mtd: powernv_flash: Remove pointless goto in driver init
[linux-2.6/btrfs-unstable.git] / net / core / dev.c
blob588b473194a8a8ce23bde368b2b0350a3317db8b
1 /*
2 * NET3 Protocol independent device support routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Derived from the non IP parts of dev.c 1.0.19
10 * Authors: Ross Biro
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
14 * Additional Authors:
15 * Florian la Roche <rzsfl@rz.uni-sb.de>
16 * Alan Cox <gw4pts@gw4pts.ampr.org>
17 * David Hinds <dahinds@users.sourceforge.net>
18 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
19 * Adam Sulmicki <adam@cfar.umd.edu>
20 * Pekka Riikonen <priikone@poesidon.pspt.fi>
22 * Changes:
23 * D.J. Barrow : Fixed bug where dev->refcnt gets set
24 * to 2 if register_netdev gets called
25 * before net_dev_init & also removed a
26 * few lines of code in the process.
27 * Alan Cox : device private ioctl copies fields back.
28 * Alan Cox : Transmit queue code does relevant
29 * stunts to keep the queue safe.
30 * Alan Cox : Fixed double lock.
31 * Alan Cox : Fixed promisc NULL pointer trap
32 * ???????? : Support the full private ioctl range
33 * Alan Cox : Moved ioctl permission check into
34 * drivers
35 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
36 * Alan Cox : 100 backlog just doesn't cut it when
37 * you start doing multicast video 8)
38 * Alan Cox : Rewrote net_bh and list manager.
39 * Alan Cox : Fix ETH_P_ALL echoback lengths.
40 * Alan Cox : Took out transmit every packet pass
41 * Saved a few bytes in the ioctl handler
42 * Alan Cox : Network driver sets packet type before
43 * calling netif_rx. Saves a function
44 * call a packet.
45 * Alan Cox : Hashed net_bh()
46 * Richard Kooijman: Timestamp fixes.
47 * Alan Cox : Wrong field in SIOCGIFDSTADDR
48 * Alan Cox : Device lock protection.
49 * Alan Cox : Fixed nasty side effect of device close
50 * changes.
51 * Rudi Cilibrasi : Pass the right thing to
52 * set_mac_address()
53 * Dave Miller : 32bit quantity for the device lock to
54 * make it work out on a Sparc.
55 * Bjorn Ekwall : Added KERNELD hack.
56 * Alan Cox : Cleaned up the backlog initialise.
57 * Craig Metz : SIOCGIFCONF fix if space for under
58 * 1 device.
59 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
60 * is no device open function.
61 * Andi Kleen : Fix error reporting for SIOCGIFCONF
62 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
63 * Cyrus Durgin : Cleaned for KMOD
64 * Adam Sulmicki : Bug Fix : Network Device Unload
65 * A network device unload needs to purge
66 * the backlog queue.
67 * Paul Rusty Russell : SIOCSIFNAME
68 * Pekka Riikonen : Netdev boot-time settings code
69 * Andrew Morton : Make unregister_netdevice wait
70 * indefinitely on dev->refcnt
71 * J Hadi Salim : - Backlog queue sampling
72 * - netif_rx() feedback
75 #include <linux/uaccess.h>
76 #include <linux/bitops.h>
77 #include <linux/capability.h>
78 #include <linux/cpu.h>
79 #include <linux/types.h>
80 #include <linux/kernel.h>
81 #include <linux/hash.h>
82 #include <linux/slab.h>
83 #include <linux/sched.h>
84 #include <linux/sched/mm.h>
85 #include <linux/mutex.h>
86 #include <linux/string.h>
87 #include <linux/mm.h>
88 #include <linux/socket.h>
89 #include <linux/sockios.h>
90 #include <linux/errno.h>
91 #include <linux/interrupt.h>
92 #include <linux/if_ether.h>
93 #include <linux/netdevice.h>
94 #include <linux/etherdevice.h>
95 #include <linux/ethtool.h>
96 #include <linux/notifier.h>
97 #include <linux/skbuff.h>
98 #include <linux/bpf.h>
99 #include <linux/bpf_trace.h>
100 #include <net/net_namespace.h>
101 #include <net/sock.h>
102 #include <net/busy_poll.h>
103 #include <linux/rtnetlink.h>
104 #include <linux/stat.h>
105 #include <net/dst.h>
106 #include <net/dst_metadata.h>
107 #include <net/pkt_sched.h>
108 #include <net/pkt_cls.h>
109 #include <net/checksum.h>
110 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
126 #include <net/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <linux/pci.h>
136 #include <linux/inetdevice.h>
137 #include <linux/cpu_rmap.h>
138 #include <linux/static_key.h>
139 #include <linux/hashtable.h>
140 #include <linux/vmalloc.h>
141 #include <linux/if_macvlan.h>
142 #include <linux/errqueue.h>
143 #include <linux/hrtimer.h>
144 #include <linux/netfilter_ingress.h>
145 #include <linux/crash_dump.h>
146 #include <linux/sctp.h>
147 #include <net/udp_tunnel.h>
149 #include "net-sysfs.h"
151 /* Instead of increasing this, you should create a hash table. */
152 #define MAX_GRO_SKBS 8
154 /* This should be increased if a protocol with a bigger head is added. */
155 #define GRO_MAX_HEAD (MAX_HEADER + 128)
157 static DEFINE_SPINLOCK(ptype_lock);
158 static DEFINE_SPINLOCK(offload_lock);
159 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
160 struct list_head ptype_all __read_mostly; /* Taps */
161 static struct list_head offload_base __read_mostly;
163 static int netif_rx_internal(struct sk_buff *skb);
164 static int call_netdevice_notifiers_info(unsigned long val,
165 struct net_device *dev,
166 struct netdev_notifier_info *info);
167 static struct napi_struct *napi_by_id(unsigned int napi_id);
170 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
171 * semaphore.
173 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
175 * Writers must hold the rtnl semaphore while they loop through the
176 * dev_base_head list, and hold dev_base_lock for writing when they do the
177 * actual updates. This allows pure readers to access the list even
178 * while a writer is preparing to update it.
180 * To put it another way, dev_base_lock is held for writing only to
181 * protect against pure readers; the rtnl semaphore provides the
182 * protection against other writers.
184 * See, for example usages, register_netdevice() and
185 * unregister_netdevice(), which must be called with the rtnl
186 * semaphore held.
188 DEFINE_RWLOCK(dev_base_lock);
189 EXPORT_SYMBOL(dev_base_lock);
191 /* protects napi_hash addition/deletion and napi_gen_id */
192 static DEFINE_SPINLOCK(napi_hash_lock);
194 static unsigned int napi_gen_id = NR_CPUS;
195 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
197 static seqcount_t devnet_rename_seq;
199 static inline void dev_base_seq_inc(struct net *net)
201 while (++net->dev_base_seq == 0)
205 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
207 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
209 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
212 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
214 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
217 static inline void rps_lock(struct softnet_data *sd)
219 #ifdef CONFIG_RPS
220 spin_lock(&sd->input_pkt_queue.lock);
221 #endif
224 static inline void rps_unlock(struct softnet_data *sd)
226 #ifdef CONFIG_RPS
227 spin_unlock(&sd->input_pkt_queue.lock);
228 #endif
231 /* Device list insertion */
232 static void list_netdevice(struct net_device *dev)
234 struct net *net = dev_net(dev);
236 ASSERT_RTNL();
238 write_lock_bh(&dev_base_lock);
239 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
240 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
241 hlist_add_head_rcu(&dev->index_hlist,
242 dev_index_hash(net, dev->ifindex));
243 write_unlock_bh(&dev_base_lock);
245 dev_base_seq_inc(net);
248 /* Device list removal
249 * caller must respect a RCU grace period before freeing/reusing dev
251 static void unlist_netdevice(struct net_device *dev)
253 ASSERT_RTNL();
255 /* Unlink dev from the device chain */
256 write_lock_bh(&dev_base_lock);
257 list_del_rcu(&dev->dev_list);
258 hlist_del_rcu(&dev->name_hlist);
259 hlist_del_rcu(&dev->index_hlist);
260 write_unlock_bh(&dev_base_lock);
262 dev_base_seq_inc(dev_net(dev));
266 * Our notifier list
269 static RAW_NOTIFIER_HEAD(netdev_chain);
272 * Device drivers call our routines to queue packets here. We empty the
273 * queue in the local softnet handler.
276 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
277 EXPORT_PER_CPU_SYMBOL(softnet_data);
279 #ifdef CONFIG_LOCKDEP
281 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
282 * according to dev->type
284 static const unsigned short netdev_lock_type[] = {
285 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
286 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
287 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
288 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
289 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
290 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
291 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
292 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
293 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
294 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
295 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
296 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
297 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
298 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
299 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
301 static const char *const netdev_lock_name[] = {
302 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
303 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
304 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
305 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
306 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
307 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
308 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
309 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
310 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
311 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
312 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
313 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
314 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
315 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
316 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
318 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
319 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
321 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
323 int i;
325 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
326 if (netdev_lock_type[i] == dev_type)
327 return i;
328 /* the last key is used by default */
329 return ARRAY_SIZE(netdev_lock_type) - 1;
332 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
333 unsigned short dev_type)
335 int i;
337 i = netdev_lock_pos(dev_type);
338 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
339 netdev_lock_name[i]);
342 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
344 int i;
346 i = netdev_lock_pos(dev->type);
347 lockdep_set_class_and_name(&dev->addr_list_lock,
348 &netdev_addr_lock_key[i],
349 netdev_lock_name[i]);
351 #else
352 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
353 unsigned short dev_type)
356 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
359 #endif
361 /*******************************************************************************
363 * Protocol management and registration routines
365 *******************************************************************************/
369 * Add a protocol ID to the list. Now that the input handler is
370 * smarter we can dispense with all the messy stuff that used to be
371 * here.
373 * BEWARE!!! Protocol handlers, mangling input packets,
374 * MUST BE last in hash buckets and checking protocol handlers
375 * MUST start from promiscuous ptype_all chain in net_bh.
376 * It is true now, do not change it.
377 * Explanation follows: if protocol handler, mangling packet, will
378 * be the first on list, it is not able to sense, that packet
379 * is cloned and should be copied-on-write, so that it will
380 * change it and subsequent readers will get broken packet.
381 * --ANK (980803)
384 static inline struct list_head *ptype_head(const struct packet_type *pt)
386 if (pt->type == htons(ETH_P_ALL))
387 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
388 else
389 return pt->dev ? &pt->dev->ptype_specific :
390 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
394 * dev_add_pack - add packet handler
395 * @pt: packet type declaration
397 * Add a protocol handler to the networking stack. The passed &packet_type
398 * is linked into kernel lists and may not be freed until it has been
399 * removed from the kernel lists.
401 * This call does not sleep therefore it can not
402 * guarantee all CPU's that are in middle of receiving packets
403 * will see the new packet type (until the next received packet).
406 void dev_add_pack(struct packet_type *pt)
408 struct list_head *head = ptype_head(pt);
410 spin_lock(&ptype_lock);
411 list_add_rcu(&pt->list, head);
412 spin_unlock(&ptype_lock);
414 EXPORT_SYMBOL(dev_add_pack);
417 * __dev_remove_pack - remove packet handler
418 * @pt: packet type declaration
420 * Remove a protocol handler that was previously added to the kernel
421 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
422 * from the kernel lists and can be freed or reused once this function
423 * returns.
425 * The packet type might still be in use by receivers
426 * and must not be freed until after all the CPU's have gone
427 * through a quiescent state.
429 void __dev_remove_pack(struct packet_type *pt)
431 struct list_head *head = ptype_head(pt);
432 struct packet_type *pt1;
434 spin_lock(&ptype_lock);
436 list_for_each_entry(pt1, head, list) {
437 if (pt == pt1) {
438 list_del_rcu(&pt->list);
439 goto out;
443 pr_warn("dev_remove_pack: %p not found\n", pt);
444 out:
445 spin_unlock(&ptype_lock);
447 EXPORT_SYMBOL(__dev_remove_pack);
450 * dev_remove_pack - remove packet handler
451 * @pt: packet type declaration
453 * Remove a protocol handler that was previously added to the kernel
454 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
455 * from the kernel lists and can be freed or reused once this function
456 * returns.
458 * This call sleeps to guarantee that no CPU is looking at the packet
459 * type after return.
461 void dev_remove_pack(struct packet_type *pt)
463 __dev_remove_pack(pt);
465 synchronize_net();
467 EXPORT_SYMBOL(dev_remove_pack);
471 * dev_add_offload - register offload handlers
472 * @po: protocol offload declaration
474 * Add protocol offload handlers to the networking stack. The passed
475 * &proto_offload is linked into kernel lists and may not be freed until
476 * it has been removed from the kernel lists.
478 * This call does not sleep therefore it can not
479 * guarantee all CPU's that are in middle of receiving packets
480 * will see the new offload handlers (until the next received packet).
482 void dev_add_offload(struct packet_offload *po)
484 struct packet_offload *elem;
486 spin_lock(&offload_lock);
487 list_for_each_entry(elem, &offload_base, list) {
488 if (po->priority < elem->priority)
489 break;
491 list_add_rcu(&po->list, elem->list.prev);
492 spin_unlock(&offload_lock);
494 EXPORT_SYMBOL(dev_add_offload);
497 * __dev_remove_offload - remove offload handler
498 * @po: packet offload declaration
500 * Remove a protocol offload handler that was previously added to the
501 * kernel offload handlers by dev_add_offload(). The passed &offload_type
502 * is removed from the kernel lists and can be freed or reused once this
503 * function returns.
505 * The packet type might still be in use by receivers
506 * and must not be freed until after all the CPU's have gone
507 * through a quiescent state.
509 static void __dev_remove_offload(struct packet_offload *po)
511 struct list_head *head = &offload_base;
512 struct packet_offload *po1;
514 spin_lock(&offload_lock);
516 list_for_each_entry(po1, head, list) {
517 if (po == po1) {
518 list_del_rcu(&po->list);
519 goto out;
523 pr_warn("dev_remove_offload: %p not found\n", po);
524 out:
525 spin_unlock(&offload_lock);
529 * dev_remove_offload - remove packet offload handler
530 * @po: packet offload declaration
532 * Remove a packet offload handler that was previously added to the kernel
533 * offload handlers by dev_add_offload(). The passed &offload_type is
534 * removed from the kernel lists and can be freed or reused once this
535 * function returns.
537 * This call sleeps to guarantee that no CPU is looking at the packet
538 * type after return.
540 void dev_remove_offload(struct packet_offload *po)
542 __dev_remove_offload(po);
544 synchronize_net();
546 EXPORT_SYMBOL(dev_remove_offload);
548 /******************************************************************************
550 * Device Boot-time Settings Routines
552 ******************************************************************************/
554 /* Boot time configuration table */
555 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
558 * netdev_boot_setup_add - add new setup entry
559 * @name: name of the device
560 * @map: configured settings for the device
562 * Adds new setup entry to the dev_boot_setup list. The function
563 * returns 0 on error and 1 on success. This is a generic routine to
564 * all netdevices.
566 static int netdev_boot_setup_add(char *name, struct ifmap *map)
568 struct netdev_boot_setup *s;
569 int i;
571 s = dev_boot_setup;
572 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
573 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
574 memset(s[i].name, 0, sizeof(s[i].name));
575 strlcpy(s[i].name, name, IFNAMSIZ);
576 memcpy(&s[i].map, map, sizeof(s[i].map));
577 break;
581 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
585 * netdev_boot_setup_check - check boot time settings
586 * @dev: the netdevice
588 * Check boot time settings for the device.
589 * The found settings are set for the device to be used
590 * later in the device probing.
591 * Returns 0 if no settings found, 1 if they are.
593 int netdev_boot_setup_check(struct net_device *dev)
595 struct netdev_boot_setup *s = dev_boot_setup;
596 int i;
598 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
599 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
600 !strcmp(dev->name, s[i].name)) {
601 dev->irq = s[i].map.irq;
602 dev->base_addr = s[i].map.base_addr;
603 dev->mem_start = s[i].map.mem_start;
604 dev->mem_end = s[i].map.mem_end;
605 return 1;
608 return 0;
610 EXPORT_SYMBOL(netdev_boot_setup_check);
614 * netdev_boot_base - get address from boot time settings
615 * @prefix: prefix for network device
616 * @unit: id for network device
618 * Check boot time settings for the base address of device.
619 * The found settings are set for the device to be used
620 * later in the device probing.
621 * Returns 0 if no settings found.
623 unsigned long netdev_boot_base(const char *prefix, int unit)
625 const struct netdev_boot_setup *s = dev_boot_setup;
626 char name[IFNAMSIZ];
627 int i;
629 sprintf(name, "%s%d", prefix, unit);
632 * If device already registered then return base of 1
633 * to indicate not to probe for this interface
635 if (__dev_get_by_name(&init_net, name))
636 return 1;
638 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
639 if (!strcmp(name, s[i].name))
640 return s[i].map.base_addr;
641 return 0;
645 * Saves at boot time configured settings for any netdevice.
647 int __init netdev_boot_setup(char *str)
649 int ints[5];
650 struct ifmap map;
652 str = get_options(str, ARRAY_SIZE(ints), ints);
653 if (!str || !*str)
654 return 0;
656 /* Save settings */
657 memset(&map, 0, sizeof(map));
658 if (ints[0] > 0)
659 map.irq = ints[1];
660 if (ints[0] > 1)
661 map.base_addr = ints[2];
662 if (ints[0] > 2)
663 map.mem_start = ints[3];
664 if (ints[0] > 3)
665 map.mem_end = ints[4];
667 /* Add new entry to the list */
668 return netdev_boot_setup_add(str, &map);
671 __setup("netdev=", netdev_boot_setup);
673 /*******************************************************************************
675 * Device Interface Subroutines
677 *******************************************************************************/
680 * dev_get_iflink - get 'iflink' value of a interface
681 * @dev: targeted interface
683 * Indicates the ifindex the interface is linked to.
684 * Physical interfaces have the same 'ifindex' and 'iflink' values.
687 int dev_get_iflink(const struct net_device *dev)
689 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
690 return dev->netdev_ops->ndo_get_iflink(dev);
692 return dev->ifindex;
694 EXPORT_SYMBOL(dev_get_iflink);
697 * dev_fill_metadata_dst - Retrieve tunnel egress information.
698 * @dev: targeted interface
699 * @skb: The packet.
701 * For better visibility of tunnel traffic OVS needs to retrieve
702 * egress tunnel information for a packet. Following API allows
703 * user to get this info.
705 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
707 struct ip_tunnel_info *info;
709 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
710 return -EINVAL;
712 info = skb_tunnel_info_unclone(skb);
713 if (!info)
714 return -ENOMEM;
715 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
716 return -EINVAL;
718 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
720 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
723 * __dev_get_by_name - find a device by its name
724 * @net: the applicable net namespace
725 * @name: name to find
727 * Find an interface by name. Must be called under RTNL semaphore
728 * or @dev_base_lock. If the name is found a pointer to the device
729 * is returned. If the name is not found then %NULL is returned. The
730 * reference counters are not incremented so the caller must be
731 * careful with locks.
734 struct net_device *__dev_get_by_name(struct net *net, const char *name)
736 struct net_device *dev;
737 struct hlist_head *head = dev_name_hash(net, name);
739 hlist_for_each_entry(dev, head, name_hlist)
740 if (!strncmp(dev->name, name, IFNAMSIZ))
741 return dev;
743 return NULL;
745 EXPORT_SYMBOL(__dev_get_by_name);
748 * dev_get_by_name_rcu - find a device by its name
749 * @net: the applicable net namespace
750 * @name: name to find
752 * Find an interface by name.
753 * If the name is found a pointer to the device is returned.
754 * If the name is not found then %NULL is returned.
755 * The reference counters are not incremented so the caller must be
756 * careful with locks. The caller must hold RCU lock.
759 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
761 struct net_device *dev;
762 struct hlist_head *head = dev_name_hash(net, name);
764 hlist_for_each_entry_rcu(dev, head, name_hlist)
765 if (!strncmp(dev->name, name, IFNAMSIZ))
766 return dev;
768 return NULL;
770 EXPORT_SYMBOL(dev_get_by_name_rcu);
773 * dev_get_by_name - find a device by its name
774 * @net: the applicable net namespace
775 * @name: name to find
777 * Find an interface by name. This can be called from any
778 * context and does its own locking. The returned handle has
779 * the usage count incremented and the caller must use dev_put() to
780 * release it when it is no longer needed. %NULL is returned if no
781 * matching device is found.
784 struct net_device *dev_get_by_name(struct net *net, const char *name)
786 struct net_device *dev;
788 rcu_read_lock();
789 dev = dev_get_by_name_rcu(net, name);
790 if (dev)
791 dev_hold(dev);
792 rcu_read_unlock();
793 return dev;
795 EXPORT_SYMBOL(dev_get_by_name);
798 * __dev_get_by_index - find a device by its ifindex
799 * @net: the applicable net namespace
800 * @ifindex: index of device
802 * Search for an interface by index. Returns %NULL if the device
803 * is not found or a pointer to the device. The device has not
804 * had its reference counter increased so the caller must be careful
805 * about locking. The caller must hold either the RTNL semaphore
806 * or @dev_base_lock.
809 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
811 struct net_device *dev;
812 struct hlist_head *head = dev_index_hash(net, ifindex);
814 hlist_for_each_entry(dev, head, index_hlist)
815 if (dev->ifindex == ifindex)
816 return dev;
818 return NULL;
820 EXPORT_SYMBOL(__dev_get_by_index);
823 * dev_get_by_index_rcu - find a device by its ifindex
824 * @net: the applicable net namespace
825 * @ifindex: index of device
827 * Search for an interface by index. Returns %NULL if the device
828 * is not found or a pointer to the device. The device has not
829 * had its reference counter increased so the caller must be careful
830 * about locking. The caller must hold RCU lock.
833 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
835 struct net_device *dev;
836 struct hlist_head *head = dev_index_hash(net, ifindex);
838 hlist_for_each_entry_rcu(dev, head, index_hlist)
839 if (dev->ifindex == ifindex)
840 return dev;
842 return NULL;
844 EXPORT_SYMBOL(dev_get_by_index_rcu);
848 * dev_get_by_index - find a device by its ifindex
849 * @net: the applicable net namespace
850 * @ifindex: index of device
852 * Search for an interface by index. Returns NULL if the device
853 * is not found or a pointer to the device. The device returned has
854 * had a reference added and the pointer is safe until the user calls
855 * dev_put to indicate they have finished with it.
858 struct net_device *dev_get_by_index(struct net *net, int ifindex)
860 struct net_device *dev;
862 rcu_read_lock();
863 dev = dev_get_by_index_rcu(net, ifindex);
864 if (dev)
865 dev_hold(dev);
866 rcu_read_unlock();
867 return dev;
869 EXPORT_SYMBOL(dev_get_by_index);
872 * dev_get_by_napi_id - find a device by napi_id
873 * @napi_id: ID of the NAPI struct
875 * Search for an interface by NAPI ID. Returns %NULL if the device
876 * is not found or a pointer to the device. The device has not had
877 * its reference counter increased so the caller must be careful
878 * about locking. The caller must hold RCU lock.
881 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
883 struct napi_struct *napi;
885 WARN_ON_ONCE(!rcu_read_lock_held());
887 if (napi_id < MIN_NAPI_ID)
888 return NULL;
890 napi = napi_by_id(napi_id);
892 return napi ? napi->dev : NULL;
894 EXPORT_SYMBOL(dev_get_by_napi_id);
897 * netdev_get_name - get a netdevice name, knowing its ifindex.
898 * @net: network namespace
899 * @name: a pointer to the buffer where the name will be stored.
900 * @ifindex: the ifindex of the interface to get the name from.
902 * The use of raw_seqcount_begin() and cond_resched() before
903 * retrying is required as we want to give the writers a chance
904 * to complete when CONFIG_PREEMPT is not set.
906 int netdev_get_name(struct net *net, char *name, int ifindex)
908 struct net_device *dev;
909 unsigned int seq;
911 retry:
912 seq = raw_seqcount_begin(&devnet_rename_seq);
913 rcu_read_lock();
914 dev = dev_get_by_index_rcu(net, ifindex);
915 if (!dev) {
916 rcu_read_unlock();
917 return -ENODEV;
920 strcpy(name, dev->name);
921 rcu_read_unlock();
922 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
923 cond_resched();
924 goto retry;
927 return 0;
931 * dev_getbyhwaddr_rcu - find a device by its hardware address
932 * @net: the applicable net namespace
933 * @type: media type of device
934 * @ha: hardware address
936 * Search for an interface by MAC address. Returns NULL if the device
937 * is not found or a pointer to the device.
938 * The caller must hold RCU or RTNL.
939 * The returned device has not had its ref count increased
940 * and the caller must therefore be careful about locking
944 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
945 const char *ha)
947 struct net_device *dev;
949 for_each_netdev_rcu(net, dev)
950 if (dev->type == type &&
951 !memcmp(dev->dev_addr, ha, dev->addr_len))
952 return dev;
954 return NULL;
956 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
958 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
960 struct net_device *dev;
962 ASSERT_RTNL();
963 for_each_netdev(net, dev)
964 if (dev->type == type)
965 return dev;
967 return NULL;
969 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
971 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
973 struct net_device *dev, *ret = NULL;
975 rcu_read_lock();
976 for_each_netdev_rcu(net, dev)
977 if (dev->type == type) {
978 dev_hold(dev);
979 ret = dev;
980 break;
982 rcu_read_unlock();
983 return ret;
985 EXPORT_SYMBOL(dev_getfirstbyhwtype);
988 * __dev_get_by_flags - find any device with given flags
989 * @net: the applicable net namespace
990 * @if_flags: IFF_* values
991 * @mask: bitmask of bits in if_flags to check
993 * Search for any interface with the given flags. Returns NULL if a device
994 * is not found or a pointer to the device. Must be called inside
995 * rtnl_lock(), and result refcount is unchanged.
998 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
999 unsigned short mask)
1001 struct net_device *dev, *ret;
1003 ASSERT_RTNL();
1005 ret = NULL;
1006 for_each_netdev(net, dev) {
1007 if (((dev->flags ^ if_flags) & mask) == 0) {
1008 ret = dev;
1009 break;
1012 return ret;
1014 EXPORT_SYMBOL(__dev_get_by_flags);
1017 * dev_valid_name - check if name is okay for network device
1018 * @name: name string
1020 * Network device names need to be valid file names to
1021 * to allow sysfs to work. We also disallow any kind of
1022 * whitespace.
1024 bool dev_valid_name(const char *name)
1026 if (*name == '\0')
1027 return false;
1028 if (strlen(name) >= IFNAMSIZ)
1029 return false;
1030 if (!strcmp(name, ".") || !strcmp(name, ".."))
1031 return false;
1033 while (*name) {
1034 if (*name == '/' || *name == ':' || isspace(*name))
1035 return false;
1036 name++;
1038 return true;
1040 EXPORT_SYMBOL(dev_valid_name);
1043 * __dev_alloc_name - allocate a name for a device
1044 * @net: network namespace to allocate the device name in
1045 * @name: name format string
1046 * @buf: scratch buffer and result name string
1048 * Passed a format string - eg "lt%d" it will try and find a suitable
1049 * id. It scans list of devices to build up a free map, then chooses
1050 * the first empty slot. The caller must hold the dev_base or rtnl lock
1051 * while allocating the name and adding the device in order to avoid
1052 * duplicates.
1053 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1054 * Returns the number of the unit assigned or a negative errno code.
1057 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1059 int i = 0;
1060 const char *p;
1061 const int max_netdevices = 8*PAGE_SIZE;
1062 unsigned long *inuse;
1063 struct net_device *d;
1065 p = strnchr(name, IFNAMSIZ-1, '%');
1066 if (p) {
1068 * Verify the string as this thing may have come from
1069 * the user. There must be either one "%d" and no other "%"
1070 * characters.
1072 if (p[1] != 'd' || strchr(p + 2, '%'))
1073 return -EINVAL;
1075 /* Use one page as a bit array of possible slots */
1076 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1077 if (!inuse)
1078 return -ENOMEM;
1080 for_each_netdev(net, d) {
1081 if (!sscanf(d->name, name, &i))
1082 continue;
1083 if (i < 0 || i >= max_netdevices)
1084 continue;
1086 /* avoid cases where sscanf is not exact inverse of printf */
1087 snprintf(buf, IFNAMSIZ, name, i);
1088 if (!strncmp(buf, d->name, IFNAMSIZ))
1089 set_bit(i, inuse);
1092 i = find_first_zero_bit(inuse, max_netdevices);
1093 free_page((unsigned long) inuse);
1096 if (buf != name)
1097 snprintf(buf, IFNAMSIZ, name, i);
1098 if (!__dev_get_by_name(net, buf))
1099 return i;
1101 /* It is possible to run out of possible slots
1102 * when the name is long and there isn't enough space left
1103 * for the digits, or if all bits are used.
1105 return -ENFILE;
1109 * dev_alloc_name - allocate a name for a device
1110 * @dev: device
1111 * @name: name format string
1113 * Passed a format string - eg "lt%d" it will try and find a suitable
1114 * id. It scans list of devices to build up a free map, then chooses
1115 * the first empty slot. The caller must hold the dev_base or rtnl lock
1116 * while allocating the name and adding the device in order to avoid
1117 * duplicates.
1118 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1119 * Returns the number of the unit assigned or a negative errno code.
1122 int dev_alloc_name(struct net_device *dev, const char *name)
1124 char buf[IFNAMSIZ];
1125 struct net *net;
1126 int ret;
1128 BUG_ON(!dev_net(dev));
1129 net = dev_net(dev);
1130 ret = __dev_alloc_name(net, name, buf);
1131 if (ret >= 0)
1132 strlcpy(dev->name, buf, IFNAMSIZ);
1133 return ret;
1135 EXPORT_SYMBOL(dev_alloc_name);
1137 static int dev_alloc_name_ns(struct net *net,
1138 struct net_device *dev,
1139 const char *name)
1141 char buf[IFNAMSIZ];
1142 int ret;
1144 ret = __dev_alloc_name(net, name, buf);
1145 if (ret >= 0)
1146 strlcpy(dev->name, buf, IFNAMSIZ);
1147 return ret;
1150 static int dev_get_valid_name(struct net *net,
1151 struct net_device *dev,
1152 const char *name)
1154 BUG_ON(!net);
1156 if (!dev_valid_name(name))
1157 return -EINVAL;
1159 if (strchr(name, '%'))
1160 return dev_alloc_name_ns(net, dev, name);
1161 else if (__dev_get_by_name(net, name))
1162 return -EEXIST;
1163 else if (dev->name != name)
1164 strlcpy(dev->name, name, IFNAMSIZ);
1166 return 0;
1170 * dev_change_name - change name of a device
1171 * @dev: device
1172 * @newname: name (or format string) must be at least IFNAMSIZ
1174 * Change name of a device, can pass format strings "eth%d".
1175 * for wildcarding.
1177 int dev_change_name(struct net_device *dev, const char *newname)
1179 unsigned char old_assign_type;
1180 char oldname[IFNAMSIZ];
1181 int err = 0;
1182 int ret;
1183 struct net *net;
1185 ASSERT_RTNL();
1186 BUG_ON(!dev_net(dev));
1188 net = dev_net(dev);
1189 if (dev->flags & IFF_UP)
1190 return -EBUSY;
1192 write_seqcount_begin(&devnet_rename_seq);
1194 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1195 write_seqcount_end(&devnet_rename_seq);
1196 return 0;
1199 memcpy(oldname, dev->name, IFNAMSIZ);
1201 err = dev_get_valid_name(net, dev, newname);
1202 if (err < 0) {
1203 write_seqcount_end(&devnet_rename_seq);
1204 return err;
1207 if (oldname[0] && !strchr(oldname, '%'))
1208 netdev_info(dev, "renamed from %s\n", oldname);
1210 old_assign_type = dev->name_assign_type;
1211 dev->name_assign_type = NET_NAME_RENAMED;
1213 rollback:
1214 ret = device_rename(&dev->dev, dev->name);
1215 if (ret) {
1216 memcpy(dev->name, oldname, IFNAMSIZ);
1217 dev->name_assign_type = old_assign_type;
1218 write_seqcount_end(&devnet_rename_seq);
1219 return ret;
1222 write_seqcount_end(&devnet_rename_seq);
1224 netdev_adjacent_rename_links(dev, oldname);
1226 write_lock_bh(&dev_base_lock);
1227 hlist_del_rcu(&dev->name_hlist);
1228 write_unlock_bh(&dev_base_lock);
1230 synchronize_rcu();
1232 write_lock_bh(&dev_base_lock);
1233 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1234 write_unlock_bh(&dev_base_lock);
1236 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1237 ret = notifier_to_errno(ret);
1239 if (ret) {
1240 /* err >= 0 after dev_alloc_name() or stores the first errno */
1241 if (err >= 0) {
1242 err = ret;
1243 write_seqcount_begin(&devnet_rename_seq);
1244 memcpy(dev->name, oldname, IFNAMSIZ);
1245 memcpy(oldname, newname, IFNAMSIZ);
1246 dev->name_assign_type = old_assign_type;
1247 old_assign_type = NET_NAME_RENAMED;
1248 goto rollback;
1249 } else {
1250 pr_err("%s: name change rollback failed: %d\n",
1251 dev->name, ret);
1255 return err;
1259 * dev_set_alias - change ifalias of a device
1260 * @dev: device
1261 * @alias: name up to IFALIASZ
1262 * @len: limit of bytes to copy from info
1264 * Set ifalias for a device,
1266 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1268 char *new_ifalias;
1270 ASSERT_RTNL();
1272 if (len >= IFALIASZ)
1273 return -EINVAL;
1275 if (!len) {
1276 kfree(dev->ifalias);
1277 dev->ifalias = NULL;
1278 return 0;
1281 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1282 if (!new_ifalias)
1283 return -ENOMEM;
1284 dev->ifalias = new_ifalias;
1285 memcpy(dev->ifalias, alias, len);
1286 dev->ifalias[len] = 0;
1288 return len;
1293 * netdev_features_change - device changes features
1294 * @dev: device to cause notification
1296 * Called to indicate a device has changed features.
1298 void netdev_features_change(struct net_device *dev)
1300 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1302 EXPORT_SYMBOL(netdev_features_change);
1305 * netdev_state_change - device changes state
1306 * @dev: device to cause notification
1308 * Called to indicate a device has changed state. This function calls
1309 * the notifier chains for netdev_chain and sends a NEWLINK message
1310 * to the routing socket.
1312 void netdev_state_change(struct net_device *dev)
1314 if (dev->flags & IFF_UP) {
1315 struct netdev_notifier_change_info change_info;
1317 change_info.flags_changed = 0;
1318 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
1319 &change_info.info);
1320 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1323 EXPORT_SYMBOL(netdev_state_change);
1326 * netdev_notify_peers - notify network peers about existence of @dev
1327 * @dev: network device
1329 * Generate traffic such that interested network peers are aware of
1330 * @dev, such as by generating a gratuitous ARP. This may be used when
1331 * a device wants to inform the rest of the network about some sort of
1332 * reconfiguration such as a failover event or virtual machine
1333 * migration.
1335 void netdev_notify_peers(struct net_device *dev)
1337 rtnl_lock();
1338 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1339 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1340 rtnl_unlock();
1342 EXPORT_SYMBOL(netdev_notify_peers);
1344 static int __dev_open(struct net_device *dev)
1346 const struct net_device_ops *ops = dev->netdev_ops;
1347 int ret;
1349 ASSERT_RTNL();
1351 if (!netif_device_present(dev))
1352 return -ENODEV;
1354 /* Block netpoll from trying to do any rx path servicing.
1355 * If we don't do this there is a chance ndo_poll_controller
1356 * or ndo_poll may be running while we open the device
1358 netpoll_poll_disable(dev);
1360 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1361 ret = notifier_to_errno(ret);
1362 if (ret)
1363 return ret;
1365 set_bit(__LINK_STATE_START, &dev->state);
1367 if (ops->ndo_validate_addr)
1368 ret = ops->ndo_validate_addr(dev);
1370 if (!ret && ops->ndo_open)
1371 ret = ops->ndo_open(dev);
1373 netpoll_poll_enable(dev);
1375 if (ret)
1376 clear_bit(__LINK_STATE_START, &dev->state);
1377 else {
1378 dev->flags |= IFF_UP;
1379 dev_set_rx_mode(dev);
1380 dev_activate(dev);
1381 add_device_randomness(dev->dev_addr, dev->addr_len);
1384 return ret;
1388 * dev_open - prepare an interface for use.
1389 * @dev: device to open
1391 * Takes a device from down to up state. The device's private open
1392 * function is invoked and then the multicast lists are loaded. Finally
1393 * the device is moved into the up state and a %NETDEV_UP message is
1394 * sent to the netdev notifier chain.
1396 * Calling this function on an active interface is a nop. On a failure
1397 * a negative errno code is returned.
1399 int dev_open(struct net_device *dev)
1401 int ret;
1403 if (dev->flags & IFF_UP)
1404 return 0;
1406 ret = __dev_open(dev);
1407 if (ret < 0)
1408 return ret;
1410 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1411 call_netdevice_notifiers(NETDEV_UP, dev);
1413 return ret;
1415 EXPORT_SYMBOL(dev_open);
1417 static void __dev_close_many(struct list_head *head)
1419 struct net_device *dev;
1421 ASSERT_RTNL();
1422 might_sleep();
1424 list_for_each_entry(dev, head, close_list) {
1425 /* Temporarily disable netpoll until the interface is down */
1426 netpoll_poll_disable(dev);
1428 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1430 clear_bit(__LINK_STATE_START, &dev->state);
1432 /* Synchronize to scheduled poll. We cannot touch poll list, it
1433 * can be even on different cpu. So just clear netif_running().
1435 * dev->stop() will invoke napi_disable() on all of it's
1436 * napi_struct instances on this device.
1438 smp_mb__after_atomic(); /* Commit netif_running(). */
1441 dev_deactivate_many(head);
1443 list_for_each_entry(dev, head, close_list) {
1444 const struct net_device_ops *ops = dev->netdev_ops;
1447 * Call the device specific close. This cannot fail.
1448 * Only if device is UP
1450 * We allow it to be called even after a DETACH hot-plug
1451 * event.
1453 if (ops->ndo_stop)
1454 ops->ndo_stop(dev);
1456 dev->flags &= ~IFF_UP;
1457 netpoll_poll_enable(dev);
1461 static void __dev_close(struct net_device *dev)
1463 LIST_HEAD(single);
1465 list_add(&dev->close_list, &single);
1466 __dev_close_many(&single);
1467 list_del(&single);
1470 void dev_close_many(struct list_head *head, bool unlink)
1472 struct net_device *dev, *tmp;
1474 /* Remove the devices that don't need to be closed */
1475 list_for_each_entry_safe(dev, tmp, head, close_list)
1476 if (!(dev->flags & IFF_UP))
1477 list_del_init(&dev->close_list);
1479 __dev_close_many(head);
1481 list_for_each_entry_safe(dev, tmp, head, close_list) {
1482 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1483 call_netdevice_notifiers(NETDEV_DOWN, dev);
1484 if (unlink)
1485 list_del_init(&dev->close_list);
1488 EXPORT_SYMBOL(dev_close_many);
1491 * dev_close - shutdown an interface.
1492 * @dev: device to shutdown
1494 * This function moves an active device into down state. A
1495 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1496 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1497 * chain.
1499 void dev_close(struct net_device *dev)
1501 if (dev->flags & IFF_UP) {
1502 LIST_HEAD(single);
1504 list_add(&dev->close_list, &single);
1505 dev_close_many(&single, true);
1506 list_del(&single);
1509 EXPORT_SYMBOL(dev_close);
1513 * dev_disable_lro - disable Large Receive Offload on a device
1514 * @dev: device
1516 * Disable Large Receive Offload (LRO) on a net device. Must be
1517 * called under RTNL. This is needed if received packets may be
1518 * forwarded to another interface.
1520 void dev_disable_lro(struct net_device *dev)
1522 struct net_device *lower_dev;
1523 struct list_head *iter;
1525 dev->wanted_features &= ~NETIF_F_LRO;
1526 netdev_update_features(dev);
1528 if (unlikely(dev->features & NETIF_F_LRO))
1529 netdev_WARN(dev, "failed to disable LRO!\n");
1531 netdev_for_each_lower_dev(dev, lower_dev, iter)
1532 dev_disable_lro(lower_dev);
1534 EXPORT_SYMBOL(dev_disable_lro);
1536 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1537 struct net_device *dev)
1539 struct netdev_notifier_info info;
1541 netdev_notifier_info_init(&info, dev);
1542 return nb->notifier_call(nb, val, &info);
1545 static int dev_boot_phase = 1;
1548 * register_netdevice_notifier - register a network notifier block
1549 * @nb: notifier
1551 * Register a notifier to be called when network device events occur.
1552 * The notifier passed is linked into the kernel structures and must
1553 * not be reused until it has been unregistered. A negative errno code
1554 * is returned on a failure.
1556 * When registered all registration and up events are replayed
1557 * to the new notifier to allow device to have a race free
1558 * view of the network device list.
1561 int register_netdevice_notifier(struct notifier_block *nb)
1563 struct net_device *dev;
1564 struct net_device *last;
1565 struct net *net;
1566 int err;
1568 rtnl_lock();
1569 err = raw_notifier_chain_register(&netdev_chain, nb);
1570 if (err)
1571 goto unlock;
1572 if (dev_boot_phase)
1573 goto unlock;
1574 for_each_net(net) {
1575 for_each_netdev(net, dev) {
1576 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1577 err = notifier_to_errno(err);
1578 if (err)
1579 goto rollback;
1581 if (!(dev->flags & IFF_UP))
1582 continue;
1584 call_netdevice_notifier(nb, NETDEV_UP, dev);
1588 unlock:
1589 rtnl_unlock();
1590 return err;
1592 rollback:
1593 last = dev;
1594 for_each_net(net) {
1595 for_each_netdev(net, dev) {
1596 if (dev == last)
1597 goto outroll;
1599 if (dev->flags & IFF_UP) {
1600 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1601 dev);
1602 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1604 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1608 outroll:
1609 raw_notifier_chain_unregister(&netdev_chain, nb);
1610 goto unlock;
1612 EXPORT_SYMBOL(register_netdevice_notifier);
1615 * unregister_netdevice_notifier - unregister a network notifier block
1616 * @nb: notifier
1618 * Unregister a notifier previously registered by
1619 * register_netdevice_notifier(). The notifier is unlinked into the
1620 * kernel structures and may then be reused. A negative errno code
1621 * is returned on a failure.
1623 * After unregistering unregister and down device events are synthesized
1624 * for all devices on the device list to the removed notifier to remove
1625 * the need for special case cleanup code.
1628 int unregister_netdevice_notifier(struct notifier_block *nb)
1630 struct net_device *dev;
1631 struct net *net;
1632 int err;
1634 rtnl_lock();
1635 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1636 if (err)
1637 goto unlock;
1639 for_each_net(net) {
1640 for_each_netdev(net, dev) {
1641 if (dev->flags & IFF_UP) {
1642 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1643 dev);
1644 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1646 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1649 unlock:
1650 rtnl_unlock();
1651 return err;
1653 EXPORT_SYMBOL(unregister_netdevice_notifier);
1656 * call_netdevice_notifiers_info - call all network notifier blocks
1657 * @val: value passed unmodified to notifier function
1658 * @dev: net_device pointer passed unmodified to notifier function
1659 * @info: notifier information data
1661 * Call all network notifier blocks. Parameters and return value
1662 * are as for raw_notifier_call_chain().
1665 static int call_netdevice_notifiers_info(unsigned long val,
1666 struct net_device *dev,
1667 struct netdev_notifier_info *info)
1669 ASSERT_RTNL();
1670 netdev_notifier_info_init(info, dev);
1671 return raw_notifier_call_chain(&netdev_chain, val, info);
1675 * call_netdevice_notifiers - call all network notifier blocks
1676 * @val: value passed unmodified to notifier function
1677 * @dev: net_device pointer passed unmodified to notifier function
1679 * Call all network notifier blocks. Parameters and return value
1680 * are as for raw_notifier_call_chain().
1683 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1685 struct netdev_notifier_info info;
1687 return call_netdevice_notifiers_info(val, dev, &info);
1689 EXPORT_SYMBOL(call_netdevice_notifiers);
1691 #ifdef CONFIG_NET_INGRESS
1692 static struct static_key ingress_needed __read_mostly;
1694 void net_inc_ingress_queue(void)
1696 static_key_slow_inc(&ingress_needed);
1698 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1700 void net_dec_ingress_queue(void)
1702 static_key_slow_dec(&ingress_needed);
1704 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1705 #endif
1707 #ifdef CONFIG_NET_EGRESS
1708 static struct static_key egress_needed __read_mostly;
1710 void net_inc_egress_queue(void)
1712 static_key_slow_inc(&egress_needed);
1714 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1716 void net_dec_egress_queue(void)
1718 static_key_slow_dec(&egress_needed);
1720 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1721 #endif
1723 static struct static_key netstamp_needed __read_mostly;
1724 #ifdef HAVE_JUMP_LABEL
1725 static atomic_t netstamp_needed_deferred;
1726 static atomic_t netstamp_wanted;
1727 static void netstamp_clear(struct work_struct *work)
1729 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1730 int wanted;
1732 wanted = atomic_add_return(deferred, &netstamp_wanted);
1733 if (wanted > 0)
1734 static_key_enable(&netstamp_needed);
1735 else
1736 static_key_disable(&netstamp_needed);
1738 static DECLARE_WORK(netstamp_work, netstamp_clear);
1739 #endif
1741 void net_enable_timestamp(void)
1743 #ifdef HAVE_JUMP_LABEL
1744 int wanted;
1746 while (1) {
1747 wanted = atomic_read(&netstamp_wanted);
1748 if (wanted <= 0)
1749 break;
1750 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1751 return;
1753 atomic_inc(&netstamp_needed_deferred);
1754 schedule_work(&netstamp_work);
1755 #else
1756 static_key_slow_inc(&netstamp_needed);
1757 #endif
1759 EXPORT_SYMBOL(net_enable_timestamp);
1761 void net_disable_timestamp(void)
1763 #ifdef HAVE_JUMP_LABEL
1764 int wanted;
1766 while (1) {
1767 wanted = atomic_read(&netstamp_wanted);
1768 if (wanted <= 1)
1769 break;
1770 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1771 return;
1773 atomic_dec(&netstamp_needed_deferred);
1774 schedule_work(&netstamp_work);
1775 #else
1776 static_key_slow_dec(&netstamp_needed);
1777 #endif
1779 EXPORT_SYMBOL(net_disable_timestamp);
1781 static inline void net_timestamp_set(struct sk_buff *skb)
1783 skb->tstamp = 0;
1784 if (static_key_false(&netstamp_needed))
1785 __net_timestamp(skb);
1788 #define net_timestamp_check(COND, SKB) \
1789 if (static_key_false(&netstamp_needed)) { \
1790 if ((COND) && !(SKB)->tstamp) \
1791 __net_timestamp(SKB); \
1794 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1796 unsigned int len;
1798 if (!(dev->flags & IFF_UP))
1799 return false;
1801 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1802 if (skb->len <= len)
1803 return true;
1805 /* if TSO is enabled, we don't care about the length as the packet
1806 * could be forwarded without being segmented before
1808 if (skb_is_gso(skb))
1809 return true;
1811 return false;
1813 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1815 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1817 int ret = ____dev_forward_skb(dev, skb);
1819 if (likely(!ret)) {
1820 skb->protocol = eth_type_trans(skb, dev);
1821 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1824 return ret;
1826 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1829 * dev_forward_skb - loopback an skb to another netif
1831 * @dev: destination network device
1832 * @skb: buffer to forward
1834 * return values:
1835 * NET_RX_SUCCESS (no congestion)
1836 * NET_RX_DROP (packet was dropped, but freed)
1838 * dev_forward_skb can be used for injecting an skb from the
1839 * start_xmit function of one device into the receive queue
1840 * of another device.
1842 * The receiving device may be in another namespace, so
1843 * we have to clear all information in the skb that could
1844 * impact namespace isolation.
1846 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1848 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1850 EXPORT_SYMBOL_GPL(dev_forward_skb);
1852 static inline int deliver_skb(struct sk_buff *skb,
1853 struct packet_type *pt_prev,
1854 struct net_device *orig_dev)
1856 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1857 return -ENOMEM;
1858 refcount_inc(&skb->users);
1859 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1862 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1863 struct packet_type **pt,
1864 struct net_device *orig_dev,
1865 __be16 type,
1866 struct list_head *ptype_list)
1868 struct packet_type *ptype, *pt_prev = *pt;
1870 list_for_each_entry_rcu(ptype, ptype_list, list) {
1871 if (ptype->type != type)
1872 continue;
1873 if (pt_prev)
1874 deliver_skb(skb, pt_prev, orig_dev);
1875 pt_prev = ptype;
1877 *pt = pt_prev;
1880 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1882 if (!ptype->af_packet_priv || !skb->sk)
1883 return false;
1885 if (ptype->id_match)
1886 return ptype->id_match(ptype, skb->sk);
1887 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1888 return true;
1890 return false;
1894 * Support routine. Sends outgoing frames to any network
1895 * taps currently in use.
1898 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1900 struct packet_type *ptype;
1901 struct sk_buff *skb2 = NULL;
1902 struct packet_type *pt_prev = NULL;
1903 struct list_head *ptype_list = &ptype_all;
1905 rcu_read_lock();
1906 again:
1907 list_for_each_entry_rcu(ptype, ptype_list, list) {
1908 /* Never send packets back to the socket
1909 * they originated from - MvS (miquels@drinkel.ow.org)
1911 if (skb_loop_sk(ptype, skb))
1912 continue;
1914 if (pt_prev) {
1915 deliver_skb(skb2, pt_prev, skb->dev);
1916 pt_prev = ptype;
1917 continue;
1920 /* need to clone skb, done only once */
1921 skb2 = skb_clone(skb, GFP_ATOMIC);
1922 if (!skb2)
1923 goto out_unlock;
1925 net_timestamp_set(skb2);
1927 /* skb->nh should be correctly
1928 * set by sender, so that the second statement is
1929 * just protection against buggy protocols.
1931 skb_reset_mac_header(skb2);
1933 if (skb_network_header(skb2) < skb2->data ||
1934 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1935 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1936 ntohs(skb2->protocol),
1937 dev->name);
1938 skb_reset_network_header(skb2);
1941 skb2->transport_header = skb2->network_header;
1942 skb2->pkt_type = PACKET_OUTGOING;
1943 pt_prev = ptype;
1946 if (ptype_list == &ptype_all) {
1947 ptype_list = &dev->ptype_all;
1948 goto again;
1950 out_unlock:
1951 if (pt_prev) {
1952 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
1953 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1954 else
1955 kfree_skb(skb2);
1957 rcu_read_unlock();
1959 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
1962 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1963 * @dev: Network device
1964 * @txq: number of queues available
1966 * If real_num_tx_queues is changed the tc mappings may no longer be
1967 * valid. To resolve this verify the tc mapping remains valid and if
1968 * not NULL the mapping. With no priorities mapping to this
1969 * offset/count pair it will no longer be used. In the worst case TC0
1970 * is invalid nothing can be done so disable priority mappings. If is
1971 * expected that drivers will fix this mapping if they can before
1972 * calling netif_set_real_num_tx_queues.
1974 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1976 int i;
1977 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1979 /* If TC0 is invalidated disable TC mapping */
1980 if (tc->offset + tc->count > txq) {
1981 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1982 dev->num_tc = 0;
1983 return;
1986 /* Invalidated prio to tc mappings set to TC0 */
1987 for (i = 1; i < TC_BITMASK + 1; i++) {
1988 int q = netdev_get_prio_tc_map(dev, i);
1990 tc = &dev->tc_to_txq[q];
1991 if (tc->offset + tc->count > txq) {
1992 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1993 i, q);
1994 netdev_set_prio_tc_map(dev, i, 0);
1999 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2001 if (dev->num_tc) {
2002 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2003 int i;
2005 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2006 if ((txq - tc->offset) < tc->count)
2007 return i;
2010 return -1;
2013 return 0;
2016 #ifdef CONFIG_XPS
2017 static DEFINE_MUTEX(xps_map_mutex);
2018 #define xmap_dereference(P) \
2019 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2021 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2022 int tci, u16 index)
2024 struct xps_map *map = NULL;
2025 int pos;
2027 if (dev_maps)
2028 map = xmap_dereference(dev_maps->cpu_map[tci]);
2029 if (!map)
2030 return false;
2032 for (pos = map->len; pos--;) {
2033 if (map->queues[pos] != index)
2034 continue;
2036 if (map->len > 1) {
2037 map->queues[pos] = map->queues[--map->len];
2038 break;
2041 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2042 kfree_rcu(map, rcu);
2043 return false;
2046 return true;
2049 static bool remove_xps_queue_cpu(struct net_device *dev,
2050 struct xps_dev_maps *dev_maps,
2051 int cpu, u16 offset, u16 count)
2053 int num_tc = dev->num_tc ? : 1;
2054 bool active = false;
2055 int tci;
2057 for (tci = cpu * num_tc; num_tc--; tci++) {
2058 int i, j;
2060 for (i = count, j = offset; i--; j++) {
2061 if (!remove_xps_queue(dev_maps, cpu, j))
2062 break;
2065 active |= i < 0;
2068 return active;
2071 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2072 u16 count)
2074 struct xps_dev_maps *dev_maps;
2075 int cpu, i;
2076 bool active = false;
2078 mutex_lock(&xps_map_mutex);
2079 dev_maps = xmap_dereference(dev->xps_maps);
2081 if (!dev_maps)
2082 goto out_no_maps;
2084 for_each_possible_cpu(cpu)
2085 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2086 offset, count);
2088 if (!active) {
2089 RCU_INIT_POINTER(dev->xps_maps, NULL);
2090 kfree_rcu(dev_maps, rcu);
2093 for (i = offset + (count - 1); count--; i--)
2094 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2095 NUMA_NO_NODE);
2097 out_no_maps:
2098 mutex_unlock(&xps_map_mutex);
2101 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2103 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2106 static struct xps_map *expand_xps_map(struct xps_map *map,
2107 int cpu, u16 index)
2109 struct xps_map *new_map;
2110 int alloc_len = XPS_MIN_MAP_ALLOC;
2111 int i, pos;
2113 for (pos = 0; map && pos < map->len; pos++) {
2114 if (map->queues[pos] != index)
2115 continue;
2116 return map;
2119 /* Need to add queue to this CPU's existing map */
2120 if (map) {
2121 if (pos < map->alloc_len)
2122 return map;
2124 alloc_len = map->alloc_len * 2;
2127 /* Need to allocate new map to store queue on this CPU's map */
2128 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2129 cpu_to_node(cpu));
2130 if (!new_map)
2131 return NULL;
2133 for (i = 0; i < pos; i++)
2134 new_map->queues[i] = map->queues[i];
2135 new_map->alloc_len = alloc_len;
2136 new_map->len = pos;
2138 return new_map;
2141 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2142 u16 index)
2144 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2145 int i, cpu, tci, numa_node_id = -2;
2146 int maps_sz, num_tc = 1, tc = 0;
2147 struct xps_map *map, *new_map;
2148 bool active = false;
2150 if (dev->num_tc) {
2151 num_tc = dev->num_tc;
2152 tc = netdev_txq_to_tc(dev, index);
2153 if (tc < 0)
2154 return -EINVAL;
2157 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2158 if (maps_sz < L1_CACHE_BYTES)
2159 maps_sz = L1_CACHE_BYTES;
2161 mutex_lock(&xps_map_mutex);
2163 dev_maps = xmap_dereference(dev->xps_maps);
2165 /* allocate memory for queue storage */
2166 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2167 if (!new_dev_maps)
2168 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2169 if (!new_dev_maps) {
2170 mutex_unlock(&xps_map_mutex);
2171 return -ENOMEM;
2174 tci = cpu * num_tc + tc;
2175 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2176 NULL;
2178 map = expand_xps_map(map, cpu, index);
2179 if (!map)
2180 goto error;
2182 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2185 if (!new_dev_maps)
2186 goto out_no_new_maps;
2188 for_each_possible_cpu(cpu) {
2189 /* copy maps belonging to foreign traffic classes */
2190 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2191 /* fill in the new device map from the old device map */
2192 map = xmap_dereference(dev_maps->cpu_map[tci]);
2193 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2196 /* We need to explicitly update tci as prevous loop
2197 * could break out early if dev_maps is NULL.
2199 tci = cpu * num_tc + tc;
2201 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2202 /* add queue to CPU maps */
2203 int pos = 0;
2205 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2206 while ((pos < map->len) && (map->queues[pos] != index))
2207 pos++;
2209 if (pos == map->len)
2210 map->queues[map->len++] = index;
2211 #ifdef CONFIG_NUMA
2212 if (numa_node_id == -2)
2213 numa_node_id = cpu_to_node(cpu);
2214 else if (numa_node_id != cpu_to_node(cpu))
2215 numa_node_id = -1;
2216 #endif
2217 } else if (dev_maps) {
2218 /* fill in the new device map from the old device map */
2219 map = xmap_dereference(dev_maps->cpu_map[tci]);
2220 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2223 /* copy maps belonging to foreign traffic classes */
2224 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2225 /* fill in the new device map from the old device map */
2226 map = xmap_dereference(dev_maps->cpu_map[tci]);
2227 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2231 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2233 /* Cleanup old maps */
2234 if (!dev_maps)
2235 goto out_no_old_maps;
2237 for_each_possible_cpu(cpu) {
2238 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2239 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2240 map = xmap_dereference(dev_maps->cpu_map[tci]);
2241 if (map && map != new_map)
2242 kfree_rcu(map, rcu);
2246 kfree_rcu(dev_maps, rcu);
2248 out_no_old_maps:
2249 dev_maps = new_dev_maps;
2250 active = true;
2252 out_no_new_maps:
2253 /* update Tx queue numa node */
2254 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2255 (numa_node_id >= 0) ? numa_node_id :
2256 NUMA_NO_NODE);
2258 if (!dev_maps)
2259 goto out_no_maps;
2261 /* removes queue from unused CPUs */
2262 for_each_possible_cpu(cpu) {
2263 for (i = tc, tci = cpu * num_tc; i--; tci++)
2264 active |= remove_xps_queue(dev_maps, tci, index);
2265 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2266 active |= remove_xps_queue(dev_maps, tci, index);
2267 for (i = num_tc - tc, tci++; --i; tci++)
2268 active |= remove_xps_queue(dev_maps, tci, index);
2271 /* free map if not active */
2272 if (!active) {
2273 RCU_INIT_POINTER(dev->xps_maps, NULL);
2274 kfree_rcu(dev_maps, rcu);
2277 out_no_maps:
2278 mutex_unlock(&xps_map_mutex);
2280 return 0;
2281 error:
2282 /* remove any maps that we added */
2283 for_each_possible_cpu(cpu) {
2284 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2285 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2286 map = dev_maps ?
2287 xmap_dereference(dev_maps->cpu_map[tci]) :
2288 NULL;
2289 if (new_map && new_map != map)
2290 kfree(new_map);
2294 mutex_unlock(&xps_map_mutex);
2296 kfree(new_dev_maps);
2297 return -ENOMEM;
2299 EXPORT_SYMBOL(netif_set_xps_queue);
2301 #endif
2302 void netdev_reset_tc(struct net_device *dev)
2304 #ifdef CONFIG_XPS
2305 netif_reset_xps_queues_gt(dev, 0);
2306 #endif
2307 dev->num_tc = 0;
2308 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2309 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2311 EXPORT_SYMBOL(netdev_reset_tc);
2313 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2315 if (tc >= dev->num_tc)
2316 return -EINVAL;
2318 #ifdef CONFIG_XPS
2319 netif_reset_xps_queues(dev, offset, count);
2320 #endif
2321 dev->tc_to_txq[tc].count = count;
2322 dev->tc_to_txq[tc].offset = offset;
2323 return 0;
2325 EXPORT_SYMBOL(netdev_set_tc_queue);
2327 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2329 if (num_tc > TC_MAX_QUEUE)
2330 return -EINVAL;
2332 #ifdef CONFIG_XPS
2333 netif_reset_xps_queues_gt(dev, 0);
2334 #endif
2335 dev->num_tc = num_tc;
2336 return 0;
2338 EXPORT_SYMBOL(netdev_set_num_tc);
2341 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2342 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
2344 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2346 int rc;
2348 if (txq < 1 || txq > dev->num_tx_queues)
2349 return -EINVAL;
2351 if (dev->reg_state == NETREG_REGISTERED ||
2352 dev->reg_state == NETREG_UNREGISTERING) {
2353 ASSERT_RTNL();
2355 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2356 txq);
2357 if (rc)
2358 return rc;
2360 if (dev->num_tc)
2361 netif_setup_tc(dev, txq);
2363 if (txq < dev->real_num_tx_queues) {
2364 qdisc_reset_all_tx_gt(dev, txq);
2365 #ifdef CONFIG_XPS
2366 netif_reset_xps_queues_gt(dev, txq);
2367 #endif
2371 dev->real_num_tx_queues = txq;
2372 return 0;
2374 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2376 #ifdef CONFIG_SYSFS
2378 * netif_set_real_num_rx_queues - set actual number of RX queues used
2379 * @dev: Network device
2380 * @rxq: Actual number of RX queues
2382 * This must be called either with the rtnl_lock held or before
2383 * registration of the net device. Returns 0 on success, or a
2384 * negative error code. If called before registration, it always
2385 * succeeds.
2387 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2389 int rc;
2391 if (rxq < 1 || rxq > dev->num_rx_queues)
2392 return -EINVAL;
2394 if (dev->reg_state == NETREG_REGISTERED) {
2395 ASSERT_RTNL();
2397 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2398 rxq);
2399 if (rc)
2400 return rc;
2403 dev->real_num_rx_queues = rxq;
2404 return 0;
2406 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2407 #endif
2410 * netif_get_num_default_rss_queues - default number of RSS queues
2412 * This routine should set an upper limit on the number of RSS queues
2413 * used by default by multiqueue devices.
2415 int netif_get_num_default_rss_queues(void)
2417 return is_kdump_kernel() ?
2418 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2420 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2422 static void __netif_reschedule(struct Qdisc *q)
2424 struct softnet_data *sd;
2425 unsigned long flags;
2427 local_irq_save(flags);
2428 sd = this_cpu_ptr(&softnet_data);
2429 q->next_sched = NULL;
2430 *sd->output_queue_tailp = q;
2431 sd->output_queue_tailp = &q->next_sched;
2432 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2433 local_irq_restore(flags);
2436 void __netif_schedule(struct Qdisc *q)
2438 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2439 __netif_reschedule(q);
2441 EXPORT_SYMBOL(__netif_schedule);
2443 struct dev_kfree_skb_cb {
2444 enum skb_free_reason reason;
2447 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2449 return (struct dev_kfree_skb_cb *)skb->cb;
2452 void netif_schedule_queue(struct netdev_queue *txq)
2454 rcu_read_lock();
2455 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2456 struct Qdisc *q = rcu_dereference(txq->qdisc);
2458 __netif_schedule(q);
2460 rcu_read_unlock();
2462 EXPORT_SYMBOL(netif_schedule_queue);
2464 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2466 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2467 struct Qdisc *q;
2469 rcu_read_lock();
2470 q = rcu_dereference(dev_queue->qdisc);
2471 __netif_schedule(q);
2472 rcu_read_unlock();
2475 EXPORT_SYMBOL(netif_tx_wake_queue);
2477 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2479 unsigned long flags;
2481 if (unlikely(!skb))
2482 return;
2484 if (likely(refcount_read(&skb->users) == 1)) {
2485 smp_rmb();
2486 refcount_set(&skb->users, 0);
2487 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2488 return;
2490 get_kfree_skb_cb(skb)->reason = reason;
2491 local_irq_save(flags);
2492 skb->next = __this_cpu_read(softnet_data.completion_queue);
2493 __this_cpu_write(softnet_data.completion_queue, skb);
2494 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2495 local_irq_restore(flags);
2497 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2499 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2501 if (in_irq() || irqs_disabled())
2502 __dev_kfree_skb_irq(skb, reason);
2503 else
2504 dev_kfree_skb(skb);
2506 EXPORT_SYMBOL(__dev_kfree_skb_any);
2510 * netif_device_detach - mark device as removed
2511 * @dev: network device
2513 * Mark device as removed from system and therefore no longer available.
2515 void netif_device_detach(struct net_device *dev)
2517 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2518 netif_running(dev)) {
2519 netif_tx_stop_all_queues(dev);
2522 EXPORT_SYMBOL(netif_device_detach);
2525 * netif_device_attach - mark device as attached
2526 * @dev: network device
2528 * Mark device as attached from system and restart if needed.
2530 void netif_device_attach(struct net_device *dev)
2532 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2533 netif_running(dev)) {
2534 netif_tx_wake_all_queues(dev);
2535 __netdev_watchdog_up(dev);
2538 EXPORT_SYMBOL(netif_device_attach);
2541 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2542 * to be used as a distribution range.
2544 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2545 unsigned int num_tx_queues)
2547 u32 hash;
2548 u16 qoffset = 0;
2549 u16 qcount = num_tx_queues;
2551 if (skb_rx_queue_recorded(skb)) {
2552 hash = skb_get_rx_queue(skb);
2553 while (unlikely(hash >= num_tx_queues))
2554 hash -= num_tx_queues;
2555 return hash;
2558 if (dev->num_tc) {
2559 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2561 qoffset = dev->tc_to_txq[tc].offset;
2562 qcount = dev->tc_to_txq[tc].count;
2565 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2567 EXPORT_SYMBOL(__skb_tx_hash);
2569 static void skb_warn_bad_offload(const struct sk_buff *skb)
2571 static const netdev_features_t null_features;
2572 struct net_device *dev = skb->dev;
2573 const char *name = "";
2575 if (!net_ratelimit())
2576 return;
2578 if (dev) {
2579 if (dev->dev.parent)
2580 name = dev_driver_string(dev->dev.parent);
2581 else
2582 name = netdev_name(dev);
2584 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2585 "gso_type=%d ip_summed=%d\n",
2586 name, dev ? &dev->features : &null_features,
2587 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2588 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2589 skb_shinfo(skb)->gso_type, skb->ip_summed);
2593 * Invalidate hardware checksum when packet is to be mangled, and
2594 * complete checksum manually on outgoing path.
2596 int skb_checksum_help(struct sk_buff *skb)
2598 __wsum csum;
2599 int ret = 0, offset;
2601 if (skb->ip_summed == CHECKSUM_COMPLETE)
2602 goto out_set_summed;
2604 if (unlikely(skb_shinfo(skb)->gso_size)) {
2605 skb_warn_bad_offload(skb);
2606 return -EINVAL;
2609 /* Before computing a checksum, we should make sure no frag could
2610 * be modified by an external entity : checksum could be wrong.
2612 if (skb_has_shared_frag(skb)) {
2613 ret = __skb_linearize(skb);
2614 if (ret)
2615 goto out;
2618 offset = skb_checksum_start_offset(skb);
2619 BUG_ON(offset >= skb_headlen(skb));
2620 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2622 offset += skb->csum_offset;
2623 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2625 if (skb_cloned(skb) &&
2626 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2627 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2628 if (ret)
2629 goto out;
2632 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2633 out_set_summed:
2634 skb->ip_summed = CHECKSUM_NONE;
2635 out:
2636 return ret;
2638 EXPORT_SYMBOL(skb_checksum_help);
2640 int skb_crc32c_csum_help(struct sk_buff *skb)
2642 __le32 crc32c_csum;
2643 int ret = 0, offset, start;
2645 if (skb->ip_summed != CHECKSUM_PARTIAL)
2646 goto out;
2648 if (unlikely(skb_is_gso(skb)))
2649 goto out;
2651 /* Before computing a checksum, we should make sure no frag could
2652 * be modified by an external entity : checksum could be wrong.
2654 if (unlikely(skb_has_shared_frag(skb))) {
2655 ret = __skb_linearize(skb);
2656 if (ret)
2657 goto out;
2659 start = skb_checksum_start_offset(skb);
2660 offset = start + offsetof(struct sctphdr, checksum);
2661 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2662 ret = -EINVAL;
2663 goto out;
2665 if (skb_cloned(skb) &&
2666 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2667 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2668 if (ret)
2669 goto out;
2671 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2672 skb->len - start, ~(__u32)0,
2673 crc32c_csum_stub));
2674 *(__le32 *)(skb->data + offset) = crc32c_csum;
2675 skb->ip_summed = CHECKSUM_NONE;
2676 skb->csum_not_inet = 0;
2677 out:
2678 return ret;
2681 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2683 __be16 type = skb->protocol;
2685 /* Tunnel gso handlers can set protocol to ethernet. */
2686 if (type == htons(ETH_P_TEB)) {
2687 struct ethhdr *eth;
2689 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2690 return 0;
2692 eth = (struct ethhdr *)skb_mac_header(skb);
2693 type = eth->h_proto;
2696 return __vlan_get_protocol(skb, type, depth);
2700 * skb_mac_gso_segment - mac layer segmentation handler.
2701 * @skb: buffer to segment
2702 * @features: features for the output path (see dev->features)
2704 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2705 netdev_features_t features)
2707 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2708 struct packet_offload *ptype;
2709 int vlan_depth = skb->mac_len;
2710 __be16 type = skb_network_protocol(skb, &vlan_depth);
2712 if (unlikely(!type))
2713 return ERR_PTR(-EINVAL);
2715 __skb_pull(skb, vlan_depth);
2717 rcu_read_lock();
2718 list_for_each_entry_rcu(ptype, &offload_base, list) {
2719 if (ptype->type == type && ptype->callbacks.gso_segment) {
2720 segs = ptype->callbacks.gso_segment(skb, features);
2721 break;
2724 rcu_read_unlock();
2726 __skb_push(skb, skb->data - skb_mac_header(skb));
2728 return segs;
2730 EXPORT_SYMBOL(skb_mac_gso_segment);
2733 /* openvswitch calls this on rx path, so we need a different check.
2735 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2737 if (tx_path)
2738 return skb->ip_summed != CHECKSUM_PARTIAL;
2740 return skb->ip_summed == CHECKSUM_NONE;
2744 * __skb_gso_segment - Perform segmentation on skb.
2745 * @skb: buffer to segment
2746 * @features: features for the output path (see dev->features)
2747 * @tx_path: whether it is called in TX path
2749 * This function segments the given skb and returns a list of segments.
2751 * It may return NULL if the skb requires no segmentation. This is
2752 * only possible when GSO is used for verifying header integrity.
2754 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2756 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2757 netdev_features_t features, bool tx_path)
2759 struct sk_buff *segs;
2761 if (unlikely(skb_needs_check(skb, tx_path))) {
2762 int err;
2764 /* We're going to init ->check field in TCP or UDP header */
2765 err = skb_cow_head(skb, 0);
2766 if (err < 0)
2767 return ERR_PTR(err);
2770 /* Only report GSO partial support if it will enable us to
2771 * support segmentation on this frame without needing additional
2772 * work.
2774 if (features & NETIF_F_GSO_PARTIAL) {
2775 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2776 struct net_device *dev = skb->dev;
2778 partial_features |= dev->features & dev->gso_partial_features;
2779 if (!skb_gso_ok(skb, features | partial_features))
2780 features &= ~NETIF_F_GSO_PARTIAL;
2783 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2784 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2786 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2787 SKB_GSO_CB(skb)->encap_level = 0;
2789 skb_reset_mac_header(skb);
2790 skb_reset_mac_len(skb);
2792 segs = skb_mac_gso_segment(skb, features);
2794 if (unlikely(skb_needs_check(skb, tx_path)))
2795 skb_warn_bad_offload(skb);
2797 return segs;
2799 EXPORT_SYMBOL(__skb_gso_segment);
2801 /* Take action when hardware reception checksum errors are detected. */
2802 #ifdef CONFIG_BUG
2803 void netdev_rx_csum_fault(struct net_device *dev)
2805 if (net_ratelimit()) {
2806 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2807 dump_stack();
2810 EXPORT_SYMBOL(netdev_rx_csum_fault);
2811 #endif
2813 /* Actually, we should eliminate this check as soon as we know, that:
2814 * 1. IOMMU is present and allows to map all the memory.
2815 * 2. No high memory really exists on this machine.
2818 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2820 #ifdef CONFIG_HIGHMEM
2821 int i;
2823 if (!(dev->features & NETIF_F_HIGHDMA)) {
2824 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2825 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2827 if (PageHighMem(skb_frag_page(frag)))
2828 return 1;
2832 if (PCI_DMA_BUS_IS_PHYS) {
2833 struct device *pdev = dev->dev.parent;
2835 if (!pdev)
2836 return 0;
2837 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2838 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2839 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2841 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2842 return 1;
2845 #endif
2846 return 0;
2849 /* If MPLS offload request, verify we are testing hardware MPLS features
2850 * instead of standard features for the netdev.
2852 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2853 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2854 netdev_features_t features,
2855 __be16 type)
2857 if (eth_p_mpls(type))
2858 features &= skb->dev->mpls_features;
2860 return features;
2862 #else
2863 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2864 netdev_features_t features,
2865 __be16 type)
2867 return features;
2869 #endif
2871 static netdev_features_t harmonize_features(struct sk_buff *skb,
2872 netdev_features_t features)
2874 int tmp;
2875 __be16 type;
2877 type = skb_network_protocol(skb, &tmp);
2878 features = net_mpls_features(skb, features, type);
2880 if (skb->ip_summed != CHECKSUM_NONE &&
2881 !can_checksum_protocol(features, type)) {
2882 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2884 if (illegal_highdma(skb->dev, skb))
2885 features &= ~NETIF_F_SG;
2887 return features;
2890 netdev_features_t passthru_features_check(struct sk_buff *skb,
2891 struct net_device *dev,
2892 netdev_features_t features)
2894 return features;
2896 EXPORT_SYMBOL(passthru_features_check);
2898 static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2899 struct net_device *dev,
2900 netdev_features_t features)
2902 return vlan_features_check(skb, features);
2905 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2906 struct net_device *dev,
2907 netdev_features_t features)
2909 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2911 if (gso_segs > dev->gso_max_segs)
2912 return features & ~NETIF_F_GSO_MASK;
2914 /* Support for GSO partial features requires software
2915 * intervention before we can actually process the packets
2916 * so we need to strip support for any partial features now
2917 * and we can pull them back in after we have partially
2918 * segmented the frame.
2920 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2921 features &= ~dev->gso_partial_features;
2923 /* Make sure to clear the IPv4 ID mangling feature if the
2924 * IPv4 header has the potential to be fragmented.
2926 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2927 struct iphdr *iph = skb->encapsulation ?
2928 inner_ip_hdr(skb) : ip_hdr(skb);
2930 if (!(iph->frag_off & htons(IP_DF)))
2931 features &= ~NETIF_F_TSO_MANGLEID;
2934 return features;
2937 netdev_features_t netif_skb_features(struct sk_buff *skb)
2939 struct net_device *dev = skb->dev;
2940 netdev_features_t features = dev->features;
2942 if (skb_is_gso(skb))
2943 features = gso_features_check(skb, dev, features);
2945 /* If encapsulation offload request, verify we are testing
2946 * hardware encapsulation features instead of standard
2947 * features for the netdev
2949 if (skb->encapsulation)
2950 features &= dev->hw_enc_features;
2952 if (skb_vlan_tagged(skb))
2953 features = netdev_intersect_features(features,
2954 dev->vlan_features |
2955 NETIF_F_HW_VLAN_CTAG_TX |
2956 NETIF_F_HW_VLAN_STAG_TX);
2958 if (dev->netdev_ops->ndo_features_check)
2959 features &= dev->netdev_ops->ndo_features_check(skb, dev,
2960 features);
2961 else
2962 features &= dflt_features_check(skb, dev, features);
2964 return harmonize_features(skb, features);
2966 EXPORT_SYMBOL(netif_skb_features);
2968 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
2969 struct netdev_queue *txq, bool more)
2971 unsigned int len;
2972 int rc;
2974 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
2975 dev_queue_xmit_nit(skb, dev);
2977 len = skb->len;
2978 trace_net_dev_start_xmit(skb, dev);
2979 rc = netdev_start_xmit(skb, dev, txq, more);
2980 trace_net_dev_xmit(skb, rc, dev, len);
2982 return rc;
2985 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
2986 struct netdev_queue *txq, int *ret)
2988 struct sk_buff *skb = first;
2989 int rc = NETDEV_TX_OK;
2991 while (skb) {
2992 struct sk_buff *next = skb->next;
2994 skb->next = NULL;
2995 rc = xmit_one(skb, dev, txq, next != NULL);
2996 if (unlikely(!dev_xmit_complete(rc))) {
2997 skb->next = next;
2998 goto out;
3001 skb = next;
3002 if (netif_xmit_stopped(txq) && skb) {
3003 rc = NETDEV_TX_BUSY;
3004 break;
3008 out:
3009 *ret = rc;
3010 return skb;
3013 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3014 netdev_features_t features)
3016 if (skb_vlan_tag_present(skb) &&
3017 !vlan_hw_offload_capable(features, skb->vlan_proto))
3018 skb = __vlan_hwaccel_push_inside(skb);
3019 return skb;
3022 int skb_csum_hwoffload_help(struct sk_buff *skb,
3023 const netdev_features_t features)
3025 if (unlikely(skb->csum_not_inet))
3026 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3027 skb_crc32c_csum_help(skb);
3029 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3031 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3033 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev)
3035 netdev_features_t features;
3037 features = netif_skb_features(skb);
3038 skb = validate_xmit_vlan(skb, features);
3039 if (unlikely(!skb))
3040 goto out_null;
3042 if (netif_needs_gso(skb, features)) {
3043 struct sk_buff *segs;
3045 segs = skb_gso_segment(skb, features);
3046 if (IS_ERR(segs)) {
3047 goto out_kfree_skb;
3048 } else if (segs) {
3049 consume_skb(skb);
3050 skb = segs;
3052 } else {
3053 if (skb_needs_linearize(skb, features) &&
3054 __skb_linearize(skb))
3055 goto out_kfree_skb;
3057 if (validate_xmit_xfrm(skb, features))
3058 goto out_kfree_skb;
3060 /* If packet is not checksummed and device does not
3061 * support checksumming for this protocol, complete
3062 * checksumming here.
3064 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3065 if (skb->encapsulation)
3066 skb_set_inner_transport_header(skb,
3067 skb_checksum_start_offset(skb));
3068 else
3069 skb_set_transport_header(skb,
3070 skb_checksum_start_offset(skb));
3071 if (skb_csum_hwoffload_help(skb, features))
3072 goto out_kfree_skb;
3076 return skb;
3078 out_kfree_skb:
3079 kfree_skb(skb);
3080 out_null:
3081 atomic_long_inc(&dev->tx_dropped);
3082 return NULL;
3085 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev)
3087 struct sk_buff *next, *head = NULL, *tail;
3089 for (; skb != NULL; skb = next) {
3090 next = skb->next;
3091 skb->next = NULL;
3093 /* in case skb wont be segmented, point to itself */
3094 skb->prev = skb;
3096 skb = validate_xmit_skb(skb, dev);
3097 if (!skb)
3098 continue;
3100 if (!head)
3101 head = skb;
3102 else
3103 tail->next = skb;
3104 /* If skb was segmented, skb->prev points to
3105 * the last segment. If not, it still contains skb.
3107 tail = skb->prev;
3109 return head;
3111 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3113 static void qdisc_pkt_len_init(struct sk_buff *skb)
3115 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3117 qdisc_skb_cb(skb)->pkt_len = skb->len;
3119 /* To get more precise estimation of bytes sent on wire,
3120 * we add to pkt_len the headers size of all segments
3122 if (shinfo->gso_size) {
3123 unsigned int hdr_len;
3124 u16 gso_segs = shinfo->gso_segs;
3126 /* mac layer + network layer */
3127 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3129 /* + transport layer */
3130 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
3131 hdr_len += tcp_hdrlen(skb);
3132 else
3133 hdr_len += sizeof(struct udphdr);
3135 if (shinfo->gso_type & SKB_GSO_DODGY)
3136 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3137 shinfo->gso_size);
3139 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3143 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3144 struct net_device *dev,
3145 struct netdev_queue *txq)
3147 spinlock_t *root_lock = qdisc_lock(q);
3148 struct sk_buff *to_free = NULL;
3149 bool contended;
3150 int rc;
3152 qdisc_calculate_pkt_len(skb, q);
3154 * Heuristic to force contended enqueues to serialize on a
3155 * separate lock before trying to get qdisc main lock.
3156 * This permits qdisc->running owner to get the lock more
3157 * often and dequeue packets faster.
3159 contended = qdisc_is_running(q);
3160 if (unlikely(contended))
3161 spin_lock(&q->busylock);
3163 spin_lock(root_lock);
3164 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3165 __qdisc_drop(skb, &to_free);
3166 rc = NET_XMIT_DROP;
3167 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3168 qdisc_run_begin(q)) {
3170 * This is a work-conserving queue; there are no old skbs
3171 * waiting to be sent out; and the qdisc is not running -
3172 * xmit the skb directly.
3175 qdisc_bstats_update(q, skb);
3177 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3178 if (unlikely(contended)) {
3179 spin_unlock(&q->busylock);
3180 contended = false;
3182 __qdisc_run(q);
3183 } else
3184 qdisc_run_end(q);
3186 rc = NET_XMIT_SUCCESS;
3187 } else {
3188 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3189 if (qdisc_run_begin(q)) {
3190 if (unlikely(contended)) {
3191 spin_unlock(&q->busylock);
3192 contended = false;
3194 __qdisc_run(q);
3197 spin_unlock(root_lock);
3198 if (unlikely(to_free))
3199 kfree_skb_list(to_free);
3200 if (unlikely(contended))
3201 spin_unlock(&q->busylock);
3202 return rc;
3205 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3206 static void skb_update_prio(struct sk_buff *skb)
3208 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
3210 if (!skb->priority && skb->sk && map) {
3211 unsigned int prioidx =
3212 sock_cgroup_prioidx(&skb->sk->sk_cgrp_data);
3214 if (prioidx < map->priomap_len)
3215 skb->priority = map->priomap[prioidx];
3218 #else
3219 #define skb_update_prio(skb)
3220 #endif
3222 DEFINE_PER_CPU(int, xmit_recursion);
3223 EXPORT_SYMBOL(xmit_recursion);
3226 * dev_loopback_xmit - loop back @skb
3227 * @net: network namespace this loopback is happening in
3228 * @sk: sk needed to be a netfilter okfn
3229 * @skb: buffer to transmit
3231 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3233 skb_reset_mac_header(skb);
3234 __skb_pull(skb, skb_network_offset(skb));
3235 skb->pkt_type = PACKET_LOOPBACK;
3236 skb->ip_summed = CHECKSUM_UNNECESSARY;
3237 WARN_ON(!skb_dst(skb));
3238 skb_dst_force(skb);
3239 netif_rx_ni(skb);
3240 return 0;
3242 EXPORT_SYMBOL(dev_loopback_xmit);
3244 #ifdef CONFIG_NET_EGRESS
3245 static struct sk_buff *
3246 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3248 struct tcf_proto *cl = rcu_dereference_bh(dev->egress_cl_list);
3249 struct tcf_result cl_res;
3251 if (!cl)
3252 return skb;
3254 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3255 qdisc_bstats_cpu_update(cl->q, skb);
3257 switch (tcf_classify(skb, cl, &cl_res, false)) {
3258 case TC_ACT_OK:
3259 case TC_ACT_RECLASSIFY:
3260 skb->tc_index = TC_H_MIN(cl_res.classid);
3261 break;
3262 case TC_ACT_SHOT:
3263 qdisc_qstats_cpu_drop(cl->q);
3264 *ret = NET_XMIT_DROP;
3265 kfree_skb(skb);
3266 return NULL;
3267 case TC_ACT_STOLEN:
3268 case TC_ACT_QUEUED:
3269 case TC_ACT_TRAP:
3270 *ret = NET_XMIT_SUCCESS;
3271 consume_skb(skb);
3272 return NULL;
3273 case TC_ACT_REDIRECT:
3274 /* No need to push/pop skb's mac_header here on egress! */
3275 skb_do_redirect(skb);
3276 *ret = NET_XMIT_SUCCESS;
3277 return NULL;
3278 default:
3279 break;
3282 return skb;
3284 #endif /* CONFIG_NET_EGRESS */
3286 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3288 #ifdef CONFIG_XPS
3289 struct xps_dev_maps *dev_maps;
3290 struct xps_map *map;
3291 int queue_index = -1;
3293 rcu_read_lock();
3294 dev_maps = rcu_dereference(dev->xps_maps);
3295 if (dev_maps) {
3296 unsigned int tci = skb->sender_cpu - 1;
3298 if (dev->num_tc) {
3299 tci *= dev->num_tc;
3300 tci += netdev_get_prio_tc_map(dev, skb->priority);
3303 map = rcu_dereference(dev_maps->cpu_map[tci]);
3304 if (map) {
3305 if (map->len == 1)
3306 queue_index = map->queues[0];
3307 else
3308 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3309 map->len)];
3310 if (unlikely(queue_index >= dev->real_num_tx_queues))
3311 queue_index = -1;
3314 rcu_read_unlock();
3316 return queue_index;
3317 #else
3318 return -1;
3319 #endif
3322 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3324 struct sock *sk = skb->sk;
3325 int queue_index = sk_tx_queue_get(sk);
3327 if (queue_index < 0 || skb->ooo_okay ||
3328 queue_index >= dev->real_num_tx_queues) {
3329 int new_index = get_xps_queue(dev, skb);
3331 if (new_index < 0)
3332 new_index = skb_tx_hash(dev, skb);
3334 if (queue_index != new_index && sk &&
3335 sk_fullsock(sk) &&
3336 rcu_access_pointer(sk->sk_dst_cache))
3337 sk_tx_queue_set(sk, new_index);
3339 queue_index = new_index;
3342 return queue_index;
3345 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3346 struct sk_buff *skb,
3347 void *accel_priv)
3349 int queue_index = 0;
3351 #ifdef CONFIG_XPS
3352 u32 sender_cpu = skb->sender_cpu - 1;
3354 if (sender_cpu >= (u32)NR_CPUS)
3355 skb->sender_cpu = raw_smp_processor_id() + 1;
3356 #endif
3358 if (dev->real_num_tx_queues != 1) {
3359 const struct net_device_ops *ops = dev->netdev_ops;
3361 if (ops->ndo_select_queue)
3362 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3363 __netdev_pick_tx);
3364 else
3365 queue_index = __netdev_pick_tx(dev, skb);
3367 if (!accel_priv)
3368 queue_index = netdev_cap_txqueue(dev, queue_index);
3371 skb_set_queue_mapping(skb, queue_index);
3372 return netdev_get_tx_queue(dev, queue_index);
3376 * __dev_queue_xmit - transmit a buffer
3377 * @skb: buffer to transmit
3378 * @accel_priv: private data used for L2 forwarding offload
3380 * Queue a buffer for transmission to a network device. The caller must
3381 * have set the device and priority and built the buffer before calling
3382 * this function. The function can be called from an interrupt.
3384 * A negative errno code is returned on a failure. A success does not
3385 * guarantee the frame will be transmitted as it may be dropped due
3386 * to congestion or traffic shaping.
3388 * -----------------------------------------------------------------------------------
3389 * I notice this method can also return errors from the queue disciplines,
3390 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3391 * be positive.
3393 * Regardless of the return value, the skb is consumed, so it is currently
3394 * difficult to retry a send to this method. (You can bump the ref count
3395 * before sending to hold a reference for retry if you are careful.)
3397 * When calling this method, interrupts MUST be enabled. This is because
3398 * the BH enable code must have IRQs enabled so that it will not deadlock.
3399 * --BLG
3401 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3403 struct net_device *dev = skb->dev;
3404 struct netdev_queue *txq;
3405 struct Qdisc *q;
3406 int rc = -ENOMEM;
3408 skb_reset_mac_header(skb);
3410 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3411 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3413 /* Disable soft irqs for various locks below. Also
3414 * stops preemption for RCU.
3416 rcu_read_lock_bh();
3418 skb_update_prio(skb);
3420 qdisc_pkt_len_init(skb);
3421 #ifdef CONFIG_NET_CLS_ACT
3422 skb->tc_at_ingress = 0;
3423 # ifdef CONFIG_NET_EGRESS
3424 if (static_key_false(&egress_needed)) {
3425 skb = sch_handle_egress(skb, &rc, dev);
3426 if (!skb)
3427 goto out;
3429 # endif
3430 #endif
3431 /* If device/qdisc don't need skb->dst, release it right now while
3432 * its hot in this cpu cache.
3434 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3435 skb_dst_drop(skb);
3436 else
3437 skb_dst_force(skb);
3439 txq = netdev_pick_tx(dev, skb, accel_priv);
3440 q = rcu_dereference_bh(txq->qdisc);
3442 trace_net_dev_queue(skb);
3443 if (q->enqueue) {
3444 rc = __dev_xmit_skb(skb, q, dev, txq);
3445 goto out;
3448 /* The device has no queue. Common case for software devices:
3449 * loopback, all the sorts of tunnels...
3451 * Really, it is unlikely that netif_tx_lock protection is necessary
3452 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3453 * counters.)
3454 * However, it is possible, that they rely on protection
3455 * made by us here.
3457 * Check this and shot the lock. It is not prone from deadlocks.
3458 *Either shot noqueue qdisc, it is even simpler 8)
3460 if (dev->flags & IFF_UP) {
3461 int cpu = smp_processor_id(); /* ok because BHs are off */
3463 if (txq->xmit_lock_owner != cpu) {
3464 if (unlikely(__this_cpu_read(xmit_recursion) >
3465 XMIT_RECURSION_LIMIT))
3466 goto recursion_alert;
3468 skb = validate_xmit_skb(skb, dev);
3469 if (!skb)
3470 goto out;
3472 HARD_TX_LOCK(dev, txq, cpu);
3474 if (!netif_xmit_stopped(txq)) {
3475 __this_cpu_inc(xmit_recursion);
3476 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3477 __this_cpu_dec(xmit_recursion);
3478 if (dev_xmit_complete(rc)) {
3479 HARD_TX_UNLOCK(dev, txq);
3480 goto out;
3483 HARD_TX_UNLOCK(dev, txq);
3484 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3485 dev->name);
3486 } else {
3487 /* Recursion is detected! It is possible,
3488 * unfortunately
3490 recursion_alert:
3491 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3492 dev->name);
3496 rc = -ENETDOWN;
3497 rcu_read_unlock_bh();
3499 atomic_long_inc(&dev->tx_dropped);
3500 kfree_skb_list(skb);
3501 return rc;
3502 out:
3503 rcu_read_unlock_bh();
3504 return rc;
3507 int dev_queue_xmit(struct sk_buff *skb)
3509 return __dev_queue_xmit(skb, NULL);
3511 EXPORT_SYMBOL(dev_queue_xmit);
3513 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3515 return __dev_queue_xmit(skb, accel_priv);
3517 EXPORT_SYMBOL(dev_queue_xmit_accel);
3520 /*************************************************************************
3521 * Receiver routines
3522 *************************************************************************/
3524 int netdev_max_backlog __read_mostly = 1000;
3525 EXPORT_SYMBOL(netdev_max_backlog);
3527 int netdev_tstamp_prequeue __read_mostly = 1;
3528 int netdev_budget __read_mostly = 300;
3529 unsigned int __read_mostly netdev_budget_usecs = 2000;
3530 int weight_p __read_mostly = 64; /* old backlog weight */
3531 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3532 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3533 int dev_rx_weight __read_mostly = 64;
3534 int dev_tx_weight __read_mostly = 64;
3536 /* Called with irq disabled */
3537 static inline void ____napi_schedule(struct softnet_data *sd,
3538 struct napi_struct *napi)
3540 list_add_tail(&napi->poll_list, &sd->poll_list);
3541 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3544 #ifdef CONFIG_RPS
3546 /* One global table that all flow-based protocols share. */
3547 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3548 EXPORT_SYMBOL(rps_sock_flow_table);
3549 u32 rps_cpu_mask __read_mostly;
3550 EXPORT_SYMBOL(rps_cpu_mask);
3552 struct static_key rps_needed __read_mostly;
3553 EXPORT_SYMBOL(rps_needed);
3554 struct static_key rfs_needed __read_mostly;
3555 EXPORT_SYMBOL(rfs_needed);
3557 static struct rps_dev_flow *
3558 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3559 struct rps_dev_flow *rflow, u16 next_cpu)
3561 if (next_cpu < nr_cpu_ids) {
3562 #ifdef CONFIG_RFS_ACCEL
3563 struct netdev_rx_queue *rxqueue;
3564 struct rps_dev_flow_table *flow_table;
3565 struct rps_dev_flow *old_rflow;
3566 u32 flow_id;
3567 u16 rxq_index;
3568 int rc;
3570 /* Should we steer this flow to a different hardware queue? */
3571 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3572 !(dev->features & NETIF_F_NTUPLE))
3573 goto out;
3574 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3575 if (rxq_index == skb_get_rx_queue(skb))
3576 goto out;
3578 rxqueue = dev->_rx + rxq_index;
3579 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3580 if (!flow_table)
3581 goto out;
3582 flow_id = skb_get_hash(skb) & flow_table->mask;
3583 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3584 rxq_index, flow_id);
3585 if (rc < 0)
3586 goto out;
3587 old_rflow = rflow;
3588 rflow = &flow_table->flows[flow_id];
3589 rflow->filter = rc;
3590 if (old_rflow->filter == rflow->filter)
3591 old_rflow->filter = RPS_NO_FILTER;
3592 out:
3593 #endif
3594 rflow->last_qtail =
3595 per_cpu(softnet_data, next_cpu).input_queue_head;
3598 rflow->cpu = next_cpu;
3599 return rflow;
3603 * get_rps_cpu is called from netif_receive_skb and returns the target
3604 * CPU from the RPS map of the receiving queue for a given skb.
3605 * rcu_read_lock must be held on entry.
3607 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3608 struct rps_dev_flow **rflowp)
3610 const struct rps_sock_flow_table *sock_flow_table;
3611 struct netdev_rx_queue *rxqueue = dev->_rx;
3612 struct rps_dev_flow_table *flow_table;
3613 struct rps_map *map;
3614 int cpu = -1;
3615 u32 tcpu;
3616 u32 hash;
3618 if (skb_rx_queue_recorded(skb)) {
3619 u16 index = skb_get_rx_queue(skb);
3621 if (unlikely(index >= dev->real_num_rx_queues)) {
3622 WARN_ONCE(dev->real_num_rx_queues > 1,
3623 "%s received packet on queue %u, but number "
3624 "of RX queues is %u\n",
3625 dev->name, index, dev->real_num_rx_queues);
3626 goto done;
3628 rxqueue += index;
3631 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3633 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3634 map = rcu_dereference(rxqueue->rps_map);
3635 if (!flow_table && !map)
3636 goto done;
3638 skb_reset_network_header(skb);
3639 hash = skb_get_hash(skb);
3640 if (!hash)
3641 goto done;
3643 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3644 if (flow_table && sock_flow_table) {
3645 struct rps_dev_flow *rflow;
3646 u32 next_cpu;
3647 u32 ident;
3649 /* First check into global flow table if there is a match */
3650 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3651 if ((ident ^ hash) & ~rps_cpu_mask)
3652 goto try_rps;
3654 next_cpu = ident & rps_cpu_mask;
3656 /* OK, now we know there is a match,
3657 * we can look at the local (per receive queue) flow table
3659 rflow = &flow_table->flows[hash & flow_table->mask];
3660 tcpu = rflow->cpu;
3663 * If the desired CPU (where last recvmsg was done) is
3664 * different from current CPU (one in the rx-queue flow
3665 * table entry), switch if one of the following holds:
3666 * - Current CPU is unset (>= nr_cpu_ids).
3667 * - Current CPU is offline.
3668 * - The current CPU's queue tail has advanced beyond the
3669 * last packet that was enqueued using this table entry.
3670 * This guarantees that all previous packets for the flow
3671 * have been dequeued, thus preserving in order delivery.
3673 if (unlikely(tcpu != next_cpu) &&
3674 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3675 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3676 rflow->last_qtail)) >= 0)) {
3677 tcpu = next_cpu;
3678 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3681 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3682 *rflowp = rflow;
3683 cpu = tcpu;
3684 goto done;
3688 try_rps:
3690 if (map) {
3691 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3692 if (cpu_online(tcpu)) {
3693 cpu = tcpu;
3694 goto done;
3698 done:
3699 return cpu;
3702 #ifdef CONFIG_RFS_ACCEL
3705 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3706 * @dev: Device on which the filter was set
3707 * @rxq_index: RX queue index
3708 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3709 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3711 * Drivers that implement ndo_rx_flow_steer() should periodically call
3712 * this function for each installed filter and remove the filters for
3713 * which it returns %true.
3715 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3716 u32 flow_id, u16 filter_id)
3718 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3719 struct rps_dev_flow_table *flow_table;
3720 struct rps_dev_flow *rflow;
3721 bool expire = true;
3722 unsigned int cpu;
3724 rcu_read_lock();
3725 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3726 if (flow_table && flow_id <= flow_table->mask) {
3727 rflow = &flow_table->flows[flow_id];
3728 cpu = ACCESS_ONCE(rflow->cpu);
3729 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3730 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3731 rflow->last_qtail) <
3732 (int)(10 * flow_table->mask)))
3733 expire = false;
3735 rcu_read_unlock();
3736 return expire;
3738 EXPORT_SYMBOL(rps_may_expire_flow);
3740 #endif /* CONFIG_RFS_ACCEL */
3742 /* Called from hardirq (IPI) context */
3743 static void rps_trigger_softirq(void *data)
3745 struct softnet_data *sd = data;
3747 ____napi_schedule(sd, &sd->backlog);
3748 sd->received_rps++;
3751 #endif /* CONFIG_RPS */
3754 * Check if this softnet_data structure is another cpu one
3755 * If yes, queue it to our IPI list and return 1
3756 * If no, return 0
3758 static int rps_ipi_queued(struct softnet_data *sd)
3760 #ifdef CONFIG_RPS
3761 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3763 if (sd != mysd) {
3764 sd->rps_ipi_next = mysd->rps_ipi_list;
3765 mysd->rps_ipi_list = sd;
3767 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3768 return 1;
3770 #endif /* CONFIG_RPS */
3771 return 0;
3774 #ifdef CONFIG_NET_FLOW_LIMIT
3775 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3776 #endif
3778 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3780 #ifdef CONFIG_NET_FLOW_LIMIT
3781 struct sd_flow_limit *fl;
3782 struct softnet_data *sd;
3783 unsigned int old_flow, new_flow;
3785 if (qlen < (netdev_max_backlog >> 1))
3786 return false;
3788 sd = this_cpu_ptr(&softnet_data);
3790 rcu_read_lock();
3791 fl = rcu_dereference(sd->flow_limit);
3792 if (fl) {
3793 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3794 old_flow = fl->history[fl->history_head];
3795 fl->history[fl->history_head] = new_flow;
3797 fl->history_head++;
3798 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3800 if (likely(fl->buckets[old_flow]))
3801 fl->buckets[old_flow]--;
3803 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3804 fl->count++;
3805 rcu_read_unlock();
3806 return true;
3809 rcu_read_unlock();
3810 #endif
3811 return false;
3815 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3816 * queue (may be a remote CPU queue).
3818 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3819 unsigned int *qtail)
3821 struct softnet_data *sd;
3822 unsigned long flags;
3823 unsigned int qlen;
3825 sd = &per_cpu(softnet_data, cpu);
3827 local_irq_save(flags);
3829 rps_lock(sd);
3830 if (!netif_running(skb->dev))
3831 goto drop;
3832 qlen = skb_queue_len(&sd->input_pkt_queue);
3833 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3834 if (qlen) {
3835 enqueue:
3836 __skb_queue_tail(&sd->input_pkt_queue, skb);
3837 input_queue_tail_incr_save(sd, qtail);
3838 rps_unlock(sd);
3839 local_irq_restore(flags);
3840 return NET_RX_SUCCESS;
3843 /* Schedule NAPI for backlog device
3844 * We can use non atomic operation since we own the queue lock
3846 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3847 if (!rps_ipi_queued(sd))
3848 ____napi_schedule(sd, &sd->backlog);
3850 goto enqueue;
3853 drop:
3854 sd->dropped++;
3855 rps_unlock(sd);
3857 local_irq_restore(flags);
3859 atomic_long_inc(&skb->dev->rx_dropped);
3860 kfree_skb(skb);
3861 return NET_RX_DROP;
3864 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3865 struct bpf_prog *xdp_prog)
3867 struct xdp_buff xdp;
3868 u32 act = XDP_DROP;
3869 void *orig_data;
3870 int hlen, off;
3871 u32 mac_len;
3873 /* Reinjected packets coming from act_mirred or similar should
3874 * not get XDP generic processing.
3876 if (skb_cloned(skb))
3877 return XDP_PASS;
3879 if (skb_linearize(skb))
3880 goto do_drop;
3882 /* The XDP program wants to see the packet starting at the MAC
3883 * header.
3885 mac_len = skb->data - skb_mac_header(skb);
3886 hlen = skb_headlen(skb) + mac_len;
3887 xdp.data = skb->data - mac_len;
3888 xdp.data_end = xdp.data + hlen;
3889 xdp.data_hard_start = skb->data - skb_headroom(skb);
3890 orig_data = xdp.data;
3892 act = bpf_prog_run_xdp(xdp_prog, &xdp);
3894 off = xdp.data - orig_data;
3895 if (off > 0)
3896 __skb_pull(skb, off);
3897 else if (off < 0)
3898 __skb_push(skb, -off);
3899 skb->mac_header += off;
3901 switch (act) {
3902 case XDP_REDIRECT:
3903 case XDP_TX:
3904 __skb_push(skb, mac_len);
3905 /* fall through */
3906 case XDP_PASS:
3907 break;
3909 default:
3910 bpf_warn_invalid_xdp_action(act);
3911 /* fall through */
3912 case XDP_ABORTED:
3913 trace_xdp_exception(skb->dev, xdp_prog, act);
3914 /* fall through */
3915 case XDP_DROP:
3916 do_drop:
3917 kfree_skb(skb);
3918 break;
3921 return act;
3924 /* When doing generic XDP we have to bypass the qdisc layer and the
3925 * network taps in order to match in-driver-XDP behavior.
3927 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
3929 struct net_device *dev = skb->dev;
3930 struct netdev_queue *txq;
3931 bool free_skb = true;
3932 int cpu, rc;
3934 txq = netdev_pick_tx(dev, skb, NULL);
3935 cpu = smp_processor_id();
3936 HARD_TX_LOCK(dev, txq, cpu);
3937 if (!netif_xmit_stopped(txq)) {
3938 rc = netdev_start_xmit(skb, dev, txq, 0);
3939 if (dev_xmit_complete(rc))
3940 free_skb = false;
3942 HARD_TX_UNLOCK(dev, txq);
3943 if (free_skb) {
3944 trace_xdp_exception(dev, xdp_prog, XDP_TX);
3945 kfree_skb(skb);
3948 EXPORT_SYMBOL_GPL(generic_xdp_tx);
3950 static struct static_key generic_xdp_needed __read_mostly;
3952 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
3954 if (xdp_prog) {
3955 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
3956 int err;
3958 if (act != XDP_PASS) {
3959 switch (act) {
3960 case XDP_REDIRECT:
3961 err = xdp_do_generic_redirect(skb->dev, skb,
3962 xdp_prog);
3963 if (err)
3964 goto out_redir;
3965 /* fallthru to submit skb */
3966 case XDP_TX:
3967 generic_xdp_tx(skb, xdp_prog);
3968 break;
3970 return XDP_DROP;
3973 return XDP_PASS;
3974 out_redir:
3975 kfree_skb(skb);
3976 return XDP_DROP;
3978 EXPORT_SYMBOL_GPL(do_xdp_generic);
3980 static int netif_rx_internal(struct sk_buff *skb)
3982 int ret;
3984 net_timestamp_check(netdev_tstamp_prequeue, skb);
3986 trace_netif_rx(skb);
3988 if (static_key_false(&generic_xdp_needed)) {
3989 int ret;
3991 preempt_disable();
3992 rcu_read_lock();
3993 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
3994 rcu_read_unlock();
3995 preempt_enable();
3997 /* Consider XDP consuming the packet a success from
3998 * the netdev point of view we do not want to count
3999 * this as an error.
4001 if (ret != XDP_PASS)
4002 return NET_RX_SUCCESS;
4005 #ifdef CONFIG_RPS
4006 if (static_key_false(&rps_needed)) {
4007 struct rps_dev_flow voidflow, *rflow = &voidflow;
4008 int cpu;
4010 preempt_disable();
4011 rcu_read_lock();
4013 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4014 if (cpu < 0)
4015 cpu = smp_processor_id();
4017 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4019 rcu_read_unlock();
4020 preempt_enable();
4021 } else
4022 #endif
4024 unsigned int qtail;
4026 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4027 put_cpu();
4029 return ret;
4033 * netif_rx - post buffer to the network code
4034 * @skb: buffer to post
4036 * This function receives a packet from a device driver and queues it for
4037 * the upper (protocol) levels to process. It always succeeds. The buffer
4038 * may be dropped during processing for congestion control or by the
4039 * protocol layers.
4041 * return values:
4042 * NET_RX_SUCCESS (no congestion)
4043 * NET_RX_DROP (packet was dropped)
4047 int netif_rx(struct sk_buff *skb)
4049 trace_netif_rx_entry(skb);
4051 return netif_rx_internal(skb);
4053 EXPORT_SYMBOL(netif_rx);
4055 int netif_rx_ni(struct sk_buff *skb)
4057 int err;
4059 trace_netif_rx_ni_entry(skb);
4061 preempt_disable();
4062 err = netif_rx_internal(skb);
4063 if (local_softirq_pending())
4064 do_softirq();
4065 preempt_enable();
4067 return err;
4069 EXPORT_SYMBOL(netif_rx_ni);
4071 static __latent_entropy void net_tx_action(struct softirq_action *h)
4073 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4075 if (sd->completion_queue) {
4076 struct sk_buff *clist;
4078 local_irq_disable();
4079 clist = sd->completion_queue;
4080 sd->completion_queue = NULL;
4081 local_irq_enable();
4083 while (clist) {
4084 struct sk_buff *skb = clist;
4086 clist = clist->next;
4088 WARN_ON(refcount_read(&skb->users));
4089 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4090 trace_consume_skb(skb);
4091 else
4092 trace_kfree_skb(skb, net_tx_action);
4094 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4095 __kfree_skb(skb);
4096 else
4097 __kfree_skb_defer(skb);
4100 __kfree_skb_flush();
4103 if (sd->output_queue) {
4104 struct Qdisc *head;
4106 local_irq_disable();
4107 head = sd->output_queue;
4108 sd->output_queue = NULL;
4109 sd->output_queue_tailp = &sd->output_queue;
4110 local_irq_enable();
4112 while (head) {
4113 struct Qdisc *q = head;
4114 spinlock_t *root_lock;
4116 head = head->next_sched;
4118 root_lock = qdisc_lock(q);
4119 spin_lock(root_lock);
4120 /* We need to make sure head->next_sched is read
4121 * before clearing __QDISC_STATE_SCHED
4123 smp_mb__before_atomic();
4124 clear_bit(__QDISC_STATE_SCHED, &q->state);
4125 qdisc_run(q);
4126 spin_unlock(root_lock);
4131 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4132 /* This hook is defined here for ATM LANE */
4133 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4134 unsigned char *addr) __read_mostly;
4135 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4136 #endif
4138 static inline struct sk_buff *
4139 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4140 struct net_device *orig_dev)
4142 #ifdef CONFIG_NET_CLS_ACT
4143 struct tcf_proto *cl = rcu_dereference_bh(skb->dev->ingress_cl_list);
4144 struct tcf_result cl_res;
4146 /* If there's at least one ingress present somewhere (so
4147 * we get here via enabled static key), remaining devices
4148 * that are not configured with an ingress qdisc will bail
4149 * out here.
4151 if (!cl)
4152 return skb;
4153 if (*pt_prev) {
4154 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4155 *pt_prev = NULL;
4158 qdisc_skb_cb(skb)->pkt_len = skb->len;
4159 skb->tc_at_ingress = 1;
4160 qdisc_bstats_cpu_update(cl->q, skb);
4162 switch (tcf_classify(skb, cl, &cl_res, false)) {
4163 case TC_ACT_OK:
4164 case TC_ACT_RECLASSIFY:
4165 skb->tc_index = TC_H_MIN(cl_res.classid);
4166 break;
4167 case TC_ACT_SHOT:
4168 qdisc_qstats_cpu_drop(cl->q);
4169 kfree_skb(skb);
4170 return NULL;
4171 case TC_ACT_STOLEN:
4172 case TC_ACT_QUEUED:
4173 case TC_ACT_TRAP:
4174 consume_skb(skb);
4175 return NULL;
4176 case TC_ACT_REDIRECT:
4177 /* skb_mac_header check was done by cls/act_bpf, so
4178 * we can safely push the L2 header back before
4179 * redirecting to another netdev
4181 __skb_push(skb, skb->mac_len);
4182 skb_do_redirect(skb);
4183 return NULL;
4184 default:
4185 break;
4187 #endif /* CONFIG_NET_CLS_ACT */
4188 return skb;
4192 * netdev_is_rx_handler_busy - check if receive handler is registered
4193 * @dev: device to check
4195 * Check if a receive handler is already registered for a given device.
4196 * Return true if there one.
4198 * The caller must hold the rtnl_mutex.
4200 bool netdev_is_rx_handler_busy(struct net_device *dev)
4202 ASSERT_RTNL();
4203 return dev && rtnl_dereference(dev->rx_handler);
4205 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4208 * netdev_rx_handler_register - register receive handler
4209 * @dev: device to register a handler for
4210 * @rx_handler: receive handler to register
4211 * @rx_handler_data: data pointer that is used by rx handler
4213 * Register a receive handler for a device. This handler will then be
4214 * called from __netif_receive_skb. A negative errno code is returned
4215 * on a failure.
4217 * The caller must hold the rtnl_mutex.
4219 * For a general description of rx_handler, see enum rx_handler_result.
4221 int netdev_rx_handler_register(struct net_device *dev,
4222 rx_handler_func_t *rx_handler,
4223 void *rx_handler_data)
4225 if (netdev_is_rx_handler_busy(dev))
4226 return -EBUSY;
4228 /* Note: rx_handler_data must be set before rx_handler */
4229 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4230 rcu_assign_pointer(dev->rx_handler, rx_handler);
4232 return 0;
4234 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4237 * netdev_rx_handler_unregister - unregister receive handler
4238 * @dev: device to unregister a handler from
4240 * Unregister a receive handler from a device.
4242 * The caller must hold the rtnl_mutex.
4244 void netdev_rx_handler_unregister(struct net_device *dev)
4247 ASSERT_RTNL();
4248 RCU_INIT_POINTER(dev->rx_handler, NULL);
4249 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4250 * section has a guarantee to see a non NULL rx_handler_data
4251 * as well.
4253 synchronize_net();
4254 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4256 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4259 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4260 * the special handling of PFMEMALLOC skbs.
4262 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4264 switch (skb->protocol) {
4265 case htons(ETH_P_ARP):
4266 case htons(ETH_P_IP):
4267 case htons(ETH_P_IPV6):
4268 case htons(ETH_P_8021Q):
4269 case htons(ETH_P_8021AD):
4270 return true;
4271 default:
4272 return false;
4276 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4277 int *ret, struct net_device *orig_dev)
4279 #ifdef CONFIG_NETFILTER_INGRESS
4280 if (nf_hook_ingress_active(skb)) {
4281 int ingress_retval;
4283 if (*pt_prev) {
4284 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4285 *pt_prev = NULL;
4288 rcu_read_lock();
4289 ingress_retval = nf_hook_ingress(skb);
4290 rcu_read_unlock();
4291 return ingress_retval;
4293 #endif /* CONFIG_NETFILTER_INGRESS */
4294 return 0;
4297 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4299 struct packet_type *ptype, *pt_prev;
4300 rx_handler_func_t *rx_handler;
4301 struct net_device *orig_dev;
4302 bool deliver_exact = false;
4303 int ret = NET_RX_DROP;
4304 __be16 type;
4306 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4308 trace_netif_receive_skb(skb);
4310 orig_dev = skb->dev;
4312 skb_reset_network_header(skb);
4313 if (!skb_transport_header_was_set(skb))
4314 skb_reset_transport_header(skb);
4315 skb_reset_mac_len(skb);
4317 pt_prev = NULL;
4319 another_round:
4320 skb->skb_iif = skb->dev->ifindex;
4322 __this_cpu_inc(softnet_data.processed);
4324 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4325 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4326 skb = skb_vlan_untag(skb);
4327 if (unlikely(!skb))
4328 goto out;
4331 if (skb_skip_tc_classify(skb))
4332 goto skip_classify;
4334 if (pfmemalloc)
4335 goto skip_taps;
4337 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4338 if (pt_prev)
4339 ret = deliver_skb(skb, pt_prev, orig_dev);
4340 pt_prev = ptype;
4343 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4344 if (pt_prev)
4345 ret = deliver_skb(skb, pt_prev, orig_dev);
4346 pt_prev = ptype;
4349 skip_taps:
4350 #ifdef CONFIG_NET_INGRESS
4351 if (static_key_false(&ingress_needed)) {
4352 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4353 if (!skb)
4354 goto out;
4356 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4357 goto out;
4359 #endif
4360 skb_reset_tc(skb);
4361 skip_classify:
4362 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4363 goto drop;
4365 if (skb_vlan_tag_present(skb)) {
4366 if (pt_prev) {
4367 ret = deliver_skb(skb, pt_prev, orig_dev);
4368 pt_prev = NULL;
4370 if (vlan_do_receive(&skb))
4371 goto another_round;
4372 else if (unlikely(!skb))
4373 goto out;
4376 rx_handler = rcu_dereference(skb->dev->rx_handler);
4377 if (rx_handler) {
4378 if (pt_prev) {
4379 ret = deliver_skb(skb, pt_prev, orig_dev);
4380 pt_prev = NULL;
4382 switch (rx_handler(&skb)) {
4383 case RX_HANDLER_CONSUMED:
4384 ret = NET_RX_SUCCESS;
4385 goto out;
4386 case RX_HANDLER_ANOTHER:
4387 goto another_round;
4388 case RX_HANDLER_EXACT:
4389 deliver_exact = true;
4390 case RX_HANDLER_PASS:
4391 break;
4392 default:
4393 BUG();
4397 if (unlikely(skb_vlan_tag_present(skb))) {
4398 if (skb_vlan_tag_get_id(skb))
4399 skb->pkt_type = PACKET_OTHERHOST;
4400 /* Note: we might in the future use prio bits
4401 * and set skb->priority like in vlan_do_receive()
4402 * For the time being, just ignore Priority Code Point
4404 skb->vlan_tci = 0;
4407 type = skb->protocol;
4409 /* deliver only exact match when indicated */
4410 if (likely(!deliver_exact)) {
4411 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4412 &ptype_base[ntohs(type) &
4413 PTYPE_HASH_MASK]);
4416 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4417 &orig_dev->ptype_specific);
4419 if (unlikely(skb->dev != orig_dev)) {
4420 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4421 &skb->dev->ptype_specific);
4424 if (pt_prev) {
4425 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4426 goto drop;
4427 else
4428 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4429 } else {
4430 drop:
4431 if (!deliver_exact)
4432 atomic_long_inc(&skb->dev->rx_dropped);
4433 else
4434 atomic_long_inc(&skb->dev->rx_nohandler);
4435 kfree_skb(skb);
4436 /* Jamal, now you will not able to escape explaining
4437 * me how you were going to use this. :-)
4439 ret = NET_RX_DROP;
4442 out:
4443 return ret;
4446 static int __netif_receive_skb(struct sk_buff *skb)
4448 int ret;
4450 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4451 unsigned int noreclaim_flag;
4454 * PFMEMALLOC skbs are special, they should
4455 * - be delivered to SOCK_MEMALLOC sockets only
4456 * - stay away from userspace
4457 * - have bounded memory usage
4459 * Use PF_MEMALLOC as this saves us from propagating the allocation
4460 * context down to all allocation sites.
4462 noreclaim_flag = memalloc_noreclaim_save();
4463 ret = __netif_receive_skb_core(skb, true);
4464 memalloc_noreclaim_restore(noreclaim_flag);
4465 } else
4466 ret = __netif_receive_skb_core(skb, false);
4468 return ret;
4471 static int generic_xdp_install(struct net_device *dev, struct netdev_xdp *xdp)
4473 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4474 struct bpf_prog *new = xdp->prog;
4475 int ret = 0;
4477 switch (xdp->command) {
4478 case XDP_SETUP_PROG:
4479 rcu_assign_pointer(dev->xdp_prog, new);
4480 if (old)
4481 bpf_prog_put(old);
4483 if (old && !new) {
4484 static_key_slow_dec(&generic_xdp_needed);
4485 } else if (new && !old) {
4486 static_key_slow_inc(&generic_xdp_needed);
4487 dev_disable_lro(dev);
4489 break;
4491 case XDP_QUERY_PROG:
4492 xdp->prog_attached = !!old;
4493 xdp->prog_id = old ? old->aux->id : 0;
4494 break;
4496 default:
4497 ret = -EINVAL;
4498 break;
4501 return ret;
4504 static int netif_receive_skb_internal(struct sk_buff *skb)
4506 int ret;
4508 net_timestamp_check(netdev_tstamp_prequeue, skb);
4510 if (skb_defer_rx_timestamp(skb))
4511 return NET_RX_SUCCESS;
4513 if (static_key_false(&generic_xdp_needed)) {
4514 int ret;
4516 preempt_disable();
4517 rcu_read_lock();
4518 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4519 rcu_read_unlock();
4520 preempt_enable();
4522 if (ret != XDP_PASS)
4523 return NET_RX_DROP;
4526 rcu_read_lock();
4527 #ifdef CONFIG_RPS
4528 if (static_key_false(&rps_needed)) {
4529 struct rps_dev_flow voidflow, *rflow = &voidflow;
4530 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4532 if (cpu >= 0) {
4533 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4534 rcu_read_unlock();
4535 return ret;
4538 #endif
4539 ret = __netif_receive_skb(skb);
4540 rcu_read_unlock();
4541 return ret;
4545 * netif_receive_skb - process receive buffer from network
4546 * @skb: buffer to process
4548 * netif_receive_skb() is the main receive data processing function.
4549 * It always succeeds. The buffer may be dropped during processing
4550 * for congestion control or by the protocol layers.
4552 * This function may only be called from softirq context and interrupts
4553 * should be enabled.
4555 * Return values (usually ignored):
4556 * NET_RX_SUCCESS: no congestion
4557 * NET_RX_DROP: packet was dropped
4559 int netif_receive_skb(struct sk_buff *skb)
4561 trace_netif_receive_skb_entry(skb);
4563 return netif_receive_skb_internal(skb);
4565 EXPORT_SYMBOL(netif_receive_skb);
4567 DEFINE_PER_CPU(struct work_struct, flush_works);
4569 /* Network device is going away, flush any packets still pending */
4570 static void flush_backlog(struct work_struct *work)
4572 struct sk_buff *skb, *tmp;
4573 struct softnet_data *sd;
4575 local_bh_disable();
4576 sd = this_cpu_ptr(&softnet_data);
4578 local_irq_disable();
4579 rps_lock(sd);
4580 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4581 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4582 __skb_unlink(skb, &sd->input_pkt_queue);
4583 kfree_skb(skb);
4584 input_queue_head_incr(sd);
4587 rps_unlock(sd);
4588 local_irq_enable();
4590 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4591 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4592 __skb_unlink(skb, &sd->process_queue);
4593 kfree_skb(skb);
4594 input_queue_head_incr(sd);
4597 local_bh_enable();
4600 static void flush_all_backlogs(void)
4602 unsigned int cpu;
4604 get_online_cpus();
4606 for_each_online_cpu(cpu)
4607 queue_work_on(cpu, system_highpri_wq,
4608 per_cpu_ptr(&flush_works, cpu));
4610 for_each_online_cpu(cpu)
4611 flush_work(per_cpu_ptr(&flush_works, cpu));
4613 put_online_cpus();
4616 static int napi_gro_complete(struct sk_buff *skb)
4618 struct packet_offload *ptype;
4619 __be16 type = skb->protocol;
4620 struct list_head *head = &offload_base;
4621 int err = -ENOENT;
4623 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4625 if (NAPI_GRO_CB(skb)->count == 1) {
4626 skb_shinfo(skb)->gso_size = 0;
4627 goto out;
4630 rcu_read_lock();
4631 list_for_each_entry_rcu(ptype, head, list) {
4632 if (ptype->type != type || !ptype->callbacks.gro_complete)
4633 continue;
4635 err = ptype->callbacks.gro_complete(skb, 0);
4636 break;
4638 rcu_read_unlock();
4640 if (err) {
4641 WARN_ON(&ptype->list == head);
4642 kfree_skb(skb);
4643 return NET_RX_SUCCESS;
4646 out:
4647 return netif_receive_skb_internal(skb);
4650 /* napi->gro_list contains packets ordered by age.
4651 * youngest packets at the head of it.
4652 * Complete skbs in reverse order to reduce latencies.
4654 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4656 struct sk_buff *skb, *prev = NULL;
4658 /* scan list and build reverse chain */
4659 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4660 skb->prev = prev;
4661 prev = skb;
4664 for (skb = prev; skb; skb = prev) {
4665 skb->next = NULL;
4667 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4668 return;
4670 prev = skb->prev;
4671 napi_gro_complete(skb);
4672 napi->gro_count--;
4675 napi->gro_list = NULL;
4677 EXPORT_SYMBOL(napi_gro_flush);
4679 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4681 struct sk_buff *p;
4682 unsigned int maclen = skb->dev->hard_header_len;
4683 u32 hash = skb_get_hash_raw(skb);
4685 for (p = napi->gro_list; p; p = p->next) {
4686 unsigned long diffs;
4688 NAPI_GRO_CB(p)->flush = 0;
4690 if (hash != skb_get_hash_raw(p)) {
4691 NAPI_GRO_CB(p)->same_flow = 0;
4692 continue;
4695 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4696 diffs |= p->vlan_tci ^ skb->vlan_tci;
4697 diffs |= skb_metadata_dst_cmp(p, skb);
4698 if (maclen == ETH_HLEN)
4699 diffs |= compare_ether_header(skb_mac_header(p),
4700 skb_mac_header(skb));
4701 else if (!diffs)
4702 diffs = memcmp(skb_mac_header(p),
4703 skb_mac_header(skb),
4704 maclen);
4705 NAPI_GRO_CB(p)->same_flow = !diffs;
4709 static void skb_gro_reset_offset(struct sk_buff *skb)
4711 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4712 const skb_frag_t *frag0 = &pinfo->frags[0];
4714 NAPI_GRO_CB(skb)->data_offset = 0;
4715 NAPI_GRO_CB(skb)->frag0 = NULL;
4716 NAPI_GRO_CB(skb)->frag0_len = 0;
4718 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4719 pinfo->nr_frags &&
4720 !PageHighMem(skb_frag_page(frag0))) {
4721 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4722 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4723 skb_frag_size(frag0),
4724 skb->end - skb->tail);
4728 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4730 struct skb_shared_info *pinfo = skb_shinfo(skb);
4732 BUG_ON(skb->end - skb->tail < grow);
4734 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4736 skb->data_len -= grow;
4737 skb->tail += grow;
4739 pinfo->frags[0].page_offset += grow;
4740 skb_frag_size_sub(&pinfo->frags[0], grow);
4742 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4743 skb_frag_unref(skb, 0);
4744 memmove(pinfo->frags, pinfo->frags + 1,
4745 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4749 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4751 struct sk_buff **pp = NULL;
4752 struct packet_offload *ptype;
4753 __be16 type = skb->protocol;
4754 struct list_head *head = &offload_base;
4755 int same_flow;
4756 enum gro_result ret;
4757 int grow;
4759 if (netif_elide_gro(skb->dev))
4760 goto normal;
4762 gro_list_prepare(napi, skb);
4764 rcu_read_lock();
4765 list_for_each_entry_rcu(ptype, head, list) {
4766 if (ptype->type != type || !ptype->callbacks.gro_receive)
4767 continue;
4769 skb_set_network_header(skb, skb_gro_offset(skb));
4770 skb_reset_mac_len(skb);
4771 NAPI_GRO_CB(skb)->same_flow = 0;
4772 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4773 NAPI_GRO_CB(skb)->free = 0;
4774 NAPI_GRO_CB(skb)->encap_mark = 0;
4775 NAPI_GRO_CB(skb)->recursion_counter = 0;
4776 NAPI_GRO_CB(skb)->is_fou = 0;
4777 NAPI_GRO_CB(skb)->is_atomic = 1;
4778 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4780 /* Setup for GRO checksum validation */
4781 switch (skb->ip_summed) {
4782 case CHECKSUM_COMPLETE:
4783 NAPI_GRO_CB(skb)->csum = skb->csum;
4784 NAPI_GRO_CB(skb)->csum_valid = 1;
4785 NAPI_GRO_CB(skb)->csum_cnt = 0;
4786 break;
4787 case CHECKSUM_UNNECESSARY:
4788 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4789 NAPI_GRO_CB(skb)->csum_valid = 0;
4790 break;
4791 default:
4792 NAPI_GRO_CB(skb)->csum_cnt = 0;
4793 NAPI_GRO_CB(skb)->csum_valid = 0;
4796 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
4797 break;
4799 rcu_read_unlock();
4801 if (&ptype->list == head)
4802 goto normal;
4804 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
4805 ret = GRO_CONSUMED;
4806 goto ok;
4809 same_flow = NAPI_GRO_CB(skb)->same_flow;
4810 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
4812 if (pp) {
4813 struct sk_buff *nskb = *pp;
4815 *pp = nskb->next;
4816 nskb->next = NULL;
4817 napi_gro_complete(nskb);
4818 napi->gro_count--;
4821 if (same_flow)
4822 goto ok;
4824 if (NAPI_GRO_CB(skb)->flush)
4825 goto normal;
4827 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
4828 struct sk_buff *nskb = napi->gro_list;
4830 /* locate the end of the list to select the 'oldest' flow */
4831 while (nskb->next) {
4832 pp = &nskb->next;
4833 nskb = *pp;
4835 *pp = NULL;
4836 nskb->next = NULL;
4837 napi_gro_complete(nskb);
4838 } else {
4839 napi->gro_count++;
4841 NAPI_GRO_CB(skb)->count = 1;
4842 NAPI_GRO_CB(skb)->age = jiffies;
4843 NAPI_GRO_CB(skb)->last = skb;
4844 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
4845 skb->next = napi->gro_list;
4846 napi->gro_list = skb;
4847 ret = GRO_HELD;
4849 pull:
4850 grow = skb_gro_offset(skb) - skb_headlen(skb);
4851 if (grow > 0)
4852 gro_pull_from_frag0(skb, grow);
4854 return ret;
4856 normal:
4857 ret = GRO_NORMAL;
4858 goto pull;
4861 struct packet_offload *gro_find_receive_by_type(__be16 type)
4863 struct list_head *offload_head = &offload_base;
4864 struct packet_offload *ptype;
4866 list_for_each_entry_rcu(ptype, offload_head, list) {
4867 if (ptype->type != type || !ptype->callbacks.gro_receive)
4868 continue;
4869 return ptype;
4871 return NULL;
4873 EXPORT_SYMBOL(gro_find_receive_by_type);
4875 struct packet_offload *gro_find_complete_by_type(__be16 type)
4877 struct list_head *offload_head = &offload_base;
4878 struct packet_offload *ptype;
4880 list_for_each_entry_rcu(ptype, offload_head, list) {
4881 if (ptype->type != type || !ptype->callbacks.gro_complete)
4882 continue;
4883 return ptype;
4885 return NULL;
4887 EXPORT_SYMBOL(gro_find_complete_by_type);
4889 static void napi_skb_free_stolen_head(struct sk_buff *skb)
4891 skb_dst_drop(skb);
4892 secpath_reset(skb);
4893 kmem_cache_free(skbuff_head_cache, skb);
4896 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
4898 switch (ret) {
4899 case GRO_NORMAL:
4900 if (netif_receive_skb_internal(skb))
4901 ret = GRO_DROP;
4902 break;
4904 case GRO_DROP:
4905 kfree_skb(skb);
4906 break;
4908 case GRO_MERGED_FREE:
4909 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4910 napi_skb_free_stolen_head(skb);
4911 else
4912 __kfree_skb(skb);
4913 break;
4915 case GRO_HELD:
4916 case GRO_MERGED:
4917 case GRO_CONSUMED:
4918 break;
4921 return ret;
4924 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4926 skb_mark_napi_id(skb, napi);
4927 trace_napi_gro_receive_entry(skb);
4929 skb_gro_reset_offset(skb);
4931 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
4933 EXPORT_SYMBOL(napi_gro_receive);
4935 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
4937 if (unlikely(skb->pfmemalloc)) {
4938 consume_skb(skb);
4939 return;
4941 __skb_pull(skb, skb_headlen(skb));
4942 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
4943 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
4944 skb->vlan_tci = 0;
4945 skb->dev = napi->dev;
4946 skb->skb_iif = 0;
4947 skb->encapsulation = 0;
4948 skb_shinfo(skb)->gso_type = 0;
4949 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
4950 secpath_reset(skb);
4952 napi->skb = skb;
4955 struct sk_buff *napi_get_frags(struct napi_struct *napi)
4957 struct sk_buff *skb = napi->skb;
4959 if (!skb) {
4960 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
4961 if (skb) {
4962 napi->skb = skb;
4963 skb_mark_napi_id(skb, napi);
4966 return skb;
4968 EXPORT_SYMBOL(napi_get_frags);
4970 static gro_result_t napi_frags_finish(struct napi_struct *napi,
4971 struct sk_buff *skb,
4972 gro_result_t ret)
4974 switch (ret) {
4975 case GRO_NORMAL:
4976 case GRO_HELD:
4977 __skb_push(skb, ETH_HLEN);
4978 skb->protocol = eth_type_trans(skb, skb->dev);
4979 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
4980 ret = GRO_DROP;
4981 break;
4983 case GRO_DROP:
4984 napi_reuse_skb(napi, skb);
4985 break;
4987 case GRO_MERGED_FREE:
4988 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4989 napi_skb_free_stolen_head(skb);
4990 else
4991 napi_reuse_skb(napi, skb);
4992 break;
4994 case GRO_MERGED:
4995 case GRO_CONSUMED:
4996 break;
4999 return ret;
5002 /* Upper GRO stack assumes network header starts at gro_offset=0
5003 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5004 * We copy ethernet header into skb->data to have a common layout.
5006 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5008 struct sk_buff *skb = napi->skb;
5009 const struct ethhdr *eth;
5010 unsigned int hlen = sizeof(*eth);
5012 napi->skb = NULL;
5014 skb_reset_mac_header(skb);
5015 skb_gro_reset_offset(skb);
5017 eth = skb_gro_header_fast(skb, 0);
5018 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5019 eth = skb_gro_header_slow(skb, hlen, 0);
5020 if (unlikely(!eth)) {
5021 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5022 __func__, napi->dev->name);
5023 napi_reuse_skb(napi, skb);
5024 return NULL;
5026 } else {
5027 gro_pull_from_frag0(skb, hlen);
5028 NAPI_GRO_CB(skb)->frag0 += hlen;
5029 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5031 __skb_pull(skb, hlen);
5034 * This works because the only protocols we care about don't require
5035 * special handling.
5036 * We'll fix it up properly in napi_frags_finish()
5038 skb->protocol = eth->h_proto;
5040 return skb;
5043 gro_result_t napi_gro_frags(struct napi_struct *napi)
5045 struct sk_buff *skb = napi_frags_skb(napi);
5047 if (!skb)
5048 return GRO_DROP;
5050 trace_napi_gro_frags_entry(skb);
5052 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5054 EXPORT_SYMBOL(napi_gro_frags);
5056 /* Compute the checksum from gro_offset and return the folded value
5057 * after adding in any pseudo checksum.
5059 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5061 __wsum wsum;
5062 __sum16 sum;
5064 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5066 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5067 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5068 if (likely(!sum)) {
5069 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5070 !skb->csum_complete_sw)
5071 netdev_rx_csum_fault(skb->dev);
5074 NAPI_GRO_CB(skb)->csum = wsum;
5075 NAPI_GRO_CB(skb)->csum_valid = 1;
5077 return sum;
5079 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5081 static void net_rps_send_ipi(struct softnet_data *remsd)
5083 #ifdef CONFIG_RPS
5084 while (remsd) {
5085 struct softnet_data *next = remsd->rps_ipi_next;
5087 if (cpu_online(remsd->cpu))
5088 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5089 remsd = next;
5091 #endif
5095 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5096 * Note: called with local irq disabled, but exits with local irq enabled.
5098 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5100 #ifdef CONFIG_RPS
5101 struct softnet_data *remsd = sd->rps_ipi_list;
5103 if (remsd) {
5104 sd->rps_ipi_list = NULL;
5106 local_irq_enable();
5108 /* Send pending IPI's to kick RPS processing on remote cpus. */
5109 net_rps_send_ipi(remsd);
5110 } else
5111 #endif
5112 local_irq_enable();
5115 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5117 #ifdef CONFIG_RPS
5118 return sd->rps_ipi_list != NULL;
5119 #else
5120 return false;
5121 #endif
5124 static int process_backlog(struct napi_struct *napi, int quota)
5126 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5127 bool again = true;
5128 int work = 0;
5130 /* Check if we have pending ipi, its better to send them now,
5131 * not waiting net_rx_action() end.
5133 if (sd_has_rps_ipi_waiting(sd)) {
5134 local_irq_disable();
5135 net_rps_action_and_irq_enable(sd);
5138 napi->weight = dev_rx_weight;
5139 while (again) {
5140 struct sk_buff *skb;
5142 while ((skb = __skb_dequeue(&sd->process_queue))) {
5143 rcu_read_lock();
5144 __netif_receive_skb(skb);
5145 rcu_read_unlock();
5146 input_queue_head_incr(sd);
5147 if (++work >= quota)
5148 return work;
5152 local_irq_disable();
5153 rps_lock(sd);
5154 if (skb_queue_empty(&sd->input_pkt_queue)) {
5156 * Inline a custom version of __napi_complete().
5157 * only current cpu owns and manipulates this napi,
5158 * and NAPI_STATE_SCHED is the only possible flag set
5159 * on backlog.
5160 * We can use a plain write instead of clear_bit(),
5161 * and we dont need an smp_mb() memory barrier.
5163 napi->state = 0;
5164 again = false;
5165 } else {
5166 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5167 &sd->process_queue);
5169 rps_unlock(sd);
5170 local_irq_enable();
5173 return work;
5177 * __napi_schedule - schedule for receive
5178 * @n: entry to schedule
5180 * The entry's receive function will be scheduled to run.
5181 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5183 void __napi_schedule(struct napi_struct *n)
5185 unsigned long flags;
5187 local_irq_save(flags);
5188 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5189 local_irq_restore(flags);
5191 EXPORT_SYMBOL(__napi_schedule);
5194 * napi_schedule_prep - check if napi can be scheduled
5195 * @n: napi context
5197 * Test if NAPI routine is already running, and if not mark
5198 * it as running. This is used as a condition variable
5199 * insure only one NAPI poll instance runs. We also make
5200 * sure there is no pending NAPI disable.
5202 bool napi_schedule_prep(struct napi_struct *n)
5204 unsigned long val, new;
5206 do {
5207 val = READ_ONCE(n->state);
5208 if (unlikely(val & NAPIF_STATE_DISABLE))
5209 return false;
5210 new = val | NAPIF_STATE_SCHED;
5212 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5213 * This was suggested by Alexander Duyck, as compiler
5214 * emits better code than :
5215 * if (val & NAPIF_STATE_SCHED)
5216 * new |= NAPIF_STATE_MISSED;
5218 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5219 NAPIF_STATE_MISSED;
5220 } while (cmpxchg(&n->state, val, new) != val);
5222 return !(val & NAPIF_STATE_SCHED);
5224 EXPORT_SYMBOL(napi_schedule_prep);
5227 * __napi_schedule_irqoff - schedule for receive
5228 * @n: entry to schedule
5230 * Variant of __napi_schedule() assuming hard irqs are masked
5232 void __napi_schedule_irqoff(struct napi_struct *n)
5234 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5236 EXPORT_SYMBOL(__napi_schedule_irqoff);
5238 bool napi_complete_done(struct napi_struct *n, int work_done)
5240 unsigned long flags, val, new;
5243 * 1) Don't let napi dequeue from the cpu poll list
5244 * just in case its running on a different cpu.
5245 * 2) If we are busy polling, do nothing here, we have
5246 * the guarantee we will be called later.
5248 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5249 NAPIF_STATE_IN_BUSY_POLL)))
5250 return false;
5252 if (n->gro_list) {
5253 unsigned long timeout = 0;
5255 if (work_done)
5256 timeout = n->dev->gro_flush_timeout;
5258 if (timeout)
5259 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5260 HRTIMER_MODE_REL_PINNED);
5261 else
5262 napi_gro_flush(n, false);
5264 if (unlikely(!list_empty(&n->poll_list))) {
5265 /* If n->poll_list is not empty, we need to mask irqs */
5266 local_irq_save(flags);
5267 list_del_init(&n->poll_list);
5268 local_irq_restore(flags);
5271 do {
5272 val = READ_ONCE(n->state);
5274 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5276 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5278 /* If STATE_MISSED was set, leave STATE_SCHED set,
5279 * because we will call napi->poll() one more time.
5280 * This C code was suggested by Alexander Duyck to help gcc.
5282 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5283 NAPIF_STATE_SCHED;
5284 } while (cmpxchg(&n->state, val, new) != val);
5286 if (unlikely(val & NAPIF_STATE_MISSED)) {
5287 __napi_schedule(n);
5288 return false;
5291 return true;
5293 EXPORT_SYMBOL(napi_complete_done);
5295 /* must be called under rcu_read_lock(), as we dont take a reference */
5296 static struct napi_struct *napi_by_id(unsigned int napi_id)
5298 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5299 struct napi_struct *napi;
5301 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5302 if (napi->napi_id == napi_id)
5303 return napi;
5305 return NULL;
5308 #if defined(CONFIG_NET_RX_BUSY_POLL)
5310 #define BUSY_POLL_BUDGET 8
5312 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5314 int rc;
5316 /* Busy polling means there is a high chance device driver hard irq
5317 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5318 * set in napi_schedule_prep().
5319 * Since we are about to call napi->poll() once more, we can safely
5320 * clear NAPI_STATE_MISSED.
5322 * Note: x86 could use a single "lock and ..." instruction
5323 * to perform these two clear_bit()
5325 clear_bit(NAPI_STATE_MISSED, &napi->state);
5326 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5328 local_bh_disable();
5330 /* All we really want here is to re-enable device interrupts.
5331 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5333 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5334 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5335 netpoll_poll_unlock(have_poll_lock);
5336 if (rc == BUSY_POLL_BUDGET)
5337 __napi_schedule(napi);
5338 local_bh_enable();
5341 void napi_busy_loop(unsigned int napi_id,
5342 bool (*loop_end)(void *, unsigned long),
5343 void *loop_end_arg)
5345 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5346 int (*napi_poll)(struct napi_struct *napi, int budget);
5347 void *have_poll_lock = NULL;
5348 struct napi_struct *napi;
5350 restart:
5351 napi_poll = NULL;
5353 rcu_read_lock();
5355 napi = napi_by_id(napi_id);
5356 if (!napi)
5357 goto out;
5359 preempt_disable();
5360 for (;;) {
5361 int work = 0;
5363 local_bh_disable();
5364 if (!napi_poll) {
5365 unsigned long val = READ_ONCE(napi->state);
5367 /* If multiple threads are competing for this napi,
5368 * we avoid dirtying napi->state as much as we can.
5370 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5371 NAPIF_STATE_IN_BUSY_POLL))
5372 goto count;
5373 if (cmpxchg(&napi->state, val,
5374 val | NAPIF_STATE_IN_BUSY_POLL |
5375 NAPIF_STATE_SCHED) != val)
5376 goto count;
5377 have_poll_lock = netpoll_poll_lock(napi);
5378 napi_poll = napi->poll;
5380 work = napi_poll(napi, BUSY_POLL_BUDGET);
5381 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5382 count:
5383 if (work > 0)
5384 __NET_ADD_STATS(dev_net(napi->dev),
5385 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5386 local_bh_enable();
5388 if (!loop_end || loop_end(loop_end_arg, start_time))
5389 break;
5391 if (unlikely(need_resched())) {
5392 if (napi_poll)
5393 busy_poll_stop(napi, have_poll_lock);
5394 preempt_enable();
5395 rcu_read_unlock();
5396 cond_resched();
5397 if (loop_end(loop_end_arg, start_time))
5398 return;
5399 goto restart;
5401 cpu_relax();
5403 if (napi_poll)
5404 busy_poll_stop(napi, have_poll_lock);
5405 preempt_enable();
5406 out:
5407 rcu_read_unlock();
5409 EXPORT_SYMBOL(napi_busy_loop);
5411 #endif /* CONFIG_NET_RX_BUSY_POLL */
5413 static void napi_hash_add(struct napi_struct *napi)
5415 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5416 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5417 return;
5419 spin_lock(&napi_hash_lock);
5421 /* 0..NR_CPUS range is reserved for sender_cpu use */
5422 do {
5423 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5424 napi_gen_id = MIN_NAPI_ID;
5425 } while (napi_by_id(napi_gen_id));
5426 napi->napi_id = napi_gen_id;
5428 hlist_add_head_rcu(&napi->napi_hash_node,
5429 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5431 spin_unlock(&napi_hash_lock);
5434 /* Warning : caller is responsible to make sure rcu grace period
5435 * is respected before freeing memory containing @napi
5437 bool napi_hash_del(struct napi_struct *napi)
5439 bool rcu_sync_needed = false;
5441 spin_lock(&napi_hash_lock);
5443 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5444 rcu_sync_needed = true;
5445 hlist_del_rcu(&napi->napi_hash_node);
5447 spin_unlock(&napi_hash_lock);
5448 return rcu_sync_needed;
5450 EXPORT_SYMBOL_GPL(napi_hash_del);
5452 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5454 struct napi_struct *napi;
5456 napi = container_of(timer, struct napi_struct, timer);
5458 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5459 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5461 if (napi->gro_list && !napi_disable_pending(napi) &&
5462 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5463 __napi_schedule_irqoff(napi);
5465 return HRTIMER_NORESTART;
5468 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5469 int (*poll)(struct napi_struct *, int), int weight)
5471 INIT_LIST_HEAD(&napi->poll_list);
5472 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5473 napi->timer.function = napi_watchdog;
5474 napi->gro_count = 0;
5475 napi->gro_list = NULL;
5476 napi->skb = NULL;
5477 napi->poll = poll;
5478 if (weight > NAPI_POLL_WEIGHT)
5479 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5480 weight, dev->name);
5481 napi->weight = weight;
5482 list_add(&napi->dev_list, &dev->napi_list);
5483 napi->dev = dev;
5484 #ifdef CONFIG_NETPOLL
5485 napi->poll_owner = -1;
5486 #endif
5487 set_bit(NAPI_STATE_SCHED, &napi->state);
5488 napi_hash_add(napi);
5490 EXPORT_SYMBOL(netif_napi_add);
5492 void napi_disable(struct napi_struct *n)
5494 might_sleep();
5495 set_bit(NAPI_STATE_DISABLE, &n->state);
5497 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5498 msleep(1);
5499 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5500 msleep(1);
5502 hrtimer_cancel(&n->timer);
5504 clear_bit(NAPI_STATE_DISABLE, &n->state);
5506 EXPORT_SYMBOL(napi_disable);
5508 /* Must be called in process context */
5509 void netif_napi_del(struct napi_struct *napi)
5511 might_sleep();
5512 if (napi_hash_del(napi))
5513 synchronize_net();
5514 list_del_init(&napi->dev_list);
5515 napi_free_frags(napi);
5517 kfree_skb_list(napi->gro_list);
5518 napi->gro_list = NULL;
5519 napi->gro_count = 0;
5521 EXPORT_SYMBOL(netif_napi_del);
5523 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5525 void *have;
5526 int work, weight;
5528 list_del_init(&n->poll_list);
5530 have = netpoll_poll_lock(n);
5532 weight = n->weight;
5534 /* This NAPI_STATE_SCHED test is for avoiding a race
5535 * with netpoll's poll_napi(). Only the entity which
5536 * obtains the lock and sees NAPI_STATE_SCHED set will
5537 * actually make the ->poll() call. Therefore we avoid
5538 * accidentally calling ->poll() when NAPI is not scheduled.
5540 work = 0;
5541 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5542 work = n->poll(n, weight);
5543 trace_napi_poll(n, work, weight);
5546 WARN_ON_ONCE(work > weight);
5548 if (likely(work < weight))
5549 goto out_unlock;
5551 /* Drivers must not modify the NAPI state if they
5552 * consume the entire weight. In such cases this code
5553 * still "owns" the NAPI instance and therefore can
5554 * move the instance around on the list at-will.
5556 if (unlikely(napi_disable_pending(n))) {
5557 napi_complete(n);
5558 goto out_unlock;
5561 if (n->gro_list) {
5562 /* flush too old packets
5563 * If HZ < 1000, flush all packets.
5565 napi_gro_flush(n, HZ >= 1000);
5568 /* Some drivers may have called napi_schedule
5569 * prior to exhausting their budget.
5571 if (unlikely(!list_empty(&n->poll_list))) {
5572 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5573 n->dev ? n->dev->name : "backlog");
5574 goto out_unlock;
5577 list_add_tail(&n->poll_list, repoll);
5579 out_unlock:
5580 netpoll_poll_unlock(have);
5582 return work;
5585 static __latent_entropy void net_rx_action(struct softirq_action *h)
5587 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5588 unsigned long time_limit = jiffies +
5589 usecs_to_jiffies(netdev_budget_usecs);
5590 int budget = netdev_budget;
5591 LIST_HEAD(list);
5592 LIST_HEAD(repoll);
5594 local_irq_disable();
5595 list_splice_init(&sd->poll_list, &list);
5596 local_irq_enable();
5598 for (;;) {
5599 struct napi_struct *n;
5601 if (list_empty(&list)) {
5602 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5603 goto out;
5604 break;
5607 n = list_first_entry(&list, struct napi_struct, poll_list);
5608 budget -= napi_poll(n, &repoll);
5610 /* If softirq window is exhausted then punt.
5611 * Allow this to run for 2 jiffies since which will allow
5612 * an average latency of 1.5/HZ.
5614 if (unlikely(budget <= 0 ||
5615 time_after_eq(jiffies, time_limit))) {
5616 sd->time_squeeze++;
5617 break;
5621 local_irq_disable();
5623 list_splice_tail_init(&sd->poll_list, &list);
5624 list_splice_tail(&repoll, &list);
5625 list_splice(&list, &sd->poll_list);
5626 if (!list_empty(&sd->poll_list))
5627 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5629 net_rps_action_and_irq_enable(sd);
5630 out:
5631 __kfree_skb_flush();
5634 struct netdev_adjacent {
5635 struct net_device *dev;
5637 /* upper master flag, there can only be one master device per list */
5638 bool master;
5640 /* counter for the number of times this device was added to us */
5641 u16 ref_nr;
5643 /* private field for the users */
5644 void *private;
5646 struct list_head list;
5647 struct rcu_head rcu;
5650 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5651 struct list_head *adj_list)
5653 struct netdev_adjacent *adj;
5655 list_for_each_entry(adj, adj_list, list) {
5656 if (adj->dev == adj_dev)
5657 return adj;
5659 return NULL;
5662 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5664 struct net_device *dev = data;
5666 return upper_dev == dev;
5670 * netdev_has_upper_dev - Check if device is linked to an upper device
5671 * @dev: device
5672 * @upper_dev: upper device to check
5674 * Find out if a device is linked to specified upper device and return true
5675 * in case it is. Note that this checks only immediate upper device,
5676 * not through a complete stack of devices. The caller must hold the RTNL lock.
5678 bool netdev_has_upper_dev(struct net_device *dev,
5679 struct net_device *upper_dev)
5681 ASSERT_RTNL();
5683 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5684 upper_dev);
5686 EXPORT_SYMBOL(netdev_has_upper_dev);
5689 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5690 * @dev: device
5691 * @upper_dev: upper device to check
5693 * Find out if a device is linked to specified upper device and return true
5694 * in case it is. Note that this checks the entire upper device chain.
5695 * The caller must hold rcu lock.
5698 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5699 struct net_device *upper_dev)
5701 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5702 upper_dev);
5704 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5707 * netdev_has_any_upper_dev - Check if device is linked to some device
5708 * @dev: device
5710 * Find out if a device is linked to an upper device and return true in case
5711 * it is. The caller must hold the RTNL lock.
5713 bool netdev_has_any_upper_dev(struct net_device *dev)
5715 ASSERT_RTNL();
5717 return !list_empty(&dev->adj_list.upper);
5719 EXPORT_SYMBOL(netdev_has_any_upper_dev);
5722 * netdev_master_upper_dev_get - Get master upper device
5723 * @dev: device
5725 * Find a master upper device and return pointer to it or NULL in case
5726 * it's not there. The caller must hold the RTNL lock.
5728 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5730 struct netdev_adjacent *upper;
5732 ASSERT_RTNL();
5734 if (list_empty(&dev->adj_list.upper))
5735 return NULL;
5737 upper = list_first_entry(&dev->adj_list.upper,
5738 struct netdev_adjacent, list);
5739 if (likely(upper->master))
5740 return upper->dev;
5741 return NULL;
5743 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5746 * netdev_has_any_lower_dev - Check if device is linked to some device
5747 * @dev: device
5749 * Find out if a device is linked to a lower device and return true in case
5750 * it is. The caller must hold the RTNL lock.
5752 static bool netdev_has_any_lower_dev(struct net_device *dev)
5754 ASSERT_RTNL();
5756 return !list_empty(&dev->adj_list.lower);
5759 void *netdev_adjacent_get_private(struct list_head *adj_list)
5761 struct netdev_adjacent *adj;
5763 adj = list_entry(adj_list, struct netdev_adjacent, list);
5765 return adj->private;
5767 EXPORT_SYMBOL(netdev_adjacent_get_private);
5770 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5771 * @dev: device
5772 * @iter: list_head ** of the current position
5774 * Gets the next device from the dev's upper list, starting from iter
5775 * position. The caller must hold RCU read lock.
5777 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5778 struct list_head **iter)
5780 struct netdev_adjacent *upper;
5782 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5784 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5786 if (&upper->list == &dev->adj_list.upper)
5787 return NULL;
5789 *iter = &upper->list;
5791 return upper->dev;
5793 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5795 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5796 struct list_head **iter)
5798 struct netdev_adjacent *upper;
5800 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5802 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5804 if (&upper->list == &dev->adj_list.upper)
5805 return NULL;
5807 *iter = &upper->list;
5809 return upper->dev;
5812 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
5813 int (*fn)(struct net_device *dev,
5814 void *data),
5815 void *data)
5817 struct net_device *udev;
5818 struct list_head *iter;
5819 int ret;
5821 for (iter = &dev->adj_list.upper,
5822 udev = netdev_next_upper_dev_rcu(dev, &iter);
5823 udev;
5824 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
5825 /* first is the upper device itself */
5826 ret = fn(udev, data);
5827 if (ret)
5828 return ret;
5830 /* then look at all of its upper devices */
5831 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
5832 if (ret)
5833 return ret;
5836 return 0;
5838 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
5841 * netdev_lower_get_next_private - Get the next ->private from the
5842 * lower neighbour list
5843 * @dev: device
5844 * @iter: list_head ** of the current position
5846 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5847 * list, starting from iter position. The caller must hold either hold the
5848 * RTNL lock or its own locking that guarantees that the neighbour lower
5849 * list will remain unchanged.
5851 void *netdev_lower_get_next_private(struct net_device *dev,
5852 struct list_head **iter)
5854 struct netdev_adjacent *lower;
5856 lower = list_entry(*iter, struct netdev_adjacent, list);
5858 if (&lower->list == &dev->adj_list.lower)
5859 return NULL;
5861 *iter = lower->list.next;
5863 return lower->private;
5865 EXPORT_SYMBOL(netdev_lower_get_next_private);
5868 * netdev_lower_get_next_private_rcu - Get the next ->private from the
5869 * lower neighbour list, RCU
5870 * variant
5871 * @dev: device
5872 * @iter: list_head ** of the current position
5874 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5875 * list, starting from iter position. The caller must hold RCU read lock.
5877 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
5878 struct list_head **iter)
5880 struct netdev_adjacent *lower;
5882 WARN_ON_ONCE(!rcu_read_lock_held());
5884 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5886 if (&lower->list == &dev->adj_list.lower)
5887 return NULL;
5889 *iter = &lower->list;
5891 return lower->private;
5893 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
5896 * netdev_lower_get_next - Get the next device from the lower neighbour
5897 * list
5898 * @dev: device
5899 * @iter: list_head ** of the current position
5901 * Gets the next netdev_adjacent from the dev's lower neighbour
5902 * list, starting from iter position. The caller must hold RTNL lock or
5903 * its own locking that guarantees that the neighbour lower
5904 * list will remain unchanged.
5906 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
5908 struct netdev_adjacent *lower;
5910 lower = list_entry(*iter, struct netdev_adjacent, list);
5912 if (&lower->list == &dev->adj_list.lower)
5913 return NULL;
5915 *iter = lower->list.next;
5917 return lower->dev;
5919 EXPORT_SYMBOL(netdev_lower_get_next);
5921 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
5922 struct list_head **iter)
5924 struct netdev_adjacent *lower;
5926 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
5928 if (&lower->list == &dev->adj_list.lower)
5929 return NULL;
5931 *iter = &lower->list;
5933 return lower->dev;
5936 int netdev_walk_all_lower_dev(struct net_device *dev,
5937 int (*fn)(struct net_device *dev,
5938 void *data),
5939 void *data)
5941 struct net_device *ldev;
5942 struct list_head *iter;
5943 int ret;
5945 for (iter = &dev->adj_list.lower,
5946 ldev = netdev_next_lower_dev(dev, &iter);
5947 ldev;
5948 ldev = netdev_next_lower_dev(dev, &iter)) {
5949 /* first is the lower device itself */
5950 ret = fn(ldev, data);
5951 if (ret)
5952 return ret;
5954 /* then look at all of its lower devices */
5955 ret = netdev_walk_all_lower_dev(ldev, fn, data);
5956 if (ret)
5957 return ret;
5960 return 0;
5962 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
5964 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
5965 struct list_head **iter)
5967 struct netdev_adjacent *lower;
5969 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5970 if (&lower->list == &dev->adj_list.lower)
5971 return NULL;
5973 *iter = &lower->list;
5975 return lower->dev;
5978 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
5979 int (*fn)(struct net_device *dev,
5980 void *data),
5981 void *data)
5983 struct net_device *ldev;
5984 struct list_head *iter;
5985 int ret;
5987 for (iter = &dev->adj_list.lower,
5988 ldev = netdev_next_lower_dev_rcu(dev, &iter);
5989 ldev;
5990 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
5991 /* first is the lower device itself */
5992 ret = fn(ldev, data);
5993 if (ret)
5994 return ret;
5996 /* then look at all of its lower devices */
5997 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
5998 if (ret)
5999 return ret;
6002 return 0;
6004 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6007 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6008 * lower neighbour list, RCU
6009 * variant
6010 * @dev: device
6012 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6013 * list. The caller must hold RCU read lock.
6015 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6017 struct netdev_adjacent *lower;
6019 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6020 struct netdev_adjacent, list);
6021 if (lower)
6022 return lower->private;
6023 return NULL;
6025 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6028 * netdev_master_upper_dev_get_rcu - Get master upper device
6029 * @dev: device
6031 * Find a master upper device and return pointer to it or NULL in case
6032 * it's not there. The caller must hold the RCU read lock.
6034 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6036 struct netdev_adjacent *upper;
6038 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6039 struct netdev_adjacent, list);
6040 if (upper && likely(upper->master))
6041 return upper->dev;
6042 return NULL;
6044 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6046 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6047 struct net_device *adj_dev,
6048 struct list_head *dev_list)
6050 char linkname[IFNAMSIZ+7];
6052 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6053 "upper_%s" : "lower_%s", adj_dev->name);
6054 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6055 linkname);
6057 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6058 char *name,
6059 struct list_head *dev_list)
6061 char linkname[IFNAMSIZ+7];
6063 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6064 "upper_%s" : "lower_%s", name);
6065 sysfs_remove_link(&(dev->dev.kobj), linkname);
6068 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6069 struct net_device *adj_dev,
6070 struct list_head *dev_list)
6072 return (dev_list == &dev->adj_list.upper ||
6073 dev_list == &dev->adj_list.lower) &&
6074 net_eq(dev_net(dev), dev_net(adj_dev));
6077 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6078 struct net_device *adj_dev,
6079 struct list_head *dev_list,
6080 void *private, bool master)
6082 struct netdev_adjacent *adj;
6083 int ret;
6085 adj = __netdev_find_adj(adj_dev, dev_list);
6087 if (adj) {
6088 adj->ref_nr += 1;
6089 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6090 dev->name, adj_dev->name, adj->ref_nr);
6092 return 0;
6095 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6096 if (!adj)
6097 return -ENOMEM;
6099 adj->dev = adj_dev;
6100 adj->master = master;
6101 adj->ref_nr = 1;
6102 adj->private = private;
6103 dev_hold(adj_dev);
6105 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6106 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6108 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6109 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6110 if (ret)
6111 goto free_adj;
6114 /* Ensure that master link is always the first item in list. */
6115 if (master) {
6116 ret = sysfs_create_link(&(dev->dev.kobj),
6117 &(adj_dev->dev.kobj), "master");
6118 if (ret)
6119 goto remove_symlinks;
6121 list_add_rcu(&adj->list, dev_list);
6122 } else {
6123 list_add_tail_rcu(&adj->list, dev_list);
6126 return 0;
6128 remove_symlinks:
6129 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6130 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6131 free_adj:
6132 kfree(adj);
6133 dev_put(adj_dev);
6135 return ret;
6138 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6139 struct net_device *adj_dev,
6140 u16 ref_nr,
6141 struct list_head *dev_list)
6143 struct netdev_adjacent *adj;
6145 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6146 dev->name, adj_dev->name, ref_nr);
6148 adj = __netdev_find_adj(adj_dev, dev_list);
6150 if (!adj) {
6151 pr_err("Adjacency does not exist for device %s from %s\n",
6152 dev->name, adj_dev->name);
6153 WARN_ON(1);
6154 return;
6157 if (adj->ref_nr > ref_nr) {
6158 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6159 dev->name, adj_dev->name, ref_nr,
6160 adj->ref_nr - ref_nr);
6161 adj->ref_nr -= ref_nr;
6162 return;
6165 if (adj->master)
6166 sysfs_remove_link(&(dev->dev.kobj), "master");
6168 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6169 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6171 list_del_rcu(&adj->list);
6172 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6173 adj_dev->name, dev->name, adj_dev->name);
6174 dev_put(adj_dev);
6175 kfree_rcu(adj, rcu);
6178 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6179 struct net_device *upper_dev,
6180 struct list_head *up_list,
6181 struct list_head *down_list,
6182 void *private, bool master)
6184 int ret;
6186 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6187 private, master);
6188 if (ret)
6189 return ret;
6191 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6192 private, false);
6193 if (ret) {
6194 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6195 return ret;
6198 return 0;
6201 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6202 struct net_device *upper_dev,
6203 u16 ref_nr,
6204 struct list_head *up_list,
6205 struct list_head *down_list)
6207 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6208 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6211 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6212 struct net_device *upper_dev,
6213 void *private, bool master)
6215 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6216 &dev->adj_list.upper,
6217 &upper_dev->adj_list.lower,
6218 private, master);
6221 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6222 struct net_device *upper_dev)
6224 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6225 &dev->adj_list.upper,
6226 &upper_dev->adj_list.lower);
6229 static int __netdev_upper_dev_link(struct net_device *dev,
6230 struct net_device *upper_dev, bool master,
6231 void *upper_priv, void *upper_info)
6233 struct netdev_notifier_changeupper_info changeupper_info;
6234 int ret = 0;
6236 ASSERT_RTNL();
6238 if (dev == upper_dev)
6239 return -EBUSY;
6241 /* To prevent loops, check if dev is not upper device to upper_dev. */
6242 if (netdev_has_upper_dev(upper_dev, dev))
6243 return -EBUSY;
6245 if (netdev_has_upper_dev(dev, upper_dev))
6246 return -EEXIST;
6248 if (master && netdev_master_upper_dev_get(dev))
6249 return -EBUSY;
6251 changeupper_info.upper_dev = upper_dev;
6252 changeupper_info.master = master;
6253 changeupper_info.linking = true;
6254 changeupper_info.upper_info = upper_info;
6256 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6257 &changeupper_info.info);
6258 ret = notifier_to_errno(ret);
6259 if (ret)
6260 return ret;
6262 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6263 master);
6264 if (ret)
6265 return ret;
6267 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6268 &changeupper_info.info);
6269 ret = notifier_to_errno(ret);
6270 if (ret)
6271 goto rollback;
6273 return 0;
6275 rollback:
6276 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6278 return ret;
6282 * netdev_upper_dev_link - Add a link to the upper device
6283 * @dev: device
6284 * @upper_dev: new upper device
6286 * Adds a link to device which is upper to this one. The caller must hold
6287 * the RTNL lock. On a failure a negative errno code is returned.
6288 * On success the reference counts are adjusted and the function
6289 * returns zero.
6291 int netdev_upper_dev_link(struct net_device *dev,
6292 struct net_device *upper_dev)
6294 return __netdev_upper_dev_link(dev, upper_dev, false, NULL, NULL);
6296 EXPORT_SYMBOL(netdev_upper_dev_link);
6299 * netdev_master_upper_dev_link - Add a master link to the upper device
6300 * @dev: device
6301 * @upper_dev: new upper device
6302 * @upper_priv: upper device private
6303 * @upper_info: upper info to be passed down via notifier
6305 * Adds a link to device which is upper to this one. In this case, only
6306 * one master upper device can be linked, although other non-master devices
6307 * might be linked as well. The caller must hold the RTNL lock.
6308 * On a failure a negative errno code is returned. On success the reference
6309 * counts are adjusted and the function returns zero.
6311 int netdev_master_upper_dev_link(struct net_device *dev,
6312 struct net_device *upper_dev,
6313 void *upper_priv, void *upper_info)
6315 return __netdev_upper_dev_link(dev, upper_dev, true,
6316 upper_priv, upper_info);
6318 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6321 * netdev_upper_dev_unlink - Removes a link to upper device
6322 * @dev: device
6323 * @upper_dev: new upper device
6325 * Removes a link to device which is upper to this one. The caller must hold
6326 * the RTNL lock.
6328 void netdev_upper_dev_unlink(struct net_device *dev,
6329 struct net_device *upper_dev)
6331 struct netdev_notifier_changeupper_info changeupper_info;
6333 ASSERT_RTNL();
6335 changeupper_info.upper_dev = upper_dev;
6336 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6337 changeupper_info.linking = false;
6339 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6340 &changeupper_info.info);
6342 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6344 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6345 &changeupper_info.info);
6347 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6350 * netdev_bonding_info_change - Dispatch event about slave change
6351 * @dev: device
6352 * @bonding_info: info to dispatch
6354 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6355 * The caller must hold the RTNL lock.
6357 void netdev_bonding_info_change(struct net_device *dev,
6358 struct netdev_bonding_info *bonding_info)
6360 struct netdev_notifier_bonding_info info;
6362 memcpy(&info.bonding_info, bonding_info,
6363 sizeof(struct netdev_bonding_info));
6364 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, dev,
6365 &info.info);
6367 EXPORT_SYMBOL(netdev_bonding_info_change);
6369 static void netdev_adjacent_add_links(struct net_device *dev)
6371 struct netdev_adjacent *iter;
6373 struct net *net = dev_net(dev);
6375 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6376 if (!net_eq(net, dev_net(iter->dev)))
6377 continue;
6378 netdev_adjacent_sysfs_add(iter->dev, dev,
6379 &iter->dev->adj_list.lower);
6380 netdev_adjacent_sysfs_add(dev, iter->dev,
6381 &dev->adj_list.upper);
6384 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6385 if (!net_eq(net, dev_net(iter->dev)))
6386 continue;
6387 netdev_adjacent_sysfs_add(iter->dev, dev,
6388 &iter->dev->adj_list.upper);
6389 netdev_adjacent_sysfs_add(dev, iter->dev,
6390 &dev->adj_list.lower);
6394 static void netdev_adjacent_del_links(struct net_device *dev)
6396 struct netdev_adjacent *iter;
6398 struct net *net = dev_net(dev);
6400 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6401 if (!net_eq(net, dev_net(iter->dev)))
6402 continue;
6403 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6404 &iter->dev->adj_list.lower);
6405 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6406 &dev->adj_list.upper);
6409 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6410 if (!net_eq(net, dev_net(iter->dev)))
6411 continue;
6412 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6413 &iter->dev->adj_list.upper);
6414 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6415 &dev->adj_list.lower);
6419 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6421 struct netdev_adjacent *iter;
6423 struct net *net = dev_net(dev);
6425 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6426 if (!net_eq(net, dev_net(iter->dev)))
6427 continue;
6428 netdev_adjacent_sysfs_del(iter->dev, oldname,
6429 &iter->dev->adj_list.lower);
6430 netdev_adjacent_sysfs_add(iter->dev, dev,
6431 &iter->dev->adj_list.lower);
6434 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6435 if (!net_eq(net, dev_net(iter->dev)))
6436 continue;
6437 netdev_adjacent_sysfs_del(iter->dev, oldname,
6438 &iter->dev->adj_list.upper);
6439 netdev_adjacent_sysfs_add(iter->dev, dev,
6440 &iter->dev->adj_list.upper);
6444 void *netdev_lower_dev_get_private(struct net_device *dev,
6445 struct net_device *lower_dev)
6447 struct netdev_adjacent *lower;
6449 if (!lower_dev)
6450 return NULL;
6451 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6452 if (!lower)
6453 return NULL;
6455 return lower->private;
6457 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6460 int dev_get_nest_level(struct net_device *dev)
6462 struct net_device *lower = NULL;
6463 struct list_head *iter;
6464 int max_nest = -1;
6465 int nest;
6467 ASSERT_RTNL();
6469 netdev_for_each_lower_dev(dev, lower, iter) {
6470 nest = dev_get_nest_level(lower);
6471 if (max_nest < nest)
6472 max_nest = nest;
6475 return max_nest + 1;
6477 EXPORT_SYMBOL(dev_get_nest_level);
6480 * netdev_lower_change - Dispatch event about lower device state change
6481 * @lower_dev: device
6482 * @lower_state_info: state to dispatch
6484 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6485 * The caller must hold the RTNL lock.
6487 void netdev_lower_state_changed(struct net_device *lower_dev,
6488 void *lower_state_info)
6490 struct netdev_notifier_changelowerstate_info changelowerstate_info;
6492 ASSERT_RTNL();
6493 changelowerstate_info.lower_state_info = lower_state_info;
6494 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, lower_dev,
6495 &changelowerstate_info.info);
6497 EXPORT_SYMBOL(netdev_lower_state_changed);
6499 static void dev_change_rx_flags(struct net_device *dev, int flags)
6501 const struct net_device_ops *ops = dev->netdev_ops;
6503 if (ops->ndo_change_rx_flags)
6504 ops->ndo_change_rx_flags(dev, flags);
6507 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6509 unsigned int old_flags = dev->flags;
6510 kuid_t uid;
6511 kgid_t gid;
6513 ASSERT_RTNL();
6515 dev->flags |= IFF_PROMISC;
6516 dev->promiscuity += inc;
6517 if (dev->promiscuity == 0) {
6519 * Avoid overflow.
6520 * If inc causes overflow, untouch promisc and return error.
6522 if (inc < 0)
6523 dev->flags &= ~IFF_PROMISC;
6524 else {
6525 dev->promiscuity -= inc;
6526 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6527 dev->name);
6528 return -EOVERFLOW;
6531 if (dev->flags != old_flags) {
6532 pr_info("device %s %s promiscuous mode\n",
6533 dev->name,
6534 dev->flags & IFF_PROMISC ? "entered" : "left");
6535 if (audit_enabled) {
6536 current_uid_gid(&uid, &gid);
6537 audit_log(current->audit_context, GFP_ATOMIC,
6538 AUDIT_ANOM_PROMISCUOUS,
6539 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6540 dev->name, (dev->flags & IFF_PROMISC),
6541 (old_flags & IFF_PROMISC),
6542 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6543 from_kuid(&init_user_ns, uid),
6544 from_kgid(&init_user_ns, gid),
6545 audit_get_sessionid(current));
6548 dev_change_rx_flags(dev, IFF_PROMISC);
6550 if (notify)
6551 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6552 return 0;
6556 * dev_set_promiscuity - update promiscuity count on a device
6557 * @dev: device
6558 * @inc: modifier
6560 * Add or remove promiscuity from a device. While the count in the device
6561 * remains above zero the interface remains promiscuous. Once it hits zero
6562 * the device reverts back to normal filtering operation. A negative inc
6563 * value is used to drop promiscuity on the device.
6564 * Return 0 if successful or a negative errno code on error.
6566 int dev_set_promiscuity(struct net_device *dev, int inc)
6568 unsigned int old_flags = dev->flags;
6569 int err;
6571 err = __dev_set_promiscuity(dev, inc, true);
6572 if (err < 0)
6573 return err;
6574 if (dev->flags != old_flags)
6575 dev_set_rx_mode(dev);
6576 return err;
6578 EXPORT_SYMBOL(dev_set_promiscuity);
6580 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6582 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6584 ASSERT_RTNL();
6586 dev->flags |= IFF_ALLMULTI;
6587 dev->allmulti += inc;
6588 if (dev->allmulti == 0) {
6590 * Avoid overflow.
6591 * If inc causes overflow, untouch allmulti and return error.
6593 if (inc < 0)
6594 dev->flags &= ~IFF_ALLMULTI;
6595 else {
6596 dev->allmulti -= inc;
6597 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6598 dev->name);
6599 return -EOVERFLOW;
6602 if (dev->flags ^ old_flags) {
6603 dev_change_rx_flags(dev, IFF_ALLMULTI);
6604 dev_set_rx_mode(dev);
6605 if (notify)
6606 __dev_notify_flags(dev, old_flags,
6607 dev->gflags ^ old_gflags);
6609 return 0;
6613 * dev_set_allmulti - update allmulti count on a device
6614 * @dev: device
6615 * @inc: modifier
6617 * Add or remove reception of all multicast frames to a device. While the
6618 * count in the device remains above zero the interface remains listening
6619 * to all interfaces. Once it hits zero the device reverts back to normal
6620 * filtering operation. A negative @inc value is used to drop the counter
6621 * when releasing a resource needing all multicasts.
6622 * Return 0 if successful or a negative errno code on error.
6625 int dev_set_allmulti(struct net_device *dev, int inc)
6627 return __dev_set_allmulti(dev, inc, true);
6629 EXPORT_SYMBOL(dev_set_allmulti);
6632 * Upload unicast and multicast address lists to device and
6633 * configure RX filtering. When the device doesn't support unicast
6634 * filtering it is put in promiscuous mode while unicast addresses
6635 * are present.
6637 void __dev_set_rx_mode(struct net_device *dev)
6639 const struct net_device_ops *ops = dev->netdev_ops;
6641 /* dev_open will call this function so the list will stay sane. */
6642 if (!(dev->flags&IFF_UP))
6643 return;
6645 if (!netif_device_present(dev))
6646 return;
6648 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6649 /* Unicast addresses changes may only happen under the rtnl,
6650 * therefore calling __dev_set_promiscuity here is safe.
6652 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6653 __dev_set_promiscuity(dev, 1, false);
6654 dev->uc_promisc = true;
6655 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6656 __dev_set_promiscuity(dev, -1, false);
6657 dev->uc_promisc = false;
6661 if (ops->ndo_set_rx_mode)
6662 ops->ndo_set_rx_mode(dev);
6665 void dev_set_rx_mode(struct net_device *dev)
6667 netif_addr_lock_bh(dev);
6668 __dev_set_rx_mode(dev);
6669 netif_addr_unlock_bh(dev);
6673 * dev_get_flags - get flags reported to userspace
6674 * @dev: device
6676 * Get the combination of flag bits exported through APIs to userspace.
6678 unsigned int dev_get_flags(const struct net_device *dev)
6680 unsigned int flags;
6682 flags = (dev->flags & ~(IFF_PROMISC |
6683 IFF_ALLMULTI |
6684 IFF_RUNNING |
6685 IFF_LOWER_UP |
6686 IFF_DORMANT)) |
6687 (dev->gflags & (IFF_PROMISC |
6688 IFF_ALLMULTI));
6690 if (netif_running(dev)) {
6691 if (netif_oper_up(dev))
6692 flags |= IFF_RUNNING;
6693 if (netif_carrier_ok(dev))
6694 flags |= IFF_LOWER_UP;
6695 if (netif_dormant(dev))
6696 flags |= IFF_DORMANT;
6699 return flags;
6701 EXPORT_SYMBOL(dev_get_flags);
6703 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6705 unsigned int old_flags = dev->flags;
6706 int ret;
6708 ASSERT_RTNL();
6711 * Set the flags on our device.
6714 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6715 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6716 IFF_AUTOMEDIA)) |
6717 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6718 IFF_ALLMULTI));
6721 * Load in the correct multicast list now the flags have changed.
6724 if ((old_flags ^ flags) & IFF_MULTICAST)
6725 dev_change_rx_flags(dev, IFF_MULTICAST);
6727 dev_set_rx_mode(dev);
6730 * Have we downed the interface. We handle IFF_UP ourselves
6731 * according to user attempts to set it, rather than blindly
6732 * setting it.
6735 ret = 0;
6736 if ((old_flags ^ flags) & IFF_UP) {
6737 if (old_flags & IFF_UP)
6738 __dev_close(dev);
6739 else
6740 ret = __dev_open(dev);
6743 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6744 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6745 unsigned int old_flags = dev->flags;
6747 dev->gflags ^= IFF_PROMISC;
6749 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6750 if (dev->flags != old_flags)
6751 dev_set_rx_mode(dev);
6754 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6755 * is important. Some (broken) drivers set IFF_PROMISC, when
6756 * IFF_ALLMULTI is requested not asking us and not reporting.
6758 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6759 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6761 dev->gflags ^= IFF_ALLMULTI;
6762 __dev_set_allmulti(dev, inc, false);
6765 return ret;
6768 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6769 unsigned int gchanges)
6771 unsigned int changes = dev->flags ^ old_flags;
6773 if (gchanges)
6774 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
6776 if (changes & IFF_UP) {
6777 if (dev->flags & IFF_UP)
6778 call_netdevice_notifiers(NETDEV_UP, dev);
6779 else
6780 call_netdevice_notifiers(NETDEV_DOWN, dev);
6783 if (dev->flags & IFF_UP &&
6784 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
6785 struct netdev_notifier_change_info change_info;
6787 change_info.flags_changed = changes;
6788 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
6789 &change_info.info);
6794 * dev_change_flags - change device settings
6795 * @dev: device
6796 * @flags: device state flags
6798 * Change settings on device based state flags. The flags are
6799 * in the userspace exported format.
6801 int dev_change_flags(struct net_device *dev, unsigned int flags)
6803 int ret;
6804 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
6806 ret = __dev_change_flags(dev, flags);
6807 if (ret < 0)
6808 return ret;
6810 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
6811 __dev_notify_flags(dev, old_flags, changes);
6812 return ret;
6814 EXPORT_SYMBOL(dev_change_flags);
6816 int __dev_set_mtu(struct net_device *dev, int new_mtu)
6818 const struct net_device_ops *ops = dev->netdev_ops;
6820 if (ops->ndo_change_mtu)
6821 return ops->ndo_change_mtu(dev, new_mtu);
6823 dev->mtu = new_mtu;
6824 return 0;
6826 EXPORT_SYMBOL(__dev_set_mtu);
6829 * dev_set_mtu - Change maximum transfer unit
6830 * @dev: device
6831 * @new_mtu: new transfer unit
6833 * Change the maximum transfer size of the network device.
6835 int dev_set_mtu(struct net_device *dev, int new_mtu)
6837 int err, orig_mtu;
6839 if (new_mtu == dev->mtu)
6840 return 0;
6842 /* MTU must be positive, and in range */
6843 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
6844 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
6845 dev->name, new_mtu, dev->min_mtu);
6846 return -EINVAL;
6849 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
6850 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
6851 dev->name, new_mtu, dev->max_mtu);
6852 return -EINVAL;
6855 if (!netif_device_present(dev))
6856 return -ENODEV;
6858 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
6859 err = notifier_to_errno(err);
6860 if (err)
6861 return err;
6863 orig_mtu = dev->mtu;
6864 err = __dev_set_mtu(dev, new_mtu);
6866 if (!err) {
6867 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6868 err = notifier_to_errno(err);
6869 if (err) {
6870 /* setting mtu back and notifying everyone again,
6871 * so that they have a chance to revert changes.
6873 __dev_set_mtu(dev, orig_mtu);
6874 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6877 return err;
6879 EXPORT_SYMBOL(dev_set_mtu);
6882 * dev_set_group - Change group this device belongs to
6883 * @dev: device
6884 * @new_group: group this device should belong to
6886 void dev_set_group(struct net_device *dev, int new_group)
6888 dev->group = new_group;
6890 EXPORT_SYMBOL(dev_set_group);
6893 * dev_set_mac_address - Change Media Access Control Address
6894 * @dev: device
6895 * @sa: new address
6897 * Change the hardware (MAC) address of the device
6899 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
6901 const struct net_device_ops *ops = dev->netdev_ops;
6902 int err;
6904 if (!ops->ndo_set_mac_address)
6905 return -EOPNOTSUPP;
6906 if (sa->sa_family != dev->type)
6907 return -EINVAL;
6908 if (!netif_device_present(dev))
6909 return -ENODEV;
6910 err = ops->ndo_set_mac_address(dev, sa);
6911 if (err)
6912 return err;
6913 dev->addr_assign_type = NET_ADDR_SET;
6914 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
6915 add_device_randomness(dev->dev_addr, dev->addr_len);
6916 return 0;
6918 EXPORT_SYMBOL(dev_set_mac_address);
6921 * dev_change_carrier - Change device carrier
6922 * @dev: device
6923 * @new_carrier: new value
6925 * Change device carrier
6927 int dev_change_carrier(struct net_device *dev, bool new_carrier)
6929 const struct net_device_ops *ops = dev->netdev_ops;
6931 if (!ops->ndo_change_carrier)
6932 return -EOPNOTSUPP;
6933 if (!netif_device_present(dev))
6934 return -ENODEV;
6935 return ops->ndo_change_carrier(dev, new_carrier);
6937 EXPORT_SYMBOL(dev_change_carrier);
6940 * dev_get_phys_port_id - Get device physical port ID
6941 * @dev: device
6942 * @ppid: port ID
6944 * Get device physical port ID
6946 int dev_get_phys_port_id(struct net_device *dev,
6947 struct netdev_phys_item_id *ppid)
6949 const struct net_device_ops *ops = dev->netdev_ops;
6951 if (!ops->ndo_get_phys_port_id)
6952 return -EOPNOTSUPP;
6953 return ops->ndo_get_phys_port_id(dev, ppid);
6955 EXPORT_SYMBOL(dev_get_phys_port_id);
6958 * dev_get_phys_port_name - Get device physical port name
6959 * @dev: device
6960 * @name: port name
6961 * @len: limit of bytes to copy to name
6963 * Get device physical port name
6965 int dev_get_phys_port_name(struct net_device *dev,
6966 char *name, size_t len)
6968 const struct net_device_ops *ops = dev->netdev_ops;
6970 if (!ops->ndo_get_phys_port_name)
6971 return -EOPNOTSUPP;
6972 return ops->ndo_get_phys_port_name(dev, name, len);
6974 EXPORT_SYMBOL(dev_get_phys_port_name);
6977 * dev_change_proto_down - update protocol port state information
6978 * @dev: device
6979 * @proto_down: new value
6981 * This info can be used by switch drivers to set the phys state of the
6982 * port.
6984 int dev_change_proto_down(struct net_device *dev, bool proto_down)
6986 const struct net_device_ops *ops = dev->netdev_ops;
6988 if (!ops->ndo_change_proto_down)
6989 return -EOPNOTSUPP;
6990 if (!netif_device_present(dev))
6991 return -ENODEV;
6992 return ops->ndo_change_proto_down(dev, proto_down);
6994 EXPORT_SYMBOL(dev_change_proto_down);
6996 u8 __dev_xdp_attached(struct net_device *dev, xdp_op_t xdp_op, u32 *prog_id)
6998 struct netdev_xdp xdp;
7000 memset(&xdp, 0, sizeof(xdp));
7001 xdp.command = XDP_QUERY_PROG;
7003 /* Query must always succeed. */
7004 WARN_ON(xdp_op(dev, &xdp) < 0);
7005 if (prog_id)
7006 *prog_id = xdp.prog_id;
7008 return xdp.prog_attached;
7011 static int dev_xdp_install(struct net_device *dev, xdp_op_t xdp_op,
7012 struct netlink_ext_ack *extack, u32 flags,
7013 struct bpf_prog *prog)
7015 struct netdev_xdp xdp;
7017 memset(&xdp, 0, sizeof(xdp));
7018 if (flags & XDP_FLAGS_HW_MODE)
7019 xdp.command = XDP_SETUP_PROG_HW;
7020 else
7021 xdp.command = XDP_SETUP_PROG;
7022 xdp.extack = extack;
7023 xdp.flags = flags;
7024 xdp.prog = prog;
7026 return xdp_op(dev, &xdp);
7030 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7031 * @dev: device
7032 * @extack: netlink extended ack
7033 * @fd: new program fd or negative value to clear
7034 * @flags: xdp-related flags
7036 * Set or clear a bpf program for a device
7038 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7039 int fd, u32 flags)
7041 const struct net_device_ops *ops = dev->netdev_ops;
7042 struct bpf_prog *prog = NULL;
7043 xdp_op_t xdp_op, xdp_chk;
7044 int err;
7046 ASSERT_RTNL();
7048 xdp_op = xdp_chk = ops->ndo_xdp;
7049 if (!xdp_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7050 return -EOPNOTSUPP;
7051 if (!xdp_op || (flags & XDP_FLAGS_SKB_MODE))
7052 xdp_op = generic_xdp_install;
7053 if (xdp_op == xdp_chk)
7054 xdp_chk = generic_xdp_install;
7056 if (fd >= 0) {
7057 if (xdp_chk && __dev_xdp_attached(dev, xdp_chk, NULL))
7058 return -EEXIST;
7059 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7060 __dev_xdp_attached(dev, xdp_op, NULL))
7061 return -EBUSY;
7063 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
7064 if (IS_ERR(prog))
7065 return PTR_ERR(prog);
7068 err = dev_xdp_install(dev, xdp_op, extack, flags, prog);
7069 if (err < 0 && prog)
7070 bpf_prog_put(prog);
7072 return err;
7076 * dev_new_index - allocate an ifindex
7077 * @net: the applicable net namespace
7079 * Returns a suitable unique value for a new device interface
7080 * number. The caller must hold the rtnl semaphore or the
7081 * dev_base_lock to be sure it remains unique.
7083 static int dev_new_index(struct net *net)
7085 int ifindex = net->ifindex;
7087 for (;;) {
7088 if (++ifindex <= 0)
7089 ifindex = 1;
7090 if (!__dev_get_by_index(net, ifindex))
7091 return net->ifindex = ifindex;
7095 /* Delayed registration/unregisteration */
7096 static LIST_HEAD(net_todo_list);
7097 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7099 static void net_set_todo(struct net_device *dev)
7101 list_add_tail(&dev->todo_list, &net_todo_list);
7102 dev_net(dev)->dev_unreg_count++;
7105 static void rollback_registered_many(struct list_head *head)
7107 struct net_device *dev, *tmp;
7108 LIST_HEAD(close_head);
7110 BUG_ON(dev_boot_phase);
7111 ASSERT_RTNL();
7113 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7114 /* Some devices call without registering
7115 * for initialization unwind. Remove those
7116 * devices and proceed with the remaining.
7118 if (dev->reg_state == NETREG_UNINITIALIZED) {
7119 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7120 dev->name, dev);
7122 WARN_ON(1);
7123 list_del(&dev->unreg_list);
7124 continue;
7126 dev->dismantle = true;
7127 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7130 /* If device is running, close it first. */
7131 list_for_each_entry(dev, head, unreg_list)
7132 list_add_tail(&dev->close_list, &close_head);
7133 dev_close_many(&close_head, true);
7135 list_for_each_entry(dev, head, unreg_list) {
7136 /* And unlink it from device chain. */
7137 unlist_netdevice(dev);
7139 dev->reg_state = NETREG_UNREGISTERING;
7141 flush_all_backlogs();
7143 synchronize_net();
7145 list_for_each_entry(dev, head, unreg_list) {
7146 struct sk_buff *skb = NULL;
7148 /* Shutdown queueing discipline. */
7149 dev_shutdown(dev);
7152 /* Notify protocols, that we are about to destroy
7153 * this device. They should clean all the things.
7155 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7157 if (!dev->rtnl_link_ops ||
7158 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7159 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7160 GFP_KERNEL);
7163 * Flush the unicast and multicast chains
7165 dev_uc_flush(dev);
7166 dev_mc_flush(dev);
7168 if (dev->netdev_ops->ndo_uninit)
7169 dev->netdev_ops->ndo_uninit(dev);
7171 if (skb)
7172 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7174 /* Notifier chain MUST detach us all upper devices. */
7175 WARN_ON(netdev_has_any_upper_dev(dev));
7176 WARN_ON(netdev_has_any_lower_dev(dev));
7178 /* Remove entries from kobject tree */
7179 netdev_unregister_kobject(dev);
7180 #ifdef CONFIG_XPS
7181 /* Remove XPS queueing entries */
7182 netif_reset_xps_queues_gt(dev, 0);
7183 #endif
7186 synchronize_net();
7188 list_for_each_entry(dev, head, unreg_list)
7189 dev_put(dev);
7192 static void rollback_registered(struct net_device *dev)
7194 LIST_HEAD(single);
7196 list_add(&dev->unreg_list, &single);
7197 rollback_registered_many(&single);
7198 list_del(&single);
7201 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7202 struct net_device *upper, netdev_features_t features)
7204 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7205 netdev_features_t feature;
7206 int feature_bit;
7208 for_each_netdev_feature(&upper_disables, feature_bit) {
7209 feature = __NETIF_F_BIT(feature_bit);
7210 if (!(upper->wanted_features & feature)
7211 && (features & feature)) {
7212 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7213 &feature, upper->name);
7214 features &= ~feature;
7218 return features;
7221 static void netdev_sync_lower_features(struct net_device *upper,
7222 struct net_device *lower, netdev_features_t features)
7224 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7225 netdev_features_t feature;
7226 int feature_bit;
7228 for_each_netdev_feature(&upper_disables, feature_bit) {
7229 feature = __NETIF_F_BIT(feature_bit);
7230 if (!(features & feature) && (lower->features & feature)) {
7231 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7232 &feature, lower->name);
7233 lower->wanted_features &= ~feature;
7234 netdev_update_features(lower);
7236 if (unlikely(lower->features & feature))
7237 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7238 &feature, lower->name);
7243 static netdev_features_t netdev_fix_features(struct net_device *dev,
7244 netdev_features_t features)
7246 /* Fix illegal checksum combinations */
7247 if ((features & NETIF_F_HW_CSUM) &&
7248 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7249 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7250 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7253 /* TSO requires that SG is present as well. */
7254 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7255 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7256 features &= ~NETIF_F_ALL_TSO;
7259 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7260 !(features & NETIF_F_IP_CSUM)) {
7261 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7262 features &= ~NETIF_F_TSO;
7263 features &= ~NETIF_F_TSO_ECN;
7266 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7267 !(features & NETIF_F_IPV6_CSUM)) {
7268 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7269 features &= ~NETIF_F_TSO6;
7272 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7273 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7274 features &= ~NETIF_F_TSO_MANGLEID;
7276 /* TSO ECN requires that TSO is present as well. */
7277 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7278 features &= ~NETIF_F_TSO_ECN;
7280 /* Software GSO depends on SG. */
7281 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7282 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7283 features &= ~NETIF_F_GSO;
7286 /* GSO partial features require GSO partial be set */
7287 if ((features & dev->gso_partial_features) &&
7288 !(features & NETIF_F_GSO_PARTIAL)) {
7289 netdev_dbg(dev,
7290 "Dropping partially supported GSO features since no GSO partial.\n");
7291 features &= ~dev->gso_partial_features;
7294 return features;
7297 int __netdev_update_features(struct net_device *dev)
7299 struct net_device *upper, *lower;
7300 netdev_features_t features;
7301 struct list_head *iter;
7302 int err = -1;
7304 ASSERT_RTNL();
7306 features = netdev_get_wanted_features(dev);
7308 if (dev->netdev_ops->ndo_fix_features)
7309 features = dev->netdev_ops->ndo_fix_features(dev, features);
7311 /* driver might be less strict about feature dependencies */
7312 features = netdev_fix_features(dev, features);
7314 /* some features can't be enabled if they're off an an upper device */
7315 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7316 features = netdev_sync_upper_features(dev, upper, features);
7318 if (dev->features == features)
7319 goto sync_lower;
7321 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7322 &dev->features, &features);
7324 if (dev->netdev_ops->ndo_set_features)
7325 err = dev->netdev_ops->ndo_set_features(dev, features);
7326 else
7327 err = 0;
7329 if (unlikely(err < 0)) {
7330 netdev_err(dev,
7331 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7332 err, &features, &dev->features);
7333 /* return non-0 since some features might have changed and
7334 * it's better to fire a spurious notification than miss it
7336 return -1;
7339 sync_lower:
7340 /* some features must be disabled on lower devices when disabled
7341 * on an upper device (think: bonding master or bridge)
7343 netdev_for_each_lower_dev(dev, lower, iter)
7344 netdev_sync_lower_features(dev, lower, features);
7346 if (!err) {
7347 netdev_features_t diff = features ^ dev->features;
7349 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7350 /* udp_tunnel_{get,drop}_rx_info both need
7351 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7352 * device, or they won't do anything.
7353 * Thus we need to update dev->features
7354 * *before* calling udp_tunnel_get_rx_info,
7355 * but *after* calling udp_tunnel_drop_rx_info.
7357 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7358 dev->features = features;
7359 udp_tunnel_get_rx_info(dev);
7360 } else {
7361 udp_tunnel_drop_rx_info(dev);
7365 dev->features = features;
7368 return err < 0 ? 0 : 1;
7372 * netdev_update_features - recalculate device features
7373 * @dev: the device to check
7375 * Recalculate dev->features set and send notifications if it
7376 * has changed. Should be called after driver or hardware dependent
7377 * conditions might have changed that influence the features.
7379 void netdev_update_features(struct net_device *dev)
7381 if (__netdev_update_features(dev))
7382 netdev_features_change(dev);
7384 EXPORT_SYMBOL(netdev_update_features);
7387 * netdev_change_features - recalculate device features
7388 * @dev: the device to check
7390 * Recalculate dev->features set and send notifications even
7391 * if they have not changed. Should be called instead of
7392 * netdev_update_features() if also dev->vlan_features might
7393 * have changed to allow the changes to be propagated to stacked
7394 * VLAN devices.
7396 void netdev_change_features(struct net_device *dev)
7398 __netdev_update_features(dev);
7399 netdev_features_change(dev);
7401 EXPORT_SYMBOL(netdev_change_features);
7404 * netif_stacked_transfer_operstate - transfer operstate
7405 * @rootdev: the root or lower level device to transfer state from
7406 * @dev: the device to transfer operstate to
7408 * Transfer operational state from root to device. This is normally
7409 * called when a stacking relationship exists between the root
7410 * device and the device(a leaf device).
7412 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7413 struct net_device *dev)
7415 if (rootdev->operstate == IF_OPER_DORMANT)
7416 netif_dormant_on(dev);
7417 else
7418 netif_dormant_off(dev);
7420 if (netif_carrier_ok(rootdev))
7421 netif_carrier_on(dev);
7422 else
7423 netif_carrier_off(dev);
7425 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7427 #ifdef CONFIG_SYSFS
7428 static int netif_alloc_rx_queues(struct net_device *dev)
7430 unsigned int i, count = dev->num_rx_queues;
7431 struct netdev_rx_queue *rx;
7432 size_t sz = count * sizeof(*rx);
7434 BUG_ON(count < 1);
7436 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7437 if (!rx)
7438 return -ENOMEM;
7440 dev->_rx = rx;
7442 for (i = 0; i < count; i++)
7443 rx[i].dev = dev;
7444 return 0;
7446 #endif
7448 static void netdev_init_one_queue(struct net_device *dev,
7449 struct netdev_queue *queue, void *_unused)
7451 /* Initialize queue lock */
7452 spin_lock_init(&queue->_xmit_lock);
7453 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7454 queue->xmit_lock_owner = -1;
7455 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7456 queue->dev = dev;
7457 #ifdef CONFIG_BQL
7458 dql_init(&queue->dql, HZ);
7459 #endif
7462 static void netif_free_tx_queues(struct net_device *dev)
7464 kvfree(dev->_tx);
7467 static int netif_alloc_netdev_queues(struct net_device *dev)
7469 unsigned int count = dev->num_tx_queues;
7470 struct netdev_queue *tx;
7471 size_t sz = count * sizeof(*tx);
7473 if (count < 1 || count > 0xffff)
7474 return -EINVAL;
7476 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7477 if (!tx)
7478 return -ENOMEM;
7480 dev->_tx = tx;
7482 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7483 spin_lock_init(&dev->tx_global_lock);
7485 return 0;
7488 void netif_tx_stop_all_queues(struct net_device *dev)
7490 unsigned int i;
7492 for (i = 0; i < dev->num_tx_queues; i++) {
7493 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7495 netif_tx_stop_queue(txq);
7498 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7501 * register_netdevice - register a network device
7502 * @dev: device to register
7504 * Take a completed network device structure and add it to the kernel
7505 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7506 * chain. 0 is returned on success. A negative errno code is returned
7507 * on a failure to set up the device, or if the name is a duplicate.
7509 * Callers must hold the rtnl semaphore. You may want
7510 * register_netdev() instead of this.
7512 * BUGS:
7513 * The locking appears insufficient to guarantee two parallel registers
7514 * will not get the same name.
7517 int register_netdevice(struct net_device *dev)
7519 int ret;
7520 struct net *net = dev_net(dev);
7522 BUG_ON(dev_boot_phase);
7523 ASSERT_RTNL();
7525 might_sleep();
7527 /* When net_device's are persistent, this will be fatal. */
7528 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7529 BUG_ON(!net);
7531 spin_lock_init(&dev->addr_list_lock);
7532 netdev_set_addr_lockdep_class(dev);
7534 ret = dev_get_valid_name(net, dev, dev->name);
7535 if (ret < 0)
7536 goto out;
7538 /* Init, if this function is available */
7539 if (dev->netdev_ops->ndo_init) {
7540 ret = dev->netdev_ops->ndo_init(dev);
7541 if (ret) {
7542 if (ret > 0)
7543 ret = -EIO;
7544 goto out;
7548 if (((dev->hw_features | dev->features) &
7549 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7550 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7551 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7552 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7553 ret = -EINVAL;
7554 goto err_uninit;
7557 ret = -EBUSY;
7558 if (!dev->ifindex)
7559 dev->ifindex = dev_new_index(net);
7560 else if (__dev_get_by_index(net, dev->ifindex))
7561 goto err_uninit;
7563 /* Transfer changeable features to wanted_features and enable
7564 * software offloads (GSO and GRO).
7566 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7567 dev->features |= NETIF_F_SOFT_FEATURES;
7569 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7570 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7571 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7574 dev->wanted_features = dev->features & dev->hw_features;
7576 if (!(dev->flags & IFF_LOOPBACK))
7577 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7579 /* If IPv4 TCP segmentation offload is supported we should also
7580 * allow the device to enable segmenting the frame with the option
7581 * of ignoring a static IP ID value. This doesn't enable the
7582 * feature itself but allows the user to enable it later.
7584 if (dev->hw_features & NETIF_F_TSO)
7585 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7586 if (dev->vlan_features & NETIF_F_TSO)
7587 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7588 if (dev->mpls_features & NETIF_F_TSO)
7589 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7590 if (dev->hw_enc_features & NETIF_F_TSO)
7591 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7593 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7595 dev->vlan_features |= NETIF_F_HIGHDMA;
7597 /* Make NETIF_F_SG inheritable to tunnel devices.
7599 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7601 /* Make NETIF_F_SG inheritable to MPLS.
7603 dev->mpls_features |= NETIF_F_SG;
7605 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7606 ret = notifier_to_errno(ret);
7607 if (ret)
7608 goto err_uninit;
7610 ret = netdev_register_kobject(dev);
7611 if (ret)
7612 goto err_uninit;
7613 dev->reg_state = NETREG_REGISTERED;
7615 __netdev_update_features(dev);
7618 * Default initial state at registry is that the
7619 * device is present.
7622 set_bit(__LINK_STATE_PRESENT, &dev->state);
7624 linkwatch_init_dev(dev);
7626 dev_init_scheduler(dev);
7627 dev_hold(dev);
7628 list_netdevice(dev);
7629 add_device_randomness(dev->dev_addr, dev->addr_len);
7631 /* If the device has permanent device address, driver should
7632 * set dev_addr and also addr_assign_type should be set to
7633 * NET_ADDR_PERM (default value).
7635 if (dev->addr_assign_type == NET_ADDR_PERM)
7636 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7638 /* Notify protocols, that a new device appeared. */
7639 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
7640 ret = notifier_to_errno(ret);
7641 if (ret) {
7642 rollback_registered(dev);
7643 dev->reg_state = NETREG_UNREGISTERED;
7646 * Prevent userspace races by waiting until the network
7647 * device is fully setup before sending notifications.
7649 if (!dev->rtnl_link_ops ||
7650 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7651 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
7653 out:
7654 return ret;
7656 err_uninit:
7657 if (dev->netdev_ops->ndo_uninit)
7658 dev->netdev_ops->ndo_uninit(dev);
7659 if (dev->priv_destructor)
7660 dev->priv_destructor(dev);
7661 goto out;
7663 EXPORT_SYMBOL(register_netdevice);
7666 * init_dummy_netdev - init a dummy network device for NAPI
7667 * @dev: device to init
7669 * This takes a network device structure and initialize the minimum
7670 * amount of fields so it can be used to schedule NAPI polls without
7671 * registering a full blown interface. This is to be used by drivers
7672 * that need to tie several hardware interfaces to a single NAPI
7673 * poll scheduler due to HW limitations.
7675 int init_dummy_netdev(struct net_device *dev)
7677 /* Clear everything. Note we don't initialize spinlocks
7678 * are they aren't supposed to be taken by any of the
7679 * NAPI code and this dummy netdev is supposed to be
7680 * only ever used for NAPI polls
7682 memset(dev, 0, sizeof(struct net_device));
7684 /* make sure we BUG if trying to hit standard
7685 * register/unregister code path
7687 dev->reg_state = NETREG_DUMMY;
7689 /* NAPI wants this */
7690 INIT_LIST_HEAD(&dev->napi_list);
7692 /* a dummy interface is started by default */
7693 set_bit(__LINK_STATE_PRESENT, &dev->state);
7694 set_bit(__LINK_STATE_START, &dev->state);
7696 /* Note : We dont allocate pcpu_refcnt for dummy devices,
7697 * because users of this 'device' dont need to change
7698 * its refcount.
7701 return 0;
7703 EXPORT_SYMBOL_GPL(init_dummy_netdev);
7707 * register_netdev - register a network device
7708 * @dev: device to register
7710 * Take a completed network device structure and add it to the kernel
7711 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7712 * chain. 0 is returned on success. A negative errno code is returned
7713 * on a failure to set up the device, or if the name is a duplicate.
7715 * This is a wrapper around register_netdevice that takes the rtnl semaphore
7716 * and expands the device name if you passed a format string to
7717 * alloc_netdev.
7719 int register_netdev(struct net_device *dev)
7721 int err;
7723 rtnl_lock();
7724 err = register_netdevice(dev);
7725 rtnl_unlock();
7726 return err;
7728 EXPORT_SYMBOL(register_netdev);
7730 int netdev_refcnt_read(const struct net_device *dev)
7732 int i, refcnt = 0;
7734 for_each_possible_cpu(i)
7735 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
7736 return refcnt;
7738 EXPORT_SYMBOL(netdev_refcnt_read);
7741 * netdev_wait_allrefs - wait until all references are gone.
7742 * @dev: target net_device
7744 * This is called when unregistering network devices.
7746 * Any protocol or device that holds a reference should register
7747 * for netdevice notification, and cleanup and put back the
7748 * reference if they receive an UNREGISTER event.
7749 * We can get stuck here if buggy protocols don't correctly
7750 * call dev_put.
7752 static void netdev_wait_allrefs(struct net_device *dev)
7754 unsigned long rebroadcast_time, warning_time;
7755 int refcnt;
7757 linkwatch_forget_dev(dev);
7759 rebroadcast_time = warning_time = jiffies;
7760 refcnt = netdev_refcnt_read(dev);
7762 while (refcnt != 0) {
7763 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
7764 rtnl_lock();
7766 /* Rebroadcast unregister notification */
7767 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7769 __rtnl_unlock();
7770 rcu_barrier();
7771 rtnl_lock();
7773 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7774 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
7775 &dev->state)) {
7776 /* We must not have linkwatch events
7777 * pending on unregister. If this
7778 * happens, we simply run the queue
7779 * unscheduled, resulting in a noop
7780 * for this device.
7782 linkwatch_run_queue();
7785 __rtnl_unlock();
7787 rebroadcast_time = jiffies;
7790 msleep(250);
7792 refcnt = netdev_refcnt_read(dev);
7794 if (time_after(jiffies, warning_time + 10 * HZ)) {
7795 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
7796 dev->name, refcnt);
7797 warning_time = jiffies;
7802 /* The sequence is:
7804 * rtnl_lock();
7805 * ...
7806 * register_netdevice(x1);
7807 * register_netdevice(x2);
7808 * ...
7809 * unregister_netdevice(y1);
7810 * unregister_netdevice(y2);
7811 * ...
7812 * rtnl_unlock();
7813 * free_netdev(y1);
7814 * free_netdev(y2);
7816 * We are invoked by rtnl_unlock().
7817 * This allows us to deal with problems:
7818 * 1) We can delete sysfs objects which invoke hotplug
7819 * without deadlocking with linkwatch via keventd.
7820 * 2) Since we run with the RTNL semaphore not held, we can sleep
7821 * safely in order to wait for the netdev refcnt to drop to zero.
7823 * We must not return until all unregister events added during
7824 * the interval the lock was held have been completed.
7826 void netdev_run_todo(void)
7828 struct list_head list;
7830 /* Snapshot list, allow later requests */
7831 list_replace_init(&net_todo_list, &list);
7833 __rtnl_unlock();
7836 /* Wait for rcu callbacks to finish before next phase */
7837 if (!list_empty(&list))
7838 rcu_barrier();
7840 while (!list_empty(&list)) {
7841 struct net_device *dev
7842 = list_first_entry(&list, struct net_device, todo_list);
7843 list_del(&dev->todo_list);
7845 rtnl_lock();
7846 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7847 __rtnl_unlock();
7849 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
7850 pr_err("network todo '%s' but state %d\n",
7851 dev->name, dev->reg_state);
7852 dump_stack();
7853 continue;
7856 dev->reg_state = NETREG_UNREGISTERED;
7858 netdev_wait_allrefs(dev);
7860 /* paranoia */
7861 BUG_ON(netdev_refcnt_read(dev));
7862 BUG_ON(!list_empty(&dev->ptype_all));
7863 BUG_ON(!list_empty(&dev->ptype_specific));
7864 WARN_ON(rcu_access_pointer(dev->ip_ptr));
7865 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
7866 WARN_ON(dev->dn_ptr);
7868 if (dev->priv_destructor)
7869 dev->priv_destructor(dev);
7870 if (dev->needs_free_netdev)
7871 free_netdev(dev);
7873 /* Report a network device has been unregistered */
7874 rtnl_lock();
7875 dev_net(dev)->dev_unreg_count--;
7876 __rtnl_unlock();
7877 wake_up(&netdev_unregistering_wq);
7879 /* Free network device */
7880 kobject_put(&dev->dev.kobj);
7884 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
7885 * all the same fields in the same order as net_device_stats, with only
7886 * the type differing, but rtnl_link_stats64 may have additional fields
7887 * at the end for newer counters.
7889 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
7890 const struct net_device_stats *netdev_stats)
7892 #if BITS_PER_LONG == 64
7893 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
7894 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
7895 /* zero out counters that only exist in rtnl_link_stats64 */
7896 memset((char *)stats64 + sizeof(*netdev_stats), 0,
7897 sizeof(*stats64) - sizeof(*netdev_stats));
7898 #else
7899 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
7900 const unsigned long *src = (const unsigned long *)netdev_stats;
7901 u64 *dst = (u64 *)stats64;
7903 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
7904 for (i = 0; i < n; i++)
7905 dst[i] = src[i];
7906 /* zero out counters that only exist in rtnl_link_stats64 */
7907 memset((char *)stats64 + n * sizeof(u64), 0,
7908 sizeof(*stats64) - n * sizeof(u64));
7909 #endif
7911 EXPORT_SYMBOL(netdev_stats_to_stats64);
7914 * dev_get_stats - get network device statistics
7915 * @dev: device to get statistics from
7916 * @storage: place to store stats
7918 * Get network statistics from device. Return @storage.
7919 * The device driver may provide its own method by setting
7920 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
7921 * otherwise the internal statistics structure is used.
7923 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
7924 struct rtnl_link_stats64 *storage)
7926 const struct net_device_ops *ops = dev->netdev_ops;
7928 if (ops->ndo_get_stats64) {
7929 memset(storage, 0, sizeof(*storage));
7930 ops->ndo_get_stats64(dev, storage);
7931 } else if (ops->ndo_get_stats) {
7932 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
7933 } else {
7934 netdev_stats_to_stats64(storage, &dev->stats);
7936 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
7937 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
7938 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
7939 return storage;
7941 EXPORT_SYMBOL(dev_get_stats);
7943 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
7945 struct netdev_queue *queue = dev_ingress_queue(dev);
7947 #ifdef CONFIG_NET_CLS_ACT
7948 if (queue)
7949 return queue;
7950 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
7951 if (!queue)
7952 return NULL;
7953 netdev_init_one_queue(dev, queue, NULL);
7954 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
7955 queue->qdisc_sleeping = &noop_qdisc;
7956 rcu_assign_pointer(dev->ingress_queue, queue);
7957 #endif
7958 return queue;
7961 static const struct ethtool_ops default_ethtool_ops;
7963 void netdev_set_default_ethtool_ops(struct net_device *dev,
7964 const struct ethtool_ops *ops)
7966 if (dev->ethtool_ops == &default_ethtool_ops)
7967 dev->ethtool_ops = ops;
7969 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
7971 void netdev_freemem(struct net_device *dev)
7973 char *addr = (char *)dev - dev->padded;
7975 kvfree(addr);
7979 * alloc_netdev_mqs - allocate network device
7980 * @sizeof_priv: size of private data to allocate space for
7981 * @name: device name format string
7982 * @name_assign_type: origin of device name
7983 * @setup: callback to initialize device
7984 * @txqs: the number of TX subqueues to allocate
7985 * @rxqs: the number of RX subqueues to allocate
7987 * Allocates a struct net_device with private data area for driver use
7988 * and performs basic initialization. Also allocates subqueue structs
7989 * for each queue on the device.
7991 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
7992 unsigned char name_assign_type,
7993 void (*setup)(struct net_device *),
7994 unsigned int txqs, unsigned int rxqs)
7996 struct net_device *dev;
7997 size_t alloc_size;
7998 struct net_device *p;
8000 BUG_ON(strlen(name) >= sizeof(dev->name));
8002 if (txqs < 1) {
8003 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8004 return NULL;
8007 #ifdef CONFIG_SYSFS
8008 if (rxqs < 1) {
8009 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8010 return NULL;
8012 #endif
8014 alloc_size = sizeof(struct net_device);
8015 if (sizeof_priv) {
8016 /* ensure 32-byte alignment of private area */
8017 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8018 alloc_size += sizeof_priv;
8020 /* ensure 32-byte alignment of whole construct */
8021 alloc_size += NETDEV_ALIGN - 1;
8023 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8024 if (!p)
8025 return NULL;
8027 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8028 dev->padded = (char *)dev - (char *)p;
8030 dev->pcpu_refcnt = alloc_percpu(int);
8031 if (!dev->pcpu_refcnt)
8032 goto free_dev;
8034 if (dev_addr_init(dev))
8035 goto free_pcpu;
8037 dev_mc_init(dev);
8038 dev_uc_init(dev);
8040 dev_net_set(dev, &init_net);
8042 dev->gso_max_size = GSO_MAX_SIZE;
8043 dev->gso_max_segs = GSO_MAX_SEGS;
8045 INIT_LIST_HEAD(&dev->napi_list);
8046 INIT_LIST_HEAD(&dev->unreg_list);
8047 INIT_LIST_HEAD(&dev->close_list);
8048 INIT_LIST_HEAD(&dev->link_watch_list);
8049 INIT_LIST_HEAD(&dev->adj_list.upper);
8050 INIT_LIST_HEAD(&dev->adj_list.lower);
8051 INIT_LIST_HEAD(&dev->ptype_all);
8052 INIT_LIST_HEAD(&dev->ptype_specific);
8053 #ifdef CONFIG_NET_SCHED
8054 hash_init(dev->qdisc_hash);
8055 #endif
8056 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8057 setup(dev);
8059 if (!dev->tx_queue_len) {
8060 dev->priv_flags |= IFF_NO_QUEUE;
8061 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8064 dev->num_tx_queues = txqs;
8065 dev->real_num_tx_queues = txqs;
8066 if (netif_alloc_netdev_queues(dev))
8067 goto free_all;
8069 #ifdef CONFIG_SYSFS
8070 dev->num_rx_queues = rxqs;
8071 dev->real_num_rx_queues = rxqs;
8072 if (netif_alloc_rx_queues(dev))
8073 goto free_all;
8074 #endif
8076 strcpy(dev->name, name);
8077 dev->name_assign_type = name_assign_type;
8078 dev->group = INIT_NETDEV_GROUP;
8079 if (!dev->ethtool_ops)
8080 dev->ethtool_ops = &default_ethtool_ops;
8082 nf_hook_ingress_init(dev);
8084 return dev;
8086 free_all:
8087 free_netdev(dev);
8088 return NULL;
8090 free_pcpu:
8091 free_percpu(dev->pcpu_refcnt);
8092 free_dev:
8093 netdev_freemem(dev);
8094 return NULL;
8096 EXPORT_SYMBOL(alloc_netdev_mqs);
8099 * free_netdev - free network device
8100 * @dev: device
8102 * This function does the last stage of destroying an allocated device
8103 * interface. The reference to the device object is released. If this
8104 * is the last reference then it will be freed.Must be called in process
8105 * context.
8107 void free_netdev(struct net_device *dev)
8109 struct napi_struct *p, *n;
8110 struct bpf_prog *prog;
8112 might_sleep();
8113 netif_free_tx_queues(dev);
8114 #ifdef CONFIG_SYSFS
8115 kvfree(dev->_rx);
8116 #endif
8118 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8120 /* Flush device addresses */
8121 dev_addr_flush(dev);
8123 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8124 netif_napi_del(p);
8126 free_percpu(dev->pcpu_refcnt);
8127 dev->pcpu_refcnt = NULL;
8129 prog = rcu_dereference_protected(dev->xdp_prog, 1);
8130 if (prog) {
8131 bpf_prog_put(prog);
8132 static_key_slow_dec(&generic_xdp_needed);
8135 /* Compatibility with error handling in drivers */
8136 if (dev->reg_state == NETREG_UNINITIALIZED) {
8137 netdev_freemem(dev);
8138 return;
8141 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8142 dev->reg_state = NETREG_RELEASED;
8144 /* will free via device release */
8145 put_device(&dev->dev);
8147 EXPORT_SYMBOL(free_netdev);
8150 * synchronize_net - Synchronize with packet receive processing
8152 * Wait for packets currently being received to be done.
8153 * Does not block later packets from starting.
8155 void synchronize_net(void)
8157 might_sleep();
8158 if (rtnl_is_locked())
8159 synchronize_rcu_expedited();
8160 else
8161 synchronize_rcu();
8163 EXPORT_SYMBOL(synchronize_net);
8166 * unregister_netdevice_queue - remove device from the kernel
8167 * @dev: device
8168 * @head: list
8170 * This function shuts down a device interface and removes it
8171 * from the kernel tables.
8172 * If head not NULL, device is queued to be unregistered later.
8174 * Callers must hold the rtnl semaphore. You may want
8175 * unregister_netdev() instead of this.
8178 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8180 ASSERT_RTNL();
8182 if (head) {
8183 list_move_tail(&dev->unreg_list, head);
8184 } else {
8185 rollback_registered(dev);
8186 /* Finish processing unregister after unlock */
8187 net_set_todo(dev);
8190 EXPORT_SYMBOL(unregister_netdevice_queue);
8193 * unregister_netdevice_many - unregister many devices
8194 * @head: list of devices
8196 * Note: As most callers use a stack allocated list_head,
8197 * we force a list_del() to make sure stack wont be corrupted later.
8199 void unregister_netdevice_many(struct list_head *head)
8201 struct net_device *dev;
8203 if (!list_empty(head)) {
8204 rollback_registered_many(head);
8205 list_for_each_entry(dev, head, unreg_list)
8206 net_set_todo(dev);
8207 list_del(head);
8210 EXPORT_SYMBOL(unregister_netdevice_many);
8213 * unregister_netdev - remove device from the kernel
8214 * @dev: device
8216 * This function shuts down a device interface and removes it
8217 * from the kernel tables.
8219 * This is just a wrapper for unregister_netdevice that takes
8220 * the rtnl semaphore. In general you want to use this and not
8221 * unregister_netdevice.
8223 void unregister_netdev(struct net_device *dev)
8225 rtnl_lock();
8226 unregister_netdevice(dev);
8227 rtnl_unlock();
8229 EXPORT_SYMBOL(unregister_netdev);
8232 * dev_change_net_namespace - move device to different nethost namespace
8233 * @dev: device
8234 * @net: network namespace
8235 * @pat: If not NULL name pattern to try if the current device name
8236 * is already taken in the destination network namespace.
8238 * This function shuts down a device interface and moves it
8239 * to a new network namespace. On success 0 is returned, on
8240 * a failure a netagive errno code is returned.
8242 * Callers must hold the rtnl semaphore.
8245 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8247 int err;
8249 ASSERT_RTNL();
8251 /* Don't allow namespace local devices to be moved. */
8252 err = -EINVAL;
8253 if (dev->features & NETIF_F_NETNS_LOCAL)
8254 goto out;
8256 /* Ensure the device has been registrered */
8257 if (dev->reg_state != NETREG_REGISTERED)
8258 goto out;
8260 /* Get out if there is nothing todo */
8261 err = 0;
8262 if (net_eq(dev_net(dev), net))
8263 goto out;
8265 /* Pick the destination device name, and ensure
8266 * we can use it in the destination network namespace.
8268 err = -EEXIST;
8269 if (__dev_get_by_name(net, dev->name)) {
8270 /* We get here if we can't use the current device name */
8271 if (!pat)
8272 goto out;
8273 if (dev_get_valid_name(net, dev, pat) < 0)
8274 goto out;
8278 * And now a mini version of register_netdevice unregister_netdevice.
8281 /* If device is running close it first. */
8282 dev_close(dev);
8284 /* And unlink it from device chain */
8285 err = -ENODEV;
8286 unlist_netdevice(dev);
8288 synchronize_net();
8290 /* Shutdown queueing discipline. */
8291 dev_shutdown(dev);
8293 /* Notify protocols, that we are about to destroy
8294 * this device. They should clean all the things.
8296 * Note that dev->reg_state stays at NETREG_REGISTERED.
8297 * This is wanted because this way 8021q and macvlan know
8298 * the device is just moving and can keep their slaves up.
8300 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8301 rcu_barrier();
8302 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8303 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U, GFP_KERNEL);
8306 * Flush the unicast and multicast chains
8308 dev_uc_flush(dev);
8309 dev_mc_flush(dev);
8311 /* Send a netdev-removed uevent to the old namespace */
8312 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8313 netdev_adjacent_del_links(dev);
8315 /* Actually switch the network namespace */
8316 dev_net_set(dev, net);
8318 /* If there is an ifindex conflict assign a new one */
8319 if (__dev_get_by_index(net, dev->ifindex))
8320 dev->ifindex = dev_new_index(net);
8322 /* Send a netdev-add uevent to the new namespace */
8323 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8324 netdev_adjacent_add_links(dev);
8326 /* Fixup kobjects */
8327 err = device_rename(&dev->dev, dev->name);
8328 WARN_ON(err);
8330 /* Add the device back in the hashes */
8331 list_netdevice(dev);
8333 /* Notify protocols, that a new device appeared. */
8334 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8337 * Prevent userspace races by waiting until the network
8338 * device is fully setup before sending notifications.
8340 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8342 synchronize_net();
8343 err = 0;
8344 out:
8345 return err;
8347 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8349 static int dev_cpu_dead(unsigned int oldcpu)
8351 struct sk_buff **list_skb;
8352 struct sk_buff *skb;
8353 unsigned int cpu;
8354 struct softnet_data *sd, *oldsd, *remsd = NULL;
8356 local_irq_disable();
8357 cpu = smp_processor_id();
8358 sd = &per_cpu(softnet_data, cpu);
8359 oldsd = &per_cpu(softnet_data, oldcpu);
8361 /* Find end of our completion_queue. */
8362 list_skb = &sd->completion_queue;
8363 while (*list_skb)
8364 list_skb = &(*list_skb)->next;
8365 /* Append completion queue from offline CPU. */
8366 *list_skb = oldsd->completion_queue;
8367 oldsd->completion_queue = NULL;
8369 /* Append output queue from offline CPU. */
8370 if (oldsd->output_queue) {
8371 *sd->output_queue_tailp = oldsd->output_queue;
8372 sd->output_queue_tailp = oldsd->output_queue_tailp;
8373 oldsd->output_queue = NULL;
8374 oldsd->output_queue_tailp = &oldsd->output_queue;
8376 /* Append NAPI poll list from offline CPU, with one exception :
8377 * process_backlog() must be called by cpu owning percpu backlog.
8378 * We properly handle process_queue & input_pkt_queue later.
8380 while (!list_empty(&oldsd->poll_list)) {
8381 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8382 struct napi_struct,
8383 poll_list);
8385 list_del_init(&napi->poll_list);
8386 if (napi->poll == process_backlog)
8387 napi->state = 0;
8388 else
8389 ____napi_schedule(sd, napi);
8392 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8393 local_irq_enable();
8395 #ifdef CONFIG_RPS
8396 remsd = oldsd->rps_ipi_list;
8397 oldsd->rps_ipi_list = NULL;
8398 #endif
8399 /* send out pending IPI's on offline CPU */
8400 net_rps_send_ipi(remsd);
8402 /* Process offline CPU's input_pkt_queue */
8403 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8404 netif_rx_ni(skb);
8405 input_queue_head_incr(oldsd);
8407 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8408 netif_rx_ni(skb);
8409 input_queue_head_incr(oldsd);
8412 return 0;
8416 * netdev_increment_features - increment feature set by one
8417 * @all: current feature set
8418 * @one: new feature set
8419 * @mask: mask feature set
8421 * Computes a new feature set after adding a device with feature set
8422 * @one to the master device with current feature set @all. Will not
8423 * enable anything that is off in @mask. Returns the new feature set.
8425 netdev_features_t netdev_increment_features(netdev_features_t all,
8426 netdev_features_t one, netdev_features_t mask)
8428 if (mask & NETIF_F_HW_CSUM)
8429 mask |= NETIF_F_CSUM_MASK;
8430 mask |= NETIF_F_VLAN_CHALLENGED;
8432 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8433 all &= one | ~NETIF_F_ALL_FOR_ALL;
8435 /* If one device supports hw checksumming, set for all. */
8436 if (all & NETIF_F_HW_CSUM)
8437 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8439 return all;
8441 EXPORT_SYMBOL(netdev_increment_features);
8443 static struct hlist_head * __net_init netdev_create_hash(void)
8445 int i;
8446 struct hlist_head *hash;
8448 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8449 if (hash != NULL)
8450 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8451 INIT_HLIST_HEAD(&hash[i]);
8453 return hash;
8456 /* Initialize per network namespace state */
8457 static int __net_init netdev_init(struct net *net)
8459 if (net != &init_net)
8460 INIT_LIST_HEAD(&net->dev_base_head);
8462 net->dev_name_head = netdev_create_hash();
8463 if (net->dev_name_head == NULL)
8464 goto err_name;
8466 net->dev_index_head = netdev_create_hash();
8467 if (net->dev_index_head == NULL)
8468 goto err_idx;
8470 return 0;
8472 err_idx:
8473 kfree(net->dev_name_head);
8474 err_name:
8475 return -ENOMEM;
8479 * netdev_drivername - network driver for the device
8480 * @dev: network device
8482 * Determine network driver for device.
8484 const char *netdev_drivername(const struct net_device *dev)
8486 const struct device_driver *driver;
8487 const struct device *parent;
8488 const char *empty = "";
8490 parent = dev->dev.parent;
8491 if (!parent)
8492 return empty;
8494 driver = parent->driver;
8495 if (driver && driver->name)
8496 return driver->name;
8497 return empty;
8500 static void __netdev_printk(const char *level, const struct net_device *dev,
8501 struct va_format *vaf)
8503 if (dev && dev->dev.parent) {
8504 dev_printk_emit(level[1] - '0',
8505 dev->dev.parent,
8506 "%s %s %s%s: %pV",
8507 dev_driver_string(dev->dev.parent),
8508 dev_name(dev->dev.parent),
8509 netdev_name(dev), netdev_reg_state(dev),
8510 vaf);
8511 } else if (dev) {
8512 printk("%s%s%s: %pV",
8513 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8514 } else {
8515 printk("%s(NULL net_device): %pV", level, vaf);
8519 void netdev_printk(const char *level, const struct net_device *dev,
8520 const char *format, ...)
8522 struct va_format vaf;
8523 va_list args;
8525 va_start(args, format);
8527 vaf.fmt = format;
8528 vaf.va = &args;
8530 __netdev_printk(level, dev, &vaf);
8532 va_end(args);
8534 EXPORT_SYMBOL(netdev_printk);
8536 #define define_netdev_printk_level(func, level) \
8537 void func(const struct net_device *dev, const char *fmt, ...) \
8539 struct va_format vaf; \
8540 va_list args; \
8542 va_start(args, fmt); \
8544 vaf.fmt = fmt; \
8545 vaf.va = &args; \
8547 __netdev_printk(level, dev, &vaf); \
8549 va_end(args); \
8551 EXPORT_SYMBOL(func);
8553 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8554 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8555 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8556 define_netdev_printk_level(netdev_err, KERN_ERR);
8557 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8558 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8559 define_netdev_printk_level(netdev_info, KERN_INFO);
8561 static void __net_exit netdev_exit(struct net *net)
8563 kfree(net->dev_name_head);
8564 kfree(net->dev_index_head);
8567 static struct pernet_operations __net_initdata netdev_net_ops = {
8568 .init = netdev_init,
8569 .exit = netdev_exit,
8572 static void __net_exit default_device_exit(struct net *net)
8574 struct net_device *dev, *aux;
8576 * Push all migratable network devices back to the
8577 * initial network namespace
8579 rtnl_lock();
8580 for_each_netdev_safe(net, dev, aux) {
8581 int err;
8582 char fb_name[IFNAMSIZ];
8584 /* Ignore unmoveable devices (i.e. loopback) */
8585 if (dev->features & NETIF_F_NETNS_LOCAL)
8586 continue;
8588 /* Leave virtual devices for the generic cleanup */
8589 if (dev->rtnl_link_ops)
8590 continue;
8592 /* Push remaining network devices to init_net */
8593 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8594 err = dev_change_net_namespace(dev, &init_net, fb_name);
8595 if (err) {
8596 pr_emerg("%s: failed to move %s to init_net: %d\n",
8597 __func__, dev->name, err);
8598 BUG();
8601 rtnl_unlock();
8604 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8606 /* Return with the rtnl_lock held when there are no network
8607 * devices unregistering in any network namespace in net_list.
8609 struct net *net;
8610 bool unregistering;
8611 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8613 add_wait_queue(&netdev_unregistering_wq, &wait);
8614 for (;;) {
8615 unregistering = false;
8616 rtnl_lock();
8617 list_for_each_entry(net, net_list, exit_list) {
8618 if (net->dev_unreg_count > 0) {
8619 unregistering = true;
8620 break;
8623 if (!unregistering)
8624 break;
8625 __rtnl_unlock();
8627 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8629 remove_wait_queue(&netdev_unregistering_wq, &wait);
8632 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8634 /* At exit all network devices most be removed from a network
8635 * namespace. Do this in the reverse order of registration.
8636 * Do this across as many network namespaces as possible to
8637 * improve batching efficiency.
8639 struct net_device *dev;
8640 struct net *net;
8641 LIST_HEAD(dev_kill_list);
8643 /* To prevent network device cleanup code from dereferencing
8644 * loopback devices or network devices that have been freed
8645 * wait here for all pending unregistrations to complete,
8646 * before unregistring the loopback device and allowing the
8647 * network namespace be freed.
8649 * The netdev todo list containing all network devices
8650 * unregistrations that happen in default_device_exit_batch
8651 * will run in the rtnl_unlock() at the end of
8652 * default_device_exit_batch.
8654 rtnl_lock_unregistering(net_list);
8655 list_for_each_entry(net, net_list, exit_list) {
8656 for_each_netdev_reverse(net, dev) {
8657 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
8658 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
8659 else
8660 unregister_netdevice_queue(dev, &dev_kill_list);
8663 unregister_netdevice_many(&dev_kill_list);
8664 rtnl_unlock();
8667 static struct pernet_operations __net_initdata default_device_ops = {
8668 .exit = default_device_exit,
8669 .exit_batch = default_device_exit_batch,
8673 * Initialize the DEV module. At boot time this walks the device list and
8674 * unhooks any devices that fail to initialise (normally hardware not
8675 * present) and leaves us with a valid list of present and active devices.
8680 * This is called single threaded during boot, so no need
8681 * to take the rtnl semaphore.
8683 static int __init net_dev_init(void)
8685 int i, rc = -ENOMEM;
8687 BUG_ON(!dev_boot_phase);
8689 if (dev_proc_init())
8690 goto out;
8692 if (netdev_kobject_init())
8693 goto out;
8695 INIT_LIST_HEAD(&ptype_all);
8696 for (i = 0; i < PTYPE_HASH_SIZE; i++)
8697 INIT_LIST_HEAD(&ptype_base[i]);
8699 INIT_LIST_HEAD(&offload_base);
8701 if (register_pernet_subsys(&netdev_net_ops))
8702 goto out;
8705 * Initialise the packet receive queues.
8708 for_each_possible_cpu(i) {
8709 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
8710 struct softnet_data *sd = &per_cpu(softnet_data, i);
8712 INIT_WORK(flush, flush_backlog);
8714 skb_queue_head_init(&sd->input_pkt_queue);
8715 skb_queue_head_init(&sd->process_queue);
8716 INIT_LIST_HEAD(&sd->poll_list);
8717 sd->output_queue_tailp = &sd->output_queue;
8718 #ifdef CONFIG_RPS
8719 sd->csd.func = rps_trigger_softirq;
8720 sd->csd.info = sd;
8721 sd->cpu = i;
8722 #endif
8724 sd->backlog.poll = process_backlog;
8725 sd->backlog.weight = weight_p;
8728 dev_boot_phase = 0;
8730 /* The loopback device is special if any other network devices
8731 * is present in a network namespace the loopback device must
8732 * be present. Since we now dynamically allocate and free the
8733 * loopback device ensure this invariant is maintained by
8734 * keeping the loopback device as the first device on the
8735 * list of network devices. Ensuring the loopback devices
8736 * is the first device that appears and the last network device
8737 * that disappears.
8739 if (register_pernet_device(&loopback_net_ops))
8740 goto out;
8742 if (register_pernet_device(&default_device_ops))
8743 goto out;
8745 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
8746 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
8748 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
8749 NULL, dev_cpu_dead);
8750 WARN_ON(rc < 0);
8751 rc = 0;
8752 out:
8753 return rc;
8756 subsys_initcall(net_dev_init);