT_SYMBOL(dev_disable_lro);
1651 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1654 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1655 * called under RTNL. This is needed if Generic XDP is installed on
1658 static void dev_disable_gro_hw(struct net_device *dev)
1660 dev->wanted_features &= ~NETIF_F_GRO_HW;
1661 netdev_update_features(dev);
1663 if (unlikely(dev->features & NETIF_F_GRO_HW))
1664 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1667 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1670 case NETDEV_##val: \
1671 return "NETDEV_" __stringify(val);
1673 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1674 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1675 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1676 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1677 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1678 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1679 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1680 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1681 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1685 return "UNKNOWN_NETDEV_EVENT";
1687 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1689 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1690 struct net_device *dev)
1692 struct netdev_notifier_info info = {
1696 return nb->notifier_call(nb, val, &info);
1699 static int call_netdevice_register_notifiers(struct notifier_block *nb,
1700 struct net_device *dev)
1704 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1705 err = notifier_to_errno(err);
1709 if (!(dev->flags & IFF_UP))
1712 call_netdevice_notifier(nb, NETDEV_UP, dev);
1716 static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
1717 struct net_device *dev)
1719 if (dev->flags & IFF_UP) {
1720 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1722 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1724 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1727 static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
1730 struct net_device *dev;
1733 for_each_netdev(net, dev) {
1734 err = call_netdevice_register_notifiers(nb, dev);
1741 for_each_netdev_continue_reverse(net, dev)
1742 call_netdevice_unregister_notifiers(nb, dev);
1746 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb,
1749 struct net_device *dev;
1751 for_each_netdev(net, dev)
1752 call_netdevice_unregister_notifiers(nb, dev);
1755 static int dev_boot_phase = 1;
1758 * register_netdevice_notifier - register a network notifier block
1761 * Register a notifier to be called when network device events occur.
1762 * The notifier passed is linked into the kernel structures and must
1763 * not be reused until it has been unregistered. A negative errno code
1764 * is returned on a failure.
1766 * When registered all registration and up events are replayed
1767 * to the new notifier to allow device to have a race free
1768 * view of the network device list.
1771 int register_netdevice_notifier(struct notifier_block *nb)
1776 /* Close race with setup_net() and cleanup_net() */
1777 down_write(&pernet_ops_rwsem);
1779 err = raw_notifier_chain_register(&netdev_chain, nb);
1785 err = call_netdevice_register_net_notifiers(nb, net);
1792 up_write(&pernet_ops_rwsem);
1796 for_each_net_continue_reverse(net)
1797 call_netdevice_unregister_net_notifiers(nb, net);
1799 raw_notifier_chain_unregister(&netdev_chain, nb);
1802 EXPORT_SYMBOL(register_netdevice_notifier);
1805 * unregister_netdevice_notifier - unregister a network notifier block
1808 * Unregister a notifier previously registered by
1809 * register_netdevice_notifier(). The notifier is unlinked into the
1810 * kernel structures and may then be reused. A negative errno code
1811 * is returned on a failure.
1813 * After unregistering unregister and down device events are synthesized
1814 * for all devices on the device list to the removed notifier to remove
1815 * the need for special case cleanup code.
1818 int unregister_netdevice_notifier(struct notifier_block *nb)
1823 /* Close race with setup_net() and cleanup_net() */
1824 down_write(&pernet_ops_rwsem);
1826 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1831 call_netdevice_unregister_net_notifiers(nb, net);
1835 up_write(&pernet_ops_rwsem);
1838 EXPORT_SYMBOL(unregister_netdevice_notifier);
1840 static int __register_netdevice_notifier_net(struct net *net,
1841 struct notifier_block *nb,
1842 bool ignore_call_fail)
1846 err = raw_notifier_chain_register(&net->netdev_chain, nb);
1852 err = call_netdevice_register_net_notifiers(nb, net);
1853 if (err && !ignore_call_fail)
1854 goto chain_unregister;
1859 raw_notifier_chain_unregister(&net->netdev_chain, nb);
1863 static int __unregister_netdevice_notifier_net(struct net *net,
1864 struct notifier_block *nb)
1868 err = raw_notifier_chain_unregister(&net->netdev_chain, nb);
1872 call_netdevice_unregister_net_notifiers(nb, net);
1877 * register_netdevice_notifier_net - register a per-netns network notifier block
1878 * @net: network namespace
1881 * Register a notifier to be called when network device events occur.
1882 * The notifier passed is linked into the kernel structures and must
1883 * not be reused until it has been unregistered. A negative errno code
1884 * is returned on a failure.
1886 * When registered all registration and up events are replayed
1887 * to the new notifier to allow device to have a race free
1888 * view of the network device list.
1891 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb)
1896 err = __register_netdevice_notifier_net(net, nb, false);
1900 EXPORT_SYMBOL(register_netdevice_notifier_net);
1903 * unregister_netdevice_notifier_net - unregister a per-netns
1904 * network notifier block
1905 * @net: network namespace
1908 * Unregister a notifier previously registered by
1909 * register_netdevice_notifier(). The notifier is unlinked into the
1910 * kernel structures and may then be reused. A negative errno code
1911 * is returned on a failure.
1913 * After unregistering unregister and down device events are synthesized
1914 * for all devices on the device list to the removed notifier to remove
1915 * the need for special case cleanup code.
1918 int unregister_netdevice_notifier_net(struct net *net,
1919 struct notifier_block *nb)
1924 err = __unregister_netdevice_notifier_net(net, nb);
1928 EXPORT_SYMBOL(unregister_netdevice_notifier_net);
1930 int register_netdevice_notifier_dev_net(struct net_device *dev,
1931 struct notifier_block *nb,
1932 struct netdev_net_notifier *nn)
1937 err = __register_netdevice_notifier_net(dev_net(dev), nb, false);
1940 list_add(&nn->list, &dev->net_notifier_list);
1945 EXPORT_SYMBOL(register_netdevice_notifier_dev_net);
1947 int unregister_netdevice_notifier_dev_net(struct net_device *dev,
1948 struct notifier_block *nb,
1949 struct netdev_net_notifier *nn)
1954 list_del(&nn->list);
1955 err = __unregister_netdevice_notifier_net(dev_net(dev), nb);
1959 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net);
1961 static void move_netdevice_notifiers_dev_net(struct net_device *dev,
1964 struct netdev_net_notifier *nn;
1966 list_for_each_entry(nn, &dev->net_notifier_list, list) {
1967 __unregister_netdevice_notifier_net(dev_net(dev), nn->nb);
1968 __register_netdevice_notifier_net(net, nn->nb, true);
1973 * call_netdevice_notifiers_info - call all network notifier blocks
1974 * @val: value passed unmodified to notifier function
1975 * @info: notifier information data
1977 * Call all network notifier blocks. Parameters and return value
1978 * are as for raw_notifier_call_chain().
1981 static int call_netdevice_notifiers_info(unsigned long val,
1982 struct netdev_notifier_info *info)
1984 struct net *net = dev_net(info->dev);
1989 /* Run per-netns notifier block chain first, then run the global one.
1990 * Hopefully, one day, the global one is going to be removed after
1991 * all notifier block registrators get converted to be per-netns.
1993 ret = raw_notifier_call_chain(&net->netdev_chain, val, info);
1994 if (ret & NOTIFY_STOP_MASK)
1996 return raw_notifier_call_chain(&netdev_chain, val, info);
1999 static int call_netdevice_notifiers_extack(unsigned long val,
2000 struct net_device *dev,
2001 struct netlink_ext_ack *extack)
2003 struct netdev_notifier_info info = {
2008 return call_netdevice_notifiers_info(val, &info);
2012 * call_netdevice_notifiers - call all network notifier blocks
2013 * @val: value passed unmodified to notifier function
2014 * @dev: net_device pointer passed unmodified to notifier function
2016 * Call all network notifier blocks. Parameters and return value
2017 * are as for raw_notifier_call_chain().
2020 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
2022 return call_netdevice_notifiers_extack(val, dev, NULL);
2024 EXPORT_SYMBOL(call_netdevice_notifiers);
2027 * call_netdevice_notifiers_mtu - call all network notifier blocks
2028 * @val: value passed unmodified to notifier function
2029 * @dev: net_device pointer passed unmodified to notifier function
2030 * @arg: additional u32 argument passed to the notifier function
2032 * Call all network notifier blocks. Parameters and return value
2033 * are as for raw_notifier_call_chain().
2035 static int call_netdevice_notifiers_mtu(unsigned long val,
2036 struct net_device *dev, u32 arg)
2038 struct netdev_notifier_info_ext info = {
2043 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0);
2045 return call_netdevice_notifiers_info(val, &info.info);
2048 #ifdef CONFIG_NET_INGRESS
2049 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
2051 void net_inc_ingress_queue(void)
2053 static_branch_inc(&ingress_needed_key);
2055 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
2057 void net_dec_ingress_queue(void)
2059 static_branch_dec(&ingress_needed_key);
2061 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
2064 #ifdef CONFIG_NET_EGRESS
2065 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
2067 void net_inc_egress_queue(void)
2069 static_branch_inc(&egress_needed_key);
2071 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
2073 void net_dec_egress_queue(void)
2075 static_branch_dec(&egress_needed_key);
2077 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
2080 static DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
2081 #ifdef CONFIG_JUMP_LABEL
2082 static atomic_t netstamp_needed_deferred;
2083 static atomic_t netstamp_wanted;
2084 static void netstamp_clear(struct work_struct *work)
2086 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
2089 wanted = atomic_add_return(deferred, &netstamp_wanted);
2091 static_branch_enable(&netstamp_needed_key);
2093 static_branch_disable(&netstamp_needed_key);
2095 static DECLARE_WORK(netstamp_work, netstamp_clear);
2098 void net_enable_timestamp(void)
2100 #ifdef CONFIG_JUMP_LABEL
2104 wanted = atomic_read(&netstamp_wanted);
2107 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
2110 atomic_inc(&netstamp_needed_deferred);
2111 schedule_work(&netstamp_work);
2113 static_branch_inc(&netstamp_needed_key);
2116 EXPORT_SYMBOL(net_enable_timestamp);
2118 void net_disable_timestamp(void)
2120 #ifdef CONFIG_JUMP_LABEL
2124 wanted = atomic_read(&netstamp_wanted);
2127 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
2130 atomic_dec(&netstamp_needed_deferred);
2131 schedule_work(&netstamp_work);
2133 static_branch_dec(&netstamp_needed_key);
2136 EXPORT_SYMBOL(net_disable_timestamp);
2138 static inline void net_timestamp_set(struct sk_buff *skb)
2141 if (static_branch_unlikely(&netstamp_needed_key))
2142 __net_timestamp(skb);
2145 #define net_timestamp_check(COND, SKB) \
2146 if (static_branch_unlikely(&netstamp_needed_key)) { \
2147 if ((COND) && !(SKB)->tstamp) \
2148 __net_timestamp(SKB); \
2151 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
2153 return __is_skb_forwardable(dev, skb, true);
2155 EXPORT_SYMBOL_GPL(is_skb_forwardable);
2157 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb,
2160 int ret = ____dev_forward_skb(dev, skb, check_mtu);
2163 skb->protocol = eth_type_trans(skb, dev);
2164 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
2170 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2172 return __dev_forward_skb2(dev, skb, true);
2174 EXPORT_SYMBOL_GPL(__dev_forward_skb);
2177 * dev_forward_skb - loopback an skb to another netif
2179 * @dev: destination network device
2180 * @skb: buffer to forward
2183 * NET_RX_SUCCESS (no congestion)
2184 * NET_RX_DROP (packet was dropped, but freed)
2186 * dev_forward_skb can be used for injecting an skb from the
2187 * start_xmit function of one device into the receive queue
2188 * of another device.
2190 * The receiving device may be in another namespace, so
2191 * we have to clear all information in the skb that could
2192 * impact namespace isolation.
2194 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2196 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
2198 EXPORT_SYMBOL_GPL(dev_forward_skb);
2200 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb)
2202 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb);
2205 static inline int deliver_skb(struct sk_buff *skb,
2206 struct packet_type *pt_prev,
2207 struct net_device *orig_dev)
2209 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
2211 refcount_inc(&skb->users);
2212 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
2215 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
2216 struct packet_type **pt,
2217 struct net_device *orig_dev,
2219 struct list_head *ptype_list)
2221 struct packet_type *ptype, *pt_prev = *pt;
2223 list_for_each_entry_rcu(ptype, ptype_list, list) {
2224 if (ptype->type != type)
2227 deliver_skb(skb, pt_prev, orig_dev);
2233 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
2235 if (!ptype->af_packet_priv || !skb->sk)
2238 if (ptype->id_match)
2239 return ptype->id_match(ptype, skb->sk);
2240 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
2247 * dev_nit_active - return true if any network interface taps are in use
2249 * @dev: network device to check for the presence of taps
2251 bool dev_nit_active(struct net_device *dev)
2253 return !list_empty(&ptype_all) || !list_empty(&dev->ptype_all);
2255 EXPORT_SYMBOL_GPL(dev_nit_active);
2258 * Support routine. Sends outgoing frames to any network
2259 * taps currently in use.
2262 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
2264 struct packet_type *ptype;
2265 struct sk_buff *skb2 = NULL;
2266 struct packet_type *pt_prev = NULL;
2267 struct list_head *ptype_list = &ptype_all;
2271 list_for_each_entry_rcu(ptype, ptype_list, list) {
2272 if (ptype->ignore_outgoing)
2275 /* Never send packets back to the socket
2276 * they originated from - MvS (miquels@drinkel.ow.org)
2278 if (skb_loop_sk(ptype, skb))
2282 deliver_skb(skb2, pt_prev, skb->dev);
2287 /* need to clone skb, done only once */
2288 skb2 = skb_clone(skb, GFP_ATOMIC);
2292 net_timestamp_set(skb2);
2294 /* skb->nh should be correctly
2295 * set by sender, so that the second statement is
2296 * just protection against buggy protocols.
2298 skb_reset_mac_header(skb2);
2300 if (skb_network_header(skb2) < skb2->data ||
2301 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2302 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2303 ntohs(skb2->protocol),
2305 skb_reset_network_header(skb2);
2308 skb2->transport_header = skb2->network_header;
2309 skb2->pkt_type = PACKET_OUTGOING;
2313 if (ptype_list == &ptype_all) {
2314 ptype_list = &dev->ptype_all;
2319 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2320 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2326 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2329 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2330 * @dev: Network device
2331 * @txq: number of queues available
2333 * If real_num_tx_queues is changed the tc mappings may no longer be
2334 * valid. To resolve this verify the tc mapping remains valid and if
2335 * not NULL the mapping. With no priorities mapping to this
2336 * offset/count pair it will no longer be used. In the worst case TC0
2337 * is invalid nothing can be done so disable priority mappings. If is
2338 * expected that drivers will fix this mapping if they can before
2339 * calling netif_set_real_num_tx_queues.
2341 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2344 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2346 /* If TC0 is invalidated disable TC mapping */
2347 if (tc->offset + tc->count > txq) {
2348 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2353 /* Invalidated prio to tc mappings set to TC0 */
2354 for (i = 1; i < TC_BITMASK + 1; i++) {
2355 int q = netdev_get_prio_tc_map(dev, i);
2357 tc = &dev->tc_to_txq[q];
2358 if (tc->offset + tc->count > txq) {
2359 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2361 netdev_set_prio_tc_map(dev, i, 0);
2366 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2369 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2372 /* walk through the TCs and see if it falls into any of them */
2373 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2374 if ((txq - tc->offset) < tc->count)
2378 /* didn't find it, just return -1 to indicate no match */
2384 EXPORT_SYMBOL(netdev_txq_to_tc);
2387 static struct static_key xps_needed __read_mostly;
2388 static struct static_key xps_rxqs_needed __read_mostly;
2389 static DEFINE_MUTEX(xps_map_mutex);
2390 #define xmap_dereference(P) \
2391 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2393 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2394 struct xps_dev_maps *old_maps, int tci, u16 index)
2396 struct xps_map *map = NULL;
2400 map = xmap_dereference(dev_maps->attr_map[tci]);
2404 for (pos = map->len; pos--;) {
2405 if (map->queues[pos] != index)
2409 map->queues[pos] = map->queues[--map->len];
2414 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL);
2415 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2416 kfree_rcu(map, rcu);
2423 static bool remove_xps_queue_cpu(struct net_device *dev,
2424 struct xps_dev_maps *dev_maps,
2425 int cpu, u16 offset, u16 count)
2427 int num_tc = dev_maps->num_tc;
2428 bool active = false;
2431 for (tci = cpu * num_tc; num_tc--; tci++) {
2434 for (i = count, j = offset; i--; j++) {
2435 if (!remove_xps_queue(dev_maps, NULL, tci, j))
2445 static void reset_xps_maps(struct net_device *dev,
2446 struct xps_dev_maps *dev_maps,
2447 enum xps_map_type type)
2449 static_key_slow_dec_cpuslocked(&xps_needed);
2450 if (type == XPS_RXQS)
2451 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2453 RCU_INIT_POINTER(dev->xps_maps[type], NULL);
2455 kfree_rcu(dev_maps, rcu);
2458 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type,
2459 u16 offset, u16 count)
2461 struct xps_dev_maps *dev_maps;
2462 bool active = false;
2465 dev_maps = xmap_dereference(dev->xps_maps[type]);
2469 for (j = 0; j < dev_maps->nr_ids; j++)
2470 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count);
2472 reset_xps_maps(dev, dev_maps, type);
2474 if (type == XPS_CPUS) {
2475 for (i = offset + (count - 1); count--; i--)
2476 netdev_queue_numa_node_write(
2477 netdev_get_tx_queue(dev, i), NUMA_NO_NODE);
2481 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2484 if (!static_key_false(&xps_needed))
2488 mutex_lock(&xps_map_mutex);
2490 if (static_key_false(&xps_rxqs_needed))
2491 clean_xps_maps(dev, XPS_RXQS, offset, count);
2493 clean_xps_maps(dev, XPS_CPUS, offset, count);
2495 mutex_unlock(&xps_map_mutex);
2499 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2501 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2504 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2505 u16 index, bool is_rxqs_map)
2507 struct xps_map *new_map;
2508 int alloc_len = XPS_MIN_MAP_ALLOC;
2511 for (pos = 0; map && pos < map->len; pos++) {
2512 if (map->queues[pos] != index)
2517 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2519 if (pos < map->alloc_len)
2522 alloc_len = map->alloc_len * 2;
2525 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2529 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2531 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2532 cpu_to_node(attr_index));
2536 for (i = 0; i < pos; i++)
2537 new_map->queues[i] = map->queues[i];
2538 new_map->alloc_len = alloc_len;
2544 /* Copy xps maps at a given index */
2545 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps,
2546 struct xps_dev_maps *new_dev_maps, int index,
2547 int tc, bool skip_tc)
2549 int i, tci = index * dev_maps->num_tc;
2550 struct xps_map *map;
2552 /* copy maps belonging to foreign traffic classes */
2553 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2554 if (i == tc && skip_tc)
2557 /* fill in the new device map from the old device map */
2558 map = xmap_dereference(dev_maps->attr_map[tci]);
2559 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2563 /* Must be called under cpus_read_lock */
2564 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2565 u16 index, enum xps_map_type type)
2567 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL;
2568 const unsigned long *online_mask = NULL;
2569 bool active = false, copy = false;
2570 int i, j, tci, numa_node_id = -2;
2571 int maps_sz, num_tc = 1, tc = 0;
2572 struct xps_map *map, *new_map;
2573 unsigned int nr_ids;
2576 /* Do not allow XPS on subordinate device directly */
2577 num_tc = dev->num_tc;
2581 /* If queue belongs to subordinate dev use its map */
2582 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2584 tc = netdev_txq_to_tc(dev, index);
2589 mutex_lock(&xps_map_mutex);
2591 dev_maps = xmap_dereference(dev->xps_maps[type]);
2592 if (type == XPS_RXQS) {
2593 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2594 nr_ids = dev->num_rx_queues;
2596 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2597 if (num_possible_cpus() > 1)
2598 online_mask = cpumask_bits(cpu_online_mask);
2599 nr_ids = nr_cpu_ids;
2602 if (maps_sz < L1_CACHE_BYTES)
2603 maps_sz = L1_CACHE_BYTES;
2605 /* The old dev_maps could be larger or smaller than the one we're
2606 * setting up now, as dev->num_tc or nr_ids could have been updated in
2607 * between. We could try to be smart, but let's be safe instead and only
2608 * copy foreign traffic classes if the two map sizes match.
2611 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids)
2614 /* allocate memory for queue storage */
2615 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2617 if (!new_dev_maps) {
2618 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2619 if (!new_dev_maps) {
2620 mutex_unlock(&xps_map_mutex);
2624 new_dev_maps->nr_ids = nr_ids;
2625 new_dev_maps->num_tc = num_tc;
2628 tci = j * num_tc + tc;
2629 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL;
2631 map = expand_xps_map(map, j, index, type == XPS_RXQS);
2635 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2639 goto out_no_new_maps;
2642 /* Increment static keys at most once per type */
2643 static_key_slow_inc_cpuslocked(&xps_needed);
2644 if (type == XPS_RXQS)
2645 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2648 for (j = 0; j < nr_ids; j++) {
2649 bool skip_tc = false;
2651 tci = j * num_tc + tc;
2652 if (netif_attr_test_mask(j, mask, nr_ids) &&
2653 netif_attr_test_online(j, online_mask, nr_ids)) {
2654 /* add tx-queue to CPU/rx-queue maps */
2659 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2660 while ((pos < map->len) && (map->queues[pos] != index))
2663 if (pos == map->len)
2664 map->queues[map->len++] = index;
2666 if (type == XPS_CPUS) {
2667 if (numa_node_id == -2)
2668 numa_node_id = cpu_to_node(j);
2669 else if (numa_node_id != cpu_to_node(j))
2676 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc,
2680 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps);
2682 /* Cleanup old maps */
2684 goto out_no_old_maps;
2686 for (j = 0; j < dev_maps->nr_ids; j++) {
2687 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) {
2688 map = xmap_dereference(dev_maps->attr_map[tci]);
2693 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2698 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2699 kfree_rcu(map, rcu);
2703 old_dev_maps = dev_maps;
2706 dev_maps = new_dev_maps;
2710 if (type == XPS_CPUS)
2711 /* update Tx queue numa node */
2712 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2713 (numa_node_id >= 0) ?
2714 numa_node_id : NUMA_NO_NODE);
2719 /* removes tx-queue from unused CPUs/rx-queues */
2720 for (j = 0; j < dev_maps->nr_ids; j++) {
2721 tci = j * dev_maps->num_tc;
2723 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2725 netif_attr_test_mask(j, mask, dev_maps->nr_ids) &&
2726 netif_attr_test_online(j, online_mask, dev_maps->nr_ids))
2729 active |= remove_xps_queue(dev_maps,
2730 copy ? old_dev_maps : NULL,
2736 kfree_rcu(old_dev_maps, rcu);
2738 /* free map if not active */
2740 reset_xps_maps(dev, dev_maps, type);
2743 mutex_unlock(&xps_map_mutex);
2747 /* remove any maps that we added */
2748 for (j = 0; j < nr_ids; j++) {
2749 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2750 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2752 xmap_dereference(dev_maps->attr_map[tci]) :
2754 if (new_map && new_map != map)
2759 mutex_unlock(&xps_map_mutex);
2761 kfree(new_dev_maps);
2764 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2766 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2772 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS);
2777 EXPORT_SYMBOL(netif_set_xps_queue);
2780 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2782 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2784 /* Unbind any subordinate channels */
2785 while (txq-- != &dev->_tx[0]) {
2787 netdev_unbind_sb_channel(dev, txq->sb_dev);
2791 void netdev_reset_tc(struct net_device *dev)
2794 netif_reset_xps_queues_gt(dev, 0);
2796 netdev_unbind_all_sb_channels(dev);
2798 /* Reset TC configuration of device */
2800 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2801 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2803 EXPORT_SYMBOL(netdev_reset_tc);
2805 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2807 if (tc >= dev->num_tc)
2811 netif_reset_xps_queues(dev, offset, count);
2813 dev->tc_to_txq[tc].count = count;
2814 dev->tc_to_txq[tc].offset = offset;
2817 EXPORT_SYMBOL(netdev_set_tc_queue);
2819 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2821 if (num_tc > TC_MAX_QUEUE)
2825 netif_reset_xps_queues_gt(dev, 0);
2827 netdev_unbind_all_sb_channels(dev);
2829 dev->num_tc = num_tc;
2832 EXPORT_SYMBOL(netdev_set_num_tc);
2834 void netdev_unbind_sb_channel(struct net_device *dev,
2835 struct net_device *sb_dev)
2837 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2840 netif_reset_xps_queues_gt(sb_dev, 0);
2842 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2843 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2845 while (txq-- != &dev->_tx[0]) {
2846 if (txq->sb_dev == sb_dev)
2850 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2852 int netdev_bind_sb_channel_queue(struct net_device *dev,
2853 struct net_device *sb_dev,
2854 u8 tc, u16 count, u16 offset)
2856 /* Make certain the sb_dev and dev are already configured */
2857 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2860 /* We cannot hand out queues we don't have */
2861 if ((offset + count) > dev->real_num_tx_queues)
2864 /* Record the mapping */
2865 sb_dev->tc_to_txq[tc].count = count;
2866 sb_dev->tc_to_txq[tc].offset = offset;
2868 /* Provide a way for Tx queue to find the tc_to_txq map or
2869 * XPS map for itself.
2872 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2876 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2878 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2880 /* Do not use a multiqueue device to represent a subordinate channel */
2881 if (netif_is_multiqueue(dev))
2884 /* We allow channels 1 - 32767 to be used for subordinate channels.
2885 * Channel 0 is meant to be "native" mode and used only to represent
2886 * the main root device. We allow writing 0 to reset the device back
2887 * to normal mode after being used as a subordinate channel.
2889 if (channel > S16_MAX)
2892 dev->num_tc = -channel;
2896 EXPORT_SYMBOL(netdev_set_sb_channel);
2899 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2900 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2902 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2907 disabling = txq < dev->real_num_tx_queues;
2909 if (txq < 1 || txq > dev->num_tx_queues)
2912 if (dev->reg_state == NETREG_REGISTERED ||
2913 dev->reg_state == NETREG_UNREGISTERING) {
2916 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2922 netif_setup_tc(dev, txq);
2924 dev_qdisc_change_real_num_tx(dev, txq);
2926 dev->real_num_tx_queues = txq;
2930 qdisc_reset_all_tx_gt(dev, txq);
2932 netif_reset_xps_queues_gt(dev, txq);
2936 dev->real_num_tx_queues = txq;
2941 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2945 * netif_set_real_num_rx_queues - set actual number of RX queues used
2946 * @dev: Network device
2947 * @rxq: Actual number of RX queues
2949 * This must be called either with the rtnl_lock held or before
2950 * registration of the net device. Returns 0 on success, or a
2951 * negative error code. If called before registration, it always
2954 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2958 if (rxq < 1 || rxq > dev->num_rx_queues)
2961 if (dev->reg_state == NETREG_REGISTERED) {
2964 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2970 dev->real_num_rx_queues = rxq;
2973 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2977 * netif_set_real_num_queues - set actual number of RX and TX queues used
2978 * @dev: Network device
2979 * @txq: Actual number of TX queues
2980 * @rxq: Actual number of RX queues
2982 * Set the real number of both TX and RX queues.
2983 * Does nothing if the number of queues is already correct.
2985 int netif_set_real_num_queues(struct net_device *dev,
2986 unsigned int txq, unsigned int rxq)
2988 unsigned int old_rxq = dev->real_num_rx_queues;
2991 if (txq < 1 || txq > dev->num_tx_queues ||
2992 rxq < 1 || rxq > dev->num_rx_queues)
2995 /* Start from increases, so the error path only does decreases -
2996 * decreases can't fail.
2998 if (rxq > dev->real_num_rx_queues) {
2999 err = netif_set_real_num_rx_queues(dev, rxq);
3003 if (txq > dev->real_num_tx_queues) {
3004 err = netif_set_real_num_tx_queues(dev, txq);
3008 if (rxq < dev->real_num_rx_queues)
3009 WARN_ON(netif_set_real_num_rx_queues(dev, rxq));
3010 if (txq < dev->real_num_tx_queues)
3011 WARN_ON(netif_set_real_num_tx_queues(dev, txq));
3015 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq));
3018 EXPORT_SYMBOL(netif_set_real_num_queues);
3021 * netif_get_num_default_rss_queues - default number of RSS queues
3023 * This routine should set an upper limit on the number of RSS queues
3024 * used by default by multiqueue devices.
3026 int netif_get_num_default_rss_queues(void)
3028 return is_kdump_kernel() ?
3029 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
3031 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
3033 static void __netif_reschedule(struct Qdisc *q)
3035 struct softnet_data *sd;
3036 unsigned long flags;
3038 local_irq_save(flags);
3039 sd = this_cpu_ptr(&softnet_data);
3040 q->next_sched = NULL;
3041 *sd->output_queue_tailp = q;
3042 sd->output_queue_tailp = &q->next_sched;
3043 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3044 local_irq_restore(flags);
3047 void __netif_schedule(struct Qdisc *q)
3049 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
3050 __netif_reschedule(q);
3052 EXPORT_SYMBOL(__netif_schedule);
3054 struct dev_kfree_skb_cb {
3055 enum skb_free_reason reason;
3058 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
3060 return (struct dev_kfree_skb_cb *)skb->cb;
3063 void netif_schedule_queue(struct netdev_queue *txq)
3066 if (!netif_xmit_stopped(txq)) {
3067 struct Qdisc *q = rcu_dereference(txq->qdisc);
3069 __netif_schedule(q);
3073 EXPORT_SYMBOL(netif_schedule_queue);
3075 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
3077 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
3081 q = rcu_dereference(dev_queue->qdisc);
3082 __netif_schedule(q);
3086 EXPORT_SYMBOL(netif_tx_wake_queue);
3088 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
3090 unsigned long flags;
3095 if (likely(refcount_read(&skb->users) == 1)) {
3097 refcount_set(&skb->users, 0);
3098 } else if (likely(!refcount_dec_and_test(&skb->users))) {
3101 get_kfree_skb_cb(skb)->reason = reason;
3102 local_irq_save(flags);
3103 skb->next = __this_cpu_read(softnet_data.completion_queue);
3104 __this_cpu_write(softnet_data.completion_queue, skb);
3105 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3106 local_irq_restore(flags);
3108 EXPORT_SYMBOL(__dev_kfree_skb_irq);
3110 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
3112 if (in_hardirq() || irqs_disabled())
3113 __dev_kfree_skb_irq(skb, reason);
3117 EXPORT_SYMBOL(__dev_kfree_skb_any);
3121 * netif_device_detach - mark device as removed
3122 * @dev: network device
3124 * Mark device as removed from system and therefore no longer available.
3126 void netif_device_detach(struct net_device *dev)
3128 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
3129 netif_running(dev)) {
3130 netif_tx_stop_all_queues(dev);
3133 EXPORT_SYMBOL(netif_device_detach);
3136 * netif_device_attach - mark device as attached
3137 * @dev: network device
3139 * Mark device as attached from system and restart if needed.
3141 void netif_device_attach(struct net_device *dev)
3143 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
3144 netif_running(dev)) {
3145 netif_tx_wake_all_queues(dev);
3146 __netdev_watchdog_up(dev);
3149 EXPORT_SYMBOL(netif_device_attach);
3152 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
3153 * to be used as a distribution range.
3155 static u16 skb_tx_hash(const struct net_device *dev,
3156 const struct net_device *sb_dev,
3157 struct sk_buff *skb)
3161 u16 qcount = dev->real_num_tx_queues;
3164 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
3166 qoffset = sb_dev->tc_to_txq[tc].offset;
3167 qcount = sb_dev->tc_to_txq[tc].count;
3168 if (unlikely(!qcount)) {
3169 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n",
3170 sb_dev->name, qoffset, tc);
3172 qcount = dev->real_num_tx_queues;
3176 if (skb_rx_queue_recorded(skb)) {
3177 hash = skb_get_rx_queue(skb);
3178 if (hash >= qoffset)
3180 while (unlikely(hash >= qcount))
3182 return hash + qoffset;
3185 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3188 static void skb_warn_bad_offload(const struct sk_buff *skb)
3190 static const netdev_features_t null_features;
3191 struct net_device *dev = skb->dev;
3192 const char *name = "";
3194 if (!net_ratelimit())
3198 if (dev->dev.parent)
3199 name = dev_driver_string(dev->dev.parent);
3201 name = netdev_name(dev);
3203 skb_dump(KERN_WARNING, skb, false);
3204 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3205 name, dev ? &dev->features : &null_features,
3206 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3210 * Invalidate hardware checksum when packet is to be mangled, and
3211 * complete checksum manually on outgoing path.
3213 int skb_checksum_help(struct sk_buff *skb)
3216 int ret = 0, offset;
3218 if (skb->ip_summed == CHECKSUM_COMPLETE)
3219 goto out_set_summed;
3221 if (unlikely(skb_is_gso(skb))) {
3222 skb_warn_bad_offload(skb);
3226 /* Before computing a checksum, we should make sure no frag could
3227 * be modified by an external entity : checksum could be wrong.
3229 if (skb_has_shared_frag(skb)) {
3230 ret = __skb_linearize(skb);
3235 offset = skb_checksum_start_offset(skb);
3236 BUG_ON(offset >= skb_headlen(skb));
3237 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3239 offset += skb->csum_offset;
3240 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
3242 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3246 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3248 skb->ip_summed = CHECKSUM_NONE;
3252 EXPORT_SYMBOL(skb_checksum_help);
3254 int skb_crc32c_csum_help(struct sk_buff *skb)
3257 int ret = 0, offset, start;
3259 if (skb->ip_summed != CHECKSUM_PARTIAL)
3262 if (unlikely(skb_is_gso(skb)))
3265 /* Before computing a checksum, we should make sure no frag could
3266 * be modified by an external entity : checksum could be wrong.
3268 if (unlikely(skb_has_shared_frag(skb))) {
3269 ret = __skb_linearize(skb);
3273 start = skb_checksum_start_offset(skb);
3274 offset = start + offsetof(struct sctphdr, checksum);
3275 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3280 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3284 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3285 skb->len - start, ~(__u32)0,
3287 *(__le32 *)(skb->data + offset) = crc32c_csum;
3288 skb->ip_summed = CHECKSUM_NONE;
3289 skb->csum_not_inet = 0;
3294 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3296 __be16 type = skb->protocol;
3298 /* Tunnel gso handlers can set protocol to ethernet. */
3299 if (type == htons(ETH_P_TEB)) {
3302 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3305 eth = (struct ethhdr *)skb->data;
3306 type = eth->h_proto;
3309 return __vlan_get_protocol(skb, type, depth);
3313 * skb_mac_gso_segment - mac layer segmentation handler.
3314 * @skb: buffer to segment
3315 * @features: features for the output path (see dev->features)
3317 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
3318 netdev_features_t features)
3320 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
3321 struct packet_offload *ptype;
3322 int vlan_depth = skb->mac_len;
3323 __be16 type = skb_network_protocol(skb, &vlan_depth);
3325 if (unlikely(!type))
3326 return ERR_PTR(-EINVAL);
3328 __skb_pull(skb, vlan_depth);
3331 list_for_each_entry_rcu(ptype, &offload_base, list) {
3332 if (ptype->type == type && ptype->callbacks.gso_segment) {
3333 segs = ptype->callbacks.gso_segment(skb, features);
3339 __skb_push(skb, skb->data - skb_mac_header(skb));
3343 EXPORT_SYMBOL(skb_mac_gso_segment);
3346 /* openvswitch calls this on rx path, so we need a different check.
3348 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
3351 return skb->ip_summed != CHECKSUM_PARTIAL &&
3352 skb->ip_summed != CHECKSUM_UNNECESSARY;
3354 return skb->ip_summed == CHECKSUM_NONE;
3358 * __skb_gso_segment - Perform segmentation on skb.
3359 * @skb: buffer to segment
3360 * @features: features for the output path (see dev->features)
3361 * @tx_path: whether it is called in TX path
3363 * This function segments the given skb and returns a list of segments.
3365 * It may return NULL if the skb requires no segmentation. This is
3366 * only possible when GSO is used for verifying header integrity.
3368 * Segmentation preserves SKB_GSO_CB_OFFSET bytes of previous skb cb.
3370 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3371 netdev_features_t features, bool tx_path)
3373 struct sk_buff *segs;
3375 if (unlikely(skb_needs_check(skb, tx_path))) {
3378 /* We're going to init ->check field in TCP or UDP header */
3379 err = skb_cow_head(skb, 0);
3381 return ERR_PTR(err);
3384 /* Only report GSO partial support if it will enable us to
3385 * support segmentation on this frame without needing additional
3388 if (features & NETIF_F_GSO_PARTIAL) {
3389 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3390 struct net_device *dev = skb->dev;
3392 partial_features |= dev->features & dev->gso_partial_features;
3393 if (!skb_gso_ok(skb, features | partial_features))
3394 features &= ~NETIF_F_GSO_PARTIAL;
3397 BUILD_BUG_ON(SKB_GSO_CB_OFFSET +
3398 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3400 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3401 SKB_GSO_CB(skb)->encap_level = 0;
3403 skb_reset_mac_header(skb);
3404 skb_reset_mac_len(skb);
3406 segs = skb_mac_gso_segment(skb, features);
3408 if (segs != skb && unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3409 skb_warn_bad_offload(skb);
3413 EXPORT_SYMBOL(__skb_gso_segment);
3415 /* Take action when hardware reception checksum errors are detected. */
3417 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3419 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
3420 skb_dump(KERN_ERR, skb, true);
3424 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3426 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3428 EXPORT_SYMBOL(netdev_rx_csum_fault);
3431 /* XXX: check that highmem exists at all on the given machine. */
3432 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3434 #ifdef CONFIG_HIGHMEM
3437 if (!(dev->features & NETIF_F_HIGHDMA)) {
3438 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3439 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3441 if (PageHighMem(skb_frag_page(frag)))
3449 /* If MPLS offload request, verify we are testing hardware MPLS features
3450 * instead of standard features for the netdev.
3452 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3453 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3454 netdev_features_t features,
3457 if (eth_p_mpls(type))
3458 features &= skb->dev->mpls_features;
3463 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3464 netdev_features_t features,
3471 static netdev_features_t harmonize_features(struct sk_buff *skb,
3472 netdev_features_t features)
3476 type = skb_network_protocol(skb, NULL);
3477 features = net_mpls_features(skb, features, type);
3479 if (skb->ip_summed != CHECKSUM_NONE &&
3480 !can_checksum_protocol(features, type)) {
3481 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3483 if (illegal_highdma(skb->dev, skb))
3484 features &= ~NETIF_F_SG;
3489 netdev_features_t passthru_features_check(struct sk_buff *skb,
3490 struct net_device *dev,
3491 netdev_features_t features)
3495 EXPORT_SYMBOL(passthru_features_check);
3497 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3498 struct net_device *dev,
3499 netdev_features_t features)
3501 return vlan_features_check(skb, features);
3504 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3505 struct net_device *dev,
3506 netdev_features_t features)
3508 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3510 if (gso_segs > dev->gso_max_segs)
3511 return features & ~NETIF_F_GSO_MASK;
3513 if (!skb_shinfo(skb)->gso_type) {
3514 skb_warn_bad_offload(skb);
3515 return features & ~NETIF_F_GSO_MASK;
3518 /* Support for GSO partial features requires software
3519 * intervention before we can actually process the packets
3520 * so we need to strip support for any partial features now
3521 * and we can pull them back in after we have partially
3522 * segmented the frame.
3524 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3525 features &= ~dev->gso_partial_features;
3527 /* Make sure to clear the IPv4 ID mangling feature if the
3528 * IPv4 header has the potential to be fragmented.
3530 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3531 struct iphdr *iph = skb->encapsulation ?
3532 inner_ip_hdr(skb) : ip_hdr(skb);
3534 if (!(iph->frag_off & htons(IP_DF)))
3535 features &= ~NETIF_F_TSO_MANGLEID;
3541 netdev_features_t netif_skb_features(struct sk_buff *skb)
3543 struct net_device *dev = skb->dev;
3544 netdev_features_t features = dev->features;
3546 if (skb_is_gso(skb))
3547 features = gso_features_check(skb, dev, features);
3549 /* If encapsulation offload request, verify we are testing
3550 * hardware encapsulation features instead of standard
3551 * features for the netdev
3553 if (skb->encapsulation)
3554 features &= dev->hw_enc_features;
3556 if (skb_vlan_tagged(skb))
3557 features = netdev_intersect_features(features,
3558 dev->vlan_features |
3559 NETIF_F_HW_VLAN_CTAG_TX |
3560 NETIF_F_HW_VLAN_STAG_TX);
3562 if (dev->netdev_ops->ndo_features_check)
3563 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3566 features &= dflt_features_check(skb, dev, features);
3568 return harmonize_features(skb, features);
3570 EXPORT_SYMBOL(netif_skb_features);
3572 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3573 struct netdev_queue *txq, bool more)
3578 if (dev_nit_active(dev))
3579 dev_queue_xmit_nit(skb, dev);
3582 PRANDOM_ADD_NOISE(skb, dev, txq, len + jiffies);
3583 trace_net_dev_start_xmit(skb, dev);
3584 rc = netdev_start_xmit(skb, dev, txq, more);
3585 trace_net_dev_xmit(skb, rc, dev, len);
3590 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3591 struct netdev_queue *txq, int *ret)
3593 struct sk_buff *skb = first;
3594 int rc = NETDEV_TX_OK;
3597 struct sk_buff *next = skb->next;
3599 skb_mark_not_on_list(skb);
3600 rc = xmit_one(skb, dev, txq, next != NULL);
3601 if (unlikely(!dev_xmit_complete(rc))) {
3607 if (netif_tx_queue_stopped(txq) && skb) {
3608 rc = NETDEV_TX_BUSY;
3618 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3619 netdev_features_t features)
3621 if (skb_vlan_tag_present(skb) &&
3622 !vlan_hw_offload_capable(features, skb->vlan_proto))
3623 skb = __vlan_hwaccel_push_inside(skb);
3627 int skb_csum_hwoffload_help(struct sk_buff *skb,
3628 const netdev_features_t features)
3630 if (unlikely(skb_csum_is_sctp(skb)))
3631 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3632 skb_crc32c_csum_help(skb);
3634 if (features & NETIF_F_HW_CSUM)
3637 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3638 switch (skb->csum_offset) {
3639 case offsetof(struct tcphdr, check):
3640 case offsetof(struct udphdr, check):
3645 return skb_checksum_help(skb);
3647 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3649 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3651 netdev_features_t features;
3653 features = netif_skb_features(skb);
3654 skb = validate_xmit_vlan(skb, features);
3658 skb = sk_validate_xmit_skb(skb, dev);
3662 if (netif_needs_gso(skb, features)) {
3663 struct sk_buff *segs;
3665 segs = skb_gso_segment(skb, features);
3673 if (skb_needs_linearize(skb, features) &&
3674 __skb_linearize(skb))
3677 /* If packet is not checksummed and device does not
3678 * support checksumming for this protocol, complete
3679 * checksumming here.
3681 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3682 if (skb->encapsulation)
3683 skb_set_inner_transport_header(skb,
3684 skb_checksum_start_offset(skb));
3686 skb_set_transport_header(skb,
3687 skb_checksum_start_offset(skb));
3688 if (skb_csum_hwoffload_help(skb, features))
3693 skb = validate_xmit_xfrm(skb, features, again);
3700 atomic_long_inc(&dev->tx_dropped);
3704 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3706 struct sk_buff *next, *head = NULL, *tail;
3708 for (; skb != NULL; skb = next) {
3710 skb_mark_not_on_list(skb);
3712 /* in case skb wont be segmented, point to itself */
3715 skb = validate_xmit_skb(skb, dev, again);
3723 /* If skb was segmented, skb->prev points to
3724 * the last segment. If not, it still contains skb.
3730 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3732 static void qdisc_pkt_len_init(struct sk_buff *skb)
3734 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3736 qdisc_skb_cb(skb)->pkt_len = skb->len;
3738 /* To get more precise estimation of bytes sent on wire,
3739 * we add to pkt_len the headers size of all segments
3741 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3742 unsigned int hdr_len;
3743 u16 gso_segs = shinfo->gso_segs;
3745 /* mac layer + network layer */
3746 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3748 /* + transport layer */
3749 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3750 const struct tcphdr *th;
3751 struct tcphdr _tcphdr;
3753 th = skb_header_pointer(skb, skb_transport_offset(skb),
3754 sizeof(_tcphdr), &_tcphdr);
3756 hdr_len += __tcp_hdrlen(th);
3758 struct udphdr _udphdr;
3760 if (skb_header_pointer(skb, skb_transport_offset(skb),
3761 sizeof(_udphdr), &_udphdr))
3762 hdr_len += sizeof(struct udphdr);
3765 if (shinfo->gso_type & SKB_GSO_DODGY)
3766 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3769 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3773 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q,
3774 struct sk_buff **to_free,
3775 struct netdev_queue *txq)
3779 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK;
3780 if (rc == NET_XMIT_SUCCESS)
3781 trace_qdisc_enqueue(q, txq, skb);
3785 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3786 struct net_device *dev,
3787 struct netdev_queue *txq)
3789 spinlock_t *root_lock = qdisc_lock(q);
3790 struct sk_buff *to_free = NULL;
3794 qdisc_calculate_pkt_len(skb, q);
3796 if (q->flags & TCQ_F_NOLOCK) {
3797 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3798 qdisc_run_begin(q)) {
3799 /* Retest nolock_qdisc_is_empty() within the protection
3800 * of q->seqlock to protect from racing with requeuing.
3802 if (unlikely(!nolock_qdisc_is_empty(q))) {
3803 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3810 qdisc_bstats_cpu_update(q, skb);
3811 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3812 !nolock_qdisc_is_empty(q))
3816 return NET_XMIT_SUCCESS;
3819 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3823 if (unlikely(to_free))
3824 kfree_skb_list(to_free);
3829 * Heuristic to force contended enqueues to serialize on a
3830 * separate lock before trying to get qdisc main lock.
3831 * This permits qdisc->running owner to get the lock more
3832 * often and dequeue packets faster.
3834 contended = qdisc_is_running(q);
3835 if (unlikely(contended))
3836 spin_lock(&q->busylock);
3838 spin_lock(root_lock);
3839 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3840 __qdisc_drop(skb, &to_free);
3842 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3843 qdisc_run_begin(q)) {
3845 * This is a work-conserving queue; there are no old skbs
3846 * waiting to be sent out; and the qdisc is not running -
3847 * xmit the skb directly.
3850 qdisc_bstats_update(q, skb);
3852 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3853 if (unlikely(contended)) {
3854 spin_unlock(&q->busylock);
3861 rc = NET_XMIT_SUCCESS;
3863 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3864 if (qdisc_run_begin(q)) {
3865 if (unlikely(contended)) {
3866 spin_unlock(&q->busylock);
3873 spin_unlock(root_lock);
3874 if (unlikely(to_free))
3875 kfree_skb_list(to_free);
3876 if (unlikely(contended))
3877 spin_unlock(&q->busylock);
3881 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3882 static void skb_update_prio(struct sk_buff *skb)
3884 const struct netprio_map *map;
3885 const struct sock *sk;
3886 unsigned int prioidx;
3890 map = rcu_dereference_bh(skb->dev->priomap);
3893 sk = skb_to_full_sk(skb);
3897 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3899 if (prioidx < map->priomap_len)
3900 skb->priority = map->priomap[prioidx];
3903 #define skb_update_prio(skb)
3907 * dev_loopback_xmit - loop back @skb
3908 * @net: network namespace this loopback is happening in
3909 * @sk: sk needed to be a netfilter okfn
3910 * @skb: buffer to transmit
3912 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3914 skb_reset_mac_header(skb);
3915 __skb_pull(skb, skb_network_offset(skb));
3916 skb->pkt_type = PACKET_LOOPBACK;
3917 if (skb->ip_summed == CHECKSUM_NONE)
3918 skb->ip_summed = CHECKSUM_UNNECESSARY;
3919 WARN_ON(!skb_dst(skb));
3924 EXPORT_SYMBOL(dev_loopback_xmit);
3926 #ifdef CONFIG_NET_EGRESS
3927 static struct sk_buff *
3928 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3930 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3931 struct tcf_result cl_res;
3936 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3937 tc_skb_cb(skb)->mru = 0;
3938 tc_skb_cb(skb)->post_ct = false;
3939 mini_qdisc_bstats_cpu_update(miniq, skb);
3941 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
3943 case TC_ACT_RECLASSIFY:
3944 skb->tc_index = TC_H_MIN(cl_res.classid);
3947 mini_qdisc_qstats_cpu_drop(miniq);
3948 *ret = NET_XMIT_DROP;
3954 *ret = NET_XMIT_SUCCESS;
3957 case TC_ACT_REDIRECT:
3958 /* No need to push/pop skb's mac_header here on egress! */
3959 skb_do_redirect(skb);
3960 *ret = NET_XMIT_SUCCESS;
3968 #endif /* CONFIG_NET_EGRESS */
3971 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
3972 struct xps_dev_maps *dev_maps, unsigned int tci)
3974 int tc = netdev_get_prio_tc_map(dev, skb->priority);
3975 struct xps_map *map;
3976 int queue_index = -1;
3978 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
3981 tci *= dev_maps->num_tc;
3984 map = rcu_dereference(dev_maps->attr_map[tci]);
3987 queue_index = map->queues[0];
3989 queue_index = map->queues[reciprocal_scale(
3990 skb_get_hash(skb), map->len)];
3991 if (unlikely(queue_index >= dev->real_num_tx_queues))
3998 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
3999 struct sk_buff *skb)
4002 struct xps_dev_maps *dev_maps;
4003 struct sock *sk = skb->sk;
4004 int queue_index = -1;
4006 if (!static_key_false(&xps_needed))
4010 if (!static_key_false(&xps_rxqs_needed))
4013 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4015 int tci = sk_rx_queue_get(sk);
4018 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4023 if (queue_index < 0) {
4024 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4026 unsigned int tci = skb->sender_cpu - 1;
4028 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4040 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4041 struct net_device *sb_dev)
4045 EXPORT_SYMBOL(dev_pick_tx_zero);
4047 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4048 struct net_device *sb_dev)
4050 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4052 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4054 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4055 struct net_device *sb_dev)
4057 struct sock *sk = skb->sk;
4058 int queue_index = sk_tx_queue_get(sk);
4060 sb_dev = sb_dev ? : dev;
4062 if (queue_index < 0 || skb->ooo_okay ||
4063 queue_index >= dev->real_num_tx_queues) {
4064 int new_index = get_xps_queue(dev, sb_dev, skb);
4067 new_index = skb_tx_hash(dev, sb_dev, skb);
4069 if (queue_index != new_index && sk &&
4071 rcu_access_pointer(sk->sk_dst_cache))
4072 sk_tx_queue_set(sk, new_index);
4074 queue_index = new_index;
4079 EXPORT_SYMBOL(netdev_pick_tx);
4081 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4082 struct sk_buff *skb,
4083 struct net_device *sb_dev)
4085 int queue_index = 0;
4088 u32 sender_cpu = skb->sender_cpu - 1;
4090 if (sender_cpu >= (u32)NR_CPUS)
4091 skb->sender_cpu = raw_smp_processor_id() + 1;
4094 if (dev->real_num_tx_queues != 1) {
4095 const struct net_device_ops *ops = dev->netdev_ops;
4097 if (ops->ndo_select_queue)
4098 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4100 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4102 queue_index = netdev_cap_txqueue(dev, queue_index);
4105 skb_set_queue_mapping(skb, queue_index);
4106 return netdev_get_tx_queue(dev, queue_index);
4110 * __dev_queue_xmit - transmit a buffer
4111 * @skb: buffer to transmit
4112 * @sb_dev: suboordinate device used for L2 forwarding offload
4114 * Queue a buffer for transmission to a network device. The caller must
4115 * have set the device and priority and built the buffer before calling
4116 * this function. The function can be called from an interrupt.
4118 * A negative errno code is returned on a failure. A success does not
4119 * guarantee the frame will be transmitted as it may be dropped due
4120 * to congestion or traffic shaping.
4122 * -----------------------------------------------------------------------------------
4123 * I notice this method can also return errors from the queue disciplines,
4124 * including NET_XMIT_DROP, which is a positive value. So, errors can also
4127 * Regardless of the return value, the skb is consumed, so it is currently
4128 * difficult to retry a send to this method. (You can bump the ref count
4129 * before sending to hold a reference for retry if you are careful.)
4131 * When calling this method, interrupts MUST be enabled. This is because
4132 * the BH enable code must have IRQs enabled so that it will not deadlock.
4135 static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4137 struct net_device *dev = skb->dev;
4138 struct netdev_queue *txq;
4143 skb_reset_mac_header(skb);
4145 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4146 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4148 /* Disable soft irqs for various locks below. Also
4149 * stops preemption for RCU.
4153 skb_update_prio(skb);
4155 qdisc_pkt_len_init(skb);
4156 #ifdef CONFIG_NET_CLS_ACT
4157 skb->tc_at_ingress = 0;
4158 # ifdef CONFIG_NET_EGRESS
4159 if (static_branch_unlikely(&egress_needed_key)) {
4160 skb = sch_handle_egress(skb, &rc, dev);
4166 /* If device/qdisc don't need skb->dst, release it right now while
4167 * its hot in this cpu cache.
4169 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4174 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4175 q = rcu_dereference_bh(txq->qdisc);
4177 trace_net_dev_queue(skb);
4179 rc = __dev_xmit_skb(skb, q, dev, txq);
4183 /* The device has no queue. Common case for software devices:
4184 * loopback, all the sorts of tunnels...
4186 * Really, it is unlikely that netif_tx_lock protection is necessary
4187 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4189 * However, it is possible, that they rely on protection
4192 * Check this and shot the lock. It is not prone from deadlocks.
4193 *Either shot noqueue qdisc, it is even simpler 8)
4195 if (dev->flags & IFF_UP) {
4196 int cpu = smp_processor_id(); /* ok because BHs are off */
4198 /* Other cpus might concurrently change txq->xmit_lock_owner
4199 * to -1 or to their cpu id, but not to our id.
4201 if (READ_ONCE(txq->xmit_lock_owner) != cpu) {
4202 if (dev_xmit_recursion())
4203 goto recursion_alert;
4205 skb = validate_xmit_skb(skb, dev, &again);
4209 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4210 HARD_TX_LOCK(dev, txq, cpu);
4212 if (!netif_xmit_stopped(txq)) {
4213 dev_xmit_recursion_inc();
4214 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4215 dev_xmit_recursion_dec();
4216 if (dev_xmit_complete(rc)) {
4217 HARD_TX_UNLOCK(dev, txq);
4221 HARD_TX_UNLOCK(dev, txq);
4222 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4225 /* Recursion is detected! It is possible,
4229 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4235 rcu_read_unlock_bh();
4237 atomic_long_inc(&dev->tx_dropped);
4238 kfree_skb_list(skb);
4241 rcu_read_unlock_bh();
4245 int dev_queue_xmit(struct sk_buff *skb)
4247 return __dev_queue_xmit(skb, NULL);
4249 EXPORT_SYMBOL(dev_queue_xmit);
4251 int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
4253 return __dev_queue_xmit(skb, sb_dev);
4255 EXPORT_SYMBOL(dev_queue_xmit_accel);
4257 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4259 struct net_device *dev = skb->dev;
4260 struct sk_buff *orig_skb = skb;
4261 struct netdev_queue *txq;
4262 int ret = NETDEV_TX_BUSY;
4265 if (unlikely(!netif_running(dev) ||
4266 !netif_carrier_ok(dev)))
4269 skb = validate_xmit_skb_list(skb, dev, &again);
4270 if (skb != orig_skb)
4273 skb_set_queue_mapping(skb, queue_id);
4274 txq = skb_get_tx_queue(dev, skb);
4275 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4279 dev_xmit_recursion_inc();
4280 HARD_TX_LOCK(dev, txq, smp_processor_id());
4281 if (!netif_xmit_frozen_or_drv_stopped(txq))
4282 ret = netdev_start_xmit(skb, dev, txq, false);
4283 HARD_TX_UNLOCK(dev, txq);
4284 dev_xmit_recursion_dec();
4289 atomic_long_inc(&dev->tx_dropped);
4290 kfree_skb_list(skb);
4291 return NET_XMIT_DROP;
4293 EXPORT_SYMBOL(__dev_direct_xmit);
4295 /*************************************************************************
4297 *************************************************************************/
4299 int netdev_max_backlog __read_mostly = 1000;
4300 EXPORT_SYMBOL(netdev_max_backlog);
4302 int netdev_tstamp_prequeue __read_mostly = 1;
4303 int netdev_budget __read_mostly = 300;
4304 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4305 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4306 int weight_p __read_mostly = 64; /* old backlog weight */
4307 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4308 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4309 int dev_rx_weight __read_mostly = 64;
4310 int dev_tx_weight __read_mostly = 64;
4311 /* Maximum number of GRO_NORMAL skbs to batch up for list-RX */
4312 int gro_normal_batch __read_mostly = 8;
4314 /* Called with irq disabled */
4315 static inline void ____napi_schedule(struct softnet_data *sd,
4316 struct napi_struct *napi)
4318 struct task_struct *thread;
4320 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4321 /* Paired with smp_mb__before_atomic() in
4322 * napi_enable()/dev_set_threaded().
4323 * Use READ_ONCE() to guarantee a complete
4324 * read on napi->thread. Only call
4325 * wake_up_process() when it's not NULL.
4327 thread = READ_ONCE(napi->thread);
4329 /* Avoid doing set_bit() if the thread is in
4330 * INTERRUPTIBLE state, cause napi_thread_wait()
4331 * makes sure to proceed with napi polling
4332 * if the thread is explicitly woken from here.
4334 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4335 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4336 wake_up_process(thread);
4341 list_add_tail(&napi->poll_list, &sd->poll_list);
4342 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4347 /* One global table that all flow-based protocols share. */
4348 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4349 EXPORT_SYMBOL(rps_sock_flow_table);
4350 u32 rps_cpu_mask __read_mostly;
4351 EXPORT_SYMBOL(rps_cpu_mask);
4353 struct static_key_false rps_needed __read_mostly;
4354 EXPORT_SYMBOL(rps_needed);
4355 struct static_key_false rfs_needed __read_mostly;
4356 EXPORT_SYMBOL(rfs_needed);
4358 static struct rps_dev_flow *
4359 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4360 struct rps_dev_flow *rflow, u16 next_cpu)
4362 if (next_cpu < nr_cpu_ids) {
4363 #ifdef CONFIG_RFS_ACCEL
4364 struct netdev_rx_queue *rxqueue;
4365 struct rps_dev_flow_table *flow_table;
4366 struct rps_dev_flow *old_rflow;
4371 /* Should we steer this flow to a different hardware queue? */
4372 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4373 !(dev->features & NETIF_F_NTUPLE))
4375 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4376 if (rxq_index == skb_get_rx_queue(skb))
4379 rxqueue = dev->_rx + rxq_index;
4380 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4383 flow_id = skb_get_hash(skb) & flow_table->mask;
4384 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4385 rxq_index, flow_id);
4389 rflow = &flow_table->flows[flow_id];
4391 if (old_rflow->filter == rflow->filter)
4392 old_rflow->filter = RPS_NO_FILTER;
4396 per_cpu(softnet_data, next_cpu).input_queue_head;
4399 rflow->cpu = next_cpu;
4404 * get_rps_cpu is called from netif_receive_skb and returns the target
4405 * CPU from the RPS map of the receiving queue for a given skb.
4406 * rcu_read_lock must be held on entry.
4408 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4409 struct rps_dev_flow **rflowp)
4411 const struct rps_sock_flow_table *sock_flow_table;
4412 struct netdev_rx_queue *rxqueue = dev->_rx;
4413 struct rps_dev_flow_table *flow_table;
4414 struct rps_map *map;
4419 if (skb_rx_queue_recorded(skb)) {
4420 u16 index = skb_get_rx_queue(skb);
4422 if (unlikely(index >= dev->real_num_rx_queues)) {
4423 WARN_ONCE(dev->real_num_rx_queues > 1,
4424 "%s received packet on queue %u, but number "
4425 "of RX queues is %u\n",
4426 dev->name, index, dev->real_num_rx_queues);
4432 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4434 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4435 map = rcu_dereference(rxqueue->rps_map);
4436 if (!flow_table && !map)
4439 skb_reset_network_header(skb);
4440 hash = skb_get_hash(skb);
4444 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4445 if (flow_table && sock_flow_table) {
4446 struct rps_dev_flow *rflow;
4450 /* First check into global flow table if there is a match */
4451 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4452 if ((ident ^ hash) & ~rps_cpu_mask)
4455 next_cpu = ident & rps_cpu_mask;
4457 /* OK, now we know there is a match,
4458 * we can look at the local (per receive queue) flow table
4460 rflow = &flow_table->flows[hash & flow_table->mask];
4464 * If the desired CPU (where last recvmsg was done) is
4465 * different from current CPU (one in the rx-queue flow
4466 * table entry), switch if one of the following holds:
4467 * - Current CPU is unset (>= nr_cpu_ids).
4468 * - Current CPU is offline.
4469 * - The current CPU's queue tail has advanced beyond the
4470 * last packet that was enqueued using this table entry.
4471 * This guarantees that all previous packets for the flow
4472 * have been dequeued, thus preserving in order delivery.
4474 if (unlikely(tcpu != next_cpu) &&
4475 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4476 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4477 rflow->last_qtail)) >= 0)) {
4479 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4482 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4492 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4493 if (cpu_online(tcpu)) {
4503 #ifdef CONFIG_RFS_ACCEL
4506 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4507 * @dev: Device on which the filter was set
4508 * @rxq_index: RX queue index
4509 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4510 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4512 * Drivers that implement ndo_rx_flow_steer() should periodically call
4513 * this function for each installed filter and remove the filters for
4514 * which it returns %true.
4516 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4517 u32 flow_id, u16 filter_id)
4519 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4520 struct rps_dev_flow_table *flow_table;
4521 struct rps_dev_flow *rflow;
4526 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4527 if (flow_table && flow_id <= flow_table->mask) {
4528 rflow = &flow_table->flows[flow_id];
4529 cpu = READ_ONCE(rflow->cpu);
4530 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4531 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4532 rflow->last_qtail) <
4533 (int)(10 * flow_table->mask)))
4539 EXPORT_SYMBOL(rps_may_expire_flow);
4541 #endif /* CONFIG_RFS_ACCEL */
4543 /* Called from hardirq (IPI) context */
4544 static void rps_trigger_softirq(void *data)
4546 struct softnet_data *sd = data;
4548 ____napi_schedule(sd, &sd->backlog);
4552 #endif /* CONFIG_RPS */
4555 * Check if this softnet_data structure is another cpu one
4556 * If yes, queue it to our IPI list and return 1
4559 static int rps_ipi_queued(struct softnet_data *sd)
4562 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4565 sd->rps_ipi_next = mysd->rps_ipi_list;
4566 mysd->rps_ipi_list = sd;
4568 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4571 #endif /* CONFIG_RPS */
4575 #ifdef CONFIG_NET_FLOW_LIMIT
4576 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4579 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4581 #ifdef CONFIG_NET_FLOW_LIMIT
4582 struct sd_flow_limit *fl;
4583 struct softnet_data *sd;
4584 unsigned int old_flow, new_flow;
4586 if (qlen < (netdev_max_backlog >> 1))
4589 sd = this_cpu_ptr(&softnet_data);
4592 fl = rcu_dereference(sd->flow_limit);
4594 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4595 old_flow = fl->history[fl->history_head];
4596 fl->history[fl->history_head] = new_flow;
4599 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4601 if (likely(fl->buckets[old_flow]))
4602 fl->buckets[old_flow]--;
4604 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4616 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4617 * queue (may be a remote CPU queue).
4619 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4620 unsigned int *qtail)
4622 struct softnet_data *sd;
4623 unsigned long flags;
4626 sd = &per_cpu(softnet_data, cpu);
4628 local_irq_save(flags);
4631 if (!netif_running(skb->dev))
4633 qlen = skb_queue_len(&sd->input_pkt_queue);
4634 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
4637 __skb_queue_tail(&sd->input_pkt_queue, skb);
4638 input_queue_tail_incr_save(sd, qtail);
4640 local_irq_restore(flags);
4641 return NET_RX_SUCCESS;
4644 /* Schedule NAPI for backlog device
4645 * We can use non atomic operation since we own the queue lock
4647 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
4648 if (!rps_ipi_queued(sd))
4649 ____napi_schedule(sd, &sd->backlog);
4658 local_irq_restore(flags);
4660 atomic_long_inc(&skb->dev->rx_dropped);
4665 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4667 struct net_device *dev = skb->dev;
4668 struct netdev_rx_queue *rxqueue;
4672 if (skb_rx_queue_recorded(skb)) {
4673 u16 index = skb_get_rx_queue(skb);
4675 if (unlikely(index >= dev->real_num_rx_queues)) {
4676 WARN_ONCE(dev->real_num_rx_queues > 1,
4677 "%s received packet on queue %u, but number "
4678 "of RX queues is %u\n",
4679 dev->name, index, dev->real_num_rx_queues);
4681 return rxqueue; /* Return first rxqueue */
4688 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp,
4689 struct bpf_prog *xdp_prog)
4691 void *orig_data, *orig_data_end, *hard_start;
4692 struct netdev_rx_queue *rxqueue;
4693 bool orig_bcast, orig_host;
4694 u32 mac_len, frame_sz;
4695 __be16 orig_eth_type;
4700 /* The XDP program wants to see the packet starting at the MAC
4703 mac_len = skb->data - skb_mac_header(skb);
4704 hard_start = skb->data - skb_headroom(skb);
4706 /* SKB "head" area always have tailroom for skb_shared_info */
4707 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4708 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4710 rxqueue = netif_get_rxqueue(skb);
4711 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4712 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4713 skb_headlen(skb) + mac_len, true);
4715 orig_data_end = xdp->data_end;
4716 orig_data = xdp->data;
4717 eth = (struct ethhdr *)xdp->data;
4718 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4719 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4720 orig_eth_type = eth->h_proto;
4722 act = bpf_prog_run_xdp(xdp_prog, xdp);
4724 /* check if bpf_xdp_adjust_head was used */
4725 off = xdp->data - orig_data;
4728 __skb_pull(skb, off);
4730 __skb_push(skb, -off);
4732 skb->mac_header += off;
4733 skb_reset_network_header(skb);
4736 /* check if bpf_xdp_adjust_tail was used */
4737 off = xdp->data_end - orig_data_end;
4739 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4740 skb->len += off; /* positive on grow, negative on shrink */
4743 /* check if XDP changed eth hdr such SKB needs update */
4744 eth = (struct ethhdr *)xdp->data;
4745 if ((orig_eth_type != eth->h_proto) ||
4746 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4747 skb->dev->dev_addr)) ||
4748 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4749 __skb_push(skb, ETH_HLEN);
4750 skb->pkt_type = PACKET_HOST;
4751 skb->protocol = eth_type_trans(skb, skb->dev);
4754 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
4755 * before calling us again on redirect path. We do not call do_redirect
4756 * as we leave that up to the caller.
4758 * Caller is responsible for managing lifetime of skb (i.e. calling
4759 * kfree_skb in response to actions it cannot handle/XDP_DROP).
4764 __skb_push(skb, mac_len);
4767 metalen = xdp->data - xdp->data_meta;
4769 skb_metadata_set(skb, metalen);
4776 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4777 struct xdp_buff *xdp,
4778 struct bpf_prog *xdp_prog)
4782 /* Reinjected packets coming from act_mirred or similar should
4783 * not get XDP generic processing.
4785 if (skb_is_redirected(skb))
4788 /* XDP packets must be linear and must have sufficient headroom
4789 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4790 * native XDP provides, thus we need to do it here as well.
4792 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4793 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4794 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4795 int troom = skb->tail + skb->data_len - skb->end;
4797 /* In case we have to go down the path and also linearize,
4798 * then lets do the pskb_expand_head() work just once here.
4800 if (pskb_expand_head(skb,
4801 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4802 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4804 if (skb_linearize(skb))
4808 act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
4815 bpf_warn_invalid_xdp_action(act);
4818 trace_xdp_exception(skb->dev, xdp_prog, act);
4829 /* When doing generic XDP we have to bypass the qdisc layer and the
4830 * network taps in order to match in-driver-XDP behavior.
4832 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4834 struct net_device *dev = skb->dev;
4835 struct netdev_queue *txq;
4836 bool free_skb = true;
4839 txq = netdev_core_pick_tx(dev, skb, NULL);
4840 cpu = smp_processor_id();
4841 HARD_TX_LOCK(dev, txq, cpu);
4842 if (!netif_xmit_stopped(txq)) {
4843 rc = netdev_start_xmit(skb, dev, txq, 0);
4844 if (dev_xmit_complete(rc))
4847 HARD_TX_UNLOCK(dev, txq);
4849 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4854 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4856 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4859 struct xdp_buff xdp;
4863 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4864 if (act != XDP_PASS) {
4867 err = xdp_do_generic_redirect(skb->dev, skb,
4873 generic_xdp_tx(skb, xdp_prog);
4884 EXPORT_SYMBOL_GPL(do_xdp_generic);
4886 static int netif_rx_internal(struct sk_buff *skb)
4890 net_timestamp_check(netdev_tstamp_prequeue, skb);
4892 trace_netif_rx(skb);
4895 if (static_branch_unlikely(&rps_needed)) {
4896 struct rps_dev_flow voidflow, *rflow = &voidflow;
4902 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4904 cpu = smp_processor_id();
4906 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4915 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4922 * netif_rx - post buffer to the network code
4923 * @skb: buffer to post
4925 * This function receives a packet from a device driver and queues it for
4926 * the upper (protocol) levels to process. It always succeeds. The buffer
4927 * may be dropped during processing for congestion control or by the
4931 * NET_RX_SUCCESS (no congestion)
4932 * NET_RX_DROP (packet was dropped)
4936 int netif_rx(struct sk_buff *skb)
4940 trace_netif_rx_entry(skb);
4942 ret = netif_rx_internal(skb);
4943 trace_netif_rx_exit(ret);
4947 EXPORT_SYMBOL(netif_rx);
4949 int netif_rx_ni(struct sk_buff *skb)
4953 trace_netif_rx_ni_entry(skb);
4956 err = netif_rx_internal(skb);
4957 if (local_softirq_pending())
4960 trace_netif_rx_ni_exit(err);
4964 EXPORT_SYMBOL(netif_rx_ni);
4966 int netif_rx_any_context(struct sk_buff *skb)
4969 * If invoked from contexts which do not invoke bottom half
4970 * processing either at return from interrupt or when softrqs are
4971 * reenabled, use netif_rx_ni() which invokes bottomhalf processing
4975 return netif_rx(skb);
4977 return netif_rx_ni(skb);
4979 EXPORT_SYMBOL(netif_rx_any_context);
4981 static __latent_entropy void net_tx_action(struct softirq_action *h)
4983 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4985 if (sd->completion_queue) {
4986 struct sk_buff *clist;
4988 local_irq_disable();
4989 clist = sd->completion_queue;
4990 sd->completion_queue = NULL;
4994 struct sk_buff *skb = clist;
4996 clist = clist->next;
4998 WARN_ON(refcount_read(&skb->users));
4999 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
5000 trace_consume_skb(skb);
5002 trace_kfree_skb(skb, net_tx_action);
5004 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
5007 __kfree_skb_defer(skb);
5011 if (sd->output_queue) {
5014 local_irq_disable();
5015 head = sd->output_queue;
5016 sd->output_queue = NULL;
5017 sd->output_queue_tailp = &sd->output_queue;
5023 struct Qdisc *q = head;
5024 spinlock_t *root_lock = NULL;
5026 head = head->next_sched;
5028 /* We need to make sure head->next_sched is read
5029 * before clearing __QDISC_STATE_SCHED
5031 smp_mb__before_atomic();
5033 if (!(q->flags & TCQ_F_NOLOCK)) {
5034 root_lock = qdisc_lock(q);
5035 spin_lock(root_lock);
5036 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5038 /* There is a synchronize_net() between
5039 * STATE_DEACTIVATED flag being set and
5040 * qdisc_reset()/some_qdisc_is_busy() in
5041 * dev_deactivate(), so we can safely bail out
5042 * early here to avoid data race between
5043 * qdisc_deactivate() and some_qdisc_is_busy()
5044 * for lockless qdisc.
5046 clear_bit(__QDISC_STATE_SCHED, &q->state);
5050 clear_bit(__QDISC_STATE_SCHED, &q->state);
5053 spin_unlock(root_lock);
5059 xfrm_dev_backlog(sd);
5062 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5063 /* This hook is defined here for ATM LANE */
5064 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5065 unsigned char *addr) __read_mostly;
5066 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5069 static inline struct sk_buff *
5070 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
5071 struct net_device *orig_dev, bool *another)
5073 #ifdef CONFIG_NET_CLS_ACT
5074 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
5075 struct tcf_result cl_res;
5077 /* If there's at least one ingress present somewhere (so
5078 * we get here via enabled static key), remaining devices
5079 * that are not configured with an ingress qdisc will bail
5086 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5090 qdisc_skb_cb(skb)->pkt_len = skb->len;
5091 tc_skb_cb(skb)->mru = 0;
5092 tc_skb_cb(skb)->post_ct = false;
5093 skb->tc_at_ingress = 1;
5094 mini_qdisc_bstats_cpu_update(miniq, skb);
5096 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
5098 case TC_ACT_RECLASSIFY:
5099 skb->tc_index = TC_H_MIN(cl_res.classid);
5102 mini_qdisc_qstats_cpu_drop(miniq);
5110 case TC_ACT_REDIRECT:
5111 /* skb_mac_header check was done by cls/act_bpf, so
5112 * we can safely push the L2 header back before
5113 * redirecting to another netdev
5115 __skb_push(skb, skb->mac_len);
5116 if (skb_do_redirect(skb) == -EAGAIN) {
5117 __skb_pull(skb, skb->mac_len);
5122 case TC_ACT_CONSUMED:
5127 #endif /* CONFIG_NET_CLS_ACT */
5132 * netdev_is_rx_handler_busy - check if receive handler is registered
5133 * @dev: device to check
5135 * Check if a receive handler is already registered for a given device.
5136 * Return true if there one.
5138 * The caller must hold the rtnl_mutex.
5140 bool netdev_is_rx_handler_busy(struct net_device *dev)
5143 return dev && rtnl_dereference(dev->rx_handler);
5145 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5148 * netdev_rx_handler_register - register receive handler
5149 * @dev: device to register a handler for
5150 * @rx_handler: receive handler to register
5151 * @rx_handler_data: data pointer that is used by rx handler
5153 * Register a receive handler for a device. This handler will then be
5154 * called from __netif_receive_skb. A negative errno code is returned
5157 * The caller must hold the rtnl_mutex.
5159 * For a general description of rx_handler, see enum rx_handler_result.
5161 int netdev_rx_handler_register(struct net_device *dev,
5162 rx_handler_func_t *rx_handler,
5163 void *rx_handler_data)
5165 if (netdev_is_rx_handler_busy(dev))
5168 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5171 /* Note: rx_handler_data must be set before rx_handler */
5172 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5173 rcu_assign_pointer(dev->rx_handler, rx_handler);
5177 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5180 * netdev_rx_handler_unregister - unregister receive handler
5181 * @dev: device to unregister a handler from
5183 * Unregister a receive handler from a device.
5185 * The caller must hold the rtnl_mutex.
5187 void netdev_rx_handler_unregister(struct net_device *dev)
5191 RCU_INIT_POINTER(dev->rx_handler, NULL);
5192 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5193 * section has a guarantee to see a non NULL rx_handler_data
5197 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5199 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5202 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5203 * the special handling of PFMEMALLOC skbs.
5205 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5207 switch (skb->protocol) {
5208 case htons(ETH_P_ARP):
5209 case htons(ETH_P_IP):
5210 case htons(ETH_P_IPV6):
5211 case htons(ETH_P_8021Q):
5212 case htons(ETH_P_8021AD):
5219 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5220 int *ret, struct net_device *orig_dev)
5222 if (nf_hook_ingress_active(skb)) {
5226 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5231 ingress_retval = nf_hook_ingress(skb);
5233 return ingress_retval;
5238 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5239 struct packet_type **ppt_prev)
5241 struct packet_type *ptype, *pt_prev;
5242 rx_handler_func_t *rx_handler;
5243 struct sk_buff *skb = *pskb;
5244 struct net_device *orig_dev;
5245 bool deliver_exact = false;
5246 int ret = NET_RX_DROP;
5249 net_timestamp_check(!netdev_tstamp_prequeue, skb);
5251 trace_netif_receive_skb(skb);
5253 orig_dev = skb->dev;
5255 skb_reset_network_header(skb);
5256 if (!skb_transport_header_was_set(skb))
5257 skb_reset_transport_header(skb);
5258 skb_reset_mac_len(skb);
5263 skb->skb_iif = skb->dev->ifindex;
5265 __this_cpu_inc(softnet_data.processed);
5267 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5271 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5274 if (ret2 != XDP_PASS) {
5280 if (eth_type_vlan(skb->protocol)) {
5281 skb = skb_vlan_untag(skb);
5286 if (skb_skip_tc_classify(skb))
5292 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5294 ret = deliver_skb(skb, pt_prev, orig_dev);
5298 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5300 ret = deliver_skb(skb, pt_prev, orig_dev);
5305 #ifdef CONFIG_NET_INGRESS
5306 if (static_branch_unlikely(&ingress_needed_key)) {
5307 bool another = false;
5309 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5316 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5320 skb_reset_redirect(skb);
5322 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5325 if (skb_vlan_tag_present(skb)) {
5327 ret = deliver_skb(skb, pt_prev, orig_dev);
5330 if (vlan_do_receive(&skb))
5332 else if (unlikely(!skb))
5336 rx_handler = rcu_dereference(skb->dev->rx_handler);
5339 ret = deliver_skb(skb, pt_prev, orig_dev);
5342 switch (rx_handler(&skb)) {
5343 case RX_HANDLER_CONSUMED:
5344 ret = NET_RX_SUCCESS;
5346 case RX_HANDLER_ANOTHER:
5348 case RX_HANDLER_EXACT:
5349 deliver_exact = true;
5351 case RX_HANDLER_PASS:
5358 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5360 if (skb_vlan_tag_get_id(skb)) {
5361 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5364 skb->pkt_type = PACKET_OTHERHOST;
5365 } else if (eth_type_vlan(skb->protocol)) {
5366 /* Outer header is 802.1P with vlan 0, inner header is
5367 * 802.1Q or 802.1AD and vlan_do_receive() above could
5368 * not find vlan dev for vlan id 0.
5370 __vlan_hwaccel_clear_tag(skb);
5371 skb = skb_vlan_untag(skb);
5374 if (vlan_do_receive(&skb))
5375 /* After stripping off 802.1P header with vlan 0
5376 * vlan dev is found for inner header.
5379 else if (unlikely(!skb))
5382 /* We have stripped outer 802.1P vlan 0 header.
5383 * But could not find vlan dev.
5384 * check again for vlan id to set OTHERHOST.
5388 /* Note: we might in the future use prio bits
5389 * and set skb->priority like in vlan_do_receive()
5390 * For the time being, just ignore Priority Code Point
5392 __vlan_hwaccel_clear_tag(skb);
5395 type = skb->protocol;
5397 /* deliver only exact match when indicated */
5398 if (likely(!deliver_exact)) {
5399 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5400 &ptype_base[ntohs(type) &
5404 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5405 &orig_dev->ptype_specific);
5407 if (unlikely(skb->dev != orig_dev)) {
5408 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5409 &skb->dev->ptype_specific);
5413 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5415 *ppt_prev = pt_prev;
5419 atomic_long_inc(&skb->dev->rx_dropped);
5421 atomic_long_inc(&skb->dev->rx_nohandler);
5423 /* Jamal, now you will not able to escape explaining
5424 * me how you were going to use this. :-)
5430 /* The invariant here is that if *ppt_prev is not NULL
5431 * then skb should also be non-NULL.
5433 * Apparently *ppt_prev assignment above holds this invariant due to
5434 * skb dereferencing near it.
5440 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5442 struct net_device *orig_dev = skb->dev;
5443 struct packet_type *pt_prev = NULL;
5446 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5448 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5449 skb->dev, pt_prev, orig_dev);
5454 * netif_receive_skb_core - special purpose version of netif_receive_skb
5455 * @skb: buffer to process
5457 * More direct receive version of netif_receive_skb(). It should
5458 * only be used by callers that have a need to skip RPS and Generic XDP.
5459 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5461 * This function may only be called from softirq context and interrupts
5462 * should be enabled.
5464 * Return values (usually ignored):
5465 * NET_RX_SUCCESS: no congestion
5466 * NET_RX_DROP: packet was dropped
5468 int netif_receive_skb_core(struct sk_buff *skb)
5473 ret = __netif_receive_skb_one_core(skb, false);
5478 EXPORT_SYMBOL(netif_receive_skb_core);
5480 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5481 struct packet_type *pt_prev,
5482 struct net_device *orig_dev)
5484 struct sk_buff *skb, *next;
5488 if (list_empty(head))
5490 if (pt_prev->list_func != NULL)
5491 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5492 ip_list_rcv, head, pt_prev, orig_dev);
5494 list_for_each_entry_safe(skb, next, head, list) {
5495 skb_list_del_init(skb);
5496 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5500 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5502 /* Fast-path assumptions:
5503 * - There is no RX handler.
5504 * - Only one packet_type matches.
5505 * If either of these fails, we will end up doing some per-packet
5506 * processing in-line, then handling the 'last ptype' for the whole
5507 * sublist. This can't cause out-of-order delivery to any single ptype,
5508 * because the 'last ptype' must be constant across the sublist, and all
5509 * other ptypes are handled per-packet.
5511 /* Current (common) ptype of sublist */
5512 struct packet_type *pt_curr = NULL;
5513 /* Current (common) orig_dev of sublist */
5514 struct net_device *od_curr = NULL;
5515 struct list_head sublist;
5516 struct sk_buff *skb, *next;
5518 INIT_LIST_HEAD(&sublist);
5519 list_for_each_entry_safe(skb, next, head, list) {
5520 struct net_device *orig_dev = skb->dev;
5521 struct packet_type *pt_prev = NULL;
5523 skb_list_del_init(skb);
5524 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5527 if (pt_curr != pt_prev || od_curr != orig_dev) {
5528 /* dispatch old sublist */
5529 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5530 /* start new sublist */
5531 INIT_LIST_HEAD(&sublist);
5535 list_add_tail(&skb->list, &sublist);
5538 /* dispatch final sublist */
5539 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5542 static int __netif_receive_skb(struct sk_buff *skb)
5546 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5547 unsigned int noreclaim_flag;
5550 * PFMEMALLOC skbs are special, they should
5551 * - be delivered to SOCK_MEMALLOC sockets only
5552 * - stay away from userspace
5553 * - have bounded memory usage
5555 * Use PF_MEMALLOC as this saves us from propagating the allocation
5556 * context down to all allocation sites.
5558 noreclaim_flag = memalloc_noreclaim_save();
5559 ret = __netif_receive_skb_one_core(skb, true);
5560 memalloc_noreclaim_restore(noreclaim_flag);
5562 ret = __netif_receive_skb_one_core(skb, false);
5567 static void __netif_receive_skb_list(struct list_head *head)
5569 unsigned long noreclaim_flag = 0;
5570 struct sk_buff *skb, *next;
5571 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5573 list_for_each_entry_safe(skb, next, head, list) {
5574 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5575 struct list_head sublist;
5577 /* Handle the previous sublist */
5578 list_cut_before(&sublist, head, &skb->list);
5579 if (!list_empty(&sublist))
5580 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5581 pfmemalloc = !pfmemalloc;
5582 /* See comments in __netif_receive_skb */
5584 noreclaim_flag = memalloc_noreclaim_save();
5586 memalloc_noreclaim_restore(noreclaim_flag);
5589 /* Handle the remaining sublist */
5590 if (!list_empty(head))
5591 __netif_receive_skb_list_core(head, pfmemalloc);
5592 /* Restore pflags */
5594 memalloc_noreclaim_restore(noreclaim_flag);
5597 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5599 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5600 struct bpf_prog *new = xdp->prog;
5603 switch (xdp->command) {
5604 case XDP_SETUP_PROG:
5605 rcu_assign_pointer(dev->xdp_prog, new);
5610 static_branch_dec(&generic_xdp_needed_key);
5611 } else if (new && !old) {
5612 static_branch_inc(&generic_xdp_needed_key);
5613 dev_disable_lro(dev);
5614 dev_disable_gro_hw(dev);
5626 static int netif_receive_skb_internal(struct sk_buff *skb)
5630 net_timestamp_check(netdev_tstamp_prequeue, skb);
5632 if (skb_defer_rx_timestamp(skb))
5633 return NET_RX_SUCCESS;
5637 if (static_branch_unlikely(&rps_needed)) {
5638 struct rps_dev_flow voidflow, *rflow = &voidflow;
5639 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5642 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5648 ret = __netif_receive_skb(skb);
5653 static void netif_receive_skb_list_internal(struct list_head *head)
5655 struct sk_buff *skb, *next;
5656 struct list_head sublist;
5658 INIT_LIST_HEAD(&sublist);
5659 list_for_each_entry_safe(skb, next, head, list) {
5660 net_timestamp_check(netdev_tstamp_prequeue, skb);
5661 skb_list_del_init(skb);
5662 if (!skb_defer_rx_timestamp(skb))
5663 list_add_tail(&skb->list, &sublist);
5665 list_splice_init(&sublist, head);
5669 if (static_branch_unlikely(&rps_needed)) {
5670 list_for_each_entry_safe(skb, next, head, list) {
5671 struct rps_dev_flow voidflow, *rflow = &voidflow;
5672 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5675 /* Will be handled, remove from list */
5676 skb_list_del_init(skb);
5677 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5682 __netif_receive_skb_list(head);
5687 * netif_receive_skb - process receive buffer from network
5688 * @skb: buffer to process
5690 * netif_receive_skb() is the main receive data processing function.
5691 * It always succeeds. The buffer may be dropped during processing
5692 * for congestion control or by the protocol layers.
5694 * This function may only be called from softirq context and interrupts
5695 * should be enabled.
5697 * Return values (usually ignored):
5698 * NET_RX_SUCCESS: no congestion
5699 * NET_RX_DROP: packet was dropped
5701 int netif_receive_skb(struct sk_buff *skb)
5705 trace_netif_receive_skb_entry(skb);
5707 ret = netif_receive_skb_internal(skb);
5708 trace_netif_receive_skb_exit(ret);
5712 EXPORT_SYMBOL(netif_receive_skb);
5715 * netif_receive_skb_list - process many receive buffers from network
5716 * @head: list of skbs to process.
5718 * Since return value of netif_receive_skb() is normally ignored, and
5719 * wouldn't be meaningful for a list, this function returns void.
5721 * This function may only be called from softirq context and interrupts
5722 * should be enabled.
5724 void netif_receive_skb_list(struct list_head *head)
5726 struct sk_buff *skb;
5728 if (list_empty(head))
5730 if (trace_netif_receive_skb_list_entry_enabled()) {
5731 list_for_each_entry(skb, head, list)
5732 trace_netif_receive_skb_list_entry(skb);
5734 netif_receive_skb_list_internal(head);
5735 trace_netif_receive_skb_list_exit(0);
5737 EXPORT_SYMBOL(netif_receive_skb_list);
5739 static DEFINE_PER_CPU(struct work_struct, flush_works);
5741 /* Network device is going away, flush any packets still pending */
5742 static void flush_backlog(struct work_struct *work)
5744 struct sk_buff *skb, *tmp;
5745 struct softnet_data *sd;
5748 sd = this_cpu_ptr(&softnet_data);
5750 local_irq_disable();
5752 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5753 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5754 __skb_unlink(skb, &sd->input_pkt_queue);
5755 dev_kfree_skb_irq(skb);
5756 input_queue_head_incr(sd);
5762 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5763 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5764 __skb_unlink(skb, &sd->process_queue);
5766 input_queue_head_incr(sd);
5772 static bool flush_required(int cpu)
5774 #if IS_ENABLED(CONFIG_RPS)
5775 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5778 local_irq_disable();
5781 /* as insertion into process_queue happens with the rps lock held,
5782 * process_queue access may race only with dequeue
5784 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5785 !skb_queue_empty_lockless(&sd->process_queue);
5791 /* without RPS we can't safely check input_pkt_queue: during a
5792 * concurrent remote skb_queue_splice() we can detect as empty both
5793 * input_pkt_queue and process_queue even if the latter could end-up
5794 * containing a lot of packets.
5799 static void flush_all_backlogs(void)
5801 static cpumask_t flush_cpus;
5804 /* since we are under rtnl lock protection we can use static data
5805 * for the cpumask and avoid allocating on stack the possibly
5812 cpumask_clear(&flush_cpus);
5813 for_each_online_cpu(cpu) {
5814 if (flush_required(cpu)) {
5815 queue_work_on(cpu, system_highpri_wq,
5816 per_cpu_ptr(&flush_works, cpu));
5817 cpumask_set_cpu(cpu, &flush_cpus);
5821 /* we can have in flight packet[s] on the cpus we are not flushing,
5822 * synchronize_net() in unregister_netdevice_many() will take care of
5825 for_each_cpu(cpu, &flush_cpus)
5826 flush_work(per_cpu_ptr(&flush_works, cpu));
5831 /* Pass the currently batched GRO_NORMAL SKBs up to the stack. */
5832 static void gro_normal_list(struct napi_struct *napi)
5834 if (!napi->rx_count)
5836 netif_receive_skb_list_internal(&napi->rx_list);
5837 INIT_LIST_HEAD(&napi->rx_list);
5841 /* Queue one GRO_NORMAL SKB up for list processing. If batch size exceeded,
5842 * pass the whole batch up to the stack.
5844 static void gro_normal_one(struct napi_struct *napi, struct sk_buff *skb, int segs)
5846 list_add_tail(&skb->list, &napi->rx_list);
5847 napi->rx_count += segs;
5848 if (napi->rx_count >= gro_normal_batch)
5849 gro_normal_list(napi);
5852 static int napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
5854 struct packet_offload *ptype;
5855 __be16 type = skb->protocol;
5856 struct list_head *head = &offload_base;
5859 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
5861 if (NAPI_GRO_CB(skb)->count == 1) {
5862 skb_shinfo(skb)->gso_size = 0;
5867 list_for_each_entry_rcu(ptype, head, list) {
5868 if (ptype->type != type || !ptype->callbacks.gro_complete)
5871 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
5872 ipv6_gro_complete, inet_gro_complete,
5879 WARN_ON(&ptype->list == head);
5881 return NET_RX_SUCCESS;
5885 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
5886 return NET_RX_SUCCESS;
5889 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
5892 struct list_head *head = &napi->gro_hash[index].list;
5893 struct sk_buff *skb, *p;
5895 list_for_each_entry_safe_reverse(skb, p, head, list) {
5896 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
5898 skb_list_del_init(skb);
5899 napi_gro_complete(napi, skb);
5900 napi->gro_hash[index].count--;
5903 if (!napi->gro_hash[index].count)
5904 __clear_bit(index, &napi->gro_bitmask);
5907 /* napi->gro_hash[].list contains packets ordered by age.
5908 * youngest packets at the head of it.
5909 * Complete skbs in reverse order to reduce latencies.
5911 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
5913 unsigned long bitmask = napi->gro_bitmask;
5914 unsigned int i, base = ~0U;
5916 while ((i = ffs(bitmask)) != 0) {
5919 __napi_gro_flush_chain(napi, base, flush_old);
5922 EXPORT_SYMBOL(napi_gro_flush);
5924 static void gro_list_prepare(const struct list_head *head,
5925 const struct sk_buff *skb)
5927 unsigned int maclen = skb->dev->hard_header_len;
5928 u32 hash = skb_get_hash_raw(skb);
5931 list_for_each_entry(p, head, list) {
5932 unsigned long diffs;
5934 NAPI_GRO_CB(p)->flush = 0;
5936 if (hash != skb_get_hash_raw(p)) {
5937 NAPI_GRO_CB(p)->same_flow = 0;
5941 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
5942 diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb);
5943 if (skb_vlan_tag_present(p))
5944 diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb);
5945 diffs |= skb_metadata_differs(p, skb);
5946 if (maclen == ETH_HLEN)
5947 diffs |= compare_ether_header(skb_mac_header(p),
5948 skb_mac_header(skb));
5950 diffs = memcmp(skb_mac_header(p),
5951 skb_mac_header(skb),
5954 /* in most common scenarions 'slow_gro' is 0
5955 * otherwise we are already on some slower paths
5956 * either skip all the infrequent tests altogether or
5957 * avoid trying too hard to skip each of them individually
5959 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
5960 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
5961 struct tc_skb_ext *skb_ext;
5962 struct tc_skb_ext *p_ext;
5965 diffs |= p->sk != skb->sk;
5966 diffs |= skb_metadata_dst_cmp(p, skb);
5967 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
5969 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
5970 skb_ext = skb_ext_find(skb, TC_SKB_EXT);
5971 p_ext = skb_ext_find(p, TC_SKB_EXT);
5973 diffs |= (!!p_ext) ^ (!!skb_ext);
5974 if (!diffs && unlikely(skb_ext))
5975 diffs |= p_ext->chain ^ skb_ext->chain;
5979 NAPI_GRO_CB(p)->same_flow = !diffs;
5983 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
5985 const struct skb_shared_info *pinfo = skb_shinfo(skb);
5986 const skb_frag_t *frag0 = &pinfo->frags[0];
5988 NAPI_GRO_CB(skb)->data_offset = 0;
5989 NAPI_GRO_CB(skb)->frag0 = NULL;
5990 NAPI_GRO_CB(skb)->frag0_len = 0;
5992 if (!skb_headlen(skb) && pinfo->nr_frags &&
5993 !PageHighMem(skb_frag_page(frag0)) &&
5994 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
5995 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
5996 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
5997 skb_frag_size(frag0),
5998 skb->end - skb->tail);
6002 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
6004 struct skb_shared_info *pinfo = skb_shinfo(skb);
6006 BUG_ON(skb->end - skb->tail < grow);
6008 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
6010 skb->data_len -= grow;
6013 skb_frag_off_add(&pinfo->frags[0], grow);
6014 skb_frag_size_sub(&pinfo->frags[0], grow);
6016 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
6017 skb_frag_unref(skb, 0);
6018 memmove(pinfo->frags, pinfo->frags + 1,
6019 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
6023 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
6025 struct sk_buff *oldest;
6027 oldest = list_last_entry(head, struct sk_buff, list);
6029 /* We are called with head length >= MAX_GRO_SKBS, so this is
6032 if (WARN_ON_ONCE(!oldest))
6035 /* Do not adjust napi->gro_hash[].count, caller is adding a new
6038 skb_list_del_init(oldest);
6039 napi_gro_complete(napi, oldest);
6042 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6044 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
6045 struct gro_list *gro_list = &napi->gro_hash[bucket];
6046 struct list_head *head = &offload_base;
6047 struct packet_offload *ptype;
6048 __be16 type = skb->protocol;
6049 struct sk_buff *pp = NULL;
6050 enum gro_result ret;
6054 if (netif_elide_gro(skb->dev))
6057 gro_list_prepare(&gro_list->list, skb);
6060 list_for_each_entry_rcu(ptype, head, list) {
6061 if (ptype->type != type || !ptype->callbacks.gro_receive)
6064 skb_set_network_header(skb, skb_gro_offset(skb));
6065 skb_reset_mac_len(skb);
6066 NAPI_GRO_CB(skb)->same_flow = 0;
6067 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
6068 NAPI_GRO_CB(skb)->free = 0;
6069 NAPI_GRO_CB(skb)->encap_mark = 0;
6070 NAPI_GRO_CB(skb)->recursion_counter = 0;
6071 NAPI_GRO_CB(skb)->is_fou = 0;
6072 NAPI_GRO_CB(skb)->is_atomic = 1;
6073 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
6075 /* Setup for GRO checksum validation */
6076 switch (skb->ip_summed) {
6077 case CHECKSUM_COMPLETE:
6078 NAPI_GRO_CB(skb)->csum = skb->csum;
6079 NAPI_GRO_CB(skb)->csum_valid = 1;
6080 NAPI_GRO_CB(skb)->csum_cnt = 0;
6082 case CHECKSUM_UNNECESSARY:
6083 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
6084 NAPI_GRO_CB(skb)->csum_valid = 0;
6087 NAPI_GRO_CB(skb)->csum_cnt = 0;
6088 NAPI_GRO_CB(skb)->csum_valid = 0;
6091 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
6092 ipv6_gro_receive, inet_gro_receive,
6093 &gro_list->list, skb);
6098 if (&ptype->list == head)
6101 if (PTR_ERR(pp) == -EINPROGRESS) {
6106 same_flow = NAPI_GRO_CB(skb)->same_flow;
6107 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
6110 skb_list_del_init(pp);
6111 napi_gro_complete(napi, pp);
6118 if (NAPI_GRO_CB(skb)->flush)
6121 if (unlikely(gro_list->count >= MAX_GRO_SKBS))
6122 gro_flush_oldest(napi, &gro_list->list);
6126 NAPI_GRO_CB(skb)->count = 1;
6127 NAPI_GRO_CB(skb)->age = jiffies;
6128 NAPI_GRO_CB(skb)->last = skb;
6129 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
6130 list_add(&skb->list, &gro_list->list);
6134 grow = skb_gro_offset(skb) - skb_headlen(skb);
6136 gro_pull_from_frag0(skb, grow);
6138 if (gro_list->count) {
6139 if (!test_bit(bucket, &napi->gro_bitmask))
6140 __set_bit(bucket, &napi->gro_bitmask);
6141 } else if (test_bit(bucket, &napi->gro_bitmask)) {
6142 __clear_bit(bucket, &napi->gro_bitmask);
6152 struct packet_offload *gro_find_receive_by_type(__be16 type)
6154 struct list_head *offload_head = &offload_base;
6155 struct packet_offload *ptype;
6157 list_for_each_entry_rcu(ptype, offload_head, list) {
6158 if (ptype->type != type || !ptype->callbacks.gro_receive)
6164 EXPORT_SYMBOL(gro_find_receive_by_type);
6166 struct packet_offload *gro_find_complete_by_type(__be16 type)
6168 struct list_head *offload_head = &offload_base;
6169 struct packet_offload *ptype;
6171 list_for_each_entry_rcu(ptype, offload_head, list) {
6172 if (ptype->type != type || !ptype->callbacks.gro_complete)
6178 EXPORT_SYMBOL(gro_find_complete_by_type);
6180 static gro_result_t napi_skb_finish(struct napi_struct *napi,
6181 struct sk_buff *skb,
6186 gro_normal_one(napi, skb, 1);
6189 case GRO_MERGED_FREE:
6190 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6191 napi_skb_free_stolen_head(skb);
6192 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
6195 __kfree_skb_defer(skb);
6207 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6211 skb_mark_napi_id(skb, napi);
6212 trace_napi_gro_receive_entry(skb);
6214 skb_gro_reset_offset(skb, 0);
6216 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
6217 trace_napi_gro_receive_exit(ret);
6221 EXPORT_SYMBOL(napi_gro_receive);
6223 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
6225 if (unlikely(skb->pfmemalloc)) {
6229 __skb_pull(skb, skb_headlen(skb));
6230 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
6231 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
6232 __vlan_hwaccel_clear_tag(skb);
6233 skb->dev = napi->dev;
6236 /* eth_type_trans() assumes pkt_type is PACKET_HOST */
6237 skb->pkt_type = PACKET_HOST;
6239 skb->encapsulation = 0;
6240 skb_shinfo(skb)->gso_type = 0;
6241 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
6242 if (unlikely(skb->slow_gro)) {
6252 struct sk_buff *napi_get_frags(struct napi_struct *napi)
6254 struct sk_buff *skb = napi->skb;
6257 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
6260 skb_mark_napi_id(skb, napi);
6265 EXPORT_SYMBOL(napi_get_frags);
6267 static gro_result_t napi_frags_finish(struct napi_struct *napi,
6268 struct sk_buff *skb,
6274 __skb_push(skb, ETH_HLEN);
6275 skb->protocol = eth_type_trans(skb, skb->dev);
6276 if (ret == GRO_NORMAL)
6277 gro_normal_one(napi, skb, 1);
6280 case GRO_MERGED_FREE:
6281 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6282 napi_skb_free_stolen_head(skb);
6284 napi_reuse_skb(napi, skb);
6295 /* Upper GRO stack assumes network header starts at gro_offset=0
6296 * Drivers could call both napi_gro_frags() and napi_gro_receive()
6297 * We copy ethernet header into skb->data to have a common layout.
6299 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
6301 struct sk_buff *skb = napi->skb;
6302 const struct ethhdr *eth;
6303 unsigned int hlen = sizeof(*eth);
6307 skb_reset_mac_header(skb);
6308 skb_gro_reset_offset(skb, hlen);
6310 if (unlikely(skb_gro_header_hard(skb, hlen))) {
6311 eth = skb_gro_header_slow(skb, hlen, 0);
6312 if (unlikely(!eth)) {
6313 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
6314 __func__, napi->dev->name);
6315 napi_reuse_skb(napi, skb);
6319 eth = (const struct ethhdr *)skb->data;
6320 gro_pull_from_frag0(skb, hlen);
6321 NAPI_GRO_CB(skb)->frag0 += hlen;
6322 NAPI_GRO_CB(skb)->frag0_len -= hlen;
6324 __skb_pull(skb, hlen);
6327 * This works because the only protocols we care about don't require
6329 * We'll fix it up properly in napi_frags_finish()
6331 skb->protocol = eth->h_proto;
6336 gro_result_t napi_gro_frags(struct napi_struct *napi)
6339 struct sk_buff *skb = napi_frags_skb(napi);
6341 trace_napi_gro_frags_entry(skb);
6343 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
6344 trace_napi_gro_frags_exit(ret);
6348 EXPORT_SYMBOL(napi_gro_frags);
6350 /* Compute the checksum from gro_offset and return the folded value
6351 * after adding in any pseudo checksum.
6353 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
6358 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
6360 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
6361 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
6362 /* See comments in __skb_checksum_complete(). */
6364 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
6365 !skb->csum_complete_sw)
6366 netdev_rx_csum_fault(skb->dev, skb);
6369 NAPI_GRO_CB(skb)->csum = wsum;
6370 NAPI_GRO_CB(skb)->csum_valid = 1;
6374 EXPORT_SYMBOL(__skb_gro_checksum_complete);
6376 static void net_rps_send_ipi(struct softnet_data *remsd)
6380 struct softnet_data *next = remsd->rps_ipi_next;
6382 if (cpu_online(remsd->cpu))
6383 smp_call_function_single_async(remsd->cpu, &remsd->csd);
6390 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
6391 * Note: called with local irq disabled, but exits with local irq enabled.
6393 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
6396 struct softnet_data *remsd = sd->rps_ipi_list;
6399 sd->rps_ipi_list = NULL;
6403 /* Send pending IPI's to kick RPS processing on remote cpus. */
6404 net_rps_send_ipi(remsd);
6410 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
6413 return sd->rps_ipi_list != NULL;
6419 static int process_backlog(struct napi_struct *napi, int quota)
6421 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
6425 /* Check if we have pending ipi, its better to send them now,
6426 * not waiting net_rx_action() end.
6428 if (sd_has_rps_ipi_waiting(sd)) {
6429 local_irq_disable();
6430 net_rps_action_and_irq_enable(sd);
6433 napi->weight = dev_rx_weight;
6435 struct sk_buff *skb;
6437 while ((skb = __skb_dequeue(&sd->process_queue))) {
6439 __netif_receive_skb(skb);
6441 input_queue_head_incr(sd);
6442 if (++work >= quota)
6447 local_irq_disable();
6449 if (skb_queue_empty(&sd->input_pkt_queue)) {
6451 * Inline a custom version of __napi_complete().
6452 * only current cpu owns and manipulates this napi,
6453 * and NAPI_STATE_SCHED is the only possible flag set
6455 * We can use a plain write instead of clear_bit(),
6456 * and we dont need an smp_mb() memory barrier.
6461 skb_queue_splice_tail_init(&sd->input_pkt_queue,
6462 &sd->process_queue);
6472 * __napi_schedule - schedule for receive
6473 * @n: entry to schedule
6475 * The entry's receive function will be scheduled to run.
6476 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
6478 void __napi_schedule(struct napi_struct *n)
6480 unsigned long flags;
6482 local_irq_save(flags);
6483 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6484 local_irq_restore(flags);
6486 EXPORT_SYMBOL(__napi_schedule);
6489 * napi_schedule_prep - check if napi can be scheduled
6492 * Test if NAPI routine is already running, and if not mark
6493 * it as running. This is used as a condition variable to
6494 * insure only one NAPI poll instance runs. We also make
6495 * sure there is no pending NAPI disable.
6497 bool napi_schedule_prep(struct napi_struct *n)
6499 unsigned long val, new;
6502 val = READ_ONCE(n->state);
6503 if (unlikely(val & NAPIF_STATE_DISABLE))
6505 new = val | NAPIF_STATE_SCHED;
6507 /* Sets STATE_MISSED bit if STATE_SCHED was already set
6508 * This was suggested by Alexander Duyck, as compiler
6509 * emits better code than :
6510 * if (val & NAPIF_STATE_SCHED)
6511 * new |= NAPIF_STATE_MISSED;
6513 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6515 } while (cmpxchg(&n->state, val, new) != val);
6517 return !(val & NAPIF_STATE_SCHED);
6519 EXPORT_SYMBOL(napi_schedule_prep);
6522 * __napi_schedule_irqoff - schedule for receive
6523 * @n: entry to schedule
6525 * Variant of __napi_schedule() assuming hard irqs are masked.
6527 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6528 * because the interrupt disabled assumption might not be true
6529 * due to force-threaded interrupts and spinlock substitution.
6531 void __napi_schedule_irqoff(struct napi_struct *n)
6533 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6534 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6538 EXPORT_SYMBOL(__napi_schedule_irqoff);
6540 bool napi_complete_done(struct napi_struct *n, int work_done)
6542 unsigned long flags, val, new, timeout = 0;
6546 * 1) Don't let napi dequeue from the cpu poll list
6547 * just in case its running on a different cpu.
6548 * 2) If we are busy polling, do nothing here, we have
6549 * the guarantee we will be called later.
6551 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6552 NAPIF_STATE_IN_BUSY_POLL)))
6557 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6558 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6560 if (n->defer_hard_irqs_count > 0) {
6561 n->defer_hard_irqs_count--;
6562 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6566 if (n->gro_bitmask) {
6567 /* When the NAPI instance uses a timeout and keeps postponing
6568 * it, we need to bound somehow the time packets are kept in
6571 napi_gro_flush(n, !!timeout);
6576 if (unlikely(!list_empty(&n->poll_list))) {
6577 /* If n->poll_list is not empty, we need to mask irqs */
6578 local_irq_save(flags);
6579 list_del_init(&n->poll_list);
6580 local_irq_restore(flags);
6584 val = READ_ONCE(n->state);
6586 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6588 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6589 NAPIF_STATE_SCHED_THREADED |
6590 NAPIF_STATE_PREFER_BUSY_POLL);
6592 /* If STATE_MISSED was set, leave STATE_SCHED set,
6593 * because we will call napi->poll() one more time.
6594 * This C code was suggested by Alexander Duyck to help gcc.
6596 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6598 } while (cmpxchg(&n->state, val, new) != val);
6600 if (unlikely(val & NAPIF_STATE_MISSED)) {
6606 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6607 HRTIMER_MODE_REL_PINNED);
6610 EXPORT_SYMBOL(napi_complete_done);
6612 /* must be called under rcu_read_lock(), as we dont take a reference */
6613 static struct napi_struct *napi_by_id(unsigned int napi_id)
6615 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6616 struct napi_struct *napi;
6618 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6619 if (napi->napi_id == napi_id)
6625 #if defined(CONFIG_NET_RX_BUSY_POLL)
6627 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6629 if (!skip_schedule) {
6630 gro_normal_list(napi);
6631 __napi_schedule(napi);
6635 if (napi->gro_bitmask) {
6636 /* flush too old packets
6637 * If HZ < 1000, flush all packets.
6639 napi_gro_flush(napi, HZ >= 1000);
6642 gro_normal_list(napi);
6643 clear_bit(NAPI_STATE_SCHED, &napi->state);
6646 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6649 bool skip_schedule = false;
6650 unsigned long timeout;
6653 /* Busy polling means there is a high chance device driver hard irq
6654 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6655 * set in napi_schedule_prep().
6656 * Since we are about to call napi->poll() once more, we can safely
6657 * clear NAPI_STATE_MISSED.
6659 * Note: x86 could use a single "lock and ..." instruction
6660 * to perform these two clear_bit()
6662 clear_bit(NAPI_STATE_MISSED, &napi->state);
6663 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6667 if (prefer_busy_poll) {
6668 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6669 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6670 if (napi->defer_hard_irqs_count && timeout) {
6671 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6672 skip_schedule = true;
6676 /* All we really want here is to re-enable device interrupts.
6677 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6679 rc = napi->poll(napi, budget);
6680 /* We can't gro_normal_list() here, because napi->poll() might have
6681 * rearmed the napi (napi_complete_done()) in which case it could
6682 * already be running on another CPU.
6684 trace_napi_poll(napi, rc, budget);
6685 netpoll_poll_unlock(have_poll_lock);
6687 __busy_poll_stop(napi, skip_schedule);
6691 void napi_busy_loop(unsigned int napi_id,
6692 bool (*loop_end)(void *, unsigned long),
6693 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6695 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6696 int (*napi_poll)(struct napi_struct *napi, int budget);
6697 void *have_poll_lock = NULL;
6698 struct napi_struct *napi;
6705 napi = napi_by_id(napi_id);
6715 unsigned long val = READ_ONCE(napi->state);
6717 /* If multiple threads are competing for this napi,
6718 * we avoid dirtying napi->state as much as we can.
6720 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6721 NAPIF_STATE_IN_BUSY_POLL)) {
6722 if (prefer_busy_poll)
6723 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6726 if (cmpxchg(&napi->state, val,
6727 val | NAPIF_STATE_IN_BUSY_POLL |
6728 NAPIF_STATE_SCHED) != val) {
6729 if (prefer_busy_poll)
6730 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6733 have_poll_lock = netpoll_poll_lock(napi);
6734 napi_poll = napi->poll;
6736 work = napi_poll(napi, budget);
6737 trace_napi_poll(napi, work, budget);
6738 gro_normal_list(napi);
6741 __NET_ADD_STATS(dev_net(napi->dev),
6742 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6745 if (!loop_end || loop_end(loop_end_arg, start_time))
6748 if (unlikely(need_resched())) {
6750 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6754 if (loop_end(loop_end_arg, start_time))
6761 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6766 EXPORT_SYMBOL(napi_busy_loop);
6768 #endif /* CONFIG_NET_RX_BUSY_POLL */
6770 static void napi_hash_add(struct napi_struct *napi)
6772 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6775 spin_lock(&napi_hash_lock);
6777 /* 0..NR_CPUS range is reserved for sender_cpu use */
6779 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6780 napi_gen_id = MIN_NAPI_ID;
6781 } while (napi_by_id(napi_gen_id));
6782 napi->napi_id = napi_gen_id;
6784 hlist_add_head_rcu(&napi->napi_hash_node,
6785 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6787 spin_unlock(&napi_hash_lock);
6790 /* Warning : caller is responsible to make sure rcu grace period
6791 * is respected before freeing memory containing @napi
6793 static void napi_hash_del(struct napi_struct *napi)
6795 spin_lock(&napi_hash_lock);
6797 hlist_del_init_rcu(&napi->napi_hash_node);
6799 spin_unlock(&napi_hash_lock);
6802 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6804 struct napi_struct *napi;
6806 napi = container_of(timer, struct napi_struct, timer);
6808 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6809 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6811 if (!napi_disable_pending(napi) &&
6812 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6813 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6814 __napi_schedule_irqoff(napi);
6817 return HRTIMER_NORESTART;
6820 static void init_gro_hash(struct napi_struct *napi)
6824 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6825 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6826 napi->gro_hash[i].count = 0;
6828 napi->gro_bitmask = 0;
6831 int dev_set_threaded(struct net_device *dev, bool threaded)
6833 struct napi_struct *napi;
6836 if (dev->threaded == threaded)
6840 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6841 if (!napi->thread) {
6842 err = napi_kthread_create(napi);
6851 dev->threaded = threaded;
6853 /* Make sure kthread is created before THREADED bit
6856 smp_mb__before_atomic();
6858 /* Setting/unsetting threaded mode on a napi might not immediately
6859 * take effect, if the current napi instance is actively being
6860 * polled. In this case, the switch between threaded mode and
6861 * softirq mode will happen in the next round of napi_schedule().
6862 * This should not cause hiccups/stalls to the live traffic.
6864 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6866 set_bit(NAPI_STATE_THREADED, &napi->state);
6868 clear_bit(NAPI_STATE_THREADED, &napi->state);
6873 EXPORT_SYMBOL(dev_set_threaded);
6875 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
6876 int (*poll)(struct napi_struct *, int), int weight)
6878 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6881 INIT_LIST_HEAD(&napi->poll_list);
6882 INIT_HLIST_NODE(&napi->napi_hash_node);
6883 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6884 napi->timer.function = napi_watchdog;
6885 init_gro_hash(napi);
6887 INIT_LIST_HEAD(&napi->rx_list);
6890 if (weight > NAPI_POLL_WEIGHT)
6891 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6893 napi->weight = weight;
6895 #ifdef CONFIG_NETPOLL
6896 napi->poll_owner = -1;
6898 set_bit(NAPI_STATE_SCHED, &napi->state);
6899 set_bit(NAPI_STATE_NPSVC, &napi->state);
6900 list_add_rcu(&napi->dev_list, &dev->napi_list);
6901 napi_hash_add(napi);
6902 /* Create kthread for this napi if dev->threaded is set.
6903 * Clear dev->threaded if kthread creation failed so that
6904 * threaded mode will not be enabled in napi_enable().
6906 if (dev->threaded && napi_kthread_create(napi))
6909 EXPORT_SYMBOL(netif_napi_add);
6911 void napi_disable(struct napi_struct *n)
6914 set_bit(NAPI_STATE_DISABLE, &n->state);
6916 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
6918 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
6921 hrtimer_cancel(&n->timer);
6923 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &n->state);
6924 clear_bit(NAPI_STATE_DISABLE, &n->state);
6925 clear_bit(NAPI_STATE_THREADED, &n->state);
6927 EXPORT_SYMBOL(napi_disable);
6930 * napi_enable - enable NAPI scheduling
6933 * Resume NAPI from being scheduled on this context.
6934 * Must be paired with napi_disable.
6936 void napi_enable(struct napi_struct *n)
6938 unsigned long val, new;
6941 val = READ_ONCE(n->state);
6942 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
6944 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
6945 if (n->dev->threaded && n->thread)
6946 new |= NAPIF_STATE_THREADED;
6947 } while (cmpxchg(&n->state, val, new) != val);
6949 EXPORT_SYMBOL(napi_enable);
6951 static void flush_gro_hash(struct napi_struct *napi)
6955 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6956 struct sk_buff *skb, *n;
6958 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6960 napi->gro_hash[i].count = 0;
6964 /* Must be called in process context */
6965 void __netif_napi_del(struct napi_struct *napi)
6967 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
6970 napi_hash_del(napi);
6971 list_del_rcu(&napi->dev_list);
6972 napi_free_frags(napi);
6974 flush_gro_hash(napi);
6975 napi->gro_bitmask = 0;
6978 kthread_stop(napi->thread);
6979 napi->thread = NULL;
6982 EXPORT_SYMBOL(__netif_napi_del);
6984 static int __napi_poll(struct napi_struct *n, bool *repoll)
6990 /* This NAPI_STATE_SCHED test is for avoiding a race
6991 * with netpoll's poll_napi(). Only the entity which
6992 * obtains the lock and sees NAPI_STATE_SCHED set will
6993 * actually make the ->poll() call. Therefore we avoid
6994 * accidentally calling ->poll() when NAPI is not scheduled.
6997 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6998 work = n->poll(n, weight);
6999 trace_napi_poll(n, work, weight);
7002 if (unlikely(work > weight))
7003 pr_err_once("NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
7004 n->poll, work, weight);
7006 if (likely(work < weight))
7009 /* Drivers must not modify the NAPI state if they
7010 * consume the entire weight. In such cases this code
7011 * still "owns" the NAPI instance and therefore can
7012 * move the instance around on the list at-will.
7014 if (unlikely(napi_disable_pending(n))) {
7019 /* The NAPI context has more processing work, but busy-polling
7020 * is preferred. Exit early.
7022 if (napi_prefer_busy_poll(n)) {
7023 if (napi_complete_done(n, work)) {
7024 /* If timeout is not set, we need to make sure
7025 * that the NAPI is re-scheduled.
7032 if (n->gro_bitmask) {
7033 /* flush too old packets
7034 * If HZ < 1000, flush all packets.
7036 napi_gro_flush(n, HZ >= 1000);
7041 /* Some drivers may have called napi_schedule
7042 * prior to exhausting their budget.
7044 if (unlikely(!list_empty(&n->poll_list))) {
7045 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
7046 n->dev ? n->dev->name : "backlog");
7055 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
7057 bool do_repoll = false;
7061 list_del_init(&n->poll_list);
7063 have = netpoll_poll_lock(n);
7065 work = __napi_poll(n, &do_repoll);
7068 list_add_tail(&n->poll_list, repoll);
7070 netpoll_poll_unlock(have);
7075 static int napi_thread_wait(struct napi_struct *napi)
7079 set_current_state(TASK_INTERRUPTIBLE);
7081 while (!kthread_should_stop()) {
7082 /* Testing SCHED_THREADED bit here to make sure the current
7083 * kthread owns this napi and could poll on this napi.
7084 * Testing SCHED bit is not enough because SCHED bit might be
7085 * set by some other busy poll thread or by napi_disable().
7087 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
7088 WARN_ON(!list_empty(&napi->poll_list));
7089 __set_current_state(TASK_RUNNING);
7094 /* woken being true indicates this thread owns this napi. */
7096 set_current_state(TASK_INTERRUPTIBLE);
7098 __set_current_state(TASK_RUNNING);
7103 static int napi_threaded_poll(void *data)
7105 struct napi_struct *napi = data;
7108 while (!napi_thread_wait(napi)) {
7110 bool repoll = false;
7114 have = netpoll_poll_lock(napi);
7115 __napi_poll(napi, &repoll);
7116 netpoll_poll_unlock(have);
7129 static __latent_entropy void net_rx_action(struct softirq_action *h)
7131 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
7132 unsigned long time_limit = jiffies +
7133 usecs_to_jiffies(netdev_budget_usecs);
7134 int budget = netdev_budget;
7138 local_irq_disable();
7139 list_splice_init(&sd->poll_list, &list);
7143 struct napi_struct *n;
7145 if (list_empty(&list)) {
7146 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
7151 n = list_first_entry(&list, struct napi_struct, poll_list);
7152 budget -= napi_poll(n, &repoll);
7154 /* If softirq window is exhausted then punt.
7155 * Allow this to run for 2 jiffies since which will allow
7156 * an average latency of 1.5/HZ.
7158 if (unlikely(budget <= 0 ||
7159 time_after_eq(jiffies, time_limit))) {
7165 local_irq_disable();
7167 list_splice_tail_init(&sd->poll_list, &list);
7168 list_splice_tail(&repoll, &list);
7169 list_splice(&list, &sd->poll_list);
7170 if (!list_empty(&sd->poll_list))
7171 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
7173 net_rps_action_and_irq_enable(sd);
7176 struct netdev_adjacent {
7177 struct net_device *dev;
7179 /* upper master flag, there can only be one master device per list */
7182 /* lookup ignore flag */
7185 /* counter for the number of times this device was added to us */
7188 /* private field for the users */
7191 struct list_head list;
7192 struct rcu_head rcu;
7195 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
7196 struct list_head *adj_list)
7198 struct netdev_adjacent *adj;
7200 list_for_each_entry(adj, adj_list, list) {
7201 if (adj->dev == adj_dev)
7207 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
7208 struct netdev_nested_priv *priv)
7210 struct net_device *dev = (struct net_device *)priv->data;
7212 return upper_dev == dev;
7216 * netdev_has_upper_dev - Check if device is linked to an upper device
7218 * @upper_dev: upper device to check
7220 * Find out if a device is linked to specified upper device and return true
7221 * in case it is. Note that this checks only immediate upper device,
7222 * not through a complete stack of devices. The caller must hold the RTNL lock.
7224 bool netdev_has_upper_dev(struct net_device *dev,
7225 struct net_device *upper_dev)
7227 struct netdev_nested_priv priv = {
7228 .data = (void *)upper_dev,
7233 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7236 EXPORT_SYMBOL(netdev_has_upper_dev);
7239 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
7241 * @upper_dev: upper device to check
7243 * Find out if a device is linked to specified upper device and return true
7244 * in case it is. Note that this checks the entire upper device chain.
7245 * The caller must hold rcu lock.
7248 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
7249 struct net_device *upper_dev)
7251 struct netdev_nested_priv priv = {
7252 .data = (void *)upper_dev,
7255 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7258 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
7261 * netdev_has_any_upper_dev - Check if device is linked to some device
7264 * Find out if a device is linked to an upper device and return true in case
7265 * it is. The caller must hold the RTNL lock.
7267 bool netdev_has_any_upper_dev(struct net_device *dev)
7271 return !list_empty(&dev->adj_list.upper);
7273 EXPORT_SYMBOL(netdev_has_any_upper_dev);
7276 * netdev_master_upper_dev_get - Get master upper device
7279 * Find a master upper device and return pointer to it or NULL in case
7280 * it's not there. The caller must hold the RTNL lock.
7282 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
7284 struct netdev_adjacent *upper;
7288 if (list_empty(&dev->adj_list.upper))
7291 upper = list_first_entry(&dev->adj_list.upper,
7292 struct netdev_adjacent, list);
7293 if (likely(upper->master))
7297 EXPORT_SYMBOL(netdev_master_upper_dev_get);
7299 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
7301 struct netdev_adjacent *upper;
7305 if (list_empty(&dev->adj_list.upper))
7308 upper = list_first_entry(&dev->adj_list.upper,
7309 struct netdev_adjacent, list);
7310 if (likely(upper->master) && !upper->ignore)
7316 * netdev_has_any_lower_dev - Check if device is linked to some device
7319 * Find out if a device is linked to a lower device and return true in case
7320 * it is. The caller must hold the RTNL lock.
7322 static bool netdev_has_any_lower_dev(struct net_device *dev)
7326 return !list_empty(&dev->adj_list.lower);
7329 void *netdev_adjacent_get_private(struct list_head *adj_list)
7331 struct netdev_adjacent *adj;
7333 adj = list_entry(adj_list, struct netdev_adjacent, list);
7335 return adj->private;
7337 EXPORT_SYMBOL(netdev_adjacent_get_private);
7340 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
7342 * @iter: list_head ** of the current position
7344 * Gets the next device from the dev's upper list, starting from iter
7345 * position. The caller must hold RCU read lock.
7347 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
7348 struct list_head **iter)
7350 struct netdev_adjacent *upper;
7352 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7354 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7356 if (&upper->list == &dev->adj_list.upper)
7359 *iter = &upper->list;
7363 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
7365 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
7366 struct list_head **iter,
7369 struct netdev_adjacent *upper;
7371 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
7373 if (&upper->list == &dev->adj_list.upper)
7376 *iter = &upper->list;
7377 *ignore = upper->ignore;
7382 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
7383 struct list_head **iter)
7385 struct netdev_adjacent *upper;
7387 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7389 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7391 if (&upper->list == &dev->adj_list.upper)
7394 *iter = &upper->list;
7399 static int __netdev_walk_all_upper_dev(struct net_device *dev,
7400 int (*fn)(struct net_device *dev,
7401 struct netdev_nested_priv *priv),
7402 struct netdev_nested_priv *priv)
7404 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7405 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7410 iter = &dev->adj_list.upper;
7414 ret = fn(now, priv);
7421 udev = __netdev_next_upper_dev(now, &iter, &ignore);
7428 niter = &udev->adj_list.upper;
7429 dev_stack[cur] = now;
7430 iter_stack[cur++] = iter;
7437 next = dev_stack[--cur];
7438 niter = iter_stack[cur];
7448 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
7449 int (*fn)(struct net_device *dev,
7450 struct netdev_nested_priv *priv),
7451 struct netdev_nested_priv *priv)
7453 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7454 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7458 iter = &dev->adj_list.upper;
7462 ret = fn(now, priv);
7469 udev = netdev_next_upper_dev_rcu(now, &iter);
7474 niter = &udev->adj_list.upper;
7475 dev_stack[cur] = now;
7476 iter_stack[cur++] = iter;
7483 next = dev_stack[--cur];
7484 niter = iter_stack[cur];
7493 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7495 static bool __netdev_has_upper_dev(struct net_device *dev,
7496 struct net_device *upper_dev)
7498 struct netdev_nested_priv priv = {
7500 .data = (void *)upper_dev,
7505 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7510 * netdev_lower_get_next_private - Get the next ->private from the
7511 * lower neighbour list
7513 * @iter: list_head ** of the current position
7515 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7516 * list, starting from iter position. The caller must hold either hold the
7517 * RTNL lock or its own locking that guarantees that the neighbour lower
7518 * list will remain unchanged.
7520 void *netdev_lower_get_next_private(struct net_device *dev,
7521 struct list_head **iter)
7523 struct netdev_adjacent *lower;
7525 lower = list_entry(*iter, struct netdev_adjacent, list);
7527 if (&lower->list == &dev->adj_list.lower)
7530 *iter = lower->list.next;
7532 return lower->private;
7534 EXPORT_SYMBOL(netdev_lower_get_next_private);
7537 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7538 * lower neighbour list, RCU
7541 * @iter: list_head ** of the current position
7543 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7544 * list, starting from iter position. The caller must hold RCU read lock.
7546 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7547 struct list_head **iter)
7549 struct netdev_adjacent *lower;
7551 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
7553 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7555 if (&lower->list == &dev->adj_list.lower)
7558 *iter = &lower->list;
7560 return lower->private;
7562 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7565 * netdev_lower_get_next - Get the next device from the lower neighbour
7568 * @iter: list_head ** of the current position
7570 * Gets the next netdev_adjacent from the dev's lower neighbour
7571 * list, starting from iter position. The caller must hold RTNL lock or
7572 * its own locking that guarantees that the neighbour lower
7573 * list will remain unchanged.
7575 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7577 struct netdev_adjacent *lower;
7579 lower = list_entry(*iter, struct netdev_adjacent, list);
7581 if (&lower->list == &dev->adj_list.lower)
7584 *iter = lower->list.next;
7588 EXPORT_SYMBOL(netdev_lower_get_next);
7590 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7591 struct list_head **iter)
7593 struct netdev_adjacent *lower;
7595 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7597 if (&lower->list == &dev->adj_list.lower)
7600 *iter = &lower->list;
7605 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7606 struct list_head **iter,
7609 struct netdev_adjacent *lower;
7611 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7613 if (&lower->list == &dev->adj_list.lower)
7616 *iter = &lower->list;
7617 *ignore = lower->ignore;
7622 int netdev_walk_all_lower_dev(struct net_device *dev,
7623 int (*fn)(struct net_device *dev,
7624 struct netdev_nested_priv *priv),
7625 struct netdev_nested_priv *priv)
7627 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7628 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7632 iter = &dev->adj_list.lower;
7636 ret = fn(now, priv);
7643 ldev = netdev_next_lower_dev(now, &iter);
7648 niter = &ldev->adj_list.lower;
7649 dev_stack[cur] = now;
7650 iter_stack[cur++] = iter;
7657 next = dev_stack[--cur];
7658 niter = iter_stack[cur];
7667 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7669 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7670 int (*fn)(struct net_device *dev,
7671 struct netdev_nested_priv *priv),
7672 struct netdev_nested_priv *priv)
7674 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7675 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7680 iter = &dev->adj_list.lower;
7684 ret = fn(now, priv);
7691 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7698 niter = &ldev->adj_list.lower;
7699 dev_stack[cur] = now;
7700 iter_stack[cur++] = iter;
7707 next = dev_stack[--cur];
7708 niter = iter_stack[cur];
7718 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7719 struct list_head **iter)
7721 struct netdev_adjacent *lower;
7723 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7724 if (&lower->list == &dev->adj_list.lower)
7727 *iter = &lower->list;
7731 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7733 static u8 __netdev_upper_depth(struct net_device *dev)
7735 struct net_device *udev;
7736 struct list_head *iter;
7740 for (iter = &dev->adj_list.upper,
7741 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7743 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7746 if (max_depth < udev->upper_level)
7747 max_depth = udev->upper_level;
7753 static u8 __netdev_lower_depth(struct net_device *dev)
7755 struct net_device *ldev;
7756 struct list_head *iter;
7760 for (iter = &dev->adj_list.lower,
7761 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7763 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7766 if (max_depth < ldev->lower_level)
7767 max_depth = ldev->lower_level;
7773 static int __netdev_update_upper_level(struct net_device *dev,
7774 struct netdev_nested_priv *__unused)
7776 dev->upper_level = __netdev_upper_depth(dev) + 1;
7780 static int __netdev_update_lower_level(struct net_device *dev,
7781 struct netdev_nested_priv *priv)
7783 dev->lower_level = __netdev_lower_depth(dev) + 1;
7785 #ifdef CONFIG_LOCKDEP
7789 if (priv->flags & NESTED_SYNC_IMM)
7790 dev->nested_level = dev->lower_level - 1;
7791 if (priv->flags & NESTED_SYNC_TODO)
7792 net_unlink_todo(dev);
7797 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7798 int (*fn)(struct net_device *dev,
7799 struct netdev_nested_priv *priv),
7800 struct netdev_nested_priv *priv)
7802 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7803 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7807 iter = &dev->adj_list.lower;
7811 ret = fn(now, priv);
7818 ldev = netdev_next_lower_dev_rcu(now, &iter);
7823 niter = &ldev->adj_list.lower;
7824 dev_stack[cur] = now;
7825 iter_stack[cur++] = iter;
7832 next = dev_stack[--cur];
7833 niter = iter_stack[cur];
7842 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7845 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7846 * lower neighbour list, RCU
7850 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7851 * list. The caller must hold RCU read lock.
7853 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7855 struct netdev_adjacent *lower;
7857 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7858 struct netdev_adjacent, list);
7860 return lower->private;
7863 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7866 * netdev_master_upper_dev_get_rcu - Get master upper device
7869 * Find a master upper device and return pointer to it or NULL in case
7870 * it's not there. The caller must hold the RCU read lock.
7872 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7874 struct netdev_adjacent *upper;
7876 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7877 struct netdev_adjacent, list);
7878 if (upper && likely(upper->master))
7882 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7884 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7885 struct net_device *adj_dev,
7886 struct list_head *dev_list)
7888 char linkname[IFNAMSIZ+7];
7890 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7891 "upper_%s" : "lower_%s", adj_dev->name);
7892 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7895 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7897 struct list_head *dev_list)
7899 char linkname[IFNAMSIZ+7];
7901 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7902 "upper_%s" : "lower_%s", name);
7903 sysfs_remove_link(&(dev->dev.kobj), linkname);
7906 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7907 struct net_device *adj_dev,
7908 struct list_head *dev_list)
7910 return (dev_list == &dev->adj_list.upper ||
7911 dev_list == &dev->adj_list.lower) &&
7912 net_eq(dev_net(dev), dev_net(adj_dev));
7915 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7916 struct net_device *adj_dev,
7917 struct list_head *dev_list,
7918 void *private, bool master)
7920 struct netdev_adjacent *adj;
7923 adj = __netdev_find_adj(adj_dev, dev_list);
7927 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7928 dev->name, adj_dev->name, adj->ref_nr);
7933 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7938 adj->master = master;
7940 adj->private = private;
7941 adj->ignore = false;
7944 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7945 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7947 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7948 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7953 /* Ensure that master link is always the first item in list. */
7955 ret = sysfs_create_link(&(dev->dev.kobj),
7956 &(adj_dev->dev.kobj), "master");
7958 goto remove_symlinks;
7960 list_add_rcu(&adj->list, dev_list);
7962 list_add_tail_rcu(&adj->list, dev_list);
7968 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7969 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7977 static void __netdev_adjacent_dev_remove(struct net_device *dev,
7978 struct net_device *adj_dev,
7980 struct list_head *dev_list)
7982 struct netdev_adjacent *adj;
7984 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
7985 dev->name, adj_dev->name, ref_nr);
7987 adj = __netdev_find_adj(adj_dev, dev_list);
7990 pr_err("Adjacency does not exist for device %s from %s\n",
7991 dev->name, adj_dev->name);
7996 if (adj->ref_nr > ref_nr) {
7997 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
7998 dev->name, adj_dev->name, ref_nr,
7999 adj->ref_nr - ref_nr);
8000 adj->ref_nr -= ref_nr;
8005 sysfs_remove_link(&(dev->dev.kobj), "master");
8007 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
8008 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
8010 list_del_rcu(&adj->list);
8011 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
8012 adj_dev->name, dev->name, adj_dev->name);
8014 kfree_rcu(adj, rcu);
8017 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
8018 struct net_device *upper_dev,
8019 struct list_head *up_list,
8020 struct list_head *down_list,
8021 void *private, bool master)
8025 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
8030 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
8033 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
8040 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
8041 struct net_device *upper_dev,
8043 struct list_head *up_list,
8044 struct list_head *down_list)
8046 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
8047 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
8050 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
8051 struct net_device *upper_dev,
8052 void *private, bool master)
8054 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
8055 &dev->adj_list.upper,
8056 &upper_dev->adj_list.lower,
8060 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
8061 struct net_device *upper_dev)
8063 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
8064 &dev->adj_list.upper,
8065 &upper_dev->adj_list.lower);
8068 static int __netdev_upper_dev_link(struct net_device *dev,
8069 struct net_device *upper_dev, bool master,
8070 void *upper_priv, void *upper_info,
8071 struct netdev_nested_priv *priv,
8072 struct netlink_ext_ack *extack)
8074 struct netdev_notifier_changeupper_info changeupper_info = {
8079 .upper_dev = upper_dev,
8082 .upper_info = upper_info,
8084 struct net_device *master_dev;
8089 if (dev == upper_dev)
8092 /* To prevent loops, check if dev is not upper device to upper_dev. */
8093 if (__netdev_has_upper_dev(upper_dev, dev))
8096 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
8100 if (__netdev_has_upper_dev(dev, upper_dev))
8103 master_dev = __netdev_master_upper_dev_get(dev);
8105 return master_dev == upper_dev ? -EEXIST : -EBUSY;
8108 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8109 &changeupper_info.info);
8110 ret = notifier_to_errno(ret);
8114 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
8119 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8120 &changeupper_info.info);
8121 ret = notifier_to_errno(ret);
8125 __netdev_update_upper_level(dev, NULL);
8126 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8128 __netdev_update_lower_level(upper_dev, priv);
8129 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8135 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8141 * netdev_upper_dev_link - Add a link to the upper device
8143 * @upper_dev: new upper device
8144 * @extack: netlink extended ack
8146 * Adds a link to device which is upper to this one. The caller must hold
8147 * the RTNL lock. On a failure a negative errno code is returned.
8148 * On success the reference counts are adjusted and the function
8151 int netdev_upper_dev_link(struct net_device *dev,
8152 struct net_device *upper_dev,
8153 struct netlink_ext_ack *extack)
8155 struct netdev_nested_priv priv = {
8156 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8160 return __netdev_upper_dev_link(dev, upper_dev, false,
8161 NULL, NULL, &priv, extack);
8163 EXPORT_SYMBOL(netdev_upper_dev_link);
8166 * netdev_master_upper_dev_link - Add a master link to the upper device
8168 * @upper_dev: new upper device
8169 * @upper_priv: upper device private
8170 * @upper_info: upper info to be passed down via notifier
8171 * @extack: netlink extended ack
8173 * Adds a link to device which is upper to this one. In this case, only
8174 * one master upper device can be linked, although other non-master devices
8175 * might be linked as well. The caller must hold the RTNL lock.
8176 * On a failure a negative errno code is returned. On success the reference
8177 * counts are adjusted and the function returns zero.
8179 int netdev_master_upper_dev_link(struct net_device *dev,
8180 struct net_device *upper_dev,
8181 void *upper_priv, void *upper_info,
8182 struct netlink_ext_ack *extack)
8184 struct netdev_nested_priv priv = {
8185 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8189 return __netdev_upper_dev_link(dev, upper_dev, true,
8190 upper_priv, upper_info, &priv, extack);
8192 EXPORT_SYMBOL(netdev_master_upper_dev_link);
8194 static void __netdev_upper_dev_unlink(struct net_device *dev,
8195 struct net_device *upper_dev,
8196 struct netdev_nested_priv *priv)
8198 struct netdev_notifier_changeupper_info changeupper_info = {
8202 .upper_dev = upper_dev,
8208 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
8210 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8211 &changeupper_info.info);
8213 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8215 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8216 &changeupper_info.info);
8218 __netdev_update_upper_level(dev, NULL);
8219 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8221 __netdev_update_lower_level(upper_dev, priv);
8222 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8227 * netdev_upper_dev_unlink - Removes a link to upper device
8229 * @upper_dev: new upper device
8231 * Removes a link to device which is upper to this one. The caller must hold
8234 void netdev_upper_dev_unlink(struct net_device *dev,
8235 struct net_device *upper_dev)
8237 struct netdev_nested_priv priv = {
8238 .flags = NESTED_SYNC_TODO,
8242 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
8244 EXPORT_SYMBOL(netdev_upper_dev_unlink);
8246 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
8247 struct net_device *lower_dev,
8250 struct netdev_adjacent *adj;
8252 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
8256 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
8261 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
8262 struct net_device *lower_dev)
8264 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
8267 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
8268 struct net_device *lower_dev)
8270 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
8273 int netdev_adjacent_change_prepare(struct net_device *old_dev,
8274 struct net_device *new_dev,
8275 struct net_device *dev,
8276 struct netlink_ext_ack *extack)
8278 struct netdev_nested_priv priv = {
8287 if (old_dev && new_dev != old_dev)
8288 netdev_adjacent_dev_disable(dev, old_dev);
8289 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
8292 if (old_dev && new_dev != old_dev)
8293 netdev_adjacent_dev_enable(dev, old_dev);
8299 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
8301 void netdev_adjacent_change_commit(struct net_device *old_dev,
8302 struct net_device *new_dev,
8303 struct net_device *dev)
8305 struct netdev_nested_priv priv = {
8306 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8310 if (!new_dev || !old_dev)
8313 if (new_dev == old_dev)
8316 netdev_adjacent_dev_enable(dev, old_dev);
8317 __netdev_upper_dev_unlink(old_dev, dev, &priv);
8319 EXPORT_SYMBOL(netdev_adjacent_change_commit);
8321 void netdev_adjacent_change_abort(struct net_device *old_dev,
8322 struct net_device *new_dev,
8323 struct net_device *dev)
8325 struct netdev_nested_priv priv = {
8333 if (old_dev && new_dev != old_dev)
8334 netdev_adjacent_dev_enable(dev, old_dev);
8336 __netdev_upper_dev_unlink(new_dev, dev, &priv);
8338 EXPORT_SYMBOL(netdev_adjacent_change_abort);
8341 * netdev_bonding_info_change - Dispatch event about slave change
8343 * @bonding_info: info to dispatch
8345 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
8346 * The caller must hold the RTNL lock.
8348 void netdev_bonding_info_change(struct net_device *dev,
8349 struct netdev_bonding_info *bonding_info)
8351 struct netdev_notifier_bonding_info info = {
8355 memcpy(&info.bonding_info, bonding_info,
8356 sizeof(struct netdev_bonding_info));
8357 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
8360 EXPORT_SYMBOL(netdev_bonding_info_change);
8363 * netdev_get_xmit_slave - Get the xmit slave of master device
8366 * @all_slaves: assume all the slaves are active
8368 * The reference counters are not incremented so the caller must be
8369 * careful with locks. The caller must hold RCU lock.
8370 * %NULL is returned if no slave is found.
8373 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8374 struct sk_buff *skb,
8377 const struct net_device_ops *ops = dev->netdev_ops;
8379 if (!ops->ndo_get_xmit_slave)
8381 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8383 EXPORT_SYMBOL(netdev_get_xmit_slave);
8385 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8388 const struct net_device_ops *ops = dev->netdev_ops;
8390 if (!ops->ndo_sk_get_lower_dev)
8392 return ops->ndo_sk_get_lower_dev(dev, sk);
8396 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8400 * %NULL is returned if no lower device is found.
8403 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8406 struct net_device *lower;
8408 lower = netdev_sk_get_lower_dev(dev, sk);
8411 lower = netdev_sk_get_lower_dev(dev, sk);
8416 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8418 static void netdev_adjacent_add_links(struct net_device *dev)
8420 struct netdev_adjacent *iter;
8422 struct net *net = dev_net(dev);
8424 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8425 if (!net_eq(net, dev_net(iter->dev)))
8427 netdev_adjacent_sysfs_add(iter->dev, dev,
8428 &iter->dev->adj_list.lower);
8429 netdev_adjacent_sysfs_add(dev, iter->dev,
8430 &dev->adj_list.upper);
8433 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8434 if (!net_eq(net, dev_net(iter->dev)))
8436 netdev_adjacent_sysfs_add(iter->dev, dev,
8437 &iter->dev->adj_list.upper);
8438 netdev_adjacent_sysfs_add(dev, iter->dev,
8439 &dev->adj_list.lower);
8443 static void netdev_adjacent_del_links(struct net_device *dev)
8445 struct netdev_adjacent *iter;
8447 struct net *net = dev_net(dev);
8449 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8450 if (!net_eq(net, dev_net(iter->dev)))
8452 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8453 &iter->dev->adj_list.lower);
8454 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8455 &dev->adj_list.upper);
8458 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8459 if (!net_eq(net, dev_net(iter->dev)))
8461 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8462 &iter->dev->adj_list.upper);
8463 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8464 &dev->adj_list.lower);
8468 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8470 struct netdev_adjacent *iter;
8472 struct net *net = dev_net(dev);
8474 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8475 if (!net_eq(net, dev_net(iter->dev)))
8477 netdev_adjacent_sysfs_del(iter->dev, oldname,
8478 &iter->dev->adj_list.lower);
8479 netdev_adjacent_sysfs_add(iter->dev, dev,
8480 &iter->dev->adj_list.lower);
8483 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8484 if (!net_eq(net, dev_net(iter->dev)))
8486 netdev_adjacent_sysfs_del(iter->dev, oldname,
8487 &iter->dev->adj_list.upper);
8488 netdev_adjacent_sysfs_add(iter->dev, dev,
8489 &iter->dev->adj_list.upper);
8493 void *netdev_lower_dev_get_private(struct net_device *dev,
8494 struct net_device *lower_dev)
8496 struct netdev_adjacent *lower;
8500 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8504 return lower->private;
8506 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8510 * netdev_lower_state_changed - Dispatch event about lower device state change
8511 * @lower_dev: device
8512 * @lower_state_info: state to dispatch
8514 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8515 * The caller must hold the RTNL lock.
8517 void netdev_lower_state_changed(struct net_device *lower_dev,
8518 void *lower_state_info)
8520 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8521 .info.dev = lower_dev,
8525 changelowerstate_info.lower_state_info = lower_state_info;
8526 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8527 &changelowerstate_info.info);
8529 EXPORT_SYMBOL(netdev_lower_state_changed);
8531 static void dev_change_rx_flags(struct net_device *dev, int flags)
8533 const struct net_device_ops *ops = dev->netdev_ops;
8535 if (ops->ndo_change_rx_flags)
8536 ops->ndo_change_rx_flags(dev, flags);
8539 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8541 unsigned int old_flags = dev->flags;
8547 dev->flags |= IFF_PROMISC;
8548 dev->promiscuity += inc;
8549 if (dev->promiscuity == 0) {
8552 * If inc causes overflow, untouch promisc and return error.
8555 dev->flags &= ~IFF_PROMISC;
8557 dev->promiscuity -= inc;
8558 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
8563 if (dev->flags != old_flags) {
8564 pr_info("device %s %s promiscuous mode\n",
8566 dev->flags & IFF_PROMISC ? "entered" : "left");
8567 if (audit_enabled) {
8568 current_uid_gid(&uid, &gid);
8569 audit_log(audit_context(), GFP_ATOMIC,
8570 AUDIT_ANOM_PROMISCUOUS,
8571 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8572 dev->name, (dev->flags & IFF_PROMISC),
8573 (old_flags & IFF_PROMISC),
8574 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8575 from_kuid(&init_user_ns, uid),
8576 from_kgid(&init_user_ns, gid),
8577 audit_get_sessionid(current));
8580 dev_change_rx_flags(dev, IFF_PROMISC);
8583 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
8588 * dev_set_promiscuity - update promiscuity count on a device
8592 * Add or remove promiscuity from a device. While the count in the device
8593 * remains above zero the interface remains promiscuous. Once it hits zero
8594 * the device reverts back to normal filtering operation. A negative inc
8595 * value is used to drop promiscuity on the device.
8596 * Return 0 if successful or a negative errno code on error.
8598 int dev_set_promiscuity(struct net_device *dev, int inc)
8600 unsigned int old_flags = dev->flags;
8603 err = __dev_set_promiscuity(dev, inc, true);
8606 if (dev->flags != old_flags)
8607 dev_set_rx_mode(dev);
8610 EXPORT_SYMBOL(dev_set_promiscuity);
8612 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8614 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8618 dev->flags |= IFF_ALLMULTI;
8619 dev->allmulti += inc;
8620 if (dev->allmulti == 0) {
8623 * If inc causes overflow, untouch allmulti and return error.
8626 dev->flags &= ~IFF_ALLMULTI;
8628 dev->allmulti -= inc;
8629 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
8634 if (dev->flags ^ old_flags) {
8635 dev_change_rx_flags(dev, IFF_ALLMULTI);
8636 dev_set_rx_mode(dev);
8638 __dev_notify_flags(dev, old_flags,
8639 dev->gflags ^ old_gflags);
8645 * dev_set_allmulti - update allmulti count on a device
8649 * Add or remove reception of all multicast frames to a device. While the
8650 * count in the device remains above zero the interface remains listening
8651 * to all interfaces. Once it hits zero the device reverts back to normal
8652 * filtering operation. A negative @inc value is used to drop the counter
8653 * when releasing a resource needing all multicasts.
8654 * Return 0 if successful or a negative errno code on error.
8657 int dev_set_allmulti(struct net_device *dev, int inc)
8659 return __dev_set_allmulti(dev, inc, true);
8661 EXPORT_SYMBOL(dev_set_allmulti);
8664 * Upload unicast and multicast address lists to device and
8665 * configure RX filtering. When the device doesn't support unicast
8666 * filtering it is put in promiscuous mode while unicast addresses
8669 void __dev_set_rx_mode(struct net_device *dev)
8671 const struct net_device_ops *ops = dev->netdev_ops;
8673 /* dev_open will call this function so the list will stay sane. */
8674 if (!(dev->flags&IFF_UP))
8677 if (!netif_device_present(dev))
8680 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8681 /* Unicast addresses changes may only happen under the rtnl,
8682 * therefore calling __dev_set_promiscuity here is safe.
8684 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8685 __dev_set_promiscuity(dev, 1, false);
8686 dev->uc_promisc = true;
8687 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8688 __dev_set_promiscuity(dev, -1, false);
8689 dev->uc_promisc = false;
8693 if (ops->ndo_set_rx_mode)
8694 ops->ndo_set_rx_mode(dev);
8697 void dev_set_rx_mode(struct net_device *dev)
8699 netif_addr_lock_bh(dev);
8700 __dev_set_rx_mode(dev);
8701 netif_addr_unlock_bh(dev);
8705 * dev_get_flags - get flags reported to userspace
8708 * Get the combination of flag bits exported through APIs to userspace.
8710 unsigned int dev_get_flags(const struct net_device *dev)
8714 flags = (dev->flags & ~(IFF_PROMISC |
8719 (dev->gflags & (IFF_PROMISC |
8722 if (netif_running(dev)) {
8723 if (netif_oper_up(dev))
8724 flags |= IFF_RUNNING;
8725 if (netif_carrier_ok(dev))
8726 flags |= IFF_LOWER_UP;
8727 if (netif_dormant(dev))
8728 flags |= IFF_DORMANT;
8733 EXPORT_SYMBOL(dev_get_flags);
8735 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8736 struct netlink_ext_ack *extack)
8738 unsigned int old_flags = dev->flags;
8744 * Set the flags on our device.
8747 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8748 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8750 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8754 * Load in the correct multicast list now the flags have changed.
8757 if ((old_flags ^ flags) & IFF_MULTICAST)
8758 dev_change_rx_flags(dev, IFF_MULTICAST);
8760 dev_set_rx_mode(dev);
8763 * Have we downed the interface. We handle IFF_UP ourselves
8764 * according to user attempts to set it, rather than blindly
8769 if ((old_flags ^ flags) & IFF_UP) {
8770 if (old_flags & IFF_UP)
8773 ret = __dev_open(dev, extack);
8776 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8777 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8778 unsigned int old_flags = dev->flags;
8780 dev->gflags ^= IFF_PROMISC;
8782 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8783 if (dev->flags != old_flags)
8784 dev_set_rx_mode(dev);
8787 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8788 * is important. Some (broken) drivers set IFF_PROMISC, when
8789 * IFF_ALLMULTI is requested not asking us and not reporting.
8791 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8792 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8794 dev->gflags ^= IFF_ALLMULTI;
8795 __dev_set_allmulti(dev, inc, false);
8801 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8802 unsigned int gchanges)
8804 unsigned int changes = dev->flags ^ old_flags;
8807 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
8809 if (changes & IFF_UP) {
8810 if (dev->flags & IFF_UP)
8811 call_netdevice_notifiers(NETDEV_UP, dev);
8813 call_netdevice_notifiers(NETDEV_DOWN, dev);
8816 if (dev->flags & IFF_UP &&
8817 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8818 struct netdev_notifier_change_info change_info = {
8822 .flags_changed = changes,
8825 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8830 * dev_change_flags - change device settings
8832 * @flags: device state flags
8833 * @extack: netlink extended ack
8835 * Change settings on device based state flags. The flags are
8836 * in the userspace exported format.
8838 int dev_change_flags(struct net_device *dev, unsigned int flags,
8839 struct netlink_ext_ack *extack)
8842 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8844 ret = __dev_change_flags(dev, flags, extack);
8848 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8849 __dev_notify_flags(dev, old_flags, changes);
8852 EXPORT_SYMBOL(dev_change_flags);
8854 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8856 const struct net_device_ops *ops = dev->netdev_ops;
8858 if (ops->ndo_change_mtu)
8859 return ops->ndo_change_mtu(dev, new_mtu);
8861 /* Pairs with all the lockless reads of dev->mtu in the stack */
8862 WRITE_ONCE(dev->mtu, new_mtu);
8865 EXPORT_SYMBOL(__dev_set_mtu);
8867 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8868 struct netlink_ext_ack *extack)
8870 /* MTU must be positive, and in range */
8871 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8872 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8876 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8877 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8884 * dev_set_mtu_ext - Change maximum transfer unit
8886 * @new_mtu: new transfer unit
8887 * @extack: netlink extended ack
8889 * Change the maximum transfer size of the network device.
8891 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8892 struct netlink_ext_ack *extack)
8896 if (new_mtu == dev->mtu)
8899 err = dev_validate_mtu(dev, new_mtu, extack);
8903 if (!netif_device_present(dev))
8906 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8907 err = notifier_to_errno(err);
8911 orig_mtu = dev->mtu;
8912 err = __dev_set_mtu(dev, new_mtu);
8915 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8917 err = notifier_to_errno(err);
8919 /* setting mtu back and notifying everyone again,
8920 * so that they have a chance to revert changes.
8922 __dev_set_mtu(dev, orig_mtu);
8923 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8930 int dev_set_mtu(struct net_device *dev, int new_mtu)
8932 struct netlink_ext_ack extack;
8935 memset(&extack, 0, sizeof(extack));
8936 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8937 if (err && extack._msg)
8938 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8941 EXPORT_SYMBOL(dev_set_mtu);
8944 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8946 * @new_len: new tx queue length
8948 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8950 unsigned int orig_len = dev->tx_queue_len;
8953 if (new_len != (unsigned int)new_len)
8956 if (new_len != orig_len) {
8957 dev->tx_queue_len = new_len;
8958 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8959 res = notifier_to_errno(res);
8962 res = dev_qdisc_change_tx_queue_len(dev);
8970 netdev_err(dev, "refused to change device tx_queue_len\n");
8971 dev->tx_queue_len = orig_len;
8976 * dev_set_group - Change group this device belongs to
8978 * @new_group: group this device should belong to
8980 void dev_set_group(struct net_device *dev, int new_group)
8982 dev->group = new_group;
8984 EXPORT_SYMBOL(dev_set_group);
8987 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
8989 * @addr: new address
8990 * @extack: netlink extended ack
8992 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
8993 struct netlink_ext_ack *extack)
8995 struct netdev_notifier_pre_changeaddr_info info = {
8997 .info.extack = extack,
9002 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
9003 return notifier_to_errno(rc);
9005 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
9008 * dev_set_mac_address - Change Media Access Control Address
9011 * @extack: netlink extended ack
9013 * Change the hardware (MAC) address of the device
9015 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
9016 struct netlink_ext_ack *extack)
9018 const struct net_device_ops *ops = dev->netdev_ops;
9021 if (!ops->ndo_set_mac_address)
9023 if (sa->sa_family != dev->type)
9025 if (!netif_device_present(dev))
9027 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
9030 err = ops->ndo_set_mac_address(dev, sa);
9033 dev->addr_assign_type = NET_ADDR_SET;
9034 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
9035 add_device_randomness(dev->dev_addr, dev->addr_len);
9038 EXPORT_SYMBOL(dev_set_mac_address);
9040 static DECLARE_RWSEM(dev_addr_sem);
9042 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
9043 struct netlink_ext_ack *extack)
9047 down_write(&dev_addr_sem);
9048 ret = dev_set_mac_address(dev, sa, extack);
9049 up_write(&dev_addr_sem);
9052 EXPORT_SYMBOL(dev_set_mac_address_user);
9054 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
9056 size_t size = sizeof(sa->sa_data);
9057 struct net_device *dev;
9060 down_read(&dev_addr_sem);
9063 dev = dev_get_by_name_rcu(net, dev_name);
9069 memset(sa->sa_data, 0, size);
9071 memcpy(sa->sa_data, dev->dev_addr,
9072 min_t(size_t, size, dev->addr_len));
9073 sa->sa_family = dev->type;
9077 up_read(&dev_addr_sem);
9080 EXPORT_SYMBOL(dev_get_mac_address);
9083 * dev_change_carrier - Change device carrier
9085 * @new_carrier: new value
9087 * Change device carrier
9089 int dev_change_carrier(struct net_device *dev, bool new_carrier)
9091 const struct net_device_ops *ops = dev->netdev_ops;
9093 if (!ops->ndo_change_carrier)
9095 if (!netif_device_present(dev))
9097 return ops->ndo_change_carrier(dev, new_carrier);
9099 EXPORT_SYMBOL(dev_change_carrier);
9102 * dev_get_phys_port_id - Get device physical port ID
9106 * Get device physical port ID
9108 int dev_get_phys_port_id(struct net_device *dev,
9109 struct netdev_phys_item_id *ppid)
9111 const struct net_device_ops *ops = dev->netdev_ops;
9113 if (!ops->ndo_get_phys_port_id)
9115 return ops->ndo_get_phys_port_id(dev, ppid);
9117 EXPORT_SYMBOL(dev_get_phys_port_id);
9120 * dev_get_phys_port_name - Get device physical port name
9123 * @len: limit of bytes to copy to name
9125 * Get device physical port name
9127 int dev_get_phys_port_name(struct net_device *dev,
9128 char *name, size_t len)
9130 const struct net_device_ops *ops = dev->netdev_ops;
9133 if (ops->ndo_get_phys_port_name) {
9134 err = ops->ndo_get_phys_port_name(dev, name, len);
9135 if (err != -EOPNOTSUPP)
9138 return devlink_compat_phys_port_name_get(dev, name, len);
9140 EXPORT_SYMBOL(dev_get_phys_port_name);
9143 * dev_get_port_parent_id - Get the device's port parent identifier
9144 * @dev: network device
9145 * @ppid: pointer to a storage for the port's parent identifier
9146 * @recurse: allow/disallow recursion to lower devices
9148 * Get the devices's port parent identifier
9150 int dev_get_port_parent_id(struct net_device *dev,
9151 struct netdev_phys_item_id *ppid,
9154 const struct net_device_ops *ops = dev->netdev_ops;
9155 struct netdev_phys_item_id first = { };
9156 struct net_device *lower_dev;
9157 struct list_head *iter;
9160 if (ops->ndo_get_port_parent_id) {
9161 err = ops->ndo_get_port_parent_id(dev, ppid);
9162 if (err != -EOPNOTSUPP)
9166 err = devlink_compat_switch_id_get(dev, ppid);
9167 if (!err || err != -EOPNOTSUPP)
9173 netdev_for_each_lower_dev(dev, lower_dev, iter) {
9174 err = dev_get_port_parent_id(lower_dev, ppid, recurse);
9179 else if (memcmp(&first, ppid, sizeof(*ppid)))
9185 EXPORT_SYMBOL(dev_get_port_parent_id);
9188 * netdev_port_same_parent_id - Indicate if two network devices have
9189 * the same port parent identifier
9190 * @a: first network device
9191 * @b: second network device
9193 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
9195 struct netdev_phys_item_id a_id = { };
9196 struct netdev_phys_item_id b_id = { };
9198 if (dev_get_port_parent_id(a, &a_id, true) ||
9199 dev_get_port_parent_id(b, &b_id, true))
9202 return netdev_phys_item_id_same(&a_id, &b_id);
9204 EXPORT_SYMBOL(netdev_port_same_parent_id);
9207 * dev_change_proto_down - update protocol port state information
9209 * @proto_down: new value
9211 * This info can be used by switch drivers to set the phys state of the
9214 int dev_change_proto_down(struct net_device *dev, bool proto_down)
9216 const struct net_device_ops *ops = dev->netdev_ops;
9218 if (!ops->ndo_change_proto_down)
9220 if (!netif_device_present(dev))
9222 return ops->ndo_change_proto_down(dev, proto_down);
9224 EXPORT_SYMBOL(dev_change_proto_down);
9227 * dev_change_proto_down_generic - generic implementation for
9228 * ndo_change_proto_down that sets carrier according to
9232 * @proto_down: new value
9234 int dev_change_proto_down_generic(struct net_device *dev, bool proto_down)
9237 netif_carrier_off(dev);
9239 netif_carrier_on(dev);
9240 dev->proto_down = proto_down;
9243 EXPORT_SYMBOL(dev_change_proto_down_generic);
9246 * dev_change_proto_down_reason - proto down reason
9249 * @mask: proto down mask
9250 * @value: proto down value
9252 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
9258 dev->proto_down_reason = value;
9260 for_each_set_bit(b, &mask, 32) {
9261 if (value & (1 << b))
9262 dev->proto_down_reason |= BIT(b);
9264 dev->proto_down_reason &= ~BIT(b);
9268 EXPORT_SYMBOL(dev_change_proto_down_reason);
9270 struct bpf_xdp_link {
9271 struct bpf_link link;
9272 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9276 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9278 if (flags & XDP_FLAGS_HW_MODE)
9280 if (flags & XDP_FLAGS_DRV_MODE)
9281 return XDP_MODE_DRV;
9282 if (flags & XDP_FLAGS_SKB_MODE)
9283 return XDP_MODE_SKB;
9284 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9287 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9291 return generic_xdp_install;
9294 return dev->netdev_ops->ndo_bpf;
9300 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9301 enum bpf_xdp_mode mode)
9303 return dev->xdp_state[mode].link;
9306 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9307 enum bpf_xdp_mode mode)
9309 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9312 return link->link.prog;
9313 return dev->xdp_state[mode].prog;
9316 u8 dev_xdp_prog_count(struct net_device *dev)
9321 for (i = 0; i < __MAX_XDP_MODE; i++)
9322 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9326 EXPORT_SYMBOL_GPL(dev_xdp_prog_count);
9328 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9330 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9332 return prog ? prog->aux->id : 0;
9335 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9336 struct bpf_xdp_link *link)
9338 dev->xdp_state[mode].link = link;
9339 dev->xdp_state[mode].prog = NULL;
9342 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9343 struct bpf_prog *prog)
9345 dev->xdp_state[mode].link = NULL;
9346 dev->xdp_state[mode].prog = prog;
9349 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9350 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9351 u32 flags, struct bpf_prog *prog)
9353 struct netdev_bpf xdp;
9356 memset(&xdp, 0, sizeof(xdp));
9357 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9358 xdp.extack = extack;
9362 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9363 * "moved" into driver), so they don't increment it on their own, but
9364 * they do decrement refcnt when program is detached or replaced.
9365 * Given net_device also owns link/prog, we need to bump refcnt here
9366 * to prevent drivers from underflowing it.
9370 err = bpf_op(dev, &xdp);
9377 if (mode != XDP_MODE_HW)
9378 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9383 static void dev_xdp_uninstall(struct net_device *dev)
9385 struct bpf_xdp_link *link;
9386 struct bpf_prog *prog;
9387 enum bpf_xdp_mode mode;
9392 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9393 prog = dev_xdp_prog(dev, mode);
9397 bpf_op = dev_xdp_bpf_op(dev, mode);
9401 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9403 /* auto-detach link from net device */
9404 link = dev_xdp_link(dev, mode);
9410 dev_xdp_set_link(dev, mode, NULL);
9414 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9415 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9416 struct bpf_prog *old_prog, u32 flags)
9418 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9419 struct bpf_prog *cur_prog;
9420 struct net_device *upper;
9421 struct list_head *iter;
9422 enum bpf_xdp_mode mode;
9428 /* either link or prog attachment, never both */
9429 if (link && (new_prog || old_prog))
9431 /* link supports only XDP mode flags */
9432 if (link && (flags & ~XDP_FLAGS_MODES)) {
9433 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9436 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9437 if (num_modes > 1) {
9438 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9441 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9442 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9443 NL_SET_ERR_MSG(extack,
9444 "More than one program loaded, unset mode is ambiguous");
9447 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9448 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9449 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9453 mode = dev_xdp_mode(dev, flags);
9454 /* can't replace attached link */
9455 if (dev_xdp_link(dev, mode)) {
9456 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9460 /* don't allow if an upper device already has a program */
9461 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
9462 if (dev_xdp_prog_count(upper) > 0) {
9463 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program");
9468 cur_prog = dev_xdp_prog(dev, mode);
9469 /* can't replace attached prog with link */
9470 if (link && cur_prog) {
9471 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9474 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9475 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9479 /* put effective new program into new_prog */
9481 new_prog = link->link.prog;
9484 bool offload = mode == XDP_MODE_HW;
9485 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9486 ? XDP_MODE_DRV : XDP_MODE_SKB;
9488 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9489 NL_SET_ERR_MSG(extack, "XDP program already attached");
9492 if (!offload && dev_xdp_prog(dev, other_mode)) {
9493 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9496 if (!offload && bpf_prog_is_dev_bound(new_prog->aux)) {
9497 NL_SET_ERR_MSG(extack, "Using device-bound program without HW_MODE flag is not supported");
9500 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9501 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9504 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9505 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9510 /* don't call drivers if the effective program didn't change */
9511 if (new_prog != cur_prog) {
9512 bpf_op = dev_xdp_bpf_op(dev, mode);
9514 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9518 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9524 dev_xdp_set_link(dev, mode, link);
9526 dev_xdp_set_prog(dev, mode, new_prog);
9528 bpf_prog_put(cur_prog);
9533 static int dev_xdp_attach_link(struct net_device *dev,
9534 struct netlink_ext_ack *extack,
9535 struct bpf_xdp_link *link)
9537 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9540 static int dev_xdp_detach_link(struct net_device *dev,
9541 struct netlink_ext_ack *extack,
9542 struct bpf_xdp_link *link)
9544 enum bpf_xdp_mode mode;
9549 mode = dev_xdp_mode(dev, link->flags);
9550 if (dev_xdp_link(dev, mode) != link)
9553 bpf_op = dev_xdp_bpf_op(dev, mode);
9554 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9555 dev_xdp_set_link(dev, mode, NULL);
9559 static void bpf_xdp_link_release(struct bpf_link *link)
9561 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9565 /* if racing with net_device's tear down, xdp_link->dev might be
9566 * already NULL, in which case link was already auto-detached
9568 if (xdp_link->dev) {
9569 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9570 xdp_link->dev = NULL;
9576 static int bpf_xdp_link_detach(struct bpf_link *link)
9578 bpf_xdp_link_release(link);
9582 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9584 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9589 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9590 struct seq_file *seq)
9592 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9597 ifindex = xdp_link->dev->ifindex;
9600 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9603 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9604 struct bpf_link_info *info)
9606 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9611 ifindex = xdp_link->dev->ifindex;
9614 info->xdp.ifindex = ifindex;
9618 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9619 struct bpf_prog *old_prog)
9621 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9622 enum bpf_xdp_mode mode;
9628 /* link might have been auto-released already, so fail */
9629 if (!xdp_link->dev) {
9634 if (old_prog && link->prog != old_prog) {
9638 old_prog = link->prog;
9639 if (old_prog->type != new_prog->type ||
9640 old_prog->expected_attach_type != new_prog->expected_attach_type) {
9645 if (old_prog == new_prog) {
9646 /* no-op, don't disturb drivers */
9647 bpf_prog_put(new_prog);
9651 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9652 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9653 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9654 xdp_link->flags, new_prog);
9658 old_prog = xchg(&link->prog, new_prog);
9659 bpf_prog_put(old_prog);
9666 static const struct bpf_link_ops bpf_xdp_link_lops = {
9667 .release = bpf_xdp_link_release,
9668 .dealloc = bpf_xdp_link_dealloc,
9669 .detach = bpf_xdp_link_detach,
9670 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9671 .fill_link_info = bpf_xdp_link_fill_link_info,
9672 .update_prog = bpf_xdp_link_update,
9675 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9677 struct net *net = current->nsproxy->net_ns;
9678 struct bpf_link_primer link_primer;
9679 struct bpf_xdp_link *link;
9680 struct net_device *dev;
9684 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9690 link = kzalloc(sizeof(*link), GFP_USER);
9696 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9698 link->flags = attr->link_create.flags;
9700 err = bpf_link_prime(&link->link, &link_primer);
9706 err = dev_xdp_attach_link(dev, NULL, link);
9711 bpf_link_cleanup(&link_primer);
9715 fd = bpf_link_settle(&link_primer);
9716 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9729 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9731 * @extack: netlink extended ack
9732 * @fd: new program fd or negative value to clear
9733 * @expected_fd: old program fd that userspace expects to replace or clear
9734 * @flags: xdp-related flags
9736 * Set or clear a bpf program for a device
9738 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9739 int fd, int expected_fd, u32 flags)
9741 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9742 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9748 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9749 mode != XDP_MODE_SKB);
9750 if (IS_ERR(new_prog))
9751 return PTR_ERR(new_prog);
9754 if (expected_fd >= 0) {
9755 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9756 mode != XDP_MODE_SKB);
9757 if (IS_ERR(old_prog)) {
9758 err = PTR_ERR(old_prog);
9764 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9767 if (err && new_prog)
9768 bpf_prog_put(new_prog);
9770 bpf_prog_put(old_prog);
9775 * dev_new_index - allocate an ifindex
9776 * @net: the applicable net namespace
9778 * Returns a suitable unique value for a new device interface
9779 * number. The caller must hold the rtnl semaphore or the
9780 * dev_base_lock to be sure it remains unique.
9782 static int dev_new_index(struct net *net)
9784 int ifindex = net->ifindex;
9789 if (!__dev_get_by_index(net, ifindex))
9790 return net->ifindex = ifindex;
9794 /* Delayed registration/unregisteration */
9795 static LIST_HEAD(net_todo_list);
9796 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9798 static void net_set_todo(struct net_device *dev)
9800 list_add_tail(&dev->todo_list, &net_todo_list);
9801 dev_net(dev)->dev_unreg_count++;
9804 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9805 struct net_device *upper, netdev_features_t features)
9807 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9808 netdev_features_t feature;
9811 for_each_netdev_feature(upper_disables, feature_bit) {
9812 feature = __NETIF_F_BIT(feature_bit);
9813 if (!(upper->wanted_features & feature)
9814 && (features & feature)) {
9815 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9816 &feature, upper->name);
9817 features &= ~feature;
9824 static void netdev_sync_lower_features(struct net_device *upper,
9825 struct net_device *lower, netdev_features_t features)
9827 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9828 netdev_features_t feature;
9831 for_each_netdev_feature(upper_disables, feature_bit) {
9832 feature = __NETIF_F_BIT(feature_bit);
9833 if (!(features & feature) && (lower->features & feature)) {
9834 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9835 &feature, lower->name);
9836 lower->wanted_features &= ~feature;
9837 __netdev_update_features(lower);
9839 if (unlikely(lower->features & feature))
9840 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9841 &feature, lower->name);
9843 netdev_features_change(lower);
9848 static netdev_features_t netdev_fix_features(struct net_device *dev,
9849 netdev_features_t features)
9851 /* Fix illegal checksum combinations */
9852 if ((features & NETIF_F_HW_CSUM) &&
9853 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9854 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9855 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9858 /* TSO requires that SG is present as well. */
9859 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9860 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9861 features &= ~NETIF_F_ALL_TSO;
9864 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9865 !(features & NETIF_F_IP_CSUM)) {
9866 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9867 features &= ~NETIF_F_TSO;
9868 features &= ~NETIF_F_TSO_ECN;
9871 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9872 !(features & NETIF_F_IPV6_CSUM)) {
9873 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9874 features &= ~NETIF_F_TSO6;
9877 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9878 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9879 features &= ~NETIF_F_TSO_MANGLEID;
9881 /* TSO ECN requires that TSO is present as well. */
9882 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9883 features &= ~NETIF_F_TSO_ECN;
9885 /* Software GSO depends on SG. */
9886 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9887 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9888 features &= ~NETIF_F_GSO;
9891 /* GSO partial features require GSO partial be set */
9892 if ((features & dev->gso_partial_features) &&
9893 !(features & NETIF_F_GSO_PARTIAL)) {
9895 "Dropping partially supported GSO features since no GSO partial.\n");
9896 features &= ~dev->gso_partial_features;
9899 if (!(features & NETIF_F_RXCSUM)) {
9900 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9901 * successfully merged by hardware must also have the
9902 * checksum verified by hardware. If the user does not
9903 * want to enable RXCSUM, logically, we should disable GRO_HW.
9905 if (features & NETIF_F_GRO_HW) {
9906 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9907 features &= ~NETIF_F_GRO_HW;
9911 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9912 if (features & NETIF_F_RXFCS) {
9913 if (features & NETIF_F_LRO) {
9914 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9915 features &= ~NETIF_F_LRO;
9918 if (features & NETIF_F_GRO_HW) {
9919 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9920 features &= ~NETIF_F_GRO_HW;
9924 if (features & NETIF_F_HW_TLS_TX) {
9925 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9926 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9927 bool hw_csum = features & NETIF_F_HW_CSUM;
9929 if (!ip_csum && !hw_csum) {
9930 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9931 features &= ~NETIF_F_HW_TLS_TX;
9935 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9936 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9937 features &= ~NETIF_F_HW_TLS_RX;
9943 int __netdev_update_features(struct net_device *dev)
9945 struct net_device *upper, *lower;
9946 netdev_features_t features;
9947 struct list_head *iter;
9952 features = netdev_get_wanted_features(dev);
9954 if (dev->netdev_ops->ndo_fix_features)
9955 features = dev->netdev_ops->ndo_fix_features(dev, features);
9957 /* driver might be less strict about feature dependencies */
9958 features = netdev_fix_features(dev, features);
9960 /* some features can't be enabled if they're off on an upper device */
9961 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9962 features = netdev_sync_upper_features(dev, upper, features);
9964 if (dev->features == features)
9967 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9968 &dev->features, &features);
9970 if (dev->netdev_ops->ndo_set_features)
9971 err = dev->netdev_ops->ndo_set_features(dev, features);
9975 if (unlikely(err < 0)) {
9977 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9978 err, &features, &dev->features);
9979 /* return non-0 since some features might have changed and
9980 * it's better to fire a spurious notification than miss it
9986 /* some features must be disabled on lower devices when disabled
9987 * on an upper device (think: bonding master or bridge)
9989 netdev_for_each_lower_dev(dev, lower, iter)
9990 netdev_sync_lower_features(dev, lower, features);
9993 netdev_features_t diff = features ^ dev->features;
9995 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
9996 /* udp_tunnel_{get,drop}_rx_info both need
9997 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
9998 * device, or they won't do anything.
9999 * Thus we need to update dev->features
10000 * *before* calling udp_tunnel_get_rx_info,
10001 * but *after* calling udp_tunnel_drop_rx_info.
10003 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
10004 dev->features = features;
10005 udp_tunnel_get_rx_info(dev);
10007 udp_tunnel_drop_rx_info(dev);
10011 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
10012 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
10013 dev->features = features;
10014 err |= vlan_get_rx_ctag_filter_info(dev);
10016 vlan_drop_rx_ctag_filter_info(dev);
10020 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
10021 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
10022 dev->features = features;
10023 err |= vlan_get_rx_stag_filter_info(dev);
10025 vlan_drop_rx_stag_filter_info(dev);
10029 dev->features = features;
10032 return err < 0 ? 0 : 1;
10036 * netdev_update_features - recalculate device features
10037 * @dev: the device to check
10039 * Recalculate dev->features set and send notifications if it
10040 * has changed. Should be called after driver or hardware dependent
10041 * conditions might have changed that influence the features.
10043 void netdev_update_features(struct net_device *dev)
10045 if (__netdev_update_features(dev))
10046 netdev_features_change(dev);
10048 EXPORT_SYMBOL(netdev_update_features);
10051 * netdev_change_features - recalculate device features
10052 * @dev: the device to check
10054 * Recalculate dev->features set and send notifications even
10055 * if they have not changed. Should be called instead of
10056 * netdev_update_features() if also dev->vlan_features might
10057 * have changed to allow the changes to be propagated to stacked
10060 void netdev_change_features(struct net_device *dev)
10062 __netdev_update_features(dev);
10063 netdev_features_change(dev);
10065 EXPORT_SYMBOL(netdev_change_features);
10068 * netif_stacked_transfer_operstate - transfer operstate
10069 * @rootdev: the root or lower level device to transfer state from
10070 * @dev: the device to transfer operstate to
10072 * Transfer operational state from root to device. This is normally
10073 * called when a stacking relationship exists between the root
10074 * device and the device(a leaf device).
10076 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
10077 struct net_device *dev)
10079 if (rootdev->operstate == IF_OPER_DORMANT)
10080 netif_dormant_on(dev);
10082 netif_dormant_off(dev);
10084 if (rootdev->operstate == IF_OPER_TESTING)
10085 netif_testing_on(dev);
10087 netif_testing_off(dev);
10089 if (netif_carrier_ok(rootdev))
10090 netif_carrier_on(dev);
10092 netif_carrier_off(dev);
10094 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
10096 static int netif_alloc_rx_queues(struct net_device *dev)
10098 unsigned int i, count = dev->num_rx_queues;
10099 struct netdev_rx_queue *rx;
10100 size_t sz = count * sizeof(*rx);
10105 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10111 for (i = 0; i < count; i++) {
10114 /* XDP RX-queue setup */
10115 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
10122 /* Rollback successful reg's and free other resources */
10124 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
10130 static void netif_free_rx_queues(struct net_device *dev)
10132 unsigned int i, count = dev->num_rx_queues;
10134 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
10138 for (i = 0; i < count; i++)
10139 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
10144 static void netdev_init_one_queue(struct net_device *dev,
10145 struct netdev_queue *queue, void *_unused)
10147 /* Initialize queue lock */
10148 spin_lock_init(&queue->_xmit_lock);
10149 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
10150 queue->xmit_lock_owner = -1;
10151 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
10154 dql_init(&queue->dql, HZ);
10158 static void netif_free_tx_queues(struct net_device *dev)
10163 static int netif_alloc_netdev_queues(struct net_device *dev)
10165 unsigned int count = dev->num_tx_queues;
10166 struct netdev_queue *tx;
10167 size_t sz = count * sizeof(*tx);
10169 if (count < 1 || count > 0xffff)
10172 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10178 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
10179 spin_lock_init(&dev->tx_global_lock);
10184 void netif_tx_stop_all_queues(struct net_device *dev)
10188 for (i = 0; i < dev->num_tx_queues; i++) {
10189 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
10191 netif_tx_stop_queue(txq);
10194 EXPORT_SYMBOL(netif_tx_stop_all_queues);
10197 * register_netdevice - register a network device
10198 * @dev: device to register
10200 * Take a completed network device structure and add it to the kernel
10201 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10202 * chain. 0 is returned on success. A negative errno code is returned
10203 * on a failure to set up the device, or if the name is a duplicate.
10205 * Callers must hold the rtnl semaphore. You may want
10206 * register_netdev() instead of this.
10209 * The locking appears insufficient to guarantee two parallel registers
10210 * will not get the same name.
10213 int register_netdevice(struct net_device *dev)
10216 struct net *net = dev_net(dev);
10218 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
10219 NETDEV_FEATURE_COUNT);
10220 BUG_ON(dev_boot_phase);
10225 /* When net_device's are persistent, this will be fatal. */
10226 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
10229 ret = ethtool_check_ops(dev->ethtool_ops);
10233 spin_lock_init(&dev->addr_list_lock);
10234 netdev_set_addr_lockdep_class(dev);
10236 ret = dev_get_valid_name(net, dev, dev->name);
10241 dev->name_node = netdev_name_node_head_alloc(dev);
10242 if (!dev->name_node)
10245 /* Init, if this function is available */
10246 if (dev->netdev_ops->ndo_init) {
10247 ret = dev->netdev_ops->ndo_init(dev);
10251 goto err_free_name;
10255 if (((dev->hw_features | dev->features) &
10256 NETIF_F_HW_VLAN_CTAG_FILTER) &&
10257 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
10258 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
10259 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10266 dev->ifindex = dev_new_index(net);
10267 else if (__dev_get_by_index(net, dev->ifindex))
10270 /* Transfer changeable features to wanted_features and enable
10271 * software offloads (GSO and GRO).
10273 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10274 dev->features |= NETIF_F_SOFT_FEATURES;
10276 if (dev->udp_tunnel_nic_info) {
10277 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10278 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10281 dev->wanted_features = dev->features & dev->hw_features;
10283 if (!(dev->flags & IFF_LOOPBACK))
10284 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10286 /* If IPv4 TCP segmentation offload is supported we should also
10287 * allow the device to enable segmenting the frame with the option
10288 * of ignoring a static IP ID value. This doesn't enable the
10289 * feature itself but allows the user to enable it later.
10291 if (dev->hw_features & NETIF_F_TSO)
10292 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10293 if (dev->vlan_features & NETIF_F_TSO)
10294 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10295 if (dev->mpls_features & NETIF_F_TSO)
10296 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10297 if (dev->hw_enc_features & NETIF_F_TSO)
10298 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10300 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10302 dev->vlan_features |= NETIF_F_HIGHDMA;
10304 /* Make NETIF_F_SG inheritable to tunnel devices.
10306 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10308 /* Make NETIF_F_SG inheritable to MPLS.
10310 dev->mpls_features |= NETIF_F_SG;
10312 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10313 ret = notifier_to_errno(ret);
10317 ret = netdev_register_kobject(dev);
10319 dev->reg_state = NETREG_UNREGISTERED;
10322 dev->reg_state = NETREG_REGISTERED;
10324 __netdev_update_features(dev);
10327 * Default initial state at registry is that the
10328 * device is present.
10331 set_bit(__LINK_STATE_PRESENT, &dev->state);
10333 linkwatch_init_dev(dev);
10335 dev_init_scheduler(dev);
10337 list_netdevice(dev);
10338 add_device_randomness(dev->dev_addr, dev->addr_len);
10340 /* If the device has permanent device address, driver should
10341 * set dev_addr and also addr_assign_type should be set to
10342 * NET_ADDR_PERM (default value).
10344 if (dev->addr_assign_type == NET_ADDR_PERM)
10345 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10347 /* Notify protocols, that a new device appeared. */
10348 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10349 ret = notifier_to_errno(ret);
10351 /* Expect explicit free_netdev() on failure */
10352 dev->needs_free_netdev = false;
10353 unregister_netdevice_queue(dev, NULL);
10357 * Prevent userspace races by waiting until the network
10358 * device is fully setup before sending notifications.
10360 if (!dev->rtnl_link_ops ||
10361 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10362 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
10368 if (dev->netdev_ops->ndo_uninit)
10369 dev->netdev_ops->ndo_uninit(dev);
10370 if (dev->priv_destructor)
10371 dev->priv_destructor(dev);
10373 netdev_name_node_free(dev->name_node);
10376 EXPORT_SYMBOL(register_netdevice);
10379 * init_dummy_netdev - init a dummy network device for NAPI
10380 * @dev: device to init
10382 * This takes a network device structure and initialize the minimum
10383 * amount of fields so it can be used to schedule NAPI polls without
10384 * registering a full blown interface. This is to be used by drivers
10385 * that need to tie several hardware interfaces to a single NAPI
10386 * poll scheduler due to HW limitations.
10388 int init_dummy_netdev(struct net_device *dev)
10390 /* Clear everything. Note we don't initialize spinlocks
10391 * are they aren't supposed to be taken by any of the
10392 * NAPI code and this dummy netdev is supposed to be
10393 * only ever used for NAPI polls
10395 memset(dev, 0, sizeof(struct net_device));
10397 /* make sure we BUG if trying to hit standard
10398 * register/unregister code path
10400 dev->reg_state = NETREG_DUMMY;
10402 /* NAPI wants this */
10403 INIT_LIST_HEAD(&dev->napi_list);
10405 /* a dummy interface is started by default */
10406 set_bit(__LINK_STATE_PRESENT, &dev->state);
10407 set_bit(__LINK_STATE_START, &dev->state);
10409 /* napi_busy_loop stats accounting wants this */
10410 dev_net_set(dev, &init_net);
10412 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10413 * because users of this 'device' dont need to change
10419 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10423 * register_netdev - register a network device
10424 * @dev: device to register
10426 * Take a completed network device structure and add it to the kernel
10427 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10428 * chain. 0 is returned on success. A negative errno code is returned
10429 * on a failure to set up the device, or if the name is a duplicate.
10431 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10432 * and expands the device name if you passed a format string to
10435 int register_netdev(struct net_device *dev)
10439 if (rtnl_lock_killable())
10441 err = register_netdevice(dev);
10445 EXPORT_SYMBOL(register_netdev);
10447 int netdev_refcnt_read(const struct net_device *dev)
10449 #ifdef CONFIG_PCPU_DEV_REFCNT
10452 for_each_possible_cpu(i)
10453 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10456 return refcount_read(&dev->dev_refcnt);
10459 EXPORT_SYMBOL(netdev_refcnt_read);
10461 int netdev_unregister_timeout_secs __read_mostly = 10;
10463 #define WAIT_REFS_MIN_MSECS 1
10464 #define WAIT_REFS_MAX_MSECS 250
10466 * netdev_wait_allrefs - wait until all references are gone.
10467 * @dev: target net_device
10469 * This is called when unregistering network devices.
10471 * Any protocol or device that holds a reference should register
10472 * for netdevice notification, and cleanup and put back the
10473 * reference if they receive an UNREGISTER event.
10474 * We can get stuck here if buggy protocols don't correctly
10477 static void netdev_wait_allrefs(struct net_device *dev)
10479 unsigned long rebroadcast_time, warning_time;
10480 int wait = 0, refcnt;
10482 linkwatch_forget_dev(dev);
10484 rebroadcast_time = warning_time = jiffies;
10485 refcnt = netdev_refcnt_read(dev);
10487 while (refcnt != 1) {
10488 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10491 /* Rebroadcast unregister notification */
10492 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10498 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10500 /* We must not have linkwatch events
10501 * pending on unregister. If this
10502 * happens, we simply run the queue
10503 * unscheduled, resulting in a noop
10506 linkwatch_run_queue();
10511 rebroadcast_time = jiffies;
10516 wait = WAIT_REFS_MIN_MSECS;
10519 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10522 refcnt = netdev_refcnt_read(dev);
10525 time_after(jiffies, warning_time +
10526 netdev_unregister_timeout_secs * HZ)) {
10527 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10528 dev->name, refcnt);
10529 warning_time = jiffies;
10534 /* The sequence is:
10538 * register_netdevice(x1);
10539 * register_netdevice(x2);
10541 * unregister_netdevice(y1);
10542 * unregister_netdevice(y2);
10548 * We are invoked by rtnl_unlock().
10549 * This allows us to deal with problems:
10550 * 1) We can delete sysfs objects which invoke hotplug
10551 * without deadlocking with linkwatch via keventd.
10552 * 2) Since we run with the RTNL semaphore not held, we can sleep
10553 * safely in order to wait for the netdev refcnt to drop to zero.
10555 * We must not return until all unregister events added during
10556 * the interval the lock was held have been completed.
10558 void netdev_run_todo(void)
10560 struct list_head list;
10561 #ifdef CONFIG_LOCKDEP
10562 struct list_head unlink_list;
10564 list_replace_init(&net_unlink_list, &unlink_list);
10566 while (!list_empty(&unlink_list)) {
10567 struct net_device *dev = list_first_entry(&unlink_list,
10570 list_del_init(&dev->unlink_list);
10571 dev->nested_level = dev->lower_level - 1;
10575 /* Snapshot list, allow later requests */
10576 list_replace_init(&net_todo_list, &list);
10581 /* Wait for rcu callbacks to finish before next phase */
10582 if (!list_empty(&list))
10585 while (!list_empty(&list)) {
10586 struct net_device *dev
10587 = list_first_entry(&list, struct net_device, todo_list);
10588 list_del(&dev->todo_list);
10590 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10591 pr_err("network todo '%s' but state %d\n",
10592 dev->name, dev->reg_state);
10597 dev->reg_state = NETREG_UNREGISTERED;
10599 netdev_wait_allrefs(dev);
10602 BUG_ON(netdev_refcnt_read(dev) != 1);
10603 BUG_ON(!list_empty(&dev->ptype_all));
10604 BUG_ON(!list_empty(&dev->ptype_specific));
10605 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10606 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10607 #if IS_ENABLED(CONFIG_DECNET)
10608 WARN_ON(dev->dn_ptr);
10610 if (dev->priv_destructor)
10611 dev->priv_destructor(dev);
10612 if (dev->needs_free_netdev)
10615 /* Report a network device has been unregistered */
10617 dev_net(dev)->dev_unreg_count--;
10619 wake_up(&netdev_unregistering_wq);
10621 /* Free network device */
10622 kobject_put(&dev->dev.kobj);
10626 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10627 * all the same fields in the same order as net_device_stats, with only
10628 * the type differing, but rtnl_link_stats64 may have additional fields
10629 * at the end for newer counters.
10631 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10632 const struct net_device_stats *netdev_stats)
10634 #if BITS_PER_LONG == 64
10635 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
10636 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
10637 /* zero out counters that only exist in rtnl_link_stats64 */
10638 memset((char *)stats64 + sizeof(*netdev_stats), 0,
10639 sizeof(*stats64) - sizeof(*netdev_stats));
10641 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
10642 const unsigned long *src = (const unsigned long *)netdev_stats;
10643 u64 *dst = (u64 *)stats64;
10645 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10646 for (i = 0; i < n; i++)
10648 /* zero out counters that only exist in rtnl_link_stats64 */
10649 memset((char *)stats64 + n * sizeof(u64), 0,
10650 sizeof(*stats64) - n * sizeof(u64));
10653 EXPORT_SYMBOL(netdev_stats_to_stats64);
10656 * dev_get_stats - get network device statistics
10657 * @dev: device to get statistics from
10658 * @storage: place to store stats
10660 * Get network statistics from device. Return @storage.
10661 * The device driver may provide its own method by setting
10662 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10663 * otherwise the internal statistics structure is used.
10665 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10666 struct rtnl_link_stats64 *storage)
10668 const struct net_device_ops *ops = dev->netdev_ops;
10670 if (ops->ndo_get_stats64) {
10671 memset(storage, 0, sizeof(*storage));
10672 ops->ndo_get_stats64(dev, storage);
10673 } else if (ops->ndo_get_stats) {
10674 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10676 netdev_stats_to_stats64(storage, &dev->stats);
10678 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
10679 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
10680 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
10683 EXPORT_SYMBOL(dev_get_stats);
10686 * dev_fetch_sw_netstats - get per-cpu network device statistics
10687 * @s: place to store stats
10688 * @netstats: per-cpu network stats to read from
10690 * Read per-cpu network statistics and populate the related fields in @s.
10692 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10693 const struct pcpu_sw_netstats __percpu *netstats)
10697 for_each_possible_cpu(cpu) {
10698 const struct pcpu_sw_netstats *stats;
10699 struct pcpu_sw_netstats tmp;
10700 unsigned int start;
10702 stats = per_cpu_ptr(netstats, cpu);
10704 start = u64_stats_fetch_begin_irq(&stats->syncp);
10705 tmp.rx_packets = stats->rx_packets;
10706 tmp.rx_bytes = stats->rx_bytes;
10707 tmp.tx_packets = stats->tx_packets;
10708 tmp.tx_bytes = stats->tx_bytes;
10709 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
10711 s->rx_packets += tmp.rx_packets;
10712 s->rx_bytes += tmp.rx_bytes;
10713 s->tx_packets += tmp.tx_packets;
10714 s->tx_bytes += tmp.tx_bytes;
10717 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10720 * dev_get_tstats64 - ndo_get_stats64 implementation
10721 * @dev: device to get statistics from
10722 * @s: place to store stats
10724 * Populate @s from dev->stats and dev->tstats. Can be used as
10725 * ndo_get_stats64() callback.
10727 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10729 netdev_stats_to_stats64(s, &dev->stats);
10730 dev_fetch_sw_netstats(s, dev->tstats);
10732 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10734 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10736 struct netdev_queue *queue = dev_ingress_queue(dev);
10738 #ifdef CONFIG_NET_CLS_ACT
10741 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10744 netdev_init_one_queue(dev, queue, NULL);
10745 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10746 queue->qdisc_sleeping = &noop_qdisc;
10747 rcu_assign_pointer(dev->ingress_queue, queue);
10752 static const struct ethtool_ops default_ethtool_ops;
10754 void netdev_set_default_ethtool_ops(struct net_device *dev,
10755 const struct ethtool_ops *ops)
10757 if (dev->ethtool_ops == &default_ethtool_ops)
10758 dev->ethtool_ops = ops;
10760 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10762 void netdev_freemem(struct net_device *dev)
10764 char *addr = (char *)dev - dev->padded;
10770 * alloc_netdev_mqs - allocate network device
10771 * @sizeof_priv: size of private data to allocate space for
10772 * @name: device name format string
10773 * @name_assign_type: origin of device name
10774 * @setup: callback to initialize device
10775 * @txqs: the number of TX subqueues to allocate
10776 * @rxqs: the number of RX subqueues to allocate
10778 * Allocates a struct net_device with private data area for driver use
10779 * and performs basic initialization. Also allocates subqueue structs
10780 * for each queue on the device.
10782 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10783 unsigned char name_assign_type,
10784 void (*setup)(struct net_device *),
10785 unsigned int txqs, unsigned int rxqs)
10787 struct net_device *dev;
10788 unsigned int alloc_size;
10789 struct net_device *p;
10791 BUG_ON(strlen(name) >= sizeof(dev->name));
10794 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10799 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10803 alloc_size = sizeof(struct net_device);
10805 /* ensure 32-byte alignment of private area */
10806 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10807 alloc_size += sizeof_priv;
10809 /* ensure 32-byte alignment of whole construct */
10810 alloc_size += NETDEV_ALIGN - 1;
10812 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10816 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10817 dev->padded = (char *)dev - (char *)p;
10819 #ifdef CONFIG_PCPU_DEV_REFCNT
10820 dev->pcpu_refcnt = alloc_percpu(int);
10821 if (!dev->pcpu_refcnt)
10825 refcount_set(&dev->dev_refcnt, 1);
10828 if (dev_addr_init(dev))
10834 dev_net_set(dev, &init_net);
10836 dev->gso_max_size = GSO_MAX_SIZE;
10837 dev->gso_max_segs = GSO_MAX_SEGS;
10838 dev->upper_level = 1;
10839 dev->lower_level = 1;
10840 #ifdef CONFIG_LOCKDEP
10841 dev->nested_level = 0;
10842 INIT_LIST_HEAD(&dev->unlink_list);
10845 INIT_LIST_HEAD(&dev->napi_list);
10846 INIT_LIST_HEAD(&dev->unreg_list);
10847 INIT_LIST_HEAD(&dev->close_list);
10848 INIT_LIST_HEAD(&dev->link_watch_list);
10849 INIT_LIST_HEAD(&dev->adj_list.upper);
10850 INIT_LIST_HEAD(&dev->adj_list.lower);
10851 INIT_LIST_HEAD(&dev->ptype_all);
10852 INIT_LIST_HEAD(&dev->ptype_specific);
10853 INIT_LIST_HEAD(&dev->net_notifier_list);
10854 #ifdef CONFIG_NET_SCHED
10855 hash_init(dev->qdisc_hash);
10857 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10860 if (!dev->tx_queue_len) {
10861 dev->priv_flags |= IFF_NO_QUEUE;
10862 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10865 dev->num_tx_queues = txqs;
10866 dev->real_num_tx_queues = txqs;
10867 if (netif_alloc_netdev_queues(dev))
10870 dev->num_rx_queues = rxqs;
10871 dev->real_num_rx_queues = rxqs;
10872 if (netif_alloc_rx_queues(dev))
10875 strcpy(dev->name, name);
10876 dev->name_assign_type = name_assign_type;
10877 dev->group = INIT_NETDEV_GROUP;
10878 if (!dev->ethtool_ops)
10879 dev->ethtool_ops = &default_ethtool_ops;
10881 nf_hook_ingress_init(dev);
10890 #ifdef CONFIG_PCPU_DEV_REFCNT
10891 free_percpu(dev->pcpu_refcnt);
10894 netdev_freemem(dev);
10897 EXPORT_SYMBOL(alloc_netdev_mqs);
10900 * free_netdev - free network device
10903 * This function does the last stage of destroying an allocated device
10904 * interface. The reference to the device object is released. If this
10905 * is the last reference then it will be freed.Must be called in process
10908 void free_netdev(struct net_device *dev)
10910 struct napi_struct *p, *n;
10914 /* When called immediately after register_netdevice() failed the unwind
10915 * handling may still be dismantling the device. Handle that case by
10916 * deferring the free.
10918 if (dev->reg_state == NETREG_UNREGISTERING) {
10920 dev->needs_free_netdev = true;
10924 netif_free_tx_queues(dev);
10925 netif_free_rx_queues(dev);
10927 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10929 /* Flush device addresses */
10930 dev_addr_flush(dev);
10932 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10935 #ifdef CONFIG_PCPU_DEV_REFCNT
10936 free_percpu(dev->pcpu_refcnt);
10937 dev->pcpu_refcnt = NULL;
10939 free_percpu(dev->xdp_bulkq);
10940 dev->xdp_bulkq = NULL;
10942 /* Compatibility with error handling in drivers */
10943 if (dev->reg_state == NETREG_UNINITIALIZED) {
10944 netdev_freemem(dev);
10948 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10949 dev->reg_state = NETREG_RELEASED;
10951 /* will free via device release */
10952 put_device(&dev->dev);
10954 EXPORT_SYMBOL(free_netdev);
10957 * synchronize_net - Synchronize with packet receive processing
10959 * Wait for packets currently being received to be done.
10960 * Does not block later packets from starting.
10962 void synchronize_net(void)
10965 if (rtnl_is_locked())
10966 synchronize_rcu_expedited();
10970 EXPORT_SYMBOL(synchronize_net);
10973 * unregister_netdevice_queue - remove device from the kernel
10977 * This function shuts down a device interface and removes it
10978 * from the kernel tables.
10979 * If head not NULL, device is queued to be unregistered later.
10981 * Callers must hold the rtnl semaphore. You may want
10982 * unregister_netdev() instead of this.
10985 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
10990 list_move_tail(&dev->unreg_list, head);
10994 list_add(&dev->unreg_list, &single);
10995 unregister_netdevice_many(&single);
10998 EXPORT_SYMBOL(unregister_netdevice_queue);
11001 * unregister_netdevice_many - unregister many devices
11002 * @head: list of devices
11004 * Note: As most callers use a stack allocated list_head,
11005 * we force a list_del() to make sure stack wont be corrupted later.
11007 void unregister_netdevice_many(struct list_head *head)
11009 struct net_device *dev, *tmp;
11010 LIST_HEAD(close_head);
11012 BUG_ON(dev_boot_phase);
11015 if (list_empty(head))
11018 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
11019 /* Some devices call without registering
11020 * for initialization unwind. Remove those
11021 * devices and proceed with the remaining.
11023 if (dev->reg_state == NETREG_UNINITIALIZED) {
11024 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
11028 list_del(&dev->unreg_list);
11031 dev->dismantle = true;
11032 BUG_ON(dev->reg_state != NETREG_REGISTERED);
11035 /* If device is running, close it first. */
11036 list_for_each_entry(dev, head, unreg_list)
11037 list_add_tail(&dev->close_list, &close_head);
11038 dev_close_many(&close_head, true);
11040 list_for_each_entry(dev, head, unreg_list) {
11041 /* And unlink it from device chain. */
11042 unlist_netdevice(dev);
11044 dev->reg_state = NETREG_UNREGISTERING;
11046 flush_all_backlogs();
11050 list_for_each_entry(dev, head, unreg_list) {
11051 struct sk_buff *skb = NULL;
11053 /* Shutdown queueing discipline. */
11056 dev_xdp_uninstall(dev);
11058 /* Notify protocols, that we are about to destroy
11059 * this device. They should clean all the things.
11061 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11063 if (!dev->rtnl_link_ops ||
11064 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
11065 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
11066 GFP_KERNEL, NULL, 0);
11069 * Flush the unicast and multicast chains
11074 netdev_name_node_alt_flush(dev);
11075 netdev_name_node_free(dev->name_node);
11077 if (dev->netdev_ops->ndo_uninit)
11078 dev->netdev_ops->ndo_uninit(dev);
11081 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
11083 /* Notifier chain MUST detach us all upper devices. */
11084 WARN_ON(netdev_has_any_upper_dev(dev));
11085 WARN_ON(netdev_has_any_lower_dev(dev));
11087 /* Remove entries from kobject tree */
11088 netdev_unregister_kobject(dev);
11090 /* Remove XPS queueing entries */
11091 netif_reset_xps_queues_gt(dev, 0);
11097 list_for_each_entry(dev, head, unreg_list) {
11104 EXPORT_SYMBOL(unregister_netdevice_many);
11107 * unregister_netdev - remove device from the kernel
11110 * This function shuts down a device interface and removes it
11111 * from the kernel tables.
11113 * This is just a wrapper for unregister_netdevice that takes
11114 * the rtnl semaphore. In general you want to use this and not
11115 * unregister_netdevice.
11117 void unregister_netdev(struct net_device *dev)
11120 unregister_netdevice(dev);
11123 EXPORT_SYMBOL(unregister_netdev);
11126 * __dev_change_net_namespace - move device to different nethost namespace
11128 * @net: network namespace
11129 * @pat: If not NULL name pattern to try if the current device name
11130 * is already taken in the destination network namespace.
11131 * @new_ifindex: If not zero, specifies device index in the target
11134 * This function shuts down a device interface and moves it
11135 * to a new network namespace. On success 0 is returned, on
11136 * a failure a netagive errno code is returned.
11138 * Callers must hold the rtnl semaphore.
11141 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
11142 const char *pat, int new_ifindex)
11144 struct net *net_old = dev_net(dev);
11149 /* Don't allow namespace local devices to be moved. */
11151 if (dev->features & NETIF_F_NETNS_LOCAL)
11154 /* Ensure the device has been registrered */
11155 if (dev->reg_state != NETREG_REGISTERED)
11158 /* Get out if there is nothing todo */
11160 if (net_eq(net_old, net))
11163 /* Pick the destination device name, and ensure
11164 * we can use it in the destination network namespace.
11167 if (__dev_get_by_name(net, dev->name)) {
11168 /* We get here if we can't use the current device name */
11171 err = dev_get_valid_name(net, dev, pat);
11176 /* Check that new_ifindex isn't used yet. */
11178 if (new_ifindex && __dev_get_by_index(net, new_ifindex))
11182 * And now a mini version of register_netdevice unregister_netdevice.
11185 /* If device is running close it first. */
11188 /* And unlink it from device chain */
11189 unlist_netdevice(dev);
11193 /* Shutdown queueing discipline. */
11196 /* Notify protocols, that we are about to destroy
11197 * this device. They should clean all the things.
11199 * Note that dev->reg_state stays at NETREG_REGISTERED.
11200 * This is wanted because this way 8021q and macvlan know
11201 * the device is just moving and can keep their slaves up.
11203 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11206 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11207 /* If there is an ifindex conflict assign a new one */
11208 if (!new_ifindex) {
11209 if (__dev_get_by_index(net, dev->ifindex))
11210 new_ifindex = dev_new_index(net);
11212 new_ifindex = dev->ifindex;
11215 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11219 * Flush the unicast and multicast chains
11224 /* Send a netdev-removed uevent to the old namespace */
11225 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11226 netdev_adjacent_del_links(dev);
11228 /* Move per-net netdevice notifiers that are following the netdevice */
11229 move_netdevice_notifiers_dev_net(dev, net);
11231 /* Actually switch the network namespace */
11232 dev_net_set(dev, net);
11233 dev->ifindex = new_ifindex;
11235 /* Send a netdev-add uevent to the new namespace */
11236 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11237 netdev_adjacent_add_links(dev);
11239 /* Fixup kobjects */
11240 err = device_rename(&dev->dev, dev->name);
11243 /* Adapt owner in case owning user namespace of target network
11244 * namespace is different from the original one.
11246 err = netdev_change_owner(dev, net_old, net);
11249 /* Add the device back in the hashes */
11250 list_netdevice(dev);
11252 /* Notify protocols, that a new device appeared. */
11253 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11256 * Prevent userspace races by waiting until the network
11257 * device is fully setup before sending notifications.
11259 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
11266 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11268 static int dev_cpu_dead(unsigned int oldcpu)
11270 struct sk_buff **list_skb;
11271 struct sk_buff *skb;
11273 struct softnet_data *sd, *oldsd, *remsd = NULL;
11275 local_irq_disable();
11276 cpu = smp_processor_id();
11277 sd = &per_cpu(softnet_data, cpu);
11278 oldsd = &per_cpu(softnet_data, oldcpu);
11280 /* Find end of our completion_queue. */
11281 list_skb = &sd->completion_queue;
11283 list_skb = &(*list_skb)->next;
11284 /* Append completion queue from offline CPU. */
11285 *list_skb = oldsd->completion_queue;
11286 oldsd->completion_queue = NULL;
11288 /* Append output queue from offline CPU. */
11289 if (oldsd->output_queue) {
11290 *sd->output_queue_tailp = oldsd->output_queue;
11291 sd->output_queue_tailp = oldsd->output_queue_tailp;
11292 oldsd->output_queue = NULL;
11293 oldsd->output_queue_tailp = &oldsd->output_queue;
11295 /* Append NAPI poll list from offline CPU, with one exception :
11296 * process_backlog() must be called by cpu owning percpu backlog.
11297 * We properly handle process_queue & input_pkt_queue later.
11299 while (!list_empty(&oldsd->poll_list)) {
11300 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11301 struct napi_struct,
11304 list_del_init(&napi->poll_list);
11305 if (napi->poll == process_backlog)
11308 ____napi_schedule(sd, napi);
11311 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11312 local_irq_enable();
11315 remsd = oldsd->rps_ipi_list;
11316 oldsd->rps_ipi_list = NULL;
11318 /* send out pending IPI's on offline CPU */
11319 net_rps_send_ipi(remsd);
11321 /* Process offline CPU's input_pkt_queue */
11322 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11324 input_queue_head_incr(oldsd);
11326 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11328 input_queue_head_incr(oldsd);
11335 * netdev_increment_features - increment feature set by one
11336 * @all: current feature set
11337 * @one: new feature set
11338 * @mask: mask feature set
11340 * Computes a new feature set after adding a device with feature set
11341 * @one to the master device with current feature set @all. Will not
11342 * enable anything that is off in @mask. Returns the new feature set.
11344 netdev_features_t netdev_increment_features(netdev_features_t all,
11345 netdev_features_t one, netdev_features_t mask)
11347 if (mask & NETIF_F_HW_CSUM)
11348 mask |= NETIF_F_CSUM_MASK;
11349 mask |= NETIF_F_VLAN_CHALLENGED;
11351 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11352 all &= one | ~NETIF_F_ALL_FOR_ALL;
11354 /* If one device supports hw checksumming, set for all. */
11355 if (all & NETIF_F_HW_CSUM)
11356 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11360 EXPORT_SYMBOL(netdev_increment_features);
11362 static struct hlist_head * __net_init netdev_create_hash(void)
11365 struct hlist_head *hash;
11367 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11369 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11370 INIT_HLIST_HEAD(&hash[i]);
11375 /* Initialize per network namespace state */
11376 static int __net_init netdev_init(struct net *net)
11378 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11379 8 * sizeof_field(struct napi_struct, gro_bitmask));
11381 INIT_LIST_HEAD(&net->dev_base_head);
11383 net->dev_name_head = netdev_create_hash();
11384 if (net->dev_name_head == NULL)
11387 net->dev_index_head = netdev_create_hash();
11388 if (net->dev_index_head == NULL)
11391 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11396 kfree(net->dev_name_head);
11402 * netdev_drivername - network driver for the device
11403 * @dev: network device
11405 * Determine network driver for device.
11407 const char *netdev_drivername(const struct net_device *dev)
11409 const struct device_driver *driver;
11410 const struct device *parent;
11411 const char *empty = "";
11413 parent = dev->dev.parent;
11417 driver = parent->driver;
11418 if (driver && driver->name)
11419 return driver->name;
11423 static void __netdev_printk(const char *level, const struct net_device *dev,
11424 struct va_format *vaf)
11426 if (dev && dev->dev.parent) {
11427 dev_printk_emit(level[1] - '0',
11430 dev_driver_string(dev->dev.parent),
11431 dev_name(dev->dev.parent),
11432 netdev_name(dev), netdev_reg_state(dev),
11435 printk("%s%s%s: %pV",
11436 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11438 printk("%s(NULL net_device): %pV", level, vaf);
11442 void netdev_printk(const char *level, const struct net_device *dev,
11443 const char *format, ...)
11445 struct va_format vaf;
11448 va_start(args, format);
11453 __netdev_printk(level, dev, &vaf);
11457 EXPORT_SYMBOL(netdev_printk);
11459 #define define_netdev_printk_level(func, level) \
11460 void func(const struct net_device *dev, const char *fmt, ...) \
11462 struct va_format vaf; \
11465 va_start(args, fmt); \
11470 __netdev_printk(level, dev, &vaf); \
11474 EXPORT_SYMBOL(func);
11476 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11477 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11478 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11479 define_netdev_printk_level(netdev_err, KERN_ERR);
11480 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11481 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11482 define_netdev_printk_level(netdev_info, KERN_INFO);
11484 static void __net_exit netdev_exit(struct net *net)
11486 kfree(net->dev_name_head);
11487 kfree(net->dev_index_head);
11488 if (net != &init_net)
11489 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11492 static struct pernet_operations __net_initdata netdev_net_ops = {
11493 .init = netdev_init,
11494 .exit = netdev_exit,
11497 static void __net_exit default_device_exit(struct net *net)
11499 struct net_device *dev, *aux;
11501 * Push all migratable network devices back to the
11502 * initial network namespace
11505 for_each_netdev_safe(net, dev, aux) {
11507 char fb_name[IFNAMSIZ];
11509 /* Ignore unmoveable devices (i.e. loopback) */
11510 if (dev->features & NETIF_F_NETNS_LOCAL)
11513 /* Leave virtual devices for the generic cleanup */
11514 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11517 /* Push remaining network devices to init_net */
11518 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11519 if (__dev_get_by_name(&init_net, fb_name))
11520 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11521 err = dev_change_net_namespace(dev, &init_net, fb_name);
11523 pr_emerg("%s: failed to move %s to init_net: %d\n",
11524 __func__, dev->name, err);
11531 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
11533 /* Return with the rtnl_lock held when there are no network
11534 * devices unregistering in any network namespace in net_list.
11537 bool unregistering;
11538 DEFINE_WAIT_FUNC(wait, woken_wake_function);
11540 add_wait_queue(&netdev_unregistering_wq, &wait);
11542 unregistering = false;
11544 list_for_each_entry(net, net_list, exit_list) {
11545 if (net->dev_unreg_count > 0) {
11546 unregistering = true;
11550 if (!unregistering)
11554 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
11556 remove_wait_queue(&netdev_unregistering_wq, &wait);
11559 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11561 /* At exit all network devices most be removed from a network
11562 * namespace. Do this in the reverse order of registration.
11563 * Do this across as many network namespaces as possible to
11564 * improve batching efficiency.
11566 struct net_device *dev;
11568 LIST_HEAD(dev_kill_list);
11570 /* To prevent network device cleanup code from dereferencing
11571 * loopback devices or network devices that have been freed
11572 * wait here for all pending unregistrations to complete,
11573 * before unregistring the loopback device and allowing the
11574 * network namespace be freed.
11576 * The netdev todo list containing all network devices
11577 * unregistrations that happen in default_device_exit_batch
11578 * will run in the rtnl_unlock() at the end of
11579 * default_device_exit_batch.
11581 rtnl_lock_unregistering(net_list);
11582 list_for_each_entry(net, net_list, exit_list) {
11583 for_each_netdev_reverse(net, dev) {
11584 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11585 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11587 unregister_netdevice_queue(dev, &dev_kill_list);
11590 unregister_netdevice_many(&dev_kill_list);
11594 static struct pernet_operations __net_initdata default_device_ops = {
11595 .exit = default_device_exit,
11596 .exit_batch = default_device_exit_batch,
11600 * Initialize the DEV module. At boot time this walks the device list and
11601 * unhooks any devices that fail to initialise (normally hardware not
11602 * present) and leaves us with a valid list of present and active devices.
11607 * This is called single threaded during boot, so no need
11608 * to take the rtnl semaphore.
11610 static int __init net_dev_init(void)
11612 int i, rc = -ENOMEM;
11614 BUG_ON(!dev_boot_phase);
11616 if (dev_proc_init())
11619 if (netdev_kobject_init())
11622 INIT_LIST_HEAD(&ptype_all);
11623 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11624 INIT_LIST_HEAD(&ptype_base[i]);
11626 INIT_LIST_HEAD(&offload_base);
11628 if (register_pernet_subsys(&netdev_net_ops))
11632 * Initialise the packet receive queues.
11635 for_each_possible_cpu(i) {
11636 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11637 struct softnet_data *sd = &per_cpu(softnet_data, i);
11639 INIT_WORK(flush, flush_backlog);
11641 skb_queue_head_init(&sd->input_pkt_queue);
11642 skb_queue_head_init(&sd->process_queue);
11643 #ifdef CONFIG_XFRM_OFFLOAD
11644 skb_queue_head_init(&sd->xfrm_backlog);
11646 INIT_LIST_HEAD(&sd->poll_list);
11647 sd->output_queue_tailp = &sd->output_queue;
11649 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11653 init_gro_hash(&sd->backlog);
11654 sd->backlog.poll = process_backlog;
11655 sd->backlog.weight = weight_p;
11658 dev_boot_phase = 0;
11660 /* The loopback device is special if any other network devices
11661 * is present in a network namespace the loopback device must
11662 * be present. Since we now dynamically allocate and free the
11663 * loopback device ensure this invariant is maintained by
11664 * keeping the loopback device as the first device on the
11665 * list of network devices. Ensuring the loopback devices
11666 * is the first device that appears and the last network device
11669 if (register_pernet_device(&loopback_net_ops))
11672 if (register_pernet_device(&default_device_ops))
11675 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11676 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11678 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11679 NULL, dev_cpu_dead);
11686 subsys_initcall(net_dev_init);