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
148#include "net-sysfs.h"
149
150/* Instead of increasing this, you should create a hash table. */
151#define MAX_GRO_SKBS 8
152
153/* This should be increased if a protocol with a bigger head is added. */
154#define GRO_MAX_HEAD (MAX_HEADER + 128)
155
156static DEFINE_SPINLOCK(ptype_lock);
157static DEFINE_SPINLOCK(offload_lock);
158struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
159struct list_head ptype_all __read_mostly; /* Taps */
160static struct list_head offload_base __read_mostly;
161
162static int netif_rx_internal(struct sk_buff *skb);
163static int call_netdevice_notifiers_info(unsigned long val,
164 struct net_device *dev,
165 struct netdev_notifier_info *info);
166static struct napi_struct *napi_by_id(unsigned int napi_id);
167
168/*
169 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
170 * semaphore.
171 *
172 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
173 *
174 * Writers must hold the rtnl semaphore while they loop through the
175 * dev_base_head list, and hold dev_base_lock for writing when they do the
176 * actual updates. This allows pure readers to access the list even
177 * while a writer is preparing to update it.
178 *
179 * To put it another way, dev_base_lock is held for writing only to
180 * protect against pure readers; the rtnl semaphore provides the
181 * protection against other writers.
182 *
183 * See, for example usages, register_netdevice() and
184 * unregister_netdevice(), which must be called with the rtnl
185 * semaphore held.
186 */
187DEFINE_RWLOCK(dev_base_lock);
188EXPORT_SYMBOL(dev_base_lock);
189
190/* protects napi_hash addition/deletion and napi_gen_id */
191static DEFINE_SPINLOCK(napi_hash_lock);
192
193static unsigned int napi_gen_id = NR_CPUS;
194static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
195
196static seqcount_t devnet_rename_seq;
197
198static inline void dev_base_seq_inc(struct net *net)
199{
200 while (++net->dev_base_seq == 0)
201 ;
202}
203
204static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
205{
206 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
207
208 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
209}
210
211static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
212{
213 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
214}
215
216static inline void rps_lock(struct softnet_data *sd)
217{
218#ifdef CONFIG_RPS
219 spin_lock(&sd->input_pkt_queue.lock);
220#endif
221}
222
223static inline void rps_unlock(struct softnet_data *sd)
224{
225#ifdef CONFIG_RPS
226 spin_unlock(&sd->input_pkt_queue.lock);
227#endif
228}
229
230/* Device list insertion */
231static void list_netdevice(struct net_device *dev)
232{
233 struct net *net = dev_net(dev);
234
235 ASSERT_RTNL();
236
237 write_lock_bh(&dev_base_lock);
238 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
239 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
240 hlist_add_head_rcu(&dev->index_hlist,
241 dev_index_hash(net, dev->ifindex));
242 write_unlock_bh(&dev_base_lock);
243
244 dev_base_seq_inc(net);
245}
246
247/* Device list removal
248 * caller must respect a RCU grace period before freeing/reusing dev
249 */
250static void unlist_netdevice(struct net_device *dev)
251{
252 ASSERT_RTNL();
253
254 /* Unlink dev from the device chain */
255 write_lock_bh(&dev_base_lock);
256 list_del_rcu(&dev->dev_list);
257 hlist_del_rcu(&dev->name_hlist);
258 hlist_del_rcu(&dev->index_hlist);
259 write_unlock_bh(&dev_base_lock);
260
261 dev_base_seq_inc(dev_net(dev));
262}
263
264/*
265 * Our notifier list
266 */
267
268static RAW_NOTIFIER_HEAD(netdev_chain);
269
270/*
271 * Device drivers call our routines to queue packets here. We empty the
272 * queue in the local softnet handler.
273 */
274
275DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
276EXPORT_PER_CPU_SYMBOL(softnet_data);
277
278#ifdef CONFIG_LOCKDEP
279/*
280 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
281 * according to dev->type
282 */
283static const unsigned short netdev_lock_type[] = {
284 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
285 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
286 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
287 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
288 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
289 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
290 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
291 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
292 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
293 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
294 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
295 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
296 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
297 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
298 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
299
300static const char *const netdev_lock_name[] = {
301 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
302 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
303 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
304 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
305 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
306 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
307 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
308 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
309 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
310 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
311 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
312 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
313 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
314 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
315 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
316
317static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
318static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
319
320static inline unsigned short netdev_lock_pos(unsigned short dev_type)
321{
322 int i;
323
324 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
325 if (netdev_lock_type[i] == dev_type)
326 return i;
327 /* the last key is used by default */
328 return ARRAY_SIZE(netdev_lock_type) - 1;
329}
330
331static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
332 unsigned short dev_type)
333{
334 int i;
335
336 i = netdev_lock_pos(dev_type);
337 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
338 netdev_lock_name[i]);
339}
340
341static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
342{
343 int i;
344
345 i = netdev_lock_pos(dev->type);
346 lockdep_set_class_and_name(&dev->addr_list_lock,
347 &netdev_addr_lock_key[i],
348 netdev_lock_name[i]);
349}
350#else
351static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
352 unsigned short dev_type)
353{
354}
355static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
356{
357}
358#endif
359
360/*******************************************************************************
361 *
362 * Protocol management and registration routines
363 *
364 *******************************************************************************/
365
366
367/*
368 * Add a protocol ID to the list. Now that the input handler is
369 * smarter we can dispense with all the messy stuff that used to be
370 * here.
371 *
372 * BEWARE!!! Protocol handlers, mangling input packets,
373 * MUST BE last in hash buckets and checking protocol handlers
374 * MUST start from promiscuous ptype_all chain in net_bh.
375 * It is true now, do not change it.
376 * Explanation follows: if protocol handler, mangling packet, will
377 * be the first on list, it is not able to sense, that packet
378 * is cloned and should be copied-on-write, so that it will
379 * change it and subsequent readers will get broken packet.
380 * --ANK (980803)
381 */
382
383static inline struct list_head *ptype_head(const struct packet_type *pt)
384{
385 if (pt->type == htons(ETH_P_ALL))
386 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
387 else
388 return pt->dev ? &pt->dev->ptype_specific :
389 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
390}
391
392/**
393 * dev_add_pack - add packet handler
394 * @pt: packet type declaration
395 *
396 * Add a protocol handler to the networking stack. The passed &packet_type
397 * is linked into kernel lists and may not be freed until it has been
398 * removed from the kernel lists.
399 *
400 * This call does not sleep therefore it can not
401 * guarantee all CPU's that are in middle of receiving packets
402 * will see the new packet type (until the next received packet).
403 */
404
405void dev_add_pack(struct packet_type *pt)
406{
407 struct list_head *head = ptype_head(pt);
408
409 spin_lock(&ptype_lock);
410 list_add_rcu(&pt->list, head);
411 spin_unlock(&ptype_lock);
412}
413EXPORT_SYMBOL(dev_add_pack);
414
415/**
416 * __dev_remove_pack - remove packet handler
417 * @pt: packet type declaration
418 *
419 * Remove a protocol handler that was previously added to the kernel
420 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
421 * from the kernel lists and can be freed or reused once this function
422 * returns.
423 *
424 * The packet type might still be in use by receivers
425 * and must not be freed until after all the CPU's have gone
426 * through a quiescent state.
427 */
428void __dev_remove_pack(struct packet_type *pt)
429{
430 struct list_head *head = ptype_head(pt);
431 struct packet_type *pt1;
432
433 spin_lock(&ptype_lock);
434
435 list_for_each_entry(pt1, head, list) {
436 if (pt == pt1) {
437 list_del_rcu(&pt->list);
438 goto out;
439 }
440 }
441
442 pr_warn("dev_remove_pack: %p not found\n", pt);
443out:
444 spin_unlock(&ptype_lock);
445}
446EXPORT_SYMBOL(__dev_remove_pack);
447
448/**
449 * dev_remove_pack - remove packet handler
450 * @pt: packet type declaration
451 *
452 * Remove a protocol handler that was previously added to the kernel
453 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
454 * from the kernel lists and can be freed or reused once this function
455 * returns.
456 *
457 * This call sleeps to guarantee that no CPU is looking at the packet
458 * type after return.
459 */
460void dev_remove_pack(struct packet_type *pt)
461{
462 __dev_remove_pack(pt);
463
464 synchronize_net();
465}
466EXPORT_SYMBOL(dev_remove_pack);
467
468
469/**
470 * dev_add_offload - register offload handlers
471 * @po: protocol offload declaration
472 *
473 * Add protocol offload handlers to the networking stack. The passed
474 * &proto_offload is linked into kernel lists and may not be freed until
475 * it has been removed from the kernel lists.
476 *
477 * This call does not sleep therefore it can not
478 * guarantee all CPU's that are in middle of receiving packets
479 * will see the new offload handlers (until the next received packet).
480 */
481void dev_add_offload(struct packet_offload *po)
482{
483 struct packet_offload *elem;
484
485 spin_lock(&offload_lock);
486 list_for_each_entry(elem, &offload_base, list) {
487 if (po->priority < elem->priority)
488 break;
489 }
490 list_add_rcu(&po->list, elem->list.prev);
491 spin_unlock(&offload_lock);
492}
493EXPORT_SYMBOL(dev_add_offload);
494
495/**
496 * __dev_remove_offload - remove offload handler
497 * @po: packet offload declaration
498 *
499 * Remove a protocol offload handler that was previously added to the
500 * kernel offload handlers by dev_add_offload(). The passed &offload_type
501 * is removed from the kernel lists and can be freed or reused once this
502 * function returns.
503 *
504 * The packet type might still be in use by receivers
505 * and must not be freed until after all the CPU's have gone
506 * through a quiescent state.
507 */
508static void __dev_remove_offload(struct packet_offload *po)
509{
510 struct list_head *head = &offload_base;
511 struct packet_offload *po1;
512
513 spin_lock(&offload_lock);
514
515 list_for_each_entry(po1, head, list) {
516 if (po == po1) {
517 list_del_rcu(&po->list);
518 goto out;
519 }
520 }
521
522 pr_warn("dev_remove_offload: %p not found\n", po);
523out:
524 spin_unlock(&offload_lock);
525}
526
527/**
528 * dev_remove_offload - remove packet offload handler
529 * @po: packet offload declaration
530 *
531 * Remove a packet offload handler that was previously added to the kernel
532 * offload handlers by dev_add_offload(). The passed &offload_type is
533 * removed from the kernel lists and can be freed or reused once this
534 * function returns.
535 *
536 * This call sleeps to guarantee that no CPU is looking at the packet
537 * type after return.
538 */
539void dev_remove_offload(struct packet_offload *po)
540{
541 __dev_remove_offload(po);
542
543 synchronize_net();
544}
545EXPORT_SYMBOL(dev_remove_offload);
546
547/******************************************************************************
548 *
549 * Device Boot-time Settings Routines
550 *
551 ******************************************************************************/
552
553/* Boot time configuration table */
554static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
555
556/**
557 * netdev_boot_setup_add - add new setup entry
558 * @name: name of the device
559 * @map: configured settings for the device
560 *
561 * Adds new setup entry to the dev_boot_setup list. The function
562 * returns 0 on error and 1 on success. This is a generic routine to
563 * all netdevices.
564 */
565static int netdev_boot_setup_add(char *name, struct ifmap *map)
566{
567 struct netdev_boot_setup *s;
568 int i;
569
570 s = dev_boot_setup;
571 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
572 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
573 memset(s[i].name, 0, sizeof(s[i].name));
574 strlcpy(s[i].name, name, IFNAMSIZ);
575 memcpy(&s[i].map, map, sizeof(s[i].map));
576 break;
577 }
578 }
579
580 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
581}
582
583/**
584 * netdev_boot_setup_check - check boot time settings
585 * @dev: the netdevice
586 *
587 * Check boot time settings for the device.
588 * The found settings are set for the device to be used
589 * later in the device probing.
590 * Returns 0 if no settings found, 1 if they are.
591 */
592int netdev_boot_setup_check(struct net_device *dev)
593{
594 struct netdev_boot_setup *s = dev_boot_setup;
595 int i;
596
597 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
598 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
599 !strcmp(dev->name, s[i].name)) {
600 dev->irq = s[i].map.irq;
601 dev->base_addr = s[i].map.base_addr;
602 dev->mem_start = s[i].map.mem_start;
603 dev->mem_end = s[i].map.mem_end;
604 return 1;
605 }
606 }
607 return 0;
608}
609EXPORT_SYMBOL(netdev_boot_setup_check);
610
611
612/**
613 * netdev_boot_base - get address from boot time settings
614 * @prefix: prefix for network device
615 * @unit: id for network device
616 *
617 * Check boot time settings for the base address of device.
618 * The found settings are set for the device to be used
619 * later in the device probing.
620 * Returns 0 if no settings found.
621 */
622unsigned long netdev_boot_base(const char *prefix, int unit)
623{
624 const struct netdev_boot_setup *s = dev_boot_setup;
625 char name[IFNAMSIZ];
626 int i;
627
628 sprintf(name, "%s%d", prefix, unit);
629
630 /*
631 * If device already registered then return base of 1
632 * to indicate not to probe for this interface
633 */
634 if (__dev_get_by_name(&init_net, name))
635 return 1;
636
637 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
638 if (!strcmp(name, s[i].name))
639 return s[i].map.base_addr;
640 return 0;
641}
642
643/*
644 * Saves at boot time configured settings for any netdevice.
645 */
646int __init netdev_boot_setup(char *str)
647{
648 int ints[5];
649 struct ifmap map;
650
651 str = get_options(str, ARRAY_SIZE(ints), ints);
652 if (!str || !*str)
653 return 0;
654
655 /* Save settings */
656 memset(&map, 0, sizeof(map));
657 if (ints[0] > 0)
658 map.irq = ints[1];
659 if (ints[0] > 1)
660 map.base_addr = ints[2];
661 if (ints[0] > 2)
662 map.mem_start = ints[3];
663 if (ints[0] > 3)
664 map.mem_end = ints[4];
665
666 /* Add new entry to the list */
667 return netdev_boot_setup_add(str, &map);
668}
669
670__setup("netdev=", netdev_boot_setup);
671
672/*******************************************************************************
673 *
674 * Device Interface Subroutines
675 *
676 *******************************************************************************/
677
678/**
679 * dev_get_iflink - get 'iflink' value of a interface
680 * @dev: targeted interface
681 *
682 * Indicates the ifindex the interface is linked to.
683 * Physical interfaces have the same 'ifindex' and 'iflink' values.
684 */
685
686int dev_get_iflink(const struct net_device *dev)
687{
688 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
689 return dev->netdev_ops->ndo_get_iflink(dev);
690
691 return dev->ifindex;
692}
693EXPORT_SYMBOL(dev_get_iflink);
694
695/**
696 * dev_fill_metadata_dst - Retrieve tunnel egress information.
697 * @dev: targeted interface
698 * @skb: The packet.
699 *
700 * For better visibility of tunnel traffic OVS needs to retrieve
701 * egress tunnel information for a packet. Following API allows
702 * user to get this info.
703 */
704int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
705{
706 struct ip_tunnel_info *info;
707
708 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
709 return -EINVAL;
710
711 info = skb_tunnel_info_unclone(skb);
712 if (!info)
713 return -ENOMEM;
714 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
715 return -EINVAL;
716
717 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
718}
719EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
720
721/**
722 * __dev_get_by_name - find a device by its name
723 * @net: the applicable net namespace
724 * @name: name to find
725 *
726 * Find an interface by name. Must be called under RTNL semaphore
727 * or @dev_base_lock. If the name is found a pointer to the device
728 * is returned. If the name is not found then %NULL is returned. The
729 * reference counters are not incremented so the caller must be
730 * careful with locks.
731 */
732
733struct net_device *__dev_get_by_name(struct net *net, const char *name)
734{
735 struct net_device *dev;
736 struct hlist_head *head = dev_name_hash(net, name);
737
738 hlist_for_each_entry(dev, head, name_hlist)
739 if (!strncmp(dev->name, name, IFNAMSIZ))
740 return dev;
741
742 return NULL;
743}
744EXPORT_SYMBOL(__dev_get_by_name);
745
746/**
747 * dev_get_by_name_rcu - find a device by its name
748 * @net: the applicable net namespace
749 * @name: name to find
750 *
751 * Find an interface by name.
752 * If the name is found a pointer to the device is returned.
753 * If the name is not found then %NULL is returned.
754 * The reference counters are not incremented so the caller must be
755 * careful with locks. The caller must hold RCU lock.
756 */
757
758struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
759{
760 struct net_device *dev;
761 struct hlist_head *head = dev_name_hash(net, name);
762
763 hlist_for_each_entry_rcu(dev, head, name_hlist)
764 if (!strncmp(dev->name, name, IFNAMSIZ))
765 return dev;
766
767 return NULL;
768}
769EXPORT_SYMBOL(dev_get_by_name_rcu);
770
771/**
772 * dev_get_by_name - find a device by its name
773 * @net: the applicable net namespace
774 * @name: name to find
775 *
776 * Find an interface by name. This can be called from any
777 * context and does its own locking. The returned handle has
778 * the usage count incremented and the caller must use dev_put() to
779 * release it when it is no longer needed. %NULL is returned if no
780 * matching device is found.
781 */
782
783struct net_device *dev_get_by_name(struct net *net, const char *name)
784{
785 struct net_device *dev;
786
787 rcu_read_lock();
788 dev = dev_get_by_name_rcu(net, name);
789 if (dev)
790 dev_hold(dev);
791 rcu_read_unlock();
792 return dev;
793}
794EXPORT_SYMBOL(dev_get_by_name);
795
796/**
797 * __dev_get_by_index - find a device by its ifindex
798 * @net: the applicable net namespace
799 * @ifindex: index of device
800 *
801 * Search for an interface by index. Returns %NULL if the device
802 * is not found or a pointer to the device. The device has not
803 * had its reference counter increased so the caller must be careful
804 * about locking. The caller must hold either the RTNL semaphore
805 * or @dev_base_lock.
806 */
807
808struct net_device *__dev_get_by_index(struct net *net, int ifindex)
809{
810 struct net_device *dev;
811 struct hlist_head *head = dev_index_hash(net, ifindex);
812
813 hlist_for_each_entry(dev, head, index_hlist)
814 if (dev->ifindex == ifindex)
815 return dev;
816
817 return NULL;
818}
819EXPORT_SYMBOL(__dev_get_by_index);
820
821/**
822 * dev_get_by_index_rcu - find a device by its ifindex
823 * @net: the applicable net namespace
824 * @ifindex: index of device
825 *
826 * Search for an interface by index. Returns %NULL if the device
827 * is not found or a pointer to the device. The device has not
828 * had its reference counter increased so the caller must be careful
829 * about locking. The caller must hold RCU lock.
830 */
831
832struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
833{
834 struct net_device *dev;
835 struct hlist_head *head = dev_index_hash(net, ifindex);
836
837 hlist_for_each_entry_rcu(dev, head, index_hlist)
838 if (dev->ifindex == ifindex)
839 return dev;
840
841 return NULL;
842}
843EXPORT_SYMBOL(dev_get_by_index_rcu);
844
845
846/**
847 * dev_get_by_index - find a device by its ifindex
848 * @net: the applicable net namespace
849 * @ifindex: index of device
850 *
851 * Search for an interface by index. Returns NULL if the device
852 * is not found or a pointer to the device. The device returned has
853 * had a reference added and the pointer is safe until the user calls
854 * dev_put to indicate they have finished with it.
855 */
856
857struct net_device *dev_get_by_index(struct net *net, int ifindex)
858{
859 struct net_device *dev;
860
861 rcu_read_lock();
862 dev = dev_get_by_index_rcu(net, ifindex);
863 if (dev)
864 dev_hold(dev);
865 rcu_read_unlock();
866 return dev;
867}
868EXPORT_SYMBOL(dev_get_by_index);
869
870/**
871 * dev_get_by_napi_id - find a device by napi_id
872 * @napi_id: ID of the NAPI struct
873 *
874 * Search for an interface by NAPI ID. Returns %NULL if the device
875 * is not found or a pointer to the device. The device has not had
876 * its reference counter increased so the caller must be careful
877 * about locking. The caller must hold RCU lock.
878 */
879
880struct net_device *dev_get_by_napi_id(unsigned int napi_id)
881{
882 struct napi_struct *napi;
883
884 WARN_ON_ONCE(!rcu_read_lock_held());
885
886 if (napi_id < MIN_NAPI_ID)
887 return NULL;
888
889 napi = napi_by_id(napi_id);
890
891 return napi ? napi->dev : NULL;
892}
893EXPORT_SYMBOL(dev_get_by_napi_id);
894
895/**
896 * netdev_get_name - get a netdevice name, knowing its ifindex.
897 * @net: network namespace
898 * @name: a pointer to the buffer where the name will be stored.
899 * @ifindex: the ifindex of the interface to get the name from.
900 *
901 * The use of raw_seqcount_begin() and cond_resched() before
902 * retrying is required as we want to give the writers a chance
903 * to complete when CONFIG_PREEMPT is not set.
904 */
905int netdev_get_name(struct net *net, char *name, int ifindex)
906{
907 struct net_device *dev;
908 unsigned int seq;
909
910retry:
911 seq = raw_seqcount_begin(&devnet_rename_seq);
912 rcu_read_lock();
913 dev = dev_get_by_index_rcu(net, ifindex);
914 if (!dev) {
915 rcu_read_unlock();
916 return -ENODEV;
917 }
918
919 strcpy(name, dev->name);
920 rcu_read_unlock();
921 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
922 cond_resched();
923 goto retry;
924 }
925
926 return 0;
927}
928
929/**
930 * dev_getbyhwaddr_rcu - find a device by its hardware address
931 * @net: the applicable net namespace
932 * @type: media type of device
933 * @ha: hardware address
934 *
935 * Search for an interface by MAC address. Returns NULL if the device
936 * is not found or a pointer to the device.
937 * The caller must hold RCU or RTNL.
938 * The returned device has not had its ref count increased
939 * and the caller must therefore be careful about locking
940 *
941 */
942
943struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
944 const char *ha)
945{
946 struct net_device *dev;
947
948 for_each_netdev_rcu(net, dev)
949 if (dev->type == type &&
950 !memcmp(dev->dev_addr, ha, dev->addr_len))
951 return dev;
952
953 return NULL;
954}
955EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
956
957struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
958{
959 struct net_device *dev;
960
961 ASSERT_RTNL();
962 for_each_netdev(net, dev)
963 if (dev->type == type)
964 return dev;
965
966 return NULL;
967}
968EXPORT_SYMBOL(__dev_getfirstbyhwtype);
969
970struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
971{
972 struct net_device *dev, *ret = NULL;
973
974 rcu_read_lock();
975 for_each_netdev_rcu(net, dev)
976 if (dev->type == type) {
977 dev_hold(dev);
978 ret = dev;
979 break;
980 }
981 rcu_read_unlock();
982 return ret;
983}
984EXPORT_SYMBOL(dev_getfirstbyhwtype);
985
986/**
987 * __dev_get_by_flags - find any device with given flags
988 * @net: the applicable net namespace
989 * @if_flags: IFF_* values
990 * @mask: bitmask of bits in if_flags to check
991 *
992 * Search for any interface with the given flags. Returns NULL if a device
993 * is not found or a pointer to the device. Must be called inside
994 * rtnl_lock(), and result refcount is unchanged.
995 */
996
997struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
998 unsigned short mask)
999{
1000 struct net_device *dev, *ret;
1001
1002 ASSERT_RTNL();
1003
1004 ret = NULL;
1005 for_each_netdev(net, dev) {
1006 if (((dev->flags ^ if_flags) & mask) == 0) {
1007 ret = dev;
1008 break;
1009 }
1010 }
1011 return ret;
1012}
1013EXPORT_SYMBOL(__dev_get_by_flags);
1014
1015/**
1016 * dev_valid_name - check if name is okay for network device
1017 * @name: name string
1018 *
1019 * Network device names need to be valid file names to
1020 * to allow sysfs to work. We also disallow any kind of
1021 * whitespace.
1022 */
1023bool dev_valid_name(const char *name)
1024{
1025 if (*name == '\0')
1026 return false;
1027 if (strlen(name) >= IFNAMSIZ)
1028 return false;
1029 if (!strcmp(name, ".") || !strcmp(name, ".."))
1030 return false;
1031
1032 while (*name) {
1033 if (*name == '/' || *name == ':' || isspace(*name))
1034 return false;
1035 name++;
1036 }
1037 return true;
1038}
1039EXPORT_SYMBOL(dev_valid_name);
1040
1041/**
1042 * __dev_alloc_name - allocate a name for a device
1043 * @net: network namespace to allocate the device name in
1044 * @name: name format string
1045 * @buf: scratch buffer and result name string
1046 *
1047 * Passed a format string - eg "lt%d" it will try and find a suitable
1048 * id. It scans list of devices to build up a free map, then chooses
1049 * the first empty slot. The caller must hold the dev_base or rtnl lock
1050 * while allocating the name and adding the device in order to avoid
1051 * duplicates.
1052 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1053 * Returns the number of the unit assigned or a negative errno code.
1054 */
1055
1056static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1057{
1058 int i = 0;
1059 const char *p;
1060 const int max_netdevices = 8*PAGE_SIZE;
1061 unsigned long *inuse;
1062 struct net_device *d;
1063
1064 p = strnchr(name, IFNAMSIZ-1, '%');
1065 if (p) {
1066 /*
1067 * Verify the string as this thing may have come from
1068 * the user. There must be either one "%d" and no other "%"
1069 * characters.
1070 */
1071 if (p[1] != 'd' || strchr(p + 2, '%'))
1072 return -EINVAL;
1073
1074 /* Use one page as a bit array of possible slots */
1075 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1076 if (!inuse)
1077 return -ENOMEM;
1078
1079 for_each_netdev(net, d) {
1080 if (!sscanf(d->name, name, &i))
1081 continue;
1082 if (i < 0 || i >= max_netdevices)
1083 continue;
1084
1085 /* avoid cases where sscanf is not exact inverse of printf */
1086 snprintf(buf, IFNAMSIZ, name, i);
1087 if (!strncmp(buf, d->name, IFNAMSIZ))
1088 set_bit(i, inuse);
1089 }
1090
1091 i = find_first_zero_bit(inuse, max_netdevices);
1092 free_page((unsigned long) inuse);
1093 }
1094
1095 if (buf != name)
1096 snprintf(buf, IFNAMSIZ, name, i);
1097 if (!__dev_get_by_name(net, buf))
1098 return i;
1099
1100 /* It is possible to run out of possible slots
1101 * when the name is long and there isn't enough space left
1102 * for the digits, or if all bits are used.
1103 */
1104 return -ENFILE;
1105}
1106
1107/**
1108 * dev_alloc_name - allocate a name for a device
1109 * @dev: device
1110 * @name: name format string
1111 *
1112 * Passed a format string - eg "lt%d" it will try and find a suitable
1113 * id. It scans list of devices to build up a free map, then chooses
1114 * the first empty slot. The caller must hold the dev_base or rtnl lock
1115 * while allocating the name and adding the device in order to avoid
1116 * duplicates.
1117 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1118 * Returns the number of the unit assigned or a negative errno code.
1119 */
1120
1121int dev_alloc_name(struct net_device *dev, const char *name)
1122{
1123 char buf[IFNAMSIZ];
1124 struct net *net;
1125 int ret;
1126
1127 BUG_ON(!dev_net(dev));
1128 net = dev_net(dev);
1129 ret = __dev_alloc_name(net, name, buf);
1130 if (ret >= 0)
1131 strlcpy(dev->name, buf, IFNAMSIZ);
1132 return ret;
1133}
1134EXPORT_SYMBOL(dev_alloc_name);
1135
1136static int dev_alloc_name_ns(struct net *net,
1137 struct net_device *dev,
1138 const char *name)
1139{
1140 char buf[IFNAMSIZ];
1141 int ret;
1142
1143 ret = __dev_alloc_name(net, name, buf);
1144 if (ret >= 0)
1145 strlcpy(dev->name, buf, IFNAMSIZ);
1146 return ret;
1147}
1148
1149static int dev_get_valid_name(struct net *net,
1150 struct net_device *dev,
1151 const char *name)
1152{
1153 BUG_ON(!net);
1154
1155 if (!dev_valid_name(name))
1156 return -EINVAL;
1157
1158 if (strchr(name, '%'))
1159 return dev_alloc_name_ns(net, dev, name);
1160 else if (__dev_get_by_name(net, name))
1161 return -EEXIST;
1162 else if (dev->name != name)
1163 strlcpy(dev->name, name, IFNAMSIZ);
1164
1165 return 0;
1166}
1167
1168/**
1169 * dev_change_name - change name of a device
1170 * @dev: device
1171 * @newname: name (or format string) must be at least IFNAMSIZ
1172 *
1173 * Change name of a device, can pass format strings "eth%d".
1174 * for wildcarding.
1175 */
1176int dev_change_name(struct net_device *dev, const char *newname)
1177{
1178 unsigned char old_assign_type;
1179 char oldname[IFNAMSIZ];
1180 int err = 0;
1181 int ret;
1182 struct net *net;
1183
1184 ASSERT_RTNL();
1185 BUG_ON(!dev_net(dev));
1186
1187 net = dev_net(dev);
1188 if (dev->flags & IFF_UP)
1189 return -EBUSY;
1190
1191 write_seqcount_begin(&devnet_rename_seq);
1192
1193 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1194 write_seqcount_end(&devnet_rename_seq);
1195 return 0;
1196 }
1197
1198 memcpy(oldname, dev->name, IFNAMSIZ);
1199
1200 err = dev_get_valid_name(net, dev, newname);
1201 if (err < 0) {
1202 write_seqcount_end(&devnet_rename_seq);
1203 return err;
1204 }
1205
1206 if (oldname[0] && !strchr(oldname, '%'))
1207 netdev_info(dev, "renamed from %s\n", oldname);
1208
1209 old_assign_type = dev->name_assign_type;
1210 dev->name_assign_type = NET_NAME_RENAMED;
1211
1212rollback:
1213 ret = device_rename(&dev->dev, dev->name);
1214 if (ret) {
1215 memcpy(dev->name, oldname, IFNAMSIZ);
1216 dev->name_assign_type = old_assign_type;
1217 write_seqcount_end(&devnet_rename_seq);
1218 return ret;
1219 }
1220
1221 write_seqcount_end(&devnet_rename_seq);
1222
1223 netdev_adjacent_rename_links(dev, oldname);
1224
1225 write_lock_bh(&dev_base_lock);
1226 hlist_del_rcu(&dev->name_hlist);
1227 write_unlock_bh(&dev_base_lock);
1228
1229 synchronize_rcu();
1230
1231 write_lock_bh(&dev_base_lock);
1232 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1233 write_unlock_bh(&dev_base_lock);
1234
1235 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1236 ret = notifier_to_errno(ret);
1237
1238 if (ret) {
1239 /* err >= 0 after dev_alloc_name() or stores the first errno */
1240 if (err >= 0) {
1241 err = ret;
1242 write_seqcount_begin(&devnet_rename_seq);
1243 memcpy(dev->name, oldname, IFNAMSIZ);
1244 memcpy(oldname, newname, IFNAMSIZ);
1245 dev->name_assign_type = old_assign_type;
1246 old_assign_type = NET_NAME_RENAMED;
1247 goto rollback;
1248 } else {
1249 pr_err("%s: name change rollback failed: %d\n",
1250 dev->name, ret);
1251 }
1252 }
1253
1254 return err;
1255}
1256
1257/**
1258 * dev_set_alias - change ifalias of a device
1259 * @dev: device
1260 * @alias: name up to IFALIASZ
1261 * @len: limit of bytes to copy from info
1262 *
1263 * Set ifalias for a device,
1264 */
1265int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1266{
1267 char *new_ifalias;
1268
1269 ASSERT_RTNL();
1270
1271 if (len >= IFALIASZ)
1272 return -EINVAL;
1273
1274 if (!len) {
1275 kfree(dev->ifalias);
1276 dev->ifalias = NULL;
1277 return 0;
1278 }
1279
1280 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1281 if (!new_ifalias)
1282 return -ENOMEM;
1283 dev->ifalias = new_ifalias;
1284 memcpy(dev->ifalias, alias, len);
1285 dev->ifalias[len] = 0;
1286
1287 return len;
1288}
1289
1290
1291/**
1292 * netdev_features_change - device changes features
1293 * @dev: device to cause notification
1294 *
1295 * Called to indicate a device has changed features.
1296 */
1297void netdev_features_change(struct net_device *dev)
1298{
1299 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1300}
1301EXPORT_SYMBOL(netdev_features_change);
1302
1303/**
1304 * netdev_state_change - device changes state
1305 * @dev: device to cause notification
1306 *
1307 * Called to indicate a device has changed state. This function calls
1308 * the notifier chains for netdev_chain and sends a NEWLINK message
1309 * to the routing socket.
1310 */
1311void netdev_state_change(struct net_device *dev)
1312{
1313 if (dev->flags & IFF_UP) {
1314 struct netdev_notifier_change_info change_info;
1315
1316 change_info.flags_changed = 0;
1317 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
1318 &change_info.info);
1319 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1320 }
1321}
1322EXPORT_SYMBOL(netdev_state_change);
1323
1324/**
1325 * netdev_notify_peers - notify network peers about existence of @dev
1326 * @dev: network device
1327 *
1328 * Generate traffic such that interested network peers are aware of
1329 * @dev, such as by generating a gratuitous ARP. This may be used when
1330 * a device wants to inform the rest of the network about some sort of
1331 * reconfiguration such as a failover event or virtual machine
1332 * migration.
1333 */
1334void netdev_notify_peers(struct net_device *dev)
1335{
1336 rtnl_lock();
1337 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1338 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1339 rtnl_unlock();
1340}
1341EXPORT_SYMBOL(netdev_notify_peers);
1342
1343static int __dev_open(struct net_device *dev)
1344{
1345 const struct net_device_ops *ops = dev->netdev_ops;
1346 int ret;
1347
1348 ASSERT_RTNL();
1349
1350 if (!netif_device_present(dev))
1351 return -ENODEV;
1352
1353 /* Block netpoll from trying to do any rx path servicing.
1354 * If we don't do this there is a chance ndo_poll_controller
1355 * or ndo_poll may be running while we open the device
1356 */
1357 netpoll_poll_disable(dev);
1358
1359 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1360 ret = notifier_to_errno(ret);
1361 if (ret)
1362 return ret;
1363
1364 set_bit(__LINK_STATE_START, &dev->state);
1365
1366 if (ops->ndo_validate_addr)
1367 ret = ops->ndo_validate_addr(dev);
1368
1369 if (!ret && ops->ndo_open)
1370 ret = ops->ndo_open(dev);
1371
1372 netpoll_poll_enable(dev);
1373
1374 if (ret)
1375 clear_bit(__LINK_STATE_START, &dev->state);
1376 else {
1377 dev->flags |= IFF_UP;
1378 dev_set_rx_mode(dev);
1379 dev_activate(dev);
1380 add_device_randomness(dev->dev_addr, dev->addr_len);
1381 }
1382
1383 return ret;
1384}
1385
1386/**
1387 * dev_open - prepare an interface for use.
1388 * @dev: device to open
1389 *
1390 * Takes a device from down to up state. The device's private open
1391 * function is invoked and then the multicast lists are loaded. Finally
1392 * the device is moved into the up state and a %NETDEV_UP message is
1393 * sent to the netdev notifier chain.
1394 *
1395 * Calling this function on an active interface is a nop. On a failure
1396 * a negative errno code is returned.
1397 */
1398int dev_open(struct net_device *dev)
1399{
1400 int ret;
1401
1402 if (dev->flags & IFF_UP)
1403 return 0;
1404
1405 ret = __dev_open(dev);
1406 if (ret < 0)
1407 return ret;
1408
1409 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1410 call_netdevice_notifiers(NETDEV_UP, dev);
1411
1412 return ret;
1413}
1414EXPORT_SYMBOL(dev_open);
1415
1416static int __dev_close_many(struct list_head *head)
1417{
1418 struct net_device *dev;
1419
1420 ASSERT_RTNL();
1421 might_sleep();
1422
1423 list_for_each_entry(dev, head, close_list) {
1424 /* Temporarily disable netpoll until the interface is down */
1425 netpoll_poll_disable(dev);
1426
1427 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1428
1429 clear_bit(__LINK_STATE_START, &dev->state);
1430
1431 /* Synchronize to scheduled poll. We cannot touch poll list, it
1432 * can be even on different cpu. So just clear netif_running().
1433 *
1434 * dev->stop() will invoke napi_disable() on all of it's
1435 * napi_struct instances on this device.
1436 */
1437 smp_mb__after_atomic(); /* Commit netif_running(). */
1438 }
1439
1440 dev_deactivate_many(head);
1441
1442 list_for_each_entry(dev, head, close_list) {
1443 const struct net_device_ops *ops = dev->netdev_ops;
1444
1445 /*
1446 * Call the device specific close. This cannot fail.
1447 * Only if device is UP
1448 *
1449 * We allow it to be called even after a DETACH hot-plug
1450 * event.
1451 */
1452 if (ops->ndo_stop)
1453 ops->ndo_stop(dev);
1454
1455 dev->flags &= ~IFF_UP;
1456 netpoll_poll_enable(dev);
1457 }
1458
1459 return 0;
1460}
1461
1462static int __dev_close(struct net_device *dev)
1463{
1464 int retval;
1465 LIST_HEAD(single);
1466
1467 list_add(&dev->close_list, &single);
1468 retval = __dev_close_many(&single);
1469 list_del(&single);
1470
1471 return retval;
1472}
1473
1474int dev_close_many(struct list_head *head, bool unlink)
1475{
1476 struct net_device *dev, *tmp;
1477
1478 /* Remove the devices that don't need to be closed */
1479 list_for_each_entry_safe(dev, tmp, head, close_list)
1480 if (!(dev->flags & IFF_UP))
1481 list_del_init(&dev->close_list);
1482
1483 __dev_close_many(head);
1484
1485 list_for_each_entry_safe(dev, tmp, head, close_list) {
1486 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1487 call_netdevice_notifiers(NETDEV_DOWN, dev);
1488 if (unlink)
1489 list_del_init(&dev->close_list);
1490 }
1491
1492 return 0;
1493}
1494EXPORT_SYMBOL(dev_close_many);
1495
1496/**
1497 * dev_close - shutdown an interface.
1498 * @dev: device to shutdown
1499 *
1500 * This function moves an active device into down state. A
1501 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1502 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1503 * chain.
1504 */
1505int dev_close(struct net_device *dev)
1506{
1507 if (dev->flags & IFF_UP) {
1508 LIST_HEAD(single);
1509
1510 list_add(&dev->close_list, &single);
1511 dev_close_many(&single, true);
1512 list_del(&single);
1513 }
1514 return 0;
1515}
1516EXPORT_SYMBOL(dev_close);
1517
1518
1519/**
1520 * dev_disable_lro - disable Large Receive Offload on a device
1521 * @dev: device
1522 *
1523 * Disable Large Receive Offload (LRO) on a net device. Must be
1524 * called under RTNL. This is needed if received packets may be
1525 * forwarded to another interface.
1526 */
1527void dev_disable_lro(struct net_device *dev)
1528{
1529 struct net_device *lower_dev;
1530 struct list_head *iter;
1531
1532 dev->wanted_features &= ~NETIF_F_LRO;
1533 netdev_update_features(dev);
1534
1535 if (unlikely(dev->features & NETIF_F_LRO))
1536 netdev_WARN(dev, "failed to disable LRO!\n");
1537
1538 netdev_for_each_lower_dev(dev, lower_dev, iter)
1539 dev_disable_lro(lower_dev);
1540}
1541EXPORT_SYMBOL(dev_disable_lro);
1542
1543static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1544 struct net_device *dev)
1545{
1546 struct netdev_notifier_info info;
1547
1548 netdev_notifier_info_init(&info, dev);
1549 return nb->notifier_call(nb, val, &info);
1550}
1551
1552static int dev_boot_phase = 1;
1553
1554/**
1555 * register_netdevice_notifier - register a network notifier block
1556 * @nb: notifier
1557 *
1558 * Register a notifier to be called when network device events occur.
1559 * The notifier passed is linked into the kernel structures and must
1560 * not be reused until it has been unregistered. A negative errno code
1561 * is returned on a failure.
1562 *
1563 * When registered all registration and up events are replayed
1564 * to the new notifier to allow device to have a race free
1565 * view of the network device list.
1566 */
1567
1568int register_netdevice_notifier(struct notifier_block *nb)
1569{
1570 struct net_device *dev;
1571 struct net_device *last;
1572 struct net *net;
1573 int err;
1574
1575 rtnl_lock();
1576 err = raw_notifier_chain_register(&netdev_chain, nb);
1577 if (err)
1578 goto unlock;
1579 if (dev_boot_phase)
1580 goto unlock;
1581 for_each_net(net) {
1582 for_each_netdev(net, dev) {
1583 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1584 err = notifier_to_errno(err);
1585 if (err)
1586 goto rollback;
1587
1588 if (!(dev->flags & IFF_UP))
1589 continue;
1590
1591 call_netdevice_notifier(nb, NETDEV_UP, dev);
1592 }
1593 }
1594
1595unlock:
1596 rtnl_unlock();
1597 return err;
1598
1599rollback:
1600 last = dev;
1601 for_each_net(net) {
1602 for_each_netdev(net, dev) {
1603 if (dev == last)
1604 goto outroll;
1605
1606 if (dev->flags & IFF_UP) {
1607 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1608 dev);
1609 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1610 }
1611 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1612 }
1613 }
1614
1615outroll:
1616 raw_notifier_chain_unregister(&netdev_chain, nb);
1617 goto unlock;
1618}
1619EXPORT_SYMBOL(register_netdevice_notifier);
1620
1621/**
1622 * unregister_netdevice_notifier - unregister a network notifier block
1623 * @nb: notifier
1624 *
1625 * Unregister a notifier previously registered by
1626 * register_netdevice_notifier(). The notifier is unlinked into the
1627 * kernel structures and may then be reused. A negative errno code
1628 * is returned on a failure.
1629 *
1630 * After unregistering unregister and down device events are synthesized
1631 * for all devices on the device list to the removed notifier to remove
1632 * the need for special case cleanup code.
1633 */
1634
1635int unregister_netdevice_notifier(struct notifier_block *nb)
1636{
1637 struct net_device *dev;
1638 struct net *net;
1639 int err;
1640
1641 rtnl_lock();
1642 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1643 if (err)
1644 goto unlock;
1645
1646 for_each_net(net) {
1647 for_each_netdev(net, dev) {
1648 if (dev->flags & IFF_UP) {
1649 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1650 dev);
1651 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1652 }
1653 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1654 }
1655 }
1656unlock:
1657 rtnl_unlock();
1658 return err;
1659}
1660EXPORT_SYMBOL(unregister_netdevice_notifier);
1661
1662/**
1663 * call_netdevice_notifiers_info - call all network notifier blocks
1664 * @val: value passed unmodified to notifier function
1665 * @dev: net_device pointer passed unmodified to notifier function
1666 * @info: notifier information data
1667 *
1668 * Call all network notifier blocks. Parameters and return value
1669 * are as for raw_notifier_call_chain().
1670 */
1671
1672static int call_netdevice_notifiers_info(unsigned long val,
1673 struct net_device *dev,
1674 struct netdev_notifier_info *info)
1675{
1676 ASSERT_RTNL();
1677 netdev_notifier_info_init(info, dev);
1678 return raw_notifier_call_chain(&netdev_chain, val, info);
1679}
1680
1681/**
1682 * call_netdevice_notifiers - call all network notifier blocks
1683 * @val: value passed unmodified to notifier function
1684 * @dev: net_device pointer passed unmodified to notifier function
1685 *
1686 * Call all network notifier blocks. Parameters and return value
1687 * are as for raw_notifier_call_chain().
1688 */
1689
1690int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1691{
1692 struct netdev_notifier_info info;
1693
1694 return call_netdevice_notifiers_info(val, dev, &info);
1695}
1696EXPORT_SYMBOL(call_netdevice_notifiers);
1697
1698#ifdef CONFIG_NET_INGRESS
1699static struct static_key ingress_needed __read_mostly;
1700
1701void net_inc_ingress_queue(void)
1702{
1703 static_key_slow_inc(&ingress_needed);
1704}
1705EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1706
1707void net_dec_ingress_queue(void)
1708{
1709 static_key_slow_dec(&ingress_needed);
1710}
1711EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1712#endif
1713
1714#ifdef CONFIG_NET_EGRESS
1715static struct static_key egress_needed __read_mostly;
1716
1717void net_inc_egress_queue(void)
1718{
1719 static_key_slow_inc(&egress_needed);
1720}
1721EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1722
1723void net_dec_egress_queue(void)
1724{
1725 static_key_slow_dec(&egress_needed);
1726}
1727EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1728#endif
1729
1730static struct static_key netstamp_needed __read_mostly;
1731#ifdef HAVE_JUMP_LABEL
1732static atomic_t netstamp_needed_deferred;
1733static atomic_t netstamp_wanted;
1734static void netstamp_clear(struct work_struct *work)
1735{
1736 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1737 int wanted;
1738
1739 wanted = atomic_add_return(deferred, &netstamp_wanted);
1740 if (wanted > 0)
1741 static_key_enable(&netstamp_needed);
1742 else
1743 static_key_disable(&netstamp_needed);
1744}
1745static DECLARE_WORK(netstamp_work, netstamp_clear);
1746#endif
1747
1748void net_enable_timestamp(void)
1749{
1750#ifdef HAVE_JUMP_LABEL
1751 int wanted;
1752
1753 while (1) {
1754 wanted = atomic_read(&netstamp_wanted);
1755 if (wanted <= 0)
1756 break;
1757 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1758 return;
1759 }
1760 atomic_inc(&netstamp_needed_deferred);
1761 schedule_work(&netstamp_work);
1762#else
1763 static_key_slow_inc(&netstamp_needed);
1764#endif
1765}
1766EXPORT_SYMBOL(net_enable_timestamp);
1767
1768void net_disable_timestamp(void)
1769{
1770#ifdef HAVE_JUMP_LABEL
1771 int wanted;
1772
1773 while (1) {
1774 wanted = atomic_read(&netstamp_wanted);
1775 if (wanted <= 1)
1776 break;
1777 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1778 return;
1779 }
1780 atomic_dec(&netstamp_needed_deferred);
1781 schedule_work(&netstamp_work);
1782#else
1783 static_key_slow_dec(&netstamp_needed);
1784#endif
1785}
1786EXPORT_SYMBOL(net_disable_timestamp);
1787
1788static inline void net_timestamp_set(struct sk_buff *skb)
1789{
1790 skb->tstamp = 0;
1791 if (static_key_false(&netstamp_needed))
1792 __net_timestamp(skb);
1793}
1794
1795#define net_timestamp_check(COND, SKB) \
1796 if (static_key_false(&netstamp_needed)) { \
1797 if ((COND) && !(SKB)->tstamp) \
1798 __net_timestamp(SKB); \
1799 } \
1800
1801bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1802{
1803 unsigned int len;
1804
1805 if (!(dev->flags & IFF_UP))
1806 return false;
1807
1808 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1809 if (skb->len <= len)
1810 return true;
1811
1812 /* if TSO is enabled, we don't care about the length as the packet
1813 * could be forwarded without being segmented before
1814 */
1815 if (skb_is_gso(skb))
1816 return true;
1817
1818 return false;
1819}
1820EXPORT_SYMBOL_GPL(is_skb_forwardable);
1821
1822int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1823{
1824 int ret = ____dev_forward_skb(dev, skb);
1825
1826 if (likely(!ret)) {
1827 skb->protocol = eth_type_trans(skb, dev);
1828 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1829 }
1830
1831 return ret;
1832}
1833EXPORT_SYMBOL_GPL(__dev_forward_skb);
1834
1835/**
1836 * dev_forward_skb - loopback an skb to another netif
1837 *
1838 * @dev: destination network device
1839 * @skb: buffer to forward
1840 *
1841 * return values:
1842 * NET_RX_SUCCESS (no congestion)
1843 * NET_RX_DROP (packet was dropped, but freed)
1844 *
1845 * dev_forward_skb can be used for injecting an skb from the
1846 * start_xmit function of one device into the receive queue
1847 * of another device.
1848 *
1849 * The receiving device may be in another namespace, so
1850 * we have to clear all information in the skb that could
1851 * impact namespace isolation.
1852 */
1853int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1854{
1855 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1856}
1857EXPORT_SYMBOL_GPL(dev_forward_skb);
1858
1859static inline int deliver_skb(struct sk_buff *skb,
1860 struct packet_type *pt_prev,
1861 struct net_device *orig_dev)
1862{
1863 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
1864 return -ENOMEM;
1865 refcount_inc(&skb->users);
1866 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1867}
1868
1869static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1870 struct packet_type **pt,
1871 struct net_device *orig_dev,
1872 __be16 type,
1873 struct list_head *ptype_list)
1874{
1875 struct packet_type *ptype, *pt_prev = *pt;
1876
1877 list_for_each_entry_rcu(ptype, ptype_list, list) {
1878 if (ptype->type != type)
1879 continue;
1880 if (pt_prev)
1881 deliver_skb(skb, pt_prev, orig_dev);
1882 pt_prev = ptype;
1883 }
1884 *pt = pt_prev;
1885}
1886
1887static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1888{
1889 if (!ptype->af_packet_priv || !skb->sk)
1890 return false;
1891
1892 if (ptype->id_match)
1893 return ptype->id_match(ptype, skb->sk);
1894 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1895 return true;
1896
1897 return false;
1898}
1899
1900/*
1901 * Support routine. Sends outgoing frames to any network
1902 * taps currently in use.
1903 */
1904
1905void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1906{
1907 struct packet_type *ptype;
1908 struct sk_buff *skb2 = NULL;
1909 struct packet_type *pt_prev = NULL;
1910 struct list_head *ptype_list = &ptype_all;
1911
1912 rcu_read_lock();
1913again:
1914 list_for_each_entry_rcu(ptype, ptype_list, list) {
1915 /* Never send packets back to the socket
1916 * they originated from - MvS (miquels@drinkel.ow.org)
1917 */
1918 if (skb_loop_sk(ptype, skb))
1919 continue;
1920
1921 if (pt_prev) {
1922 deliver_skb(skb2, pt_prev, skb->dev);
1923 pt_prev = ptype;
1924 continue;
1925 }
1926
1927 /* need to clone skb, done only once */
1928 skb2 = skb_clone(skb, GFP_ATOMIC);
1929 if (!skb2)
1930 goto out_unlock;
1931
1932 net_timestamp_set(skb2);
1933
1934 /* skb->nh should be correctly
1935 * set by sender, so that the second statement is
1936 * just protection against buggy protocols.
1937 */
1938 skb_reset_mac_header(skb2);
1939
1940 if (skb_network_header(skb2) < skb2->data ||
1941 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1942 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1943 ntohs(skb2->protocol),
1944 dev->name);
1945 skb_reset_network_header(skb2);
1946 }
1947
1948 skb2->transport_header = skb2->network_header;
1949 skb2->pkt_type = PACKET_OUTGOING;
1950 pt_prev = ptype;
1951 }
1952
1953 if (ptype_list == &ptype_all) {
1954 ptype_list = &dev->ptype_all;
1955 goto again;
1956 }
1957out_unlock:
1958 if (pt_prev)
1959 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1960 rcu_read_unlock();
1961}
1962EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
1963
1964/**
1965 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1966 * @dev: Network device
1967 * @txq: number of queues available
1968 *
1969 * If real_num_tx_queues is changed the tc mappings may no longer be
1970 * valid. To resolve this verify the tc mapping remains valid and if
1971 * not NULL the mapping. With no priorities mapping to this
1972 * offset/count pair it will no longer be used. In the worst case TC0
1973 * is invalid nothing can be done so disable priority mappings. If is
1974 * expected that drivers will fix this mapping if they can before
1975 * calling netif_set_real_num_tx_queues.
1976 */
1977static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1978{
1979 int i;
1980 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1981
1982 /* If TC0 is invalidated disable TC mapping */
1983 if (tc->offset + tc->count > txq) {
1984 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1985 dev->num_tc = 0;
1986 return;
1987 }
1988
1989 /* Invalidated prio to tc mappings set to TC0 */
1990 for (i = 1; i < TC_BITMASK + 1; i++) {
1991 int q = netdev_get_prio_tc_map(dev, i);
1992
1993 tc = &dev->tc_to_txq[q];
1994 if (tc->offset + tc->count > txq) {
1995 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1996 i, q);
1997 netdev_set_prio_tc_map(dev, i, 0);
1998 }
1999 }
2000}
2001
2002int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2003{
2004 if (dev->num_tc) {
2005 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2006 int i;
2007
2008 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2009 if ((txq - tc->offset) < tc->count)
2010 return i;
2011 }
2012
2013 return -1;
2014 }
2015
2016 return 0;
2017}
2018
2019#ifdef CONFIG_XPS
2020static DEFINE_MUTEX(xps_map_mutex);
2021#define xmap_dereference(P) \
2022 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2023
2024static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2025 int tci, u16 index)
2026{
2027 struct xps_map *map = NULL;
2028 int pos;
2029
2030 if (dev_maps)
2031 map = xmap_dereference(dev_maps->cpu_map[tci]);
2032 if (!map)
2033 return false;
2034
2035 for (pos = map->len; pos--;) {
2036 if (map->queues[pos] != index)
2037 continue;
2038
2039 if (map->len > 1) {
2040 map->queues[pos] = map->queues[--map->len];
2041 break;
2042 }
2043
2044 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2045 kfree_rcu(map, rcu);
2046 return false;
2047 }
2048
2049 return true;
2050}
2051
2052static bool remove_xps_queue_cpu(struct net_device *dev,
2053 struct xps_dev_maps *dev_maps,
2054 int cpu, u16 offset, u16 count)
2055{
2056 int num_tc = dev->num_tc ? : 1;
2057 bool active = false;
2058 int tci;
2059
2060 for (tci = cpu * num_tc; num_tc--; tci++) {
2061 int i, j;
2062
2063 for (i = count, j = offset; i--; j++) {
2064 if (!remove_xps_queue(dev_maps, cpu, j))
2065 break;
2066 }
2067
2068 active |= i < 0;
2069 }
2070
2071 return active;
2072}
2073
2074static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2075 u16 count)
2076{
2077 struct xps_dev_maps *dev_maps;
2078 int cpu, i;
2079 bool active = false;
2080
2081 mutex_lock(&xps_map_mutex);
2082 dev_maps = xmap_dereference(dev->xps_maps);
2083
2084 if (!dev_maps)
2085 goto out_no_maps;
2086
2087 for_each_possible_cpu(cpu)
2088 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2089 offset, count);
2090
2091 if (!active) {
2092 RCU_INIT_POINTER(dev->xps_maps, NULL);
2093 kfree_rcu(dev_maps, rcu);
2094 }
2095
2096 for (i = offset + (count - 1); count--; i--)
2097 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2098 NUMA_NO_NODE);
2099
2100out_no_maps:
2101 mutex_unlock(&xps_map_mutex);
2102}
2103
2104static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2105{
2106 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2107}
2108
2109static struct xps_map *expand_xps_map(struct xps_map *map,
2110 int cpu, u16 index)
2111{
2112 struct xps_map *new_map;
2113 int alloc_len = XPS_MIN_MAP_ALLOC;
2114 int i, pos;
2115
2116 for (pos = 0; map && pos < map->len; pos++) {
2117 if (map->queues[pos] != index)
2118 continue;
2119 return map;
2120 }
2121
2122 /* Need to add queue to this CPU's existing map */
2123 if (map) {
2124 if (pos < map->alloc_len)
2125 return map;
2126
2127 alloc_len = map->alloc_len * 2;
2128 }
2129
2130 /* Need to allocate new map to store queue on this CPU's map */
2131 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2132 cpu_to_node(cpu));
2133 if (!new_map)
2134 return NULL;
2135
2136 for (i = 0; i < pos; i++)
2137 new_map->queues[i] = map->queues[i];
2138 new_map->alloc_len = alloc_len;
2139 new_map->len = pos;
2140
2141 return new_map;
2142}
2143
2144int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2145 u16 index)
2146{
2147 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2148 int i, cpu, tci, numa_node_id = -2;
2149 int maps_sz, num_tc = 1, tc = 0;
2150 struct xps_map *map, *new_map;
2151 bool active = false;
2152
2153 if (dev->num_tc) {
2154 num_tc = dev->num_tc;
2155 tc = netdev_txq_to_tc(dev, index);
2156 if (tc < 0)
2157 return -EINVAL;
2158 }
2159
2160 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2161 if (maps_sz < L1_CACHE_BYTES)
2162 maps_sz = L1_CACHE_BYTES;
2163
2164 mutex_lock(&xps_map_mutex);
2165
2166 dev_maps = xmap_dereference(dev->xps_maps);
2167
2168 /* allocate memory for queue storage */
2169 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2170 if (!new_dev_maps)
2171 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2172 if (!new_dev_maps) {
2173 mutex_unlock(&xps_map_mutex);
2174 return -ENOMEM;
2175 }
2176
2177 tci = cpu * num_tc + tc;
2178 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2179 NULL;
2180
2181 map = expand_xps_map(map, cpu, index);
2182 if (!map)
2183 goto error;
2184
2185 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2186 }
2187
2188 if (!new_dev_maps)
2189 goto out_no_new_maps;
2190
2191 for_each_possible_cpu(cpu) {
2192 /* copy maps belonging to foreign traffic classes */
2193 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2194 /* fill in the new device map from the old device map */
2195 map = xmap_dereference(dev_maps->cpu_map[tci]);
2196 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2197 }
2198
2199 /* We need to explicitly update tci as prevous loop
2200 * could break out early if dev_maps is NULL.
2201 */
2202 tci = cpu * num_tc + tc;
2203
2204 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2205 /* add queue to CPU maps */
2206 int pos = 0;
2207
2208 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2209 while ((pos < map->len) && (map->queues[pos] != index))
2210 pos++;
2211
2212 if (pos == map->len)
2213 map->queues[map->len++] = index;
2214#ifdef CONFIG_NUMA
2215 if (numa_node_id == -2)
2216 numa_node_id = cpu_to_node(cpu);
2217 else if (numa_node_id != cpu_to_node(cpu))
2218 numa_node_id = -1;
2219#endif
2220 } else if (dev_maps) {
2221 /* fill in the new device map from the old device map */
2222 map = xmap_dereference(dev_maps->cpu_map[tci]);
2223 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2224 }
2225
2226 /* copy maps belonging to foreign traffic classes */
2227 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2228 /* fill in the new device map from the old device map */
2229 map = xmap_dereference(dev_maps->cpu_map[tci]);
2230 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2231 }
2232 }
2233
2234 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2235
2236 /* Cleanup old maps */
2237 if (!dev_maps)
2238 goto out_no_old_maps;
2239
2240 for_each_possible_cpu(cpu) {
2241 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2242 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2243 map = xmap_dereference(dev_maps->cpu_map[tci]);
2244 if (map && map != new_map)
2245 kfree_rcu(map, rcu);
2246 }
2247 }
2248
2249 kfree_rcu(dev_maps, rcu);
2250
2251out_no_old_maps:
2252 dev_maps = new_dev_maps;
2253 active = true;
2254
2255out_no_new_maps:
2256 /* update Tx queue numa node */
2257 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2258 (numa_node_id >= 0) ? numa_node_id :
2259 NUMA_NO_NODE);
2260
2261 if (!dev_maps)
2262 goto out_no_maps;
2263
2264 /* removes queue from unused CPUs */
2265 for_each_possible_cpu(cpu) {
2266 for (i = tc, tci = cpu * num_tc; i--; tci++)
2267 active |= remove_xps_queue(dev_maps, tci, index);
2268 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2269 active |= remove_xps_queue(dev_maps, tci, index);
2270 for (i = num_tc - tc, tci++; --i; tci++)
2271 active |= remove_xps_queue(dev_maps, tci, index);
2272 }
2273
2274 /* free map if not active */
2275 if (!active) {
2276 RCU_INIT_POINTER(dev->xps_maps, NULL);
2277 kfree_rcu(dev_maps, rcu);
2278 }
2279
2280out_no_maps:
2281 mutex_unlock(&xps_map_mutex);
2282
2283 return 0;
2284error:
2285 /* remove any maps that we added */
2286 for_each_possible_cpu(cpu) {
2287 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2288 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2289 map = dev_maps ?
2290 xmap_dereference(dev_maps->cpu_map[tci]) :
2291 NULL;
2292 if (new_map && new_map != map)
2293 kfree(new_map);
2294 }
2295 }
2296
2297 mutex_unlock(&xps_map_mutex);
2298
2299 kfree(new_dev_maps);
2300 return -ENOMEM;
2301}
2302EXPORT_SYMBOL(netif_set_xps_queue);
2303
2304#endif
2305void netdev_reset_tc(struct net_device *dev)
2306{
2307#ifdef CONFIG_XPS
2308 netif_reset_xps_queues_gt(dev, 0);
2309#endif
2310 dev->num_tc = 0;
2311 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2312 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2313}
2314EXPORT_SYMBOL(netdev_reset_tc);
2315
2316int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2317{
2318 if (tc >= dev->num_tc)
2319 return -EINVAL;
2320
2321#ifdef CONFIG_XPS
2322 netif_reset_xps_queues(dev, offset, count);
2323#endif
2324 dev->tc_to_txq[tc].count = count;
2325 dev->tc_to_txq[tc].offset = offset;
2326 return 0;
2327}
2328EXPORT_SYMBOL(netdev_set_tc_queue);
2329
2330int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2331{
2332 if (num_tc > TC_MAX_QUEUE)
2333 return -EINVAL;
2334
2335#ifdef CONFIG_XPS
2336 netif_reset_xps_queues_gt(dev, 0);
2337#endif
2338 dev->num_tc = num_tc;
2339 return 0;
2340}
2341EXPORT_SYMBOL(netdev_set_num_tc);
2342
2343/*
2344 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2345 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
2346 */
2347int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2348{
2349 int rc;
2350
2351 if (txq < 1 || txq > dev->num_tx_queues)
2352 return -EINVAL;
2353
2354 if (dev->reg_state == NETREG_REGISTERED ||
2355 dev->reg_state == NETREG_UNREGISTERING) {
2356 ASSERT_RTNL();
2357
2358 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2359 txq);
2360 if (rc)
2361 return rc;
2362
2363 if (dev->num_tc)
2364 netif_setup_tc(dev, txq);
2365
2366 if (txq < dev->real_num_tx_queues) {
2367 qdisc_reset_all_tx_gt(dev, txq);
2368#ifdef CONFIG_XPS
2369 netif_reset_xps_queues_gt(dev, txq);
2370#endif
2371 }
2372 }
2373
2374 dev->real_num_tx_queues = txq;
2375 return 0;
2376}
2377EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2378
2379#ifdef CONFIG_SYSFS
2380/**
2381 * netif_set_real_num_rx_queues - set actual number of RX queues used
2382 * @dev: Network device
2383 * @rxq: Actual number of RX queues
2384 *
2385 * This must be called either with the rtnl_lock held or before
2386 * registration of the net device. Returns 0 on success, or a
2387 * negative error code. If called before registration, it always
2388 * succeeds.
2389 */
2390int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2391{
2392 int rc;
2393
2394 if (rxq < 1 || rxq > dev->num_rx_queues)
2395 return -EINVAL;
2396
2397 if (dev->reg_state == NETREG_REGISTERED) {
2398 ASSERT_RTNL();
2399
2400 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2401 rxq);
2402 if (rc)
2403 return rc;
2404 }
2405
2406 dev->real_num_rx_queues = rxq;
2407 return 0;
2408}
2409EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2410#endif
2411
2412/**
2413 * netif_get_num_default_rss_queues - default number of RSS queues
2414 *
2415 * This routine should set an upper limit on the number of RSS queues
2416 * used by default by multiqueue devices.
2417 */
2418int netif_get_num_default_rss_queues(void)
2419{
2420 return is_kdump_kernel() ?
2421 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2422}
2423EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2424
2425static void __netif_reschedule(struct Qdisc *q)
2426{
2427 struct softnet_data *sd;
2428 unsigned long flags;
2429
2430 local_irq_save(flags);
2431 sd = this_cpu_ptr(&softnet_data);
2432 q->next_sched = NULL;
2433 *sd->output_queue_tailp = q;
2434 sd->output_queue_tailp = &q->next_sched;
2435 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2436 local_irq_restore(flags);
2437}
2438
2439void __netif_schedule(struct Qdisc *q)
2440{
2441 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2442 __netif_reschedule(q);
2443}
2444EXPORT_SYMBOL(__netif_schedule);
2445
2446struct dev_kfree_skb_cb {
2447 enum skb_free_reason reason;
2448};
2449
2450static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2451{
2452 return (struct dev_kfree_skb_cb *)skb->cb;
2453}
2454
2455void netif_schedule_queue(struct netdev_queue *txq)
2456{
2457 rcu_read_lock();
2458 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2459 struct Qdisc *q = rcu_dereference(txq->qdisc);
2460
2461 __netif_schedule(q);
2462 }
2463 rcu_read_unlock();
2464}
2465EXPORT_SYMBOL(netif_schedule_queue);
2466
2467void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2468{
2469 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2470 struct Qdisc *q;
2471
2472 rcu_read_lock();
2473 q = rcu_dereference(dev_queue->qdisc);
2474 __netif_schedule(q);
2475 rcu_read_unlock();
2476 }
2477}
2478EXPORT_SYMBOL(netif_tx_wake_queue);
2479
2480void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2481{
2482 unsigned long flags;
2483
2484 if (unlikely(!skb))
2485 return;
2486
2487 if (likely(refcount_read(&skb->users) == 1)) {
2488 smp_rmb();
2489 refcount_set(&skb->users, 0);
2490 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2491 return;
2492 }
2493 get_kfree_skb_cb(skb)->reason = reason;
2494 local_irq_save(flags);
2495 skb->next = __this_cpu_read(softnet_data.completion_queue);
2496 __this_cpu_write(softnet_data.completion_queue, skb);
2497 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2498 local_irq_restore(flags);
2499}
2500EXPORT_SYMBOL(__dev_kfree_skb_irq);
2501
2502void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2503{
2504 if (in_irq() || irqs_disabled())
2505 __dev_kfree_skb_irq(skb, reason);
2506 else
2507 dev_kfree_skb(skb);
2508}
2509EXPORT_SYMBOL(__dev_kfree_skb_any);
2510
2511
2512/**
2513 * netif_device_detach - mark device as removed
2514 * @dev: network device
2515 *
2516 * Mark device as removed from system and therefore no longer available.
2517 */
2518void netif_device_detach(struct net_device *dev)
2519{
2520 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2521 netif_running(dev)) {
2522 netif_tx_stop_all_queues(dev);
2523 }
2524}
2525EXPORT_SYMBOL(netif_device_detach);
2526
2527/**
2528 * netif_device_attach - mark device as attached
2529 * @dev: network device
2530 *
2531 * Mark device as attached from system and restart if needed.
2532 */
2533void netif_device_attach(struct net_device *dev)
2534{
2535 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2536 netif_running(dev)) {
2537 netif_tx_wake_all_queues(dev);
2538 __netdev_watchdog_up(dev);
2539 }
2540}
2541EXPORT_SYMBOL(netif_device_attach);
2542
2543/*
2544 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2545 * to be used as a distribution range.
2546 */
2547u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2548 unsigned int num_tx_queues)
2549{
2550 u32 hash;
2551 u16 qoffset = 0;
2552 u16 qcount = num_tx_queues;
2553
2554 if (skb_rx_queue_recorded(skb)) {
2555 hash = skb_get_rx_queue(skb);
2556 while (unlikely(hash >= num_tx_queues))
2557 hash -= num_tx_queues;
2558 return hash;
2559 }
2560
2561 if (dev->num_tc) {
2562 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2563
2564 qoffset = dev->tc_to_txq[tc].offset;
2565 qcount = dev->tc_to_txq[tc].count;
2566 }
2567
2568 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2569}
2570EXPORT_SYMBOL(__skb_tx_hash);
2571
2572static void skb_warn_bad_offload(const struct sk_buff *skb)
2573{
2574 static const netdev_features_t null_features;
2575 struct net_device *dev = skb->dev;
2576 const char *name = "";
2577
2578 if (!net_ratelimit())
2579 return;
2580
2581 if (dev) {
2582 if (dev->dev.parent)
2583 name = dev_driver_string(dev->dev.parent);
2584 else
2585 name = netdev_name(dev);
2586 }
2587 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2588 "gso_type=%d ip_summed=%d\n",
2589 name, dev ? &dev->features : &null_features,
2590 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2591 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2592 skb_shinfo(skb)->gso_type, skb->ip_summed);
2593}
2594
2595/*
2596 * Invalidate hardware checksum when packet is to be mangled, and
2597 * complete checksum manually on outgoing path.
2598 */
2599int skb_checksum_help(struct sk_buff *skb)
2600{
2601 __wsum csum;
2602 int ret = 0, offset;
2603
2604 if (skb->ip_summed == CHECKSUM_COMPLETE)
2605 goto out_set_summed;
2606
2607 if (unlikely(skb_shinfo(skb)->gso_size)) {
2608 skb_warn_bad_offload(skb);
2609 return -EINVAL;
2610 }
2611
2612 /* Before computing a checksum, we should make sure no frag could
2613 * be modified by an external entity : checksum could be wrong.
2614 */
2615 if (skb_has_shared_frag(skb)) {
2616 ret = __skb_linearize(skb);
2617 if (ret)
2618 goto out;
2619 }
2620
2621 offset = skb_checksum_start_offset(skb);
2622 BUG_ON(offset >= skb_headlen(skb));
2623 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2624
2625 offset += skb->csum_offset;
2626 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2627
2628 if (skb_cloned(skb) &&
2629 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2630 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2631 if (ret)
2632 goto out;
2633 }
2634
2635 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2636out_set_summed:
2637 skb->ip_summed = CHECKSUM_NONE;
2638out:
2639 return ret;
2640}
2641EXPORT_SYMBOL(skb_checksum_help);
2642
2643int skb_crc32c_csum_help(struct sk_buff *skb)
2644{
2645 __le32 crc32c_csum;
2646 int ret = 0, offset, start;
2647
2648 if (skb->ip_summed != CHECKSUM_PARTIAL)
2649 goto out;
2650
2651 if (unlikely(skb_is_gso(skb)))
2652 goto out;
2653
2654 /* Before computing a checksum, we should make sure no frag could
2655 * be modified by an external entity : checksum could be wrong.
2656 */
2657 if (unlikely(skb_has_shared_frag(skb))) {
2658 ret = __skb_linearize(skb);
2659 if (ret)
2660 goto out;
2661 }
2662 start = skb_checksum_start_offset(skb);
2663 offset = start + offsetof(struct sctphdr, checksum);
2664 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2665 ret = -EINVAL;
2666 goto out;
2667 }
2668 if (skb_cloned(skb) &&
2669 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2670 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2671 if (ret)
2672 goto out;
2673 }
2674 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2675 skb->len - start, ~(__u32)0,
2676 crc32c_csum_stub));
2677 *(__le32 *)(skb->data + offset) = crc32c_csum;
2678 skb->ip_summed = CHECKSUM_NONE;
2679 skb->csum_not_inet = 0;
2680out:
2681 return ret;
2682}
2683
2684__be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2685{
2686 __be16 type = skb->protocol;
2687
2688 /* Tunnel gso handlers can set protocol to ethernet. */
2689 if (type == htons(ETH_P_TEB)) {
2690 struct ethhdr *eth;
2691
2692 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2693 return 0;
2694
2695 eth = (struct ethhdr *)skb_mac_header(skb);
2696 type = eth->h_proto;
2697 }
2698
2699 return __vlan_get_protocol(skb, type, depth);
2700}
2701
2702/**
2703 * skb_mac_gso_segment - mac layer segmentation handler.
2704 * @skb: buffer to segment
2705 * @features: features for the output path (see dev->features)
2706 */
2707struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2708 netdev_features_t features)
2709{
2710 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2711 struct packet_offload *ptype;
2712 int vlan_depth = skb->mac_len;
2713 __be16 type = skb_network_protocol(skb, &vlan_depth);
2714
2715 if (unlikely(!type))
2716 return ERR_PTR(-EINVAL);
2717
2718 __skb_pull(skb, vlan_depth);
2719
2720 rcu_read_lock();
2721 list_for_each_entry_rcu(ptype, &offload_base, list) {
2722 if (ptype->type == type && ptype->callbacks.gso_segment) {
2723 segs = ptype->callbacks.gso_segment(skb, features);
2724 break;
2725 }
2726 }
2727 rcu_read_unlock();
2728
2729 __skb_push(skb, skb->data - skb_mac_header(skb));
2730
2731 return segs;
2732}
2733EXPORT_SYMBOL(skb_mac_gso_segment);
2734
2735
2736/* openvswitch calls this on rx path, so we need a different check.
2737 */
2738static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2739{
2740 if (tx_path)
2741 return skb->ip_summed != CHECKSUM_PARTIAL &&
2742 skb->ip_summed != CHECKSUM_UNNECESSARY;
2743
2744 return skb->ip_summed == CHECKSUM_NONE;
2745}
2746
2747/**
2748 * __skb_gso_segment - Perform segmentation on skb.
2749 * @skb: buffer to segment
2750 * @features: features for the output path (see dev->features)
2751 * @tx_path: whether it is called in TX path
2752 *
2753 * This function segments the given skb and returns a list of segments.
2754 *
2755 * It may return NULL if the skb requires no segmentation. This is
2756 * only possible when GSO is used for verifying header integrity.
2757 *
2758 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2759 */
2760struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2761 netdev_features_t features, bool tx_path)
2762{
2763 struct sk_buff *segs;
2764
2765 if (unlikely(skb_needs_check(skb, tx_path))) {
2766 int err;
2767
2768 /* We're going to init ->check field in TCP or UDP header */
2769 err = skb_cow_head(skb, 0);
2770 if (err < 0)
2771 return ERR_PTR(err);
2772 }
2773
2774 /* Only report GSO partial support if it will enable us to
2775 * support segmentation on this frame without needing additional
2776 * work.
2777 */
2778 if (features & NETIF_F_GSO_PARTIAL) {
2779 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2780 struct net_device *dev = skb->dev;
2781
2782 partial_features |= dev->features & dev->gso_partial_features;
2783 if (!skb_gso_ok(skb, features | partial_features))
2784 features &= ~NETIF_F_GSO_PARTIAL;
2785 }
2786
2787 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2788 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2789
2790 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2791 SKB_GSO_CB(skb)->encap_level = 0;
2792
2793 skb_reset_mac_header(skb);
2794 skb_reset_mac_len(skb);
2795
2796 segs = skb_mac_gso_segment(skb, features);
2797
2798 if (unlikely(skb_needs_check(skb, tx_path)))
2799 skb_warn_bad_offload(skb);
2800
2801 return segs;
2802}
2803EXPORT_SYMBOL(__skb_gso_segment);
2804
2805/* Take action when hardware reception checksum errors are detected. */
2806#ifdef CONFIG_BUG
2807void netdev_rx_csum_fault(struct net_device *dev)
2808{
2809 if (net_ratelimit()) {
2810 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2811 dump_stack();
2812 }
2813}
2814EXPORT_SYMBOL(netdev_rx_csum_fault);
2815#endif
2816
2817/* Actually, we should eliminate this check as soon as we know, that:
2818 * 1. IOMMU is present and allows to map all the memory.
2819 * 2. No high memory really exists on this machine.
2820 */
2821
2822static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2823{
2824#ifdef CONFIG_HIGHMEM
2825 int i;
2826
2827 if (!(dev->features & NETIF_F_HIGHDMA)) {
2828 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2829 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2830
2831 if (PageHighMem(skb_frag_page(frag)))
2832 return 1;
2833 }
2834 }
2835
2836 if (PCI_DMA_BUS_IS_PHYS) {
2837 struct device *pdev = dev->dev.parent;
2838
2839 if (!pdev)
2840 return 0;
2841 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2842 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2843 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2844
2845 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2846 return 1;
2847 }
2848 }
2849#endif
2850 return 0;
2851}
2852
2853/* If MPLS offload request, verify we are testing hardware MPLS features
2854 * instead of standard features for the netdev.
2855 */
2856#if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2857static netdev_features_t net_mpls_features(struct sk_buff *skb,
2858 netdev_features_t features,
2859 __be16 type)
2860{
2861 if (eth_p_mpls(type))
2862 features &= skb->dev->mpls_features;
2863
2864 return features;
2865}
2866#else
2867static netdev_features_t net_mpls_features(struct sk_buff *skb,
2868 netdev_features_t features,
2869 __be16 type)
2870{
2871 return features;
2872}
2873#endif
2874
2875static netdev_features_t harmonize_features(struct sk_buff *skb,
2876 netdev_features_t features)
2877{
2878 int tmp;
2879 __be16 type;
2880
2881 type = skb_network_protocol(skb, &tmp);
2882 features = net_mpls_features(skb, features, type);
2883
2884 if (skb->ip_summed != CHECKSUM_NONE &&
2885 !can_checksum_protocol(features, type)) {
2886 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2887 }
2888 if (illegal_highdma(skb->dev, skb))
2889 features &= ~NETIF_F_SG;
2890
2891 return features;
2892}
2893
2894netdev_features_t passthru_features_check(struct sk_buff *skb,
2895 struct net_device *dev,
2896 netdev_features_t features)
2897{
2898 return features;
2899}
2900EXPORT_SYMBOL(passthru_features_check);
2901
2902static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2903 struct net_device *dev,
2904 netdev_features_t features)
2905{
2906 return vlan_features_check(skb, features);
2907}
2908
2909static netdev_features_t gso_features_check(const struct sk_buff *skb,
2910 struct net_device *dev,
2911 netdev_features_t features)
2912{
2913 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2914
2915 if (gso_segs > dev->gso_max_segs)
2916 return features & ~NETIF_F_GSO_MASK;
2917
2918 /* Support for GSO partial features requires software
2919 * intervention before we can actually process the packets
2920 * so we need to strip support for any partial features now
2921 * and we can pull them back in after we have partially
2922 * segmented the frame.
2923 */
2924 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2925 features &= ~dev->gso_partial_features;
2926
2927 /* Make sure to clear the IPv4 ID mangling feature if the
2928 * IPv4 header has the potential to be fragmented.
2929 */
2930 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2931 struct iphdr *iph = skb->encapsulation ?
2932 inner_ip_hdr(skb) : ip_hdr(skb);
2933
2934 if (!(iph->frag_off & htons(IP_DF)))
2935 features &= ~NETIF_F_TSO_MANGLEID;
2936 }
2937
2938 return features;
2939}
2940
2941netdev_features_t netif_skb_features(struct sk_buff *skb)
2942{
2943 struct net_device *dev = skb->dev;
2944 netdev_features_t features = dev->features;
2945
2946 if (skb_is_gso(skb))
2947 features = gso_features_check(skb, dev, features);
2948
2949 /* If encapsulation offload request, verify we are testing
2950 * hardware encapsulation features instead of standard
2951 * features for the netdev
2952 */
2953 if (skb->encapsulation)
2954 features &= dev->hw_enc_features;
2955
2956 if (skb_vlan_tagged(skb))
2957 features = netdev_intersect_features(features,
2958 dev->vlan_features |
2959 NETIF_F_HW_VLAN_CTAG_TX |
2960 NETIF_F_HW_VLAN_STAG_TX);
2961
2962 if (dev->netdev_ops->ndo_features_check)
2963 features &= dev->netdev_ops->ndo_features_check(skb, dev,
2964 features);
2965 else
2966 features &= dflt_features_check(skb, dev, features);
2967
2968 return harmonize_features(skb, features);
2969}
2970EXPORT_SYMBOL(netif_skb_features);
2971
2972static int xmit_one(struct sk_buff *skb, struct net_device *dev,
2973 struct netdev_queue *txq, bool more)
2974{
2975 unsigned int len;
2976 int rc;
2977
2978 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
2979 dev_queue_xmit_nit(skb, dev);
2980
2981 len = skb->len;
2982 trace_net_dev_start_xmit(skb, dev);
2983 rc = netdev_start_xmit(skb, dev, txq, more);
2984 trace_net_dev_xmit(skb, rc, dev, len);
2985
2986 return rc;
2987}
2988
2989struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
2990 struct netdev_queue *txq, int *ret)
2991{
2992 struct sk_buff *skb = first;
2993 int rc = NETDEV_TX_OK;
2994
2995 while (skb) {
2996 struct sk_buff *next = skb->next;
2997
2998 skb->next = NULL;
2999 rc = xmit_one(skb, dev, txq, next != NULL);
3000 if (unlikely(!dev_xmit_complete(rc))) {
3001 skb->next = next;
3002 goto out;
3003 }
3004
3005 skb = next;
3006 if (netif_xmit_stopped(txq) && skb) {
3007 rc = NETDEV_TX_BUSY;
3008 break;
3009 }
3010 }
3011
3012out:
3013 *ret = rc;
3014 return skb;
3015}
3016
3017static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3018 netdev_features_t features)
3019{
3020 if (skb_vlan_tag_present(skb) &&
3021 !vlan_hw_offload_capable(features, skb->vlan_proto))
3022 skb = __vlan_hwaccel_push_inside(skb);
3023 return skb;
3024}
3025
3026int skb_csum_hwoffload_help(struct sk_buff *skb,
3027 const netdev_features_t features)
3028{
3029 if (unlikely(skb->csum_not_inet))
3030 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3031 skb_crc32c_csum_help(skb);
3032
3033 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3034}
3035EXPORT_SYMBOL(skb_csum_hwoffload_help);
3036
3037static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev)
3038{
3039 netdev_features_t features;
3040
3041 features = netif_skb_features(skb);
3042 skb = validate_xmit_vlan(skb, features);
3043 if (unlikely(!skb))
3044 goto out_null;
3045
3046 if (netif_needs_gso(skb, features)) {
3047 struct sk_buff *segs;
3048
3049 segs = skb_gso_segment(skb, features);
3050 if (IS_ERR(segs)) {
3051 goto out_kfree_skb;
3052 } else if (segs) {
3053 consume_skb(skb);
3054 skb = segs;
3055 }
3056 } else {
3057 if (skb_needs_linearize(skb, features) &&
3058 __skb_linearize(skb))
3059 goto out_kfree_skb;
3060
3061 if (validate_xmit_xfrm(skb, features))
3062 goto out_kfree_skb;
3063
3064 /* If packet is not checksummed and device does not
3065 * support checksumming for this protocol, complete
3066 * checksumming here.
3067 */
3068 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3069 if (skb->encapsulation)
3070 skb_set_inner_transport_header(skb,
3071 skb_checksum_start_offset(skb));
3072 else
3073 skb_set_transport_header(skb,
3074 skb_checksum_start_offset(skb));
3075 if (skb_csum_hwoffload_help(skb, features))
3076 goto out_kfree_skb;
3077 }
3078 }
3079
3080 return skb;
3081
3082out_kfree_skb:
3083 kfree_skb(skb);
3084out_null:
3085 atomic_long_inc(&dev->tx_dropped);
3086 return NULL;
3087}
3088
3089struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev)
3090{
3091 struct sk_buff *next, *head = NULL, *tail;
3092
3093 for (; skb != NULL; skb = next) {
3094 next = skb->next;
3095 skb->next = NULL;
3096
3097 /* in case skb wont be segmented, point to itself */
3098 skb->prev = skb;
3099
3100 skb = validate_xmit_skb(skb, dev);
3101 if (!skb)
3102 continue;
3103
3104 if (!head)
3105 head = skb;
3106 else
3107 tail->next = skb;
3108 /* If skb was segmented, skb->prev points to
3109 * the last segment. If not, it still contains skb.
3110 */
3111 tail = skb->prev;
3112 }
3113 return head;
3114}
3115EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3116
3117static void qdisc_pkt_len_init(struct sk_buff *skb)
3118{
3119 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3120
3121 qdisc_skb_cb(skb)->pkt_len = skb->len;
3122
3123 /* To get more precise estimation of bytes sent on wire,
3124 * we add to pkt_len the headers size of all segments
3125 */
3126 if (shinfo->gso_size) {
3127 unsigned int hdr_len;
3128 u16 gso_segs = shinfo->gso_segs;
3129
3130 /* mac layer + network layer */
3131 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3132
3133 /* + transport layer */
3134 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
3135 hdr_len += tcp_hdrlen(skb);
3136 else
3137 hdr_len += sizeof(struct udphdr);
3138
3139 if (shinfo->gso_type & SKB_GSO_DODGY)
3140 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3141 shinfo->gso_size);
3142
3143 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3144 }
3145}
3146
3147static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3148 struct net_device *dev,
3149 struct netdev_queue *txq)
3150{
3151 spinlock_t *root_lock = qdisc_lock(q);
3152 struct sk_buff *to_free = NULL;
3153 bool contended;
3154 int rc;
3155
3156 qdisc_calculate_pkt_len(skb, q);
3157 /*
3158 * Heuristic to force contended enqueues to serialize on a
3159 * separate lock before trying to get qdisc main lock.
3160 * This permits qdisc->running owner to get the lock more
3161 * often and dequeue packets faster.
3162 */
3163 contended = qdisc_is_running(q);
3164 if (unlikely(contended))
3165 spin_lock(&q->busylock);
3166
3167 spin_lock(root_lock);
3168 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3169 __qdisc_drop(skb, &to_free);
3170 rc = NET_XMIT_DROP;
3171 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3172 qdisc_run_begin(q)) {
3173 /*
3174 * This is a work-conserving queue; there are no old skbs
3175 * waiting to be sent out; and the qdisc is not running -
3176 * xmit the skb directly.
3177 */
3178
3179 qdisc_bstats_update(q, skb);
3180
3181 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3182 if (unlikely(contended)) {
3183 spin_unlock(&q->busylock);
3184 contended = false;
3185 }
3186 __qdisc_run(q);
3187 } else
3188 qdisc_run_end(q);
3189
3190 rc = NET_XMIT_SUCCESS;
3191 } else {
3192 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3193 if (qdisc_run_begin(q)) {
3194 if (unlikely(contended)) {
3195 spin_unlock(&q->busylock);
3196 contended = false;
3197 }
3198 __qdisc_run(q);
3199 }
3200 }
3201 spin_unlock(root_lock);
3202 if (unlikely(to_free))
3203 kfree_skb_list(to_free);
3204 if (unlikely(contended))
3205 spin_unlock(&q->busylock);
3206 return rc;
3207}
3208
3209#if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3210static void skb_update_prio(struct sk_buff *skb)
3211{
3212 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
3213
3214 if (!skb->priority && skb->sk && map) {
3215 unsigned int prioidx =
3216 sock_cgroup_prioidx(&skb->sk->sk_cgrp_data);
3217
3218 if (prioidx < map->priomap_len)
3219 skb->priority = map->priomap[prioidx];
3220 }
3221}
3222#else
3223#define skb_update_prio(skb)
3224#endif
3225
3226DEFINE_PER_CPU(int, xmit_recursion);
3227EXPORT_SYMBOL(xmit_recursion);
3228
3229/**
3230 * dev_loopback_xmit - loop back @skb
3231 * @net: network namespace this loopback is happening in
3232 * @sk: sk needed to be a netfilter okfn
3233 * @skb: buffer to transmit
3234 */
3235int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3236{
3237 skb_reset_mac_header(skb);
3238 __skb_pull(skb, skb_network_offset(skb));
3239 skb->pkt_type = PACKET_LOOPBACK;
3240 skb->ip_summed = CHECKSUM_UNNECESSARY;
3241 WARN_ON(!skb_dst(skb));
3242 skb_dst_force(skb);
3243 netif_rx_ni(skb);
3244 return 0;
3245}
3246EXPORT_SYMBOL(dev_loopback_xmit);
3247
3248#ifdef CONFIG_NET_EGRESS
3249static struct sk_buff *
3250sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3251{
3252 struct tcf_proto *cl = rcu_dereference_bh(dev->egress_cl_list);
3253 struct tcf_result cl_res;
3254
3255 if (!cl)
3256 return skb;
3257
3258 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3259 qdisc_bstats_cpu_update(cl->q, skb);
3260
3261 switch (tcf_classify(skb, cl, &cl_res, false)) {
3262 case TC_ACT_OK:
3263 case TC_ACT_RECLASSIFY:
3264 skb->tc_index = TC_H_MIN(cl_res.classid);
3265 break;
3266 case TC_ACT_SHOT:
3267 qdisc_qstats_cpu_drop(cl->q);
3268 *ret = NET_XMIT_DROP;
3269 kfree_skb(skb);
3270 return NULL;
3271 case TC_ACT_STOLEN:
3272 case TC_ACT_QUEUED:
3273 case TC_ACT_TRAP:
3274 *ret = NET_XMIT_SUCCESS;
3275 consume_skb(skb);
3276 return NULL;
3277 case TC_ACT_REDIRECT:
3278 /* No need to push/pop skb's mac_header here on egress! */
3279 skb_do_redirect(skb);
3280 *ret = NET_XMIT_SUCCESS;
3281 return NULL;
3282 default:
3283 break;
3284 }
3285
3286 return skb;
3287}
3288#endif /* CONFIG_NET_EGRESS */
3289
3290static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3291{
3292#ifdef CONFIG_XPS
3293 struct xps_dev_maps *dev_maps;
3294 struct xps_map *map;
3295 int queue_index = -1;
3296
3297 rcu_read_lock();
3298 dev_maps = rcu_dereference(dev->xps_maps);
3299 if (dev_maps) {
3300 unsigned int tci = skb->sender_cpu - 1;
3301
3302 if (dev->num_tc) {
3303 tci *= dev->num_tc;
3304 tci += netdev_get_prio_tc_map(dev, skb->priority);
3305 }
3306
3307 map = rcu_dereference(dev_maps->cpu_map[tci]);
3308 if (map) {
3309 if (map->len == 1)
3310 queue_index = map->queues[0];
3311 else
3312 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3313 map->len)];
3314 if (unlikely(queue_index >= dev->real_num_tx_queues))
3315 queue_index = -1;
3316 }
3317 }
3318 rcu_read_unlock();
3319
3320 return queue_index;
3321#else
3322 return -1;
3323#endif
3324}
3325
3326static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3327{
3328 struct sock *sk = skb->sk;
3329 int queue_index = sk_tx_queue_get(sk);
3330
3331 if (queue_index < 0 || skb->ooo_okay ||
3332 queue_index >= dev->real_num_tx_queues) {
3333 int new_index = get_xps_queue(dev, skb);
3334
3335 if (new_index < 0)
3336 new_index = skb_tx_hash(dev, skb);
3337
3338 if (queue_index != new_index && sk &&
3339 sk_fullsock(sk) &&
3340 rcu_access_pointer(sk->sk_dst_cache))
3341 sk_tx_queue_set(sk, new_index);
3342
3343 queue_index = new_index;
3344 }
3345
3346 return queue_index;
3347}
3348
3349struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3350 struct sk_buff *skb,
3351 void *accel_priv)
3352{
3353 int queue_index = 0;
3354
3355#ifdef CONFIG_XPS
3356 u32 sender_cpu = skb->sender_cpu - 1;
3357
3358 if (sender_cpu >= (u32)NR_CPUS)
3359 skb->sender_cpu = raw_smp_processor_id() + 1;
3360#endif
3361
3362 if (dev->real_num_tx_queues != 1) {
3363 const struct net_device_ops *ops = dev->netdev_ops;
3364
3365 if (ops->ndo_select_queue)
3366 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3367 __netdev_pick_tx);
3368 else
3369 queue_index = __netdev_pick_tx(dev, skb);
3370
3371 if (!accel_priv)
3372 queue_index = netdev_cap_txqueue(dev, queue_index);
3373 }
3374
3375 skb_set_queue_mapping(skb, queue_index);
3376 return netdev_get_tx_queue(dev, queue_index);
3377}
3378
3379/**
3380 * __dev_queue_xmit - transmit a buffer
3381 * @skb: buffer to transmit
3382 * @accel_priv: private data used for L2 forwarding offload
3383 *
3384 * Queue a buffer for transmission to a network device. The caller must
3385 * have set the device and priority and built the buffer before calling
3386 * this function. The function can be called from an interrupt.
3387 *
3388 * A negative errno code is returned on a failure. A success does not
3389 * guarantee the frame will be transmitted as it may be dropped due
3390 * to congestion or traffic shaping.
3391 *
3392 * -----------------------------------------------------------------------------------
3393 * I notice this method can also return errors from the queue disciplines,
3394 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3395 * be positive.
3396 *
3397 * Regardless of the return value, the skb is consumed, so it is currently
3398 * difficult to retry a send to this method. (You can bump the ref count
3399 * before sending to hold a reference for retry if you are careful.)
3400 *
3401 * When calling this method, interrupts MUST be enabled. This is because
3402 * the BH enable code must have IRQs enabled so that it will not deadlock.
3403 * --BLG
3404 */
3405static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3406{
3407 struct net_device *dev = skb->dev;
3408 struct netdev_queue *txq;
3409 struct Qdisc *q;
3410 int rc = -ENOMEM;
3411
3412 skb_reset_mac_header(skb);
3413
3414 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3415 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3416
3417 /* Disable soft irqs for various locks below. Also
3418 * stops preemption for RCU.
3419 */
3420 rcu_read_lock_bh();
3421
3422 skb_update_prio(skb);
3423
3424 qdisc_pkt_len_init(skb);
3425#ifdef CONFIG_NET_CLS_ACT
3426 skb->tc_at_ingress = 0;
3427# ifdef CONFIG_NET_EGRESS
3428 if (static_key_false(&egress_needed)) {
3429 skb = sch_handle_egress(skb, &rc, dev);
3430 if (!skb)
3431 goto out;
3432 }
3433# endif
3434#endif
3435 /* If device/qdisc don't need skb->dst, release it right now while
3436 * its hot in this cpu cache.
3437 */
3438 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3439 skb_dst_drop(skb);
3440 else
3441 skb_dst_force(skb);
3442
3443 txq = netdev_pick_tx(dev, skb, accel_priv);
3444 q = rcu_dereference_bh(txq->qdisc);
3445
3446 trace_net_dev_queue(skb);
3447 if (q->enqueue) {
3448 rc = __dev_xmit_skb(skb, q, dev, txq);
3449 goto out;
3450 }
3451
3452 /* The device has no queue. Common case for software devices:
3453 * loopback, all the sorts of tunnels...
3454
3455 * Really, it is unlikely that netif_tx_lock protection is necessary
3456 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3457 * counters.)
3458 * However, it is possible, that they rely on protection
3459 * made by us here.
3460
3461 * Check this and shot the lock. It is not prone from deadlocks.
3462 *Either shot noqueue qdisc, it is even simpler 8)
3463 */
3464 if (dev->flags & IFF_UP) {
3465 int cpu = smp_processor_id(); /* ok because BHs are off */
3466
3467 if (txq->xmit_lock_owner != cpu) {
3468 if (unlikely(__this_cpu_read(xmit_recursion) >
3469 XMIT_RECURSION_LIMIT))
3470 goto recursion_alert;
3471
3472 skb = validate_xmit_skb(skb, dev);
3473 if (!skb)
3474 goto out;
3475
3476 HARD_TX_LOCK(dev, txq, cpu);
3477
3478 if (!netif_xmit_stopped(txq)) {
3479 __this_cpu_inc(xmit_recursion);
3480 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3481 __this_cpu_dec(xmit_recursion);
3482 if (dev_xmit_complete(rc)) {
3483 HARD_TX_UNLOCK(dev, txq);
3484 goto out;
3485 }
3486 }
3487 HARD_TX_UNLOCK(dev, txq);
3488 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3489 dev->name);
3490 } else {
3491 /* Recursion is detected! It is possible,
3492 * unfortunately
3493 */
3494recursion_alert:
3495 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3496 dev->name);
3497 }
3498 }
3499
3500 rc = -ENETDOWN;
3501 rcu_read_unlock_bh();
3502
3503 atomic_long_inc(&dev->tx_dropped);
3504 kfree_skb_list(skb);
3505 return rc;
3506out:
3507 rcu_read_unlock_bh();
3508 return rc;
3509}
3510
3511int dev_queue_xmit(struct sk_buff *skb)
3512{
3513 return __dev_queue_xmit(skb, NULL);
3514}
3515EXPORT_SYMBOL(dev_queue_xmit);
3516
3517int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3518{
3519 return __dev_queue_xmit(skb, accel_priv);
3520}
3521EXPORT_SYMBOL(dev_queue_xmit_accel);
3522
3523
3524/*************************************************************************
3525 * Receiver routines
3526 *************************************************************************/
3527
3528int netdev_max_backlog __read_mostly = 1000;
3529EXPORT_SYMBOL(netdev_max_backlog);
3530
3531int netdev_tstamp_prequeue __read_mostly = 1;
3532int netdev_budget __read_mostly = 300;
3533unsigned int __read_mostly netdev_budget_usecs = 2000;
3534int weight_p __read_mostly = 64; /* old backlog weight */
3535int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3536int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3537int dev_rx_weight __read_mostly = 64;
3538int dev_tx_weight __read_mostly = 64;
3539
3540/* Called with irq disabled */
3541static inline void ____napi_schedule(struct softnet_data *sd,
3542 struct napi_struct *napi)
3543{
3544 list_add_tail(&napi->poll_list, &sd->poll_list);
3545 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3546}
3547
3548#ifdef CONFIG_RPS
3549
3550/* One global table that all flow-based protocols share. */
3551struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3552EXPORT_SYMBOL(rps_sock_flow_table);
3553u32 rps_cpu_mask __read_mostly;
3554EXPORT_SYMBOL(rps_cpu_mask);
3555
3556struct static_key rps_needed __read_mostly;
3557EXPORT_SYMBOL(rps_needed);
3558struct static_key rfs_needed __read_mostly;
3559EXPORT_SYMBOL(rfs_needed);
3560
3561static struct rps_dev_flow *
3562set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3563 struct rps_dev_flow *rflow, u16 next_cpu)
3564{
3565 if (next_cpu < nr_cpu_ids) {
3566#ifdef CONFIG_RFS_ACCEL
3567 struct netdev_rx_queue *rxqueue;
3568 struct rps_dev_flow_table *flow_table;
3569 struct rps_dev_flow *old_rflow;
3570 u32 flow_id;
3571 u16 rxq_index;
3572 int rc;
3573
3574 /* Should we steer this flow to a different hardware queue? */
3575 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3576 !(dev->features & NETIF_F_NTUPLE))
3577 goto out;
3578 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3579 if (rxq_index == skb_get_rx_queue(skb))
3580 goto out;
3581
3582 rxqueue = dev->_rx + rxq_index;
3583 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3584 if (!flow_table)
3585 goto out;
3586 flow_id = skb_get_hash(skb) & flow_table->mask;
3587 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3588 rxq_index, flow_id);
3589 if (rc < 0)
3590 goto out;
3591 old_rflow = rflow;
3592 rflow = &flow_table->flows[flow_id];
3593 rflow->filter = rc;
3594 if (old_rflow->filter == rflow->filter)
3595 old_rflow->filter = RPS_NO_FILTER;
3596 out:
3597#endif
3598 rflow->last_qtail =
3599 per_cpu(softnet_data, next_cpu).input_queue_head;
3600 }
3601
3602 rflow->cpu = next_cpu;
3603 return rflow;
3604}
3605
3606/*
3607 * get_rps_cpu is called from netif_receive_skb and returns the target
3608 * CPU from the RPS map of the receiving queue for a given skb.
3609 * rcu_read_lock must be held on entry.
3610 */
3611static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3612 struct rps_dev_flow **rflowp)
3613{
3614 const struct rps_sock_flow_table *sock_flow_table;
3615 struct netdev_rx_queue *rxqueue = dev->_rx;
3616 struct rps_dev_flow_table *flow_table;
3617 struct rps_map *map;
3618 int cpu = -1;
3619 u32 tcpu;
3620 u32 hash;
3621
3622 if (skb_rx_queue_recorded(skb)) {
3623 u16 index = skb_get_rx_queue(skb);
3624
3625 if (unlikely(index >= dev->real_num_rx_queues)) {
3626 WARN_ONCE(dev->real_num_rx_queues > 1,
3627 "%s received packet on queue %u, but number "
3628 "of RX queues is %u\n",
3629 dev->name, index, dev->real_num_rx_queues);
3630 goto done;
3631 }
3632 rxqueue += index;
3633 }
3634
3635 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3636
3637 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3638 map = rcu_dereference(rxqueue->rps_map);
3639 if (!flow_table && !map)
3640 goto done;
3641
3642 skb_reset_network_header(skb);
3643 hash = skb_get_hash(skb);
3644 if (!hash)
3645 goto done;
3646
3647 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3648 if (flow_table && sock_flow_table) {
3649 struct rps_dev_flow *rflow;
3650 u32 next_cpu;
3651 u32 ident;
3652
3653 /* First check into global flow table if there is a match */
3654 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3655 if ((ident ^ hash) & ~rps_cpu_mask)
3656 goto try_rps;
3657
3658 next_cpu = ident & rps_cpu_mask;
3659
3660 /* OK, now we know there is a match,
3661 * we can look at the local (per receive queue) flow table
3662 */
3663 rflow = &flow_table->flows[hash & flow_table->mask];
3664 tcpu = rflow->cpu;
3665
3666 /*
3667 * If the desired CPU (where last recvmsg was done) is
3668 * different from current CPU (one in the rx-queue flow
3669 * table entry), switch if one of the following holds:
3670 * - Current CPU is unset (>= nr_cpu_ids).
3671 * - Current CPU is offline.
3672 * - The current CPU's queue tail has advanced beyond the
3673 * last packet that was enqueued using this table entry.
3674 * This guarantees that all previous packets for the flow
3675 * have been dequeued, thus preserving in order delivery.
3676 */
3677 if (unlikely(tcpu != next_cpu) &&
3678 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3679 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3680 rflow->last_qtail)) >= 0)) {
3681 tcpu = next_cpu;
3682 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3683 }
3684
3685 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3686 *rflowp = rflow;
3687 cpu = tcpu;
3688 goto done;
3689 }
3690 }
3691
3692try_rps:
3693
3694 if (map) {
3695 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3696 if (cpu_online(tcpu)) {
3697 cpu = tcpu;
3698 goto done;
3699 }
3700 }
3701
3702done:
3703 return cpu;
3704}
3705
3706#ifdef CONFIG_RFS_ACCEL
3707
3708/**
3709 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3710 * @dev: Device on which the filter was set
3711 * @rxq_index: RX queue index
3712 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3713 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3714 *
3715 * Drivers that implement ndo_rx_flow_steer() should periodically call
3716 * this function for each installed filter and remove the filters for
3717 * which it returns %true.
3718 */
3719bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3720 u32 flow_id, u16 filter_id)
3721{
3722 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3723 struct rps_dev_flow_table *flow_table;
3724 struct rps_dev_flow *rflow;
3725 bool expire = true;
3726 unsigned int cpu;
3727
3728 rcu_read_lock();
3729 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3730 if (flow_table && flow_id <= flow_table->mask) {
3731 rflow = &flow_table->flows[flow_id];
3732 cpu = ACCESS_ONCE(rflow->cpu);
3733 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3734 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3735 rflow->last_qtail) <
3736 (int)(10 * flow_table->mask)))
3737 expire = false;
3738 }
3739 rcu_read_unlock();
3740 return expire;
3741}
3742EXPORT_SYMBOL(rps_may_expire_flow);
3743
3744#endif /* CONFIG_RFS_ACCEL */
3745
3746/* Called from hardirq (IPI) context */
3747static void rps_trigger_softirq(void *data)
3748{
3749 struct softnet_data *sd = data;
3750
3751 ____napi_schedule(sd, &sd->backlog);
3752 sd->received_rps++;
3753}
3754
3755#endif /* CONFIG_RPS */
3756
3757/*
3758 * Check if this softnet_data structure is another cpu one
3759 * If yes, queue it to our IPI list and return 1
3760 * If no, return 0
3761 */
3762static int rps_ipi_queued(struct softnet_data *sd)
3763{
3764#ifdef CONFIG_RPS
3765 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3766
3767 if (sd != mysd) {
3768 sd->rps_ipi_next = mysd->rps_ipi_list;
3769 mysd->rps_ipi_list = sd;
3770
3771 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3772 return 1;
3773 }
3774#endif /* CONFIG_RPS */
3775 return 0;
3776}
3777
3778#ifdef CONFIG_NET_FLOW_LIMIT
3779int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3780#endif
3781
3782static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3783{
3784#ifdef CONFIG_NET_FLOW_LIMIT
3785 struct sd_flow_limit *fl;
3786 struct softnet_data *sd;
3787 unsigned int old_flow, new_flow;
3788
3789 if (qlen < (netdev_max_backlog >> 1))
3790 return false;
3791
3792 sd = this_cpu_ptr(&softnet_data);
3793
3794 rcu_read_lock();
3795 fl = rcu_dereference(sd->flow_limit);
3796 if (fl) {
3797 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3798 old_flow = fl->history[fl->history_head];
3799 fl->history[fl->history_head] = new_flow;
3800
3801 fl->history_head++;
3802 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3803
3804 if (likely(fl->buckets[old_flow]))
3805 fl->buckets[old_flow]--;
3806
3807 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3808 fl->count++;
3809 rcu_read_unlock();
3810 return true;
3811 }
3812 }
3813 rcu_read_unlock();
3814#endif
3815 return false;
3816}
3817
3818/*
3819 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3820 * queue (may be a remote CPU queue).
3821 */
3822static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3823 unsigned int *qtail)
3824{
3825 struct softnet_data *sd;
3826 unsigned long flags;
3827 unsigned int qlen;
3828
3829 sd = &per_cpu(softnet_data, cpu);
3830
3831 local_irq_save(flags);
3832
3833 rps_lock(sd);
3834 if (!netif_running(skb->dev))
3835 goto drop;
3836 qlen = skb_queue_len(&sd->input_pkt_queue);
3837 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3838 if (qlen) {
3839enqueue:
3840 __skb_queue_tail(&sd->input_pkt_queue, skb);
3841 input_queue_tail_incr_save(sd, qtail);
3842 rps_unlock(sd);
3843 local_irq_restore(flags);
3844 return NET_RX_SUCCESS;
3845 }
3846
3847 /* Schedule NAPI for backlog device
3848 * We can use non atomic operation since we own the queue lock
3849 */
3850 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3851 if (!rps_ipi_queued(sd))
3852 ____napi_schedule(sd, &sd->backlog);
3853 }
3854 goto enqueue;
3855 }
3856
3857drop:
3858 sd->dropped++;
3859 rps_unlock(sd);
3860
3861 local_irq_restore(flags);
3862
3863 atomic_long_inc(&skb->dev->rx_dropped);
3864 kfree_skb(skb);
3865 return NET_RX_DROP;
3866}
3867
3868static int netif_rx_internal(struct sk_buff *skb)
3869{
3870 int ret;
3871
3872 net_timestamp_check(netdev_tstamp_prequeue, skb);
3873
3874 trace_netif_rx(skb);
3875#ifdef CONFIG_RPS
3876 if (static_key_false(&rps_needed)) {
3877 struct rps_dev_flow voidflow, *rflow = &voidflow;
3878 int cpu;
3879
3880 preempt_disable();
3881 rcu_read_lock();
3882
3883 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3884 if (cpu < 0)
3885 cpu = smp_processor_id();
3886
3887 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3888
3889 rcu_read_unlock();
3890 preempt_enable();
3891 } else
3892#endif
3893 {
3894 unsigned int qtail;
3895
3896 ret = enqueue_to_backlog(skb,