1/*
2 * NET3 Protocol independent device support routines.
3 *
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.
8 *
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>
13 *
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>
21 *
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
73 */
74
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>
148#include <linux/net_namespace.h>
149
150#include "net-sysfs.h"
151
152/* Instead of increasing this, you should create a hash table. */
153#define MAX_GRO_SKBS 8
154
155/* This should be increased if a protocol with a bigger head is added. */
156#define GRO_MAX_HEAD (MAX_HEADER + 128)
157
158static DEFINE_SPINLOCK(ptype_lock);
159static DEFINE_SPINLOCK(offload_lock);
160struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
161struct list_head ptype_all __read_mostly; /* Taps */
162static struct list_head offload_base __read_mostly;
163
164static int netif_rx_internal(struct sk_buff *skb);
165static int call_netdevice_notifiers_info(unsigned long val,
166 struct netdev_notifier_info *info);
167static struct napi_struct *napi_by_id(unsigned int napi_id);
168
169/*
170 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
171 * semaphore.
172 *
173 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
174 *
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.
179 *
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.
183 *
184 * See, for example usages, register_netdevice() and
185 * unregister_netdevice(), which must be called with the rtnl
186 * semaphore held.
187 */
188DEFINE_RWLOCK(dev_base_lock);
189EXPORT_SYMBOL(dev_base_lock);
190
191static DEFINE_MUTEX(ifalias_mutex);
192
193/* protects napi_hash addition/deletion and napi_gen_id */
194static DEFINE_SPINLOCK(napi_hash_lock);
195
196static unsigned int napi_gen_id = NR_CPUS;
197static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
198
199static seqcount_t devnet_rename_seq;
200
201static inline void dev_base_seq_inc(struct net *net)
202{
203 while (++net->dev_base_seq == 0)
204 ;
205}
206
207static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
208{
209 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
210
211 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
212}
213
214static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
215{
216 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
217}
218
219static inline void rps_lock(struct softnet_data *sd)
220{
221#ifdef CONFIG_RPS
222 spin_lock(&sd->input_pkt_queue.lock);
223#endif
224}
225
226static inline void rps_unlock(struct softnet_data *sd)
227{
228#ifdef CONFIG_RPS
229 spin_unlock(&sd->input_pkt_queue.lock);
230#endif
231}
232
233/* Device list insertion */
234static void list_netdevice(struct net_device *dev)
235{
236 struct net *net = dev_net(dev);
237
238 ASSERT_RTNL();
239
240 write_lock_bh(&dev_base_lock);
241 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
242 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
243 hlist_add_head_rcu(&dev->index_hlist,
244 dev_index_hash(net, dev->ifindex));
245 write_unlock_bh(&dev_base_lock);
246
247 dev_base_seq_inc(net);
248}
249
250/* Device list removal
251 * caller must respect a RCU grace period before freeing/reusing dev
252 */
253static void unlist_netdevice(struct net_device *dev)
254{
255 ASSERT_RTNL();
256
257 /* Unlink dev from the device chain */
258 write_lock_bh(&dev_base_lock);
259 list_del_rcu(&dev->dev_list);
260 hlist_del_rcu(&dev->name_hlist);
261 hlist_del_rcu(&dev->index_hlist);
262 write_unlock_bh(&dev_base_lock);
263
264 dev_base_seq_inc(dev_net(dev));
265}
266
267/*
268 * Our notifier list
269 */
270
271static RAW_NOTIFIER_HEAD(netdev_chain);
272
273/*
274 * Device drivers call our routines to queue packets here. We empty the
275 * queue in the local softnet handler.
276 */
277
278DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
279EXPORT_PER_CPU_SYMBOL(softnet_data);
280
281#ifdef CONFIG_LOCKDEP
282/*
283 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
284 * according to dev->type
285 */
286static const unsigned short netdev_lock_type[] = {
287 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
288 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
289 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
290 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
291 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
292 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
293 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
294 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
295 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
296 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
297 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
298 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
299 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
300 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
301 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
302
303static const char *const netdev_lock_name[] = {
304 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
305 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
306 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
307 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
308 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
309 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
310 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
311 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
312 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
313 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
314 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
315 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
316 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
317 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
318 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
319
320static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
321static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
322
323static inline unsigned short netdev_lock_pos(unsigned short dev_type)
324{
325 int i;
326
327 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
328 if (netdev_lock_type[i] == dev_type)
329 return i;
330 /* the last key is used by default */
331 return ARRAY_SIZE(netdev_lock_type) - 1;
332}
333
334static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
335 unsigned short dev_type)
336{
337 int i;
338
339 i = netdev_lock_pos(dev_type);
340 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
341 netdev_lock_name[i]);
342}
343
344static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
345{
346 int i;
347
348 i = netdev_lock_pos(dev->type);
349 lockdep_set_class_and_name(&dev->addr_list_lock,
350 &netdev_addr_lock_key[i],
351 netdev_lock_name[i]);
352}
353#else
354static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
355 unsigned short dev_type)
356{
357}
358static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
359{
360}
361#endif
362
363/*******************************************************************************
364 *
365 * Protocol management and registration routines
366 *
367 *******************************************************************************/
368
369
370/*
371 * Add a protocol ID to the list. Now that the input handler is
372 * smarter we can dispense with all the messy stuff that used to be
373 * here.
374 *
375 * BEWARE!!! Protocol handlers, mangling input packets,
376 * MUST BE last in hash buckets and checking protocol handlers
377 * MUST start from promiscuous ptype_all chain in net_bh.
378 * It is true now, do not change it.
379 * Explanation follows: if protocol handler, mangling packet, will
380 * be the first on list, it is not able to sense, that packet
381 * is cloned and should be copied-on-write, so that it will
382 * change it and subsequent readers will get broken packet.
383 * --ANK (980803)
384 */
385
386static inline struct list_head *ptype_head(const struct packet_type *pt)
387{
388 if (pt->type == htons(ETH_P_ALL))
389 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
390 else
391 return pt->dev ? &pt->dev->ptype_specific :
392 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
393}
394
395/**
396 * dev_add_pack - add packet handler
397 * @pt: packet type declaration
398 *
399 * Add a protocol handler to the networking stack. The passed &packet_type
400 * is linked into kernel lists and may not be freed until it has been
401 * removed from the kernel lists.
402 *
403 * This call does not sleep therefore it can not
404 * guarantee all CPU's that are in middle of receiving packets
405 * will see the new packet type (until the next received packet).
406 */
407
408void dev_add_pack(struct packet_type *pt)
409{
410 struct list_head *head = ptype_head(pt);
411
412 spin_lock(&ptype_lock);
413 list_add_rcu(&pt->list, head);
414 spin_unlock(&ptype_lock);
415}
416EXPORT_SYMBOL(dev_add_pack);
417
418/**
419 * __dev_remove_pack - remove packet handler
420 * @pt: packet type declaration
421 *
422 * Remove a protocol handler that was previously added to the kernel
423 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
424 * from the kernel lists and can be freed or reused once this function
425 * returns.
426 *
427 * The packet type might still be in use by receivers
428 * and must not be freed until after all the CPU's have gone
429 * through a quiescent state.
430 */
431void __dev_remove_pack(struct packet_type *pt)
432{
433 struct list_head *head = ptype_head(pt);
434 struct packet_type *pt1;
435
436 spin_lock(&ptype_lock);
437
438 list_for_each_entry(pt1, head, list) {
439 if (pt == pt1) {
440 list_del_rcu(&pt->list);
441 goto out;
442 }
443 }
444
445 pr_warn("dev_remove_pack: %p not found\n", pt);
446out:
447 spin_unlock(&ptype_lock);
448}
449EXPORT_SYMBOL(__dev_remove_pack);
450
451/**
452 * dev_remove_pack - remove packet handler
453 * @pt: packet type declaration
454 *
455 * Remove a protocol handler that was previously added to the kernel
456 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
457 * from the kernel lists and can be freed or reused once this function
458 * returns.
459 *
460 * This call sleeps to guarantee that no CPU is looking at the packet
461 * type after return.
462 */
463void dev_remove_pack(struct packet_type *pt)
464{
465 __dev_remove_pack(pt);
466
467 synchronize_net();
468}
469EXPORT_SYMBOL(dev_remove_pack);
470
471
472/**
473 * dev_add_offload - register offload handlers
474 * @po: protocol offload declaration
475 *
476 * Add protocol offload handlers to the networking stack. The passed
477 * &proto_offload is linked into kernel lists and may not be freed until
478 * it has been removed from the kernel lists.
479 *
480 * This call does not sleep therefore it can not
481 * guarantee all CPU's that are in middle of receiving packets
482 * will see the new offload handlers (until the next received packet).
483 */
484void dev_add_offload(struct packet_offload *po)
485{
486 struct packet_offload *elem;
487
488 spin_lock(&offload_lock);
489 list_for_each_entry(elem, &offload_base, list) {
490 if (po->priority < elem->priority)
491 break;
492 }
493 list_add_rcu(&po->list, elem->list.prev);
494 spin_unlock(&offload_lock);
495}
496EXPORT_SYMBOL(dev_add_offload);
497
498/**
499 * __dev_remove_offload - remove offload handler
500 * @po: packet offload declaration
501 *
502 * Remove a protocol offload handler that was previously added to the
503 * kernel offload handlers by dev_add_offload(). The passed &offload_type
504 * is removed from the kernel lists and can be freed or reused once this
505 * function returns.
506 *
507 * The packet type might still be in use by receivers
508 * and must not be freed until after all the CPU's have gone
509 * through a quiescent state.
510 */
511static void __dev_remove_offload(struct packet_offload *po)
512{
513 struct list_head *head = &offload_base;
514 struct packet_offload *po1;
515
516 spin_lock(&offload_lock);
517
518 list_for_each_entry(po1, head, list) {
519 if (po == po1) {
520 list_del_rcu(&po->list);
521 goto out;
522 }
523 }
524
525 pr_warn("dev_remove_offload: %p not found\n", po);
526out:
527 spin_unlock(&offload_lock);
528}
529
530/**
531 * dev_remove_offload - remove packet offload handler
532 * @po: packet offload declaration
533 *
534 * Remove a packet offload handler that was previously added to the kernel
535 * offload handlers by dev_add_offload(). The passed &offload_type is
536 * removed from the kernel lists and can be freed or reused once this
537 * function returns.
538 *
539 * This call sleeps to guarantee that no CPU is looking at the packet
540 * type after return.
541 */
542void dev_remove_offload(struct packet_offload *po)
543{
544 __dev_remove_offload(po);
545
546 synchronize_net();
547}
548EXPORT_SYMBOL(dev_remove_offload);
549
550/******************************************************************************
551 *
552 * Device Boot-time Settings Routines
553 *
554 ******************************************************************************/
555
556/* Boot time configuration table */
557static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
558
559/**
560 * netdev_boot_setup_add - add new setup entry
561 * @name: name of the device
562 * @map: configured settings for the device
563 *
564 * Adds new setup entry to the dev_boot_setup list. The function
565 * returns 0 on error and 1 on success. This is a generic routine to
566 * all netdevices.
567 */
568static int netdev_boot_setup_add(char *name, struct ifmap *map)
569{
570 struct netdev_boot_setup *s;
571 int i;
572
573 s = dev_boot_setup;
574 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
575 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
576 memset(s[i].name, 0, sizeof(s[i].name));
577 strlcpy(s[i].name, name, IFNAMSIZ);
578 memcpy(&s[i].map, map, sizeof(s[i].map));
579 break;
580 }
581 }
582
583 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
584}
585
586/**
587 * netdev_boot_setup_check - check boot time settings
588 * @dev: the netdevice
589 *
590 * Check boot time settings for the device.
591 * The found settings are set for the device to be used
592 * later in the device probing.
593 * Returns 0 if no settings found, 1 if they are.
594 */
595int netdev_boot_setup_check(struct net_device *dev)
596{
597 struct netdev_boot_setup *s = dev_boot_setup;
598 int i;
599
600 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
601 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
602 !strcmp(dev->name, s[i].name)) {
603 dev->irq = s[i].map.irq;
604 dev->base_addr = s[i].map.base_addr;
605 dev->mem_start = s[i].map.mem_start;
606 dev->mem_end = s[i].map.mem_end;
607 return 1;
608 }
609 }
610 return 0;
611}
612EXPORT_SYMBOL(netdev_boot_setup_check);
613
614
615/**
616 * netdev_boot_base - get address from boot time settings
617 * @prefix: prefix for network device
618 * @unit: id for network device
619 *
620 * Check boot time settings for the base address of device.
621 * The found settings are set for the device to be used
622 * later in the device probing.
623 * Returns 0 if no settings found.
624 */
625unsigned long netdev_boot_base(const char *prefix, int unit)
626{
627 const struct netdev_boot_setup *s = dev_boot_setup;
628 char name[IFNAMSIZ];
629 int i;
630
631 sprintf(name, "%s%d", prefix, unit);
632
633 /*
634 * If device already registered then return base of 1
635 * to indicate not to probe for this interface
636 */
637 if (__dev_get_by_name(&init_net, name))
638 return 1;
639
640 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
641 if (!strcmp(name, s[i].name))
642 return s[i].map.base_addr;
643 return 0;
644}
645
646/*
647 * Saves at boot time configured settings for any netdevice.
648 */
649int __init netdev_boot_setup(char *str)
650{
651 int ints[5];
652 struct ifmap map;
653
654 str = get_options(str, ARRAY_SIZE(ints), ints);
655 if (!str || !*str)
656 return 0;
657
658 /* Save settings */
659 memset(&map, 0, sizeof(map));
660 if (ints[0] > 0)
661 map.irq = ints[1];
662 if (ints[0] > 1)
663 map.base_addr = ints[2];
664 if (ints[0] > 2)
665 map.mem_start = ints[3];
666 if (ints[0] > 3)
667 map.mem_end = ints[4];
668
669 /* Add new entry to the list */
670 return netdev_boot_setup_add(str, &map);
671}
672
673__setup("netdev=", netdev_boot_setup);
674
675/*******************************************************************************
676 *
677 * Device Interface Subroutines
678 *
679 *******************************************************************************/
680
681/**
682 * dev_get_iflink - get 'iflink' value of a interface
683 * @dev: targeted interface
684 *
685 * Indicates the ifindex the interface is linked to.
686 * Physical interfaces have the same 'ifindex' and 'iflink' values.
687 */
688
689int dev_get_iflink(const struct net_device *dev)
690{
691 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
692 return dev->netdev_ops->ndo_get_iflink(dev);
693
694 return dev->ifindex;
695}
696EXPORT_SYMBOL(dev_get_iflink);
697
698/**
699 * dev_fill_metadata_dst - Retrieve tunnel egress information.
700 * @dev: targeted interface
701 * @skb: The packet.
702 *
703 * For better visibility of tunnel traffic OVS needs to retrieve
704 * egress tunnel information for a packet. Following API allows
705 * user to get this info.
706 */
707int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
708{
709 struct ip_tunnel_info *info;
710
711 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
712 return -EINVAL;
713
714 info = skb_tunnel_info_unclone(skb);
715 if (!info)
716 return -ENOMEM;
717 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
718 return -EINVAL;
719
720 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
721}
722EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
723
724/**
725 * __dev_get_by_name - find a device by its name
726 * @net: the applicable net namespace
727 * @name: name to find
728 *
729 * Find an interface by name. Must be called under RTNL semaphore
730 * or @dev_base_lock. If the name is found a pointer to the device
731 * is returned. If the name is not found then %NULL is returned. The
732 * reference counters are not incremented so the caller must be
733 * careful with locks.
734 */
735
736struct net_device *__dev_get_by_name(struct net *net, const char *name)
737{
738 struct net_device *dev;
739 struct hlist_head *head = dev_name_hash(net, name);
740
741 hlist_for_each_entry(dev, head, name_hlist)
742 if (!strncmp(dev->name, name, IFNAMSIZ))
743 return dev;
744
745 return NULL;
746}
747EXPORT_SYMBOL(__dev_get_by_name);
748
749/**
750 * dev_get_by_name_rcu - find a device by its name
751 * @net: the applicable net namespace
752 * @name: name to find
753 *
754 * Find an interface by name.
755 * If the name is found a pointer to the device is returned.
756 * If the name is not found then %NULL is returned.
757 * The reference counters are not incremented so the caller must be
758 * careful with locks. The caller must hold RCU lock.
759 */
760
761struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
762{
763 struct net_device *dev;
764 struct hlist_head *head = dev_name_hash(net, name);
765
766 hlist_for_each_entry_rcu(dev, head, name_hlist)
767 if (!strncmp(dev->name, name, IFNAMSIZ))
768 return dev;
769
770 return NULL;
771}
772EXPORT_SYMBOL(dev_get_by_name_rcu);
773
774/**
775 * dev_get_by_name - find a device by its name
776 * @net: the applicable net namespace
777 * @name: name to find
778 *
779 * Find an interface by name. This can be called from any
780 * context and does its own locking. The returned handle has
781 * the usage count incremented and the caller must use dev_put() to
782 * release it when it is no longer needed. %NULL is returned if no
783 * matching device is found.
784 */
785
786struct net_device *dev_get_by_name(struct net *net, const char *name)
787{
788 struct net_device *dev;
789
790 rcu_read_lock();
791 dev = dev_get_by_name_rcu(net, name);
792 if (dev)
793 dev_hold(dev);
794 rcu_read_unlock();
795 return dev;
796}
797EXPORT_SYMBOL(dev_get_by_name);
798
799/**
800 * __dev_get_by_index - find a device by its ifindex
801 * @net: the applicable net namespace
802 * @ifindex: index of device
803 *
804 * Search for an interface by index. Returns %NULL if the device
805 * is not found or a pointer to the device. The device has not
806 * had its reference counter increased so the caller must be careful
807 * about locking. The caller must hold either the RTNL semaphore
808 * or @dev_base_lock.
809 */
810
811struct net_device *__dev_get_by_index(struct net *net, int ifindex)
812{
813 struct net_device *dev;
814 struct hlist_head *head = dev_index_hash(net, ifindex);
815
816 hlist_for_each_entry(dev, head, index_hlist)
817 if (dev->ifindex == ifindex)
818 return dev;
819
820 return NULL;
821}
822EXPORT_SYMBOL(__dev_get_by_index);
823
824/**
825 * dev_get_by_index_rcu - find a device by its ifindex
826 * @net: the applicable net namespace
827 * @ifindex: index of device
828 *
829 * Search for an interface by index. Returns %NULL if the device
830 * is not found or a pointer to the device. The device has not
831 * had its reference counter increased so the caller must be careful
832 * about locking. The caller must hold RCU lock.
833 */
834
835struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
836{
837 struct net_device *dev;
838 struct hlist_head *head = dev_index_hash(net, ifindex);
839
840 hlist_for_each_entry_rcu(dev, head, index_hlist)
841 if (dev->ifindex == ifindex)
842 return dev;
843
844 return NULL;
845}
846EXPORT_SYMBOL(dev_get_by_index_rcu);
847
848
849/**
850 * dev_get_by_index - find a device by its ifindex
851 * @net: the applicable net namespace
852 * @ifindex: index of device
853 *
854 * Search for an interface by index. Returns NULL if the device
855 * is not found or a pointer to the device. The device returned has
856 * had a reference added and the pointer is safe until the user calls
857 * dev_put to indicate they have finished with it.
858 */
859
860struct net_device *dev_get_by_index(struct net *net, int ifindex)
861{
862 struct net_device *dev;
863
864 rcu_read_lock();
865 dev = dev_get_by_index_rcu(net, ifindex);
866 if (dev)
867 dev_hold(dev);
868 rcu_read_unlock();
869 return dev;
870}
871EXPORT_SYMBOL(dev_get_by_index);
872
873/**
874 * dev_get_by_napi_id - find a device by napi_id
875 * @napi_id: ID of the NAPI struct
876 *
877 * Search for an interface by NAPI ID. Returns %NULL if the device
878 * is not found or a pointer to the device. The device has not had
879 * its reference counter increased so the caller must be careful
880 * about locking. The caller must hold RCU lock.
881 */
882
883struct net_device *dev_get_by_napi_id(unsigned int napi_id)
884{
885 struct napi_struct *napi;
886
887 WARN_ON_ONCE(!rcu_read_lock_held());
888
889 if (napi_id < MIN_NAPI_ID)
890 return NULL;
891
892 napi = napi_by_id(napi_id);
893
894 return napi ? napi->dev : NULL;
895}
896EXPORT_SYMBOL(dev_get_by_napi_id);
897
898/**
899 * netdev_get_name - get a netdevice name, knowing its ifindex.
900 * @net: network namespace
901 * @name: a pointer to the buffer where the name will be stored.
902 * @ifindex: the ifindex of the interface to get the name from.
903 *
904 * The use of raw_seqcount_begin() and cond_resched() before
905 * retrying is required as we want to give the writers a chance
906 * to complete when CONFIG_PREEMPT is not set.
907 */
908int netdev_get_name(struct net *net, char *name, int ifindex)
909{
910 struct net_device *dev;
911 unsigned int seq;
912
913retry:
914 seq = raw_seqcount_begin(&devnet_rename_seq);
915 rcu_read_lock();
916 dev = dev_get_by_index_rcu(net, ifindex);
917 if (!dev) {
918 rcu_read_unlock();
919 return -ENODEV;
920 }
921
922 strcpy(name, dev->name);
923 rcu_read_unlock();
924 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
925 cond_resched();
926 goto retry;
927 }
928
929 return 0;
930}
931
932/**
933 * dev_getbyhwaddr_rcu - find a device by its hardware address
934 * @net: the applicable net namespace
935 * @type: media type of device
936 * @ha: hardware address
937 *
938 * Search for an interface by MAC address. Returns NULL if the device
939 * is not found or a pointer to the device.
940 * The caller must hold RCU or RTNL.
941 * The returned device has not had its ref count increased
942 * and the caller must therefore be careful about locking
943 *
944 */
945
946struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
947 const char *ha)
948{
949 struct net_device *dev;
950
951 for_each_netdev_rcu(net, dev)
952 if (dev->type == type &&
953 !memcmp(dev->dev_addr, ha, dev->addr_len))
954 return dev;
955
956 return NULL;
957}
958EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
959
960struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
961{
962 struct net_device *dev;
963
964 ASSERT_RTNL();
965 for_each_netdev(net, dev)
966 if (dev->type == type)
967 return dev;
968
969 return NULL;
970}
971EXPORT_SYMBOL(__dev_getfirstbyhwtype);
972
973struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
974{
975 struct net_device *dev, *ret = NULL;
976
977 rcu_read_lock();
978 for_each_netdev_rcu(net, dev)
979 if (dev->type == type) {
980 dev_hold(dev);
981 ret = dev;
982 break;
983 }
984 rcu_read_unlock();
985 return ret;
986}
987EXPORT_SYMBOL(dev_getfirstbyhwtype);
988
989/**
990 * __dev_get_by_flags - find any device with given flags
991 * @net: the applicable net namespace
992 * @if_flags: IFF_* values
993 * @mask: bitmask of bits in if_flags to check
994 *
995 * Search for any interface with the given flags. Returns NULL if a device
996 * is not found or a pointer to the device. Must be called inside
997 * rtnl_lock(), and result refcount is unchanged.
998 */
999
1000struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1001 unsigned short mask)
1002{
1003 struct net_device *dev, *ret;
1004
1005 ASSERT_RTNL();
1006
1007 ret = NULL;
1008 for_each_netdev(net, dev) {
1009 if (((dev->flags ^ if_flags) & mask) == 0) {
1010 ret = dev;
1011 break;
1012 }
1013 }
1014 return ret;
1015}
1016EXPORT_SYMBOL(__dev_get_by_flags);
1017
1018/**
1019 * dev_valid_name - check if name is okay for network device
1020 * @name: name string
1021 *
1022 * Network device names need to be valid file names to
1023 * to allow sysfs to work. We also disallow any kind of
1024 * whitespace.
1025 */
1026bool dev_valid_name(const char *name)
1027{
1028 if (*name == '\0')
1029 return false;
1030 if (strlen(name) >= IFNAMSIZ)
1031 return false;
1032 if (!strcmp(name, ".") || !strcmp(name, ".."))
1033 return false;
1034
1035 while (*name) {
1036 if (*name == '/' || *name == ':' || isspace(*name))
1037 return false;
1038 name++;
1039 }
1040 return true;
1041}
1042EXPORT_SYMBOL(dev_valid_name);
1043
1044/**
1045 * __dev_alloc_name - allocate a name for a device
1046 * @net: network namespace to allocate the device name in
1047 * @name: name format string
1048 * @buf: scratch buffer and result name string
1049 *
1050 * Passed a format string - eg "lt%d" it will try and find a suitable
1051 * id. It scans list of devices to build up a free map, then chooses
1052 * the first empty slot. The caller must hold the dev_base or rtnl lock
1053 * while allocating the name and adding the device in order to avoid
1054 * duplicates.
1055 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1056 * Returns the number of the unit assigned or a negative errno code.
1057 */
1058
1059static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1060{
1061 int i = 0;
1062 const char *p;
1063 const int max_netdevices = 8*PAGE_SIZE;
1064 unsigned long *inuse;
1065 struct net_device *d;
1066
1067 if (!dev_valid_name(name))
1068 return -EINVAL;
1069
1070 p = strchr(name, '%');
1071 if (p) {
1072 /*
1073 * Verify the string as this thing may have come from
1074 * the user. There must be either one "%d" and no other "%"
1075 * characters.
1076 */
1077 if (p[1] != 'd' || strchr(p + 2, '%'))
1078 return -EINVAL;
1079
1080 /* Use one page as a bit array of possible slots */
1081 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1082 if (!inuse)
1083 return -ENOMEM;
1084
1085 for_each_netdev(net, d) {
1086 if (!sscanf(d->name, name, &i))
1087 continue;
1088 if (i < 0 || i >= max_netdevices)
1089 continue;
1090
1091 /* avoid cases where sscanf is not exact inverse of printf */
1092 snprintf(buf, IFNAMSIZ, name, i);
1093 if (!strncmp(buf, d->name, IFNAMSIZ))
1094 set_bit(i, inuse);
1095 }
1096
1097 i = find_first_zero_bit(inuse, max_netdevices);
1098 free_page((unsigned long) inuse);
1099 }
1100
1101 snprintf(buf, IFNAMSIZ, name, i);
1102 if (!__dev_get_by_name(net, buf))
1103 return i;
1104
1105 /* It is possible to run out of possible slots
1106 * when the name is long and there isn't enough space left
1107 * for the digits, or if all bits are used.
1108 */
1109 return p ? -ENFILE : -EEXIST;
1110}
1111
1112static int dev_alloc_name_ns(struct net *net,
1113 struct net_device *dev,
1114 const char *name)
1115{
1116 char buf[IFNAMSIZ];
1117 int ret;
1118
1119 BUG_ON(!net);
1120 ret = __dev_alloc_name(net, name, buf);
1121 if (ret >= 0)
1122 strlcpy(dev->name, buf, IFNAMSIZ);
1123 return ret;
1124}
1125
1126/**
1127 * dev_alloc_name - allocate a name for a device
1128 * @dev: device
1129 * @name: name format string
1130 *
1131 * Passed a format string - eg "lt%d" it will try and find a suitable
1132 * id. It scans list of devices to build up a free map, then chooses
1133 * the first empty slot. The caller must hold the dev_base or rtnl lock
1134 * while allocating the name and adding the device in order to avoid
1135 * duplicates.
1136 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1137 * Returns the number of the unit assigned or a negative errno code.
1138 */
1139
1140int dev_alloc_name(struct net_device *dev, const char *name)
1141{
1142 return dev_alloc_name_ns(dev_net(dev), dev, name);
1143}
1144EXPORT_SYMBOL(dev_alloc_name);
1145
1146int dev_get_valid_name(struct net *net, struct net_device *dev,
1147 const char *name)
1148{
1149 return dev_alloc_name_ns(net, dev, name);
1150}
1151EXPORT_SYMBOL(dev_get_valid_name);
1152
1153/**
1154 * dev_change_name - change name of a device
1155 * @dev: device
1156 * @newname: name (or format string) must be at least IFNAMSIZ
1157 *
1158 * Change name of a device, can pass format strings "eth%d".
1159 * for wildcarding.
1160 */
1161int dev_change_name(struct net_device *dev, const char *newname)
1162{
1163 unsigned char old_assign_type;
1164 char oldname[IFNAMSIZ];
1165 int err = 0;
1166 int ret;
1167 struct net *net;
1168
1169 ASSERT_RTNL();
1170 BUG_ON(!dev_net(dev));
1171
1172 net = dev_net(dev);
1173 if (dev->flags & IFF_UP)
1174 return -EBUSY;
1175
1176 write_seqcount_begin(&devnet_rename_seq);
1177
1178 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1179 write_seqcount_end(&devnet_rename_seq);
1180 return 0;
1181 }
1182
1183 memcpy(oldname, dev->name, IFNAMSIZ);
1184
1185 err = dev_get_valid_name(net, dev, newname);
1186 if (err < 0) {
1187 write_seqcount_end(&devnet_rename_seq);
1188 return err;
1189 }
1190
1191 if (oldname[0] && !strchr(oldname, '%'))
1192 netdev_info(dev, "renamed from %s\n", oldname);
1193
1194 old_assign_type = dev->name_assign_type;
1195 dev->name_assign_type = NET_NAME_RENAMED;
1196
1197rollback:
1198 ret = device_rename(&dev->dev, dev->name);
1199 if (ret) {
1200 memcpy(dev->name, oldname, IFNAMSIZ);
1201 dev->name_assign_type = old_assign_type;
1202 write_seqcount_end(&devnet_rename_seq);
1203 return ret;
1204 }
1205
1206 write_seqcount_end(&devnet_rename_seq);
1207
1208 netdev_adjacent_rename_links(dev, oldname);
1209
1210 write_lock_bh(&dev_base_lock);
1211 hlist_del_rcu(&dev->name_hlist);
1212 write_unlock_bh(&dev_base_lock);
1213
1214 synchronize_rcu();
1215
1216 write_lock_bh(&dev_base_lock);
1217 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1218 write_unlock_bh(&dev_base_lock);
1219
1220 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1221 ret = notifier_to_errno(ret);
1222
1223 if (ret) {
1224 /* err >= 0 after dev_alloc_name() or stores the first errno */
1225 if (err >= 0) {
1226 err = ret;
1227 write_seqcount_begin(&devnet_rename_seq);
1228 memcpy(dev->name, oldname, IFNAMSIZ);
1229 memcpy(oldname, newname, IFNAMSIZ);
1230 dev->name_assign_type = old_assign_type;
1231 old_assign_type = NET_NAME_RENAMED;
1232 goto rollback;
1233 } else {
1234 pr_err("%s: name change rollback failed: %d\n",
1235 dev->name, ret);
1236 }
1237 }
1238
1239 return err;
1240}
1241
1242/**
1243 * dev_set_alias - change ifalias of a device
1244 * @dev: device
1245 * @alias: name up to IFALIASZ
1246 * @len: limit of bytes to copy from info
1247 *
1248 * Set ifalias for a device,
1249 */
1250int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1251{
1252 struct dev_ifalias *new_alias = NULL;
1253
1254 if (len >= IFALIASZ)
1255 return -EINVAL;
1256
1257 if (len) {
1258 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1259 if (!new_alias)
1260 return -ENOMEM;
1261
1262 memcpy(new_alias->ifalias, alias, len);
1263 new_alias->ifalias[len] = 0;
1264 }
1265
1266 mutex_lock(&ifalias_mutex);
1267 rcu_swap_protected(dev->ifalias, new_alias,
1268 mutex_is_locked(&ifalias_mutex));
1269 mutex_unlock(&ifalias_mutex);
1270
1271 if (new_alias)
1272 kfree_rcu(new_alias, rcuhead);
1273
1274 return len;
1275}
1276
1277/**
1278 * dev_get_alias - get ifalias of a device
1279 * @dev: device
1280 * @name: buffer to store name of ifalias
1281 * @len: size of buffer
1282 *
1283 * get ifalias for a device. Caller must make sure dev cannot go
1284 * away, e.g. rcu read lock or own a reference count to device.
1285 */
1286int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1287{
1288 const struct dev_ifalias *alias;
1289 int ret = 0;
1290
1291 rcu_read_lock();
1292 alias = rcu_dereference(dev->ifalias);
1293 if (alias)
1294 ret = snprintf(name, len, "%s", alias->ifalias);
1295 rcu_read_unlock();
1296
1297 return ret;
1298}
1299
1300/**
1301 * netdev_features_change - device changes features
1302 * @dev: device to cause notification
1303 *
1304 * Called to indicate a device has changed features.
1305 */
1306void netdev_features_change(struct net_device *dev)
1307{
1308 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1309}
1310EXPORT_SYMBOL(netdev_features_change);
1311
1312/**
1313 * netdev_state_change - device changes state
1314 * @dev: device to cause notification
1315 *
1316 * Called to indicate a device has changed state. This function calls
1317 * the notifier chains for netdev_chain and sends a NEWLINK message
1318 * to the routing socket.
1319 */
1320void netdev_state_change(struct net_device *dev)
1321{
1322 if (dev->flags & IFF_UP) {
1323 struct netdev_notifier_change_info change_info = {
1324 .info.dev = dev,
1325 };
1326
1327 call_netdevice_notifiers_info(NETDEV_CHANGE,
1328 &change_info.info);
1329 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1330 }
1331}
1332EXPORT_SYMBOL(netdev_state_change);
1333
1334/**
1335 * netdev_notify_peers - notify network peers about existence of @dev
1336 * @dev: network device
1337 *
1338 * Generate traffic such that interested network peers are aware of
1339 * @dev, such as by generating a gratuitous ARP. This may be used when
1340 * a device wants to inform the rest of the network about some sort of
1341 * reconfiguration such as a failover event or virtual machine
1342 * migration.
1343 */
1344void netdev_notify_peers(struct net_device *dev)
1345{
1346 rtnl_lock();
1347 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1348 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1349 rtnl_unlock();
1350}
1351EXPORT_SYMBOL(netdev_notify_peers);
1352
1353static int __dev_open(struct net_device *dev)
1354{
1355 const struct net_device_ops *ops = dev->netdev_ops;
1356 int ret;
1357
1358 ASSERT_RTNL();
1359
1360 if (!netif_device_present(dev))
1361 return -ENODEV;
1362
1363 /* Block netpoll from trying to do any rx path servicing.
1364 * If we don't do this there is a chance ndo_poll_controller
1365 * or ndo_poll may be running while we open the device
1366 */
1367 netpoll_poll_disable(dev);
1368
1369 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1370 ret = notifier_to_errno(ret);
1371 if (ret)
1372 return ret;
1373
1374 set_bit(__LINK_STATE_START, &dev->state);
1375
1376 if (ops->ndo_validate_addr)
1377 ret = ops->ndo_validate_addr(dev);
1378
1379 if (!ret && ops->ndo_open)
1380 ret = ops->ndo_open(dev);
1381
1382 netpoll_poll_enable(dev);
1383
1384 if (ret)
1385 clear_bit(__LINK_STATE_START, &dev->state);
1386 else {
1387 dev->flags |= IFF_UP;
1388 dev_set_rx_mode(dev);
1389 dev_activate(dev);
1390 add_device_randomness(dev->dev_addr, dev->addr_len);
1391 }
1392
1393 return ret;
1394}
1395
1396/**
1397 * dev_open - prepare an interface for use.
1398 * @dev: device to open
1399 *
1400 * Takes a device from down to up state. The device's private open
1401 * function is invoked and then the multicast lists are loaded. Finally
1402 * the device is moved into the up state and a %NETDEV_UP message is
1403 * sent to the netdev notifier chain.
1404 *
1405 * Calling this function on an active interface is a nop. On a failure
1406 * a negative errno code is returned.
1407 */
1408int dev_open(struct net_device *dev)
1409{
1410 int ret;
1411
1412 if (dev->flags & IFF_UP)
1413 return 0;
1414
1415 ret = __dev_open(dev);
1416 if (ret < 0)
1417 return ret;
1418
1419 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1420 call_netdevice_notifiers(NETDEV_UP, dev);
1421
1422 return ret;
1423}
1424EXPORT_SYMBOL(dev_open);
1425
1426static void __dev_close_many(struct list_head *head)
1427{
1428 struct net_device *dev;
1429
1430 ASSERT_RTNL();
1431 might_sleep();
1432
1433 list_for_each_entry(dev, head, close_list) {
1434 /* Temporarily disable netpoll until the interface is down */
1435 netpoll_poll_disable(dev);
1436
1437 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1438
1439 clear_bit(__LINK_STATE_START, &dev->state);
1440
1441 /* Synchronize to scheduled poll. We cannot touch poll list, it
1442 * can be even on different cpu. So just clear netif_running().
1443 *
1444 * dev->stop() will invoke napi_disable() on all of it's
1445 * napi_struct instances on this device.
1446 */
1447 smp_mb__after_atomic(); /* Commit netif_running(). */
1448 }
1449
1450 dev_deactivate_many(head);
1451
1452 list_for_each_entry(dev, head, close_list) {
1453 const struct net_device_ops *ops = dev->netdev_ops;
1454
1455 /*
1456 * Call the device specific close. This cannot fail.
1457 * Only if device is UP
1458 *
1459 * We allow it to be called even after a DETACH hot-plug
1460 * event.
1461 */
1462 if (ops->ndo_stop)
1463 ops->ndo_stop(dev);
1464
1465 dev->flags &= ~IFF_UP;
1466 netpoll_poll_enable(dev);
1467 }
1468}
1469
1470static void __dev_close(struct net_device *dev)
1471{
1472 LIST_HEAD(single);
1473
1474 list_add(&dev->close_list, &single);
1475 __dev_close_many(&single);
1476 list_del(&single);
1477}
1478
1479void dev_close_many(struct list_head *head, bool unlink)
1480{
1481 struct net_device *dev, *tmp;
1482
1483 /* Remove the devices that don't need to be closed */
1484 list_for_each_entry_safe(dev, tmp, head, close_list)
1485 if (!(dev->flags & IFF_UP))
1486 list_del_init(&dev->close_list);
1487
1488 __dev_close_many(head);
1489
1490 list_for_each_entry_safe(dev, tmp, head, close_list) {
1491 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1492 call_netdevice_notifiers(NETDEV_DOWN, dev);
1493 if (unlink)
1494 list_del_init(&dev->close_list);
1495 }
1496}
1497EXPORT_SYMBOL(dev_close_many);
1498
1499/**
1500 * dev_close - shutdown an interface.
1501 * @dev: device to shutdown
1502 *
1503 * This function moves an active device into down state. A
1504 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1505 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1506 * chain.
1507 */
1508void dev_close(struct net_device *dev)
1509{
1510 if (dev->flags & IFF_UP) {
1511 LIST_HEAD(single);
1512
1513 list_add(&dev->close_list, &single);
1514 dev_close_many(&single, true);
1515 list_del(&single);
1516 }
1517}
1518EXPORT_SYMBOL(dev_close);
1519
1520
1521/**
1522 * dev_disable_lro - disable Large Receive Offload on a device
1523 * @dev: device
1524 *
1525 * Disable Large Receive Offload (LRO) on a net device. Must be
1526 * called under RTNL. This is needed if received packets may be
1527 * forwarded to another interface.
1528 */
1529void dev_disable_lro(struct net_device *dev)
1530{
1531 struct net_device *lower_dev;
1532 struct list_head *iter;
1533
1534 dev->wanted_features &= ~NETIF_F_LRO;
1535 netdev_update_features(dev);
1536
1537 if (unlikely(dev->features & NETIF_F_LRO))
1538 netdev_WARN(dev, "failed to disable LRO!\n");
1539
1540 netdev_for_each_lower_dev(dev, lower_dev, iter)
1541 dev_disable_lro(lower_dev);
1542}
1543EXPORT_SYMBOL(dev_disable_lro);
1544
1545static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1546 struct net_device *dev)
1547{
1548 struct netdev_notifier_info info = {
1549 .dev = dev,
1550 };
1551
1552 return nb->notifier_call(nb, val, &info);
1553}
1554
1555static int dev_boot_phase = 1;
1556
1557/**
1558 * register_netdevice_notifier - register a network notifier block
1559 * @nb: notifier
1560 *
1561 * Register a notifier to be called when network device events occur.
1562 * The notifier passed is linked into the kernel structures and must
1563 * not be reused until it has been unregistered. A negative errno code
1564 * is returned on a failure.
1565 *
1566 * When registered all registration and up events are replayed
1567 * to the new notifier to allow device to have a race free
1568 * view of the network device list.
1569 */
1570
1571int register_netdevice_notifier(struct notifier_block *nb)
1572{
1573 struct net_device *dev;
1574 struct net_device *last;
1575 struct net *net;
1576 int err;
1577
1578 rtnl_lock();
1579 err = raw_notifier_chain_register(&netdev_chain, nb);
1580 if (err)
1581 goto unlock;
1582 if (dev_boot_phase)
1583 goto unlock;
1584 for_each_net(net) {
1585 for_each_netdev(net, dev) {
1586 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1587 err = notifier_to_errno(err);
1588 if (err)
1589 goto rollback;
1590
1591 if (!(dev->flags & IFF_UP))
1592 continue;
1593
1594 call_netdevice_notifier(nb, NETDEV_UP, dev);
1595 }
1596 }
1597
1598unlock:
1599 rtnl_unlock();
1600 return err;
1601
1602rollback:
1603 last = dev;
1604 for_each_net(net) {
1605 for_each_netdev(net, dev) {
1606 if (dev == last)
1607 goto outroll;
1608
1609 if (dev->flags & IFF_UP) {
1610 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1611 dev);
1612 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1613 }
1614 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1615 }
1616 }
1617
1618outroll:
1619 raw_notifier_chain_unregister(&netdev_chain, nb);
1620 goto unlock;
1621}
1622EXPORT_SYMBOL(register_netdevice_notifier);
1623
1624/**
1625 * unregister_netdevice_notifier - unregister a network notifier block
1626 * @nb: notifier
1627 *
1628 * Unregister a notifier previously registered by
1629 * register_netdevice_notifier(). The notifier is unlinked into the
1630 * kernel structures and may then be reused. A negative errno code
1631 * is returned on a failure.
1632 *
1633 * After unregistering unregister and down device events are synthesized
1634 * for all devices on the device list to the removed notifier to remove
1635 * the need for special case cleanup code.
1636 */
1637
1638int unregister_netdevice_notifier(struct notifier_block *nb)
1639{
1640 struct net_device *dev;
1641 struct net *net;
1642 int err;
1643
1644 rtnl_lock();
1645 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1646 if (err)
1647 goto unlock;
1648
1649 for_each_net(net) {
1650 for_each_netdev(net, dev) {
1651 if (dev->flags & IFF_UP) {
1652 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1653 dev);
1654 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1655 }
1656 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1657 }
1658 }
1659unlock:
1660 rtnl_unlock();
1661 return err;
1662}
1663EXPORT_SYMBOL(unregister_netdevice_notifier);
1664
1665/**
1666 * call_netdevice_notifiers_info - call all network notifier blocks
1667 * @val: value passed unmodified to notifier function
1668 * @dev: net_device pointer passed unmodified to notifier function
1669 * @info: notifier information data
1670 *
1671 * Call all network notifier blocks. Parameters and return value
1672 * are as for raw_notifier_call_chain().
1673 */
1674
1675static int call_netdevice_notifiers_info(unsigned long val,
1676 struct netdev_notifier_info *info)
1677{
1678 ASSERT_RTNL();
1679 return raw_notifier_call_chain(&netdev_chain, val, info);
1680}
1681
1682/**
1683 * call_netdevice_notifiers - call all network notifier blocks
1684 * @val: value passed unmodified to notifier function
1685 * @dev: net_device pointer passed unmodified to notifier function
1686 *
1687 * Call all network notifier blocks. Parameters and return value
1688 * are as for raw_notifier_call_chain().
1689 */
1690
1691int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1692{
1693 struct netdev_notifier_info info = {
1694 .dev = dev,
1695 };
1696
1697 return call_netdevice_notifiers_info(val, &info);
1698}
1699EXPORT_SYMBOL(call_netdevice_notifiers);
1700
1701#ifdef CONFIG_NET_INGRESS
1702static struct static_key ingress_needed __read_mostly;
1703
1704void net_inc_ingress_queue(void)
1705{
1706 static_key_slow_inc(&ingress_needed);
1707}
1708EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1709
1710void net_dec_ingress_queue(void)
1711{
1712 static_key_slow_dec(&ingress_needed);
1713}
1714EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1715#endif
1716
1717#ifdef CONFIG_NET_EGRESS
1718static struct static_key egress_needed __read_mostly;
1719
1720void net_inc_egress_queue(void)
1721{
1722 static_key_slow_inc(&egress_needed);
1723}
1724EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1725
1726void net_dec_egress_queue(void)
1727{
1728 static_key_slow_dec(&egress_needed);
1729}
1730EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1731#endif
1732
1733static struct static_key netstamp_needed __read_mostly;
1734#ifdef HAVE_JUMP_LABEL
1735static atomic_t netstamp_needed_deferred;
1736static atomic_t netstamp_wanted;
1737static void netstamp_clear(struct work_struct *work)
1738{
1739 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1740 int wanted;
1741
1742 wanted = atomic_add_return(deferred, &netstamp_wanted);
1743 if (wanted > 0)
1744 static_key_enable(&netstamp_needed);
1745 else
1746 static_key_disable(&netstamp_needed);
1747}
1748static DECLARE_WORK(netstamp_work, netstamp_clear);
1749#endif
1750
1751void net_enable_timestamp(void)
1752{
1753#ifdef HAVE_JUMP_LABEL
1754 int wanted;
1755
1756 while (1) {
1757 wanted = atomic_read(&netstamp_wanted);
1758 if (wanted <= 0)
1759 break;
1760 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1761 return;
1762 }
1763 atomic_inc(&netstamp_needed_deferred);
1764 schedule_work(&netstamp_work);
1765#else
1766 static_key_slow_inc(&netstamp_needed);
1767#endif
1768}
1769EXPORT_SYMBOL(net_enable_timestamp);
1770
1771void net_disable_timestamp(void)
1772{
1773#ifdef HAVE_JUMP_LABEL
1774 int wanted;
1775
1776 while (1) {
1777 wanted = atomic_read(&netstamp_wanted);
1778 if (wanted <= 1)
1779 break;
1780 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1781 return;
1782 }
1783 atomic_dec(&netstamp_needed_deferred);
1784 schedule_work(&netstamp_work);
1785#else
1786 static_key_slow_dec(&netstamp_needed);
1787#endif
1788}
1789EXPORT_SYMBOL(net_disable_timestamp);
1790
1791static inline void net_timestamp_set(struct sk_buff *skb)
1792{
1793 skb->tstamp = 0;
1794 if (static_key_false(&netstamp_needed))
1795 __net_timestamp(skb);
1796}
1797
1798#define net_timestamp_check(COND, SKB) \
1799 if (static_key_false(&netstamp_needed)) { \
1800 if ((COND) && !(SKB)->tstamp) \
1801 __net_timestamp(SKB); \
1802 } \
1803
1804bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1805{
1806 unsigned int len;
1807
1808 if (!(dev->flags & IFF_UP))
1809 return false;
1810
1811 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1812 if (skb->len <= len)
1813 return true;
1814
1815 /* if TSO is enabled, we don't care about the length as the packet
1816 * could be forwarded without being segmented before
1817 */
1818 if (skb_is_gso(skb))
1819 return true;
1820
1821 return false;
1822}
1823EXPORT_SYMBOL_GPL(is_skb_forwardable);
1824
1825int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1826{
1827 int ret = ____dev_forward_skb(dev, skb);
1828
1829 if (likely(!ret)) {
1830 skb->protocol = eth_type_trans(skb, dev);
1831 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1832 }
1833
1834 return ret;
1835}
1836EXPORT_SYMBOL_GPL(__dev_forward_skb);
1837
1838/**
1839 * dev_forward_skb - loopback an skb to another netif
1840 *
1841 * @dev: destination network device
1842 * @skb: buffer to forward
1843 *
1844 * return values:
1845 * NET_RX_SUCCESS (no congestion)
1846 * NET_RX_DROP (packet was dropped, but freed)
1847 *
1848 * dev_forward_skb can be used for injecting an skb from the
1849 * start_xmit function of one device into the receive queue
1850 * of another device.
1851 *
1852 * The receiving device may be in another namespace, so
1853 * we have to clear all information in the skb that could
1854 * impact namespace isolation.
1855 */
1856int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1857{
1858 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1859}
1860EXPORT_SYMBOL_GPL(dev_forward_skb);
1861
1862static inline int deliver_skb(struct sk_buff *skb,
1863 struct packet_type *pt_prev,
1864 struct net_device *orig_dev)
1865{
1866 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1867 return -ENOMEM;
1868 refcount_inc(&skb->users);
1869 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1870}
1871
1872static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1873 struct packet_type **pt,
1874 struct net_device *orig_dev,
1875 __be16 type,
1876 struct list_head *ptype_list)
1877{
1878 struct packet_type *ptype, *pt_prev = *pt;
1879
1880 list_for_each_entry_rcu(ptype, ptype_list, list) {
1881 if (ptype->type != type)
1882 continue;
1883 if (pt_prev)
1884 deliver_skb(skb, pt_prev, orig_dev);
1885 pt_prev = ptype;
1886 }
1887 *pt = pt_prev;
1888}
1889
1890static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1891{
1892 if (!ptype->af_packet_priv || !skb->sk)
1893 return false;
1894
1895 if (ptype->id_match)
1896 return ptype->id_match(ptype, skb->sk);
1897 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1898 return true;
1899
1900 return false;
1901}
1902
1903/*
1904 * Support routine. Sends outgoing frames to any network
1905 * taps currently in use.
1906 */
1907
1908void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1909{
1910 struct packet_type *ptype;
1911 struct sk_buff *skb2 = NULL;
1912 struct packet_type *pt_prev = NULL;
1913 struct list_head *ptype_list = &ptype_all;
1914
1915 rcu_read_lock();
1916again:
1917 list_for_each_entry_rcu(ptype, ptype_list, list) {
1918 /* Never send packets back to the socket
1919 * they originated from - MvS (miquels@drinkel.ow.org)
1920 */
1921 if (skb_loop_sk(ptype, skb))
1922 continue;
1923
1924 if (pt_prev) {
1925 deliver_skb(skb2, pt_prev, skb->dev);
1926 pt_prev = ptype;
1927 continue;
1928 }
1929
1930 /* need to clone skb, done only once */
1931 skb2 = skb_clone(skb, GFP_ATOMIC);
1932 if (!skb2)
1933 goto out_unlock;
1934
1935 net_timestamp_set(skb2);
1936
1937 /* skb->nh should be correctly
1938 * set by sender, so that the second statement is
1939 * just protection against buggy protocols.
1940 */
1941 skb_reset_mac_header(skb2);
1942
1943 if (skb_network_header(skb2) < skb2->data ||
1944 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1945 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1946 ntohs(skb2->protocol),
1947 dev->name);
1948 skb_reset_network_header(skb2);
1949 }
1950
1951 skb2->transport_header = skb2->network_header;
1952 skb2->pkt_type = PACKET_OUTGOING;
1953 pt_prev = ptype;
1954 }
1955
1956 if (ptype_list == &ptype_all) {
1957 ptype_list = &dev->ptype_all;
1958 goto again;
1959 }
1960out_unlock:
1961 if (pt_prev) {
1962 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
1963 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1964 else
1965 kfree_skb(skb2);
1966 }
1967 rcu_read_unlock();
1968}
1969EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
1970
1971/**
1972 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1973 * @dev: Network device
1974 * @txq: number of queues available
1975 *
1976 * If real_num_tx_queues is changed the tc mappings may no longer be
1977 * valid. To resolve this verify the tc mapping remains valid and if
1978 * not NULL the mapping. With no priorities mapping to this
1979 * offset/count pair it will no longer be used. In the worst case TC0
1980 * is invalid nothing can be done so disable priority mappings. If is
1981 * expected that drivers will fix this mapping if they can before
1982 * calling netif_set_real_num_tx_queues.
1983 */
1984static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1985{
1986 int i;
1987 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1988
1989 /* If TC0 is invalidated disable TC mapping */
1990 if (tc->offset + tc->count > txq) {
1991 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1992 dev->num_tc = 0;
1993 return;
1994 }
1995
1996 /* Invalidated prio to tc mappings set to TC0 */
1997 for (i = 1; i < TC_BITMASK + 1; i++) {
1998 int q = netdev_get_prio_tc_map(dev, i);
1999
2000 tc = &dev->tc_to_txq[q];
2001 if (tc->offset + tc->count > txq) {
2002 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2003 i, q);
2004 netdev_set_prio_tc_map(dev, i, 0);
2005 }
2006 }
2007}
2008
2009int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2010{
2011 if (dev->num_tc) {
2012 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2013 int i;
2014
2015 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2016 if ((txq - tc->offset) < tc->count)
2017 return i;
2018 }
2019
2020 return -1;
2021 }
2022
2023 return 0;
2024}
2025EXPORT_SYMBOL(netdev_txq_to_tc);
2026
2027#ifdef CONFIG_XPS
2028static DEFINE_MUTEX(xps_map_mutex);
2029#define xmap_dereference(P) \
2030 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2031
2032static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2033 int tci, u16 index)
2034{
2035 struct xps_map *map = NULL;
2036 int pos;
2037
2038 if (dev_maps)
2039 map = xmap_dereference(dev_maps->cpu_map[tci]);
2040 if (!map)
2041 return false;
2042
2043 for (pos = map->len; pos--;) {
2044 if (map->queues[pos] != index)
2045 continue;
2046
2047 if (map->len > 1) {
2048 map->queues[pos] = map->queues[--map->len];
2049 break;
2050 }
2051
2052 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2053 kfree_rcu(map, rcu);
2054 return false;
2055 }
2056
2057 return true;
2058}
2059
2060static bool remove_xps_queue_cpu(struct net_device *dev,
2061 struct xps_dev_maps *dev_maps,
2062 int cpu, u16 offset, u16 count)
2063{
2064 int num_tc = dev->num_tc ? : 1;
2065 bool active = false;
2066 int tci;
2067
2068 for (tci = cpu * num_tc; num_tc--; tci++) {
2069 int i, j;
2070
2071 for (i = count, j = offset; i--; j++) {
2072 if (!remove_xps_queue(dev_maps, cpu, j))
2073 break;
2074 }
2075
2076 active |= i < 0;
2077 }
2078
2079 return active;
2080}
2081
2082static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2083 u16 count)
2084{
2085 struct xps_dev_maps *dev_maps;
2086 int cpu, i;
2087 bool active = false;
2088
2089 mutex_lock(&xps_map_mutex);
2090 dev_maps = xmap_dereference(dev->xps_maps);
2091
2092 if (!dev_maps)
2093 goto out_no_maps;
2094
2095 for_each_possible_cpu(cpu)
2096 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2097 offset, count);
2098
2099 if (!active) {
2100 RCU_INIT_POINTER(dev->xps_maps, NULL);
2101 kfree_rcu(dev_maps, rcu);
2102 }
2103
2104 for (i = offset + (count - 1); count--; i--)
2105 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2106 NUMA_NO_NODE);
2107
2108out_no_maps:
2109 mutex_unlock(&xps_map_mutex);
2110}
2111
2112static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2113{
2114 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2115}
2116
2117static struct xps_map *expand_xps_map(struct xps_map *map,
2118 int cpu, u16 index)
2119{
2120 struct xps_map *new_map;
2121 int alloc_len = XPS_MIN_MAP_ALLOC;
2122 int i, pos;
2123
2124 for (pos = 0; map && pos < map->len; pos++) {
2125 if (map->queues[pos] != index)
2126 continue;
2127 return map;
2128 }
2129
2130 /* Need to add queue to this CPU's existing map */
2131 if (map) {
2132 if (pos < map->alloc_len)
2133 return map;
2134
2135 alloc_len = map->alloc_len * 2;
2136 }
2137
2138 /* Need to allocate new map to store queue on this CPU's map */
2139 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2140 cpu_to_node(cpu));
2141 if (!new_map)
2142 return NULL;
2143
2144 for (i = 0; i < pos; i++)
2145 new_map->queues[i] = map->queues[i];
2146 new_map->alloc_len = alloc_len;
2147 new_map->len = pos;
2148
2149 return new_map;
2150}
2151
2152int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2153 u16 index)
2154{
2155 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2156 int i, cpu, tci, numa_node_id = -2;
2157 int maps_sz, num_tc = 1, tc = 0;
2158 struct xps_map *map, *new_map;
2159 bool active = false;
2160
2161 if (dev->num_tc) {
2162 num_tc = dev->num_tc;
2163 tc = netdev_txq_to_tc(dev, index);
2164 if (tc < 0)
2165 return -EINVAL;
2166 }
2167
2168 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2169 if (maps_sz < L1_CACHE_BYTES)
2170 maps_sz = L1_CACHE_BYTES;
2171
2172 mutex_lock(&xps_map_mutex);
2173
2174 dev_maps = xmap_dereference(dev->xps_maps);
2175
2176 /* allocate memory for queue storage */
2177 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2178 if (!new_dev_maps)
2179 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2180 if (!new_dev_maps) {
2181 mutex_unlock(&xps_map_mutex);
2182 return -ENOMEM;
2183 }
2184
2185 tci = cpu * num_tc + tc;
2186 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2187 NULL;
2188
2189 map = expand_xps_map(map, cpu, index);
2190 if (!map)
2191 goto error;
2192
2193 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2194 }
2195
2196 if (!new_dev_maps)
2197 goto out_no_new_maps;
2198
2199 for_each_possible_cpu(cpu) {
2200 /* copy maps belonging to foreign traffic classes */
2201 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2202 /* fill in the new device map from the old device map */
2203 map = xmap_dereference(dev_maps->cpu_map[tci]);
2204 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2205 }
2206
2207 /* We need to explicitly update tci as prevous loop
2208 * could break out early if dev_maps is NULL.
2209 */
2210 tci = cpu * num_tc + tc;
2211
2212 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2213 /* add queue to CPU maps */
2214 int pos = 0;
2215
2216 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2217 while ((pos < map->len) && (map->queues[pos] != index))
2218 pos++;
2219
2220 if (pos == map->len)
2221 map->queues[map->len++] = index;
2222#ifdef CONFIG_NUMA
2223 if (numa_node_id == -2)
2224 numa_node_id = cpu_to_node(cpu);
2225 else if (numa_node_id != cpu_to_node(cpu))
2226 numa_node_id = -1;
2227#endif
2228 } else if (dev_maps) {
2229 /* fill in the new device map from the old device map */
2230 map = xmap_dereference(dev_maps->cpu_map[tci]);
2231 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2232 }
2233
2234 /* copy maps belonging to foreign traffic classes */
2235 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2236 /* fill in the new device map from the old device map */
2237 map = xmap_dereference(dev_maps->cpu_map[tci]);
2238 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2239 }
2240 }
2241
2242 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2243
2244 /* Cleanup old maps */
2245 if (!dev_maps)
2246 goto out_no_old_maps;
2247
2248 for_each_possible_cpu(cpu) {
2249 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2250 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2251 map = xmap_dereference(dev_maps->cpu_map[tci]);
2252 if (map && map != new_map)
2253 kfree_rcu(map, rcu);
2254 }
2255 }
2256
2257 kfree_rcu(dev_maps, rcu);
2258
2259out_no_old_maps:
2260 dev_maps = new_dev_maps;
2261 active = true;
2262
2263out_no_new_maps:
2264 /* update Tx queue numa node */
2265 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2266 (numa_node_id >= 0) ? numa_node_id :
2267 NUMA_NO_NODE);
2268
2269 if (!dev_maps)
2270 goto out_no_maps;
2271
2272 /* removes queue from unused CPUs */
2273 for_each_possible_cpu(cpu) {
2274 for (i = tc, tci = cpu * num_tc; i--; tci++)
2275 active |= remove_xps_queue(dev_maps, tci, index);
2276 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2277 active |= remove_xps_queue(dev_maps, tci, index);
2278 for (i = num_tc - tc, tci++; --i; tci++)
2279 active |= remove_xps_queue(dev_maps, tci, index);
2280 }
2281
2282 /* free map if not active */
2283 if (!active) {
2284 RCU_INIT_POINTER(dev->xps_maps, NULL);
2285 kfree_rcu(dev_maps, rcu);
2286 }
2287
2288out_no_maps:
2289 mutex_unlock(&xps_map_mutex);
2290
2291 return 0;
2292error:
2293 /* remove any maps that we added */
2294 for_each_possible_cpu(cpu) {
2295 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2296 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2297 map = dev_maps ?
2298 xmap_dereference(dev_maps->cpu_map[tci]) :
2299 NULL;
2300 if (new_map && new_map != map)
2301 kfree(new_map);
2302 }
2303 }
2304
2305 mutex_unlock(&xps_map_mutex);
2306
2307 kfree(new_dev_maps);
2308 return -ENOMEM;
2309}
2310EXPORT_SYMBOL(netif_set_xps_queue);
2311
2312#endif
2313void netdev_reset_tc(struct net_device *dev)
2314{
2315#ifdef CONFIG_XPS
2316 netif_reset_xps_queues_gt(dev, 0);
2317#endif
2318 dev->num_tc = 0;
2319 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2320 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2321}
2322EXPORT_SYMBOL(netdev_reset_tc);
2323
2324int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2325{
2326 if (tc >= dev->num_tc)
2327 return -EINVAL;
2328
2329#ifdef CONFIG_XPS
2330 netif_reset_xps_queues(dev, offset, count);
2331#endif
2332 dev->tc_to_txq[tc].count = count;
2333 dev->tc_to_txq[tc].offset = offset;
2334 return 0;
2335}
2336EXPORT_SYMBOL(netdev_set_tc_queue);
2337
2338int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2339{
2340 if (num_tc > TC_MAX_QUEUE)
2341 return -EINVAL;
2342
2343#ifdef CONFIG_XPS
2344 netif_reset_xps_queues_gt(dev, 0);
2345#endif
2346 dev->num_tc = num_tc;
2347 return 0;
2348}
2349EXPORT_SYMBOL(netdev_set_num_tc);
2350
2351/*
2352 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2353 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
2354 */
2355int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2356{
2357 int rc;
2358
2359 if (txq < 1 || txq > dev->num_tx_queues)
2360 return -EINVAL;
2361
2362 if (dev->reg_state == NETREG_REGISTERED ||
2363 dev->reg_state == NETREG_UNREGISTERING) {
2364 ASSERT_RTNL();
2365
2366 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2367 txq);
2368 if (rc)
2369 return rc;
2370
2371 if (dev->num_tc)
2372 netif_setup_tc(dev, txq);
2373
2374 if (txq < dev->real_num_tx_queues) {
2375 qdisc_reset_all_tx_gt(dev, txq);
2376#ifdef CONFIG_XPS
2377 netif_reset_xps_queues_gt(dev, txq);
2378#endif
2379 }
2380 }
2381
2382 dev->real_num_tx_queues = txq;
2383 return 0;
2384}
2385EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2386
2387#ifdef CONFIG_SYSFS
2388/**
2389 * netif_set_real_num_rx_queues - set actual number of RX queues used
2390 * @dev: Network device
2391 * @rxq: Actual number of RX queues
2392 *
2393 * This must be called either with the rtnl_lock held or before
2394 * registration of the net device. Returns 0 on success, or a
2395 * negative error code. If called before registration, it always
2396 * succeeds.
2397 */
2398int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2399{
2400 int rc;
2401
2402 if (rxq < 1 || rxq > dev->num_rx_queues)
2403 return -EINVAL;
2404
2405 if (dev->reg_state == NETREG_REGISTERED) {
2406 ASSERT_RTNL();
2407
2408 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2409 rxq);
2410 if (rc)
2411 return rc;
2412 }
2413
2414 dev->real_num_rx_queues = rxq;
2415 return 0;
2416}
2417EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2418#endif
2419
2420/**
2421 * netif_get_num_default_rss_queues - default number of RSS queues
2422 *
2423 * This routine should set an upper limit on the number of RSS queues
2424 * used by default by multiqueue devices.
2425 */
2426int netif_get_num_default_rss_queues(void)
2427{
2428 return is_kdump_kernel() ?
2429 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2430}
2431EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2432
2433static void __netif_reschedule(struct Qdisc *q)
2434{
2435 struct softnet_data *sd;
2436 unsigned long flags;
2437
2438 local_irq_save(flags);
2439 sd = this_cpu_ptr(&softnet_data);
2440 q->next_sched = NULL;
2441 *sd->output_queue_tailp = q;
2442 sd->output_queue_tailp = &q->next_sched;
2443 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2444 local_irq_restore(flags);
2445}
2446
2447void __netif_schedule(struct Qdisc *q)
2448{
2449 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2450 __netif_reschedule(q);
2451}
2452EXPORT_SYMBOL(__netif_schedule);
2453
2454struct dev_kfree_skb_cb {
2455 enum skb_free_reason reason;
2456};
2457
2458static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2459{
2460 return (struct dev_kfree_skb_cb *)skb->cb;
2461}
2462
2463void netif_schedule_queue(struct netdev_queue *txq)
2464{
2465 rcu_read_lock();
2466 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2467 struct Qdisc *q = rcu_dereference(txq->qdisc);
2468
2469 __netif_schedule(q);
2470 }
2471 rcu_read_unlock();
2472}
2473EXPORT_SYMBOL(netif_schedule_queue);
2474
2475void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2476{
2477 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2478 struct Qdisc *q;
2479
2480 rcu_read_lock();
2481 q = rcu_dereference(dev_queue->qdisc);
2482 __netif_schedule(q);
2483 rcu_read_unlock();
2484 }
2485}
2486EXPORT_SYMBOL(netif_tx_wake_queue);
2487
2488void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2489{
2490 unsigned long flags;
2491
2492 if (unlikely(!skb))
2493 return;
2494
2495 if (likely(refcount_read(&skb->users) == 1)) {
2496 smp_rmb();
2497 refcount_set(&skb->users, 0);
2498 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2499 return;
2500 }
2501 get_kfree_skb_cb(skb)->reason = reason;
2502 local_irq_save(flags);
2503 skb->next = __this_cpu_read(softnet_data.completion_queue);
2504 __this_cpu_write(softnet_data.completion_queue, skb);
2505 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2506 local_irq_restore(flags);
2507}
2508EXPORT_SYMBOL(__dev_kfree_skb_irq);
2509
2510void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2511{
2512 if (in_irq() || irqs_disabled())
2513 __dev_kfree_skb_irq(skb, reason);
2514 else
2515 dev_kfree_skb(skb);
2516}
2517EXPORT_SYMBOL(__dev_kfree_skb_any);
2518
2519
2520/**
2521 * netif_device_detach - mark device as removed
2522 * @dev: network device
2523 *
2524 * Mark device as removed from system and therefore no longer available.
2525 */
2526void netif_device_detach(struct net_device *dev)
2527{
2528 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2529 netif_running(dev)) {
2530 netif_tx_stop_all_queues(dev);
2531 }
2532}
2533EXPORT_SYMBOL(netif_device_detach);
2534
2535/**
2536 * netif_device_attach - mark device as attached
2537 * @dev: network device
2538 *
2539 * Mark device as attached from system and restart if needed.
2540 */
2541void netif_device_attach(struct net_device *dev)
2542{
2543 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2544 netif_running(dev)) {
2545 netif_tx_wake_all_queues(dev);
2546 __netdev_watchdog_up(dev);
2547 }
2548}
2549EXPORT_SYMBOL(netif_device_attach);
2550
2551/*
2552 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2553 * to be used as a distribution range.
2554 */
2555u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2556 unsigned int num_tx_queues)
2557{
2558 u32 hash;
2559 u16 qoffset = 0;
2560 u16 qcount = num_tx_queues;
2561
2562 if (skb_rx_queue_recorded(skb)) {
2563 hash = skb_get_rx_queue(skb);
2564 while (unlikely(hash >= num_tx_queues))
2565 hash -= num_tx_queues;
2566 return hash;
2567 }
2568
2569 if (dev->num_tc) {
2570 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2571
2572 qoffset = dev->tc_to_txq[tc].offset;
2573 qcount = dev->tc_to_txq[tc].count;
2574 }
2575
2576 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2577}
2578EXPORT_SYMBOL(__skb_tx_hash);
2579
2580static void skb_warn_bad_offload(const struct sk_buff *skb)
2581{
2582 static const netdev_features_t null_features;
2583 struct net_device *dev = skb->dev;
2584 const char *name = "";
2585
2586 if (!net_ratelimit())
2587 return;
2588
2589 if (dev) {
2590 if (dev->dev.parent)
2591 name = dev_driver_string(dev->dev.parent);
2592 else
2593 name = netdev_name(dev);
2594 }
2595 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2596 "gso_type=%d ip_summed=%d\n",
2597 name, dev ? &dev->features : &null_features,
2598 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2599 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2600 skb_shinfo(skb)->gso_type, skb->ip_summed);
2601}
2602
2603/*
2604 * Invalidate hardware checksum when packet is to be mangled, and
2605 * complete checksum manually on outgoing path.
2606 */
2607int skb_checksum_help(struct sk_buff *skb)
2608{
2609 __wsum csum;
2610 int ret = 0, offset;
2611
2612 if (skb->ip_summed == CHECKSUM_COMPLETE)
2613 goto out_set_summed;
2614
2615 if (unlikely(skb_shinfo(skb)->gso_size)) {
2616 skb_warn_bad_offload(skb);
2617 return -EINVAL;
2618 }
2619
2620 /* Before computing a checksum, we should make sure no frag could
2621 * be modified by an external entity : checksum could be wrong.
2622 */
2623 if (skb_has_shared_frag(skb)) {
2624 ret = __skb_linearize(skb);
2625 if (ret)
2626 goto out;
2627 }
2628
2629 offset = skb_checksum_start_offset(skb);
2630 BUG_ON(offset >= skb_headlen(skb));
2631 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2632
2633 offset += skb->csum_offset;
2634 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2635
2636 if (skb_cloned(skb) &&
2637 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2638 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2639 if (ret)
2640 goto out;
2641 }
2642
2643 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2644out_set_summed:
2645 skb->ip_summed = CHECKSUM_NONE;
2646out:
2647 return ret;
2648}
2649EXPORT_SYMBOL(skb_checksum_help);
2650
2651int skb_crc32c_csum_help(struct sk_buff *skb)
2652{
2653 __le32 crc32c_csum;
2654 int ret = 0, offset, start;
2655
2656 if (skb->ip_summed != CHECKSUM_PARTIAL)
2657 goto out;
2658
2659 if (unlikely(skb_is_gso(skb)))
2660 goto out;
2661
2662 /* Before computing a checksum, we should make sure no frag could
2663 * be modified by an external entity : checksum could be wrong.
2664 */
2665 if (unlikely(skb_has_shared_frag(skb))) {
2666 ret = __skb_linearize(skb);
2667 if (ret)
2668 goto out;
2669 }
2670 start = skb_checksum_start_offset(skb);
2671 offset = start + offsetof(struct sctphdr, checksum);
2672 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2673 ret = -EINVAL;
2674 goto out;
2675 }
2676 if (skb_cloned(skb) &&
2677 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2678 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2679 if (ret)
2680 goto out;
2681 }
2682 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2683 skb->len - start, ~(__u32)0,
2684 crc32c_csum_stub));
2685 *(__le32 *)(skb->data + offset) = crc32c_csum;
2686 skb->ip_summed = CHECKSUM_NONE;
2687 skb->csum_not_inet = 0;
2688out:
2689 return ret;
2690}
2691
2692__be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2693{
2694 __be16 type = skb->protocol;
2695
2696 /* Tunnel gso handlers can set protocol to ethernet. */
2697 if (type == htons(ETH_P_TEB)) {
2698 struct ethhdr *eth;
2699
2700 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2701 return 0;
2702
2703 eth = (struct ethhdr *)skb_mac_header(skb);
2704 type = eth->h_proto;
2705 }
2706
2707 return __vlan_get_protocol(skb, type, depth);
2708}
2709
2710/**
2711 * skb_mac_gso_segment - mac layer segmentation handler.
2712 * @skb: buffer to segment
2713 * @features: features for the output path (see dev->features)
2714 */
2715struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2716 netdev_features_t features)
2717{
2718 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2719 struct packet_offload *ptype;
2720 int vlan_depth = skb->mac_len;
2721 __be16 type = skb_network_protocol(skb, &vlan_depth);
2722
2723 if (unlikely(!type))
2724 return ERR_PTR(-EINVAL);
2725
2726 __skb_pull(skb, vlan_depth);
2727
2728 rcu_read_lock();
2729 list_for_each_entry_rcu(ptype, &offload_base, list) {
2730 if (ptype->type == type && ptype->callbacks.gso_segment) {
2731 segs = ptype->callbacks.gso_segment(skb, features);
2732 break;
2733 }
2734 }
2735 rcu_read_unlock();
2736
2737 __skb_push(skb, skb->data - skb_mac_header(skb));
2738
2739 return segs;
2740}
2741EXPORT_SYMBOL(skb_mac_gso_segment);
2742
2743
2744/* openvswitch calls this on rx path, so we need a different check.
2745 */
2746static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2747{
2748 if (tx_path)
2749 return skb->ip_summed != CHECKSUM_PARTIAL &&
2750 skb->ip_summed != CHECKSUM_UNNECESSARY;
2751
2752 return skb->ip_summed == CHECKSUM_NONE;
2753}
2754
2755/**
2756 * __skb_gso_segment - Perform segmentation on skb.
2757 * @skb: buffer to segment
2758 * @features: features for the output path (see dev->features)
2759 * @tx_path: whether it is called in TX path
2760 *
2761 * This function segments the given skb and returns a list of segments.
2762 *
2763 * It may return NULL if the skb requires no segmentation. This is
2764 * only possible when GSO is used for verifying header integrity.
2765 *
2766 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2767 */
2768struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2769 netdev_features_t features, bool tx_path)
2770{
2771 struct sk_buff *segs;
2772
2773 if (unlikely(skb_needs_check(skb, tx_path))) {
2774 int err;
2775
2776 /* We're going to init ->check field in TCP or UDP header */
2777 err = skb_cow_head(skb, 0);
2778 if (err < 0)
2779 return ERR_PTR(err);
2780 }
2781
2782 /* Only report GSO partial support if it will enable us to
2783 * support segmentation on this frame without needing additional
2784 * work.
2785 */
2786 if (features & NETIF_F_GSO_PARTIAL) {
2787 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2788 struct net_device *dev = skb->dev;
2789
2790 partial_features |= dev->features & dev->gso_partial_features;
2791 if (!skb_gso_ok(skb, features | partial_features))
2792 features &= ~NETIF_F_GSO_PARTIAL;
2793 }
2794
2795 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2796 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2797
2798 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2799 SKB_GSO_CB(skb)->encap_level = 0;
2800
2801 skb_reset_mac_header(skb);
2802 skb_reset_mac_len(skb);
2803
2804 segs = skb_mac_gso_segment(skb, features);
2805
2806 if (unlikely(skb_needs_check(skb, tx_path)))
2807 skb_warn_bad_offload(skb);
2808
2809 return segs;
2810}
2811EXPORT_SYMBOL(__skb_gso_segment);
2812
2813/* Take action when hardware reception checksum errors are detected. */
2814#ifdef CONFIG_BUG
2815void netdev_rx_csum_fault(struct net_device *dev)
2816{
2817 if (net_ratelimit()) {
2818 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2819 dump_stack();
2820 }
2821}
2822EXPORT_SYMBOL(netdev_rx_csum_fault);
2823#endif
2824
2825/* Actually, we should eliminate this check as soon as we know, that:
2826 * 1. IOMMU is present and allows to map all the memory.
2827 * 2. No high memory really exists on this machine.
2828 */
2829
2830static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2831{
2832#ifdef CONFIG_HIGHMEM
2833 int i;
2834
2835 if (!(dev->features & NETIF_F_HIGHDMA)) {
2836 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2837 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2838
2839 if (PageHighMem(skb_frag_page(frag)))
2840 return 1;
2841 }
2842 }
2843
2844 if (PCI_DMA_BUS_IS_PHYS) {
2845 struct device *pdev = dev->dev.parent;
2846
2847 if (!pdev)
2848 return 0;
2849 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2850 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2851 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2852
2853 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2854 return 1;
2855 }
2856 }
2857#endif
2858 return 0;
2859}
2860
2861/* If MPLS offload request, verify we are testing hardware MPLS features
2862 * instead of standard features for the netdev.
2863 */
2864#if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2865static netdev_features_t net_mpls_features(struct sk_buff *skb,
2866 netdev_features_t features,
2867 __be16 type)
2868{
2869 if (eth_p_mpls(type))
2870 features &= skb->dev->mpls_features;
2871
2872 return features;
2873}
2874#else
2875static netdev_features_t net_mpls_features(struct sk_buff *skb,
2876 netdev_features_t features,
2877 __be16 type)
2878{
2879 return features;
2880}
2881#endif
2882
2883static netdev_features_t harmonize_features(struct sk_buff *skb,
2884 netdev_features_t features)
2885{
2886 int tmp;
2887 __be16 type;
2888
2889 type = skb_network_protocol(skb, &tmp);
2890 features = net_mpls_features(skb, features, type);
2891
2892 if (skb->ip_summed != CHECKSUM_NONE &&
2893 !can_checksum_protocol(features, type)) {
2894 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2895 }
2896 if (illegal_highdma(skb->dev, skb))
2897 features &= ~NETIF_F_SG;
2898
2899 return features;
2900}
2901
2902netdev_features_t passthru_features_check(struct sk_buff *skb,
2903 struct net_device *dev,
2904 netdev_features_t features)
2905{
2906 return features;
2907}
2908EXPORT_SYMBOL(passthru_features_check);
2909
2910static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2911 struct net_device *dev,
2912 netdev_features_t features)
2913{
2914 return vlan_features_check(skb, features);
2915}
2916
2917static netdev_features_t gso_features_check(const struct sk_buff *skb,
2918 struct net_device *dev,
2919 netdev_features_t features)
2920{
2921 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2922
2923 if (gso_segs > dev->gso_max_segs)
2924 return features & ~NETIF_F_GSO_MASK;
2925
2926 /* Support for GSO partial features requires software
2927 * intervention before we can actually process the packets
2928 * so we need to strip support for any partial features now
2929 * and we can pull them back in after we have partially
2930 * segmented the frame.
2931 */
2932 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2933 features &= ~dev->gso_partial_features;
2934
2935 /* Make sure to clear the IPv4 ID mangling feature if the
2936 * IPv4 header has the potential to be fragmented.
2937 */
2938 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2939 struct iphdr *iph = skb->encapsulation ?
2940 inner_ip_hdr(skb) : ip_hdr(skb);
2941
2942 if (!(iph->frag_off & htons(IP_DF)))
2943 features &= ~NETIF_F_TSO_MANGLEID;
2944 }
2945
2946 return features;
2947}
2948
2949netdev_features_t netif_skb_features(struct sk_buff *skb)
2950{
2951 struct net_device *dev = skb->dev;
2952 netdev_features_t features = dev->features;
2953
2954 if (skb_is_gso(skb))
2955 features = gso_features_check(skb, dev, features);
2956
2957 /* If encapsulation offload request, verify we are testing
2958 * hardware encapsulation features instead of standard
2959 * features for the netdev
2960 */
2961 if (skb->encapsulation)
2962 features &= dev->hw_enc_features;
2963
2964 if (skb_vlan_tagged(skb))
2965 features = netdev_intersect_features(features,
2966 dev->vlan_features |
2967 NETIF_F_HW_VLAN_CTAG_TX |
2968 NETIF_F_HW_VLAN_STAG_TX);
2969
2970 if (dev->netdev_ops->ndo_features_check)
2971 features &= dev->netdev_ops->ndo_features_check(skb, dev,
2972 features);
2973 else
2974 features &= dflt_features_check(skb, dev, features);
2975
2976 return harmonize_features(skb, features);
2977}
2978EXPORT_SYMBOL(netif_skb_features);
2979
2980static int xmit_one(struct sk_buff *skb, struct net_device *dev,
2981 struct netdev_queue *txq, bool more)
2982{
2983 unsigned int len;
2984 int rc;
2985
2986 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
2987 dev_queue_xmit_nit(skb, dev);
2988
2989 len = skb->len;
2990 trace_net_dev_start_xmit(skb, dev);
2991 rc = netdev_start_xmit(skb, dev, txq, more);
2992 trace_net_dev_xmit(skb, rc, dev, len);
2993
2994 return rc;
2995}
2996
2997struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
2998 struct netdev_queue *txq, int *ret)
2999{
3000 struct sk_buff *skb = first;
3001 int rc = NETDEV_TX_OK;
3002
3003 while (skb) {
3004 struct sk_buff *next = skb->next;
3005
3006 skb->next = NULL;
3007 rc = xmit_one(skb, dev, txq, next != NULL);
3008 if (unlikely(!dev_xmit_complete(rc))) {
3009 skb->next = next;
3010 goto out;
3011 }
3012
3013 skb = next;
3014 if (netif_xmit_stopped(txq) && skb) {
3015 rc = NETDEV_TX_BUSY;
3016 break;
3017 }
3018 }
3019
3020out:
3021 *ret = rc;
3022 return skb;
3023}
3024
3025static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3026 netdev_features_t features)
3027{
3028 if (skb_vlan_tag_present(skb) &&
3029 !vlan_hw_offload_capable(features, skb->vlan_proto))
3030 skb = __vlan_hwaccel_push_inside(skb);
3031 return skb;
3032}
3033
3034int skb_csum_hwoffload_help(struct sk_buff *skb,
3035 const netdev_features_t features)
3036{
3037 if (unlikely(skb->csum_not_inet))
3038 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3039 skb_crc32c_csum_help(skb);
3040
3041 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3042}
3043EXPORT_SYMBOL(skb_csum_hwoffload_help);
3044
3045static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev)
3046{
3047 netdev_features_t features;
3048
3049 features = netif_skb_features(skb);
3050 skb = validate_xmit_vlan(skb, features);
3051 if (unlikely(!skb))
3052 goto out_null;
3053
3054 if (netif_needs_gso(skb, features)) {
3055 struct sk_buff *segs;
3056
3057 segs = skb_gso_segment(skb, features);
3058 if (IS_ERR(segs)) {
3059 goto out_kfree_skb;
3060 } else if (segs) {
3061 consume_skb(skb);
3062 skb = segs;
3063 }
3064 } else {
3065 if (skb_needs_linearize(skb, features) &&
3066 __skb_linearize(skb))
3067 goto out_kfree_skb;
3068
3069 if (validate_xmit_xfrm(skb, features))
3070 goto out_kfree_skb;
3071
3072 /* If packet is not checksummed and device does not
3073 * support checksumming for this protocol, complete
3074 * checksumming here.
3075 */
3076 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3077 if (skb->encapsulation)
3078 skb_set_inner_transport_header(skb,
3079 skb_checksum_start_offset(skb));
3080 else
3081 skb_set_transport_header(skb,
3082 skb_checksum_start_offset(skb));
3083 if (skb_csum_hwoffload_help(skb, features))
3084 goto out_kfree_skb;
3085 }
3086 }
3087
3088 return skb;
3089
3090out_kfree_skb:
3091 kfree_skb(skb);
3092out_null:
3093 atomic_long_inc(&dev->tx_dropped);
3094 return NULL;
3095}
3096
3097struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev)
3098{
3099 struct sk_buff *next, *head = NULL, *tail;
3100
3101 for (; skb != NULL; skb = next) {
3102 next = skb->next;
3103 skb->next = NULL;
3104
3105 /* in case skb wont be segmented, point to itself */
3106 skb->prev = skb;
3107
3108 skb = validate_xmit_skb(skb, dev);
3109 if (!skb)
3110 continue;
3111
3112 if (!head)
3113 head = skb;
3114 else
3115 tail->next = skb;
3116 /* If skb was segmented, skb->prev points to
3117 * the last segment. If not, it still contains skb.
3118 */
3119 tail = skb->prev;
3120 }
3121 return head;
3122}
3123EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3124
3125static void qdisc_pkt_len_init(struct sk_buff *skb)
3126{
3127 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3128
3129 qdisc_skb_cb(skb)->pkt_len = skb->len;
3130
3131 /* To get more precise estimation of bytes sent on wire,
3132 * we add to pkt_len the headers size of all segments
3133 */
3134 if (shinfo->gso_size) {
3135 unsigned int hdr_len;
3136 u16 gso_segs = shinfo->gso_segs;
3137
3138 /* mac layer + network layer */
3139 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3140
3141 /* + transport layer */
3142 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
3143 hdr_len += tcp_hdrlen(skb);
3144 else
3145 hdr_len += sizeof(struct udphdr);
3146
3147 if (shinfo->gso_type & SKB_GSO_DODGY)
3148 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3149 shinfo->gso_size);
3150
3151 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3152 }
3153}
3154
3155static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3156 struct net_device *dev,
3157 struct netdev_queue *txq)
3158{
3159 spinlock_t *root_lock = qdisc_lock(q);
3160 struct sk_buff *to_free = NULL;
3161 bool contended;
3162 int rc;
3163
3164 qdisc_calculate_pkt_len(skb, q);
3165 /*
3166 * Heuristic to force contended enqueues to serialize on a
3167 * separate lock before trying to get qdisc main lock.
3168 * This permits qdisc->running owner to get the lock more
3169 * often and dequeue packets faster.
3170 */
3171 contended = qdisc_is_running(q);
3172 if (unlikely(contended))
3173 spin_lock(&q->busylock);
3174
3175 spin_lock(root_lock);
3176 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3177 __qdisc_drop(skb, &to_free);
3178 rc = NET_XMIT_DROP;
3179 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3180 qdisc_run_begin(q)) {
3181 /*
3182 * This is a work-conserving queue; there are no old skbs
3183 * waiting to be sent out; and the qdisc is not running -
3184 * xmit the skb directly.
3185 */
3186
3187 qdisc_bstats_update(q, skb);
3188
3189 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3190 if (unlikely(contended)) {
3191 spin_unlock(&q->busylock);
3192 contended = false;
3193 }
3194 __qdisc_run(q);
3195 } else
3196 qdisc_run_end(q);
3197
3198 rc = NET_XMIT_SUCCESS;
3199 } else {
3200 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3201 if (qdisc_run_begin(q)) {
3202 if (unlikely(contended)) {
3203 spin_unlock(&q->busylock);
3204 contended = false;
3205 }
3206 __qdisc_run(q);
3207 }
3208 }
3209 spin_unlock(root_lock);
3210 if (unlikely(to_free))
3211 kfree_skb_list(to_free);
3212 if (unlikely(contended))
3213 spin_unlock(&q->busylock);
3214 return rc;
3215}
3216
3217#if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3218static void skb_update_prio(struct sk_buff *skb)
3219{
3220 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
3221
3222 if (!skb->priority && skb->sk && map) {
3223 unsigned int prioidx =
3224 sock_cgroup_prioidx(&skb->sk->sk_cgrp_data);
3225
3226 if (prioidx < map->priomap_len)
3227 skb->priority = map->priomap[prioidx];
3228 }
3229}
3230#else
3231#define skb_update_prio(skb)
3232#endif
3233
3234DEFINE_PER_CPU(int, xmit_recursion);
3235EXPORT_SYMBOL(xmit_recursion);
3236
3237/**
3238 * dev_loopback_xmit - loop back @skb
3239 * @net: network namespace this loopback is happening in
3240 * @sk: sk needed to be a netfilter okfn
3241 * @skb: buffer to transmit
3242 */
3243int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3244{
3245 skb_reset_mac_header(skb);
3246 __skb_pull(skb, skb_network_offset(skb));
3247 skb->pkt_type = PACKET_LOOPBACK;
3248 skb->ip_summed = CHECKSUM_UNNECESSARY;
3249 WARN_ON(!skb_dst(skb));
3250 skb_dst_force(skb);
3251 netif_rx_ni(skb);
3252 return 0;
3253}
3254EXPORT_SYMBOL(dev_loopback_xmit);
3255
3256#ifdef CONFIG_NET_EGRESS
3257static struct sk_buff *
3258sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3259{
3260 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3261 struct tcf_result cl_res;
3262
3263 if (!miniq)
3264 return skb;
3265
3266 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3267 mini_qdisc_bstats_cpu_update(miniq, skb);
3268
3269 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3270 case TC_ACT_OK:
3271 case TC_ACT_RECLASSIFY:
3272 skb->tc_index = TC_H_MIN(cl_res.classid);
3273 break;
3274 case TC_ACT_SHOT:
3275 mini_qdisc_qstats_cpu_drop(miniq);
3276 *ret = NET_XMIT_DROP;
3277 kfree_skb(skb);
3278 return NULL;
3279 case TC_ACT_STOLEN:
3280 case TC_ACT_QUEUED:
3281 case TC_ACT_TRAP:
3282 *ret = NET_XMIT_SUCCESS;
3283 consume_skb(skb);
3284 return NULL;
3285 case TC_ACT_REDIRECT:
3286 /* No need to push/pop skb's mac_header here on egress! */
3287 skb_do_redirect(skb);
3288 *ret = NET_XMIT_SUCCESS;
3289 return NULL;
3290 default:
3291 break;
3292 }
3293
3294 return skb;
3295}
3296#endif /* CONFIG_NET_EGRESS */
3297
3298static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3299{
3300#ifdef CONFIG_XPS
3301 struct xps_dev_maps *dev_maps;
3302 struct xps_map *map;
3303 int queue_index = -1;
3304
3305 rcu_read_lock();
3306 dev_maps = rcu_dereference(dev->xps_maps);
3307 if (dev_maps) {
3308 unsigned int tci = skb->sender_cpu - 1;
3309
3310 if (dev->num_tc) {
3311 tci *= dev->num_tc;
3312 tci += netdev_get_prio_tc_map(dev, skb->priority);
3313 }
3314
3315 map = rcu_dereference(dev_maps->cpu_map[tci]);
3316 if (map) {
3317 if (map->len == 1)
3318 queue_index = map->queues[0];
3319 else
3320 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3321 map->len)];
3322 if (unlikely(queue_index >= dev->real_num_tx_queues))
3323 queue_index = -1;
3324 }
3325 }
3326 rcu_read_unlock();
3327
3328 return queue_index;
3329#else
3330 return -1;
3331#endif
3332}
3333
3334static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3335{
3336 struct sock *sk = skb->sk;
3337 int queue_index = sk_tx_queue_get(sk);
3338
3339 if (queue_index < 0 || skb->ooo_okay ||
3340 queue_index >= dev->real_num_tx_queues) {
3341 int new_index = get_xps_queue(dev, skb);
3342
3343 if (new_index < 0)
3344 new_index = skb_tx_hash(dev, skb);
3345
3346 if (queue_index != new_index && sk &&
3347 sk_fullsock(sk) &&
3348 rcu_access_pointer(sk->sk_dst_cache))
3349 sk_tx_queue_set(sk, new_index);
3350
3351 queue_index = new_index;
3352 }
3353
3354 return queue_index;
3355}
3356
3357struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3358 struct sk_buff *skb,
3359 void *accel_priv)
3360{
3361 int queue_index = 0;
3362
3363#ifdef CONFIG_XPS
3364 u32 sender_cpu = skb->sender_cpu - 1;
3365
3366 if (sender_cpu >= (u32)NR_CPUS)
3367 skb->sender_cpu = raw_smp_processor_id() + 1;
3368#endif
3369
3370 if (dev->real_num_tx_queues != 1) {
3371 const struct net_device_ops *ops = dev->netdev_ops;
3372
3373 if (ops->ndo_select_queue)
3374 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3375 __netdev_pick_tx);
3376 else
3377 queue_index = __netdev_pick_tx(dev, skb);
3378
3379 if (!accel_priv)
3380 queue_index = netdev_cap_txqueue(dev, queue_index);
3381 }
3382
3383 skb_set_queue_mapping(skb, queue_index);
3384 return netdev_get_tx_queue(dev, queue_index);
3385}
3386
3387/**
3388 * __dev_queue_xmit - transmit a buffer
3389 * @skb: buffer to transmit
3390 * @accel_priv: private data used for L2 forwarding offload
3391 *
3392 * Queue a buffer for transmission to a network device. The caller must
3393 * have set the device and priority and built the buffer before calling
3394 * this function. The function can be called from an interrupt.
3395 *
3396 * A negative errno code is returned on a failure. A success does not
3397 * guarantee the frame will be transmitted as it may be dropped due
3398 * to congestion or traffic shaping.
3399 *
3400 * -----------------------------------------------------------------------------------
3401 * I notice this method can also return errors from the queue disciplines,
3402 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3403 * be positive.
3404 *
3405 * Regardless of the return value, the skb is consumed, so it is currently
3406 * difficult to retry a send to this method. (You can bump the ref count
3407 * before sending to hold a reference for retry if you are careful.)
3408 *
3409 * When calling this method, interrupts MUST be enabled. This is because
3410 * the BH enable code must have IRQs enabled so that it will not deadlock.
3411 * --BLG
3412 */
3413static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3414{
3415 struct net_device *dev = skb->dev;
3416 struct netdev_queue *txq;
3417 struct Qdisc *q;
3418 int rc = -ENOMEM;
3419
3420 skb_reset_mac_header(skb);
3421
3422 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3423 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3424
3425 /* Disable soft irqs for various locks below. Also
3426 * stops preemption for RCU.
3427 */
3428 rcu_read_lock_bh();
3429
3430 skb_update_prio(skb);
3431
3432 qdisc_pkt_len_init(skb);
3433#ifdef CONFIG_NET_CLS_ACT
3434 skb->tc_at_ingress = 0;
3435# ifdef CONFIG_NET_EGRESS
3436 if (static_key_false(&egress_needed)) {
3437 skb = sch_handle_egress(skb, &rc, dev);
3438 if (!skb)
3439 goto out;
3440 }
3441# endif
3442#endif
3443 /* If device/qdisc don't need skb->dst, release it right now while
3444 * its hot in this cpu cache.
3445 */
3446 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3447 skb_dst_drop(skb);
3448 else
3449 skb_dst_force(skb);
3450
3451 txq = netdev_pick_tx(dev, skb, accel_priv);
3452 q = rcu_dereference_bh(txq->qdisc);
3453
3454 trace_net_dev_queue(skb);
3455 if (q->enqueue) {
3456 rc = __dev_xmit_skb(skb, q, dev, txq);
3457 goto out;
3458 }
3459
3460 /* The device has no queue. Common case for software devices:
3461 * loopback, all the sorts of tunnels...
3462
3463 * Really, it is unlikely that netif_tx_lock protection is necessary
3464 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3465 * counters.)
3466 * However, it is possible, that they rely on protection
3467 * made by us here.
3468
3469 * Check this and shot the lock. It is not prone from deadlocks.
3470 *Either shot noqueue qdisc, it is even simpler 8)
3471 */
3472 if (dev->flags & IFF_UP) {
3473 int cpu = smp_processor_id(); /* ok because BHs are off */
3474
3475 if (txq->xmit_lock_owner != cpu) {
3476 if (unlikely(__this_cpu_read(xmit_recursion) >
3477 XMIT_RECURSION_LIMIT))
3478 goto recursion_alert;
3479
3480 skb = validate_xmit_skb(skb, dev);
3481 if (!skb)
3482 goto out;
3483
3484 HARD_TX_LOCK(dev, txq, cpu);
3485
3486 if (!netif_xmit_stopped(txq)) {
3487 __this_cpu_inc(xmit_recursion);
3488 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3489 __this_cpu_dec(xmit_recursion);
3490 if (dev_xmit_complete(rc)) {
3491 HARD_TX_UNLOCK(dev, txq);
3492 goto out;
3493 }
3494 }
3495 HARD_TX_UNLOCK(dev, txq);
3496 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3497 dev->name);
3498 } else {
3499 /* Recursion is detected! It is possible,
3500 * unfortunately
3501 */
3502recursion_alert:
3503 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3504 dev->name);
3505 }
3506 }
3507
3508 rc = -ENETDOWN;
3509 rcu_read_unlock_bh();
3510
3511 atomic_long_inc(&dev->tx_dropped);
3512 kfree_skb_list(skb);
3513 return rc;
3514out:
3515 rcu_read_unlock_bh();
3516 return rc;
3517}
3518
3519int dev_queue_xmit(struct sk_buff *skb)
3520{
3521 return __dev_queue_xmit(skb, NULL);
3522}
3523EXPORT_SYMBOL(dev_queue_xmit);
3524
3525int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3526{
3527 return __dev_queue_xmit(skb, accel_priv);
3528}
3529EXPORT_SYMBOL(dev_queue_xmit_accel);
3530
3531
3532/*************************************************************************
3533 * Receiver routines
3534 *************************************************************************/
3535
3536int netdev_max_backlog __read_mostly = 1000;
3537EXPORT_SYMBOL(netdev_max_backlog);
3538
3539int netdev_tstamp_prequeue __read_mostly = 1;
3540int netdev_budget __read_mostly = 300;
3541unsigned int __read_mostly netdev_budget_usecs = 2000;
3542int weight_p __read_mostly = 64; /* old backlog weight */
3543int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3544int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3545int dev_rx_weight __read_mostly = 64;
3546int dev_tx_weight __read_mostly = 64;
3547
3548/* Called with irq disabled */
3549static inline void ____napi_schedule(struct softnet_data *sd,
3550 struct napi_struct *napi)
3551{
3552 list_add_tail(&napi->poll_list, &sd->poll_list);
3553 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3554}
3555
3556#ifdef CONFIG_RPS
3557
3558/* One global table that all flow-based protocols share. */
3559struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3560EXPORT_SYMBOL(rps_sock_flow_table);
3561u32 rps_cpu_mask __read_mostly;
3562EXPORT_SYMBOL(rps_cpu_mask);
3563
3564struct static_key rps_needed __read_mostly;
3565EXPORT_SYMBOL(rps_needed);
3566struct static_key rfs_needed __read_mostly;
3567EXPORT_SYMBOL(rfs_needed);
3568
3569static struct rps_dev_flow *
3570set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3571 struct rps_dev_flow *rflow, u16 next_cpu)
3572{
3573 if (next_cpu < nr_cpu_ids) {
3574#ifdef CONFIG_RFS_ACCEL
3575 struct netdev_rx_queue *rxqueue;
3576 struct rps_dev_flow_table *flow_table;
3577 struct rps_dev_flow *old_rflow;
3578 u32 flow_id;
3579 u16 rxq_index;
3580 int rc;
3581
3582 /* Should we steer this flow to a different hardware queue? */
3583 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3584 !(dev->features & NETIF_F_NTUPLE))
3585 goto out;
3586 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3587 if (rxq_index == skb_get_rx_queue(skb))
3588 goto out;
3589
3590 rxqueue = dev->_rx + rxq_index;
3591 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3592 if (!flow_table)
3593 goto out;
3594 flow_id = skb_get_hash(skb) & flow_table->mask;
3595 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3596 rxq_index, flow_id);
3597 if (rc < 0)
3598 goto out;
3599 old_rflow = rflow;
3600 rflow = &flow_table->flows[flow_id];
3601 rflow->filter = rc;
3602 if (old_rflow->filter == rflow->filter)
3603 old_rflow->filter = RPS_NO_FILTER;
3604 out:
3605#endif
3606 rflow->last_qtail =
3607 per_cpu(softnet_data, next_cpu).input_queue_head;
3608 }
3609
3610 rflow->cpu = next_cpu;
3611 return rflow;
3612}
3613
3614/*
3615 * get_rps_cpu is called from netif_receive_skb and returns the target
3616 * CPU from the RPS map of the receiving queue for a given skb.
3617 * rcu_read_lock must be held on entry.
3618 */
3619static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3620 struct rps_dev_flow **rflowp)
3621{
3622 const struct rps_sock_flow_table *sock_flow_table;
3623 struct netdev_rx_queue *rxqueue = dev->_rx;
3624 struct rps_dev_flow_table *flow_table;
3625 struct rps_map *map;
3626 int cpu = -1;
3627 u32 tcpu;
3628 u32 hash;
3629
3630 if (skb_rx_queue_recorded(skb)) {
3631 u16 index = skb_get_rx_queue(skb);
3632
3633 if (unlikely(index >= dev->real_num_rx_queues)) {
3634 WARN_ONCE(dev->real_num_rx_queues > 1,
3635 "%s received packet on queue %u, but number "
3636 "of RX queues is %u\n",
3637 dev->name, index, dev->real_num_rx_queues);
3638 goto done;
3639 }
3640 rxqueue += index;
3641 }
3642
3643 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3644
3645 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3646 map = rcu_dereference(rxqueue->rps_map);
3647 if (!flow_table && !map)
3648 goto done;
3649
3650 skb_reset_network_header(skb);
3651 hash = skb_get_hash(skb);
3652 if (!hash)
3653 goto done;
3654
3655 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3656 if (flow_table && sock_flow_table) {
3657 struct rps_dev_flow *rflow;
3658 u32 next_cpu;
3659 u32 ident;
3660
3661 /* First check into global flow table if there is a match */
3662 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3663 if ((ident ^ hash) & ~rps_cpu_mask)
3664 goto try_rps;
3665
3666 next_cpu = ident & rps_cpu_mask;
3667
3668 /* OK, now we know there is a match,
3669 * we can look at the local (per receive queue) flow table
3670 */
3671 rflow = &flow_table->flows[hash & flow_table->mask];
3672 tcpu = rflow->cpu;
3673
3674 /*
3675 * If the desired CPU (where last recvmsg was done) is
3676 * different from current CPU (one in the rx-queue flow
3677 * table entry), switch if one of the following holds:
3678 * - Current CPU is unset (>= nr_cpu_ids).
3679 * - Current CPU is offline.
3680 * - The current CPU's queue tail has advanced beyond the
3681 * last packet that was enqueued using this table entry.
3682 * This guarantees that all previous packets for the flow
3683 * have been dequeued, thus preserving in order delivery.
3684 */
3685 if (unlikely(tcpu != next_cpu) &&
3686 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3687 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3688 rflow->last_qtail)) >= 0)) {
3689 tcpu = next_cpu;
3690 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3691 }
3692
3693 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3694 *rflowp = rflow;
3695 cpu = tcpu;
3696 goto done;
3697 }
3698 }
3699
3700try_rps:
3701
3702 if (map) {
3703 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3704 if (cpu_online(tcpu)) {
3705 cpu = tcpu;
3706 goto done;
3707 }
3708 }
3709
3710done:
3711 return cpu;
3712}
3713
3714#ifdef CONFIG_RFS_ACCEL
3715
3716/**
3717 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3718 * @dev: Device on which the filter was set
3719 * @rxq_index: RX queue index
3720 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3721 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3722 *
3723 * Drivers that implement ndo_rx_flow_steer() should periodically call
3724 * this function for each installed filter and remove the filters for
3725 * which it returns %true.
3726 */
3727bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3728 u32 flow_id, u16 filter_id)
3729{
3730 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3731 struct rps_dev_flow_table *flow_table;
3732 struct rps_dev_flow *rflow;
3733 bool expire = true;
3734 unsigned int cpu;
3735
3736 rcu_read_lock();
3737 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3738 if (flow_table && flow_id <= flow_table->mask) {
3739 rflow = &flow_table->flows[flow_id];
3740 cpu = READ_ONCE(rflow->cpu);
3741 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3742 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3743 rflow->last_qtail) <
3744 (int)(10 * flow_table->mask)))
3745 expire = false;
3746 }
3747 rcu_read_unlock();
3748 return expire;
3749}
3750EXPORT_SYMBOL(rps_may_expire_flow);
3751
3752#endif /* CONFIG_RFS_ACCEL */
3753
3754/* Called from hardirq (IPI) context */
3755static void rps_trigger_softirq(void *data)
3756{
3757 struct softnet_data *sd = data;
3758
3759 ____napi_schedule(sd, &sd->backlog);
3760 sd->received_rps++;
3761}
3762
3763#endif /* CONFIG_RPS */
3764
3765/*
3766 * Check if this softnet_data structure is another cpu one
3767 * If yes, queue it to our IPI list and return 1
3768 * If no, return 0
3769 */
3770static int rps_ipi_queued(struct softnet_data *sd)
3771{
3772#ifdef CONFIG_RPS
3773 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3774
3775 if (sd != mysd) {
3776 sd->rps_ipi_next = mysd->rps_ipi_list;
3777 mysd->rps_ipi_list = sd;
3778
3779 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3780 return 1;
3781 }
3782#endif /* CONFIG_RPS */
3783 return 0;
3784}
3785
3786#ifdef CONFIG_NET_FLOW_LIMIT
3787int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3788#endif
3789
3790static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3791{
3792#ifdef CONFIG_NET_FLOW_LIMIT
3793 struct sd_flow_limit *fl;
3794 struct softnet_data *sd;
3795 unsigned int old_flow, new_flow;
3796
3797 if (qlen < (netdev_max_backlog >> 1))
3798 return false;
3799
3800 sd = this_cpu_ptr(&softnet_data);
3801
3802 rcu_read_lock();
3803 fl = rcu_dereference(sd->flow_limit);
3804 if (fl) {
3805 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3806 old_flow = fl->history[fl->history_head];
3807 fl->history[fl->history_head] = new_flow;
3808
3809 fl->history_head++;
3810 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3811
3812 if (likely(fl->buckets[old_flow]))
3813 fl->buckets[old_flow]--;
3814
3815 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3816 fl->count++;
3817 rcu_read_unlock();
3818 return true;
3819 }
3820 }
3821 rcu_read_unlock();
3822#endif
3823 return false;
3824}
3825
3826/*
3827 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3828 * queue (may be a remote CPU queue).
3829 */
3830static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3831 unsigned int *qtail)
3832{
3833 struct softnet_data *sd;
3834 unsigned long flags;
3835 unsigned int qlen;
3836
3837 sd = &per_cpu(softnet_data, cpu);
3838
3839 local_irq_save(flags);
3840
3841 rps_lock(sd);
3842 if (!netif_running(skb->dev))
3843 goto drop;
3844 qlen = skb_queue_len(&sd->input_pkt_queue);
3845 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3846 if (qlen) {
3847enqueue:
3848 __skb_queue_tail(&sd->input_pkt_queue, skb);
3849 input_queue_tail_incr_save(sd, qtail);
3850 rps_unlock(sd);
3851 local_irq_restore(flags);
3852 return NET_RX_SUCCESS;
3853 }
3854
3855 /* Schedule NAPI for backlog device
3856 * We can use non atomic operation since we own the queue lock
3857 */
3858 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3859 if (!rps_ipi_queued(sd))
3860 ____napi_schedule(sd, &sd->backlog);
3861 }
3862 goto enqueue;
3863 }
3864
3865drop:
3866 sd->dropped++;
3867 rps_unlock(sd);
3868
3869 local_irq_restore(flags);
3870
3871 atomic_long_inc(&skb->dev->rx_dropped);
3872 kfree_skb(skb);
3873 return NET_RX_DROP;
3874}
3875
3876static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3877 struct bpf_prog *xdp_prog)
3878{
3879 u32 metalen, act = XDP_DROP;
3880 struct xdp_buff xdp;
3881 void *orig_data;
3882 int hlen, off;
3883 u32 mac_len;
3884
3885 /* Reinjected packets coming from act_mirred or similar should
3886 * not get XDP generic processing.
3887 */
3888 if (skb_cloned(skb))
3889 return XDP_PASS;
3890
3891 /* XDP packets must be linear and must have sufficient headroom
3892 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
3893 * native XDP provides, thus we need to do it here as well.
3894 */
3895 if (skb_is_nonlinear(skb) ||
3896 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
3897 int hroom = XDP_PA