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;
3170 if (skb_rx_queue_recorded(skb)) {
3171 hash = skb_get_rx_queue(skb);
3172 if (hash >= qoffset)
3174 while (unlikely(hash >= qcount))
3176 return hash + qoffset;
3179 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3182 static void skb_warn_bad_offload(const struct sk_buff *skb)
3184 static const netdev_features_t null_features;
3185 struct net_device *dev = skb->dev;
3186 const char *name = "";
3188 if (!net_ratelimit())
3192 if (dev->dev.parent)
3193 name = dev_driver_string(dev->dev.parent);
3195 name = netdev_name(dev);
3197 skb_dump(KERN_WARNING, skb, false);
3198 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3199 name, dev ? &dev->features : &null_features,
3200 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3204 * Invalidate hardware checksum when packet is to be mangled, and
3205 * complete checksum manually on outgoing path.
3207 int skb_checksum_help(struct sk_buff *skb)
3210 int ret = 0, offset;
3212 if (skb->ip_summed == CHECKSUM_COMPLETE)
3213 goto out_set_summed;
3215 if (unlikely(skb_is_gso(skb))) {
3216 skb_warn_bad_offload(skb);
3220 /* Before computing a checksum, we should make sure no frag could
3221 * be modified by an external entity : checksum could be wrong.
3223 if (skb_has_shared_frag(skb)) {
3224 ret = __skb_linearize(skb);
3229 offset = skb_checksum_start_offset(skb);
3230 BUG_ON(offset >= skb_headlen(skb));
3231 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3233 offset += skb->csum_offset;
3234 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
3236 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3240 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3242 skb->ip_summed = CHECKSUM_NONE;
3246 EXPORT_SYMBOL(skb_checksum_help);
3248 int skb_crc32c_csum_help(struct sk_buff *skb)
3251 int ret = 0, offset, start;
3253 if (skb->ip_summed != CHECKSUM_PARTIAL)
3256 if (unlikely(skb_is_gso(skb)))
3259 /* Before computing a checksum, we should make sure no frag could
3260 * be modified by an external entity : checksum could be wrong.
3262 if (unlikely(skb_has_shared_frag(skb))) {
3263 ret = __skb_linearize(skb);
3267 start = skb_checksum_start_offset(skb);
3268 offset = start + offsetof(struct sctphdr, checksum);
3269 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3274 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3278 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3279 skb->len - start, ~(__u32)0,
3281 *(__le32 *)(skb->data + offset) = crc32c_csum;
3282 skb->ip_summed = CHECKSUM_NONE;
3283 skb->csum_not_inet = 0;
3288 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3290 __be16 type = skb->protocol;
3292 /* Tunnel gso handlers can set protocol to ethernet. */
3293 if (type == htons(ETH_P_TEB)) {
3296 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3299 eth = (struct ethhdr *)skb->data;
3300 type = eth->h_proto;
3303 return __vlan_get_protocol(skb, type, depth);
3307 * skb_mac_gso_segment - mac layer segmentation handler.
3308 * @skb: buffer to segment
3309 * @features: features for the output path (see dev->features)
3311 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
3312 netdev_features_t features)
3314 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
3315 struct packet_offload *ptype;
3316 int vlan_depth = skb->mac_len;
3317 __be16 type = skb_network_protocol(skb, &vlan_depth);
3319 if (unlikely(!type))
3320 return ERR_PTR(-EINVAL);
3322 __skb_pull(skb, vlan_depth);
3325 list_for_each_entry_rcu(ptype, &offload_base, list) {
3326 if (ptype->type == type && ptype->callbacks.gso_segment) {
3327 segs = ptype->callbacks.gso_segment(skb, features);
3333 __skb_push(skb, skb->data - skb_mac_header(skb));
3337 EXPORT_SYMBOL(skb_mac_gso_segment);
3340 /* openvswitch calls this on rx path, so we need a different check.
3342 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
3345 return skb->ip_summed != CHECKSUM_PARTIAL &&
3346 skb->ip_summed != CHECKSUM_UNNECESSARY;
3348 return skb->ip_summed == CHECKSUM_NONE;
3352 * __skb_gso_segment - Perform segmentation on skb.
3353 * @skb: buffer to segment
3354 * @features: features for the output path (see dev->features)
3355 * @tx_path: whether it is called in TX path
3357 * This function segments the given skb and returns a list of segments.
3359 * It may return NULL if the skb requires no segmentation. This is
3360 * only possible when GSO is used for verifying header integrity.
3362 * Segmentation preserves SKB_GSO_CB_OFFSET bytes of previous skb cb.
3364 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3365 netdev_features_t features, bool tx_path)
3367 struct sk_buff *segs;
3369 if (unlikely(skb_needs_check(skb, tx_path))) {
3372 /* We're going to init ->check field in TCP or UDP header */
3373 err = skb_cow_head(skb, 0);
3375 return ERR_PTR(err);
3378 /* Only report GSO partial support if it will enable us to
3379 * support segmentation on this frame without needing additional
3382 if (features & NETIF_F_GSO_PARTIAL) {
3383 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3384 struct net_device *dev = skb->dev;
3386 partial_features |= dev->features & dev->gso_partial_features;
3387 if (!skb_gso_ok(skb, features | partial_features))
3388 features &= ~NETIF_F_GSO_PARTIAL;
3391 BUILD_BUG_ON(SKB_GSO_CB_OFFSET +
3392 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3394 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3395 SKB_GSO_CB(skb)->encap_level = 0;
3397 skb_reset_mac_header(skb);
3398 skb_reset_mac_len(skb);
3400 segs = skb_mac_gso_segment(skb, features);
3402 if (segs != skb && unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3403 skb_warn_bad_offload(skb);
3407 EXPORT_SYMBOL(__skb_gso_segment);
3409 /* Take action when hardware reception checksum errors are detected. */
3411 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3413 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
3414 skb_dump(KERN_ERR, skb, true);
3418 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3420 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3422 EXPORT_SYMBOL(netdev_rx_csum_fault);
3425 /* XXX: check that highmem exists at all on the given machine. */
3426 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3428 #ifdef CONFIG_HIGHMEM
3431 if (!(dev->features & NETIF_F_HIGHDMA)) {
3432 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3433 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3435 if (PageHighMem(skb_frag_page(frag)))
3443 /* If MPLS offload request, verify we are testing hardware MPLS features
3444 * instead of standard features for the netdev.
3446 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3447 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3448 netdev_features_t features,
3451 if (eth_p_mpls(type))
3452 features &= skb->dev->mpls_features;
3457 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3458 netdev_features_t features,
3465 static netdev_features_t harmonize_features(struct sk_buff *skb,
3466 netdev_features_t features)
3470 type = skb_network_protocol(skb, NULL);
3471 features = net_mpls_features(skb, features, type);
3473 if (skb->ip_summed != CHECKSUM_NONE &&
3474 !can_checksum_protocol(features, type)) {
3475 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3477 if (illegal_highdma(skb->dev, skb))
3478 features &= ~NETIF_F_SG;
3483 netdev_features_t passthru_features_check(struct sk_buff *skb,
3484 struct net_device *dev,
3485 netdev_features_t features)
3489 EXPORT_SYMBOL(passthru_features_check);
3491 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3492 struct net_device *dev,
3493 netdev_features_t features)
3495 return vlan_features_check(skb, features);
3498 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3499 struct net_device *dev,
3500 netdev_features_t features)
3502 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3504 if (gso_segs > dev->gso_max_segs)
3505 return features & ~NETIF_F_GSO_MASK;
3507 if (!skb_shinfo(skb)->gso_type) {
3508 skb_warn_bad_offload(skb);
3509 return features & ~NETIF_F_GSO_MASK;
3512 /* Support for GSO partial features requires software
3513 * intervention before we can actually process the packets
3514 * so we need to strip support for any partial features now
3515 * and we can pull them back in after we have partially
3516 * segmented the frame.
3518 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3519 features &= ~dev->gso_partial_features;
3521 /* Make sure to clear the IPv4 ID mangling feature if the
3522 * IPv4 header has the potential to be fragmented.
3524 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3525 struct iphdr *iph = skb->encapsulation ?
3526 inner_ip_hdr(skb) : ip_hdr(skb);
3528 if (!(iph->frag_off & htons(IP_DF)))
3529 features &= ~NETIF_F_TSO_MANGLEID;
3535 netdev_features_t netif_skb_features(struct sk_buff *skb)
3537 struct net_device *dev = skb->dev;
3538 netdev_features_t features = dev->features;
3540 if (skb_is_gso(skb))
3541 features = gso_features_check(skb, dev, features);
3543 /* If encapsulation offload request, verify we are testing
3544 * hardware encapsulation features instead of standard
3545 * features for the netdev
3547 if (skb->encapsulation)
3548 features &= dev->hw_enc_features;
3550 if (skb_vlan_tagged(skb))
3551 features = netdev_intersect_features(features,
3552 dev->vlan_features |
3553 NETIF_F_HW_VLAN_CTAG_TX |
3554 NETIF_F_HW_VLAN_STAG_TX);
3556 if (dev->netdev_ops->ndo_features_check)
3557 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3560 features &= dflt_features_check(skb, dev, features);
3562 return harmonize_features(skb, features);
3564 EXPORT_SYMBOL(netif_skb_features);
3566 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3567 struct netdev_queue *txq, bool more)
3572 if (dev_nit_active(dev))
3573 dev_queue_xmit_nit(skb, dev);
3576 PRANDOM_ADD_NOISE(skb, dev, txq, len + jiffies);
3577 trace_net_dev_start_xmit(skb, dev);
3578 rc = netdev_start_xmit(skb, dev, txq, more);
3579 trace_net_dev_xmit(skb, rc, dev, len);
3584 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3585 struct netdev_queue *txq, int *ret)
3587 struct sk_buff *skb = first;
3588 int rc = NETDEV_TX_OK;
3591 struct sk_buff *next = skb->next;
3593 skb_mark_not_on_list(skb);
3594 rc = xmit_one(skb, dev, txq, next != NULL);
3595 if (unlikely(!dev_xmit_complete(rc))) {
3601 if (netif_tx_queue_stopped(txq) && skb) {
3602 rc = NETDEV_TX_BUSY;
3612 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3613 netdev_features_t features)
3615 if (skb_vlan_tag_present(skb) &&
3616 !vlan_hw_offload_capable(features, skb->vlan_proto))
3617 skb = __vlan_hwaccel_push_inside(skb);
3621 int skb_csum_hwoffload_help(struct sk_buff *skb,
3622 const netdev_features_t features)
3624 if (unlikely(skb_csum_is_sctp(skb)))
3625 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3626 skb_crc32c_csum_help(skb);
3628 if (features & NETIF_F_HW_CSUM)
3631 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3632 switch (skb->csum_offset) {
3633 case offsetof(struct tcphdr, check):
3634 case offsetof(struct udphdr, check):
3639 return skb_checksum_help(skb);
3641 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3643 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3645 netdev_features_t features;
3647 features = netif_skb_features(skb);
3648 skb = validate_xmit_vlan(skb, features);
3652 skb = sk_validate_xmit_skb(skb, dev);
3656 if (netif_needs_gso(skb, features)) {
3657 struct sk_buff *segs;
3659 segs = skb_gso_segment(skb, features);
3667 if (skb_needs_linearize(skb, features) &&
3668 __skb_linearize(skb))
3671 /* If packet is not checksummed and device does not
3672 * support checksumming for this protocol, complete
3673 * checksumming here.
3675 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3676 if (skb->encapsulation)
3677 skb_set_inner_transport_header(skb,
3678 skb_checksum_start_offset(skb));
3680 skb_set_transport_header(skb,
3681 skb_checksum_start_offset(skb));
3682 if (skb_csum_hwoffload_help(skb, features))
3687 skb = validate_xmit_xfrm(skb, features, again);
3694 atomic_long_inc(&dev->tx_dropped);
3698 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3700 struct sk_buff *next, *head = NULL, *tail;
3702 for (; skb != NULL; skb = next) {
3704 skb_mark_not_on_list(skb);
3706 /* in case skb wont be segmented, point to itself */
3709 skb = validate_xmit_skb(skb, dev, again);
3717 /* If skb was segmented, skb->prev points to
3718 * the last segment. If not, it still contains skb.
3724 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3726 static void qdisc_pkt_len_init(struct sk_buff *skb)
3728 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3730 qdisc_skb_cb(skb)->pkt_len = skb->len;
3732 /* To get more precise estimation of bytes sent on wire,
3733 * we add to pkt_len the headers size of all segments
3735 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3736 unsigned int hdr_len;
3737 u16 gso_segs = shinfo->gso_segs;
3739 /* mac layer + network layer */
3740 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3742 /* + transport layer */
3743 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3744 const struct tcphdr *th;
3745 struct tcphdr _tcphdr;
3747 th = skb_header_pointer(skb, skb_transport_offset(skb),
3748 sizeof(_tcphdr), &_tcphdr);
3750 hdr_len += __tcp_hdrlen(th);
3752 struct udphdr _udphdr;
3754 if (skb_header_pointer(skb, skb_transport_offset(skb),
3755 sizeof(_udphdr), &_udphdr))
3756 hdr_len += sizeof(struct udphdr);
3759 if (shinfo->gso_type & SKB_GSO_DODGY)
3760 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3763 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3767 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q,
3768 struct sk_buff **to_free,
3769 struct netdev_queue *txq)
3773 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK;
3774 if (rc == NET_XMIT_SUCCESS)
3775 trace_qdisc_enqueue(q, txq, skb);
3779 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3780 struct net_device *dev,
3781 struct netdev_queue *txq)
3783 spinlock_t *root_lock = qdisc_lock(q);
3784 struct sk_buff *to_free = NULL;
3788 qdisc_calculate_pkt_len(skb, q);
3790 if (q->flags & TCQ_F_NOLOCK) {
3791 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3792 qdisc_run_begin(q)) {
3793 /* Retest nolock_qdisc_is_empty() within the protection
3794 * of q->seqlock to protect from racing with requeuing.
3796 if (unlikely(!nolock_qdisc_is_empty(q))) {
3797 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3804 qdisc_bstats_cpu_update(q, skb);
3805 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3806 !nolock_qdisc_is_empty(q))
3810 return NET_XMIT_SUCCESS;
3813 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3817 if (unlikely(to_free))
3818 kfree_skb_list(to_free);
3823 * Heuristic to force contended enqueues to serialize on a
3824 * separate lock before trying to get qdisc main lock.
3825 * This permits qdisc->running owner to get the lock more
3826 * often and dequeue packets faster.
3828 contended = qdisc_is_running(q);
3829 if (unlikely(contended))
3830 spin_lock(&q->busylock);
3832 spin_lock(root_lock);
3833 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3834 __qdisc_drop(skb, &to_free);
3836 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3837 qdisc_run_begin(q)) {
3839 * This is a work-conserving queue; there are no old skbs
3840 * waiting to be sent out; and the qdisc is not running -
3841 * xmit the skb directly.
3844 qdisc_bstats_update(q, skb);
3846 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3847 if (unlikely(contended)) {
3848 spin_unlock(&q->busylock);
3855 rc = NET_XMIT_SUCCESS;
3857 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3858 if (qdisc_run_begin(q)) {
3859 if (unlikely(contended)) {
3860 spin_unlock(&q->busylock);
3867 spin_unlock(root_lock);
3868 if (unlikely(to_free))
3869 kfree_skb_list(to_free);
3870 if (unlikely(contended))
3871 spin_unlock(&q->busylock);
3875 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3876 static void skb_update_prio(struct sk_buff *skb)
3878 const struct netprio_map *map;
3879 const struct sock *sk;
3880 unsigned int prioidx;
3884 map = rcu_dereference_bh(skb->dev->priomap);
3887 sk = skb_to_full_sk(skb);
3891 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3893 if (prioidx < map->priomap_len)
3894 skb->priority = map->priomap[prioidx];
3897 #define skb_update_prio(skb)
3901 * dev_loopback_xmit - loop back @skb
3902 * @net: network namespace this loopback is happening in
3903 * @sk: sk needed to be a netfilter okfn
3904 * @skb: buffer to transmit
3906 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3908 skb_reset_mac_header(skb);
3909 __skb_pull(skb, skb_network_offset(skb));
3910 skb->pkt_type = PACKET_LOOPBACK;
3911 skb->ip_summed = CHECKSUM_UNNECESSARY;
3912 WARN_ON(!skb_dst(skb));
3917 EXPORT_SYMBOL(dev_loopback_xmit);
3919 #ifdef CONFIG_NET_EGRESS
3920 static struct sk_buff *
3921 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3923 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3924 struct tcf_result cl_res;
3929 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3930 qdisc_skb_cb(skb)->mru = 0;
3931 qdisc_skb_cb(skb)->post_ct = false;
3932 mini_qdisc_bstats_cpu_update(miniq, skb);
3934 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
3936 case TC_ACT_RECLASSIFY:
3937 skb->tc_index = TC_H_MIN(cl_res.classid);
3940 mini_qdisc_qstats_cpu_drop(miniq);
3941 *ret = NET_XMIT_DROP;
3947 *ret = NET_XMIT_SUCCESS;
3950 case TC_ACT_REDIRECT:
3951 /* No need to push/pop skb's mac_header here on egress! */
3952 skb_do_redirect(skb);
3953 *ret = NET_XMIT_SUCCESS;
3961 #endif /* CONFIG_NET_EGRESS */
3964 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
3965 struct xps_dev_maps *dev_maps, unsigned int tci)
3967 int tc = netdev_get_prio_tc_map(dev, skb->priority);
3968 struct xps_map *map;
3969 int queue_index = -1;
3971 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
3974 tci *= dev_maps->num_tc;
3977 map = rcu_dereference(dev_maps->attr_map[tci]);
3980 queue_index = map->queues[0];
3982 queue_index = map->queues[reciprocal_scale(
3983 skb_get_hash(skb), map->len)];
3984 if (unlikely(queue_index >= dev->real_num_tx_queues))
3991 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
3992 struct sk_buff *skb)
3995 struct xps_dev_maps *dev_maps;
3996 struct sock *sk = skb->sk;
3997 int queue_index = -1;
3999 if (!static_key_false(&xps_needed))
4003 if (!static_key_false(&xps_rxqs_needed))
4006 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4008 int tci = sk_rx_queue_get(sk);
4011 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4016 if (queue_index < 0) {
4017 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4019 unsigned int tci = skb->sender_cpu - 1;
4021 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4033 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4034 struct net_device *sb_dev)
4038 EXPORT_SYMBOL(dev_pick_tx_zero);
4040 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4041 struct net_device *sb_dev)
4043 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4045 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4047 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4048 struct net_device *sb_dev)
4050 struct sock *sk = skb->sk;
4051 int queue_index = sk_tx_queue_get(sk);
4053 sb_dev = sb_dev ? : dev;
4055 if (queue_index < 0 || skb->ooo_okay ||
4056 queue_index >= dev->real_num_tx_queues) {
4057 int new_index = get_xps_queue(dev, sb_dev, skb);
4060 new_index = skb_tx_hash(dev, sb_dev, skb);
4062 if (queue_index != new_index && sk &&
4064 rcu_access_pointer(sk->sk_dst_cache))
4065 sk_tx_queue_set(sk, new_index);
4067 queue_index = new_index;
4072 EXPORT_SYMBOL(netdev_pick_tx);
4074 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4075 struct sk_buff *skb,
4076 struct net_device *sb_dev)
4078 int queue_index = 0;
4081 u32 sender_cpu = skb->sender_cpu - 1;
4083 if (sender_cpu >= (u32)NR_CPUS)
4084 skb->sender_cpu = raw_smp_processor_id() + 1;
4087 if (dev->real_num_tx_queues != 1) {
4088 const struct net_device_ops *ops = dev->netdev_ops;
4090 if (ops->ndo_select_queue)
4091 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4093 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4095 queue_index = netdev_cap_txqueue(dev, queue_index);
4098 skb_set_queue_mapping(skb, queue_index);
4099 return netdev_get_tx_queue(dev, queue_index);
4103 * __dev_queue_xmit - transmit a buffer
4104 * @skb: buffer to transmit
4105 * @sb_dev: suboordinate device used for L2 forwarding offload
4107 * Queue a buffer for transmission to a network device. The caller must
4108 * have set the device and priority and built the buffer before calling
4109 * this function. The function can be called from an interrupt.
4111 * A negative errno code is returned on a failure. A success does not
4112 * guarantee the frame will be transmitted as it may be dropped due
4113 * to congestion or traffic shaping.
4115 * -----------------------------------------------------------------------------------
4116 * I notice this method can also return errors from the queue disciplines,
4117 * including NET_XMIT_DROP, which is a positive value. So, errors can also
4120 * Regardless of the return value, the skb is consumed, so it is currently
4121 * difficult to retry a send to this method. (You can bump the ref count
4122 * before sending to hold a reference for retry if you are careful.)
4124 * When calling this method, interrupts MUST be enabled. This is because
4125 * the BH enable code must have IRQs enabled so that it will not deadlock.
4128 static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4130 struct net_device *dev = skb->dev;
4131 struct netdev_queue *txq;
4136 skb_reset_mac_header(skb);
4138 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4139 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4141 /* Disable soft irqs for various locks below. Also
4142 * stops preemption for RCU.
4146 skb_update_prio(skb);
4148 qdisc_pkt_len_init(skb);
4149 #ifdef CONFIG_NET_CLS_ACT
4150 skb->tc_at_ingress = 0;
4151 # ifdef CONFIG_NET_EGRESS
4152 if (static_branch_unlikely(&egress_needed_key)) {
4153 skb = sch_handle_egress(skb, &rc, dev);
4159 /* If device/qdisc don't need skb->dst, release it right now while
4160 * its hot in this cpu cache.
4162 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4167 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4168 q = rcu_dereference_bh(txq->qdisc);
4170 trace_net_dev_queue(skb);
4172 rc = __dev_xmit_skb(skb, q, dev, txq);
4176 /* The device has no queue. Common case for software devices:
4177 * loopback, all the sorts of tunnels...
4179 * Really, it is unlikely that netif_tx_lock protection is necessary
4180 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4182 * However, it is possible, that they rely on protection
4185 * Check this and shot the lock. It is not prone from deadlocks.
4186 *Either shot noqueue qdisc, it is even simpler 8)
4188 if (dev->flags & IFF_UP) {
4189 int cpu = smp_processor_id(); /* ok because BHs are off */
4191 if (txq->xmit_lock_owner != cpu) {
4192 if (dev_xmit_recursion())
4193 goto recursion_alert;
4195 skb = validate_xmit_skb(skb, dev, &again);
4199 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4200 HARD_TX_LOCK(dev, txq, cpu);
4202 if (!netif_xmit_stopped(txq)) {
4203 dev_xmit_recursion_inc();
4204 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4205 dev_xmit_recursion_dec();
4206 if (dev_xmit_complete(rc)) {
4207 HARD_TX_UNLOCK(dev, txq);
4211 HARD_TX_UNLOCK(dev, txq);
4212 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4215 /* Recursion is detected! It is possible,
4219 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4225 rcu_read_unlock_bh();
4227 atomic_long_inc(&dev->tx_dropped);
4228 kfree_skb_list(skb);
4231 rcu_read_unlock_bh();
4235 int dev_queue_xmit(struct sk_buff *skb)
4237 return __dev_queue_xmit(skb, NULL);
4239 EXPORT_SYMBOL(dev_queue_xmit);
4241 int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
4243 return __dev_queue_xmit(skb, sb_dev);
4245 EXPORT_SYMBOL(dev_queue_xmit_accel);
4247 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4249 struct net_device *dev = skb->dev;
4250 struct sk_buff *orig_skb = skb;
4251 struct netdev_queue *txq;
4252 int ret = NETDEV_TX_BUSY;
4255 if (unlikely(!netif_running(dev) ||
4256 !netif_carrier_ok(dev)))
4259 skb = validate_xmit_skb_list(skb, dev, &again);
4260 if (skb != orig_skb)
4263 skb_set_queue_mapping(skb, queue_id);
4264 txq = skb_get_tx_queue(dev, skb);
4265 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4269 dev_xmit_recursion_inc();
4270 HARD_TX_LOCK(dev, txq, smp_processor_id());
4271 if (!netif_xmit_frozen_or_drv_stopped(txq))
4272 ret = netdev_start_xmit(skb, dev, txq, false);
4273 HARD_TX_UNLOCK(dev, txq);
4274 dev_xmit_recursion_dec();
4279 atomic_long_inc(&dev->tx_dropped);
4280 kfree_skb_list(skb);
4281 return NET_XMIT_DROP;
4283 EXPORT_SYMBOL(__dev_direct_xmit);
4285 /*************************************************************************
4287 *************************************************************************/
4289 int netdev_max_backlog __read_mostly = 1000;
4290 EXPORT_SYMBOL(netdev_max_backlog);
4292 int netdev_tstamp_prequeue __read_mostly = 1;
4293 int netdev_budget __read_mostly = 300;
4294 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4295 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4296 int weight_p __read_mostly = 64; /* old backlog weight */
4297 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4298 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4299 int dev_rx_weight __read_mostly = 64;
4300 int dev_tx_weight __read_mostly = 64;
4301 /* Maximum number of GRO_NORMAL skbs to batch up for list-RX */
4302 int gro_normal_batch __read_mostly = 8;
4304 /* Called with irq disabled */
4305 static inline void ____napi_schedule(struct softnet_data *sd,
4306 struct napi_struct *napi)
4308 struct task_struct *thread;
4310 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4311 /* Paired with smp_mb__before_atomic() in
4312 * napi_enable()/dev_set_threaded().
4313 * Use READ_ONCE() to guarantee a complete
4314 * read on napi->thread. Only call
4315 * wake_up_process() when it's not NULL.
4317 thread = READ_ONCE(napi->thread);
4319 /* Avoid doing set_bit() if the thread is in
4320 * INTERRUPTIBLE state, cause napi_thread_wait()
4321 * makes sure to proceed with napi polling
4322 * if the thread is explicitly woken from here.
4324 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4325 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4326 wake_up_process(thread);
4331 list_add_tail(&napi->poll_list, &sd->poll_list);
4332 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4337 /* One global table that all flow-based protocols share. */
4338 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4339 EXPORT_SYMBOL(rps_sock_flow_table);
4340 u32 rps_cpu_mask __read_mostly;
4341 EXPORT_SYMBOL(rps_cpu_mask);
4343 struct static_key_false rps_needed __read_mostly;
4344 EXPORT_SYMBOL(rps_needed);
4345 struct static_key_false rfs_needed __read_mostly;
4346 EXPORT_SYMBOL(rfs_needed);
4348 static struct rps_dev_flow *
4349 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4350 struct rps_dev_flow *rflow, u16 next_cpu)
4352 if (next_cpu < nr_cpu_ids) {
4353 #ifdef CONFIG_RFS_ACCEL
4354 struct netdev_rx_queue *rxqueue;
4355 struct rps_dev_flow_table *flow_table;
4356 struct rps_dev_flow *old_rflow;
4361 /* Should we steer this flow to a different hardware queue? */
4362 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4363 !(dev->features & NETIF_F_NTUPLE))
4365 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4366 if (rxq_index == skb_get_rx_queue(skb))
4369 rxqueue = dev->_rx + rxq_index;
4370 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4373 flow_id = skb_get_hash(skb) & flow_table->mask;
4374 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4375 rxq_index, flow_id);
4379 rflow = &flow_table->flows[flow_id];
4381 if (old_rflow->filter == rflow->filter)
4382 old_rflow->filter = RPS_NO_FILTER;
4386 per_cpu(softnet_data, next_cpu).input_queue_head;
4389 rflow->cpu = next_cpu;
4394 * get_rps_cpu is called from netif_receive_skb and returns the target
4395 * CPU from the RPS map of the receiving queue for a given skb.
4396 * rcu_read_lock must be held on entry.
4398 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4399 struct rps_dev_flow **rflowp)
4401 const struct rps_sock_flow_table *sock_flow_table;
4402 struct netdev_rx_queue *rxqueue = dev->_rx;
4403 struct rps_dev_flow_table *flow_table;
4404 struct rps_map *map;
4409 if (skb_rx_queue_recorded(skb)) {
4410 u16 index = skb_get_rx_queue(skb);
4412 if (unlikely(index >= dev->real_num_rx_queues)) {
4413 WARN_ONCE(dev->real_num_rx_queues > 1,
4414 "%s received packet on queue %u, but number "
4415 "of RX queues is %u\n",
4416 dev->name, index, dev->real_num_rx_queues);
4422 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4424 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4425 map = rcu_dereference(rxqueue->rps_map);
4426 if (!flow_table && !map)
4429 skb_reset_network_header(skb);
4430 hash = skb_get_hash(skb);
4434 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4435 if (flow_table && sock_flow_table) {
4436 struct rps_dev_flow *rflow;
4440 /* First check into global flow table if there is a match */
4441 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4442 if ((ident ^ hash) & ~rps_cpu_mask)
4445 next_cpu = ident & rps_cpu_mask;
4447 /* OK, now we know there is a match,
4448 * we can look at the local (per receive queue) flow table
4450 rflow = &flow_table->flows[hash & flow_table->mask];
4454 * If the desired CPU (where last recvmsg was done) is
4455 * different from current CPU (one in the rx-queue flow
4456 * table entry), switch if one of the following holds:
4457 * - Current CPU is unset (>= nr_cpu_ids).
4458 * - Current CPU is offline.
4459 * - The current CPU's queue tail has advanced beyond the
4460 * last packet that was enqueued using this table entry.
4461 * This guarantees that all previous packets for the flow
4462 * have been dequeued, thus preserving in order delivery.
4464 if (unlikely(tcpu != next_cpu) &&
4465 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4466 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4467 rflow->last_qtail)) >= 0)) {
4469 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4472 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4482 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4483 if (cpu_online(tcpu)) {
4493 #ifdef CONFIG_RFS_ACCEL
4496 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4497 * @dev: Device on which the filter was set
4498 * @rxq_index: RX queue index
4499 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4500 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4502 * Drivers that implement ndo_rx_flow_steer() should periodically call
4503 * this function for each installed filter and remove the filters for
4504 * which it returns %true.
4506 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4507 u32 flow_id, u16 filter_id)
4509 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4510 struct rps_dev_flow_table *flow_table;
4511 struct rps_dev_flow *rflow;
4516 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4517 if (flow_table && flow_id <= flow_table->mask) {
4518 rflow = &flow_table->flows[flow_id];
4519 cpu = READ_ONCE(rflow->cpu);
4520 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4521 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4522 rflow->last_qtail) <
4523 (int)(10 * flow_table->mask)))
4529 EXPORT_SYMBOL(rps_may_expire_flow);
4531 #endif /* CONFIG_RFS_ACCEL */
4533 /* Called from hardirq (IPI) context */
4534 static void rps_trigger_softirq(void *data)
4536 struct softnet_data *sd = data;
4538 ____napi_schedule(sd, &sd->backlog);
4542 #endif /* CONFIG_RPS */
4545 * Check if this softnet_data structure is another cpu one
4546 * If yes, queue it to our IPI list and return 1
4549 static int rps_ipi_queued(struct softnet_data *sd)
4552 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4555 sd->rps_ipi_next = mysd->rps_ipi_list;
4556 mysd->rps_ipi_list = sd;
4558 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4561 #endif /* CONFIG_RPS */
4565 #ifdef CONFIG_NET_FLOW_LIMIT
4566 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4569 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4571 #ifdef CONFIG_NET_FLOW_LIMIT
4572 struct sd_flow_limit *fl;
4573 struct softnet_data *sd;
4574 unsigned int old_flow, new_flow;
4576 if (qlen < (netdev_max_backlog >> 1))
4579 sd = this_cpu_ptr(&softnet_data);
4582 fl = rcu_dereference(sd->flow_limit);
4584 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4585 old_flow = fl->history[fl->history_head];
4586 fl->history[fl->history_head] = new_flow;
4589 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4591 if (likely(fl->buckets[old_flow]))
4592 fl->buckets[old_flow]--;
4594 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4606 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4607 * queue (may be a remote CPU queue).
4609 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4610 unsigned int *qtail)
4612 struct softnet_data *sd;
4613 unsigned long flags;
4616 sd = &per_cpu(softnet_data, cpu);
4618 local_irq_save(flags);
4621 if (!netif_running(skb->dev))
4623 qlen = skb_queue_len(&sd->input_pkt_queue);
4624 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
4627 __skb_queue_tail(&sd->input_pkt_queue, skb);
4628 input_queue_tail_incr_save(sd, qtail);
4630 local_irq_restore(flags);
4631 return NET_RX_SUCCESS;
4634 /* Schedule NAPI for backlog device
4635 * We can use non atomic operation since we own the queue lock
4637 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
4638 if (!rps_ipi_queued(sd))
4639 ____napi_schedule(sd, &sd->backlog);
4648 local_irq_restore(flags);
4650 atomic_long_inc(&skb->dev->rx_dropped);
4655 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4657 struct net_device *dev = skb->dev;
4658 struct netdev_rx_queue *rxqueue;
4662 if (skb_rx_queue_recorded(skb)) {
4663 u16 index = skb_get_rx_queue(skb);
4665 if (unlikely(index >= dev->real_num_rx_queues)) {
4666 WARN_ONCE(dev->real_num_rx_queues > 1,
4667 "%s received packet on queue %u, but number "
4668 "of RX queues is %u\n",
4669 dev->name, index, dev->real_num_rx_queues);
4671 return rxqueue; /* Return first rxqueue */
4678 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp,
4679 struct bpf_prog *xdp_prog)
4681 void *orig_data, *orig_data_end, *hard_start;
4682 struct netdev_rx_queue *rxqueue;
4683 bool orig_bcast, orig_host;
4684 u32 mac_len, frame_sz;
4685 __be16 orig_eth_type;
4690 /* The XDP program wants to see the packet starting at the MAC
4693 mac_len = skb->data - skb_mac_header(skb);
4694 hard_start = skb->data - skb_headroom(skb);
4696 /* SKB "head" area always have tailroom for skb_shared_info */
4697 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4698 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4700 rxqueue = netif_get_rxqueue(skb);
4701 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4702 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4703 skb_headlen(skb) + mac_len, true);
4705 orig_data_end = xdp->data_end;
4706 orig_data = xdp->data;
4707 eth = (struct ethhdr *)xdp->data;
4708 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4709 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4710 orig_eth_type = eth->h_proto;
4712 act = bpf_prog_run_xdp(xdp_prog, xdp);
4714 /* check if bpf_xdp_adjust_head was used */
4715 off = xdp->data - orig_data;
4718 __skb_pull(skb, off);
4720 __skb_push(skb, -off);
4722 skb->mac_header += off;
4723 skb_reset_network_header(skb);
4726 /* check if bpf_xdp_adjust_tail was used */
4727 off = xdp->data_end - orig_data_end;
4729 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4730 skb->len += off; /* positive on grow, negative on shrink */
4733 /* check if XDP changed eth hdr such SKB needs update */
4734 eth = (struct ethhdr *)xdp->data;
4735 if ((orig_eth_type != eth->h_proto) ||
4736 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4737 skb->dev->dev_addr)) ||
4738 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4739 __skb_push(skb, ETH_HLEN);
4740 skb->pkt_type = PACKET_HOST;
4741 skb->protocol = eth_type_trans(skb, skb->dev);
4744 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
4745 * before calling us again on redirect path. We do not call do_redirect
4746 * as we leave that up to the caller.
4748 * Caller is responsible for managing lifetime of skb (i.e. calling
4749 * kfree_skb in response to actions it cannot handle/XDP_DROP).
4754 __skb_push(skb, mac_len);
4757 metalen = xdp->data - xdp->data_meta;
4759 skb_metadata_set(skb, metalen);
4766 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4767 struct xdp_buff *xdp,
4768 struct bpf_prog *xdp_prog)
4772 /* Reinjected packets coming from act_mirred or similar should
4773 * not get XDP generic processing.
4775 if (skb_is_redirected(skb))
4778 /* XDP packets must be linear and must have sufficient headroom
4779 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4780 * native XDP provides, thus we need to do it here as well.
4782 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4783 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4784 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4785 int troom = skb->tail + skb->data_len - skb->end;
4787 /* In case we have to go down the path and also linearize,
4788 * then lets do the pskb_expand_head() work just once here.
4790 if (pskb_expand_head(skb,
4791 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4792 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4794 if (skb_linearize(skb))
4798 act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
4805 bpf_warn_invalid_xdp_action(act);
4808 trace_xdp_exception(skb->dev, xdp_prog, act);
4819 /* When doing generic XDP we have to bypass the qdisc layer and the
4820 * network taps in order to match in-driver-XDP behavior.
4822 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4824 struct net_device *dev = skb->dev;
4825 struct netdev_queue *txq;
4826 bool free_skb = true;
4829 txq = netdev_core_pick_tx(dev, skb, NULL);
4830 cpu = smp_processor_id();
4831 HARD_TX_LOCK(dev, txq, cpu);
4832 if (!netif_xmit_stopped(txq)) {
4833 rc = netdev_start_xmit(skb, dev, txq, 0);
4834 if (dev_xmit_complete(rc))
4837 HARD_TX_UNLOCK(dev, txq);
4839 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4844 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4846 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4849 struct xdp_buff xdp;
4853 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4854 if (act != XDP_PASS) {
4857 err = xdp_do_generic_redirect(skb->dev, skb,
4863 generic_xdp_tx(skb, xdp_prog);
4874 EXPORT_SYMBOL_GPL(do_xdp_generic);
4876 static int netif_rx_internal(struct sk_buff *skb)
4880 net_timestamp_check(netdev_tstamp_prequeue, skb);
4882 trace_netif_rx(skb);
4885 if (static_branch_unlikely(&rps_needed)) {
4886 struct rps_dev_flow voidflow, *rflow = &voidflow;
4892 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4894 cpu = smp_processor_id();
4896 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4905 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4912 * netif_rx - post buffer to the network code
4913 * @skb: buffer to post
4915 * This function receives a packet from a device driver and queues it for
4916 * the upper (protocol) levels to process. It always succeeds. The buffer
4917 * may be dropped during processing for congestion control or by the
4921 * NET_RX_SUCCESS (no congestion)
4922 * NET_RX_DROP (packet was dropped)
4926 int netif_rx(struct sk_buff *skb)
4930 trace_netif_rx_entry(skb);
4932 ret = netif_rx_internal(skb);
4933 trace_netif_rx_exit(ret);
4937 EXPORT_SYMBOL(netif_rx);
4939 int netif_rx_ni(struct sk_buff *skb)
4943 trace_netif_rx_ni_entry(skb);
4946 err = netif_rx_internal(skb);
4947 if (local_softirq_pending())
4950 trace_netif_rx_ni_exit(err);
4954 EXPORT_SYMBOL(netif_rx_ni);
4956 int netif_rx_any_context(struct sk_buff *skb)
4959 * If invoked from contexts which do not invoke bottom half
4960 * processing either at return from interrupt or when softrqs are
4961 * reenabled, use netif_rx_ni() which invokes bottomhalf processing
4965 return netif_rx(skb);
4967 return netif_rx_ni(skb);
4969 EXPORT_SYMBOL(netif_rx_any_context);
4971 static __latent_entropy void net_tx_action(struct softirq_action *h)
4973 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4975 if (sd->completion_queue) {
4976 struct sk_buff *clist;
4978 local_irq_disable();
4979 clist = sd->completion_queue;
4980 sd->completion_queue = NULL;
4984 struct sk_buff *skb = clist;
4986 clist = clist->next;
4988 WARN_ON(refcount_read(&skb->users));
4989 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4990 trace_consume_skb(skb);
4992 trace_kfree_skb(skb, net_tx_action);
4994 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4997 __kfree_skb_defer(skb);
5001 if (sd->output_queue) {
5004 local_irq_disable();
5005 head = sd->output_queue;
5006 sd->output_queue = NULL;
5007 sd->output_queue_tailp = &sd->output_queue;
5013 struct Qdisc *q = head;
5014 spinlock_t *root_lock = NULL;
5016 head = head->next_sched;
5018 /* We need to make sure head->next_sched is read
5019 * before clearing __QDISC_STATE_SCHED
5021 smp_mb__before_atomic();
5023 if (!(q->flags & TCQ_F_NOLOCK)) {
5024 root_lock = qdisc_lock(q);
5025 spin_lock(root_lock);
5026 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5028 /* There is a synchronize_net() between
5029 * STATE_DEACTIVATED flag being set and
5030 * qdisc_reset()/some_qdisc_is_busy() in
5031 * dev_deactivate(), so we can safely bail out
5032 * early here to avoid data race between
5033 * qdisc_deactivate() and some_qdisc_is_busy()
5034 * for lockless qdisc.
5036 clear_bit(__QDISC_STATE_SCHED, &q->state);
5040 clear_bit(__QDISC_STATE_SCHED, &q->state);
5043 spin_unlock(root_lock);
5049 xfrm_dev_backlog(sd);
5052 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5053 /* This hook is defined here for ATM LANE */
5054 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5055 unsigned char *addr) __read_mostly;
5056 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5059 static inline struct sk_buff *
5060 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
5061 struct net_device *orig_dev, bool *another)
5063 #ifdef CONFIG_NET_CLS_ACT
5064 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
5065 struct tcf_result cl_res;
5067 /* If there's at least one ingress present somewhere (so
5068 * we get here via enabled static key), remaining devices
5069 * that are not configured with an ingress qdisc will bail
5076 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5080 qdisc_skb_cb(skb)->pkt_len = skb->len;
5081 qdisc_skb_cb(skb)->mru = 0;
5082 qdisc_skb_cb(skb)->post_ct = false;
5083 skb->tc_at_ingress = 1;
5084 mini_qdisc_bstats_cpu_update(miniq, skb);
5086 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
5088 case TC_ACT_RECLASSIFY:
5089 skb->tc_index = TC_H_MIN(cl_res.classid);
5092 mini_qdisc_qstats_cpu_drop(miniq);
5100 case TC_ACT_REDIRECT:
5101 /* skb_mac_header check was done by cls/act_bpf, so
5102 * we can safely push the L2 header back before
5103 * redirecting to another netdev
5105 __skb_push(skb, skb->mac_len);
5106 if (skb_do_redirect(skb) == -EAGAIN) {
5107 __skb_pull(skb, skb->mac_len);
5112 case TC_ACT_CONSUMED:
5117 #endif /* CONFIG_NET_CLS_ACT */
5122 * netdev_is_rx_handler_busy - check if receive handler is registered
5123 * @dev: device to check
5125 * Check if a receive handler is already registered for a given device.
5126 * Return true if there one.
5128 * The caller must hold the rtnl_mutex.
5130 bool netdev_is_rx_handler_busy(struct net_device *dev)
5133 return dev && rtnl_dereference(dev->rx_handler);
5135 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5138 * netdev_rx_handler_register - register receive handler
5139 * @dev: device to register a handler for
5140 * @rx_handler: receive handler to register
5141 * @rx_handler_data: data pointer that is used by rx handler
5143 * Register a receive handler for a device. This handler will then be
5144 * called from __netif_receive_skb. A negative errno code is returned
5147 * The caller must hold the rtnl_mutex.
5149 * For a general description of rx_handler, see enum rx_handler_result.
5151 int netdev_rx_handler_register(struct net_device *dev,
5152 rx_handler_func_t *rx_handler,
5153 void *rx_handler_data)
5155 if (netdev_is_rx_handler_busy(dev))
5158 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5161 /* Note: rx_handler_data must be set before rx_handler */
5162 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5163 rcu_assign_pointer(dev->rx_handler, rx_handler);
5167 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5170 * netdev_rx_handler_unregister - unregister receive handler
5171 * @dev: device to unregister a handler from
5173 * Unregister a receive handler from a device.
5175 * The caller must hold the rtnl_mutex.
5177 void netdev_rx_handler_unregister(struct net_device *dev)
5181 RCU_INIT_POINTER(dev->rx_handler, NULL);
5182 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5183 * section has a guarantee to see a non NULL rx_handler_data
5187 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5189 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5192 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5193 * the special handling of PFMEMALLOC skbs.
5195 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5197 switch (skb->protocol) {
5198 case htons(ETH_P_ARP):
5199 case htons(ETH_P_IP):
5200 case htons(ETH_P_IPV6):
5201 case htons(ETH_P_8021Q):
5202 case htons(ETH_P_8021AD):
5209 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5210 int *ret, struct net_device *orig_dev)
5212 if (nf_hook_ingress_active(skb)) {
5216 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5221 ingress_retval = nf_hook_ingress(skb);
5223 return ingress_retval;
5228 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5229 struct packet_type **ppt_prev)
5231 struct packet_type *ptype, *pt_prev;
5232 rx_handler_func_t *rx_handler;
5233 struct sk_buff *skb = *pskb;
5234 struct net_device *orig_dev;
5235 bool deliver_exact = false;
5236 int ret = NET_RX_DROP;
5239 net_timestamp_check(!netdev_tstamp_prequeue, skb);
5241 trace_netif_receive_skb(skb);
5243 orig_dev = skb->dev;
5245 skb_reset_network_header(skb);
5246 if (!skb_transport_header_was_set(skb))
5247 skb_reset_transport_header(skb);
5248 skb_reset_mac_len(skb);
5253 skb->skb_iif = skb->dev->ifindex;
5255 __this_cpu_inc(softnet_data.processed);
5257 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5261 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5264 if (ret2 != XDP_PASS) {
5270 if (eth_type_vlan(skb->protocol)) {
5271 skb = skb_vlan_untag(skb);
5276 if (skb_skip_tc_classify(skb))
5282 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5284 ret = deliver_skb(skb, pt_prev, orig_dev);
5288 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5290 ret = deliver_skb(skb, pt_prev, orig_dev);
5295 #ifdef CONFIG_NET_INGRESS
5296 if (static_branch_unlikely(&ingress_needed_key)) {
5297 bool another = false;
5299 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5306 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5310 skb_reset_redirect(skb);
5312 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5315 if (skb_vlan_tag_present(skb)) {
5317 ret = deliver_skb(skb, pt_prev, orig_dev);
5320 if (vlan_do_receive(&skb))
5322 else if (unlikely(!skb))
5326 rx_handler = rcu_dereference(skb->dev->rx_handler);
5329 ret = deliver_skb(skb, pt_prev, orig_dev);
5332 switch (rx_handler(&skb)) {
5333 case RX_HANDLER_CONSUMED:
5334 ret = NET_RX_SUCCESS;
5336 case RX_HANDLER_ANOTHER:
5338 case RX_HANDLER_EXACT:
5339 deliver_exact = true;
5341 case RX_HANDLER_PASS:
5348 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5350 if (skb_vlan_tag_get_id(skb)) {
5351 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5354 skb->pkt_type = PACKET_OTHERHOST;
5355 } else if (eth_type_vlan(skb->protocol)) {
5356 /* Outer header is 802.1P with vlan 0, inner header is
5357 * 802.1Q or 802.1AD and vlan_do_receive() above could
5358 * not find vlan dev for vlan id 0.
5360 __vlan_hwaccel_clear_tag(skb);
5361 skb = skb_vlan_untag(skb);
5364 if (vlan_do_receive(&skb))
5365 /* After stripping off 802.1P header with vlan 0
5366 * vlan dev is found for inner header.
5369 else if (unlikely(!skb))
5372 /* We have stripped outer 802.1P vlan 0 header.
5373 * But could not find vlan dev.
5374 * check again for vlan id to set OTHERHOST.
5378 /* Note: we might in the future use prio bits
5379 * and set skb->priority like in vlan_do_receive()
5380 * For the time being, just ignore Priority Code Point
5382 __vlan_hwaccel_clear_tag(skb);
5385 type = skb->protocol;
5387 /* deliver only exact match when indicated */
5388 if (likely(!deliver_exact)) {
5389 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5390 &ptype_base[ntohs(type) &
5394 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5395 &orig_dev->ptype_specific);
5397 if (unlikely(skb->dev != orig_dev)) {
5398 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5399 &skb->dev->ptype_specific);
5403 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5405 *ppt_prev = pt_prev;
5409 atomic_long_inc(&skb->dev->rx_dropped);
5411 atomic_long_inc(&skb->dev->rx_nohandler);
5413 /* Jamal, now you will not able to escape explaining
5414 * me how you were going to use this. :-)
5420 /* The invariant here is that if *ppt_prev is not NULL
5421 * then skb should also be non-NULL.
5423 * Apparently *ppt_prev assignment above holds this invariant due to
5424 * skb dereferencing near it.
5430 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5432 struct net_device *orig_dev = skb->dev;
5433 struct packet_type *pt_prev = NULL;
5436 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5438 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5439 skb->dev, pt_prev, orig_dev);
5444 * netif_receive_skb_core - special purpose version of netif_receive_skb
5445 * @skb: buffer to process
5447 * More direct receive version of netif_receive_skb(). It should
5448 * only be used by callers that have a need to skip RPS and Generic XDP.
5449 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5451 * This function may only be called from softirq context and interrupts
5452 * should be enabled.
5454 * Return values (usually ignored):
5455 * NET_RX_SUCCESS: no congestion
5456 * NET_RX_DROP: packet was dropped
5458 int netif_receive_skb_core(struct sk_buff *skb)
5463 ret = __netif_receive_skb_one_core(skb, false);
5468 EXPORT_SYMBOL(netif_receive_skb_core);
5470 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5471 struct packet_type *pt_prev,
5472 struct net_device *orig_dev)
5474 struct sk_buff *skb, *next;
5478 if (list_empty(head))
5480 if (pt_prev->list_func != NULL)
5481 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5482 ip_list_rcv, head, pt_prev, orig_dev);
5484 list_for_each_entry_safe(skb, next, head, list) {
5485 skb_list_del_init(skb);
5486 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5490 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5492 /* Fast-path assumptions:
5493 * - There is no RX handler.
5494 * - Only one packet_type matches.
5495 * If either of these fails, we will end up doing some per-packet
5496 * processing in-line, then handling the 'last ptype' for the whole
5497 * sublist. This can't cause out-of-order delivery to any single ptype,
5498 * because the 'last ptype' must be constant across the sublist, and all
5499 * other ptypes are handled per-packet.
5501 /* Current (common) ptype of sublist */
5502 struct packet_type *pt_curr = NULL;
5503 /* Current (common) orig_dev of sublist */
5504 struct net_device *od_curr = NULL;
5505 struct list_head sublist;
5506 struct sk_buff *skb, *next;
5508 INIT_LIST_HEAD(&sublist);
5509 list_for_each_entry_safe(skb, next, head, list) {
5510 struct net_device *orig_dev = skb->dev;
5511 struct packet_type *pt_prev = NULL;
5513 skb_list_del_init(skb);
5514 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5517 if (pt_curr != pt_prev || od_curr != orig_dev) {
5518 /* dispatch old sublist */
5519 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5520 /* start new sublist */
5521 INIT_LIST_HEAD(&sublist);
5525 list_add_tail(&skb->list, &sublist);
5528 /* dispatch final sublist */
5529 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5532 static int __netif_receive_skb(struct sk_buff *skb)
5536 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5537 unsigned int noreclaim_flag;
5540 * PFMEMALLOC skbs are special, they should
5541 * - be delivered to SOCK_MEMALLOC sockets only
5542 * - stay away from userspace
5543 * - have bounded memory usage
5545 * Use PF_MEMALLOC as this saves us from propagating the allocation
5546 * context down to all allocation sites.
5548 noreclaim_flag = memalloc_noreclaim_save();
5549 ret = __netif_receive_skb_one_core(skb, true);
5550 memalloc_noreclaim_restore(noreclaim_flag);
5552 ret = __netif_receive_skb_one_core(skb, false);
5557 static void __netif_receive_skb_list(struct list_head *head)
5559 unsigned long noreclaim_flag = 0;
5560 struct sk_buff *skb, *next;
5561 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5563 list_for_each_entry_safe(skb, next, head, list) {
5564 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5565 struct list_head sublist;
5567 /* Handle the previous sublist */
5568 list_cut_before(&sublist, head, &skb->list);
5569 if (!list_empty(&sublist))
5570 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5571 pfmemalloc = !pfmemalloc;
5572 /* See comments in __netif_receive_skb */
5574 noreclaim_flag = memalloc_noreclaim_save();
5576 memalloc_noreclaim_restore(noreclaim_flag);
5579 /* Handle the remaining sublist */
5580 if (!list_empty(head))
5581 __netif_receive_skb_list_core(head, pfmemalloc);
5582 /* Restore pflags */
5584 memalloc_noreclaim_restore(noreclaim_flag);
5587 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5589 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5590 struct bpf_prog *new = xdp->prog;
5593 switch (xdp->command) {
5594 case XDP_SETUP_PROG:
5595 rcu_assign_pointer(dev->xdp_prog, new);
5600 static_branch_dec(&generic_xdp_needed_key);
5601 } else if (new && !old) {
5602 static_branch_inc(&generic_xdp_needed_key);
5603 dev_disable_lro(dev);
5604 dev_disable_gro_hw(dev);
5616 static int netif_receive_skb_internal(struct sk_buff *skb)
5620 net_timestamp_check(netdev_tstamp_prequeue, skb);
5622 if (skb_defer_rx_timestamp(skb))
5623 return NET_RX_SUCCESS;
5627 if (static_branch_unlikely(&rps_needed)) {
5628 struct rps_dev_flow voidflow, *rflow = &voidflow;
5629 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5632 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5638 ret = __netif_receive_skb(skb);
5643 static void netif_receive_skb_list_internal(struct list_head *head)
5645 struct sk_buff *skb, *next;
5646 struct list_head sublist;
5648 INIT_LIST_HEAD(&sublist);
5649 list_for_each_entry_safe(skb, next, head, list) {
5650 net_timestamp_check(netdev_tstamp_prequeue, skb);
5651 skb_list_del_init(skb);
5652 if (!skb_defer_rx_timestamp(skb))
5653 list_add_tail(&skb->list, &sublist);
5655 list_splice_init(&sublist, head);
5659 if (static_branch_unlikely(&rps_needed)) {
5660 list_for_each_entry_safe(skb, next, head, list) {
5661 struct rps_dev_flow voidflow, *rflow = &voidflow;
5662 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5665 /* Will be handled, remove from list */
5666 skb_list_del_init(skb);
5667 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5672 __netif_receive_skb_list(head);
5677 * netif_receive_skb - process receive buffer from network
5678 * @skb: buffer to process
5680 * netif_receive_skb() is the main receive data processing function.
5681 * It always succeeds. The buffer may be dropped during processing
5682 * for congestion control or by the protocol layers.
5684 * This function may only be called from softirq context and interrupts
5685 * should be enabled.
5687 * Return values (usually ignored):
5688 * NET_RX_SUCCESS: no congestion
5689 * NET_RX_DROP: packet was dropped
5691 int netif_receive_skb(struct sk_buff *skb)
5695 trace_netif_receive_skb_entry(skb);
5697 ret = netif_receive_skb_internal(skb);
5698 trace_netif_receive_skb_exit(ret);
5702 EXPORT_SYMBOL(netif_receive_skb);
5705 * netif_receive_skb_list - process many receive buffers from network
5706 * @head: list of skbs to process.
5708 * Since return value of netif_receive_skb() is normally ignored, and
5709 * wouldn't be meaningful for a list, this function returns void.
5711 * This function may only be called from softirq context and interrupts
5712 * should be enabled.
5714 void netif_receive_skb_list(struct list_head *head)
5716 struct sk_buff *skb;
5718 if (list_empty(head))
5720 if (trace_netif_receive_skb_list_entry_enabled()) {
5721 list_for_each_entry(skb, head, list)
5722 trace_netif_receive_skb_list_entry(skb);
5724 netif_receive_skb_list_internal(head);
5725 trace_netif_receive_skb_list_exit(0);
5727 EXPORT_SYMBOL(netif_receive_skb_list);
5729 static DEFINE_PER_CPU(struct work_struct, flush_works);
5731 /* Network device is going away, flush any packets still pending */
5732 static void flush_backlog(struct work_struct *work)
5734 struct sk_buff *skb, *tmp;
5735 struct softnet_data *sd;
5738 sd = this_cpu_ptr(&softnet_data);
5740 local_irq_disable();
5742 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5743 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5744 __skb_unlink(skb, &sd->input_pkt_queue);
5745 dev_kfree_skb_irq(skb);
5746 input_queue_head_incr(sd);
5752 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5753 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5754 __skb_unlink(skb, &sd->process_queue);
5756 input_queue_head_incr(sd);
5762 static bool flush_required(int cpu)
5764 #if IS_ENABLED(CONFIG_RPS)
5765 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5768 local_irq_disable();
5771 /* as insertion into process_queue happens with the rps lock held,
5772 * process_queue access may race only with dequeue
5774 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5775 !skb_queue_empty_lockless(&sd->process_queue);
5781 /* without RPS we can't safely check input_pkt_queue: during a
5782 * concurrent remote skb_queue_splice() we can detect as empty both
5783 * input_pkt_queue and process_queue even if the latter could end-up
5784 * containing a lot of packets.
5789 static void flush_all_backlogs(void)
5791 static cpumask_t flush_cpus;
5794 /* since we are under rtnl lock protection we can use static data
5795 * for the cpumask and avoid allocating on stack the possibly
5802 cpumask_clear(&flush_cpus);
5803 for_each_online_cpu(cpu) {
5804 if (flush_required(cpu)) {
5805 queue_work_on(cpu, system_highpri_wq,
5806 per_cpu_ptr(&flush_works, cpu));
5807 cpumask_set_cpu(cpu, &flush_cpus);
5811 /* we can have in flight packet[s] on the cpus we are not flushing,
5812 * synchronize_net() in unregister_netdevice_many() will take care of
5815 for_each_cpu(cpu, &flush_cpus)
5816 flush_work(per_cpu_ptr(&flush_works, cpu));
5821 /* Pass the currently batched GRO_NORMAL SKBs up to the stack. */
5822 static void gro_normal_list(struct napi_struct *napi)
5824 if (!napi->rx_count)
5826 netif_receive_skb_list_internal(&napi->rx_list);
5827 INIT_LIST_HEAD(&napi->rx_list);
5831 /* Queue one GRO_NORMAL SKB up for list processing. If batch size exceeded,
5832 * pass the whole batch up to the stack.
5834 static void gro_normal_one(struct napi_struct *napi, struct sk_buff *skb, int segs)
5836 list_add_tail(&skb->list, &napi->rx_list);
5837 napi->rx_count += segs;
5838 if (napi->rx_count >= gro_normal_batch)
5839 gro_normal_list(napi);
5842 static int napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
5844 struct packet_offload *ptype;
5845 __be16 type = skb->protocol;
5846 struct list_head *head = &offload_base;
5849 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
5851 if (NAPI_GRO_CB(skb)->count == 1) {
5852 skb_shinfo(skb)->gso_size = 0;
5857 list_for_each_entry_rcu(ptype, head, list) {
5858 if (ptype->type != type || !ptype->callbacks.gro_complete)
5861 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
5862 ipv6_gro_complete, inet_gro_complete,
5869 WARN_ON(&ptype->list == head);
5871 return NET_RX_SUCCESS;
5875 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
5876 return NET_RX_SUCCESS;
5879 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
5882 struct list_head *head = &napi->gro_hash[index].list;
5883 struct sk_buff *skb, *p;
5885 list_for_each_entry_safe_reverse(skb, p, head, list) {
5886 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
5888 skb_list_del_init(skb);
5889 napi_gro_complete(napi, skb);
5890 napi->gro_hash[index].count--;
5893 if (!napi->gro_hash[index].count)
5894 __clear_bit(index, &napi->gro_bitmask);
5897 /* napi->gro_hash[].list contains packets ordered by age.
5898 * youngest packets at the head of it.
5899 * Complete skbs in reverse order to reduce latencies.
5901 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
5903 unsigned long bitmask = napi->gro_bitmask;
5904 unsigned int i, base = ~0U;
5906 while ((i = ffs(bitmask)) != 0) {
5909 __napi_gro_flush_chain(napi, base, flush_old);
5912 EXPORT_SYMBOL(napi_gro_flush);
5914 static void gro_list_prepare(const struct list_head *head,
5915 const struct sk_buff *skb)
5917 unsigned int maclen = skb->dev->hard_header_len;
5918 u32 hash = skb_get_hash_raw(skb);
5921 list_for_each_entry(p, head, list) {
5922 unsigned long diffs;
5924 NAPI_GRO_CB(p)->flush = 0;
5926 if (hash != skb_get_hash_raw(p)) {
5927 NAPI_GRO_CB(p)->same_flow = 0;
5931 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
5932 diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb);
5933 if (skb_vlan_tag_present(p))
5934 diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb);
5935 diffs |= skb_metadata_differs(p, skb);
5936 if (maclen == ETH_HLEN)
5937 diffs |= compare_ether_header(skb_mac_header(p),
5938 skb_mac_header(skb));
5940 diffs = memcmp(skb_mac_header(p),
5941 skb_mac_header(skb),
5944 /* in most common scenarions 'slow_gro' is 0
5945 * otherwise we are already on some slower paths
5946 * either skip all the infrequent tests altogether or
5947 * avoid trying too hard to skip each of them individually
5949 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
5950 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
5951 struct tc_skb_ext *skb_ext;
5952 struct tc_skb_ext *p_ext;
5955 diffs |= p->sk != skb->sk;
5956 diffs |= skb_metadata_dst_cmp(p, skb);
5957 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
5959 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
5960 skb_ext = skb_ext_find(skb, TC_SKB_EXT);
5961 p_ext = skb_ext_find(p, TC_SKB_EXT);
5963 diffs |= (!!p_ext) ^ (!!skb_ext);
5964 if (!diffs && unlikely(skb_ext))
5965 diffs |= p_ext->chain ^ skb_ext->chain;
5969 NAPI_GRO_CB(p)->same_flow = !diffs;
5973 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
5975 const struct skb_shared_info *pinfo = skb_shinfo(skb);
5976 const skb_frag_t *frag0 = &pinfo->frags[0];
5978 NAPI_GRO_CB(skb)->data_offset = 0;
5979 NAPI_GRO_CB(skb)->frag0 = NULL;
5980 NAPI_GRO_CB(skb)->frag0_len = 0;
5982 if (!skb_headlen(skb) && pinfo->nr_frags &&
5983 !PageHighMem(skb_frag_page(frag0)) &&
5984 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
5985 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
5986 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
5987 skb_frag_size(frag0),
5988 skb->end - skb->tail);
5992 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
5994 struct skb_shared_info *pinfo = skb_shinfo(skb);
5996 BUG_ON(skb->end - skb->tail < grow);
5998 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
6000 skb->data_len -= grow;
6003 skb_frag_off_add(&pinfo->frags[0], grow);
6004 skb_frag_size_sub(&pinfo->frags[0], grow);
6006 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
6007 skb_frag_unref(skb, 0);
6008 memmove(pinfo->frags, pinfo->frags + 1,
6009 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
6013 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
6015 struct sk_buff *oldest;
6017 oldest = list_last_entry(head, struct sk_buff, list);
6019 /* We are called with head length >= MAX_GRO_SKBS, so this is
6022 if (WARN_ON_ONCE(!oldest))
6025 /* Do not adjust napi->gro_hash[].count, caller is adding a new
6028 skb_list_del_init(oldest);
6029 napi_gro_complete(napi, oldest);
6032 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6034 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
6035 struct gro_list *gro_list = &napi->gro_hash[bucket];
6036 struct list_head *head = &offload_base;
6037 struct packet_offload *ptype;
6038 __be16 type = skb->protocol;
6039 struct sk_buff *pp = NULL;
6040 enum gro_result ret;
6044 if (netif_elide_gro(skb->dev))
6047 gro_list_prepare(&gro_list->list, skb);
6050 list_for_each_entry_rcu(ptype, head, list) {
6051 if (ptype->type != type || !ptype->callbacks.gro_receive)
6054 skb_set_network_header(skb, skb_gro_offset(skb));
6055 skb_reset_mac_len(skb);
6056 NAPI_GRO_CB(skb)->same_flow = 0;
6057 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
6058 NAPI_GRO_CB(skb)->free = 0;
6059 NAPI_GRO_CB(skb)->encap_mark = 0;
6060 NAPI_GRO_CB(skb)->recursion_counter = 0;
6061 NAPI_GRO_CB(skb)->is_fou = 0;
6062 NAPI_GRO_CB(skb)->is_atomic = 1;
6063 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
6065 /* Setup for GRO checksum validation */
6066 switch (skb->ip_summed) {
6067 case CHECKSUM_COMPLETE:
6068 NAPI_GRO_CB(skb)->csum = skb->csum;
6069 NAPI_GRO_CB(skb)->csum_valid = 1;
6070 NAPI_GRO_CB(skb)->csum_cnt = 0;
6072 case CHECKSUM_UNNECESSARY:
6073 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
6074 NAPI_GRO_CB(skb)->csum_valid = 0;
6077 NAPI_GRO_CB(skb)->csum_cnt = 0;
6078 NAPI_GRO_CB(skb)->csum_valid = 0;
6081 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
6082 ipv6_gro_receive, inet_gro_receive,
6083 &gro_list->list, skb);
6088 if (&ptype->list == head)
6091 if (PTR_ERR(pp) == -EINPROGRESS) {
6096 same_flow = NAPI_GRO_CB(skb)->same_flow;
6097 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
6100 skb_list_del_init(pp);
6101 napi_gro_complete(napi, pp);
6108 if (NAPI_GRO_CB(skb)->flush)
6111 if (unlikely(gro_list->count >= MAX_GRO_SKBS))
6112 gro_flush_oldest(napi, &gro_list->list);
6116 NAPI_GRO_CB(skb)->count = 1;
6117 NAPI_GRO_CB(skb)->age = jiffies;
6118 NAPI_GRO_CB(skb)->last = skb;
6119 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
6120 list_add(&skb->list, &gro_list->list);
6124 grow = skb_gro_offset(skb) - skb_headlen(skb);
6126 gro_pull_from_frag0(skb, grow);
6128 if (gro_list->count) {
6129 if (!test_bit(bucket, &napi->gro_bitmask))
6130 __set_bit(bucket, &napi->gro_bitmask);
6131 } else if (test_bit(bucket, &napi->gro_bitmask)) {
6132 __clear_bit(bucket, &napi->gro_bitmask);
6142 struct packet_offload *gro_find_receive_by_type(__be16 type)
6144 struct list_head *offload_head = &offload_base;
6145 struct packet_offload *ptype;
6147 list_for_each_entry_rcu(ptype, offload_head, list) {
6148 if (ptype->type != type || !ptype->callbacks.gro_receive)
6154 EXPORT_SYMBOL(gro_find_receive_by_type);
6156 struct packet_offload *gro_find_complete_by_type(__be16 type)
6158 struct list_head *offload_head = &offload_base;
6159 struct packet_offload *ptype;
6161 list_for_each_entry_rcu(ptype, offload_head, list) {
6162 if (ptype->type != type || !ptype->callbacks.gro_complete)
6168 EXPORT_SYMBOL(gro_find_complete_by_type);
6170 static gro_result_t napi_skb_finish(struct napi_struct *napi,
6171 struct sk_buff *skb,
6176 gro_normal_one(napi, skb, 1);
6179 case GRO_MERGED_FREE:
6180 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6181 napi_skb_free_stolen_head(skb);
6182 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
6185 __kfree_skb_defer(skb);
6197 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6201 skb_mark_napi_id(skb, napi);
6202 trace_napi_gro_receive_entry(skb);
6204 skb_gro_reset_offset(skb, 0);
6206 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
6207 trace_napi_gro_receive_exit(ret);
6211 EXPORT_SYMBOL(napi_gro_receive);
6213 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
6215 if (unlikely(skb->pfmemalloc)) {
6219 __skb_pull(skb, skb_headlen(skb));
6220 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
6221 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
6222 __vlan_hwaccel_clear_tag(skb);
6223 skb->dev = napi->dev;
6226 /* eth_type_trans() assumes pkt_type is PACKET_HOST */
6227 skb->pkt_type = PACKET_HOST;
6229 skb->encapsulation = 0;
6230 skb_shinfo(skb)->gso_type = 0;
6231 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
6232 if (unlikely(skb->slow_gro)) {
6242 struct sk_buff *napi_get_frags(struct napi_struct *napi)
6244 struct sk_buff *skb = napi->skb;
6247 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
6250 skb_mark_napi_id(skb, napi);
6255 EXPORT_SYMBOL(napi_get_frags);
6257 static gro_result_t napi_frags_finish(struct napi_struct *napi,
6258 struct sk_buff *skb,
6264 __skb_push(skb, ETH_HLEN);
6265 skb->protocol = eth_type_trans(skb, skb->dev);
6266 if (ret == GRO_NORMAL)
6267 gro_normal_one(napi, skb, 1);
6270 case GRO_MERGED_FREE:
6271 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6272 napi_skb_free_stolen_head(skb);
6274 napi_reuse_skb(napi, skb);
6285 /* Upper GRO stack assumes network header starts at gro_offset=0
6286 * Drivers could call both napi_gro_frags() and napi_gro_receive()
6287 * We copy ethernet header into skb->data to have a common layout.
6289 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
6291 struct sk_buff *skb = napi->skb;
6292 const struct ethhdr *eth;
6293 unsigned int hlen = sizeof(*eth);
6297 skb_reset_mac_header(skb);
6298 skb_gro_reset_offset(skb, hlen);
6300 if (unlikely(skb_gro_header_hard(skb, hlen))) {
6301 eth = skb_gro_header_slow(skb, hlen, 0);
6302 if (unlikely(!eth)) {
6303 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
6304 __func__, napi->dev->name);
6305 napi_reuse_skb(napi, skb);
6309 eth = (const struct ethhdr *)skb->data;
6310 gro_pull_from_frag0(skb, hlen);
6311 NAPI_GRO_CB(skb)->frag0 += hlen;
6312 NAPI_GRO_CB(skb)->frag0_len -= hlen;
6314 __skb_pull(skb, hlen);
6317 * This works because the only protocols we care about don't require
6319 * We'll fix it up properly in napi_frags_finish()
6321 skb->protocol = eth->h_proto;
6326 gro_result_t napi_gro_frags(struct napi_struct *napi)
6329 struct sk_buff *skb = napi_frags_skb(napi);
6331 trace_napi_gro_frags_entry(skb);
6333 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
6334 trace_napi_gro_frags_exit(ret);
6338 EXPORT_SYMBOL(napi_gro_frags);
6340 /* Compute the checksum from gro_offset and return the folded value
6341 * after adding in any pseudo checksum.
6343 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
6348 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
6350 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
6351 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
6352 /* See comments in __skb_checksum_complete(). */
6354 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
6355 !skb->csum_complete_sw)
6356 netdev_rx_csum_fault(skb->dev, skb);
6359 NAPI_GRO_CB(skb)->csum = wsum;
6360 NAPI_GRO_CB(skb)->csum_valid = 1;
6364 EXPORT_SYMBOL(__skb_gro_checksum_complete);
6366 static void net_rps_send_ipi(struct softnet_data *remsd)
6370 struct softnet_data *next = remsd->rps_ipi_next;
6372 if (cpu_online(remsd->cpu))
6373 smp_call_function_single_async(remsd->cpu, &remsd->csd);
6380 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
6381 * Note: called with local irq disabled, but exits with local irq enabled.
6383 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
6386 struct softnet_data *remsd = sd->rps_ipi_list;
6389 sd->rps_ipi_list = NULL;
6393 /* Send pending IPI's to kick RPS processing on remote cpus. */
6394 net_rps_send_ipi(remsd);
6400 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
6403 return sd->rps_ipi_list != NULL;
6409 static int process_backlog(struct napi_struct *napi, int quota)
6411 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
6415 /* Check if we have pending ipi, its better to send them now,
6416 * not waiting net_rx_action() end.
6418 if (sd_has_rps_ipi_waiting(sd)) {
6419 local_irq_disable();
6420 net_rps_action_and_irq_enable(sd);
6423 napi->weight = dev_rx_weight;
6425 struct sk_buff *skb;
6427 while ((skb = __skb_dequeue(&sd->process_queue))) {
6429 __netif_receive_skb(skb);
6431 input_queue_head_incr(sd);
6432 if (++work >= quota)
6437 local_irq_disable();
6439 if (skb_queue_empty(&sd->input_pkt_queue)) {
6441 * Inline a custom version of __napi_complete().
6442 * only current cpu owns and manipulates this napi,
6443 * and NAPI_STATE_SCHED is the only possible flag set
6445 * We can use a plain write instead of clear_bit(),
6446 * and we dont need an smp_mb() memory barrier.
6451 skb_queue_splice_tail_init(&sd->input_pkt_queue,
6452 &sd->process_queue);
6462 * __napi_schedule - schedule for receive
6463 * @n: entry to schedule
6465 * The entry's receive function will be scheduled to run.
6466 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
6468 void __napi_schedule(struct napi_struct *n)
6470 unsigned long flags;
6472 local_irq_save(flags);
6473 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6474 local_irq_restore(flags);
6476 EXPORT_SYMBOL(__napi_schedule);
6479 * napi_schedule_prep - check if napi can be scheduled
6482 * Test if NAPI routine is already running, and if not mark
6483 * it as running. This is used as a condition variable to
6484 * insure only one NAPI poll instance runs. We also make
6485 * sure there is no pending NAPI disable.
6487 bool napi_schedule_prep(struct napi_struct *n)
6489 unsigned long val, new;
6492 val = READ_ONCE(n->state);
6493 if (unlikely(val & NAPIF_STATE_DISABLE))
6495 new = val | NAPIF_STATE_SCHED;
6497 /* Sets STATE_MISSED bit if STATE_SCHED was already set
6498 * This was suggested by Alexander Duyck, as compiler
6499 * emits better code than :
6500 * if (val & NAPIF_STATE_SCHED)
6501 * new |= NAPIF_STATE_MISSED;
6503 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6505 } while (cmpxchg(&n->state, val, new) != val);
6507 return !(val & NAPIF_STATE_SCHED);
6509 EXPORT_SYMBOL(napi_schedule_prep);
6512 * __napi_schedule_irqoff - schedule for receive
6513 * @n: entry to schedule
6515 * Variant of __napi_schedule() assuming hard irqs are masked.
6517 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6518 * because the interrupt disabled assumption might not be true
6519 * due to force-threaded interrupts and spinlock substitution.
6521 void __napi_schedule_irqoff(struct napi_struct *n)
6523 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6524 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6528 EXPORT_SYMBOL(__napi_schedule_irqoff);
6530 bool napi_complete_done(struct napi_struct *n, int work_done)
6532 unsigned long flags, val, new, timeout = 0;
6536 * 1) Don't let napi dequeue from the cpu poll list
6537 * just in case its running on a different cpu.
6538 * 2) If we are busy polling, do nothing here, we have
6539 * the guarantee we will be called later.
6541 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6542 NAPIF_STATE_IN_BUSY_POLL)))
6547 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6548 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6550 if (n->defer_hard_irqs_count > 0) {
6551 n->defer_hard_irqs_count--;
6552 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6556 if (n->gro_bitmask) {
6557 /* When the NAPI instance uses a timeout and keeps postponing
6558 * it, we need to bound somehow the time packets are kept in
6561 napi_gro_flush(n, !!timeout);
6566 if (unlikely(!list_empty(&n->poll_list))) {
6567 /* If n->poll_list is not empty, we need to mask irqs */
6568 local_irq_save(flags);
6569 list_del_init(&n->poll_list);
6570 local_irq_restore(flags);
6574 val = READ_ONCE(n->state);
6576 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6578 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6579 NAPIF_STATE_SCHED_THREADED |
6580 NAPIF_STATE_PREFER_BUSY_POLL);
6582 /* If STATE_MISSED was set, leave STATE_SCHED set,
6583 * because we will call napi->poll() one more time.
6584 * This C code was suggested by Alexander Duyck to help gcc.
6586 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6588 } while (cmpxchg(&n->state, val, new) != val);
6590 if (unlikely(val & NAPIF_STATE_MISSED)) {
6596 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6597 HRTIMER_MODE_REL_PINNED);
6600 EXPORT_SYMBOL(napi_complete_done);
6602 /* must be called under rcu_read_lock(), as we dont take a reference */
6603 static struct napi_struct *napi_by_id(unsigned int napi_id)
6605 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6606 struct napi_struct *napi;
6608 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6609 if (napi->napi_id == napi_id)
6615 #if defined(CONFIG_NET_RX_BUSY_POLL)
6617 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6619 if (!skip_schedule) {
6620 gro_normal_list(napi);
6621 __napi_schedule(napi);
6625 if (napi->gro_bitmask) {
6626 /* flush too old packets
6627 * If HZ < 1000, flush all packets.
6629 napi_gro_flush(napi, HZ >= 1000);
6632 gro_normal_list(napi);
6633 clear_bit(NAPI_STATE_SCHED, &napi->state);
6636 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6639 bool skip_schedule = false;
6640 unsigned long timeout;
6643 /* Busy polling means there is a high chance device driver hard irq
6644 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6645 * set in napi_schedule_prep().
6646 * Since we are about to call napi->poll() once more, we can safely
6647 * clear NAPI_STATE_MISSED.
6649 * Note: x86 could use a single "lock and ..." instruction
6650 * to perform these two clear_bit()
6652 clear_bit(NAPI_STATE_MISSED, &napi->state);
6653 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6657 if (prefer_busy_poll) {
6658 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6659 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6660 if (napi->defer_hard_irqs_count && timeout) {
6661 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6662 skip_schedule = true;
6666 /* All we really want here is to re-enable device interrupts.
6667 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6669 rc = napi->poll(napi, budget);
6670 /* We can't gro_normal_list() here, because napi->poll() might have
6671 * rearmed the napi (napi_complete_done()) in which case it could
6672 * already be running on another CPU.
6674 trace_napi_poll(napi, rc, budget);
6675 netpoll_poll_unlock(have_poll_lock);
6677 __busy_poll_stop(napi, skip_schedule);
6681 void napi_busy_loop(unsigned int napi_id,
6682 bool (*loop_end)(void *, unsigned long),
6683 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6685 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6686 int (*napi_poll)(struct napi_struct *napi, int budget);
6687 void *have_poll_lock = NULL;
6688 struct napi_struct *napi;
6695 napi = napi_by_id(napi_id);
6705 unsigned long val = READ_ONCE(napi->state);
6707 /* If multiple threads are competing for this napi,
6708 * we avoid dirtying napi->state as much as we can.
6710 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6711 NAPIF_STATE_IN_BUSY_POLL)) {
6712 if (prefer_busy_poll)
6713 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6716 if (cmpxchg(&napi->state, val,
6717 val | NAPIF_STATE_IN_BUSY_POLL |
6718 NAPIF_STATE_SCHED) != val) {
6719 if (prefer_busy_poll)
6720 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6723 have_poll_lock = netpoll_poll_lock(napi);
6724 napi_poll = napi->poll;
6726 work = napi_poll(napi, budget);
6727 trace_napi_poll(napi, work, budget);
6728 gro_normal_list(napi);
6731 __NET_ADD_STATS(dev_net(napi->dev),
6732 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6735 if (!loop_end || loop_end(loop_end_arg, start_time))
6738 if (unlikely(need_resched())) {
6740 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6744 if (loop_end(loop_end_arg, start_time))
6751 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6756 EXPORT_SYMBOL(napi_busy_loop);
6758 #endif /* CONFIG_NET_RX_BUSY_POLL */
6760 static void napi_hash_add(struct napi_struct *napi)
6762 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6765 spin_lock(&napi_hash_lock);
6767 /* 0..NR_CPUS range is reserved for sender_cpu use */
6769 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6770 napi_gen_id = MIN_NAPI_ID;
6771 } while (napi_by_id(napi_gen_id));
6772 napi->napi_id = napi_gen_id;
6774 hlist_add_head_rcu(&napi->napi_hash_node,
6775 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6777 spin_unlock(&napi_hash_lock);
6780 /* Warning : caller is responsible to make sure rcu grace period
6781 * is respected before freeing memory containing @napi
6783 static void napi_hash_del(struct napi_struct *napi)
6785 spin_lock(&napi_hash_lock);
6787 hlist_del_init_rcu(&napi->napi_hash_node);
6789 spin_unlock(&napi_hash_lock);
6792 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6794 struct napi_struct *napi;
6796 napi = container_of(timer, struct napi_struct, timer);
6798 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6799 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6801 if (!napi_disable_pending(napi) &&
6802 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6803 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6804 __napi_schedule_irqoff(napi);
6807 return HRTIMER_NORESTART;
6810 static void init_gro_hash(struct napi_struct *napi)
6814 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6815 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6816 napi->gro_hash[i].count = 0;
6818 napi->gro_bitmask = 0;
6821 int dev_set_threaded(struct net_device *dev, bool threaded)
6823 struct napi_struct *napi;
6826 if (dev->threaded == threaded)
6830 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6831 if (!napi->thread) {
6832 err = napi_kthread_create(napi);
6841 dev->threaded = threaded;
6843 /* Make sure kthread is created before THREADED bit
6846 smp_mb__before_atomic();
6848 /* Setting/unsetting threaded mode on a napi might not immediately
6849 * take effect, if the current napi instance is actively being
6850 * polled. In this case, the switch between threaded mode and
6851 * softirq mode will happen in the next round of napi_schedule().
6852 * This should not cause hiccups/stalls to the live traffic.
6854 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6856 set_bit(NAPI_STATE_THREADED, &napi->state);
6858 clear_bit(NAPI_STATE_THREADED, &napi->state);
6863 EXPORT_SYMBOL(dev_set_threaded);
6865 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
6866 int (*poll)(struct napi_struct *, int), int weight)
6868 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6871 INIT_LIST_HEAD(&napi->poll_list);
6872 INIT_HLIST_NODE(&napi->napi_hash_node);
6873 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6874 napi->timer.function = napi_watchdog;
6875 init_gro_hash(napi);
6877 INIT_LIST_HEAD(&napi->rx_list);
6880 if (weight > NAPI_POLL_WEIGHT)
6881 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6883 napi->weight = weight;
6885 #ifdef CONFIG_NETPOLL
6886 napi->poll_owner = -1;
6888 set_bit(NAPI_STATE_SCHED, &napi->state);
6889 set_bit(NAPI_STATE_NPSVC, &napi->state);
6890 list_add_rcu(&napi->dev_list, &dev->napi_list);
6891 napi_hash_add(napi);
6892 /* Create kthread for this napi if dev->threaded is set.
6893 * Clear dev->threaded if kthread creation failed so that
6894 * threaded mode will not be enabled in napi_enable().
6896 if (dev->threaded && napi_kthread_create(napi))
6899 EXPORT_SYMBOL(netif_napi_add);
6901 void napi_disable(struct napi_struct *n)
6903 unsigned long val, new;
6906 set_bit(NAPI_STATE_DISABLE, &n->state);
6909 val = READ_ONCE(n->state);
6910 if (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
6911 usleep_range(20, 200);
6915 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC;
6916 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL);
6917 } while (cmpxchg(&n->state, val, new) != val);
6919 hrtimer_cancel(&n->timer);
6921 clear_bit(NAPI_STATE_DISABLE, &n->state);
6923 EXPORT_SYMBOL(napi_disable);
6926 * napi_enable - enable NAPI scheduling
6929 * Resume NAPI from being scheduled on this context.
6930 * Must be paired with napi_disable.
6932 void napi_enable(struct napi_struct *n)
6934 unsigned long val, new;
6937 val = READ_ONCE(n->state);
6938 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
6940 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
6941 if (n->dev->threaded && n->thread)
6942 new |= NAPIF_STATE_THREADED;
6943 } while (cmpxchg(&n->state, val, new) != val);
6945 EXPORT_SYMBOL(napi_enable);
6947 static void flush_gro_hash(struct napi_struct *napi)
6951 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6952 struct sk_buff *skb, *n;
6954 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6956 napi->gro_hash[i].count = 0;
6960 /* Must be called in process context */
6961 void __netif_napi_del(struct napi_struct *napi)
6963 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
6966 napi_hash_del(napi);
6967 list_del_rcu(&napi->dev_list);
6968 napi_free_frags(napi);
6970 flush_gro_hash(napi);
6971 napi->gro_bitmask = 0;
6974 kthread_stop(napi->thread);
6975 napi->thread = NULL;
6978 EXPORT_SYMBOL(__netif_napi_del);
6980 static int __napi_poll(struct napi_struct *n, bool *repoll)
6986 /* This NAPI_STATE_SCHED test is for avoiding a race
6987 * with netpoll's poll_napi(). Only the entity which
6988 * obtains the lock and sees NAPI_STATE_SCHED set will
6989 * actually make the ->poll() call. Therefore we avoid
6990 * accidentally calling ->poll() when NAPI is not scheduled.
6993 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6994 work = n->poll(n, weight);
6995 trace_napi_poll(n, work, weight);
6998 if (unlikely(work > weight))
6999 pr_err_once("NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
7000 n->poll, work, weight);
7002 if (likely(work < weight))
7005 /* Drivers must not modify the NAPI state if they
7006 * consume the entire weight. In such cases this code
7007 * still "owns" the NAPI instance and therefore can
7008 * move the instance around on the list at-will.
7010 if (unlikely(napi_disable_pending(n))) {
7015 /* The NAPI context has more processing work, but busy-polling
7016 * is preferred. Exit early.
7018 if (napi_prefer_busy_poll(n)) {
7019 if (napi_complete_done(n, work)) {
7020 /* If timeout is not set, we need to make sure
7021 * that the NAPI is re-scheduled.
7028 if (n->gro_bitmask) {
7029 /* flush too old packets
7030 * If HZ < 1000, flush all packets.
7032 napi_gro_flush(n, HZ >= 1000);
7037 /* Some drivers may have called napi_schedule
7038 * prior to exhausting their budget.
7040 if (unlikely(!list_empty(&n->poll_list))) {
7041 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
7042 n->dev ? n->dev->name : "backlog");
7051 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
7053 bool do_repoll = false;
7057 list_del_init(&n->poll_list);
7059 have = netpoll_poll_lock(n);
7061 work = __napi_poll(n, &do_repoll);
7064 list_add_tail(&n->poll_list, repoll);
7066 netpoll_poll_unlock(have);
7071 static int napi_thread_wait(struct napi_struct *napi)
7075 set_current_state(TASK_INTERRUPTIBLE);
7077 while (!kthread_should_stop()) {
7078 /* Testing SCHED_THREADED bit here to make sure the current
7079 * kthread owns this napi and could poll on this napi.
7080 * Testing SCHED bit is not enough because SCHED bit might be
7081 * set by some other busy poll thread or by napi_disable().
7083 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
7084 WARN_ON(!list_empty(&napi->poll_list));
7085 __set_current_state(TASK_RUNNING);
7090 /* woken being true indicates this thread owns this napi. */
7092 set_current_state(TASK_INTERRUPTIBLE);
7094 __set_current_state(TASK_RUNNING);
7099 static int napi_threaded_poll(void *data)
7101 struct napi_struct *napi = data;
7104 while (!napi_thread_wait(napi)) {
7106 bool repoll = false;
7110 have = netpoll_poll_lock(napi);
7111 __napi_poll(napi, &repoll);
7112 netpoll_poll_unlock(have);
7125 static __latent_entropy void net_rx_action(struct softirq_action *h)
7127 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
7128 unsigned long time_limit = jiffies +
7129 usecs_to_jiffies(netdev_budget_usecs);
7130 int budget = netdev_budget;
7134 local_irq_disable();
7135 list_splice_init(&sd->poll_list, &list);
7139 struct napi_struct *n;
7141 if (list_empty(&list)) {
7142 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
7147 n = list_first_entry(&list, struct napi_struct, poll_list);
7148 budget -= napi_poll(n, &repoll);
7150 /* If softirq window is exhausted then punt.
7151 * Allow this to run for 2 jiffies since which will allow
7152 * an average latency of 1.5/HZ.
7154 if (unlikely(budget <= 0 ||
7155 time_after_eq(jiffies, time_limit))) {
7161 local_irq_disable();
7163 list_splice_tail_init(&sd->poll_list, &list);
7164 list_splice_tail(&repoll, &list);
7165 list_splice(&list, &sd->poll_list);
7166 if (!list_empty(&sd->poll_list))
7167 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
7169 net_rps_action_and_irq_enable(sd);
7172 struct netdev_adjacent {
7173 struct net_device *dev;
7175 /* upper master flag, there can only be one master device per list */
7178 /* lookup ignore flag */
7181 /* counter for the number of times this device was added to us */
7184 /* private field for the users */
7187 struct list_head list;
7188 struct rcu_head rcu;
7191 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
7192 struct list_head *adj_list)
7194 struct netdev_adjacent *adj;
7196 list_for_each_entry(adj, adj_list, list) {
7197 if (adj->dev == adj_dev)
7203 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
7204 struct netdev_nested_priv *priv)
7206 struct net_device *dev = (struct net_device *)priv->data;
7208 return upper_dev == dev;
7212 * netdev_has_upper_dev - Check if device is linked to an upper device
7214 * @upper_dev: upper device to check
7216 * Find out if a device is linked to specified upper device and return true
7217 * in case it is. Note that this checks only immediate upper device,
7218 * not through a complete stack of devices. The caller must hold the RTNL lock.
7220 bool netdev_has_upper_dev(struct net_device *dev,
7221 struct net_device *upper_dev)
7223 struct netdev_nested_priv priv = {
7224 .data = (void *)upper_dev,
7229 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7232 EXPORT_SYMBOL(netdev_has_upper_dev);
7235 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
7237 * @upper_dev: upper device to check
7239 * Find out if a device is linked to specified upper device and return true
7240 * in case it is. Note that this checks the entire upper device chain.
7241 * The caller must hold rcu lock.
7244 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
7245 struct net_device *upper_dev)
7247 struct netdev_nested_priv priv = {
7248 .data = (void *)upper_dev,
7251 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7254 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
7257 * netdev_has_any_upper_dev - Check if device is linked to some device
7260 * Find out if a device is linked to an upper device and return true in case
7261 * it is. The caller must hold the RTNL lock.
7263 bool netdev_has_any_upper_dev(struct net_device *dev)
7267 return !list_empty(&dev->adj_list.upper);
7269 EXPORT_SYMBOL(netdev_has_any_upper_dev);
7272 * netdev_master_upper_dev_get - Get master upper device
7275 * Find a master upper device and return pointer to it or NULL in case
7276 * it's not there. The caller must hold the RTNL lock.
7278 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
7280 struct netdev_adjacent *upper;
7284 if (list_empty(&dev->adj_list.upper))
7287 upper = list_first_entry(&dev->adj_list.upper,
7288 struct netdev_adjacent, list);
7289 if (likely(upper->master))
7293 EXPORT_SYMBOL(netdev_master_upper_dev_get);
7295 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
7297 struct netdev_adjacent *upper;
7301 if (list_empty(&dev->adj_list.upper))
7304 upper = list_first_entry(&dev->adj_list.upper,
7305 struct netdev_adjacent, list);
7306 if (likely(upper->master) && !upper->ignore)
7312 * netdev_has_any_lower_dev - Check if device is linked to some device
7315 * Find out if a device is linked to a lower device and return true in case
7316 * it is. The caller must hold the RTNL lock.
7318 static bool netdev_has_any_lower_dev(struct net_device *dev)
7322 return !list_empty(&dev->adj_list.lower);
7325 void *netdev_adjacent_get_private(struct list_head *adj_list)
7327 struct netdev_adjacent *adj;
7329 adj = list_entry(adj_list, struct netdev_adjacent, list);
7331 return adj->private;
7333 EXPORT_SYMBOL(netdev_adjacent_get_private);
7336 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
7338 * @iter: list_head ** of the current position
7340 * Gets the next device from the dev's upper list, starting from iter
7341 * position. The caller must hold RCU read lock.
7343 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
7344 struct list_head **iter)
7346 struct netdev_adjacent *upper;
7348 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7350 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7352 if (&upper->list == &dev->adj_list.upper)
7355 *iter = &upper->list;
7359 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
7361 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
7362 struct list_head **iter,
7365 struct netdev_adjacent *upper;
7367 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
7369 if (&upper->list == &dev->adj_list.upper)
7372 *iter = &upper->list;
7373 *ignore = upper->ignore;
7378 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
7379 struct list_head **iter)
7381 struct netdev_adjacent *upper;
7383 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7385 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7387 if (&upper->list == &dev->adj_list.upper)
7390 *iter = &upper->list;
7395 static int __netdev_walk_all_upper_dev(struct net_device *dev,
7396 int (*fn)(struct net_device *dev,
7397 struct netdev_nested_priv *priv),
7398 struct netdev_nested_priv *priv)
7400 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7401 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7406 iter = &dev->adj_list.upper;
7410 ret = fn(now, priv);
7417 udev = __netdev_next_upper_dev(now, &iter, &ignore);
7424 niter = &udev->adj_list.upper;
7425 dev_stack[cur] = now;
7426 iter_stack[cur++] = iter;
7433 next = dev_stack[--cur];
7434 niter = iter_stack[cur];
7444 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
7445 int (*fn)(struct net_device *dev,
7446 struct netdev_nested_priv *priv),
7447 struct netdev_nested_priv *priv)
7449 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7450 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7454 iter = &dev->adj_list.upper;
7458 ret = fn(now, priv);
7465 udev = netdev_next_upper_dev_rcu(now, &iter);
7470 niter = &udev->adj_list.upper;
7471 dev_stack[cur] = now;
7472 iter_stack[cur++] = iter;
7479 next = dev_stack[--cur];
7480 niter = iter_stack[cur];
7489 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7491 static bool __netdev_has_upper_dev(struct net_device *dev,
7492 struct net_device *upper_dev)
7494 struct netdev_nested_priv priv = {
7496 .data = (void *)upper_dev,
7501 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7506 * netdev_lower_get_next_private - Get the next ->private from the
7507 * lower neighbour list
7509 * @iter: list_head ** of the current position
7511 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7512 * list, starting from iter position. The caller must hold either hold the
7513 * RTNL lock or its own locking that guarantees that the neighbour lower
7514 * list will remain unchanged.
7516 void *netdev_lower_get_next_private(struct net_device *dev,
7517 struct list_head **iter)
7519 struct netdev_adjacent *lower;
7521 lower = list_entry(*iter, struct netdev_adjacent, list);
7523 if (&lower->list == &dev->adj_list.lower)
7526 *iter = lower->list.next;
7528 return lower->private;
7530 EXPORT_SYMBOL(netdev_lower_get_next_private);
7533 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7534 * lower neighbour list, RCU
7537 * @iter: list_head ** of the current position
7539 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7540 * list, starting from iter position. The caller must hold RCU read lock.
7542 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7543 struct list_head **iter)
7545 struct netdev_adjacent *lower;
7547 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
7549 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7551 if (&lower->list == &dev->adj_list.lower)
7554 *iter = &lower->list;
7556 return lower->private;
7558 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7561 * netdev_lower_get_next - Get the next device from the lower neighbour
7564 * @iter: list_head ** of the current position
7566 * Gets the next netdev_adjacent from the dev's lower neighbour
7567 * list, starting from iter position. The caller must hold RTNL lock or
7568 * its own locking that guarantees that the neighbour lower
7569 * list will remain unchanged.
7571 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7573 struct netdev_adjacent *lower;
7575 lower = list_entry(*iter, struct netdev_adjacent, list);
7577 if (&lower->list == &dev->adj_list.lower)
7580 *iter = lower->list.next;
7584 EXPORT_SYMBOL(netdev_lower_get_next);
7586 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7587 struct list_head **iter)
7589 struct netdev_adjacent *lower;
7591 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7593 if (&lower->list == &dev->adj_list.lower)
7596 *iter = &lower->list;
7601 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7602 struct list_head **iter,
7605 struct netdev_adjacent *lower;
7607 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7609 if (&lower->list == &dev->adj_list.lower)
7612 *iter = &lower->list;
7613 *ignore = lower->ignore;
7618 int netdev_walk_all_lower_dev(struct net_device *dev,
7619 int (*fn)(struct net_device *dev,
7620 struct netdev_nested_priv *priv),
7621 struct netdev_nested_priv *priv)
7623 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7624 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7628 iter = &dev->adj_list.lower;
7632 ret = fn(now, priv);
7639 ldev = netdev_next_lower_dev(now, &iter);
7644 niter = &ldev->adj_list.lower;
7645 dev_stack[cur] = now;
7646 iter_stack[cur++] = iter;
7653 next = dev_stack[--cur];
7654 niter = iter_stack[cur];
7663 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7665 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7666 int (*fn)(struct net_device *dev,
7667 struct netdev_nested_priv *priv),
7668 struct netdev_nested_priv *priv)
7670 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7671 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7676 iter = &dev->adj_list.lower;
7680 ret = fn(now, priv);
7687 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7694 niter = &ldev->adj_list.lower;
7695 dev_stack[cur] = now;
7696 iter_stack[cur++] = iter;
7703 next = dev_stack[--cur];
7704 niter = iter_stack[cur];
7714 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7715 struct list_head **iter)
7717 struct netdev_adjacent *lower;
7719 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7720 if (&lower->list == &dev->adj_list.lower)
7723 *iter = &lower->list;
7727 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7729 static u8 __netdev_upper_depth(struct net_device *dev)
7731 struct net_device *udev;
7732 struct list_head *iter;
7736 for (iter = &dev->adj_list.upper,
7737 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7739 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7742 if (max_depth < udev->upper_level)
7743 max_depth = udev->upper_level;
7749 static u8 __netdev_lower_depth(struct net_device *dev)
7751 struct net_device *ldev;
7752 struct list_head *iter;
7756 for (iter = &dev->adj_list.lower,
7757 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7759 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7762 if (max_depth < ldev->lower_level)
7763 max_depth = ldev->lower_level;
7769 static int __netdev_update_upper_level(struct net_device *dev,
7770 struct netdev_nested_priv *__unused)
7772 dev->upper_level = __netdev_upper_depth(dev) + 1;
7776 static int __netdev_update_lower_level(struct net_device *dev,
7777 struct netdev_nested_priv *priv)
7779 dev->lower_level = __netdev_lower_depth(dev) + 1;
7781 #ifdef CONFIG_LOCKDEP
7785 if (priv->flags & NESTED_SYNC_IMM)
7786 dev->nested_level = dev->lower_level - 1;
7787 if (priv->flags & NESTED_SYNC_TODO)
7788 net_unlink_todo(dev);
7793 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7794 int (*fn)(struct net_device *dev,
7795 struct netdev_nested_priv *priv),
7796 struct netdev_nested_priv *priv)
7798 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7799 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7803 iter = &dev->adj_list.lower;
7807 ret = fn(now, priv);
7814 ldev = netdev_next_lower_dev_rcu(now, &iter);
7819 niter = &ldev->adj_list.lower;
7820 dev_stack[cur] = now;
7821 iter_stack[cur++] = iter;
7828 next = dev_stack[--cur];
7829 niter = iter_stack[cur];
7838 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7841 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7842 * lower neighbour list, RCU
7846 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7847 * list. The caller must hold RCU read lock.
7849 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7851 struct netdev_adjacent *lower;
7853 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7854 struct netdev_adjacent, list);
7856 return lower->private;
7859 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7862 * netdev_master_upper_dev_get_rcu - Get master upper device
7865 * Find a master upper device and return pointer to it or NULL in case
7866 * it's not there. The caller must hold the RCU read lock.
7868 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7870 struct netdev_adjacent *upper;
7872 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7873 struct netdev_adjacent, list);
7874 if (upper && likely(upper->master))
7878 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7880 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7881 struct net_device *adj_dev,
7882 struct list_head *dev_list)
7884 char linkname[IFNAMSIZ+7];
7886 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7887 "upper_%s" : "lower_%s", adj_dev->name);
7888 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7891 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7893 struct list_head *dev_list)
7895 char linkname[IFNAMSIZ+7];
7897 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7898 "upper_%s" : "lower_%s", name);
7899 sysfs_remove_link(&(dev->dev.kobj), linkname);
7902 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7903 struct net_device *adj_dev,
7904 struct list_head *dev_list)
7906 return (dev_list == &dev->adj_list.upper ||
7907 dev_list == &dev->adj_list.lower) &&
7908 net_eq(dev_net(dev), dev_net(adj_dev));
7911 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7912 struct net_device *adj_dev,
7913 struct list_head *dev_list,
7914 void *private, bool master)
7916 struct netdev_adjacent *adj;
7919 adj = __netdev_find_adj(adj_dev, dev_list);
7923 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7924 dev->name, adj_dev->name, adj->ref_nr);
7929 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7934 adj->master = master;
7936 adj->private = private;
7937 adj->ignore = false;
7940 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7941 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7943 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7944 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7949 /* Ensure that master link is always the first item in list. */
7951 ret = sysfs_create_link(&(dev->dev.kobj),
7952 &(adj_dev->dev.kobj), "master");
7954 goto remove_symlinks;
7956 list_add_rcu(&adj->list, dev_list);
7958 list_add_tail_rcu(&adj->list, dev_list);
7964 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7965 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7973 static void __netdev_adjacent_dev_remove(struct net_device *dev,
7974 struct net_device *adj_dev,
7976 struct list_head *dev_list)
7978 struct netdev_adjacent *adj;
7980 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
7981 dev->name, adj_dev->name, ref_nr);
7983 adj = __netdev_find_adj(adj_dev, dev_list);
7986 pr_err("Adjacency does not exist for device %s from %s\n",
7987 dev->name, adj_dev->name);
7992 if (adj->ref_nr > ref_nr) {
7993 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
7994 dev->name, adj_dev->name, ref_nr,
7995 adj->ref_nr - ref_nr);
7996 adj->ref_nr -= ref_nr;
8001 sysfs_remove_link(&(dev->dev.kobj), "master");
8003 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
8004 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
8006 list_del_rcu(&adj->list);
8007 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
8008 adj_dev->name, dev->name, adj_dev->name);
8010 kfree_rcu(adj, rcu);
8013 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
8014 struct net_device *upper_dev,
8015 struct list_head *up_list,
8016 struct list_head *down_list,
8017 void *private, bool master)
8021 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
8026 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
8029 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
8036 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
8037 struct net_device *upper_dev,
8039 struct list_head *up_list,
8040 struct list_head *down_list)
8042 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
8043 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
8046 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
8047 struct net_device *upper_dev,
8048 void *private, bool master)
8050 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
8051 &dev->adj_list.upper,
8052 &upper_dev->adj_list.lower,
8056 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
8057 struct net_device *upper_dev)
8059 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
8060 &dev->adj_list.upper,
8061 &upper_dev->adj_list.lower);
8064 static int __netdev_upper_dev_link(struct net_device *dev,
8065 struct net_device *upper_dev, bool master,
8066 void *upper_priv, void *upper_info,
8067 struct netdev_nested_priv *priv,
8068 struct netlink_ext_ack *extack)
8070 struct netdev_notifier_changeupper_info changeupper_info = {
8075 .upper_dev = upper_dev,
8078 .upper_info = upper_info,
8080 struct net_device *master_dev;
8085 if (dev == upper_dev)
8088 /* To prevent loops, check if dev is not upper device to upper_dev. */
8089 if (__netdev_has_upper_dev(upper_dev, dev))
8092 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
8096 if (__netdev_has_upper_dev(dev, upper_dev))
8099 master_dev = __netdev_master_upper_dev_get(dev);
8101 return master_dev == upper_dev ? -EEXIST : -EBUSY;
8104 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8105 &changeupper_info.info);
8106 ret = notifier_to_errno(ret);
8110 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
8115 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8116 &changeupper_info.info);
8117 ret = notifier_to_errno(ret);
8121 __netdev_update_upper_level(dev, NULL);
8122 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8124 __netdev_update_lower_level(upper_dev, priv);
8125 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8131 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8137 * netdev_upper_dev_link - Add a link to the upper device
8139 * @upper_dev: new upper device
8140 * @extack: netlink extended ack
8142 * Adds a link to device which is upper to this one. The caller must hold
8143 * the RTNL lock. On a failure a negative errno code is returned.
8144 * On success the reference counts are adjusted and the function
8147 int netdev_upper_dev_link(struct net_device *dev,
8148 struct net_device *upper_dev,
8149 struct netlink_ext_ack *extack)
8151 struct netdev_nested_priv priv = {
8152 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8156 return __netdev_upper_dev_link(dev, upper_dev, false,
8157 NULL, NULL, &priv, extack);
8159 EXPORT_SYMBOL(netdev_upper_dev_link);
8162 * netdev_master_upper_dev_link - Add a master link to the upper device
8164 * @upper_dev: new upper device
8165 * @upper_priv: upper device private
8166 * @upper_info: upper info to be passed down via notifier
8167 * @extack: netlink extended ack
8169 * Adds a link to device which is upper to this one. In this case, only
8170 * one master upper device can be linked, although other non-master devices
8171 * might be linked as well. The caller must hold the RTNL lock.
8172 * On a failure a negative errno code is returned. On success the reference
8173 * counts are adjusted and the function returns zero.
8175 int netdev_master_upper_dev_link(struct net_device *dev,
8176 struct net_device *upper_dev,
8177 void *upper_priv, void *upper_info,
8178 struct netlink_ext_ack *extack)
8180 struct netdev_nested_priv priv = {
8181 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8185 return __netdev_upper_dev_link(dev, upper_dev, true,
8186 upper_priv, upper_info, &priv, extack);
8188 EXPORT_SYMBOL(netdev_master_upper_dev_link);
8190 static void __netdev_upper_dev_unlink(struct net_device *dev,
8191 struct net_device *upper_dev,
8192 struct netdev_nested_priv *priv)
8194 struct netdev_notifier_changeupper_info changeupper_info = {
8198 .upper_dev = upper_dev,
8204 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
8206 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8207 &changeupper_info.info);
8209 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8211 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8212 &changeupper_info.info);
8214 __netdev_update_upper_level(dev, NULL);
8215 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8217 __netdev_update_lower_level(upper_dev, priv);
8218 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8223 * netdev_upper_dev_unlink - Removes a link to upper device
8225 * @upper_dev: new upper device
8227 * Removes a link to device which is upper to this one. The caller must hold
8230 void netdev_upper_dev_unlink(struct net_device *dev,
8231 struct net_device *upper_dev)
8233 struct netdev_nested_priv priv = {
8234 .flags = NESTED_SYNC_TODO,
8238 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
8240 EXPORT_SYMBOL(netdev_upper_dev_unlink);
8242 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
8243 struct net_device *lower_dev,
8246 struct netdev_adjacent *adj;
8248 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
8252 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
8257 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
8258 struct net_device *lower_dev)
8260 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
8263 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
8264 struct net_device *lower_dev)
8266 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
8269 int netdev_adjacent_change_prepare(struct net_device *old_dev,
8270 struct net_device *new_dev,
8271 struct net_device *dev,
8272 struct netlink_ext_ack *extack)
8274 struct netdev_nested_priv priv = {
8283 if (old_dev && new_dev != old_dev)
8284 netdev_adjacent_dev_disable(dev, old_dev);
8285 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
8288 if (old_dev && new_dev != old_dev)
8289 netdev_adjacent_dev_enable(dev, old_dev);
8295 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
8297 void netdev_adjacent_change_commit(struct net_device *old_dev,
8298 struct net_device *new_dev,
8299 struct net_device *dev)
8301 struct netdev_nested_priv priv = {
8302 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8306 if (!new_dev || !old_dev)
8309 if (new_dev == old_dev)
8312 netdev_adjacent_dev_enable(dev, old_dev);
8313 __netdev_upper_dev_unlink(old_dev, dev, &priv);
8315 EXPORT_SYMBOL(netdev_adjacent_change_commit);
8317 void netdev_adjacent_change_abort(struct net_device *old_dev,
8318 struct net_device *new_dev,
8319 struct net_device *dev)
8321 struct netdev_nested_priv priv = {
8329 if (old_dev && new_dev != old_dev)
8330 netdev_adjacent_dev_enable(dev, old_dev);
8332 __netdev_upper_dev_unlink(new_dev, dev, &priv);
8334 EXPORT_SYMBOL(netdev_adjacent_change_abort);
8337 * netdev_bonding_info_change - Dispatch event about slave change
8339 * @bonding_info: info to dispatch
8341 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
8342 * The caller must hold the RTNL lock.
8344 void netdev_bonding_info_change(struct net_device *dev,
8345 struct netdev_bonding_info *bonding_info)
8347 struct netdev_notifier_bonding_info info = {
8351 memcpy(&info.bonding_info, bonding_info,
8352 sizeof(struct netdev_bonding_info));
8353 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
8356 EXPORT_SYMBOL(netdev_bonding_info_change);
8359 * netdev_get_xmit_slave - Get the xmit slave of master device
8362 * @all_slaves: assume all the slaves are active
8364 * The reference counters are not incremented so the caller must be
8365 * careful with locks. The caller must hold RCU lock.
8366 * %NULL is returned if no slave is found.
8369 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8370 struct sk_buff *skb,
8373 const struct net_device_ops *ops = dev->netdev_ops;
8375 if (!ops->ndo_get_xmit_slave)
8377 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8379 EXPORT_SYMBOL(netdev_get_xmit_slave);
8381 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8384 const struct net_device_ops *ops = dev->netdev_ops;
8386 if (!ops->ndo_sk_get_lower_dev)
8388 return ops->ndo_sk_get_lower_dev(dev, sk);
8392 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8396 * %NULL is returned if no lower device is found.
8399 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8402 struct net_device *lower;
8404 lower = netdev_sk_get_lower_dev(dev, sk);
8407 lower = netdev_sk_get_lower_dev(dev, sk);
8412 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8414 static void netdev_adjacent_add_links(struct net_device *dev)
8416 struct netdev_adjacent *iter;
8418 struct net *net = dev_net(dev);
8420 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8421 if (!net_eq(net, dev_net(iter->dev)))
8423 netdev_adjacent_sysfs_add(iter->dev, dev,
8424 &iter->dev->adj_list.lower);
8425 netdev_adjacent_sysfs_add(dev, iter->dev,
8426 &dev->adj_list.upper);
8429 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8430 if (!net_eq(net, dev_net(iter->dev)))
8432 netdev_adjacent_sysfs_add(iter->dev, dev,
8433 &iter->dev->adj_list.upper);
8434 netdev_adjacent_sysfs_add(dev, iter->dev,
8435 &dev->adj_list.lower);
8439 static void netdev_adjacent_del_links(struct net_device *dev)
8441 struct netdev_adjacent *iter;
8443 struct net *net = dev_net(dev);
8445 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8446 if (!net_eq(net, dev_net(iter->dev)))
8448 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8449 &iter->dev->adj_list.lower);
8450 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8451 &dev->adj_list.upper);
8454 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8455 if (!net_eq(net, dev_net(iter->dev)))
8457 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8458 &iter->dev->adj_list.upper);
8459 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8460 &dev->adj_list.lower);
8464 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8466 struct netdev_adjacent *iter;
8468 struct net *net = dev_net(dev);
8470 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8471 if (!net_eq(net, dev_net(iter->dev)))
8473 netdev_adjacent_sysfs_del(iter->dev, oldname,
8474 &iter->dev->adj_list.lower);
8475 netdev_adjacent_sysfs_add(iter->dev, dev,
8476 &iter->dev->adj_list.lower);
8479 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8480 if (!net_eq(net, dev_net(iter->dev)))
8482 netdev_adjacent_sysfs_del(iter->dev, oldname,
8483 &iter->dev->adj_list.upper);
8484 netdev_adjacent_sysfs_add(iter->dev, dev,
8485 &iter->dev->adj_list.upper);
8489 void *netdev_lower_dev_get_private(struct net_device *dev,
8490 struct net_device *lower_dev)
8492 struct netdev_adjacent *lower;
8496 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8500 return lower->private;
8502 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8506 * netdev_lower_state_changed - Dispatch event about lower device state change
8507 * @lower_dev: device
8508 * @lower_state_info: state to dispatch
8510 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8511 * The caller must hold the RTNL lock.
8513 void netdev_lower_state_changed(struct net_device *lower_dev,
8514 void *lower_state_info)
8516 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8517 .info.dev = lower_dev,
8521 changelowerstate_info.lower_state_info = lower_state_info;
8522 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8523 &changelowerstate_info.info);
8525 EXPORT_SYMBOL(netdev_lower_state_changed);
8527 static void dev_change_rx_flags(struct net_device *dev, int flags)
8529 const struct net_device_ops *ops = dev->netdev_ops;
8531 if (ops->ndo_change_rx_flags)
8532 ops->ndo_change_rx_flags(dev, flags);
8535 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8537 unsigned int old_flags = dev->flags;
8543 dev->flags |= IFF_PROMISC;
8544 dev->promiscuity += inc;
8545 if (dev->promiscuity == 0) {
8548 * If inc causes overflow, untouch promisc and return error.
8551 dev->flags &= ~IFF_PROMISC;
8553 dev->promiscuity -= inc;
8554 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
8559 if (dev->flags != old_flags) {
8560 pr_info("device %s %s promiscuous mode\n",
8562 dev->flags & IFF_PROMISC ? "entered" : "left");
8563 if (audit_enabled) {
8564 current_uid_gid(&uid, &gid);
8565 audit_log(audit_context(), GFP_ATOMIC,
8566 AUDIT_ANOM_PROMISCUOUS,
8567 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8568 dev->name, (dev->flags & IFF_PROMISC),
8569 (old_flags & IFF_PROMISC),
8570 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8571 from_kuid(&init_user_ns, uid),
8572 from_kgid(&init_user_ns, gid),
8573 audit_get_sessionid(current));
8576 dev_change_rx_flags(dev, IFF_PROMISC);
8579 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
8584 * dev_set_promiscuity - update promiscuity count on a device
8588 * Add or remove promiscuity from a device. While the count in the device
8589 * remains above zero the interface remains promiscuous. Once it hits zero
8590 * the device reverts back to normal filtering operation. A negative inc
8591 * value is used to drop promiscuity on the device.
8592 * Return 0 if successful or a negative errno code on error.
8594 int dev_set_promiscuity(struct net_device *dev, int inc)
8596 unsigned int old_flags = dev->flags;
8599 err = __dev_set_promiscuity(dev, inc, true);
8602 if (dev->flags != old_flags)
8603 dev_set_rx_mode(dev);
8606 EXPORT_SYMBOL(dev_set_promiscuity);
8608 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8610 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8614 dev->flags |= IFF_ALLMULTI;
8615 dev->allmulti += inc;
8616 if (dev->allmulti == 0) {
8619 * If inc causes overflow, untouch allmulti and return error.
8622 dev->flags &= ~IFF_ALLMULTI;
8624 dev->allmulti -= inc;
8625 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
8630 if (dev->flags ^ old_flags) {
8631 dev_change_rx_flags(dev, IFF_ALLMULTI);
8632 dev_set_rx_mode(dev);
8634 __dev_notify_flags(dev, old_flags,
8635 dev->gflags ^ old_gflags);
8641 * dev_set_allmulti - update allmulti count on a device
8645 * Add or remove reception of all multicast frames to a device. While the
8646 * count in the device remains above zero the interface remains listening
8647 * to all interfaces. Once it hits zero the device reverts back to normal
8648 * filtering operation. A negative @inc value is used to drop the counter
8649 * when releasing a resource needing all multicasts.
8650 * Return 0 if successful or a negative errno code on error.
8653 int dev_set_allmulti(struct net_device *dev, int inc)
8655 return __dev_set_allmulti(dev, inc, true);
8657 EXPORT_SYMBOL(dev_set_allmulti);
8660 * Upload unicast and multicast address lists to device and
8661 * configure RX filtering. When the device doesn't support unicast
8662 * filtering it is put in promiscuous mode while unicast addresses
8665 void __dev_set_rx_mode(struct net_device *dev)
8667 const struct net_device_ops *ops = dev->netdev_ops;
8669 /* dev_open will call this function so the list will stay sane. */
8670 if (!(dev->flags&IFF_UP))
8673 if (!netif_device_present(dev))
8676 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8677 /* Unicast addresses changes may only happen under the rtnl,
8678 * therefore calling __dev_set_promiscuity here is safe.
8680 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8681 __dev_set_promiscuity(dev, 1, false);
8682 dev->uc_promisc = true;
8683 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8684 __dev_set_promiscuity(dev, -1, false);
8685 dev->uc_promisc = false;
8689 if (ops->ndo_set_rx_mode)
8690 ops->ndo_set_rx_mode(dev);
8693 void dev_set_rx_mode(struct net_device *dev)
8695 netif_addr_lock_bh(dev);
8696 __dev_set_rx_mode(dev);
8697 netif_addr_unlock_bh(dev);
8701 * dev_get_flags - get flags reported to userspace
8704 * Get the combination of flag bits exported through APIs to userspace.
8706 unsigned int dev_get_flags(const struct net_device *dev)
8710 flags = (dev->flags & ~(IFF_PROMISC |
8715 (dev->gflags & (IFF_PROMISC |
8718 if (netif_running(dev)) {
8719 if (netif_oper_up(dev))
8720 flags |= IFF_RUNNING;
8721 if (netif_carrier_ok(dev))
8722 flags |= IFF_LOWER_UP;
8723 if (netif_dormant(dev))
8724 flags |= IFF_DORMANT;
8729 EXPORT_SYMBOL(dev_get_flags);
8731 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8732 struct netlink_ext_ack *extack)
8734 unsigned int old_flags = dev->flags;
8740 * Set the flags on our device.
8743 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8744 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8746 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8750 * Load in the correct multicast list now the flags have changed.
8753 if ((old_flags ^ flags) & IFF_MULTICAST)
8754 dev_change_rx_flags(dev, IFF_MULTICAST);
8756 dev_set_rx_mode(dev);
8759 * Have we downed the interface. We handle IFF_UP ourselves
8760 * according to user attempts to set it, rather than blindly
8765 if ((old_flags ^ flags) & IFF_UP) {
8766 if (old_flags & IFF_UP)
8769 ret = __dev_open(dev, extack);
8772 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8773 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8774 unsigned int old_flags = dev->flags;
8776 dev->gflags ^= IFF_PROMISC;
8778 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8779 if (dev->flags != old_flags)
8780 dev_set_rx_mode(dev);
8783 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8784 * is important. Some (broken) drivers set IFF_PROMISC, when
8785 * IFF_ALLMULTI is requested not asking us and not reporting.
8787 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8788 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8790 dev->gflags ^= IFF_ALLMULTI;
8791 __dev_set_allmulti(dev, inc, false);
8797 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8798 unsigned int gchanges)
8800 unsigned int changes = dev->flags ^ old_flags;
8803 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
8805 if (changes & IFF_UP) {
8806 if (dev->flags & IFF_UP)
8807 call_netdevice_notifiers(NETDEV_UP, dev);
8809 call_netdevice_notifiers(NETDEV_DOWN, dev);
8812 if (dev->flags & IFF_UP &&
8813 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8814 struct netdev_notifier_change_info change_info = {
8818 .flags_changed = changes,
8821 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8826 * dev_change_flags - change device settings
8828 * @flags: device state flags
8829 * @extack: netlink extended ack
8831 * Change settings on device based state flags. The flags are
8832 * in the userspace exported format.
8834 int dev_change_flags(struct net_device *dev, unsigned int flags,
8835 struct netlink_ext_ack *extack)
8838 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8840 ret = __dev_change_flags(dev, flags, extack);
8844 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8845 __dev_notify_flags(dev, old_flags, changes);
8848 EXPORT_SYMBOL(dev_change_flags);
8850 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8852 const struct net_device_ops *ops = dev->netdev_ops;
8854 if (ops->ndo_change_mtu)
8855 return ops->ndo_change_mtu(dev, new_mtu);
8857 /* Pairs with all the lockless reads of dev->mtu in the stack */
8858 WRITE_ONCE(dev->mtu, new_mtu);
8861 EXPORT_SYMBOL(__dev_set_mtu);
8863 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8864 struct netlink_ext_ack *extack)
8866 /* MTU must be positive, and in range */
8867 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8868 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8872 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8873 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8880 * dev_set_mtu_ext - Change maximum transfer unit
8882 * @new_mtu: new transfer unit
8883 * @extack: netlink extended ack
8885 * Change the maximum transfer size of the network device.
8887 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8888 struct netlink_ext_ack *extack)
8892 if (new_mtu == dev->mtu)
8895 err = dev_validate_mtu(dev, new_mtu, extack);
8899 if (!netif_device_present(dev))
8902 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8903 err = notifier_to_errno(err);
8907 orig_mtu = dev->mtu;
8908 err = __dev_set_mtu(dev, new_mtu);
8911 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8913 err = notifier_to_errno(err);
8915 /* setting mtu back and notifying everyone again,
8916 * so that they have a chance to revert changes.
8918 __dev_set_mtu(dev, orig_mtu);
8919 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8926 int dev_set_mtu(struct net_device *dev, int new_mtu)
8928 struct netlink_ext_ack extack;
8931 memset(&extack, 0, sizeof(extack));
8932 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8933 if (err && extack._msg)
8934 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8937 EXPORT_SYMBOL(dev_set_mtu);
8940 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8942 * @new_len: new tx queue length
8944 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8946 unsigned int orig_len = dev->tx_queue_len;
8949 if (new_len != (unsigned int)new_len)
8952 if (new_len != orig_len) {
8953 dev->tx_queue_len = new_len;
8954 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8955 res = notifier_to_errno(res);
8958 res = dev_qdisc_change_tx_queue_len(dev);
8966 netdev_err(dev, "refused to change device tx_queue_len\n");
8967 dev->tx_queue_len = orig_len;
8972 * dev_set_group - Change group this device belongs to
8974 * @new_group: group this device should belong to
8976 void dev_set_group(struct net_device *dev, int new_group)
8978 dev->group = new_group;
8980 EXPORT_SYMBOL(dev_set_group);
8983 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
8985 * @addr: new address
8986 * @extack: netlink extended ack
8988 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
8989 struct netlink_ext_ack *extack)
8991 struct netdev_notifier_pre_changeaddr_info info = {
8993 .info.extack = extack,
8998 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
8999 return notifier_to_errno(rc);
9001 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
9004 * dev_set_mac_address - Change Media Access Control Address
9007 * @extack: netlink extended ack
9009 * Change the hardware (MAC) address of the device
9011 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
9012 struct netlink_ext_ack *extack)
9014 const struct net_device_ops *ops = dev->netdev_ops;
9017 if (!ops->ndo_set_mac_address)
9019 if (sa->sa_family != dev->type)
9021 if (!netif_device_present(dev))
9023 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
9026 err = ops->ndo_set_mac_address(dev, sa);
9029 dev->addr_assign_type = NET_ADDR_SET;
9030 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
9031 add_device_randomness(dev->dev_addr, dev->addr_len);
9034 EXPORT_SYMBOL(dev_set_mac_address);
9036 static DECLARE_RWSEM(dev_addr_sem);
9038 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
9039 struct netlink_ext_ack *extack)
9043 down_write(&dev_addr_sem);
9044 ret = dev_set_mac_address(dev, sa, extack);
9045 up_write(&dev_addr_sem);
9048 EXPORT_SYMBOL(dev_set_mac_address_user);
9050 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
9052 size_t size = sizeof(sa->sa_data);
9053 struct net_device *dev;
9056 down_read(&dev_addr_sem);
9059 dev = dev_get_by_name_rcu(net, dev_name);
9065 memset(sa->sa_data, 0, size);
9067 memcpy(sa->sa_data, dev->dev_addr,
9068 min_t(size_t, size, dev->addr_len));
9069 sa->sa_family = dev->type;
9073 up_read(&dev_addr_sem);
9076 EXPORT_SYMBOL(dev_get_mac_address);
9079 * dev_change_carrier - Change device carrier
9081 * @new_carrier: new value
9083 * Change device carrier
9085 int dev_change_carrier(struct net_device *dev, bool new_carrier)
9087 const struct net_device_ops *ops = dev->netdev_ops;
9089 if (!ops->ndo_change_carrier)
9091 if (!netif_device_present(dev))
9093 return ops->ndo_change_carrier(dev, new_carrier);
9095 EXPORT_SYMBOL(dev_change_carrier);
9098 * dev_get_phys_port_id - Get device physical port ID
9102 * Get device physical port ID
9104 int dev_get_phys_port_id(struct net_device *dev,
9105 struct netdev_phys_item_id *ppid)
9107 const struct net_device_ops *ops = dev->netdev_ops;
9109 if (!ops->ndo_get_phys_port_id)
9111 return ops->ndo_get_phys_port_id(dev, ppid);
9113 EXPORT_SYMBOL(dev_get_phys_port_id);
9116 * dev_get_phys_port_name - Get device physical port name
9119 * @len: limit of bytes to copy to name
9121 * Get device physical port name
9123 int dev_get_phys_port_name(struct net_device *dev,
9124 char *name, size_t len)
9126 const struct net_device_ops *ops = dev->netdev_ops;
9129 if (ops->ndo_get_phys_port_name) {
9130 err = ops->ndo_get_phys_port_name(dev, name, len);
9131 if (err != -EOPNOTSUPP)
9134 return devlink_compat_phys_port_name_get(dev, name, len);
9136 EXPORT_SYMBOL(dev_get_phys_port_name);
9139 * dev_get_port_parent_id - Get the device's port parent identifier
9140 * @dev: network device
9141 * @ppid: pointer to a storage for the port's parent identifier
9142 * @recurse: allow/disallow recursion to lower devices
9144 * Get the devices's port parent identifier
9146 int dev_get_port_parent_id(struct net_device *dev,
9147 struct netdev_phys_item_id *ppid,
9150 const struct net_device_ops *ops = dev->netdev_ops;
9151 struct netdev_phys_item_id first = { };
9152 struct net_device *lower_dev;
9153 struct list_head *iter;
9156 if (ops->ndo_get_port_parent_id) {
9157 err = ops->ndo_get_port_parent_id(dev, ppid);
9158 if (err != -EOPNOTSUPP)
9162 err = devlink_compat_switch_id_get(dev, ppid);
9163 if (!err || err != -EOPNOTSUPP)
9169 netdev_for_each_lower_dev(dev, lower_dev, iter) {
9170 err = dev_get_port_parent_id(lower_dev, ppid, recurse);
9175 else if (memcmp(&first, ppid, sizeof(*ppid)))
9181 EXPORT_SYMBOL(dev_get_port_parent_id);
9184 * netdev_port_same_parent_id - Indicate if two network devices have
9185 * the same port parent identifier
9186 * @a: first network device
9187 * @b: second network device
9189 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
9191 struct netdev_phys_item_id a_id = { };
9192 struct netdev_phys_item_id b_id = { };
9194 if (dev_get_port_parent_id(a, &a_id, true) ||
9195 dev_get_port_parent_id(b, &b_id, true))
9198 return netdev_phys_item_id_same(&a_id, &b_id);
9200 EXPORT_SYMBOL(netdev_port_same_parent_id);
9203 * dev_change_proto_down - update protocol port state information
9205 * @proto_down: new value
9207 * This info can be used by switch drivers to set the phys state of the
9210 int dev_change_proto_down(struct net_device *dev, bool proto_down)
9212 const struct net_device_ops *ops = dev->netdev_ops;
9214 if (!ops->ndo_change_proto_down)
9216 if (!netif_device_present(dev))
9218 return ops->ndo_change_proto_down(dev, proto_down);
9220 EXPORT_SYMBOL(dev_change_proto_down);
9223 * dev_change_proto_down_generic - generic implementation for
9224 * ndo_change_proto_down that sets carrier according to
9228 * @proto_down: new value
9230 int dev_change_proto_down_generic(struct net_device *dev, bool proto_down)
9233 netif_carrier_off(dev);
9235 netif_carrier_on(dev);
9236 dev->proto_down = proto_down;
9239 EXPORT_SYMBOL(dev_change_proto_down_generic);
9242 * dev_change_proto_down_reason - proto down reason
9245 * @mask: proto down mask
9246 * @value: proto down value
9248 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
9254 dev->proto_down_reason = value;
9256 for_each_set_bit(b, &mask, 32) {
9257 if (value & (1 << b))
9258 dev->proto_down_reason |= BIT(b);
9260 dev->proto_down_reason &= ~BIT(b);
9264 EXPORT_SYMBOL(dev_change_proto_down_reason);
9266 struct bpf_xdp_link {
9267 struct bpf_link link;
9268 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9272 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9274 if (flags & XDP_FLAGS_HW_MODE)
9276 if (flags & XDP_FLAGS_DRV_MODE)
9277 return XDP_MODE_DRV;
9278 if (flags & XDP_FLAGS_SKB_MODE)
9279 return XDP_MODE_SKB;
9280 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9283 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9287 return generic_xdp_install;
9290 return dev->netdev_ops->ndo_bpf;
9296 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9297 enum bpf_xdp_mode mode)
9299 return dev->xdp_state[mode].link;
9302 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9303 enum bpf_xdp_mode mode)
9305 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9308 return link->link.prog;
9309 return dev->xdp_state[mode].prog;
9312 u8 dev_xdp_prog_count(struct net_device *dev)
9317 for (i = 0; i < __MAX_XDP_MODE; i++)
9318 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9322 EXPORT_SYMBOL_GPL(dev_xdp_prog_count);
9324 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9326 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9328 return prog ? prog->aux->id : 0;
9331 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9332 struct bpf_xdp_link *link)
9334 dev->xdp_state[mode].link = link;
9335 dev->xdp_state[mode].prog = NULL;
9338 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9339 struct bpf_prog *prog)
9341 dev->xdp_state[mode].link = NULL;
9342 dev->xdp_state[mode].prog = prog;
9345 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9346 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9347 u32 flags, struct bpf_prog *prog)
9349 struct netdev_bpf xdp;
9352 memset(&xdp, 0, sizeof(xdp));
9353 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9354 xdp.extack = extack;
9358 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9359 * "moved" into driver), so they don't increment it on their own, but
9360 * they do decrement refcnt when program is detached or replaced.
9361 * Given net_device also owns link/prog, we need to bump refcnt here
9362 * to prevent drivers from underflowing it.
9366 err = bpf_op(dev, &xdp);
9373 if (mode != XDP_MODE_HW)
9374 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9379 static void dev_xdp_uninstall(struct net_device *dev)
9381 struct bpf_xdp_link *link;
9382 struct bpf_prog *prog;
9383 enum bpf_xdp_mode mode;
9388 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9389 prog = dev_xdp_prog(dev, mode);
9393 bpf_op = dev_xdp_bpf_op(dev, mode);
9397 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9399 /* auto-detach link from net device */
9400 link = dev_xdp_link(dev, mode);
9406 dev_xdp_set_link(dev, mode, NULL);
9410 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9411 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9412 struct bpf_prog *old_prog, u32 flags)
9414 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9415 struct bpf_prog *cur_prog;
9416 struct net_device *upper;
9417 struct list_head *iter;
9418 enum bpf_xdp_mode mode;
9424 /* either link or prog attachment, never both */
9425 if (link && (new_prog || old_prog))
9427 /* link supports only XDP mode flags */
9428 if (link && (flags & ~XDP_FLAGS_MODES)) {
9429 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9432 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9433 if (num_modes > 1) {
9434 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9437 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9438 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9439 NL_SET_ERR_MSG(extack,
9440 "More than one program loaded, unset mode is ambiguous");
9443 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9444 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9445 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9449 mode = dev_xdp_mode(dev, flags);
9450 /* can't replace attached link */
9451 if (dev_xdp_link(dev, mode)) {
9452 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9456 /* don't allow if an upper device already has a program */
9457 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
9458 if (dev_xdp_prog_count(upper) > 0) {
9459 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program");
9464 cur_prog = dev_xdp_prog(dev, mode);
9465 /* can't replace attached prog with link */
9466 if (link && cur_prog) {
9467 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9470 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9471 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9475 /* put effective new program into new_prog */
9477 new_prog = link->link.prog;
9480 bool offload = mode == XDP_MODE_HW;
9481 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9482 ? XDP_MODE_DRV : XDP_MODE_SKB;
9484 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9485 NL_SET_ERR_MSG(extack, "XDP program already attached");
9488 if (!offload && dev_xdp_prog(dev, other_mode)) {
9489 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9492 if (!offload && bpf_prog_is_dev_bound(new_prog->aux)) {
9493 NL_SET_ERR_MSG(extack, "Using device-bound program without HW_MODE flag is not supported");
9496 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9497 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9500 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9501 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9506 /* don't call drivers if the effective program didn't change */
9507 if (new_prog != cur_prog) {
9508 bpf_op = dev_xdp_bpf_op(dev, mode);
9510 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9514 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9520 dev_xdp_set_link(dev, mode, link);
9522 dev_xdp_set_prog(dev, mode, new_prog);
9524 bpf_prog_put(cur_prog);
9529 static int dev_xdp_attach_link(struct net_device *dev,
9530 struct netlink_ext_ack *extack,
9531 struct bpf_xdp_link *link)
9533 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9536 static int dev_xdp_detach_link(struct net_device *dev,
9537 struct netlink_ext_ack *extack,
9538 struct bpf_xdp_link *link)
9540 enum bpf_xdp_mode mode;
9545 mode = dev_xdp_mode(dev, link->flags);
9546 if (dev_xdp_link(dev, mode) != link)
9549 bpf_op = dev_xdp_bpf_op(dev, mode);
9550 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9551 dev_xdp_set_link(dev, mode, NULL);
9555 static void bpf_xdp_link_release(struct bpf_link *link)
9557 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9561 /* if racing with net_device's tear down, xdp_link->dev might be
9562 * already NULL, in which case link was already auto-detached
9564 if (xdp_link->dev) {
9565 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9566 xdp_link->dev = NULL;
9572 static int bpf_xdp_link_detach(struct bpf_link *link)
9574 bpf_xdp_link_release(link);
9578 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9580 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9585 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9586 struct seq_file *seq)
9588 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9593 ifindex = xdp_link->dev->ifindex;
9596 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9599 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9600 struct bpf_link_info *info)
9602 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9607 ifindex = xdp_link->dev->ifindex;
9610 info->xdp.ifindex = ifindex;
9614 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9615 struct bpf_prog *old_prog)
9617 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9618 enum bpf_xdp_mode mode;
9624 /* link might have been auto-released already, so fail */
9625 if (!xdp_link->dev) {
9630 if (old_prog && link->prog != old_prog) {
9634 old_prog = link->prog;
9635 if (old_prog == new_prog) {
9636 /* no-op, don't disturb drivers */
9637 bpf_prog_put(new_prog);
9641 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9642 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9643 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9644 xdp_link->flags, new_prog);
9648 old_prog = xchg(&link->prog, new_prog);
9649 bpf_prog_put(old_prog);
9656 static const struct bpf_link_ops bpf_xdp_link_lops = {
9657 .release = bpf_xdp_link_release,
9658 .dealloc = bpf_xdp_link_dealloc,
9659 .detach = bpf_xdp_link_detach,
9660 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9661 .fill_link_info = bpf_xdp_link_fill_link_info,
9662 .update_prog = bpf_xdp_link_update,
9665 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9667 struct net *net = current->nsproxy->net_ns;
9668 struct bpf_link_primer link_primer;
9669 struct bpf_xdp_link *link;
9670 struct net_device *dev;
9674 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9680 link = kzalloc(sizeof(*link), GFP_USER);
9686 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9688 link->flags = attr->link_create.flags;
9690 err = bpf_link_prime(&link->link, &link_primer);
9696 err = dev_xdp_attach_link(dev, NULL, link);
9701 bpf_link_cleanup(&link_primer);
9705 fd = bpf_link_settle(&link_primer);
9706 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9719 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9721 * @extack: netlink extended ack
9722 * @fd: new program fd or negative value to clear
9723 * @expected_fd: old program fd that userspace expects to replace or clear
9724 * @flags: xdp-related flags
9726 * Set or clear a bpf program for a device
9728 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9729 int fd, int expected_fd, u32 flags)
9731 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9732 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9738 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9739 mode != XDP_MODE_SKB);
9740 if (IS_ERR(new_prog))
9741 return PTR_ERR(new_prog);
9744 if (expected_fd >= 0) {
9745 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9746 mode != XDP_MODE_SKB);
9747 if (IS_ERR(old_prog)) {
9748 err = PTR_ERR(old_prog);
9754 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9757 if (err && new_prog)
9758 bpf_prog_put(new_prog);
9760 bpf_prog_put(old_prog);
9765 * dev_new_index - allocate an ifindex
9766 * @net: the applicable net namespace
9768 * Returns a suitable unique value for a new device interface
9769 * number. The caller must hold the rtnl semaphore or the
9770 * dev_base_lock to be sure it remains unique.
9772 static int dev_new_index(struct net *net)
9774 int ifindex = net->ifindex;
9779 if (!__dev_get_by_index(net, ifindex))
9780 return net->ifindex = ifindex;
9784 /* Delayed registration/unregisteration */
9785 static LIST_HEAD(net_todo_list);
9786 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9788 static void net_set_todo(struct net_device *dev)
9790 list_add_tail(&dev->todo_list, &net_todo_list);
9791 dev_net(dev)->dev_unreg_count++;
9794 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9795 struct net_device *upper, netdev_features_t features)
9797 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9798 netdev_features_t feature;
9801 for_each_netdev_feature(upper_disables, feature_bit) {
9802 feature = __NETIF_F_BIT(feature_bit);
9803 if (!(upper->wanted_features & feature)
9804 && (features & feature)) {
9805 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9806 &feature, upper->name);
9807 features &= ~feature;
9814 static void netdev_sync_lower_features(struct net_device *upper,
9815 struct net_device *lower, netdev_features_t features)
9817 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9818 netdev_features_t feature;
9821 for_each_netdev_feature(upper_disables, feature_bit) {
9822 feature = __NETIF_F_BIT(feature_bit);
9823 if (!(features & feature) && (lower->features & feature)) {
9824 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9825 &feature, lower->name);
9826 lower->wanted_features &= ~feature;
9827 __netdev_update_features(lower);
9829 if (unlikely(lower->features & feature))
9830 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9831 &feature, lower->name);
9833 netdev_features_change(lower);
9838 static netdev_features_t netdev_fix_features(struct net_device *dev,
9839 netdev_features_t features)
9841 /* Fix illegal checksum combinations */
9842 if ((features & NETIF_F_HW_CSUM) &&
9843 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9844 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9845 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9848 /* TSO requires that SG is present as well. */
9849 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9850 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9851 features &= ~NETIF_F_ALL_TSO;
9854 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9855 !(features & NETIF_F_IP_CSUM)) {
9856 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9857 features &= ~NETIF_F_TSO;
9858 features &= ~NETIF_F_TSO_ECN;
9861 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9862 !(features & NETIF_F_IPV6_CSUM)) {
9863 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9864 features &= ~NETIF_F_TSO6;
9867 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9868 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9869 features &= ~NETIF_F_TSO_MANGLEID;
9871 /* TSO ECN requires that TSO is present as well. */
9872 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9873 features &= ~NETIF_F_TSO_ECN;
9875 /* Software GSO depends on SG. */
9876 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9877 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9878 features &= ~NETIF_F_GSO;
9881 /* GSO partial features require GSO partial be set */
9882 if ((features & dev->gso_partial_features) &&
9883 !(features & NETIF_F_GSO_PARTIAL)) {
9885 "Dropping partially supported GSO features since no GSO partial.\n");
9886 features &= ~dev->gso_partial_features;
9889 if (!(features & NETIF_F_RXCSUM)) {
9890 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9891 * successfully merged by hardware must also have the
9892 * checksum verified by hardware. If the user does not
9893 * want to enable RXCSUM, logically, we should disable GRO_HW.
9895 if (features & NETIF_F_GRO_HW) {
9896 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9897 features &= ~NETIF_F_GRO_HW;
9901 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9902 if (features & NETIF_F_RXFCS) {
9903 if (features & NETIF_F_LRO) {
9904 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9905 features &= ~NETIF_F_LRO;
9908 if (features & NETIF_F_GRO_HW) {
9909 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9910 features &= ~NETIF_F_GRO_HW;
9914 if (features & NETIF_F_HW_TLS_TX) {
9915 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9916 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9917 bool hw_csum = features & NETIF_F_HW_CSUM;
9919 if (!ip_csum && !hw_csum) {
9920 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9921 features &= ~NETIF_F_HW_TLS_TX;
9925 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9926 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9927 features &= ~NETIF_F_HW_TLS_RX;
9933 int __netdev_update_features(struct net_device *dev)
9935 struct net_device *upper, *lower;
9936 netdev_features_t features;
9937 struct list_head *iter;
9942 features = netdev_get_wanted_features(dev);
9944 if (dev->netdev_ops->ndo_fix_features)
9945 features = dev->netdev_ops->ndo_fix_features(dev, features);
9947 /* driver might be less strict about feature dependencies */
9948 features = netdev_fix_features(dev, features);
9950 /* some features can't be enabled if they're off on an upper device */
9951 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9952 features = netdev_sync_upper_features(dev, upper, features);
9954 if (dev->features == features)
9957 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9958 &dev->features, &features);
9960 if (dev->netdev_ops->ndo_set_features)
9961 err = dev->netdev_ops->ndo_set_features(dev, features);
9965 if (unlikely(err < 0)) {
9967 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9968 err, &features, &dev->features);
9969 /* return non-0 since some features might have changed and
9970 * it's better to fire a spurious notification than miss it
9976 /* some features must be disabled on lower devices when disabled
9977 * on an upper device (think: bonding master or bridge)
9979 netdev_for_each_lower_dev(dev, lower, iter)
9980 netdev_sync_lower_features(dev, lower, features);
9983 netdev_features_t diff = features ^ dev->features;
9985 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
9986 /* udp_tunnel_{get,drop}_rx_info both need
9987 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
9988 * device, or they won't do anything.
9989 * Thus we need to update dev->features
9990 * *before* calling udp_tunnel_get_rx_info,
9991 * but *after* calling udp_tunnel_drop_rx_info.
9993 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
9994 dev->features = features;
9995 udp_tunnel_get_rx_info(dev);
9997 udp_tunnel_drop_rx_info(dev);
10001 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
10002 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
10003 dev->features = features;
10004 err |= vlan_get_rx_ctag_filter_info(dev);
10006 vlan_drop_rx_ctag_filter_info(dev);
10010 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
10011 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
10012 dev->features = features;
10013 err |= vlan_get_rx_stag_filter_info(dev);
10015 vlan_drop_rx_stag_filter_info(dev);
10019 dev->features = features;
10022 return err < 0 ? 0 : 1;
10026 * netdev_update_features - recalculate device features
10027 * @dev: the device to check
10029 * Recalculate dev->features set and send notifications if it
10030 * has changed. Should be called after driver or hardware dependent
10031 * conditions might have changed that influence the features.
10033 void netdev_update_features(struct net_device *dev)
10035 if (__netdev_update_features(dev))
10036 netdev_features_change(dev);
10038 EXPORT_SYMBOL(netdev_update_features);
10041 * netdev_change_features - recalculate device features
10042 * @dev: the device to check
10044 * Recalculate dev->features set and send notifications even
10045 * if they have not changed. Should be called instead of
10046 * netdev_update_features() if also dev->vlan_features might
10047 * have changed to allow the changes to be propagated to stacked
10050 void netdev_change_features(struct net_device *dev)
10052 __netdev_update_features(dev);
10053 netdev_features_change(dev);
10055 EXPORT_SYMBOL(netdev_change_features);
10058 * netif_stacked_transfer_operstate - transfer operstate
10059 * @rootdev: the root or lower level device to transfer state from
10060 * @dev: the device to transfer operstate to
10062 * Transfer operational state from root to device. This is normally
10063 * called when a stacking relationship exists between the root
10064 * device and the device(a leaf device).
10066 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
10067 struct net_device *dev)
10069 if (rootdev->operstate == IF_OPER_DORMANT)
10070 netif_dormant_on(dev);
10072 netif_dormant_off(dev);
10074 if (rootdev->operstate == IF_OPER_TESTING)
10075 netif_testing_on(dev);
10077 netif_testing_off(dev);
10079 if (netif_carrier_ok(rootdev))
10080 netif_carrier_on(dev);
10082 netif_carrier_off(dev);
10084 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
10086 static int netif_alloc_rx_queues(struct net_device *dev)
10088 unsigned int i, count = dev->num_rx_queues;
10089 struct netdev_rx_queue *rx;
10090 size_t sz = count * sizeof(*rx);
10095 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10101 for (i = 0; i < count; i++) {
10104 /* XDP RX-queue setup */
10105 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
10112 /* Rollback successful reg's and free other resources */
10114 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
10120 static void netif_free_rx_queues(struct net_device *dev)
10122 unsigned int i, count = dev->num_rx_queues;
10124 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
10128 for (i = 0; i < count; i++)
10129 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
10134 static void netdev_init_one_queue(struct net_device *dev,
10135 struct netdev_queue *queue, void *_unused)
10137 /* Initialize queue lock */
10138 spin_lock_init(&queue->_xmit_lock);
10139 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
10140 queue->xmit_lock_owner = -1;
10141 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
10144 dql_init(&queue->dql, HZ);
10148 static void netif_free_tx_queues(struct net_device *dev)
10153 static int netif_alloc_netdev_queues(struct net_device *dev)
10155 unsigned int count = dev->num_tx_queues;
10156 struct netdev_queue *tx;
10157 size_t sz = count * sizeof(*tx);
10159 if (count < 1 || count > 0xffff)
10162 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10168 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
10169 spin_lock_init(&dev->tx_global_lock);
10174 void netif_tx_stop_all_queues(struct net_device *dev)
10178 for (i = 0; i < dev->num_tx_queues; i++) {
10179 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
10181 netif_tx_stop_queue(txq);
10184 EXPORT_SYMBOL(netif_tx_stop_all_queues);
10187 * register_netdevice - register a network device
10188 * @dev: device to register
10190 * Take a completed network device structure and add it to the kernel
10191 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10192 * chain. 0 is returned on success. A negative errno code is returned
10193 * on a failure to set up the device, or if the name is a duplicate.
10195 * Callers must hold the rtnl semaphore. You may want
10196 * register_netdev() instead of this.
10199 * The locking appears insufficient to guarantee two parallel registers
10200 * will not get the same name.
10203 int register_netdevice(struct net_device *dev)
10206 struct net *net = dev_net(dev);
10208 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
10209 NETDEV_FEATURE_COUNT);
10210 BUG_ON(dev_boot_phase);
10215 /* When net_device's are persistent, this will be fatal. */
10216 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
10219 ret = ethtool_check_ops(dev->ethtool_ops);
10223 spin_lock_init(&dev->addr_list_lock);
10224 netdev_set_addr_lockdep_class(dev);
10226 ret = dev_get_valid_name(net, dev, dev->name);
10231 dev->name_node = netdev_name_node_head_alloc(dev);
10232 if (!dev->name_node)
10235 /* Init, if this function is available */
10236 if (dev->netdev_ops->ndo_init) {
10237 ret = dev->netdev_ops->ndo_init(dev);
10241 goto err_free_name;
10245 if (((dev->hw_features | dev->features) &
10246 NETIF_F_HW_VLAN_CTAG_FILTER) &&
10247 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
10248 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
10249 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10256 dev->ifindex = dev_new_index(net);
10257 else if (__dev_get_by_index(net, dev->ifindex))
10260 /* Transfer changeable features to wanted_features and enable
10261 * software offloads (GSO and GRO).
10263 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10264 dev->features |= NETIF_F_SOFT_FEATURES;
10266 if (dev->udp_tunnel_nic_info) {
10267 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10268 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10271 dev->wanted_features = dev->features & dev->hw_features;
10273 if (!(dev->flags & IFF_LOOPBACK))
10274 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10276 /* If IPv4 TCP segmentation offload is supported we should also
10277 * allow the device to enable segmenting the frame with the option
10278 * of ignoring a static IP ID value. This doesn't enable the
10279 * feature itself but allows the user to enable it later.
10281 if (dev->hw_features & NETIF_F_TSO)
10282 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10283 if (dev->vlan_features & NETIF_F_TSO)
10284 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10285 if (dev->mpls_features & NETIF_F_TSO)
10286 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10287 if (dev->hw_enc_features & NETIF_F_TSO)
10288 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10290 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10292 dev->vlan_features |= NETIF_F_HIGHDMA;
10294 /* Make NETIF_F_SG inheritable to tunnel devices.
10296 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10298 /* Make NETIF_F_SG inheritable to MPLS.
10300 dev->mpls_features |= NETIF_F_SG;
10302 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10303 ret = notifier_to_errno(ret);
10307 ret = netdev_register_kobject(dev);
10309 dev->reg_state = NETREG_UNREGISTERED;
10312 dev->reg_state = NETREG_REGISTERED;
10314 __netdev_update_features(dev);
10317 * Default initial state at registry is that the
10318 * device is present.
10321 set_bit(__LINK_STATE_PRESENT, &dev->state);
10323 linkwatch_init_dev(dev);
10325 dev_init_scheduler(dev);
10327 list_netdevice(dev);
10328 add_device_randomness(dev->dev_addr, dev->addr_len);
10330 /* If the device has permanent device address, driver should
10331 * set dev_addr and also addr_assign_type should be set to
10332 * NET_ADDR_PERM (default value).
10334 if (dev->addr_assign_type == NET_ADDR_PERM)
10335 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10337 /* Notify protocols, that a new device appeared. */
10338 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10339 ret = notifier_to_errno(ret);
10341 /* Expect explicit free_netdev() on failure */
10342 dev->needs_free_netdev = false;
10343 unregister_netdevice_queue(dev, NULL);
10347 * Prevent userspace races by waiting until the network
10348 * device is fully setup before sending notifications.
10350 if (!dev->rtnl_link_ops ||
10351 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10352 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
10358 if (dev->netdev_ops->ndo_uninit)
10359 dev->netdev_ops->ndo_uninit(dev);
10360 if (dev->priv_destructor)
10361 dev->priv_destructor(dev);
10363 netdev_name_node_free(dev->name_node);
10366 EXPORT_SYMBOL(register_netdevice);
10369 * init_dummy_netdev - init a dummy network device for NAPI
10370 * @dev: device to init
10372 * This takes a network device structure and initialize the minimum
10373 * amount of fields so it can be used to schedule NAPI polls without
10374 * registering a full blown interface. This is to be used by drivers
10375 * that need to tie several hardware interfaces to a single NAPI
10376 * poll scheduler due to HW limitations.
10378 int init_dummy_netdev(struct net_device *dev)
10380 /* Clear everything. Note we don't initialize spinlocks
10381 * are they aren't supposed to be taken by any of the
10382 * NAPI code and this dummy netdev is supposed to be
10383 * only ever used for NAPI polls
10385 memset(dev, 0, sizeof(struct net_device));
10387 /* make sure we BUG if trying to hit standard
10388 * register/unregister code path
10390 dev->reg_state = NETREG_DUMMY;
10392 /* NAPI wants this */
10393 INIT_LIST_HEAD(&dev->napi_list);
10395 /* a dummy interface is started by default */
10396 set_bit(__LINK_STATE_PRESENT, &dev->state);
10397 set_bit(__LINK_STATE_START, &dev->state);
10399 /* napi_busy_loop stats accounting wants this */
10400 dev_net_set(dev, &init_net);
10402 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10403 * because users of this 'device' dont need to change
10409 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10413 * register_netdev - register a network device
10414 * @dev: device to register
10416 * Take a completed network device structure and add it to the kernel
10417 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10418 * chain. 0 is returned on success. A negative errno code is returned
10419 * on a failure to set up the device, or if the name is a duplicate.
10421 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10422 * and expands the device name if you passed a format string to
10425 int register_netdev(struct net_device *dev)
10429 if (rtnl_lock_killable())
10431 err = register_netdevice(dev);
10435 EXPORT_SYMBOL(register_netdev);
10437 int netdev_refcnt_read(const struct net_device *dev)
10439 #ifdef CONFIG_PCPU_DEV_REFCNT
10442 for_each_possible_cpu(i)
10443 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10446 return refcount_read(&dev->dev_refcnt);
10449 EXPORT_SYMBOL(netdev_refcnt_read);
10451 int netdev_unregister_timeout_secs __read_mostly = 10;
10453 #define WAIT_REFS_MIN_MSECS 1
10454 #define WAIT_REFS_MAX_MSECS 250
10456 * netdev_wait_allrefs - wait until all references are gone.
10457 * @dev: target net_device
10459 * This is called when unregistering network devices.
10461 * Any protocol or device that holds a reference should register
10462 * for netdevice notification, and cleanup and put back the
10463 * reference if they receive an UNREGISTER event.
10464 * We can get stuck here if buggy protocols don't correctly
10467 static void netdev_wait_allrefs(struct net_device *dev)
10469 unsigned long rebroadcast_time, warning_time;
10470 int wait = 0, refcnt;
10472 linkwatch_forget_dev(dev);
10474 rebroadcast_time = warning_time = jiffies;
10475 refcnt = netdev_refcnt_read(dev);
10477 while (refcnt != 1) {
10478 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10481 /* Rebroadcast unregister notification */
10482 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10488 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10490 /* We must not have linkwatch events
10491 * pending on unregister. If this
10492 * happens, we simply run the queue
10493 * unscheduled, resulting in a noop
10496 linkwatch_run_queue();
10501 rebroadcast_time = jiffies;
10506 wait = WAIT_REFS_MIN_MSECS;
10509 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10512 refcnt = netdev_refcnt_read(dev);
10515 time_after(jiffies, warning_time +
10516 netdev_unregister_timeout_secs * HZ)) {
10517 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10518 dev->name, refcnt);
10519 warning_time = jiffies;
10524 /* The sequence is:
10528 * register_netdevice(x1);
10529 * register_netdevice(x2);
10531 * unregister_netdevice(y1);
10532 * unregister_netdevice(y2);
10538 * We are invoked by rtnl_unlock().
10539 * This allows us to deal with problems:
10540 * 1) We can delete sysfs objects which invoke hotplug
10541 * without deadlocking with linkwatch via keventd.
10542 * 2) Since we run with the RTNL semaphore not held, we can sleep
10543 * safely in order to wait for the netdev refcnt to drop to zero.
10545 * We must not return until all unregister events added during
10546 * the interval the lock was held have been completed.
10548 void netdev_run_todo(void)
10550 struct list_head list;
10551 #ifdef CONFIG_LOCKDEP
10552 struct list_head unlink_list;
10554 list_replace_init(&net_unlink_list, &unlink_list);
10556 while (!list_empty(&unlink_list)) {
10557 struct net_device *dev = list_first_entry(&unlink_list,
10560 list_del_init(&dev->unlink_list);
10561 dev->nested_level = dev->lower_level - 1;
10565 /* Snapshot list, allow later requests */
10566 list_replace_init(&net_todo_list, &list);
10571 /* Wait for rcu callbacks to finish before next phase */
10572 if (!list_empty(&list))
10575 while (!list_empty(&list)) {
10576 struct net_device *dev
10577 = list_first_entry(&list, struct net_device, todo_list);
10578 list_del(&dev->todo_list);
10580 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10581 pr_err("network todo '%s' but state %d\n",
10582 dev->name, dev->reg_state);
10587 dev->reg_state = NETREG_UNREGISTERED;
10589 netdev_wait_allrefs(dev);
10592 BUG_ON(netdev_refcnt_read(dev) != 1);
10593 BUG_ON(!list_empty(&dev->ptype_all));
10594 BUG_ON(!list_empty(&dev->ptype_specific));
10595 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10596 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10597 #if IS_ENABLED(CONFIG_DECNET)
10598 WARN_ON(dev->dn_ptr);
10600 if (dev->priv_destructor)
10601 dev->priv_destructor(dev);
10602 if (dev->needs_free_netdev)
10605 /* Report a network device has been unregistered */
10607 dev_net(dev)->dev_unreg_count--;
10609 wake_up(&netdev_unregistering_wq);
10611 /* Free network device */
10612 kobject_put(&dev->dev.kobj);
10616 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10617 * all the same fields in the same order as net_device_stats, with only
10618 * the type differing, but rtnl_link_stats64 may have additional fields
10619 * at the end for newer counters.
10621 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10622 const struct net_device_stats *netdev_stats)
10624 #if BITS_PER_LONG == 64
10625 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
10626 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
10627 /* zero out counters that only exist in rtnl_link_stats64 */
10628 memset((char *)stats64 + sizeof(*netdev_stats), 0,
10629 sizeof(*stats64) - sizeof(*netdev_stats));
10631 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
10632 const unsigned long *src = (const unsigned long *)netdev_stats;
10633 u64 *dst = (u64 *)stats64;
10635 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10636 for (i = 0; i < n; i++)
10638 /* zero out counters that only exist in rtnl_link_stats64 */
10639 memset((char *)stats64 + n * sizeof(u64), 0,
10640 sizeof(*stats64) - n * sizeof(u64));
10643 EXPORT_SYMBOL(netdev_stats_to_stats64);
10646 * dev_get_stats - get network device statistics
10647 * @dev: device to get statistics from
10648 * @storage: place to store stats
10650 * Get network statistics from device. Return @storage.
10651 * The device driver may provide its own method by setting
10652 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10653 * otherwise the internal statistics structure is used.
10655 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10656 struct rtnl_link_stats64 *storage)
10658 const struct net_device_ops *ops = dev->netdev_ops;
10660 if (ops->ndo_get_stats64) {
10661 memset(storage, 0, sizeof(*storage));
10662 ops->ndo_get_stats64(dev, storage);
10663 } else if (ops->ndo_get_stats) {
10664 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10666 netdev_stats_to_stats64(storage, &dev->stats);
10668 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
10669 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
10670 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
10673 EXPORT_SYMBOL(dev_get_stats);
10676 * dev_fetch_sw_netstats - get per-cpu network device statistics
10677 * @s: place to store stats
10678 * @netstats: per-cpu network stats to read from
10680 * Read per-cpu network statistics and populate the related fields in @s.
10682 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10683 const struct pcpu_sw_netstats __percpu *netstats)
10687 for_each_possible_cpu(cpu) {
10688 const struct pcpu_sw_netstats *stats;
10689 struct pcpu_sw_netstats tmp;
10690 unsigned int start;
10692 stats = per_cpu_ptr(netstats, cpu);
10694 start = u64_stats_fetch_begin_irq(&stats->syncp);
10695 tmp.rx_packets = stats->rx_packets;
10696 tmp.rx_bytes = stats->rx_bytes;
10697 tmp.tx_packets = stats->tx_packets;
10698 tmp.tx_bytes = stats->tx_bytes;
10699 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
10701 s->rx_packets += tmp.rx_packets;
10702 s->rx_bytes += tmp.rx_bytes;
10703 s->tx_packets += tmp.tx_packets;
10704 s->tx_bytes += tmp.tx_bytes;
10707 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10710 * dev_get_tstats64 - ndo_get_stats64 implementation
10711 * @dev: device to get statistics from
10712 * @s: place to store stats
10714 * Populate @s from dev->stats and dev->tstats. Can be used as
10715 * ndo_get_stats64() callback.
10717 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10719 netdev_stats_to_stats64(s, &dev->stats);
10720 dev_fetch_sw_netstats(s, dev->tstats);
10722 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10724 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10726 struct netdev_queue *queue = dev_ingress_queue(dev);
10728 #ifdef CONFIG_NET_CLS_ACT
10731 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10734 netdev_init_one_queue(dev, queue, NULL);
10735 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10736 queue->qdisc_sleeping = &noop_qdisc;
10737 rcu_assign_pointer(dev->ingress_queue, queue);
10742 static const struct ethtool_ops default_ethtool_ops;
10744 void netdev_set_default_ethtool_ops(struct net_device *dev,
10745 const struct ethtool_ops *ops)
10747 if (dev->ethtool_ops == &default_ethtool_ops)
10748 dev->ethtool_ops = ops;
10750 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10752 void netdev_freemem(struct net_device *dev)
10754 char *addr = (char *)dev - dev->padded;
10760 * alloc_netdev_mqs - allocate network device
10761 * @sizeof_priv: size of private data to allocate space for
10762 * @name: device name format string
10763 * @name_assign_type: origin of device name
10764 * @setup: callback to initialize device
10765 * @txqs: the number of TX subqueues to allocate
10766 * @rxqs: the number of RX subqueues to allocate
10768 * Allocates a struct net_device with private data area for driver use
10769 * and performs basic initialization. Also allocates subqueue structs
10770 * for each queue on the device.
10772 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10773 unsigned char name_assign_type,
10774 void (*setup)(struct net_device *),
10775 unsigned int txqs, unsigned int rxqs)
10777 struct net_device *dev;
10778 unsigned int alloc_size;
10779 struct net_device *p;
10781 BUG_ON(strlen(name) >= sizeof(dev->name));
10784 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10789 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10793 alloc_size = sizeof(struct net_device);
10795 /* ensure 32-byte alignment of private area */
10796 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10797 alloc_size += sizeof_priv;
10799 /* ensure 32-byte alignment of whole construct */
10800 alloc_size += NETDEV_ALIGN - 1;
10802 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10806 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10807 dev->padded = (char *)dev - (char *)p;
10809 #ifdef CONFIG_PCPU_DEV_REFCNT
10810 dev->pcpu_refcnt = alloc_percpu(int);
10811 if (!dev->pcpu_refcnt)
10815 refcount_set(&dev->dev_refcnt, 1);
10818 if (dev_addr_init(dev))
10824 dev_net_set(dev, &init_net);
10826 dev->gso_max_size = GSO_MAX_SIZE;
10827 dev->gso_max_segs = GSO_MAX_SEGS;
10828 dev->upper_level = 1;
10829 dev->lower_level = 1;
10830 #ifdef CONFIG_LOCKDEP
10831 dev->nested_level = 0;
10832 INIT_LIST_HEAD(&dev->unlink_list);
10835 INIT_LIST_HEAD(&dev->napi_list);
10836 INIT_LIST_HEAD(&dev->unreg_list);
10837 INIT_LIST_HEAD(&dev->close_list);
10838 INIT_LIST_HEAD(&dev->link_watch_list);
10839 INIT_LIST_HEAD(&dev->adj_list.upper);
10840 INIT_LIST_HEAD(&dev->adj_list.lower);
10841 INIT_LIST_HEAD(&dev->ptype_all);
10842 INIT_LIST_HEAD(&dev->ptype_specific);
10843 INIT_LIST_HEAD(&dev->net_notifier_list);
10844 #ifdef CONFIG_NET_SCHED
10845 hash_init(dev->qdisc_hash);
10847 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10850 if (!dev->tx_queue_len) {
10851 dev->priv_flags |= IFF_NO_QUEUE;
10852 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10855 dev->num_tx_queues = txqs;
10856 dev->real_num_tx_queues = txqs;
10857 if (netif_alloc_netdev_queues(dev))
10860 dev->num_rx_queues = rxqs;
10861 dev->real_num_rx_queues = rxqs;
10862 if (netif_alloc_rx_queues(dev))
10865 strcpy(dev->name, name);
10866 dev->name_assign_type = name_assign_type;
10867 dev->group = INIT_NETDEV_GROUP;
10868 if (!dev->ethtool_ops)
10869 dev->ethtool_ops = &default_ethtool_ops;
10871 nf_hook_ingress_init(dev);
10880 #ifdef CONFIG_PCPU_DEV_REFCNT
10881 free_percpu(dev->pcpu_refcnt);
10884 netdev_freemem(dev);
10887 EXPORT_SYMBOL(alloc_netdev_mqs);
10890 * free_netdev - free network device
10893 * This function does the last stage of destroying an allocated device
10894 * interface. The reference to the device object is released. If this
10895 * is the last reference then it will be freed.Must be called in process
10898 void free_netdev(struct net_device *dev)
10900 struct napi_struct *p, *n;
10904 /* When called immediately after register_netdevice() failed the unwind
10905 * handling may still be dismantling the device. Handle that case by
10906 * deferring the free.
10908 if (dev->reg_state == NETREG_UNREGISTERING) {
10910 dev->needs_free_netdev = true;
10914 netif_free_tx_queues(dev);
10915 netif_free_rx_queues(dev);
10917 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10919 /* Flush device addresses */
10920 dev_addr_flush(dev);
10922 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10925 #ifdef CONFIG_PCPU_DEV_REFCNT
10926 free_percpu(dev->pcpu_refcnt);
10927 dev->pcpu_refcnt = NULL;
10929 free_percpu(dev->xdp_bulkq);
10930 dev->xdp_bulkq = NULL;
10932 /* Compatibility with error handling in drivers */
10933 if (dev->reg_state == NETREG_UNINITIALIZED) {
10934 netdev_freemem(dev);
10938 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10939 dev->reg_state = NETREG_RELEASED;
10941 /* will free via device release */
10942 put_device(&dev->dev);
10944 EXPORT_SYMBOL(free_netdev);
10947 * synchronize_net - Synchronize with packet receive processing
10949 * Wait for packets currently being received to be done.
10950 * Does not block later packets from starting.
10952 void synchronize_net(void)
10955 if (rtnl_is_locked())
10956 synchronize_rcu_expedited();
10960 EXPORT_SYMBOL(synchronize_net);
10963 * unregister_netdevice_queue - remove device from the kernel
10967 * This function shuts down a device interface and removes it
10968 * from the kernel tables.
10969 * If head not NULL, device is queued to be unregistered later.
10971 * Callers must hold the rtnl semaphore. You may want
10972 * unregister_netdev() instead of this.
10975 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
10980 list_move_tail(&dev->unreg_list, head);
10984 list_add(&dev->unreg_list, &single);
10985 unregister_netdevice_many(&single);
10988 EXPORT_SYMBOL(unregister_netdevice_queue);
10991 * unregister_netdevice_many - unregister many devices
10992 * @head: list of devices
10994 * Note: As most callers use a stack allocated list_head,
10995 * we force a list_del() to make sure stack wont be corrupted later.
10997 void unregister_netdevice_many(struct list_head *head)
10999 struct net_device *dev, *tmp;
11000 LIST_HEAD(close_head);
11002 BUG_ON(dev_boot_phase);
11005 if (list_empty(head))
11008 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
11009 /* Some devices call without registering
11010 * for initialization unwind. Remove those
11011 * devices and proceed with the remaining.
11013 if (dev->reg_state == NETREG_UNINITIALIZED) {
11014 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
11018 list_del(&dev->unreg_list);
11021 dev->dismantle = true;
11022 BUG_ON(dev->reg_state != NETREG_REGISTERED);
11025 /* If device is running, close it first. */
11026 list_for_each_entry(dev, head, unreg_list)
11027 list_add_tail(&dev->close_list, &close_head);
11028 dev_close_many(&close_head, true);
11030 list_for_each_entry(dev, head, unreg_list) {
11031 /* And unlink it from device chain. */
11032 unlist_netdevice(dev);
11034 dev->reg_state = NETREG_UNREGISTERING;
11036 flush_all_backlogs();
11040 list_for_each_entry(dev, head, unreg_list) {
11041 struct sk_buff *skb = NULL;
11043 /* Shutdown queueing discipline. */
11046 dev_xdp_uninstall(dev);
11048 /* Notify protocols, that we are about to destroy
11049 * this device. They should clean all the things.
11051 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11053 if (!dev->rtnl_link_ops ||
11054 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
11055 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
11056 GFP_KERNEL, NULL, 0);
11059 * Flush the unicast and multicast chains
11064 netdev_name_node_alt_flush(dev);
11065 netdev_name_node_free(dev->name_node);
11067 if (dev->netdev_ops->ndo_uninit)
11068 dev->netdev_ops->ndo_uninit(dev);
11071 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
11073 /* Notifier chain MUST detach us all upper devices. */
11074 WARN_ON(netdev_has_any_upper_dev(dev));
11075 WARN_ON(netdev_has_any_lower_dev(dev));
11077 /* Remove entries from kobject tree */
11078 netdev_unregister_kobject(dev);
11080 /* Remove XPS queueing entries */
11081 netif_reset_xps_queues_gt(dev, 0);
11087 list_for_each_entry(dev, head, unreg_list) {
11094 EXPORT_SYMBOL(unregister_netdevice_many);
11097 * unregister_netdev - remove device from the kernel
11100 * This function shuts down a device interface and removes it
11101 * from the kernel tables.
11103 * This is just a wrapper for unregister_netdevice that takes
11104 * the rtnl semaphore. In general you want to use this and not
11105 * unregister_netdevice.
11107 void unregister_netdev(struct net_device *dev)
11110 unregister_netdevice(dev);
11113 EXPORT_SYMBOL(unregister_netdev);
11116 * __dev_change_net_namespace - move device to different nethost namespace
11118 * @net: network namespace
11119 * @pat: If not NULL name pattern to try if the current device name
11120 * is already taken in the destination network namespace.
11121 * @new_ifindex: If not zero, specifies device index in the target
11124 * This function shuts down a device interface and moves it
11125 * to a new network namespace. On success 0 is returned, on
11126 * a failure a netagive errno code is returned.
11128 * Callers must hold the rtnl semaphore.
11131 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
11132 const char *pat, int new_ifindex)
11134 struct net *net_old = dev_net(dev);
11139 /* Don't allow namespace local devices to be moved. */
11141 if (dev->features & NETIF_F_NETNS_LOCAL)
11144 /* Ensure the device has been registrered */
11145 if (dev->reg_state != NETREG_REGISTERED)
11148 /* Get out if there is nothing todo */
11150 if (net_eq(net_old, net))
11153 /* Pick the destination device name, and ensure
11154 * we can use it in the destination network namespace.
11157 if (__dev_get_by_name(net, dev->name)) {
11158 /* We get here if we can't use the current device name */
11161 err = dev_get_valid_name(net, dev, pat);
11166 /* Check that new_ifindex isn't used yet. */
11168 if (new_ifindex && __dev_get_by_index(net, new_ifindex))
11172 * And now a mini version of register_netdevice unregister_netdevice.
11175 /* If device is running close it first. */
11178 /* And unlink it from device chain */
11179 unlist_netdevice(dev);
11183 /* Shutdown queueing discipline. */
11186 /* Notify protocols, that we are about to destroy
11187 * this device. They should clean all the things.
11189 * Note that dev->reg_state stays at NETREG_REGISTERED.
11190 * This is wanted because this way 8021q and macvlan know
11191 * the device is just moving and can keep their slaves up.
11193 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11196 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11197 /* If there is an ifindex conflict assign a new one */
11198 if (!new_ifindex) {
11199 if (__dev_get_by_index(net, dev->ifindex))
11200 new_ifindex = dev_new_index(net);
11202 new_ifindex = dev->ifindex;
11205 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11209 * Flush the unicast and multicast chains
11214 /* Send a netdev-removed uevent to the old namespace */
11215 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11216 netdev_adjacent_del_links(dev);
11218 /* Move per-net netdevice notifiers that are following the netdevice */
11219 move_netdevice_notifiers_dev_net(dev, net);
11221 /* Actually switch the network namespace */
11222 dev_net_set(dev, net);
11223 dev->ifindex = new_ifindex;
11225 /* Send a netdev-add uevent to the new namespace */
11226 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11227 netdev_adjacent_add_links(dev);
11229 /* Fixup kobjects */
11230 err = device_rename(&dev->dev, dev->name);
11233 /* Adapt owner in case owning user namespace of target network
11234 * namespace is different from the original one.
11236 err = netdev_change_owner(dev, net_old, net);
11239 /* Add the device back in the hashes */
11240 list_netdevice(dev);
11242 /* Notify protocols, that a new device appeared. */
11243 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11246 * Prevent userspace races by waiting until the network
11247 * device is fully setup before sending notifications.
11249 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
11256 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11258 static int dev_cpu_dead(unsigned int oldcpu)
11260 struct sk_buff **list_skb;
11261 struct sk_buff *skb;
11263 struct softnet_data *sd, *oldsd, *remsd = NULL;
11265 local_irq_disable();
11266 cpu = smp_processor_id();
11267 sd = &per_cpu(softnet_data, cpu);
11268 oldsd = &per_cpu(softnet_data, oldcpu);
11270 /* Find end of our completion_queue. */
11271 list_skb = &sd->completion_queue;
11273 list_skb = &(*list_skb)->next;
11274 /* Append completion queue from offline CPU. */
11275 *list_skb = oldsd->completion_queue;
11276 oldsd->completion_queue = NULL;
11278 /* Append output queue from offline CPU. */
11279 if (oldsd->output_queue) {
11280 *sd->output_queue_tailp = oldsd->output_queue;
11281 sd->output_queue_tailp = oldsd->output_queue_tailp;
11282 oldsd->output_queue = NULL;
11283 oldsd->output_queue_tailp = &oldsd->output_queue;
11285 /* Append NAPI poll list from offline CPU, with one exception :
11286 * process_backlog() must be called by cpu owning percpu backlog.
11287 * We properly handle process_queue & input_pkt_queue later.
11289 while (!list_empty(&oldsd->poll_list)) {
11290 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11291 struct napi_struct,
11294 list_del_init(&napi->poll_list);
11295 if (napi->poll == process_backlog)
11298 ____napi_schedule(sd, napi);
11301 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11302 local_irq_enable();
11305 remsd = oldsd->rps_ipi_list;
11306 oldsd->rps_ipi_list = NULL;
11308 /* send out pending IPI's on offline CPU */
11309 net_rps_send_ipi(remsd);
11311 /* Process offline CPU's input_pkt_queue */
11312 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11314 input_queue_head_incr(oldsd);
11316 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11318 input_queue_head_incr(oldsd);
11325 * netdev_increment_features - increment feature set by one
11326 * @all: current feature set
11327 * @one: new feature set
11328 * @mask: mask feature set
11330 * Computes a new feature set after adding a device with feature set
11331 * @one to the master device with current feature set @all. Will not
11332 * enable anything that is off in @mask. Returns the new feature set.
11334 netdev_features_t netdev_increment_features(netdev_features_t all,
11335 netdev_features_t one, netdev_features_t mask)
11337 if (mask & NETIF_F_HW_CSUM)
11338 mask |= NETIF_F_CSUM_MASK;
11339 mask |= NETIF_F_VLAN_CHALLENGED;
11341 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11342 all &= one | ~NETIF_F_ALL_FOR_ALL;
11344 /* If one device supports hw checksumming, set for all. */
11345 if (all & NETIF_F_HW_CSUM)
11346 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11350 EXPORT_SYMBOL(netdev_increment_features);
11352 static struct hlist_head * __net_init netdev_create_hash(void)
11355 struct hlist_head *hash;
11357 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11359 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11360 INIT_HLIST_HEAD(&hash[i]);
11365 /* Initialize per network namespace state */
11366 static int __net_init netdev_init(struct net *net)
11368 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11369 8 * sizeof_field(struct napi_struct, gro_bitmask));
11371 if (net != &init_net)
11372 INIT_LIST_HEAD(&net->dev_base_head);
11374 net->dev_name_head = netdev_create_hash();
11375 if (net->dev_name_head == NULL)
11378 net->dev_index_head = netdev_create_hash();
11379 if (net->dev_index_head == NULL)
11382 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11387 kfree(net->dev_name_head);
11393 * netdev_drivername - network driver for the device
11394 * @dev: network device
11396 * Determine network driver for device.
11398 const char *netdev_drivername(const struct net_device *dev)
11400 const struct device_driver *driver;
11401 const struct device *parent;
11402 const char *empty = "";
11404 parent = dev->dev.parent;
11408 driver = parent->driver;
11409 if (driver && driver->name)
11410 return driver->name;
11414 static void __netdev_printk(const char *level, const struct net_device *dev,
11415 struct va_format *vaf)
11417 if (dev && dev->dev.parent) {
11418 dev_printk_emit(level[1] - '0',
11421 dev_driver_string(dev->dev.parent),
11422 dev_name(dev->dev.parent),
11423 netdev_name(dev), netdev_reg_state(dev),
11426 printk("%s%s%s: %pV",
11427 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11429 printk("%s(NULL net_device): %pV", level, vaf);
11433 void netdev_printk(const char *level, const struct net_device *dev,
11434 const char *format, ...)
11436 struct va_format vaf;
11439 va_start(args, format);
11444 __netdev_printk(level, dev, &vaf);
11448 EXPORT_SYMBOL(netdev_printk);
11450 #define define_netdev_printk_level(func, level) \
11451 void func(const struct net_device *dev, const char *fmt, ...) \
11453 struct va_format vaf; \
11456 va_start(args, fmt); \
11461 __netdev_printk(level, dev, &vaf); \
11465 EXPORT_SYMBOL(func);
11467 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11468 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11469 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11470 define_netdev_printk_level(netdev_err, KERN_ERR);
11471 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11472 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11473 define_netdev_printk_level(netdev_info, KERN_INFO);
11475 static void __net_exit netdev_exit(struct net *net)
11477 kfree(net->dev_name_head);
11478 kfree(net->dev_index_head);
11479 if (net != &init_net)
11480 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11483 static struct pernet_operations __net_initdata netdev_net_ops = {
11484 .init = netdev_init,
11485 .exit = netdev_exit,
11488 static void __net_exit default_device_exit(struct net *net)
11490 struct net_device *dev, *aux;
11492 * Push all migratable network devices back to the
11493 * initial network namespace
11496 for_each_netdev_safe(net, dev, aux) {
11498 char fb_name[IFNAMSIZ];
11500 /* Ignore unmoveable devices (i.e. loopback) */
11501 if (dev->features & NETIF_F_NETNS_LOCAL)
11504 /* Leave virtual devices for the generic cleanup */
11505 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11508 /* Push remaining network devices to init_net */
11509 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11510 if (__dev_get_by_name(&init_net, fb_name))
11511 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11512 err = dev_change_net_namespace(dev, &init_net, fb_name);
11514 pr_emerg("%s: failed to move %s to init_net: %d\n",
11515 __func__, dev->name, err);
11522 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
11524 /* Return with the rtnl_lock held when there are no network
11525 * devices unregistering in any network namespace in net_list.
11528 bool unregistering;
11529 DEFINE_WAIT_FUNC(wait, woken_wake_function);
11531 add_wait_queue(&netdev_unregistering_wq, &wait);
11533 unregistering = false;
11535 list_for_each_entry(net, net_list, exit_list) {
11536 if (net->dev_unreg_count > 0) {
11537 unregistering = true;
11541 if (!unregistering)
11545 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
11547 remove_wait_queue(&netdev_unregistering_wq, &wait);
11550 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11552 /* At exit all network devices most be removed from a network
11553 * namespace. Do this in the reverse order of registration.
11554 * Do this across as many network namespaces as possible to
11555 * improve batching efficiency.
11557 struct net_device *dev;
11559 LIST_HEAD(dev_kill_list);
11561 /* To prevent network device cleanup code from dereferencing
11562 * loopback devices or network devices that have been freed
11563 * wait here for all pending unregistrations to complete,
11564 * before unregistring the loopback device and allowing the
11565 * network namespace be freed.
11567 * The netdev todo list containing all network devices
11568 * unregistrations that happen in default_device_exit_batch
11569 * will run in the rtnl_unlock() at the end of
11570 * default_device_exit_batch.
11572 rtnl_lock_unregistering(net_list);
11573 list_for_each_entry(net, net_list, exit_list) {
11574 for_each_netdev_reverse(net, dev) {
11575 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11576 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11578 unregister_netdevice_queue(dev, &dev_kill_list);
11581 unregister_netdevice_many(&dev_kill_list);
11585 static struct pernet_operations __net_initdata default_device_ops = {
11586 .exit = default_device_exit,
11587 .exit_batch = default_device_exit_batch,
11591 * Initialize the DEV module. At boot time this walks the device list and
11592 * unhooks any devices that fail to initialise (normally hardware not
11593 * present) and leaves us with a valid list of present and active devices.
11598 * This is called single threaded during boot, so no need
11599 * to take the rtnl semaphore.
11601 static int __init net_dev_init(void)
11603 int i, rc = -ENOMEM;
11605 BUG_ON(!dev_boot_phase);
11607 if (dev_proc_init())
11610 if (netdev_kobject_init())
11613 INIT_LIST_HEAD(&ptype_all);
11614 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11615 INIT_LIST_HEAD(&ptype_base[i]);
11617 INIT_LIST_HEAD(&offload_base);
11619 if (register_pernet_subsys(&netdev_net_ops))
11623 * Initialise the packet receive queues.
11626 for_each_possible_cpu(i) {
11627 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11628 struct softnet_data *sd = &per_cpu(softnet_data, i);
11630 INIT_WORK(flush, flush_backlog);
11632 skb_queue_head_init(&sd->input_pkt_queue);
11633 skb_queue_head_init(&sd->process_queue);
11634 #ifdef CONFIG_XFRM_OFFLOAD
11635 skb_queue_head_init(&sd->xfrm_backlog);
11637 INIT_LIST_HEAD(&sd->poll_list);
11638 sd->output_queue_tailp = &sd->output_queue;
11640 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11644 init_gro_hash(&sd->backlog);
11645 sd->backlog.poll = process_backlog;
11646 sd->backlog.weight = weight_p;
11649 dev_boot_phase = 0;
11651 /* The loopback device is special if any other network devices
11652 * is present in a network namespace the loopback device must
11653 * be present. Since we now dynamically allocate and free the
11654 * loopback device ensure this invariant is maintained by
11655 * keeping the loopback device as the first device on the
11656 * list of network devices. Ensuring the loopback devices
11657 * is the first device that appears and the last network device
11660 if (register_pernet_device(&loopback_net_ops))
11663 if (register_pernet_device(&default_device_ops))
11666 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11667 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11669 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11670 NULL, dev_cpu_dead);
11677 subsys_initcall(net_dev_init);