Loading...
1/*
2 * Routines having to do with the 'struct sk_buff' memory handlers.
3 *
4 * Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>
5 * Florian La Roche <rzsfl@rz.uni-sb.de>
6 *
7 * Fixes:
8 * Alan Cox : Fixed the worst of the load
9 * balancer bugs.
10 * Dave Platt : Interrupt stacking fix.
11 * Richard Kooijman : Timestamp fixes.
12 * Alan Cox : Changed buffer format.
13 * Alan Cox : destructor hook for AF_UNIX etc.
14 * Linus Torvalds : Better skb_clone.
15 * Alan Cox : Added skb_copy.
16 * Alan Cox : Added all the changed routines Linus
17 * only put in the headers
18 * Ray VanTassle : Fixed --skb->lock in free
19 * Alan Cox : skb_copy copy arp field
20 * Andi Kleen : slabified it.
21 * Robert Olsson : Removed skb_head_pool
22 *
23 * NOTE:
24 * The __skb_ routines should be called with interrupts
25 * disabled, or you better be *real* sure that the operation is atomic
26 * with respect to whatever list is being frobbed (e.g. via lock_sock()
27 * or via disabling bottom half handlers, etc).
28 *
29 * This program is free software; you can redistribute it and/or
30 * modify it under the terms of the GNU General Public License
31 * as published by the Free Software Foundation; either version
32 * 2 of the License, or (at your option) any later version.
33 */
34
35/*
36 * The functions in this file will not compile correctly with gcc 2.4.x
37 */
38
39#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
40
41#include <linux/module.h>
42#include <linux/types.h>
43#include <linux/kernel.h>
44#include <linux/mm.h>
45#include <linux/interrupt.h>
46#include <linux/in.h>
47#include <linux/inet.h>
48#include <linux/slab.h>
49#include <linux/tcp.h>
50#include <linux/udp.h>
51#include <linux/sctp.h>
52#include <linux/netdevice.h>
53#ifdef CONFIG_NET_CLS_ACT
54#include <net/pkt_sched.h>
55#endif
56#include <linux/string.h>
57#include <linux/skbuff.h>
58#include <linux/splice.h>
59#include <linux/cache.h>
60#include <linux/rtnetlink.h>
61#include <linux/init.h>
62#include <linux/scatterlist.h>
63#include <linux/errqueue.h>
64#include <linux/prefetch.h>
65#include <linux/if_vlan.h>
66
67#include <net/protocol.h>
68#include <net/dst.h>
69#include <net/sock.h>
70#include <net/checksum.h>
71#include <net/ip6_checksum.h>
72#include <net/xfrm.h>
73
74#include <linux/uaccess.h>
75#include <trace/events/skb.h>
76#include <linux/highmem.h>
77#include <linux/capability.h>
78#include <linux/user_namespace.h>
79
80struct kmem_cache *skbuff_head_cache __ro_after_init;
81static struct kmem_cache *skbuff_fclone_cache __ro_after_init;
82int sysctl_max_skb_frags __read_mostly = MAX_SKB_FRAGS;
83EXPORT_SYMBOL(sysctl_max_skb_frags);
84
85/**
86 * skb_panic - private function for out-of-line support
87 * @skb: buffer
88 * @sz: size
89 * @addr: address
90 * @msg: skb_over_panic or skb_under_panic
91 *
92 * Out-of-line support for skb_put() and skb_push().
93 * Called via the wrapper skb_over_panic() or skb_under_panic().
94 * Keep out of line to prevent kernel bloat.
95 * __builtin_return_address is not used because it is not always reliable.
96 */
97static void skb_panic(struct sk_buff *skb, unsigned int sz, void *addr,
98 const char msg[])
99{
100 pr_emerg("%s: text:%p len:%d put:%d head:%p data:%p tail:%#lx end:%#lx dev:%s\n",
101 msg, addr, skb->len, sz, skb->head, skb->data,
102 (unsigned long)skb->tail, (unsigned long)skb->end,
103 skb->dev ? skb->dev->name : "<NULL>");
104 BUG();
105}
106
107static void skb_over_panic(struct sk_buff *skb, unsigned int sz, void *addr)
108{
109 skb_panic(skb, sz, addr, __func__);
110}
111
112static void skb_under_panic(struct sk_buff *skb, unsigned int sz, void *addr)
113{
114 skb_panic(skb, sz, addr, __func__);
115}
116
117/*
118 * kmalloc_reserve is a wrapper around kmalloc_node_track_caller that tells
119 * the caller if emergency pfmemalloc reserves are being used. If it is and
120 * the socket is later found to be SOCK_MEMALLOC then PFMEMALLOC reserves
121 * may be used. Otherwise, the packet data may be discarded until enough
122 * memory is free
123 */
124#define kmalloc_reserve(size, gfp, node, pfmemalloc) \
125 __kmalloc_reserve(size, gfp, node, _RET_IP_, pfmemalloc)
126
127static void *__kmalloc_reserve(size_t size, gfp_t flags, int node,
128 unsigned long ip, bool *pfmemalloc)
129{
130 void *obj;
131 bool ret_pfmemalloc = false;
132
133 /*
134 * Try a regular allocation, when that fails and we're not entitled
135 * to the reserves, fail.
136 */
137 obj = kmalloc_node_track_caller(size,
138 flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
139 node);
140 if (obj || !(gfp_pfmemalloc_allowed(flags)))
141 goto out;
142
143 /* Try again but now we are using pfmemalloc reserves */
144 ret_pfmemalloc = true;
145 obj = kmalloc_node_track_caller(size, flags, node);
146
147out:
148 if (pfmemalloc)
149 *pfmemalloc = ret_pfmemalloc;
150
151 return obj;
152}
153
154/* Allocate a new skbuff. We do this ourselves so we can fill in a few
155 * 'private' fields and also do memory statistics to find all the
156 * [BEEP] leaks.
157 *
158 */
159
160/**
161 * __alloc_skb - allocate a network buffer
162 * @size: size to allocate
163 * @gfp_mask: allocation mask
164 * @flags: If SKB_ALLOC_FCLONE is set, allocate from fclone cache
165 * instead of head cache and allocate a cloned (child) skb.
166 * If SKB_ALLOC_RX is set, __GFP_MEMALLOC will be used for
167 * allocations in case the data is required for writeback
168 * @node: numa node to allocate memory on
169 *
170 * Allocate a new &sk_buff. The returned buffer has no headroom and a
171 * tail room of at least size bytes. The object has a reference count
172 * of one. The return is the buffer. On a failure the return is %NULL.
173 *
174 * Buffers may only be allocated from interrupts using a @gfp_mask of
175 * %GFP_ATOMIC.
176 */
177struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
178 int flags, int node)
179{
180 struct kmem_cache *cache;
181 struct skb_shared_info *shinfo;
182 struct sk_buff *skb;
183 u8 *data;
184 bool pfmemalloc;
185
186 cache = (flags & SKB_ALLOC_FCLONE)
187 ? skbuff_fclone_cache : skbuff_head_cache;
188
189 if (sk_memalloc_socks() && (flags & SKB_ALLOC_RX))
190 gfp_mask |= __GFP_MEMALLOC;
191
192 /* Get the HEAD */
193 skb = kmem_cache_alloc_node(cache, gfp_mask & ~__GFP_DMA, node);
194 if (!skb)
195 goto out;
196 prefetchw(skb);
197
198 /* We do our best to align skb_shared_info on a separate cache
199 * line. It usually works because kmalloc(X > SMP_CACHE_BYTES) gives
200 * aligned memory blocks, unless SLUB/SLAB debug is enabled.
201 * Both skb->head and skb_shared_info are cache line aligned.
202 */
203 size = SKB_DATA_ALIGN(size);
204 size += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
205 data = kmalloc_reserve(size, gfp_mask, node, &pfmemalloc);
206 if (!data)
207 goto nodata;
208 /* kmalloc(size) might give us more room than requested.
209 * Put skb_shared_info exactly at the end of allocated zone,
210 * to allow max possible filling before reallocation.
211 */
212 size = SKB_WITH_OVERHEAD(ksize(data));
213 prefetchw(data + size);
214
215 /*
216 * Only clear those fields we need to clear, not those that we will
217 * actually initialise below. Hence, don't put any more fields after
218 * the tail pointer in struct sk_buff!
219 */
220 memset(skb, 0, offsetof(struct sk_buff, tail));
221 /* Account for allocated memory : skb + skb->head */
222 skb->truesize = SKB_TRUESIZE(size);
223 skb->pfmemalloc = pfmemalloc;
224 refcount_set(&skb->users, 1);
225 skb->head = data;
226 skb->data = data;
227 skb_reset_tail_pointer(skb);
228 skb->end = skb->tail + size;
229 skb->mac_header = (typeof(skb->mac_header))~0U;
230 skb->transport_header = (typeof(skb->transport_header))~0U;
231
232 /* make sure we initialize shinfo sequentially */
233 shinfo = skb_shinfo(skb);
234 memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
235 atomic_set(&shinfo->dataref, 1);
236
237 if (flags & SKB_ALLOC_FCLONE) {
238 struct sk_buff_fclones *fclones;
239
240 fclones = container_of(skb, struct sk_buff_fclones, skb1);
241
242 skb->fclone = SKB_FCLONE_ORIG;
243 refcount_set(&fclones->fclone_ref, 1);
244
245 fclones->skb2.fclone = SKB_FCLONE_CLONE;
246 }
247out:
248 return skb;
249nodata:
250 kmem_cache_free(cache, skb);
251 skb = NULL;
252 goto out;
253}
254EXPORT_SYMBOL(__alloc_skb);
255
256/**
257 * __build_skb - build a network buffer
258 * @data: data buffer provided by caller
259 * @frag_size: size of data, or 0 if head was kmalloced
260 *
261 * Allocate a new &sk_buff. Caller provides space holding head and
262 * skb_shared_info. @data must have been allocated by kmalloc() only if
263 * @frag_size is 0, otherwise data should come from the page allocator
264 * or vmalloc()
265 * The return is the new skb buffer.
266 * On a failure the return is %NULL, and @data is not freed.
267 * Notes :
268 * Before IO, driver allocates only data buffer where NIC put incoming frame
269 * Driver should add room at head (NET_SKB_PAD) and
270 * MUST add room at tail (SKB_DATA_ALIGN(skb_shared_info))
271 * After IO, driver calls build_skb(), to allocate sk_buff and populate it
272 * before giving packet to stack.
273 * RX rings only contains data buffers, not full skbs.
274 */
275struct sk_buff *__build_skb(void *data, unsigned int frag_size)
276{
277 struct skb_shared_info *shinfo;
278 struct sk_buff *skb;
279 unsigned int size = frag_size ? : ksize(data);
280
281 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
282 if (!skb)
283 return NULL;
284
285 size -= SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
286
287 memset(skb, 0, offsetof(struct sk_buff, tail));
288 skb->truesize = SKB_TRUESIZE(size);
289 refcount_set(&skb->users, 1);
290 skb->head = data;
291 skb->data = data;
292 skb_reset_tail_pointer(skb);
293 skb->end = skb->tail + size;
294 skb->mac_header = (typeof(skb->mac_header))~0U;
295 skb->transport_header = (typeof(skb->transport_header))~0U;
296
297 /* make sure we initialize shinfo sequentially */
298 shinfo = skb_shinfo(skb);
299 memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
300 atomic_set(&shinfo->dataref, 1);
301
302 return skb;
303}
304
305/* build_skb() is wrapper over __build_skb(), that specifically
306 * takes care of skb->head and skb->pfmemalloc
307 * This means that if @frag_size is not zero, then @data must be backed
308 * by a page fragment, not kmalloc() or vmalloc()
309 */
310struct sk_buff *build_skb(void *data, unsigned int frag_size)
311{
312 struct sk_buff *skb = __build_skb(data, frag_size);
313
314 if (skb && frag_size) {
315 skb->head_frag = 1;
316 if (page_is_pfmemalloc(virt_to_head_page(data)))
317 skb->pfmemalloc = 1;
318 }
319 return skb;
320}
321EXPORT_SYMBOL(build_skb);
322
323#define NAPI_SKB_CACHE_SIZE 64
324
325struct napi_alloc_cache {
326 struct page_frag_cache page;
327 unsigned int skb_count;
328 void *skb_cache[NAPI_SKB_CACHE_SIZE];
329};
330
331static DEFINE_PER_CPU(struct page_frag_cache, netdev_alloc_cache);
332static DEFINE_PER_CPU(struct napi_alloc_cache, napi_alloc_cache);
333
334static void *__netdev_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
335{
336 struct page_frag_cache *nc;
337 unsigned long flags;
338 void *data;
339
340 local_irq_save(flags);
341 nc = this_cpu_ptr(&netdev_alloc_cache);
342 data = page_frag_alloc(nc, fragsz, gfp_mask);
343 local_irq_restore(flags);
344 return data;
345}
346
347/**
348 * netdev_alloc_frag - allocate a page fragment
349 * @fragsz: fragment size
350 *
351 * Allocates a frag from a page for receive buffer.
352 * Uses GFP_ATOMIC allocations.
353 */
354void *netdev_alloc_frag(unsigned int fragsz)
355{
356 return __netdev_alloc_frag(fragsz, GFP_ATOMIC);
357}
358EXPORT_SYMBOL(netdev_alloc_frag);
359
360static void *__napi_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
361{
362 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
363
364 return page_frag_alloc(&nc->page, fragsz, gfp_mask);
365}
366
367void *napi_alloc_frag(unsigned int fragsz)
368{
369 return __napi_alloc_frag(fragsz, GFP_ATOMIC);
370}
371EXPORT_SYMBOL(napi_alloc_frag);
372
373/**
374 * __netdev_alloc_skb - allocate an skbuff for rx on a specific device
375 * @dev: network device to receive on
376 * @len: length to allocate
377 * @gfp_mask: get_free_pages mask, passed to alloc_skb
378 *
379 * Allocate a new &sk_buff and assign it a usage count of one. The
380 * buffer has NET_SKB_PAD headroom built in. Users should allocate
381 * the headroom they think they need without accounting for the
382 * built in space. The built in space is used for optimisations.
383 *
384 * %NULL is returned if there is no free memory.
385 */
386struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
387 gfp_t gfp_mask)
388{
389 struct page_frag_cache *nc;
390 unsigned long flags;
391 struct sk_buff *skb;
392 bool pfmemalloc;
393 void *data;
394
395 len += NET_SKB_PAD;
396
397 if ((len > SKB_WITH_OVERHEAD(PAGE_SIZE)) ||
398 (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
399 skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
400 if (!skb)
401 goto skb_fail;
402 goto skb_success;
403 }
404
405 len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
406 len = SKB_DATA_ALIGN(len);
407
408 if (sk_memalloc_socks())
409 gfp_mask |= __GFP_MEMALLOC;
410
411 local_irq_save(flags);
412
413 nc = this_cpu_ptr(&netdev_alloc_cache);
414 data = page_frag_alloc(nc, len, gfp_mask);
415 pfmemalloc = nc->pfmemalloc;
416
417 local_irq_restore(flags);
418
419 if (unlikely(!data))
420 return NULL;
421
422 skb = __build_skb(data, len);
423 if (unlikely(!skb)) {
424 skb_free_frag(data);
425 return NULL;
426 }
427
428 /* use OR instead of assignment to avoid clearing of bits in mask */
429 if (pfmemalloc)
430 skb->pfmemalloc = 1;
431 skb->head_frag = 1;
432
433skb_success:
434 skb_reserve(skb, NET_SKB_PAD);
435 skb->dev = dev;
436
437skb_fail:
438 return skb;
439}
440EXPORT_SYMBOL(__netdev_alloc_skb);
441
442/**
443 * __napi_alloc_skb - allocate skbuff for rx in a specific NAPI instance
444 * @napi: napi instance this buffer was allocated for
445 * @len: length to allocate
446 * @gfp_mask: get_free_pages mask, passed to alloc_skb and alloc_pages
447 *
448 * Allocate a new sk_buff for use in NAPI receive. This buffer will
449 * attempt to allocate the head from a special reserved region used
450 * only for NAPI Rx allocation. By doing this we can save several
451 * CPU cycles by avoiding having to disable and re-enable IRQs.
452 *
453 * %NULL is returned if there is no free memory.
454 */
455struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
456 gfp_t gfp_mask)
457{
458 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
459 struct sk_buff *skb;
460 void *data;
461
462 len += NET_SKB_PAD + NET_IP_ALIGN;
463
464 if ((len > SKB_WITH_OVERHEAD(PAGE_SIZE)) ||
465 (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
466 skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
467 if (!skb)
468 goto skb_fail;
469 goto skb_success;
470 }
471
472 len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
473 len = SKB_DATA_ALIGN(len);
474
475 if (sk_memalloc_socks())
476 gfp_mask |= __GFP_MEMALLOC;
477
478 data = page_frag_alloc(&nc->page, len, gfp_mask);
479 if (unlikely(!data))
480 return NULL;
481
482 skb = __build_skb(data, len);
483 if (unlikely(!skb)) {
484 skb_free_frag(data);
485 return NULL;
486 }
487
488 /* use OR instead of assignment to avoid clearing of bits in mask */
489 if (nc->page.pfmemalloc)
490 skb->pfmemalloc = 1;
491 skb->head_frag = 1;
492
493skb_success:
494 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
495 skb->dev = napi->dev;
496
497skb_fail:
498 return skb;
499}
500EXPORT_SYMBOL(__napi_alloc_skb);
501
502void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
503 int size, unsigned int truesize)
504{
505 skb_fill_page_desc(skb, i, page, off, size);
506 skb->len += size;
507 skb->data_len += size;
508 skb->truesize += truesize;
509}
510EXPORT_SYMBOL(skb_add_rx_frag);
511
512void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
513 unsigned int truesize)
514{
515 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
516
517 skb_frag_size_add(frag, size);
518 skb->len += size;
519 skb->data_len += size;
520 skb->truesize += truesize;
521}
522EXPORT_SYMBOL(skb_coalesce_rx_frag);
523
524static void skb_drop_list(struct sk_buff **listp)
525{
526 kfree_skb_list(*listp);
527 *listp = NULL;
528}
529
530static inline void skb_drop_fraglist(struct sk_buff *skb)
531{
532 skb_drop_list(&skb_shinfo(skb)->frag_list);
533}
534
535static void skb_clone_fraglist(struct sk_buff *skb)
536{
537 struct sk_buff *list;
538
539 skb_walk_frags(skb, list)
540 skb_get(list);
541}
542
543static void skb_free_head(struct sk_buff *skb)
544{
545 unsigned char *head = skb->head;
546
547 if (skb->head_frag)
548 skb_free_frag(head);
549 else
550 kfree(head);
551}
552
553static void skb_release_data(struct sk_buff *skb)
554{
555 struct skb_shared_info *shinfo = skb_shinfo(skb);
556 int i;
557
558 if (skb->cloned &&
559 atomic_sub_return(skb->nohdr ? (1 << SKB_DATAREF_SHIFT) + 1 : 1,
560 &shinfo->dataref))
561 return;
562
563 for (i = 0; i < shinfo->nr_frags; i++)
564 __skb_frag_unref(&shinfo->frags[i]);
565
566 if (shinfo->frag_list)
567 kfree_skb_list(shinfo->frag_list);
568
569 skb_zcopy_clear(skb, true);
570 skb_free_head(skb);
571}
572
573/*
574 * Free an skbuff by memory without cleaning the state.
575 */
576static void kfree_skbmem(struct sk_buff *skb)
577{
578 struct sk_buff_fclones *fclones;
579
580 switch (skb->fclone) {
581 case SKB_FCLONE_UNAVAILABLE:
582 kmem_cache_free(skbuff_head_cache, skb);
583 return;
584
585 case SKB_FCLONE_ORIG:
586 fclones = container_of(skb, struct sk_buff_fclones, skb1);
587
588 /* We usually free the clone (TX completion) before original skb
589 * This test would have no chance to be true for the clone,
590 * while here, branch prediction will be good.
591 */
592 if (refcount_read(&fclones->fclone_ref) == 1)
593 goto fastpath;
594 break;
595
596 default: /* SKB_FCLONE_CLONE */
597 fclones = container_of(skb, struct sk_buff_fclones, skb2);
598 break;
599 }
600 if (!refcount_dec_and_test(&fclones->fclone_ref))
601 return;
602fastpath:
603 kmem_cache_free(skbuff_fclone_cache, fclones);
604}
605
606void skb_release_head_state(struct sk_buff *skb)
607{
608 skb_dst_drop(skb);
609 secpath_reset(skb);
610 if (skb->destructor) {
611 WARN_ON(in_irq());
612 skb->destructor(skb);
613 }
614#if IS_ENABLED(CONFIG_NF_CONNTRACK)
615 nf_conntrack_put(skb_nfct(skb));
616#endif
617#if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
618 nf_bridge_put(skb->nf_bridge);
619#endif
620}
621
622/* Free everything but the sk_buff shell. */
623static void skb_release_all(struct sk_buff *skb)
624{
625 skb_release_head_state(skb);
626 if (likely(skb->head))
627 skb_release_data(skb);
628}
629
630/**
631 * __kfree_skb - private function
632 * @skb: buffer
633 *
634 * Free an sk_buff. Release anything attached to the buffer.
635 * Clean the state. This is an internal helper function. Users should
636 * always call kfree_skb
637 */
638
639void __kfree_skb(struct sk_buff *skb)
640{
641 skb_release_all(skb);
642 kfree_skbmem(skb);
643}
644EXPORT_SYMBOL(__kfree_skb);
645
646/**
647 * kfree_skb - free an sk_buff
648 * @skb: buffer to free
649 *
650 * Drop a reference to the buffer and free it if the usage count has
651 * hit zero.
652 */
653void kfree_skb(struct sk_buff *skb)
654{
655 if (!skb_unref(skb))
656 return;
657
658 trace_kfree_skb(skb, __builtin_return_address(0));
659 __kfree_skb(skb);
660}
661EXPORT_SYMBOL(kfree_skb);
662
663void kfree_skb_list(struct sk_buff *segs)
664{
665 while (segs) {
666 struct sk_buff *next = segs->next;
667
668 kfree_skb(segs);
669 segs = next;
670 }
671}
672EXPORT_SYMBOL(kfree_skb_list);
673
674/**
675 * skb_tx_error - report an sk_buff xmit error
676 * @skb: buffer that triggered an error
677 *
678 * Report xmit error if a device callback is tracking this skb.
679 * skb must be freed afterwards.
680 */
681void skb_tx_error(struct sk_buff *skb)
682{
683 skb_zcopy_clear(skb, true);
684}
685EXPORT_SYMBOL(skb_tx_error);
686
687/**
688 * consume_skb - free an skbuff
689 * @skb: buffer to free
690 *
691 * Drop a ref to the buffer and free it if the usage count has hit zero
692 * Functions identically to kfree_skb, but kfree_skb assumes that the frame
693 * is being dropped after a failure and notes that
694 */
695void consume_skb(struct sk_buff *skb)
696{
697 if (!skb_unref(skb))
698 return;
699
700 trace_consume_skb(skb);
701 __kfree_skb(skb);
702}
703EXPORT_SYMBOL(consume_skb);
704
705/**
706 * consume_stateless_skb - free an skbuff, assuming it is stateless
707 * @skb: buffer to free
708 *
709 * Alike consume_skb(), but this variant assumes that this is the last
710 * skb reference and all the head states have been already dropped
711 */
712void __consume_stateless_skb(struct sk_buff *skb)
713{
714 trace_consume_skb(skb);
715 skb_release_data(skb);
716 kfree_skbmem(skb);
717}
718
719void __kfree_skb_flush(void)
720{
721 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
722
723 /* flush skb_cache if containing objects */
724 if (nc->skb_count) {
725 kmem_cache_free_bulk(skbuff_head_cache, nc->skb_count,
726 nc->skb_cache);
727 nc->skb_count = 0;
728 }
729}
730
731static inline void _kfree_skb_defer(struct sk_buff *skb)
732{
733 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
734
735 /* drop skb->head and call any destructors for packet */
736 skb_release_all(skb);
737
738 /* record skb to CPU local list */
739 nc->skb_cache[nc->skb_count++] = skb;
740
741#ifdef CONFIG_SLUB
742 /* SLUB writes into objects when freeing */
743 prefetchw(skb);
744#endif
745
746 /* flush skb_cache if it is filled */
747 if (unlikely(nc->skb_count == NAPI_SKB_CACHE_SIZE)) {
748 kmem_cache_free_bulk(skbuff_head_cache, NAPI_SKB_CACHE_SIZE,
749 nc->skb_cache);
750 nc->skb_count = 0;
751 }
752}
753void __kfree_skb_defer(struct sk_buff *skb)
754{
755 _kfree_skb_defer(skb);
756}
757
758void napi_consume_skb(struct sk_buff *skb, int budget)
759{
760 if (unlikely(!skb))
761 return;
762
763 /* Zero budget indicate non-NAPI context called us, like netpoll */
764 if (unlikely(!budget)) {
765 dev_consume_skb_any(skb);
766 return;
767 }
768
769 if (!skb_unref(skb))
770 return;
771
772 /* if reaching here SKB is ready to free */
773 trace_consume_skb(skb);
774
775 /* if SKB is a clone, don't handle this case */
776 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) {
777 __kfree_skb(skb);
778 return;
779 }
780
781 _kfree_skb_defer(skb);
782}
783EXPORT_SYMBOL(napi_consume_skb);
784
785/* Make sure a field is enclosed inside headers_start/headers_end section */
786#define CHECK_SKB_FIELD(field) \
787 BUILD_BUG_ON(offsetof(struct sk_buff, field) < \
788 offsetof(struct sk_buff, headers_start)); \
789 BUILD_BUG_ON(offsetof(struct sk_buff, field) > \
790 offsetof(struct sk_buff, headers_end)); \
791
792static void __copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
793{
794 new->tstamp = old->tstamp;
795 /* We do not copy old->sk */
796 new->dev = old->dev;
797 memcpy(new->cb, old->cb, sizeof(old->cb));
798 skb_dst_copy(new, old);
799#ifdef CONFIG_XFRM
800 new->sp = secpath_get(old->sp);
801#endif
802 __nf_copy(new, old, false);
803
804 /* Note : this field could be in headers_start/headers_end section
805 * It is not yet because we do not want to have a 16 bit hole
806 */
807 new->queue_mapping = old->queue_mapping;
808
809 memcpy(&new->headers_start, &old->headers_start,
810 offsetof(struct sk_buff, headers_end) -
811 offsetof(struct sk_buff, headers_start));
812 CHECK_SKB_FIELD(protocol);
813 CHECK_SKB_FIELD(csum);
814 CHECK_SKB_FIELD(hash);
815 CHECK_SKB_FIELD(priority);
816 CHECK_SKB_FIELD(skb_iif);
817 CHECK_SKB_FIELD(vlan_proto);
818 CHECK_SKB_FIELD(vlan_tci);
819 CHECK_SKB_FIELD(transport_header);
820 CHECK_SKB_FIELD(network_header);
821 CHECK_SKB_FIELD(mac_header);
822 CHECK_SKB_FIELD(inner_protocol);
823 CHECK_SKB_FIELD(inner_transport_header);
824 CHECK_SKB_FIELD(inner_network_header);
825 CHECK_SKB_FIELD(inner_mac_header);
826 CHECK_SKB_FIELD(mark);
827#ifdef CONFIG_NETWORK_SECMARK
828 CHECK_SKB_FIELD(secmark);
829#endif
830#ifdef CONFIG_NET_RX_BUSY_POLL
831 CHECK_SKB_FIELD(napi_id);
832#endif
833#ifdef CONFIG_XPS
834 CHECK_SKB_FIELD(sender_cpu);
835#endif
836#ifdef CONFIG_NET_SCHED
837 CHECK_SKB_FIELD(tc_index);
838#endif
839
840}
841
842/*
843 * You should not add any new code to this function. Add it to
844 * __copy_skb_header above instead.
845 */
846static struct sk_buff *__skb_clone(struct sk_buff *n, struct sk_buff *skb)
847{
848#define C(x) n->x = skb->x
849
850 n->next = n->prev = NULL;
851 n->sk = NULL;
852 __copy_skb_header(n, skb);
853
854 C(len);
855 C(data_len);
856 C(mac_len);
857 n->hdr_len = skb->nohdr ? skb_headroom(skb) : skb->hdr_len;
858 n->cloned = 1;
859 n->nohdr = 0;
860 n->peeked = 0;
861 n->destructor = NULL;
862 C(tail);
863 C(end);
864 C(head);
865 C(head_frag);
866 C(data);
867 C(truesize);
868 refcount_set(&n->users, 1);
869
870 atomic_inc(&(skb_shinfo(skb)->dataref));
871 skb->cloned = 1;
872
873 return n;
874#undef C
875}
876
877/**
878 * skb_morph - morph one skb into another
879 * @dst: the skb to receive the contents
880 * @src: the skb to supply the contents
881 *
882 * This is identical to skb_clone except that the target skb is
883 * supplied by the user.
884 *
885 * The target skb is returned upon exit.
886 */
887struct sk_buff *skb_morph(struct sk_buff *dst, struct sk_buff *src)
888{
889 skb_release_all(dst);
890 return __skb_clone(dst, src);
891}
892EXPORT_SYMBOL_GPL(skb_morph);
893
894int mm_account_pinned_pages(struct mmpin *mmp, size_t size)
895{
896 unsigned long max_pg, num_pg, new_pg, old_pg;
897 struct user_struct *user;
898
899 if (capable(CAP_IPC_LOCK) || !size)
900 return 0;
901
902 num_pg = (size >> PAGE_SHIFT) + 2; /* worst case */
903 max_pg = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
904 user = mmp->user ? : current_user();
905
906 do {
907 old_pg = atomic_long_read(&user->locked_vm);
908 new_pg = old_pg + num_pg;
909 if (new_pg > max_pg)
910 return -ENOBUFS;
911 } while (atomic_long_cmpxchg(&user->locked_vm, old_pg, new_pg) !=
912 old_pg);
913
914 if (!mmp->user) {
915 mmp->user = get_uid(user);
916 mmp->num_pg = num_pg;
917 } else {
918 mmp->num_pg += num_pg;
919 }
920
921 return 0;
922}
923EXPORT_SYMBOL_GPL(mm_account_pinned_pages);
924
925void mm_unaccount_pinned_pages(struct mmpin *mmp)
926{
927 if (mmp->user) {
928 atomic_long_sub(mmp->num_pg, &mmp->user->locked_vm);
929 free_uid(mmp->user);
930 }
931}
932EXPORT_SYMBOL_GPL(mm_unaccount_pinned_pages);
933
934struct ubuf_info *sock_zerocopy_alloc(struct sock *sk, size_t size)
935{
936 struct ubuf_info *uarg;
937 struct sk_buff *skb;
938
939 WARN_ON_ONCE(!in_task());
940
941 if (!sock_flag(sk, SOCK_ZEROCOPY))
942 return NULL;
943
944 skb = sock_omalloc(sk, 0, GFP_KERNEL);
945 if (!skb)
946 return NULL;
947
948 BUILD_BUG_ON(sizeof(*uarg) > sizeof(skb->cb));
949 uarg = (void *)skb->cb;
950 uarg->mmp.user = NULL;
951
952 if (mm_account_pinned_pages(&uarg->mmp, size)) {
953 kfree_skb(skb);
954 return NULL;
955 }
956
957 uarg->callback = sock_zerocopy_callback;
958 uarg->id = ((u32)atomic_inc_return(&sk->sk_zckey)) - 1;
959 uarg->len = 1;
960 uarg->bytelen = size;
961 uarg->zerocopy = 1;
962 refcount_set(&uarg->refcnt, 1);
963 sock_hold(sk);
964
965 return uarg;
966}
967EXPORT_SYMBOL_GPL(sock_zerocopy_alloc);
968
969static inline struct sk_buff *skb_from_uarg(struct ubuf_info *uarg)
970{
971 return container_of((void *)uarg, struct sk_buff, cb);
972}
973
974struct ubuf_info *sock_zerocopy_realloc(struct sock *sk, size_t size,
975 struct ubuf_info *uarg)
976{
977 if (uarg) {
978 const u32 byte_limit = 1 << 19; /* limit to a few TSO */
979 u32 bytelen, next;
980
981 /* realloc only when socket is locked (TCP, UDP cork),
982 * so uarg->len and sk_zckey access is serialized
983 */
984 if (!sock_owned_by_user(sk)) {
985 WARN_ON_ONCE(1);
986 return NULL;
987 }
988
989 bytelen = uarg->bytelen + size;
990 if (uarg->len == USHRT_MAX - 1 || bytelen > byte_limit) {
991 /* TCP can create new skb to attach new uarg */
992 if (sk->sk_type == SOCK_STREAM)
993 goto new_alloc;
994 return NULL;
995 }
996
997 next = (u32)atomic_read(&sk->sk_zckey);
998 if ((u32)(uarg->id + uarg->len) == next) {
999 if (mm_account_pinned_pages(&uarg->mmp, size))
1000 return NULL;
1001 uarg->len++;
1002 uarg->bytelen = bytelen;
1003 atomic_set(&sk->sk_zckey, ++next);
1004 sock_zerocopy_get(uarg);
1005 return uarg;
1006 }
1007 }
1008
1009new_alloc:
1010 return sock_zerocopy_alloc(sk, size);
1011}
1012EXPORT_SYMBOL_GPL(sock_zerocopy_realloc);
1013
1014static bool skb_zerocopy_notify_extend(struct sk_buff *skb, u32 lo, u16 len)
1015{
1016 struct sock_exterr_skb *serr = SKB_EXT_ERR(skb);
1017 u32 old_lo, old_hi;
1018 u64 sum_len;
1019
1020 old_lo = serr->ee.ee_info;
1021 old_hi = serr->ee.ee_data;
1022 sum_len = old_hi - old_lo + 1ULL + len;
1023
1024 if (sum_len >= (1ULL << 32))
1025 return false;
1026
1027 if (lo != old_hi + 1)
1028 return false;
1029
1030 serr->ee.ee_data += len;
1031 return true;
1032}
1033
1034void sock_zerocopy_callback(struct ubuf_info *uarg, bool success)
1035{
1036 struct sk_buff *tail, *skb = skb_from_uarg(uarg);
1037 struct sock_exterr_skb *serr;
1038 struct sock *sk = skb->sk;
1039 struct sk_buff_head *q;
1040 unsigned long flags;
1041 u32 lo, hi;
1042 u16 len;
1043
1044 mm_unaccount_pinned_pages(&uarg->mmp);
1045
1046 /* if !len, there was only 1 call, and it was aborted
1047 * so do not queue a completion notification
1048 */
1049 if (!uarg->len || sock_flag(sk, SOCK_DEAD))
1050 goto release;
1051
1052 len = uarg->len;
1053 lo = uarg->id;
1054 hi = uarg->id + len - 1;
1055
1056 serr = SKB_EXT_ERR(skb);
1057 memset(serr, 0, sizeof(*serr));
1058 serr->ee.ee_errno = 0;
1059 serr->ee.ee_origin = SO_EE_ORIGIN_ZEROCOPY;
1060 serr->ee.ee_data = hi;
1061 serr->ee.ee_info = lo;
1062 if (!success)
1063 serr->ee.ee_code |= SO_EE_CODE_ZEROCOPY_COPIED;
1064
1065 q = &sk->sk_error_queue;
1066 spin_lock_irqsave(&q->lock, flags);
1067 tail = skb_peek_tail(q);
1068 if (!tail || SKB_EXT_ERR(tail)->ee.ee_origin != SO_EE_ORIGIN_ZEROCOPY ||
1069 !skb_zerocopy_notify_extend(tail, lo, len)) {
1070 __skb_queue_tail(q, skb);
1071 skb = NULL;
1072 }
1073 spin_unlock_irqrestore(&q->lock, flags);
1074
1075 sk->sk_error_report(sk);
1076
1077release:
1078 consume_skb(skb);
1079 sock_put(sk);
1080}
1081EXPORT_SYMBOL_GPL(sock_zerocopy_callback);
1082
1083void sock_zerocopy_put(struct ubuf_info *uarg)
1084{
1085 if (uarg && refcount_dec_and_test(&uarg->refcnt)) {
1086 if (uarg->callback)
1087 uarg->callback(uarg, uarg->zerocopy);
1088 else
1089 consume_skb(skb_from_uarg(uarg));
1090 }
1091}
1092EXPORT_SYMBOL_GPL(sock_zerocopy_put);
1093
1094void sock_zerocopy_put_abort(struct ubuf_info *uarg)
1095{
1096 if (uarg) {
1097 struct sock *sk = skb_from_uarg(uarg)->sk;
1098
1099 atomic_dec(&sk->sk_zckey);
1100 uarg->len--;
1101
1102 sock_zerocopy_put(uarg);
1103 }
1104}
1105EXPORT_SYMBOL_GPL(sock_zerocopy_put_abort);
1106
1107extern int __zerocopy_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1108 struct iov_iter *from, size_t length);
1109
1110int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
1111 struct msghdr *msg, int len,
1112 struct ubuf_info *uarg)
1113{
1114 struct ubuf_info *orig_uarg = skb_zcopy(skb);
1115 struct iov_iter orig_iter = msg->msg_iter;
1116 int err, orig_len = skb->len;
1117
1118 /* An skb can only point to one uarg. This edge case happens when
1119 * TCP appends to an skb, but zerocopy_realloc triggered a new alloc.
1120 */
1121 if (orig_uarg && uarg != orig_uarg)
1122 return -EEXIST;
1123
1124 err = __zerocopy_sg_from_iter(sk, skb, &msg->msg_iter, len);
1125 if (err == -EFAULT || (err == -EMSGSIZE && skb->len == orig_len)) {
1126 struct sock *save_sk = skb->sk;
1127
1128 /* Streams do not free skb on error. Reset to prev state. */
1129 msg->msg_iter = orig_iter;
1130 skb->sk = sk;
1131 ___pskb_trim(skb, orig_len);
1132 skb->sk = save_sk;
1133 return err;
1134 }
1135
1136 skb_zcopy_set(skb, uarg);
1137 return skb->len - orig_len;
1138}
1139EXPORT_SYMBOL_GPL(skb_zerocopy_iter_stream);
1140
1141static int skb_zerocopy_clone(struct sk_buff *nskb, struct sk_buff *orig,
1142 gfp_t gfp_mask)
1143{
1144 if (skb_zcopy(orig)) {
1145 if (skb_zcopy(nskb)) {
1146 /* !gfp_mask callers are verified to !skb_zcopy(nskb) */
1147 if (!gfp_mask) {
1148 WARN_ON_ONCE(1);
1149 return -ENOMEM;
1150 }
1151 if (skb_uarg(nskb) == skb_uarg(orig))
1152 return 0;
1153 if (skb_copy_ubufs(nskb, GFP_ATOMIC))
1154 return -EIO;
1155 }
1156 skb_zcopy_set(nskb, skb_uarg(orig));
1157 }
1158 return 0;
1159}
1160
1161/**
1162 * skb_copy_ubufs - copy userspace skb frags buffers to kernel
1163 * @skb: the skb to modify
1164 * @gfp_mask: allocation priority
1165 *
1166 * This must be called on SKBTX_DEV_ZEROCOPY skb.
1167 * It will copy all frags into kernel and drop the reference
1168 * to userspace pages.
1169 *
1170 * If this function is called from an interrupt gfp_mask() must be
1171 * %GFP_ATOMIC.
1172 *
1173 * Returns 0 on success or a negative error code on failure
1174 * to allocate kernel memory to copy to.
1175 */
1176int skb_copy_ubufs(struct sk_buff *skb, gfp_t gfp_mask)
1177{
1178 int num_frags = skb_shinfo(skb)->nr_frags;
1179 struct page *page, *head = NULL;
1180 int i, new_frags;
1181 u32 d_off;
1182
1183 if (skb_shared(skb) || skb_unclone(skb, gfp_mask))
1184 return -EINVAL;
1185
1186 if (!num_frags)
1187 goto release;
1188
1189 new_frags = (__skb_pagelen(skb) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1190 for (i = 0; i < new_frags; i++) {
1191 page = alloc_page(gfp_mask);
1192 if (!page) {
1193 while (head) {
1194 struct page *next = (struct page *)page_private(head);
1195 put_page(head);
1196 head = next;
1197 }
1198 return -ENOMEM;
1199 }
1200 set_page_private(page, (unsigned long)head);
1201 head = page;
1202 }
1203
1204 page = head;
1205 d_off = 0;
1206 for (i = 0; i < num_frags; i++) {
1207 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
1208 u32 p_off, p_len, copied;
1209 struct page *p;
1210 u8 *vaddr;
1211
1212 skb_frag_foreach_page(f, f->page_offset, skb_frag_size(f),
1213 p, p_off, p_len, copied) {
1214 u32 copy, done = 0;
1215 vaddr = kmap_atomic(p);
1216
1217 while (done < p_len) {
1218 if (d_off == PAGE_SIZE) {
1219 d_off = 0;
1220 page = (struct page *)page_private(page);
1221 }
1222 copy = min_t(u32, PAGE_SIZE - d_off, p_len - done);
1223 memcpy(page_address(page) + d_off,
1224 vaddr + p_off + done, copy);
1225 done += copy;
1226 d_off += copy;
1227 }
1228 kunmap_atomic(vaddr);
1229 }
1230 }
1231
1232 /* skb frags release userspace buffers */
1233 for (i = 0; i < num_frags; i++)
1234 skb_frag_unref(skb, i);
1235
1236 /* skb frags point to kernel buffers */
1237 for (i = 0; i < new_frags - 1; i++) {
1238 __skb_fill_page_desc(skb, i, head, 0, PAGE_SIZE);
1239 head = (struct page *)page_private(head);
1240 }
1241 __skb_fill_page_desc(skb, new_frags - 1, head, 0, d_off);
1242 skb_shinfo(skb)->nr_frags = new_frags;
1243
1244release:
1245 skb_zcopy_clear(skb, false);
1246 return 0;
1247}
1248EXPORT_SYMBOL_GPL(skb_copy_ubufs);
1249
1250/**
1251 * skb_clone - duplicate an sk_buff
1252 * @skb: buffer to clone
1253 * @gfp_mask: allocation priority
1254 *
1255 * Duplicate an &sk_buff. The new one is not owned by a socket. Both
1256 * copies share the same packet data but not structure. The new
1257 * buffer has a reference count of 1. If the allocation fails the
1258 * function returns %NULL otherwise the new buffer is returned.
1259 *
1260 * If this function is called from an interrupt gfp_mask() must be
1261 * %GFP_ATOMIC.
1262 */
1263
1264struct sk_buff *skb_clone(struct sk_buff *skb, gfp_t gfp_mask)
1265{
1266 struct sk_buff_fclones *fclones = container_of(skb,
1267 struct sk_buff_fclones,
1268 skb1);
1269 struct sk_buff *n;
1270
1271 if (skb_orphan_frags(skb, gfp_mask))
1272 return NULL;
1273
1274 if (skb->fclone == SKB_FCLONE_ORIG &&
1275 refcount_read(&fclones->fclone_ref) == 1) {
1276 n = &fclones->skb2;
1277 refcount_set(&fclones->fclone_ref, 2);
1278 } else {
1279 if (skb_pfmemalloc(skb))
1280 gfp_mask |= __GFP_MEMALLOC;
1281
1282 n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
1283 if (!n)
1284 return NULL;
1285
1286 n->fclone = SKB_FCLONE_UNAVAILABLE;
1287 }
1288
1289 return __skb_clone(n, skb);
1290}
1291EXPORT_SYMBOL(skb_clone);
1292
1293static void skb_headers_offset_update(struct sk_buff *skb, int off)
1294{
1295 /* Only adjust this if it actually is csum_start rather than csum */
1296 if (skb->ip_summed == CHECKSUM_PARTIAL)
1297 skb->csum_start += off;
1298 /* {transport,network,mac}_header and tail are relative to skb->head */
1299 skb->transport_header += off;
1300 skb->network_header += off;
1301 if (skb_mac_header_was_set(skb))
1302 skb->mac_header += off;
1303 skb->inner_transport_header += off;
1304 skb->inner_network_header += off;
1305 skb->inner_mac_header += off;
1306}
1307
1308static void copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
1309{
1310 __copy_skb_header(new, old);
1311
1312 skb_shinfo(new)->gso_size = skb_shinfo(old)->gso_size;
1313 skb_shinfo(new)->gso_segs = skb_shinfo(old)->gso_segs;
1314 skb_shinfo(new)->gso_type = skb_shinfo(old)->gso_type;
1315}
1316
1317static inline int skb_alloc_rx_flag(const struct sk_buff *skb)
1318{
1319 if (skb_pfmemalloc(skb))
1320 return SKB_ALLOC_RX;
1321 return 0;
1322}
1323
1324/**
1325 * skb_copy - create private copy of an sk_buff
1326 * @skb: buffer to copy
1327 * @gfp_mask: allocation priority
1328 *
1329 * Make a copy of both an &sk_buff and its data. This is used when the
1330 * caller wishes to modify the data and needs a private copy of the
1331 * data to alter. Returns %NULL on failure or the pointer to the buffer
1332 * on success. The returned buffer has a reference count of 1.
1333 *
1334 * As by-product this function converts non-linear &sk_buff to linear
1335 * one, so that &sk_buff becomes completely private and caller is allowed
1336 * to modify all the data of returned buffer. This means that this
1337 * function is not recommended for use in circumstances when only
1338 * header is going to be modified. Use pskb_copy() instead.
1339 */
1340
1341struct sk_buff *skb_copy(const struct sk_buff *skb, gfp_t gfp_mask)
1342{
1343 int headerlen = skb_headroom(skb);
1344 unsigned int size = skb_end_offset(skb) + skb->data_len;
1345 struct sk_buff *n = __alloc_skb(size, gfp_mask,
1346 skb_alloc_rx_flag(skb), NUMA_NO_NODE);
1347
1348 if (!n)
1349 return NULL;
1350
1351 /* Set the data pointer */
1352 skb_reserve(n, headerlen);
1353 /* Set the tail pointer and length */
1354 skb_put(n, skb->len);
1355
1356 BUG_ON(skb_copy_bits(skb, -headerlen, n->head, headerlen + skb->len));
1357
1358 copy_skb_header(n, skb);
1359 return n;
1360}
1361EXPORT_SYMBOL(skb_copy);
1362
1363/**
1364 * __pskb_copy_fclone - create copy of an sk_buff with private head.
1365 * @skb: buffer to copy
1366 * @headroom: headroom of new skb
1367 * @gfp_mask: allocation priority
1368 * @fclone: if true allocate the copy of the skb from the fclone
1369 * cache instead of the head cache; it is recommended to set this
1370 * to true for the cases where the copy will likely be cloned
1371 *
1372 * Make a copy of both an &sk_buff and part of its data, located
1373 * in header. Fragmented data remain shared. This is used when
1374 * the caller wishes to modify only header of &sk_buff and needs
1375 * private copy of the header to alter. Returns %NULL on failure
1376 * or the pointer to the buffer on success.
1377 * The returned buffer has a reference count of 1.
1378 */
1379
1380struct sk_buff *__pskb_copy_fclone(struct sk_buff *skb, int headroom,
1381 gfp_t gfp_mask, bool fclone)
1382{
1383 unsigned int size = skb_headlen(skb) + headroom;
1384 int flags = skb_alloc_rx_flag(skb) | (fclone ? SKB_ALLOC_FCLONE : 0);
1385 struct sk_buff *n = __alloc_skb(size, gfp_mask, flags, NUMA_NO_NODE);
1386
1387 if (!n)
1388 goto out;
1389
1390 /* Set the data pointer */
1391 skb_reserve(n, headroom);
1392 /* Set the tail pointer and length */
1393 skb_put(n, skb_headlen(skb));
1394 /* Copy the bytes */
1395 skb_copy_from_linear_data(skb, n->data, n->len);
1396
1397 n->truesize += skb->data_len;
1398 n->data_len = skb->data_len;
1399 n->len = skb->len;
1400
1401 if (skb_shinfo(skb)->nr_frags) {
1402 int i;
1403
1404 if (skb_orphan_frags(skb, gfp_mask) ||
1405 skb_zerocopy_clone(n, skb, gfp_mask)) {
1406 kfree_skb(n);
1407 n = NULL;
1408 goto out;
1409 }
1410 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1411 skb_shinfo(n)->frags[i] = skb_shinfo(skb)->frags[i];
1412 skb_frag_ref(skb, i);
1413 }
1414 skb_shinfo(n)->nr_frags = i;
1415 }
1416
1417 if (skb_has_frag_list(skb)) {
1418 skb_shinfo(n)->frag_list = skb_shinfo(skb)->frag_list;
1419 skb_clone_fraglist(n);
1420 }
1421
1422 copy_skb_header(n, skb);
1423out:
1424 return n;
1425}
1426EXPORT_SYMBOL(__pskb_copy_fclone);
1427
1428/**
1429 * pskb_expand_head - reallocate header of &sk_buff
1430 * @skb: buffer to reallocate
1431 * @nhead: room to add at head
1432 * @ntail: room to add at tail
1433 * @gfp_mask: allocation priority
1434 *
1435 * Expands (or creates identical copy, if @nhead and @ntail are zero)
1436 * header of @skb. &sk_buff itself is not changed. &sk_buff MUST have
1437 * reference count of 1. Returns zero in the case of success or error,
1438 * if expansion failed. In the last case, &sk_buff is not changed.
1439 *
1440 * All the pointers pointing into skb header may change and must be
1441 * reloaded after call to this function.
1442 */
1443
1444int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
1445 gfp_t gfp_mask)
1446{
1447 int i, osize = skb_end_offset(skb);
1448 int size = osize + nhead + ntail;
1449 long off;
1450 u8 *data;
1451
1452 BUG_ON(nhead < 0);
1453
1454 BUG_ON(skb_shared(skb));
1455
1456 size = SKB_DATA_ALIGN(size);
1457
1458 if (skb_pfmemalloc(skb))
1459 gfp_mask |= __GFP_MEMALLOC;
1460 data = kmalloc_reserve(size + SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
1461 gfp_mask, NUMA_NO_NODE, NULL);
1462 if (!data)
1463 goto nodata;
1464 size = SKB_WITH_OVERHEAD(ksize(data));
1465
1466 /* Copy only real data... and, alas, header. This should be
1467 * optimized for the cases when header is void.
1468 */
1469 memcpy(data + nhead, skb->head, skb_tail_pointer(skb) - skb->head);
1470
1471 memcpy((struct skb_shared_info *)(data + size),
1472 skb_shinfo(skb),
1473 offsetof(struct skb_shared_info, frags[skb_shinfo(skb)->nr_frags]));
1474
1475 /*
1476 * if shinfo is shared we must drop the old head gracefully, but if it
1477 * is not we can just drop the old head and let the existing refcount
1478 * be since all we did is relocate the values
1479 */
1480 if (skb_cloned(skb)) {
1481 if (skb_orphan_frags(skb, gfp_mask))
1482 goto nofrags;
1483 if (skb_zcopy(skb))
1484 refcount_inc(&skb_uarg(skb)->refcnt);
1485 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1486 skb_frag_ref(skb, i);
1487
1488 if (skb_has_frag_list(skb))
1489 skb_clone_fraglist(skb);
1490
1491 skb_release_data(skb);
1492 } else {
1493 skb_free_head(skb);
1494 }
1495 off = (data + nhead) - skb->head;
1496
1497 skb->head = data;
1498 skb->head_frag = 0;
1499 skb->data += off;
1500#ifdef NET_SKBUFF_DATA_USES_OFFSET
1501 skb->end = size;
1502 off = nhead;
1503#else
1504 skb->end = skb->head + size;
1505#endif
1506 skb->tail += off;
1507 skb_headers_offset_update(skb, nhead);
1508 skb->cloned = 0;
1509 skb->hdr_len = 0;
1510 skb->nohdr = 0;
1511 atomic_set(&skb_shinfo(skb)->dataref, 1);
1512
1513 skb_metadata_clear(skb);
1514
1515 /* It is not generally safe to change skb->truesize.
1516 * For the moment, we really care of rx path, or
1517 * when skb is orphaned (not attached to a socket).
1518 */
1519 if (!skb->sk || skb->destructor == sock_edemux)
1520 skb->truesize += size - osize;
1521
1522 return 0;
1523
1524nofrags:
1525 kfree(data);
1526nodata:
1527 return -ENOMEM;
1528}
1529EXPORT_SYMBOL(pskb_expand_head);
1530
1531/* Make private copy of skb with writable head and some headroom */
1532
1533struct sk_buff *skb_realloc_headroom(struct sk_buff *skb, unsigned int headroom)
1534{
1535 struct sk_buff *skb2;
1536 int delta = headroom - skb_headroom(skb);
1537
1538 if (delta <= 0)
1539 skb2 = pskb_copy(skb, GFP_ATOMIC);
1540 else {
1541 skb2 = skb_clone(skb, GFP_ATOMIC);
1542 if (skb2 && pskb_expand_head(skb2, SKB_DATA_ALIGN(delta), 0,
1543 GFP_ATOMIC)) {
1544 kfree_skb(skb2);
1545 skb2 = NULL;
1546 }
1547 }
1548 return skb2;
1549}
1550EXPORT_SYMBOL(skb_realloc_headroom);
1551
1552/**
1553 * skb_copy_expand - copy and expand sk_buff
1554 * @skb: buffer to copy
1555 * @newheadroom: new free bytes at head
1556 * @newtailroom: new free bytes at tail
1557 * @gfp_mask: allocation priority
1558 *
1559 * Make a copy of both an &sk_buff and its data and while doing so
1560 * allocate additional space.
1561 *
1562 * This is used when the caller wishes to modify the data and needs a
1563 * private copy of the data to alter as well as more space for new fields.
1564 * Returns %NULL on failure or the pointer to the buffer
1565 * on success. The returned buffer has a reference count of 1.
1566 *
1567 * You must pass %GFP_ATOMIC as the allocation priority if this function
1568 * is called from an interrupt.
1569 */
1570struct sk_buff *skb_copy_expand(const struct sk_buff *skb,
1571 int newheadroom, int newtailroom,
1572 gfp_t gfp_mask)
1573{
1574 /*
1575 * Allocate the copy buffer
1576 */
1577 struct sk_buff *n = __alloc_skb(newheadroom + skb->len + newtailroom,
1578 gfp_mask, skb_alloc_rx_flag(skb),
1579 NUMA_NO_NODE);
1580 int oldheadroom = skb_headroom(skb);
1581 int head_copy_len, head_copy_off;
1582
1583 if (!n)
1584 return NULL;
1585
1586 skb_reserve(n, newheadroom);
1587
1588 /* Set the tail pointer and length */
1589 skb_put(n, skb->len);
1590
1591 head_copy_len = oldheadroom;
1592 head_copy_off = 0;
1593 if (newheadroom <= head_copy_len)
1594 head_copy_len = newheadroom;
1595 else
1596 head_copy_off = newheadroom - head_copy_len;
1597
1598 /* Copy the linear header and data. */
1599 BUG_ON(skb_copy_bits(skb, -head_copy_len, n->head + head_copy_off,
1600 skb->len + head_copy_len));
1601
1602 copy_skb_header(n, skb);
1603
1604 skb_headers_offset_update(n, newheadroom - oldheadroom);
1605
1606 return n;
1607}
1608EXPORT_SYMBOL(skb_copy_expand);
1609
1610/**
1611 * __skb_pad - zero pad the tail of an skb
1612 * @skb: buffer to pad
1613 * @pad: space to pad
1614 * @free_on_error: free buffer on error
1615 *
1616 * Ensure that a buffer is followed by a padding area that is zero
1617 * filled. Used by network drivers which may DMA or transfer data
1618 * beyond the buffer end onto the wire.
1619 *
1620 * May return error in out of memory cases. The skb is freed on error
1621 * if @free_on_error is true.
1622 */
1623
1624int __skb_pad(struct sk_buff *skb, int pad, bool free_on_error)
1625{
1626 int err;
1627 int ntail;
1628
1629 /* If the skbuff is non linear tailroom is always zero.. */
1630 if (!skb_cloned(skb) && skb_tailroom(skb) >= pad) {
1631 memset(skb->data+skb->len, 0, pad);
1632 return 0;
1633 }
1634
1635 ntail = skb->data_len + pad - (skb->end - skb->tail);
1636 if (likely(skb_cloned(skb) || ntail > 0)) {
1637 err = pskb_expand_head(skb, 0, ntail, GFP_ATOMIC);
1638 if (unlikely(err))
1639 goto free_skb;
1640 }
1641
1642 /* FIXME: The use of this function with non-linear skb's really needs
1643 * to be audited.
1644 */
1645 err = skb_linearize(skb);
1646 if (unlikely(err))
1647 goto free_skb;
1648
1649 memset(skb->data + skb->len, 0, pad);
1650 return 0;
1651
1652free_skb:
1653 if (free_on_error)
1654 kfree_skb(skb);
1655 return err;
1656}
1657EXPORT_SYMBOL(__skb_pad);
1658
1659/**
1660 * pskb_put - add data to the tail of a potentially fragmented buffer
1661 * @skb: start of the buffer to use
1662 * @tail: tail fragment of the buffer to use
1663 * @len: amount of data to add
1664 *
1665 * This function extends the used data area of the potentially
1666 * fragmented buffer. @tail must be the last fragment of @skb -- or
1667 * @skb itself. If this would exceed the total buffer size the kernel
1668 * will panic. A pointer to the first byte of the extra data is
1669 * returned.
1670 */
1671
1672void *pskb_put(struct sk_buff *skb, struct sk_buff *tail, int len)
1673{
1674 if (tail != skb) {
1675 skb->data_len += len;
1676 skb->len += len;
1677 }
1678 return skb_put(tail, len);
1679}
1680EXPORT_SYMBOL_GPL(pskb_put);
1681
1682/**
1683 * skb_put - add data to a buffer
1684 * @skb: buffer to use
1685 * @len: amount of data to add
1686 *
1687 * This function extends the used data area of the buffer. If this would
1688 * exceed the total buffer size the kernel will panic. A pointer to the
1689 * first byte of the extra data is returned.
1690 */
1691void *skb_put(struct sk_buff *skb, unsigned int len)
1692{
1693 void *tmp = skb_tail_pointer(skb);
1694 SKB_LINEAR_ASSERT(skb);
1695 skb->tail += len;
1696 skb->len += len;
1697 if (unlikely(skb->tail > skb->end))
1698 skb_over_panic(skb, len, __builtin_return_address(0));
1699 return tmp;
1700}
1701EXPORT_SYMBOL(skb_put);
1702
1703/**
1704 * skb_push - add data to the start of a buffer
1705 * @skb: buffer to use
1706 * @len: amount of data to add
1707 *
1708 * This function extends the used data area of the buffer at the buffer
1709 * start. If this would exceed the total buffer headroom the kernel will
1710 * panic. A pointer to the first byte of the extra data is returned.
1711 */
1712void *skb_push(struct sk_buff *skb, unsigned int len)
1713{
1714 skb->data -= len;
1715 skb->len += len;
1716 if (unlikely(skb->data<skb->head))
1717 skb_under_panic(skb, len, __builtin_return_address(0));
1718 return skb->data;
1719}
1720EXPORT_SYMBOL(skb_push);
1721
1722/**
1723 * skb_pull - remove data from the start of a buffer
1724 * @skb: buffer to use
1725 * @len: amount of data to remove
1726 *
1727 * This function removes data from the start of a buffer, returning
1728 * the memory to the headroom. A pointer to the next data in the buffer
1729 * is returned. Once the data has been pulled future pushes will overwrite
1730 * the old data.
1731 */
1732void *skb_pull(struct sk_buff *skb, unsigned int len)
1733{
1734 return skb_pull_inline(skb, len);
1735}
1736EXPORT_SYMBOL(skb_pull);
1737
1738/**
1739 * skb_trim - remove end from a buffer
1740 * @skb: buffer to alter
1741 * @len: new length
1742 *
1743 * Cut the length of a buffer down by removing data from the tail. If
1744 * the buffer is already under the length specified it is not modified.
1745 * The skb must be linear.
1746 */
1747void skb_trim(struct sk_buff *skb, unsigned int len)
1748{
1749 if (skb->len > len)
1750 __skb_trim(skb, len);
1751}
1752EXPORT_SYMBOL(skb_trim);
1753
1754/* Trims skb to length len. It can change skb pointers.
1755 */
1756
1757int ___pskb_trim(struct sk_buff *skb, unsigned int len)
1758{
1759 struct sk_buff **fragp;
1760 struct sk_buff *frag;
1761 int offset = skb_headlen(skb);
1762 int nfrags = skb_shinfo(skb)->nr_frags;
1763 int i;
1764 int err;
1765
1766 if (skb_cloned(skb) &&
1767 unlikely((err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC))))
1768 return err;
1769
1770 i = 0;
1771 if (offset >= len)
1772 goto drop_pages;
1773
1774 for (; i < nfrags; i++) {
1775 int end = offset + skb_frag_size(&skb_shinfo(skb)->frags[i]);
1776
1777 if (end < len) {
1778 offset = end;
1779 continue;
1780 }
1781
1782 skb_frag_size_set(&skb_shinfo(skb)->frags[i++], len - offset);
1783
1784drop_pages:
1785 skb_shinfo(skb)->nr_frags = i;
1786
1787 for (; i < nfrags; i++)
1788 skb_frag_unref(skb, i);
1789
1790 if (skb_has_frag_list(skb))
1791 skb_drop_fraglist(skb);
1792 goto done;
1793 }
1794
1795 for (fragp = &skb_shinfo(skb)->frag_list; (frag = *fragp);
1796 fragp = &frag->next) {
1797 int end = offset + frag->len;
1798
1799 if (skb_shared(frag)) {
1800 struct sk_buff *nfrag;
1801
1802 nfrag = skb_clone(frag, GFP_ATOMIC);
1803 if (unlikely(!nfrag))
1804 return -ENOMEM;
1805
1806 nfrag->next = frag->next;
1807 consume_skb(frag);
1808 frag = nfrag;
1809 *fragp = frag;
1810 }
1811
1812 if (end < len) {
1813 offset = end;
1814 continue;
1815 }
1816
1817 if (end > len &&
1818 unlikely((err = pskb_trim(frag, len - offset))))
1819 return err;
1820
1821 if (frag->next)
1822 skb_drop_list(&frag->next);
1823 break;
1824 }
1825
1826done:
1827 if (len > skb_headlen(skb)) {
1828 skb->data_len -= skb->len - len;
1829 skb->len = len;
1830 } else {
1831 skb->len = len;
1832 skb->data_len = 0;
1833 skb_set_tail_pointer(skb, len);
1834 }
1835
1836 if (!skb->sk || skb->destructor == sock_edemux)
1837 skb_condense(skb);
1838 return 0;
1839}
1840EXPORT_SYMBOL(___pskb_trim);
1841
1842/**
1843 * __pskb_pull_tail - advance tail of skb header
1844 * @skb: buffer to reallocate
1845 * @delta: number of bytes to advance tail
1846 *
1847 * The function makes a sense only on a fragmented &sk_buff,
1848 * it expands header moving its tail forward and copying necessary
1849 * data from fragmented part.
1850 *
1851 * &sk_buff MUST have reference count of 1.
1852 *
1853 * Returns %NULL (and &sk_buff does not change) if pull failed
1854 * or value of new tail of skb in the case of success.
1855 *
1856 * All the pointers pointing into skb header may change and must be
1857 * reloaded after call to this function.
1858 */
1859
1860/* Moves tail of skb head forward, copying data from fragmented part,
1861 * when it is necessary.
1862 * 1. It may fail due to malloc failure.
1863 * 2. It may change skb pointers.
1864 *
1865 * It is pretty complicated. Luckily, it is called only in exceptional cases.
1866 */
1867void *__pskb_pull_tail(struct sk_buff *skb, int delta)
1868{
1869 /* If skb has not enough free space at tail, get new one
1870 * plus 128 bytes for future expansions. If we have enough
1871 * room at tail, reallocate without expansion only if skb is cloned.
1872 */
1873 int i, k, eat = (skb->tail + delta) - skb->end;
1874
1875 if (eat > 0 || skb_cloned(skb)) {
1876 if (pskb_expand_head(skb, 0, eat > 0 ? eat + 128 : 0,
1877 GFP_ATOMIC))
1878 return NULL;
1879 }
1880
1881 BUG_ON(skb_copy_bits(skb, skb_headlen(skb),
1882 skb_tail_pointer(skb), delta));
1883
1884 /* Optimization: no fragments, no reasons to preestimate
1885 * size of pulled pages. Superb.
1886 */
1887 if (!skb_has_frag_list(skb))
1888 goto pull_pages;
1889
1890 /* Estimate size of pulled pages. */
1891 eat = delta;
1892 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1893 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1894
1895 if (size >= eat)
1896 goto pull_pages;
1897 eat -= size;
1898 }
1899
1900 /* If we need update frag list, we are in troubles.
1901 * Certainly, it is possible to add an offset to skb data,
1902 * but taking into account that pulling is expected to
1903 * be very rare operation, it is worth to fight against
1904 * further bloating skb head and crucify ourselves here instead.
1905 * Pure masohism, indeed. 8)8)
1906 */
1907 if (eat) {
1908 struct sk_buff *list = skb_shinfo(skb)->frag_list;
1909 struct sk_buff *clone = NULL;
1910 struct sk_buff *insp = NULL;
1911
1912 do {
1913 BUG_ON(!list);
1914
1915 if (list->len <= eat) {
1916 /* Eaten as whole. */
1917 eat -= list->len;
1918 list = list->next;
1919 insp = list;
1920 } else {
1921 /* Eaten partially. */
1922
1923 if (skb_shared(list)) {
1924 /* Sucks! We need to fork list. :-( */
1925 clone = skb_clone(list, GFP_ATOMIC);
1926 if (!clone)
1927 return NULL;
1928 insp = list->next;
1929 list = clone;
1930 } else {
1931 /* This may be pulled without
1932 * problems. */
1933 insp = list;
1934 }
1935 if (!pskb_pull(list, eat)) {
1936 kfree_skb(clone);
1937 return NULL;
1938 }
1939 break;
1940 }
1941 } while (eat);
1942
1943 /* Free pulled out fragments. */
1944 while ((list = skb_shinfo(skb)->frag_list) != insp) {
1945 skb_shinfo(skb)->frag_list = list->next;
1946 kfree_skb(list);
1947 }
1948 /* And insert new clone at head. */
1949 if (clone) {
1950 clone->next = list;
1951 skb_shinfo(skb)->frag_list = clone;
1952 }
1953 }
1954 /* Success! Now we may commit changes to skb data. */
1955
1956pull_pages:
1957 eat = delta;
1958 k = 0;
1959 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1960 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1961
1962 if (size <= eat) {
1963 skb_frag_unref(skb, i);
1964 eat -= size;
1965 } else {
1966 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1967 if (eat) {
1968 skb_shinfo(skb)->frags[k].page_offset += eat;
1969 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1970 if (!i)
1971 goto end;
1972 eat = 0;
1973 }
1974 k++;
1975 }
1976 }
1977 skb_shinfo(skb)->nr_frags = k;
1978
1979end:
1980 skb->tail += delta;
1981 skb->data_len -= delta;
1982
1983 if (!skb->data_len)
1984 skb_zcopy_clear(skb, false);
1985
1986 return skb_tail_pointer(skb);
1987}
1988EXPORT_SYMBOL(__pskb_pull_tail);
1989
1990/**
1991 * skb_copy_bits - copy bits from skb to kernel buffer
1992 * @skb: source skb
1993 * @offset: offset in source
1994 * @to: destination buffer
1995 * @len: number of bytes to copy
1996 *
1997 * Copy the specified number of bytes from the source skb to the
1998 * destination buffer.
1999 *
2000 * CAUTION ! :
2001 * If its prototype is ever changed,
2002 * check arch/{*}/net/{*}.S files,
2003 * since it is called from BPF assembly code.
2004 */
2005int skb_copy_bits(const struct sk_buff *skb, int offset, void *to, int len)
2006{
2007 int start = skb_headlen(skb);
2008 struct sk_buff *frag_iter;
2009 int i, copy;
2010
2011 if (offset > (int)skb->len - len)
2012 goto fault;
2013
2014 /* Copy header. */
2015 if ((copy = start - offset) > 0) {
2016 if (copy > len)
2017 copy = len;
2018 skb_copy_from_linear_data_offset(skb, offset, to, copy);
2019 if ((len -= copy) == 0)
2020 return 0;
2021 offset += copy;
2022 to += copy;
2023 }
2024
2025 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2026 int end;
2027 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
2028
2029 WARN_ON(start > offset + len);
2030
2031 end = start + skb_frag_size(f);
2032 if ((copy = end - offset) > 0) {
2033 u32 p_off, p_len, copied;
2034 struct page *p;
2035 u8 *vaddr;
2036
2037 if (copy > len)
2038 copy = len;
2039
2040 skb_frag_foreach_page(f,
2041 f->page_offset + offset - start,
2042 copy, p, p_off, p_len, copied) {
2043 vaddr = kmap_atomic(p);
2044 memcpy(to + copied, vaddr + p_off, p_len);
2045 kunmap_atomic(vaddr);
2046 }
2047
2048 if ((len -= copy) == 0)
2049 return 0;
2050 offset += copy;
2051 to += copy;
2052 }
2053 start = end;
2054 }
2055
2056 skb_walk_frags(skb, frag_iter) {
2057 int end;
2058
2059 WARN_ON(start > offset + len);
2060
2061 end = start + frag_iter->len;
2062 if ((copy = end - offset) > 0) {
2063 if (copy > len)
2064 copy = len;
2065 if (skb_copy_bits(frag_iter, offset - start, to, copy))
2066 goto fault;
2067 if ((len -= copy) == 0)
2068 return 0;
2069 offset += copy;
2070 to += copy;
2071 }
2072 start = end;
2073 }
2074
2075 if (!len)
2076 return 0;
2077
2078fault:
2079 return -EFAULT;
2080}
2081EXPORT_SYMBOL(skb_copy_bits);
2082
2083/*
2084 * Callback from splice_to_pipe(), if we need to release some pages
2085 * at the end of the spd in case we error'ed out in filling the pipe.
2086 */
2087static void sock_spd_release(struct splice_pipe_desc *spd, unsigned int i)
2088{
2089 put_page(spd->pages[i]);
2090}
2091
2092static struct page *linear_to_page(struct page *page, unsigned int *len,
2093 unsigned int *offset,
2094 struct sock *sk)
2095{
2096 struct page_frag *pfrag = sk_page_frag(sk);
2097
2098 if (!sk_page_frag_refill(sk, pfrag))
2099 return NULL;
2100
2101 *len = min_t(unsigned int, *len, pfrag->size - pfrag->offset);
2102
2103 memcpy(page_address(pfrag->page) + pfrag->offset,
2104 page_address(page) + *offset, *len);
2105 *offset = pfrag->offset;
2106 pfrag->offset += *len;
2107
2108 return pfrag->page;
2109}
2110
2111static bool spd_can_coalesce(const struct splice_pipe_desc *spd,
2112 struct page *page,
2113 unsigned int offset)
2114{
2115 return spd->nr_pages &&
2116 spd->pages[spd->nr_pages - 1] == page &&
2117 (spd->partial[spd->nr_pages - 1].offset +
2118 spd->partial[spd->nr_pages - 1].len == offset);
2119}
2120
2121/*
2122 * Fill page/offset/length into spd, if it can hold more pages.
2123 */
2124static bool spd_fill_page(struct splice_pipe_desc *spd,
2125 struct pipe_inode_info *pipe, struct page *page,
2126 unsigned int *len, unsigned int offset,
2127 bool linear,
2128 struct sock *sk)
2129{
2130 if (unlikely(spd->nr_pages == MAX_SKB_FRAGS))
2131 return true;
2132
2133 if (linear) {
2134 page = linear_to_page(page, len, &offset, sk);
2135 if (!page)
2136 return true;
2137 }
2138 if (spd_can_coalesce(spd, page, offset)) {
2139 spd->partial[spd->nr_pages - 1].len += *len;
2140 return false;
2141 }
2142 get_page(page);
2143 spd->pages[spd->nr_pages] = page;
2144 spd->partial[spd->nr_pages].len = *len;
2145 spd->partial[spd->nr_pages].offset = offset;
2146 spd->nr_pages++;
2147
2148 return false;
2149}
2150
2151static bool __splice_segment(struct page *page, unsigned int poff,
2152 unsigned int plen, unsigned int *off,
2153 unsigned int *len,
2154 struct splice_pipe_desc *spd, bool linear,
2155 struct sock *sk,
2156 struct pipe_inode_info *pipe)
2157{
2158 if (!*len)
2159 return true;
2160
2161 /* skip this segment if already processed */
2162 if (*off >= plen) {
2163 *off -= plen;
2164 return false;
2165 }
2166
2167 /* ignore any bits we already processed */
2168 poff += *off;
2169 plen -= *off;
2170 *off = 0;
2171
2172 do {
2173 unsigned int flen = min(*len, plen);
2174
2175 if (spd_fill_page(spd, pipe, page, &flen, poff,
2176 linear, sk))
2177 return true;
2178 poff += flen;
2179 plen -= flen;
2180 *len -= flen;
2181 } while (*len && plen);
2182
2183 return false;
2184}
2185
2186/*
2187 * Map linear and fragment data from the skb to spd. It reports true if the
2188 * pipe is full or if we already spliced the requested length.
2189 */
2190static bool __skb_splice_bits(struct sk_buff *skb, struct pipe_inode_info *pipe,
2191 unsigned int *offset, unsigned int *len,
2192 struct splice_pipe_desc *spd, struct sock *sk)
2193{
2194 int seg;
2195 struct sk_buff *iter;
2196
2197 /* map the linear part :
2198 * If skb->head_frag is set, this 'linear' part is backed by a
2199 * fragment, and if the head is not shared with any clones then
2200 * we can avoid a copy since we own the head portion of this page.
2201 */
2202 if (__splice_segment(virt_to_page(skb->data),
2203 (unsigned long) skb->data & (PAGE_SIZE - 1),
2204 skb_headlen(skb),
2205 offset, len, spd,
2206 skb_head_is_locked(skb),
2207 sk, pipe))
2208 return true;
2209
2210 /*
2211 * then map the fragments
2212 */
2213 for (seg = 0; seg < skb_shinfo(skb)->nr_frags; seg++) {
2214 const skb_frag_t *f = &skb_shinfo(skb)->frags[seg];
2215
2216 if (__splice_segment(skb_frag_page(f),
2217 f->page_offset, skb_frag_size(f),
2218 offset, len, spd, false, sk, pipe))
2219 return true;
2220 }
2221
2222 skb_walk_frags(skb, iter) {
2223 if (*offset >= iter->len) {
2224 *offset -= iter->len;
2225 continue;
2226 }
2227 /* __skb_splice_bits() only fails if the output has no room
2228 * left, so no point in going over the frag_list for the error
2229 * case.
2230 */
2231 if (__skb_splice_bits(iter, pipe, offset, len, spd, sk))
2232 return true;
2233 }
2234
2235 return false;
2236}
2237
2238/*
2239 * Map data from the skb to a pipe. Should handle both the linear part,
2240 * the fragments, and the frag list.
2241 */
2242int skb_splice_bits(struct sk_buff *skb, struct sock *sk, unsigned int offset,
2243 struct pipe_inode_info *pipe, unsigned int tlen,
2244 unsigned int flags)
2245{
2246 struct partial_page partial[MAX_SKB_FRAGS];
2247 struct page *pages[MAX_SKB_FRAGS];
2248 struct splice_pipe_desc spd = {
2249 .pages = pages,
2250 .partial = partial,
2251 .nr_pages_max = MAX_SKB_FRAGS,
2252 .ops = &nosteal_pipe_buf_ops,
2253 .spd_release = sock_spd_release,
2254 };
2255 int ret = 0;
2256
2257 __skb_splice_bits(skb, pipe, &offset, &tlen, &spd, sk);
2258
2259 if (spd.nr_pages)
2260 ret = splice_to_pipe(pipe, &spd);
2261
2262 return ret;
2263}
2264EXPORT_SYMBOL_GPL(skb_splice_bits);
2265
2266/* Send skb data on a socket. Socket must be locked. */
2267int skb_send_sock_locked(struct sock *sk, struct sk_buff *skb, int offset,
2268 int len)
2269{
2270 unsigned int orig_len = len;
2271 struct sk_buff *head = skb;
2272 unsigned short fragidx;
2273 int slen, ret;
2274
2275do_frag_list:
2276
2277 /* Deal with head data */
2278 while (offset < skb_headlen(skb) && len) {
2279 struct kvec kv;
2280 struct msghdr msg;
2281
2282 slen = min_t(int, len, skb_headlen(skb) - offset);
2283 kv.iov_base = skb->data + offset;
2284 kv.iov_len = slen;
2285 memset(&msg, 0, sizeof(msg));
2286
2287 ret = kernel_sendmsg_locked(sk, &msg, &kv, 1, slen);
2288 if (ret <= 0)
2289 goto error;
2290
2291 offset += ret;
2292 len -= ret;
2293 }
2294
2295 /* All the data was skb head? */
2296 if (!len)
2297 goto out;
2298
2299 /* Make offset relative to start of frags */
2300 offset -= skb_headlen(skb);
2301
2302 /* Find where we are in frag list */
2303 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags; fragidx++) {
2304 skb_frag_t *frag = &skb_shinfo(skb)->frags[fragidx];
2305
2306 if (offset < frag->size)
2307 break;
2308
2309 offset -= frag->size;
2310 }
2311
2312 for (; len && fragidx < skb_shinfo(skb)->nr_frags; fragidx++) {
2313 skb_frag_t *frag = &skb_shinfo(skb)->frags[fragidx];
2314
2315 slen = min_t(size_t, len, frag->size - offset);
2316
2317 while (slen) {
2318 ret = kernel_sendpage_locked(sk, frag->page.p,
2319 frag->page_offset + offset,
2320 slen, MSG_DONTWAIT);
2321 if (ret <= 0)
2322 goto error;
2323
2324 len -= ret;
2325 offset += ret;
2326 slen -= ret;
2327 }
2328
2329 offset = 0;
2330 }
2331
2332 if (len) {
2333 /* Process any frag lists */
2334
2335 if (skb == head) {
2336 if (skb_has_frag_list(skb)) {
2337 skb = skb_shinfo(skb)->frag_list;
2338 goto do_frag_list;
2339 }
2340 } else if (skb->next) {
2341 skb = skb->next;
2342 goto do_frag_list;
2343 }
2344 }
2345
2346out:
2347 return orig_len - len;
2348
2349error:
2350 return orig_len == len ? ret : orig_len - len;
2351}
2352EXPORT_SYMBOL_GPL(skb_send_sock_locked);
2353
2354/* Send skb data on a socket. */
2355int skb_send_sock(struct sock *sk, struct sk_buff *skb, int offset, int len)
2356{
2357 int ret = 0;
2358
2359 lock_sock(sk);
2360 ret = skb_send_sock_locked(sk, skb, offset, len);
2361 release_sock(sk);
2362
2363 return ret;
2364}
2365EXPORT_SYMBOL_GPL(skb_send_sock);
2366
2367/**
2368 * skb_store_bits - store bits from kernel buffer to skb
2369 * @skb: destination buffer
2370 * @offset: offset in destination
2371 * @from: source buffer
2372 * @len: number of bytes to copy
2373 *
2374 * Copy the specified number of bytes from the source buffer to the
2375 * destination skb. This function handles all the messy bits of
2376 * traversing fragment lists and such.
2377 */
2378
2379int skb_store_bits(struct sk_buff *skb, int offset, const void *from, int len)
2380{
2381 int start = skb_headlen(skb);
2382 struct sk_buff *frag_iter;
2383 int i, copy;
2384
2385 if (offset > (int)skb->len - len)
2386 goto fault;
2387
2388 if ((copy = start - offset) > 0) {
2389 if (copy > len)
2390 copy = len;
2391 skb_copy_to_linear_data_offset(skb, offset, from, copy);
2392 if ((len -= copy) == 0)
2393 return 0;
2394 offset += copy;
2395 from += copy;
2396 }
2397
2398 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2399 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2400 int end;
2401
2402 WARN_ON(start > offset + len);
2403
2404 end = start + skb_frag_size(frag);
2405 if ((copy = end - offset) > 0) {
2406 u32 p_off, p_len, copied;
2407 struct page *p;
2408 u8 *vaddr;
2409
2410 if (copy > len)
2411 copy = len;
2412
2413 skb_frag_foreach_page(frag,
2414 frag->page_offset + offset - start,
2415 copy, p, p_off, p_len, copied) {
2416 vaddr = kmap_atomic(p);
2417 memcpy(vaddr + p_off, from + copied, p_len);
2418 kunmap_atomic(vaddr);
2419 }
2420
2421 if ((len -= copy) == 0)
2422 return 0;
2423 offset += copy;
2424 from += copy;
2425 }
2426 start = end;
2427 }
2428
2429 skb_walk_frags(skb, frag_iter) {
2430 int end;
2431
2432 WARN_ON(start > offset + len);
2433
2434 end = start + frag_iter->len;
2435 if ((copy = end - offset) > 0) {
2436 if (copy > len)
2437 copy = len;
2438 if (skb_store_bits(frag_iter, offset - start,
2439 from, copy))
2440 goto fault;
2441 if ((len -= copy) == 0)
2442 return 0;
2443 offset += copy;
2444 from += copy;
2445 }
2446 start = end;
2447 }
2448 if (!len)
2449 return 0;
2450
2451fault:
2452 return -EFAULT;
2453}
2454EXPORT_SYMBOL(skb_store_bits);
2455
2456/* Checksum skb data. */
2457__wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
2458 __wsum csum, const struct skb_checksum_ops *ops)
2459{
2460 int start = skb_headlen(skb);
2461 int i, copy = start - offset;
2462 struct sk_buff *frag_iter;
2463 int pos = 0;
2464
2465 /* Checksum header. */
2466 if (copy > 0) {
2467 if (copy > len)
2468 copy = len;
2469 csum = ops->update(skb->data + offset, copy, csum);
2470 if ((len -= copy) == 0)
2471 return csum;
2472 offset += copy;
2473 pos = copy;
2474 }
2475
2476 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2477 int end;
2478 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2479
2480 WARN_ON(start > offset + len);
2481
2482 end = start + skb_frag_size(frag);
2483 if ((copy = end - offset) > 0) {
2484 u32 p_off, p_len, copied;
2485 struct page *p;
2486 __wsum csum2;
2487 u8 *vaddr;
2488
2489 if (copy > len)
2490 copy = len;
2491
2492 skb_frag_foreach_page(frag,
2493 frag->page_offset + offset - start,
2494 copy, p, p_off, p_len, copied) {
2495 vaddr = kmap_atomic(p);
2496 csum2 = ops->update(vaddr + p_off, p_len, 0);
2497 kunmap_atomic(vaddr);
2498 csum = ops->combine(csum, csum2, pos, p_len);
2499 pos += p_len;
2500 }
2501
2502 if (!(len -= copy))
2503 return csum;
2504 offset += copy;
2505 }
2506 start = end;
2507 }
2508
2509 skb_walk_frags(skb, frag_iter) {
2510 int end;
2511
2512 WARN_ON(start > offset + len);
2513
2514 end = start + frag_iter->len;
2515 if ((copy = end - offset) > 0) {
2516 __wsum csum2;
2517 if (copy > len)
2518 copy = len;
2519 csum2 = __skb_checksum(frag_iter, offset - start,
2520 copy, 0, ops);
2521 csum = ops->combine(csum, csum2, pos, copy);
2522 if ((len -= copy) == 0)
2523 return csum;
2524 offset += copy;
2525 pos += copy;
2526 }
2527 start = end;
2528 }
2529 BUG_ON(len);
2530
2531 return csum;
2532}
2533EXPORT_SYMBOL(__skb_checksum);
2534
2535__wsum skb_checksum(const struct sk_buff *skb, int offset,
2536 int len, __wsum csum)
2537{
2538 const struct skb_checksum_ops ops = {
2539 .update = csum_partial_ext,
2540 .combine = csum_block_add_ext,
2541 };
2542
2543 return __skb_checksum(skb, offset, len, csum, &ops);
2544}
2545EXPORT_SYMBOL(skb_checksum);
2546
2547/* Both of above in one bottle. */
2548
2549__wsum skb_copy_and_csum_bits(const struct sk_buff *skb, int offset,
2550 u8 *to, int len, __wsum csum)
2551{
2552 int start = skb_headlen(skb);
2553 int i, copy = start - offset;
2554 struct sk_buff *frag_iter;
2555 int pos = 0;
2556
2557 /* Copy header. */
2558 if (copy > 0) {
2559 if (copy > len)
2560 copy = len;
2561 csum = csum_partial_copy_nocheck(skb->data + offset, to,
2562 copy, csum);
2563 if ((len -= copy) == 0)
2564 return csum;
2565 offset += copy;
2566 to += copy;
2567 pos = copy;
2568 }
2569
2570 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2571 int end;
2572
2573 WARN_ON(start > offset + len);
2574
2575 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
2576 if ((copy = end - offset) > 0) {
2577 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2578 u32 p_off, p_len, copied;
2579 struct page *p;
2580 __wsum csum2;
2581 u8 *vaddr;
2582
2583 if (copy > len)
2584 copy = len;
2585
2586 skb_frag_foreach_page(frag,
2587 frag->page_offset + offset - start,
2588 copy, p, p_off, p_len, copied) {
2589 vaddr = kmap_atomic(p);
2590 csum2 = csum_partial_copy_nocheck(vaddr + p_off,
2591 to + copied,
2592 p_len, 0);
2593 kunmap_atomic(vaddr);
2594 csum = csum_block_add(csum, csum2, pos);
2595 pos += p_len;
2596 }
2597
2598 if (!(len -= copy))
2599 return csum;
2600 offset += copy;
2601 to += copy;
2602 }
2603 start = end;
2604 }
2605
2606 skb_walk_frags(skb, frag_iter) {
2607 __wsum csum2;
2608 int end;
2609
2610 WARN_ON(start > offset + len);
2611
2612 end = start + frag_iter->len;
2613 if ((copy = end - offset) > 0) {
2614 if (copy > len)
2615 copy = len;
2616 csum2 = skb_copy_and_csum_bits(frag_iter,
2617 offset - start,
2618 to, copy, 0);
2619 csum = csum_block_add(csum, csum2, pos);
2620 if ((len -= copy) == 0)
2621 return csum;
2622 offset += copy;
2623 to += copy;
2624 pos += copy;
2625 }
2626 start = end;
2627 }
2628 BUG_ON(len);
2629 return csum;
2630}
2631EXPORT_SYMBOL(skb_copy_and_csum_bits);
2632
2633static __wsum warn_crc32c_csum_update(const void *buff, int len, __wsum sum)
2634{
2635 net_warn_ratelimited(
2636 "%s: attempt to compute crc32c without libcrc32c.ko\n",
2637 __func__);
2638 return 0;
2639}
2640
2641static __wsum warn_crc32c_csum_combine(__wsum csum, __wsum csum2,
2642 int offset, int len)
2643{
2644 net_warn_ratelimited(
2645 "%s: attempt to compute crc32c without libcrc32c.ko\n",
2646 __func__);
2647 return 0;
2648}
2649
2650static const struct skb_checksum_ops default_crc32c_ops = {
2651 .update = warn_crc32c_csum_update,
2652 .combine = warn_crc32c_csum_combine,
2653};
2654
2655const struct skb_checksum_ops *crc32c_csum_stub __read_mostly =
2656 &default_crc32c_ops;
2657EXPORT_SYMBOL(crc32c_csum_stub);
2658
2659 /**
2660 * skb_zerocopy_headlen - Calculate headroom needed for skb_zerocopy()
2661 * @from: source buffer
2662 *
2663 * Calculates the amount of linear headroom needed in the 'to' skb passed
2664 * into skb_zerocopy().
2665 */
2666unsigned int
2667skb_zerocopy_headlen(const struct sk_buff *from)
2668{
2669 unsigned int hlen = 0;
2670
2671 if (!from->head_frag ||
2672 skb_headlen(from) < L1_CACHE_BYTES ||
2673 skb_shinfo(from)->nr_frags >= MAX_SKB_FRAGS)
2674 hlen = skb_headlen(from);
2675
2676 if (skb_has_frag_list(from))
2677 hlen = from->len;
2678
2679 return hlen;
2680}
2681EXPORT_SYMBOL_GPL(skb_zerocopy_headlen);
2682
2683/**
2684 * skb_zerocopy - Zero copy skb to skb
2685 * @to: destination buffer
2686 * @from: source buffer
2687 * @len: number of bytes to copy from source buffer
2688 * @hlen: size of linear headroom in destination buffer
2689 *
2690 * Copies up to `len` bytes from `from` to `to` by creating references
2691 * to the frags in the source buffer.
2692 *
2693 * The `hlen` as calculated by skb_zerocopy_headlen() specifies the
2694 * headroom in the `to` buffer.
2695 *
2696 * Return value:
2697 * 0: everything is OK
2698 * -ENOMEM: couldn't orphan frags of @from due to lack of memory
2699 * -EFAULT: skb_copy_bits() found some problem with skb geometry
2700 */
2701int
2702skb_zerocopy(struct sk_buff *to, struct sk_buff *from, int len, int hlen)
2703{
2704 int i, j = 0;
2705 int plen = 0; /* length of skb->head fragment */
2706 int ret;
2707 struct page *page;
2708 unsigned int offset;
2709
2710 BUG_ON(!from->head_frag && !hlen);
2711
2712 /* dont bother with small payloads */
2713 if (len <= skb_tailroom(to))
2714 return skb_copy_bits(from, 0, skb_put(to, len), len);
2715
2716 if (hlen) {
2717 ret = skb_copy_bits(from, 0, skb_put(to, hlen), hlen);
2718 if (unlikely(ret))
2719 return ret;
2720 len -= hlen;
2721 } else {
2722 plen = min_t(int, skb_headlen(from), len);
2723 if (plen) {
2724 page = virt_to_head_page(from->head);
2725 offset = from->data - (unsigned char *)page_address(page);
2726 __skb_fill_page_desc(to, 0, page, offset, plen);
2727 get_page(page);
2728 j = 1;
2729 len -= plen;
2730 }
2731 }
2732
2733 to->truesize += len + plen;
2734 to->len += len + plen;
2735 to->data_len += len + plen;
2736
2737 if (unlikely(skb_orphan_frags(from, GFP_ATOMIC))) {
2738 skb_tx_error(from);
2739 return -ENOMEM;
2740 }
2741 skb_zerocopy_clone(to, from, GFP_ATOMIC);
2742
2743 for (i = 0; i < skb_shinfo(from)->nr_frags; i++) {
2744 if (!len)
2745 break;
2746 skb_shinfo(to)->frags[j] = skb_shinfo(from)->frags[i];
2747 skb_shinfo(to)->frags[j].size = min_t(int, skb_shinfo(to)->frags[j].size, len);
2748 len -= skb_shinfo(to)->frags[j].size;
2749 skb_frag_ref(to, j);
2750 j++;
2751 }
2752 skb_shinfo(to)->nr_frags = j;
2753
2754 return 0;
2755}
2756EXPORT_SYMBOL_GPL(skb_zerocopy);
2757
2758void skb_copy_and_csum_dev(const struct sk_buff *skb, u8 *to)
2759{
2760 __wsum csum;
2761 long csstart;
2762
2763 if (skb->ip_summed == CHECKSUM_PARTIAL)
2764 csstart = skb_checksum_start_offset(skb);
2765 else
2766 csstart = skb_headlen(skb);
2767
2768 BUG_ON(csstart > skb_headlen(skb));
2769
2770 skb_copy_from_linear_data(skb, to, csstart);
2771
2772 csum = 0;
2773 if (csstart != skb->len)
2774 csum = skb_copy_and_csum_bits(skb, csstart, to + csstart,
2775 skb->len - csstart, 0);
2776
2777 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2778 long csstuff = csstart + skb->csum_offset;
2779
2780 *((__sum16 *)(to + csstuff)) = csum_fold(csum);
2781 }
2782}
2783EXPORT_SYMBOL(skb_copy_and_csum_dev);
2784
2785/**
2786 * skb_dequeue - remove from the head of the queue
2787 * @list: list to dequeue from
2788 *
2789 * Remove the head of the list. The list lock is taken so the function
2790 * may be used safely with other locking list functions. The head item is
2791 * returned or %NULL if the list is empty.
2792 */
2793
2794struct sk_buff *skb_dequeue(struct sk_buff_head *list)
2795{
2796 unsigned long flags;
2797 struct sk_buff *result;
2798
2799 spin_lock_irqsave(&list->lock, flags);
2800 result = __skb_dequeue(list);
2801 spin_unlock_irqrestore(&list->lock, flags);
2802 return result;
2803}
2804EXPORT_SYMBOL(skb_dequeue);
2805
2806/**
2807 * skb_dequeue_tail - remove from the tail of the queue
2808 * @list: list to dequeue from
2809 *
2810 * Remove the tail of the list. The list lock is taken so the function
2811 * may be used safely with other locking list functions. The tail item is
2812 * returned or %NULL if the list is empty.
2813 */
2814struct sk_buff *skb_dequeue_tail(struct sk_buff_head *list)
2815{
2816 unsigned long flags;
2817 struct sk_buff *result;
2818
2819 spin_lock_irqsave(&list->lock, flags);
2820 result = __skb_dequeue_tail(list);
2821 spin_unlock_irqrestore(&list->lock, flags);
2822 return result;
2823}
2824EXPORT_SYMBOL(skb_dequeue_tail);
2825
2826/**
2827 * skb_queue_purge - empty a list
2828 * @list: list to empty
2829 *
2830 * Delete all buffers on an &sk_buff list. Each buffer is removed from
2831 * the list and one reference dropped. This function takes the list
2832 * lock and is atomic with respect to other list locking functions.
2833 */
2834void skb_queue_purge(struct sk_buff_head *list)
2835{
2836 struct sk_buff *skb;
2837 while ((skb = skb_dequeue(list)) != NULL)
2838 kfree_skb(skb);
2839}
2840EXPORT_SYMBOL(skb_queue_purge);
2841
2842/**
2843 * skb_rbtree_purge - empty a skb rbtree
2844 * @root: root of the rbtree to empty
2845 *
2846 * Delete all buffers on an &sk_buff rbtree. Each buffer is removed from
2847 * the list and one reference dropped. This function does not take
2848 * any lock. Synchronization should be handled by the caller (e.g., TCP
2849 * out-of-order queue is protected by the socket lock).
2850 */
2851void skb_rbtree_purge(struct rb_root *root)
2852{
2853 struct rb_node *p = rb_first(root);
2854
2855 while (p) {
2856 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
2857
2858 p = rb_next(p);
2859 rb_erase(&skb->rbnode, root);
2860 kfree_skb(skb);
2861 }
2862}
2863
2864/**
2865 * skb_queue_head - queue a buffer at the list head
2866 * @list: list to use
2867 * @newsk: buffer to queue
2868 *
2869 * Queue a buffer at the start of the list. This function takes the
2870 * list lock and can be used safely with other locking &sk_buff functions
2871 * safely.
2872 *
2873 * A buffer cannot be placed on two lists at the same time.
2874 */
2875void skb_queue_head(struct sk_buff_head *list, struct sk_buff *newsk)
2876{
2877 unsigned long flags;
2878
2879 spin_lock_irqsave(&list->lock, flags);
2880 __skb_queue_head(list, newsk);
2881 spin_unlock_irqrestore(&list->lock, flags);
2882}
2883EXPORT_SYMBOL(skb_queue_head);
2884
2885/**
2886 * skb_queue_tail - queue a buffer at the list tail
2887 * @list: list to use
2888 * @newsk: buffer to queue
2889 *
2890 * Queue a buffer at the tail of the list. This function takes the
2891 * list lock and can be used safely with other locking &sk_buff functions
2892 * safely.
2893 *
2894 * A buffer cannot be placed on two lists at the same time.
2895 */
2896void skb_queue_tail(struct sk_buff_head *list, struct sk_buff *newsk)
2897{
2898 unsigned long flags;
2899
2900 spin_lock_irqsave(&list->lock, flags);
2901 __skb_queue_tail(list, newsk);
2902 spin_unlock_irqrestore(&list->lock, flags);
2903}
2904EXPORT_SYMBOL(skb_queue_tail);
2905
2906/**
2907 * skb_unlink - remove a buffer from a list
2908 * @skb: buffer to remove
2909 * @list: list to use
2910 *
2911 * Remove a packet from a list. The list locks are taken and this
2912 * function is atomic with respect to other list locked calls
2913 *
2914 * You must know what list the SKB is on.
2915 */
2916void skb_unlink(struct sk_buff *skb, struct sk_buff_head *list)
2917{
2918 unsigned long flags;
2919
2920 spin_lock_irqsave(&list->lock, flags);
2921 __skb_unlink(skb, list);
2922 spin_unlock_irqrestore(&list->lock, flags);
2923}
2924EXPORT_SYMBOL(skb_unlink);
2925
2926/**
2927 * skb_append - append a buffer
2928 * @old: buffer to insert after
2929 * @newsk: buffer to insert
2930 * @list: list to use
2931 *
2932 * Place a packet after a given packet in a list. The list locks are taken
2933 * and this function is atomic with respect to other list locked calls.
2934 * A buffer cannot be placed on two lists at the same time.
2935 */
2936void skb_append(struct sk_buff *old, struct sk_buff *newsk, struct sk_buff_head *list)
2937{
2938 unsigned long flags;
2939
2940 spin_lock_irqsave(&list->lock, flags);
2941 __skb_queue_after(list, old, newsk);
2942 spin_unlock_irqrestore(&list->lock, flags);
2943}
2944EXPORT_SYMBOL(skb_append);
2945
2946/**
2947 * skb_insert - insert a buffer
2948 * @old: buffer to insert before
2949 * @newsk: buffer to insert
2950 * @list: list to use
2951 *
2952 * Place a packet before a given packet in a list. The list locks are
2953 * taken and this function is atomic with respect to other list locked
2954 * calls.
2955 *
2956 * A buffer cannot be placed on two lists at the same time.
2957 */
2958void skb_insert(struct sk_buff *old, struct sk_buff *newsk, struct sk_buff_head *list)
2959{
2960 unsigned long flags;
2961
2962 spin_lock_irqsave(&list->lock, flags);
2963 __skb_insert(newsk, old->prev, old, list);
2964 spin_unlock_irqrestore(&list->lock, flags);
2965}
2966EXPORT_SYMBOL(skb_insert);
2967
2968static inline void skb_split_inside_header(struct sk_buff *skb,
2969 struct sk_buff* skb1,
2970 const u32 len, const int pos)
2971{
2972 int i;
2973
2974 skb_copy_from_linear_data_offset(skb, len, skb_put(skb1, pos - len),
2975 pos - len);
2976 /* And move data appendix as is. */
2977 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
2978 skb_shinfo(skb1)->frags[i] = skb_shinfo(skb)->frags[i];
2979
2980 skb_shinfo(skb1)->nr_frags = skb_shinfo(skb)->nr_frags;
2981 skb_shinfo(skb)->nr_frags = 0;
2982 skb1->data_len = skb->data_len;
2983 skb1->len += skb1->data_len;
2984 skb->data_len = 0;
2985 skb->len = len;
2986 skb_set_tail_pointer(skb, len);
2987}
2988
2989static inline void skb_split_no_header(struct sk_buff *skb,
2990 struct sk_buff* skb1,
2991 const u32 len, int pos)
2992{
2993 int i, k = 0;
2994 const int nfrags = skb_shinfo(skb)->nr_frags;
2995
2996 skb_shinfo(skb)->nr_frags = 0;
2997 skb1->len = skb1->data_len = skb->len - len;
2998 skb->len = len;
2999 skb->data_len = len - pos;
3000
3001 for (i = 0; i < nfrags; i++) {
3002 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
3003
3004 if (pos + size > len) {
3005 skb_shinfo(skb1)->frags[k] = skb_shinfo(skb)->frags[i];
3006
3007 if (pos < len) {
3008 /* Split frag.
3009 * We have two variants in this case:
3010 * 1. Move all the frag to the second
3011 * part, if it is possible. F.e.
3012 * this approach is mandatory for TUX,
3013 * where splitting is expensive.
3014 * 2. Split is accurately. We make this.
3015 */
3016 skb_frag_ref(skb, i);
3017 skb_shinfo(skb1)->frags[0].page_offset += len - pos;
3018 skb_frag_size_sub(&skb_shinfo(skb1)->frags[0], len - pos);
3019 skb_frag_size_set(&skb_shinfo(skb)->frags[i], len - pos);
3020 skb_shinfo(skb)->nr_frags++;
3021 }
3022 k++;
3023 } else
3024 skb_shinfo(skb)->nr_frags++;
3025 pos += size;
3026 }
3027 skb_shinfo(skb1)->nr_frags = k;
3028}
3029
3030/**
3031 * skb_split - Split fragmented skb to two parts at length len.
3032 * @skb: the buffer to split
3033 * @skb1: the buffer to receive the second part
3034 * @len: new length for skb
3035 */
3036void skb_split(struct sk_buff *skb, struct sk_buff *skb1, const u32 len)
3037{
3038 int pos = skb_headlen(skb);
3039
3040 skb_shinfo(skb1)->tx_flags |= skb_shinfo(skb)->tx_flags &
3041 SKBTX_SHARED_FRAG;
3042 skb_zerocopy_clone(skb1, skb, 0);
3043 if (len < pos) /* Split line is inside header. */
3044 skb_split_inside_header(skb, skb1, len, pos);
3045 else /* Second chunk has no header, nothing to copy. */
3046 skb_split_no_header(skb, skb1, len, pos);
3047}
3048EXPORT_SYMBOL(skb_split);
3049
3050/* Shifting from/to a cloned skb is a no-go.
3051 *
3052 * Caller cannot keep skb_shinfo related pointers past calling here!
3053 */
3054static int skb_prepare_for_shift(struct sk_buff *skb)
3055{
3056 return skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
3057}
3058
3059/**
3060 * skb_shift - Shifts paged data partially from skb to another
3061 * @tgt: buffer into which tail data gets added
3062 * @skb: buffer from which the paged data comes from
3063 * @shiftlen: shift up to this many bytes
3064 *
3065 * Attempts to shift up to shiftlen worth of bytes, which may be less than
3066 * the length of the skb, from skb to tgt. Returns number bytes shifted.
3067 * It's up to caller to free skb if everything was shifted.
3068 *
3069 * If @tgt runs out of frags, the whole operation is aborted.
3070 *
3071 * Skb cannot include anything else but paged data while tgt is allowed
3072 * to have non-paged data as well.
3073 *
3074 * TODO: full sized shift could be optimized but that would need
3075 * specialized skb free'er to handle frags without up-to-date nr_frags.
3076 */
3077int skb_shift(struct sk_buff *tgt, struct sk_buff *skb, int shiftlen)
3078{
3079 int from, to, merge, todo;
3080 struct skb_frag_struct *fragfrom, *fragto;
3081
3082 BUG_ON(shiftlen > skb->len);
3083
3084 if (skb_headlen(skb))
3085 return 0;
3086 if (skb_zcopy(tgt) || skb_zcopy(skb))
3087 return 0;
3088
3089 todo = shiftlen;
3090 from = 0;
3091 to = skb_shinfo(tgt)->nr_frags;
3092 fragfrom = &skb_shinfo(skb)->frags[from];
3093
3094 /* Actual merge is delayed until the point when we know we can
3095 * commit all, so that we don't have to undo partial changes
3096 */
3097 if (!to ||
3098 !skb_can_coalesce(tgt, to, skb_frag_page(fragfrom),
3099 fragfrom->page_offset)) {
3100 merge = -1;
3101 } else {
3102 merge = to - 1;
3103
3104 todo -= skb_frag_size(fragfrom);
3105 if (todo < 0) {
3106 if (skb_prepare_for_shift(skb) ||
3107 skb_prepare_for_shift(tgt))
3108 return 0;
3109
3110 /* All previous frag pointers might be stale! */
3111 fragfrom = &skb_shinfo(skb)->frags[from];
3112 fragto = &skb_shinfo(tgt)->frags[merge];
3113
3114 skb_frag_size_add(fragto, shiftlen);
3115 skb_frag_size_sub(fragfrom, shiftlen);
3116 fragfrom->page_offset += shiftlen;
3117
3118 goto onlymerged;
3119 }
3120
3121 from++;
3122 }
3123
3124 /* Skip full, not-fitting skb to avoid expensive operations */
3125 if ((shiftlen == skb->len) &&
3126 (skb_shinfo(skb)->nr_frags - from) > (MAX_SKB_FRAGS - to))
3127 return 0;
3128
3129 if (skb_prepare_for_shift(skb) || skb_prepare_for_shift(tgt))
3130 return 0;
3131
3132 while ((todo > 0) && (from < skb_shinfo(skb)->nr_frags)) {
3133 if (to == MAX_SKB_FRAGS)
3134 return 0;
3135
3136 fragfrom = &skb_shinfo(skb)->frags[from];
3137 fragto = &skb_shinfo(tgt)->frags[to];
3138
3139 if (todo >= skb_frag_size(fragfrom)) {
3140 *fragto = *fragfrom;
3141 todo -= skb_frag_size(fragfrom);
3142 from++;
3143 to++;
3144
3145 } else {
3146 __skb_frag_ref(fragfrom);
3147 fragto->page = fragfrom->page;
3148 fragto->page_offset = fragfrom->page_offset;
3149 skb_frag_size_set(fragto, todo);
3150
3151 fragfrom->page_offset += todo;
3152 skb_frag_size_sub(fragfrom, todo);
3153 todo = 0;
3154
3155 to++;
3156 break;
3157 }
3158 }
3159
3160 /* Ready to "commit" this state change to tgt */
3161 skb_shinfo(tgt)->nr_frags = to;
3162
3163 if (merge >= 0) {
3164 fragfrom = &skb_shinfo(skb)->frags[0];
3165 fragto = &skb_shinfo(tgt)->frags[merge];
3166
3167 skb_frag_size_add(fragto, skb_frag_size(fragfrom));
3168 __skb_frag_unref(fragfrom);
3169 }
3170
3171 /* Reposition in the original skb */
3172 to = 0;
3173 while (from < skb_shinfo(skb)->nr_frags)
3174 skb_shinfo(skb)->frags[to++] = skb_shinfo(skb)->frags[from++];
3175 skb_shinfo(skb)->nr_frags = to;
3176
3177 BUG_ON(todo > 0 && !skb_shinfo(skb)->nr_frags);
3178
3179onlymerged:
3180 /* Most likely the tgt won't ever need its checksum anymore, skb on
3181 * the other hand might need it if it needs to be resent
3182 */
3183 tgt->ip_summed = CHECKSUM_PARTIAL;
3184 skb->ip_summed = CHECKSUM_PARTIAL;
3185
3186 /* Yak, is it really working this way? Some helper please? */
3187 skb->len -= shiftlen;
3188 skb->data_len -= shiftlen;
3189 skb->truesize -= shiftlen;
3190 tgt->len += shiftlen;
3191 tgt->data_len += shiftlen;
3192 tgt->truesize += shiftlen;
3193
3194 return shiftlen;
3195}
3196
3197/**
3198 * skb_prepare_seq_read - Prepare a sequential read of skb data
3199 * @skb: the buffer to read
3200 * @from: lower offset of data to be read
3201 * @to: upper offset of data to be read
3202 * @st: state variable
3203 *
3204 * Initializes the specified state variable. Must be called before
3205 * invoking skb_seq_read() for the first time.
3206 */
3207void skb_prepare_seq_read(struct sk_buff *skb, unsigned int from,
3208 unsigned int to, struct skb_seq_state *st)
3209{
3210 st->lower_offset = from;
3211 st->upper_offset = to;
3212 st->root_skb = st->cur_skb = skb;
3213 st->frag_idx = st->stepped_offset = 0;
3214 st->frag_data = NULL;
3215}
3216EXPORT_SYMBOL(skb_prepare_seq_read);
3217
3218/**
3219 * skb_seq_read - Sequentially read skb data
3220 * @consumed: number of bytes consumed by the caller so far
3221 * @data: destination pointer for data to be returned
3222 * @st: state variable
3223 *
3224 * Reads a block of skb data at @consumed relative to the
3225 * lower offset specified to skb_prepare_seq_read(). Assigns
3226 * the head of the data block to @data and returns the length
3227 * of the block or 0 if the end of the skb data or the upper
3228 * offset has been reached.
3229 *
3230 * The caller is not required to consume all of the data
3231 * returned, i.e. @consumed is typically set to the number
3232 * of bytes already consumed and the next call to
3233 * skb_seq_read() will return the remaining part of the block.
3234 *
3235 * Note 1: The size of each block of data returned can be arbitrary,
3236 * this limitation is the cost for zerocopy sequential
3237 * reads of potentially non linear data.
3238 *
3239 * Note 2: Fragment lists within fragments are not implemented
3240 * at the moment, state->root_skb could be replaced with
3241 * a stack for this purpose.
3242 */
3243unsigned int skb_seq_read(unsigned int consumed, const u8 **data,
3244 struct skb_seq_state *st)
3245{
3246 unsigned int block_limit, abs_offset = consumed + st->lower_offset;
3247 skb_frag_t *frag;
3248
3249 if (unlikely(abs_offset >= st->upper_offset)) {
3250 if (st->frag_data) {
3251 kunmap_atomic(st->frag_data);
3252 st->frag_data = NULL;
3253 }
3254 return 0;
3255 }
3256
3257next_skb:
3258 block_limit = skb_headlen(st->cur_skb) + st->stepped_offset;
3259
3260 if (abs_offset < block_limit && !st->frag_data) {
3261 *data = st->cur_skb->data + (abs_offset - st->stepped_offset);
3262 return block_limit - abs_offset;
3263 }
3264
3265 if (st->frag_idx == 0 && !st->frag_data)
3266 st->stepped_offset += skb_headlen(st->cur_skb);
3267
3268 while (st->frag_idx < skb_shinfo(st->cur_skb)->nr_frags) {
3269 frag = &skb_shinfo(st->cur_skb)->frags[st->frag_idx];
3270 block_limit = skb_frag_size(frag) + st->stepped_offset;
3271
3272 if (abs_offset < block_limit) {
3273 if (!st->frag_data)
3274 st->frag_data = kmap_atomic(skb_frag_page(frag));
3275
3276 *data = (u8 *) st->frag_data + frag->page_offset +
3277 (abs_offset - st->stepped_offset);
3278
3279 return block_limit - abs_offset;
3280 }
3281
3282 if (st->frag_data) {
3283 kunmap_atomic(st->frag_data);
3284 st->frag_data = NULL;
3285 }
3286
3287 st->frag_idx++;
3288 st->stepped_offset += skb_frag_size(frag);
3289 }
3290
3291 if (st->frag_data) {
3292 kunmap_atomic(st->frag_data);
3293 st->frag_data = NULL;
3294 }
3295
3296 if (st->root_skb == st->cur_skb && skb_has_frag_list(st->root_skb)) {
3297 st->cur_skb = skb_shinfo(st->root_skb)->frag_list;
3298 st->frag_idx = 0;
3299 goto next_skb;
3300 } else if (st->cur_skb->next) {
3301 st->cur_skb = st->cur_skb->next;
3302 st->frag_idx = 0;
3303 goto next_skb;
3304 }
3305
3306 return 0;
3307}
3308EXPORT_SYMBOL(skb_seq_read);
3309
3310/**
3311 * skb_abort_seq_read - Abort a sequential read of skb data
3312 * @st: state variable
3313 *
3314 * Must be called if skb_seq_read() was not called until it
3315 * returned 0.
3316 */
3317void skb_abort_seq_read(struct skb_seq_state *st)
3318{
3319 if (st->frag_data)
3320 kunmap_atomic(st->frag_data);
3321}
3322EXPORT_SYMBOL(skb_abort_seq_read);
3323
3324#define TS_SKB_CB(state) ((struct skb_seq_state *) &((state)->cb))
3325
3326static unsigned int skb_ts_get_next_block(unsigned int offset, const u8 **text,
3327 struct ts_config *conf,
3328 struct ts_state *state)
3329{
3330 return skb_seq_read(offset, text, TS_SKB_CB(state));
3331}
3332
3333static void skb_ts_finish(struct ts_config *conf, struct ts_state *state)
3334{
3335 skb_abort_seq_read(TS_SKB_CB(state));
3336}
3337
3338/**
3339 * skb_find_text - Find a text pattern in skb data
3340 * @skb: the buffer to look in
3341 * @from: search offset
3342 * @to: search limit
3343 * @config: textsearch configuration
3344 *
3345 * Finds a pattern in the skb data according to the specified
3346 * textsearch configuration. Use textsearch_next() to retrieve
3347 * subsequent occurrences of the pattern. Returns the offset
3348 * to the first occurrence or UINT_MAX if no match was found.
3349 */
3350unsigned int skb_find_text(struct sk_buff *skb, unsigned int from,
3351 unsigned int to, struct ts_config *config)
3352{
3353 struct ts_state state;
3354 unsigned int ret;
3355
3356 config->get_next_block = skb_ts_get_next_block;
3357 config->finish = skb_ts_finish;
3358
3359 skb_prepare_seq_read(skb, from, to, TS_SKB_CB(&state));
3360
3361 ret = textsearch_find(config, &state);
3362 return (ret <= to - from ? ret : UINT_MAX);
3363}
3364EXPORT_SYMBOL(skb_find_text);
3365
3366/**
3367 * skb_append_datato_frags - append the user data to a skb
3368 * @sk: sock structure
3369 * @skb: skb structure to be appended with user data.
3370 * @getfrag: call back function to be used for getting the user data
3371 * @from: pointer to user message iov
3372 * @length: length of the iov message
3373 *
3374 * Description: This procedure append the user data in the fragment part
3375 * of the skb if any page alloc fails user this procedure returns -ENOMEM
3376 */
3377int skb_append_datato_frags(struct sock *sk, struct sk_buff *skb,
3378 int (*getfrag)(void *from, char *to, int offset,
3379 int len, int odd, struct sk_buff *skb),
3380 void *from, int length)
3381{
3382 int frg_cnt = skb_shinfo(skb)->nr_frags;
3383 int copy;
3384 int offset = 0;
3385 int ret;
3386 struct page_frag *pfrag = ¤t->task_frag;
3387
3388 do {
3389 /* Return error if we don't have space for new frag */
3390 if (frg_cnt >= MAX_SKB_FRAGS)
3391 return -EMSGSIZE;
3392
3393 if (!sk_page_frag_refill(sk, pfrag))
3394 return -ENOMEM;
3395
3396 /* copy the user data to page */
3397 copy = min_t(int, length, pfrag->size - pfrag->offset);
3398
3399 ret = getfrag(from, page_address(pfrag->page) + pfrag->offset,
3400 offset, copy, 0, skb);
3401 if (ret < 0)
3402 return -EFAULT;
3403
3404 /* copy was successful so update the size parameters */
3405 skb_fill_page_desc(skb, frg_cnt, pfrag->page, pfrag->offset,
3406 copy);
3407 frg_cnt++;
3408 pfrag->offset += copy;
3409 get_page(pfrag->page);
3410
3411 skb->truesize += copy;
3412 refcount_add(copy, &sk->sk_wmem_alloc);
3413 skb->len += copy;
3414 skb->data_len += copy;
3415 offset += copy;
3416 length -= copy;
3417
3418 } while (length > 0);
3419
3420 return 0;
3421}
3422EXPORT_SYMBOL(skb_append_datato_frags);
3423
3424int skb_append_pagefrags(struct sk_buff *skb, struct page *page,
3425 int offset, size_t size)
3426{
3427 int i = skb_shinfo(skb)->nr_frags;
3428
3429 if (skb_can_coalesce(skb, i, page, offset)) {
3430 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
3431 } else if (i < MAX_SKB_FRAGS) {
3432 get_page(page);
3433 skb_fill_page_desc(skb, i, page, offset, size);
3434 } else {
3435 return -EMSGSIZE;
3436 }
3437
3438 return 0;
3439}
3440EXPORT_SYMBOL_GPL(skb_append_pagefrags);
3441
3442/**
3443 * skb_pull_rcsum - pull skb and update receive checksum
3444 * @skb: buffer to update
3445 * @len: length of data pulled
3446 *
3447 * This function performs an skb_pull on the packet and updates
3448 * the CHECKSUM_COMPLETE checksum. It should be used on
3449 * receive path processing instead of skb_pull unless you know
3450 * that the checksum difference is zero (e.g., a valid IP header)
3451 * or you are setting ip_summed to CHECKSUM_NONE.
3452 */
3453void *skb_pull_rcsum(struct sk_buff *skb, unsigned int len)
3454{
3455 unsigned char *data = skb->data;
3456
3457 BUG_ON(len > skb->len);
3458 __skb_pull(skb, len);
3459 skb_postpull_rcsum(skb, data, len);
3460 return skb->data;
3461}
3462EXPORT_SYMBOL_GPL(skb_pull_rcsum);
3463
3464static inline skb_frag_t skb_head_frag_to_page_desc(struct sk_buff *frag_skb)
3465{
3466 skb_frag_t head_frag;
3467 struct page *page;
3468
3469 page = virt_to_head_page(frag_skb->head);
3470 head_frag.page.p = page;
3471 head_frag.page_offset = frag_skb->data -
3472 (unsigned char *)page_address(page);
3473 head_frag.size = skb_headlen(frag_skb);
3474 return head_frag;
3475}
3476
3477/**
3478 * skb_segment - Perform protocol segmentation on skb.
3479 * @head_skb: buffer to segment
3480 * @features: features for the output path (see dev->features)
3481 *
3482 * This function performs segmentation on the given skb. It returns
3483 * a pointer to the first in a list of new skbs for the segments.
3484 * In case of error it returns ERR_PTR(err).
3485 */
3486struct sk_buff *skb_segment(struct sk_buff *head_skb,
3487 netdev_features_t features)
3488{
3489 struct sk_buff *segs = NULL;
3490 struct sk_buff *tail = NULL;
3491 struct sk_buff *list_skb = skb_shinfo(head_skb)->frag_list;
3492 skb_frag_t *frag = skb_shinfo(head_skb)->frags;
3493 unsigned int mss = skb_shinfo(head_skb)->gso_size;
3494 unsigned int doffset = head_skb->data - skb_mac_header(head_skb);
3495 struct sk_buff *frag_skb = head_skb;
3496 unsigned int offset = doffset;
3497 unsigned int tnl_hlen = skb_tnl_header_len(head_skb);
3498 unsigned int partial_segs = 0;
3499 unsigned int headroom;
3500 unsigned int len = head_skb->len;
3501 __be16 proto;
3502 bool csum, sg;
3503 int nfrags = skb_shinfo(head_skb)->nr_frags;
3504 int err = -ENOMEM;
3505 int i = 0;
3506 int pos;
3507 int dummy;
3508
3509 __skb_push(head_skb, doffset);
3510 proto = skb_network_protocol(head_skb, &dummy);
3511 if (unlikely(!proto))
3512 return ERR_PTR(-EINVAL);
3513
3514 sg = !!(features & NETIF_F_SG);
3515 csum = !!can_checksum_protocol(features, proto);
3516
3517 if (sg && csum && (mss != GSO_BY_FRAGS)) {
3518 if (!(features & NETIF_F_GSO_PARTIAL)) {
3519 struct sk_buff *iter;
3520 unsigned int frag_len;
3521
3522 if (!list_skb ||
3523 !net_gso_ok(features, skb_shinfo(head_skb)->gso_type))
3524 goto normal;
3525
3526 /* If we get here then all the required
3527 * GSO features except frag_list are supported.
3528 * Try to split the SKB to multiple GSO SKBs
3529 * with no frag_list.
3530 * Currently we can do that only when the buffers don't
3531 * have a linear part and all the buffers except
3532 * the last are of the same length.
3533 */
3534 frag_len = list_skb->len;
3535 skb_walk_frags(head_skb, iter) {
3536 if (frag_len != iter->len && iter->next)
3537 goto normal;
3538 if (skb_headlen(iter) && !iter->head_frag)
3539 goto normal;
3540
3541 len -= iter->len;
3542 }
3543
3544 if (len != frag_len)
3545 goto normal;
3546 }
3547
3548 /* GSO partial only requires that we trim off any excess that
3549 * doesn't fit into an MSS sized block, so take care of that
3550 * now.
3551 */
3552 partial_segs = len / mss;
3553 if (partial_segs > 1)
3554 mss *= partial_segs;
3555 else
3556 partial_segs = 0;
3557 }
3558
3559normal:
3560 headroom = skb_headroom(head_skb);
3561 pos = skb_headlen(head_skb);
3562
3563 do {
3564 struct sk_buff *nskb;
3565 skb_frag_t *nskb_frag;
3566 int hsize;
3567 int size;
3568
3569 if (unlikely(mss == GSO_BY_FRAGS)) {
3570 len = list_skb->len;
3571 } else {
3572 len = head_skb->len - offset;
3573 if (len > mss)
3574 len = mss;
3575 }
3576
3577 hsize = skb_headlen(head_skb) - offset;
3578 if (hsize < 0)
3579 hsize = 0;
3580 if (hsize > len || !sg)
3581 hsize = len;
3582
3583 if (!hsize && i >= nfrags && skb_headlen(list_skb) &&
3584 (skb_headlen(list_skb) == len || sg)) {
3585 BUG_ON(skb_headlen(list_skb) > len);
3586
3587 i = 0;
3588 nfrags = skb_shinfo(list_skb)->nr_frags;
3589 frag = skb_shinfo(list_skb)->frags;
3590 frag_skb = list_skb;
3591 pos += skb_headlen(list_skb);
3592
3593 while (pos < offset + len) {
3594 BUG_ON(i >= nfrags);
3595
3596 size = skb_frag_size(frag);
3597 if (pos + size > offset + len)
3598 break;
3599
3600 i++;
3601 pos += size;
3602 frag++;
3603 }
3604
3605 nskb = skb_clone(list_skb, GFP_ATOMIC);
3606 list_skb = list_skb->next;
3607
3608 if (unlikely(!nskb))
3609 goto err;
3610
3611 if (unlikely(pskb_trim(nskb, len))) {
3612 kfree_skb(nskb);
3613 goto err;
3614 }
3615
3616 hsize = skb_end_offset(nskb);
3617 if (skb_cow_head(nskb, doffset + headroom)) {
3618 kfree_skb(nskb);
3619 goto err;
3620 }
3621
3622 nskb->truesize += skb_end_offset(nskb) - hsize;
3623 skb_release_head_state(nskb);
3624 __skb_push(nskb, doffset);
3625 } else {
3626 nskb = __alloc_skb(hsize + doffset + headroom,
3627 GFP_ATOMIC, skb_alloc_rx_flag(head_skb),
3628 NUMA_NO_NODE);
3629
3630 if (unlikely(!nskb))
3631 goto err;
3632
3633 skb_reserve(nskb, headroom);
3634 __skb_put(nskb, doffset);
3635 }
3636
3637 if (segs)
3638 tail->next = nskb;
3639 else
3640 segs = nskb;
3641 tail = nskb;
3642
3643 __copy_skb_header(nskb, head_skb);
3644
3645 skb_headers_offset_update(nskb, skb_headroom(nskb) - headroom);
3646 skb_reset_mac_len(nskb);
3647
3648 skb_copy_from_linear_data_offset(head_skb, -tnl_hlen,
3649 nskb->data - tnl_hlen,
3650 doffset + tnl_hlen);
3651
3652 if (nskb->len == len + doffset)
3653 goto perform_csum_check;
3654
3655 if (!sg) {
3656 if (!nskb->remcsum_offload)
3657 nskb->ip_summed = CHECKSUM_NONE;
3658 SKB_GSO_CB(nskb)->csum =
3659 skb_copy_and_csum_bits(head_skb, offset,
3660 skb_put(nskb, len),
3661 len, 0);
3662 SKB_GSO_CB(nskb)->csum_start =
3663 skb_headroom(nskb) + doffset;
3664 continue;
3665 }
3666
3667 nskb_frag = skb_shinfo(nskb)->frags;
3668
3669 skb_copy_from_linear_data_offset(head_skb, offset,
3670 skb_put(nskb, hsize), hsize);
3671
3672 skb_shinfo(nskb)->tx_flags |= skb_shinfo(head_skb)->tx_flags &
3673 SKBTX_SHARED_FRAG;
3674
3675 if (skb_orphan_frags(frag_skb, GFP_ATOMIC) ||
3676 skb_zerocopy_clone(nskb, frag_skb, GFP_ATOMIC))
3677 goto err;
3678
3679 while (pos < offset + len) {
3680 if (i >= nfrags) {
3681 i = 0;
3682 nfrags = skb_shinfo(list_skb)->nr_frags;
3683 frag = skb_shinfo(list_skb)->frags;
3684 frag_skb = list_skb;
3685 if (!skb_headlen(list_skb)) {
3686 BUG_ON(!nfrags);
3687 } else {
3688 BUG_ON(!list_skb->head_frag);
3689
3690 /* to make room for head_frag. */
3691 i--;
3692 frag--;
3693 }
3694 if (skb_orphan_frags(frag_skb, GFP_ATOMIC) ||
3695 skb_zerocopy_clone(nskb, frag_skb,
3696 GFP_ATOMIC))
3697 goto err;
3698
3699 list_skb = list_skb->next;
3700 }
3701
3702 if (unlikely(skb_shinfo(nskb)->nr_frags >=
3703 MAX_SKB_FRAGS)) {
3704 net_warn_ratelimited(
3705 "skb_segment: too many frags: %u %u\n",
3706 pos, mss);
3707 goto err;
3708 }
3709
3710 *nskb_frag = (i < 0) ? skb_head_frag_to_page_desc(frag_skb) : *frag;
3711 __skb_frag_ref(nskb_frag);
3712 size = skb_frag_size(nskb_frag);
3713
3714 if (pos < offset) {
3715 nskb_frag->page_offset += offset - pos;
3716 skb_frag_size_sub(nskb_frag, offset - pos);
3717 }
3718
3719 skb_shinfo(nskb)->nr_frags++;
3720
3721 if (pos + size <= offset + len) {
3722 i++;
3723 frag++;
3724 pos += size;
3725 } else {
3726 skb_frag_size_sub(nskb_frag, pos + size - (offset + len));
3727 goto skip_fraglist;
3728 }
3729
3730 nskb_frag++;
3731 }
3732
3733skip_fraglist:
3734 nskb->data_len = len - hsize;
3735 nskb->len += nskb->data_len;
3736 nskb->truesize += nskb->data_len;
3737
3738perform_csum_check:
3739 if (!csum) {
3740 if (skb_has_shared_frag(nskb)) {
3741 err = __skb_linearize(nskb);
3742 if (err)
3743 goto err;
3744 }
3745 if (!nskb->remcsum_offload)
3746 nskb->ip_summed = CHECKSUM_NONE;
3747 SKB_GSO_CB(nskb)->csum =
3748 skb_checksum(nskb, doffset,
3749 nskb->len - doffset, 0);
3750 SKB_GSO_CB(nskb)->csum_start =
3751 skb_headroom(nskb) + doffset;
3752 }
3753 } while ((offset += len) < head_skb->len);
3754
3755 /* Some callers want to get the end of the list.
3756 * Put it in segs->prev to avoid walking the list.
3757 * (see validate_xmit_skb_list() for example)
3758 */
3759 segs->prev = tail;
3760
3761 if (partial_segs) {
3762 struct sk_buff *iter;
3763 int type = skb_shinfo(head_skb)->gso_type;
3764 unsigned short gso_size = skb_shinfo(head_skb)->gso_size;
3765
3766 /* Update type to add partial and then remove dodgy if set */
3767 type |= (features & NETIF_F_GSO_PARTIAL) / NETIF_F_GSO_PARTIAL * SKB_GSO_PARTIAL;
3768 type &= ~SKB_GSO_DODGY;
3769
3770 /* Update GSO info and prepare to start updating headers on
3771 * our way back down the stack of protocols.
3772 */
3773 for (iter = segs; iter; iter = iter->next) {
3774 skb_shinfo(iter)->gso_size = gso_size;
3775 skb_shinfo(iter)->gso_segs = partial_segs;
3776 skb_shinfo(iter)->gso_type = type;
3777 SKB_GSO_CB(iter)->data_offset = skb_headroom(iter) + doffset;
3778 }
3779
3780 if (tail->len - doffset <= gso_size)
3781 skb_shinfo(tail)->gso_size = 0;
3782 else if (tail != segs)
3783 skb_shinfo(tail)->gso_segs = DIV_ROUND_UP(tail->len - doffset, gso_size);
3784 }
3785
3786 /* Following permits correct backpressure, for protocols
3787 * using skb_set_owner_w().
3788 * Idea is to tranfert ownership from head_skb to last segment.
3789 */
3790 if (head_skb->destructor == sock_wfree) {
3791 swap(tail->truesize, head_skb->truesize);
3792 swap(tail->destructor, head_skb->destructor);
3793 swap(tail->sk, head_skb->sk);
3794 }
3795 return segs;
3796
3797err:
3798 kfree_skb_list(segs);
3799 return ERR_PTR(err);
3800}
3801EXPORT_SYMBOL_GPL(skb_segment);
3802
3803int skb_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3804{
3805 struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
3806 unsigned int offset = skb_gro_offset(skb);
3807 unsigned int headlen = skb_headlen(skb);
3808 unsigned int len = skb_gro_len(skb);
3809 struct sk_buff *lp, *p = *head;
3810 unsigned int delta_truesize;
3811
3812 if (unlikely(p->len + len >= 65536))
3813 return -E2BIG;
3814
3815 lp = NAPI_GRO_CB(p)->last;
3816 pinfo = skb_shinfo(lp);
3817
3818 if (headlen <= offset) {
3819 skb_frag_t *frag;
3820 skb_frag_t *frag2;
3821 int i = skbinfo->nr_frags;
3822 int nr_frags = pinfo->nr_frags + i;
3823
3824 if (nr_frags > MAX_SKB_FRAGS)
3825 goto merge;
3826
3827 offset -= headlen;
3828 pinfo->nr_frags = nr_frags;
3829 skbinfo->nr_frags = 0;
3830
3831 frag = pinfo->frags + nr_frags;
3832 frag2 = skbinfo->frags + i;
3833 do {
3834 *--frag = *--frag2;
3835 } while (--i);
3836
3837 frag->page_offset += offset;
3838 skb_frag_size_sub(frag, offset);
3839
3840 /* all fragments truesize : remove (head size + sk_buff) */
3841 delta_truesize = skb->truesize -
3842 SKB_TRUESIZE(skb_end_offset(skb));
3843
3844 skb->truesize -= skb->data_len;
3845 skb->len -= skb->data_len;
3846 skb->data_len = 0;
3847
3848 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
3849 goto done;
3850 } else if (skb->head_frag) {
3851 int nr_frags = pinfo->nr_frags;
3852 skb_frag_t *frag = pinfo->frags + nr_frags;
3853 struct page *page = virt_to_head_page(skb->head);
3854 unsigned int first_size = headlen - offset;
3855 unsigned int first_offset;
3856
3857 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
3858 goto merge;
3859
3860 first_offset = skb->data -
3861 (unsigned char *)page_address(page) +
3862 offset;
3863
3864 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
3865
3866 frag->page.p = page;
3867 frag->page_offset = first_offset;
3868 skb_frag_size_set(frag, first_size);
3869
3870 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
3871 /* We dont need to clear skbinfo->nr_frags here */
3872
3873 delta_truesize = skb->truesize - SKB_DATA_ALIGN(sizeof(struct sk_buff));
3874 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
3875 goto done;
3876 }
3877
3878merge:
3879 delta_truesize = skb->truesize;
3880 if (offset > headlen) {
3881 unsigned int eat = offset - headlen;
3882
3883 skbinfo->frags[0].page_offset += eat;
3884 skb_frag_size_sub(&skbinfo->frags[0], eat);
3885 skb->data_len -= eat;
3886 skb->len -= eat;
3887 offset = headlen;
3888 }
3889
3890 __skb_pull(skb, offset);
3891
3892 if (NAPI_GRO_CB(p)->last == p)
3893 skb_shinfo(p)->frag_list = skb;
3894 else
3895 NAPI_GRO_CB(p)->last->next = skb;
3896 NAPI_GRO_CB(p)->last = skb;
3897 __skb_header_release(skb);
3898 lp = p;
3899
3900done:
3901 NAPI_GRO_CB(p)->count++;
3902 p->data_len += len;
3903 p->truesize += delta_truesize;
3904 p->len += len;
3905 if (lp != p) {
3906 lp->data_len += len;
3907 lp->truesize += delta_truesize;
3908 lp->len += len;
3909 }
3910 NAPI_GRO_CB(skb)->same_flow = 1;
3911 return 0;
3912}
3913EXPORT_SYMBOL_GPL(skb_gro_receive);
3914
3915void __init skb_init(void)
3916{
3917 skbuff_head_cache = kmem_cache_create_usercopy("skbuff_head_cache",
3918 sizeof(struct sk_buff),
3919 0,
3920 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
3921 offsetof(struct sk_buff, cb),
3922 sizeof_field(struct sk_buff, cb),
3923 NULL);
3924 skbuff_fclone_cache = kmem_cache_create("skbuff_fclone_cache",
3925 sizeof(struct sk_buff_fclones),
3926 0,
3927 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
3928 NULL);
3929}
3930
3931static int
3932__skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len,
3933 unsigned int recursion_level)
3934{
3935 int start = skb_headlen(skb);
3936 int i, copy = start - offset;
3937 struct sk_buff *frag_iter;
3938 int elt = 0;
3939
3940 if (unlikely(recursion_level >= 24))
3941 return -EMSGSIZE;
3942
3943 if (copy > 0) {
3944 if (copy > len)
3945 copy = len;
3946 sg_set_buf(sg, skb->data + offset, copy);
3947 elt++;
3948 if ((len -= copy) == 0)
3949 return elt;
3950 offset += copy;
3951 }
3952
3953 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3954 int end;
3955
3956 WARN_ON(start > offset + len);
3957
3958 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
3959 if ((copy = end - offset) > 0) {
3960 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3961 if (unlikely(elt && sg_is_last(&sg[elt - 1])))
3962 return -EMSGSIZE;
3963
3964 if (copy > len)
3965 copy = len;
3966 sg_set_page(&sg[elt], skb_frag_page(frag), copy,
3967 frag->page_offset+offset-start);
3968 elt++;
3969 if (!(len -= copy))
3970 return elt;
3971 offset += copy;
3972 }
3973 start = end;
3974 }
3975
3976 skb_walk_frags(skb, frag_iter) {
3977 int end, ret;
3978
3979 WARN_ON(start > offset + len);
3980
3981 end = start + frag_iter->len;
3982 if ((copy = end - offset) > 0) {
3983 if (unlikely(elt && sg_is_last(&sg[elt - 1])))
3984 return -EMSGSIZE;
3985
3986 if (copy > len)
3987 copy = len;
3988 ret = __skb_to_sgvec(frag_iter, sg+elt, offset - start,
3989 copy, recursion_level + 1);
3990 if (unlikely(ret < 0))
3991 return ret;
3992 elt += ret;
3993 if ((len -= copy) == 0)
3994 return elt;
3995 offset += copy;
3996 }
3997 start = end;
3998 }
3999 BUG_ON(len);
4000 return elt;
4001}
4002
4003/**
4004 * skb_to_sgvec - Fill a scatter-gather list from a socket buffer
4005 * @skb: Socket buffer containing the buffers to be mapped
4006 * @sg: The scatter-gather list to map into
4007 * @offset: The offset into the buffer's contents to start mapping
4008 * @len: Length of buffer space to be mapped
4009 *
4010 * Fill the specified scatter-gather list with mappings/pointers into a
4011 * region of the buffer space attached to a socket buffer. Returns either
4012 * the number of scatterlist items used, or -EMSGSIZE if the contents
4013 * could not fit.
4014 */
4015int skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len)
4016{
4017 int nsg = __skb_to_sgvec(skb, sg, offset, len, 0);
4018
4019 if (nsg <= 0)
4020 return nsg;
4021
4022 sg_mark_end(&sg[nsg - 1]);
4023
4024 return nsg;
4025}
4026EXPORT_SYMBOL_GPL(skb_to_sgvec);
4027
4028/* As compared with skb_to_sgvec, skb_to_sgvec_nomark only map skb to given
4029 * sglist without mark the sg which contain last skb data as the end.
4030 * So the caller can mannipulate sg list as will when padding new data after
4031 * the first call without calling sg_unmark_end to expend sg list.
4032 *
4033 * Scenario to use skb_to_sgvec_nomark:
4034 * 1. sg_init_table
4035 * 2. skb_to_sgvec_nomark(payload1)
4036 * 3. skb_to_sgvec_nomark(payload2)
4037 *
4038 * This is equivalent to:
4039 * 1. sg_init_table
4040 * 2. skb_to_sgvec(payload1)
4041 * 3. sg_unmark_end
4042 * 4. skb_to_sgvec(payload2)
4043 *
4044 * When mapping mutilple payload conditionally, skb_to_sgvec_nomark
4045 * is more preferable.
4046 */
4047int skb_to_sgvec_nomark(struct sk_buff *skb, struct scatterlist *sg,
4048 int offset, int len)
4049{
4050 return __skb_to_sgvec(skb, sg, offset, len, 0);
4051}
4052EXPORT_SYMBOL_GPL(skb_to_sgvec_nomark);
4053
4054
4055
4056/**
4057 * skb_cow_data - Check that a socket buffer's data buffers are writable
4058 * @skb: The socket buffer to check.
4059 * @tailbits: Amount of trailing space to be added
4060 * @trailer: Returned pointer to the skb where the @tailbits space begins
4061 *
4062 * Make sure that the data buffers attached to a socket buffer are
4063 * writable. If they are not, private copies are made of the data buffers
4064 * and the socket buffer is set to use these instead.
4065 *
4066 * If @tailbits is given, make sure that there is space to write @tailbits
4067 * bytes of data beyond current end of socket buffer. @trailer will be
4068 * set to point to the skb in which this space begins.
4069 *
4070 * The number of scatterlist elements required to completely map the
4071 * COW'd and extended socket buffer will be returned.
4072 */
4073int skb_cow_data(struct sk_buff *skb, int tailbits, struct sk_buff **trailer)
4074{
4075 int copyflag;
4076 int elt;
4077 struct sk_buff *skb1, **skb_p;
4078
4079 /* If skb is cloned or its head is paged, reallocate
4080 * head pulling out all the pages (pages are considered not writable
4081 * at the moment even if they are anonymous).
4082 */
4083 if ((skb_cloned(skb) || skb_shinfo(skb)->nr_frags) &&
4084 __pskb_pull_tail(skb, skb_pagelen(skb)-skb_headlen(skb)) == NULL)
4085 return -ENOMEM;
4086
4087 /* Easy case. Most of packets will go this way. */
4088 if (!skb_has_frag_list(skb)) {
4089 /* A little of trouble, not enough of space for trailer.
4090 * This should not happen, when stack is tuned to generate
4091 * good frames. OK, on miss we reallocate and reserve even more
4092 * space, 128 bytes is fair. */
4093
4094 if (skb_tailroom(skb) < tailbits &&
4095 pskb_expand_head(skb, 0, tailbits-skb_tailroom(skb)+128, GFP_ATOMIC))
4096 return -ENOMEM;
4097
4098 /* Voila! */
4099 *trailer = skb;
4100 return 1;
4101 }
4102
4103 /* Misery. We are in troubles, going to mincer fragments... */
4104
4105 elt = 1;
4106 skb_p = &skb_shinfo(skb)->frag_list;
4107 copyflag = 0;
4108
4109 while ((skb1 = *skb_p) != NULL) {
4110 int ntail = 0;
4111
4112 /* The fragment is partially pulled by someone,
4113 * this can happen on input. Copy it and everything
4114 * after it. */
4115
4116 if (skb_shared(skb1))
4117 copyflag = 1;
4118
4119 /* If the skb is the last, worry about trailer. */
4120
4121 if (skb1->next == NULL && tailbits) {
4122 if (skb_shinfo(skb1)->nr_frags ||
4123 skb_has_frag_list(skb1) ||
4124 skb_tailroom(skb1) < tailbits)
4125 ntail = tailbits + 128;
4126 }
4127
4128 if (copyflag ||
4129 skb_cloned(skb1) ||
4130 ntail ||
4131 skb_shinfo(skb1)->nr_frags ||
4132 skb_has_frag_list(skb1)) {
4133 struct sk_buff *skb2;
4134
4135 /* Fuck, we are miserable poor guys... */
4136 if (ntail == 0)
4137 skb2 = skb_copy(skb1, GFP_ATOMIC);
4138 else
4139 skb2 = skb_copy_expand(skb1,
4140 skb_headroom(skb1),
4141 ntail,
4142 GFP_ATOMIC);
4143 if (unlikely(skb2 == NULL))
4144 return -ENOMEM;
4145
4146 if (skb1->sk)
4147 skb_set_owner_w(skb2, skb1->sk);
4148
4149 /* Looking around. Are we still alive?
4150 * OK, link new skb, drop old one */
4151
4152 skb2->next = skb1->next;
4153 *skb_p = skb2;
4154 kfree_skb(skb1);
4155 skb1 = skb2;
4156 }
4157 elt++;
4158 *trailer = skb1;
4159 skb_p = &skb1->next;
4160 }
4161
4162 return elt;
4163}
4164EXPORT_SYMBOL_GPL(skb_cow_data);
4165
4166static void sock_rmem_free(struct sk_buff *skb)
4167{
4168 struct sock *sk = skb->sk;
4169
4170 atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
4171}
4172
4173static void skb_set_err_queue(struct sk_buff *skb)
4174{
4175 /* pkt_type of skbs received on local sockets is never PACKET_OUTGOING.
4176 * So, it is safe to (mis)use it to mark skbs on the error queue.
4177 */
4178 skb->pkt_type = PACKET_OUTGOING;
4179 BUILD_BUG_ON(PACKET_OUTGOING == 0);
4180}
4181
4182/*
4183 * Note: We dont mem charge error packets (no sk_forward_alloc changes)
4184 */
4185int sock_queue_err_skb(struct sock *sk, struct sk_buff *skb)
4186{
4187 if (atomic_read(&sk->sk_rmem_alloc) + skb->truesize >=
4188 (unsigned int)sk->sk_rcvbuf)
4189 return -ENOMEM;
4190
4191 skb_orphan(skb);
4192 skb->sk = sk;
4193 skb->destructor = sock_rmem_free;
4194 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
4195 skb_set_err_queue(skb);
4196
4197 /* before exiting rcu section, make sure dst is refcounted */
4198 skb_dst_force(skb);
4199
4200 skb_queue_tail(&sk->sk_error_queue, skb);
4201 if (!sock_flag(sk, SOCK_DEAD))
4202 sk->sk_error_report(sk);
4203 return 0;
4204}
4205EXPORT_SYMBOL(sock_queue_err_skb);
4206
4207static bool is_icmp_err_skb(const struct sk_buff *skb)
4208{
4209 return skb && (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP ||
4210 SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP6);
4211}
4212
4213struct sk_buff *sock_dequeue_err_skb(struct sock *sk)
4214{
4215 struct sk_buff_head *q = &sk->sk_error_queue;
4216 struct sk_buff *skb, *skb_next = NULL;
4217 bool icmp_next = false;
4218 unsigned long flags;
4219
4220 spin_lock_irqsave(&q->lock, flags);
4221 skb = __skb_dequeue(q);
4222 if (skb && (skb_next = skb_peek(q))) {
4223 icmp_next = is_icmp_err_skb(skb_next);
4224 if (icmp_next)
4225 sk->sk_err = SKB_EXT_ERR(skb_next)->ee.ee_origin;
4226 }
4227 spin_unlock_irqrestore(&q->lock, flags);
4228
4229 if (is_icmp_err_skb(skb) && !icmp_next)
4230 sk->sk_err = 0;
4231
4232 if (skb_next)
4233 sk->sk_error_report(sk);
4234
4235 return skb;
4236}
4237EXPORT_SYMBOL(sock_dequeue_err_skb);
4238
4239/**
4240 * skb_clone_sk - create clone of skb, and take reference to socket
4241 * @skb: the skb to clone
4242 *
4243 * This function creates a clone of a buffer that holds a reference on
4244 * sk_refcnt. Buffers created via this function are meant to be
4245 * returned using sock_queue_err_skb, or free via kfree_skb.
4246 *
4247 * When passing buffers allocated with this function to sock_queue_err_skb
4248 * it is necessary to wrap the call with sock_hold/sock_put in order to
4249 * prevent the socket from being released prior to being enqueued on
4250 * the sk_error_queue.
4251 */
4252struct sk_buff *skb_clone_sk(struct sk_buff *skb)
4253{
4254 struct sock *sk = skb->sk;
4255 struct sk_buff *clone;
4256
4257 if (!sk || !refcount_inc_not_zero(&sk->sk_refcnt))
4258 return NULL;
4259
4260 clone = skb_clone(skb, GFP_ATOMIC);
4261 if (!clone) {
4262 sock_put(sk);
4263 return NULL;
4264 }
4265
4266 clone->sk = sk;
4267 clone->destructor = sock_efree;
4268
4269 return clone;
4270}
4271EXPORT_SYMBOL(skb_clone_sk);
4272
4273static void __skb_complete_tx_timestamp(struct sk_buff *skb,
4274 struct sock *sk,
4275 int tstype,
4276 bool opt_stats)
4277{
4278 struct sock_exterr_skb *serr;
4279 int err;
4280
4281 BUILD_BUG_ON(sizeof(struct sock_exterr_skb) > sizeof(skb->cb));
4282
4283 serr = SKB_EXT_ERR(skb);
4284 memset(serr, 0, sizeof(*serr));
4285 serr->ee.ee_errno = ENOMSG;
4286 serr->ee.ee_origin = SO_EE_ORIGIN_TIMESTAMPING;
4287 serr->ee.ee_info = tstype;
4288 serr->opt_stats = opt_stats;
4289 serr->header.h4.iif = skb->dev ? skb->dev->ifindex : 0;
4290 if (sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID) {
4291 serr->ee.ee_data = skb_shinfo(skb)->tskey;
4292 if (sk->sk_protocol == IPPROTO_TCP &&
4293 sk->sk_type == SOCK_STREAM)
4294 serr->ee.ee_data -= sk->sk_tskey;
4295 }
4296
4297 err = sock_queue_err_skb(sk, skb);
4298
4299 if (err)
4300 kfree_skb(skb);
4301}
4302
4303static bool skb_may_tx_timestamp(struct sock *sk, bool tsonly)
4304{
4305 bool ret;
4306
4307 if (likely(sysctl_tstamp_allow_data || tsonly))
4308 return true;
4309
4310 read_lock_bh(&sk->sk_callback_lock);
4311 ret = sk->sk_socket && sk->sk_socket->file &&
4312 file_ns_capable(sk->sk_socket->file, &init_user_ns, CAP_NET_RAW);
4313 read_unlock_bh(&sk->sk_callback_lock);
4314 return ret;
4315}
4316
4317void skb_complete_tx_timestamp(struct sk_buff *skb,
4318 struct skb_shared_hwtstamps *hwtstamps)
4319{
4320 struct sock *sk = skb->sk;
4321
4322 if (!skb_may_tx_timestamp(sk, false))
4323 goto err;
4324
4325 /* Take a reference to prevent skb_orphan() from freeing the socket,
4326 * but only if the socket refcount is not zero.
4327 */
4328 if (likely(refcount_inc_not_zero(&sk->sk_refcnt))) {
4329 *skb_hwtstamps(skb) = *hwtstamps;
4330 __skb_complete_tx_timestamp(skb, sk, SCM_TSTAMP_SND, false);
4331 sock_put(sk);
4332 return;
4333 }
4334
4335err:
4336 kfree_skb(skb);
4337}
4338EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp);
4339
4340void __skb_tstamp_tx(struct sk_buff *orig_skb,
4341 struct skb_shared_hwtstamps *hwtstamps,
4342 struct sock *sk, int tstype)
4343{
4344 struct sk_buff *skb;
4345 bool tsonly, opt_stats = false;
4346
4347 if (!sk)
4348 return;
4349
4350 if (!hwtstamps && !(sk->sk_tsflags & SOF_TIMESTAMPING_OPT_TX_SWHW) &&
4351 skb_shinfo(orig_skb)->tx_flags & SKBTX_IN_PROGRESS)
4352 return;
4353
4354 tsonly = sk->sk_tsflags & SOF_TIMESTAMPING_OPT_TSONLY;
4355 if (!skb_may_tx_timestamp(sk, tsonly))
4356 return;
4357
4358 if (tsonly) {
4359#ifdef CONFIG_INET
4360 if ((sk->sk_tsflags & SOF_TIMESTAMPING_OPT_STATS) &&
4361 sk->sk_protocol == IPPROTO_TCP &&
4362 sk->sk_type == SOCK_STREAM) {
4363 skb = tcp_get_timestamping_opt_stats(sk);
4364 opt_stats = true;
4365 } else
4366#endif
4367 skb = alloc_skb(0, GFP_ATOMIC);
4368 } else {
4369 skb = skb_clone(orig_skb, GFP_ATOMIC);
4370 }
4371 if (!skb)
4372 return;
4373
4374 if (tsonly) {
4375 skb_shinfo(skb)->tx_flags |= skb_shinfo(orig_skb)->tx_flags &
4376 SKBTX_ANY_TSTAMP;
4377 skb_shinfo(skb)->tskey = skb_shinfo(orig_skb)->tskey;
4378 }
4379
4380 if (hwtstamps)
4381 *skb_hwtstamps(skb) = *hwtstamps;
4382 else
4383 skb->tstamp = ktime_get_real();
4384
4385 __skb_complete_tx_timestamp(skb, sk, tstype, opt_stats);
4386}
4387EXPORT_SYMBOL_GPL(__skb_tstamp_tx);
4388
4389void skb_tstamp_tx(struct sk_buff *orig_skb,
4390 struct skb_shared_hwtstamps *hwtstamps)
4391{
4392 return __skb_tstamp_tx(orig_skb, hwtstamps, orig_skb->sk,
4393 SCM_TSTAMP_SND);
4394}
4395EXPORT_SYMBOL_GPL(skb_tstamp_tx);
4396
4397void skb_complete_wifi_ack(struct sk_buff *skb, bool acked)
4398{
4399 struct sock *sk = skb->sk;
4400 struct sock_exterr_skb *serr;
4401 int err = 1;
4402
4403 skb->wifi_acked_valid = 1;
4404 skb->wifi_acked = acked;
4405
4406 serr = SKB_EXT_ERR(skb);
4407 memset(serr, 0, sizeof(*serr));
4408 serr->ee.ee_errno = ENOMSG;
4409 serr->ee.ee_origin = SO_EE_ORIGIN_TXSTATUS;
4410
4411 /* Take a reference to prevent skb_orphan() from freeing the socket,
4412 * but only if the socket refcount is not zero.
4413 */
4414 if (likely(refcount_inc_not_zero(&sk->sk_refcnt))) {
4415 err = sock_queue_err_skb(sk, skb);
4416 sock_put(sk);
4417 }
4418 if (err)
4419 kfree_skb(skb);
4420}
4421EXPORT_SYMBOL_GPL(skb_complete_wifi_ack);
4422
4423/**
4424 * skb_partial_csum_set - set up and verify partial csum values for packet
4425 * @skb: the skb to set
4426 * @start: the number of bytes after skb->data to start checksumming.
4427 * @off: the offset from start to place the checksum.
4428 *
4429 * For untrusted partially-checksummed packets, we need to make sure the values
4430 * for skb->csum_start and skb->csum_offset are valid so we don't oops.
4431 *
4432 * This function checks and sets those values and skb->ip_summed: if this
4433 * returns false you should drop the packet.
4434 */
4435bool skb_partial_csum_set(struct sk_buff *skb, u16 start, u16 off)
4436{
4437 if (unlikely(start > skb_headlen(skb)) ||
4438 unlikely((int)start + off > skb_headlen(skb) - 2)) {
4439 net_warn_ratelimited("bad partial csum: csum=%u/%u len=%u\n",
4440 start, off, skb_headlen(skb));
4441 return false;
4442 }
4443 skb->ip_summed = CHECKSUM_PARTIAL;
4444 skb->csum_start = skb_headroom(skb) + start;
4445 skb->csum_offset = off;
4446 skb_set_transport_header(skb, start);
4447 return true;
4448}
4449EXPORT_SYMBOL_GPL(skb_partial_csum_set);
4450
4451static int skb_maybe_pull_tail(struct sk_buff *skb, unsigned int len,
4452 unsigned int max)
4453{
4454 if (skb_headlen(skb) >= len)
4455 return 0;
4456
4457 /* If we need to pullup then pullup to the max, so we
4458 * won't need to do it again.
4459 */
4460 if (max > skb->len)
4461 max = skb->len;
4462
4463 if (__pskb_pull_tail(skb, max - skb_headlen(skb)) == NULL)
4464 return -ENOMEM;
4465
4466 if (skb_headlen(skb) < len)
4467 return -EPROTO;
4468
4469 return 0;
4470}
4471
4472#define MAX_TCP_HDR_LEN (15 * 4)
4473
4474static __sum16 *skb_checksum_setup_ip(struct sk_buff *skb,
4475 typeof(IPPROTO_IP) proto,
4476 unsigned int off)
4477{
4478 switch (proto) {
4479 int err;
4480
4481 case IPPROTO_TCP:
4482 err = skb_maybe_pull_tail(skb, off + sizeof(struct tcphdr),
4483 off + MAX_TCP_HDR_LEN);
4484 if (!err && !skb_partial_csum_set(skb, off,
4485 offsetof(struct tcphdr,
4486 check)))
4487 err = -EPROTO;
4488 return err ? ERR_PTR(err) : &tcp_hdr(skb)->check;
4489
4490 case IPPROTO_UDP:
4491 err = skb_maybe_pull_tail(skb, off + sizeof(struct udphdr),
4492 off + sizeof(struct udphdr));
4493 if (!err && !skb_partial_csum_set(skb, off,
4494 offsetof(struct udphdr,
4495 check)))
4496 err = -EPROTO;
4497 return err ? ERR_PTR(err) : &udp_hdr(skb)->check;
4498 }
4499
4500 return ERR_PTR(-EPROTO);
4501}
4502
4503/* This value should be large enough to cover a tagged ethernet header plus
4504 * maximally sized IP and TCP or UDP headers.
4505 */
4506#define MAX_IP_HDR_LEN 128
4507
4508static int skb_checksum_setup_ipv4(struct sk_buff *skb, bool recalculate)
4509{
4510 unsigned int off;
4511 bool fragment;
4512 __sum16 *csum;
4513 int err;
4514
4515 fragment = false;
4516
4517 err = skb_maybe_pull_tail(skb,
4518 sizeof(struct iphdr),
4519 MAX_IP_HDR_LEN);
4520 if (err < 0)
4521 goto out;
4522
4523 if (ip_hdr(skb)->frag_off & htons(IP_OFFSET | IP_MF))
4524 fragment = true;
4525
4526 off = ip_hdrlen(skb);
4527
4528 err = -EPROTO;
4529
4530 if (fragment)
4531 goto out;
4532
4533 csum = skb_checksum_setup_ip(skb, ip_hdr(skb)->protocol, off);
4534 if (IS_ERR(csum))
4535 return PTR_ERR(csum);
4536
4537 if (recalculate)
4538 *csum = ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
4539 ip_hdr(skb)->daddr,
4540 skb->len - off,
4541 ip_hdr(skb)->protocol, 0);
4542 err = 0;
4543
4544out:
4545 return err;
4546}
4547
4548/* This value should be large enough to cover a tagged ethernet header plus
4549 * an IPv6 header, all options, and a maximal TCP or UDP header.
4550 */
4551#define MAX_IPV6_HDR_LEN 256
4552
4553#define OPT_HDR(type, skb, off) \
4554 (type *)(skb_network_header(skb) + (off))
4555
4556static int skb_checksum_setup_ipv6(struct sk_buff *skb, bool recalculate)
4557{
4558 int err;
4559 u8 nexthdr;
4560 unsigned int off;
4561 unsigned int len;
4562 bool fragment;
4563 bool done;
4564 __sum16 *csum;
4565
4566 fragment = false;
4567 done = false;
4568
4569 off = sizeof(struct ipv6hdr);
4570
4571 err = skb_maybe_pull_tail(skb, off, MAX_IPV6_HDR_LEN);
4572 if (err < 0)
4573 goto out;
4574
4575 nexthdr = ipv6_hdr(skb)->nexthdr;
4576
4577 len = sizeof(struct ipv6hdr) + ntohs(ipv6_hdr(skb)->payload_len);
4578 while (off <= len && !done) {
4579 switch (nexthdr) {
4580 case IPPROTO_DSTOPTS:
4581 case IPPROTO_HOPOPTS:
4582 case IPPROTO_ROUTING: {
4583 struct ipv6_opt_hdr *hp;
4584
4585 err = skb_maybe_pull_tail(skb,
4586 off +
4587 sizeof(struct ipv6_opt_hdr),
4588 MAX_IPV6_HDR_LEN);
4589 if (err < 0)
4590 goto out;
4591
4592 hp = OPT_HDR(struct ipv6_opt_hdr, skb, off);
4593 nexthdr = hp->nexthdr;
4594 off += ipv6_optlen(hp);
4595 break;
4596 }
4597 case IPPROTO_AH: {
4598 struct ip_auth_hdr *hp;
4599
4600 err = skb_maybe_pull_tail(skb,
4601 off +
4602 sizeof(struct ip_auth_hdr),
4603 MAX_IPV6_HDR_LEN);
4604 if (err < 0)
4605 goto out;
4606
4607 hp = OPT_HDR(struct ip_auth_hdr, skb, off);
4608 nexthdr = hp->nexthdr;
4609 off += ipv6_authlen(hp);
4610 break;
4611 }
4612 case IPPROTO_FRAGMENT: {
4613 struct frag_hdr *hp;
4614
4615 err = skb_maybe_pull_tail(skb,
4616 off +
4617 sizeof(struct frag_hdr),
4618 MAX_IPV6_HDR_LEN);
4619 if (err < 0)
4620 goto out;
4621
4622 hp = OPT_HDR(struct frag_hdr, skb, off);
4623
4624 if (hp->frag_off & htons(IP6_OFFSET | IP6_MF))
4625 fragment = true;
4626
4627 nexthdr = hp->nexthdr;
4628 off += sizeof(struct frag_hdr);
4629 break;
4630 }
4631 default:
4632 done = true;
4633 break;
4634 }
4635 }
4636
4637 err = -EPROTO;
4638
4639 if (!done || fragment)
4640 goto out;
4641
4642 csum = skb_checksum_setup_ip(skb, nexthdr, off);
4643 if (IS_ERR(csum))
4644 return PTR_ERR(csum);
4645
4646 if (recalculate)
4647 *csum = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
4648 &ipv6_hdr(skb)->daddr,
4649 skb->len - off, nexthdr, 0);
4650 err = 0;
4651
4652out:
4653 return err;
4654}
4655
4656/**
4657 * skb_checksum_setup - set up partial checksum offset
4658 * @skb: the skb to set up
4659 * @recalculate: if true the pseudo-header checksum will be recalculated
4660 */
4661int skb_checksum_setup(struct sk_buff *skb, bool recalculate)
4662{
4663 int err;
4664
4665 switch (skb->protocol) {
4666 case htons(ETH_P_IP):
4667 err = skb_checksum_setup_ipv4(skb, recalculate);
4668 break;
4669
4670 case htons(ETH_P_IPV6):
4671 err = skb_checksum_setup_ipv6(skb, recalculate);
4672 break;
4673
4674 default:
4675 err = -EPROTO;
4676 break;
4677 }
4678
4679 return err;
4680}
4681EXPORT_SYMBOL(skb_checksum_setup);
4682
4683/**
4684 * skb_checksum_maybe_trim - maybe trims the given skb
4685 * @skb: the skb to check
4686 * @transport_len: the data length beyond the network header
4687 *
4688 * Checks whether the given skb has data beyond the given transport length.
4689 * If so, returns a cloned skb trimmed to this transport length.
4690 * Otherwise returns the provided skb. Returns NULL in error cases
4691 * (e.g. transport_len exceeds skb length or out-of-memory).
4692 *
4693 * Caller needs to set the skb transport header and free any returned skb if it
4694 * differs from the provided skb.
4695 */
4696static struct sk_buff *skb_checksum_maybe_trim(struct sk_buff *skb,
4697 unsigned int transport_len)
4698{
4699 struct sk_buff *skb_chk;
4700 unsigned int len = skb_transport_offset(skb) + transport_len;
4701 int ret;
4702
4703 if (skb->len < len)
4704 return NULL;
4705 else if (skb->len == len)
4706 return skb;
4707
4708 skb_chk = skb_clone(skb, GFP_ATOMIC);
4709 if (!skb_chk)
4710 return NULL;
4711
4712 ret = pskb_trim_rcsum(skb_chk, len);
4713 if (ret) {
4714 kfree_skb(skb_chk);
4715 return NULL;
4716 }
4717
4718 return skb_chk;
4719}
4720
4721/**
4722 * skb_checksum_trimmed - validate checksum of an skb
4723 * @skb: the skb to check
4724 * @transport_len: the data length beyond the network header
4725 * @skb_chkf: checksum function to use
4726 *
4727 * Applies the given checksum function skb_chkf to the provided skb.
4728 * Returns a checked and maybe trimmed skb. Returns NULL on error.
4729 *
4730 * If the skb has data beyond the given transport length, then a
4731 * trimmed & cloned skb is checked and returned.
4732 *
4733 * Caller needs to set the skb transport header and free any returned skb if it
4734 * differs from the provided skb.
4735 */
4736struct sk_buff *skb_checksum_trimmed(struct sk_buff *skb,
4737 unsigned int transport_len,
4738 __sum16(*skb_chkf)(struct sk_buff *skb))
4739{
4740 struct sk_buff *skb_chk;
4741 unsigned int offset = skb_transport_offset(skb);
4742 __sum16 ret;
4743
4744 skb_chk = skb_checksum_maybe_trim(skb, transport_len);
4745 if (!skb_chk)
4746 goto err;
4747
4748 if (!pskb_may_pull(skb_chk, offset))
4749 goto err;
4750
4751 skb_pull_rcsum(skb_chk, offset);
4752 ret = skb_chkf(skb_chk);
4753 skb_push_rcsum(skb_chk, offset);
4754
4755 if (ret)
4756 goto err;
4757
4758 return skb_chk;
4759
4760err:
4761 if (skb_chk && skb_chk != skb)
4762 kfree_skb(skb_chk);
4763
4764 return NULL;
4765
4766}
4767EXPORT_SYMBOL(skb_checksum_trimmed);
4768
4769void __skb_warn_lro_forwarding(const struct sk_buff *skb)
4770{
4771 net_warn_ratelimited("%s: received packets cannot be forwarded while LRO is enabled\n",
4772 skb->dev->name);
4773}
4774EXPORT_SYMBOL(__skb_warn_lro_forwarding);
4775
4776void kfree_skb_partial(struct sk_buff *skb, bool head_stolen)
4777{
4778 if (head_stolen) {
4779 skb_release_head_state(skb);
4780 kmem_cache_free(skbuff_head_cache, skb);
4781 } else {
4782 __kfree_skb(skb);
4783 }
4784}
4785EXPORT_SYMBOL(kfree_skb_partial);
4786
4787/**
4788 * skb_try_coalesce - try to merge skb to prior one
4789 * @to: prior buffer
4790 * @from: buffer to add
4791 * @fragstolen: pointer to boolean
4792 * @delta_truesize: how much more was allocated than was requested
4793 */
4794bool skb_try_coalesce(struct sk_buff *to, struct sk_buff *from,
4795 bool *fragstolen, int *delta_truesize)
4796{
4797 struct skb_shared_info *to_shinfo, *from_shinfo;
4798 int i, delta, len = from->len;
4799
4800 *fragstolen = false;
4801
4802 if (skb_cloned(to))
4803 return false;
4804
4805 if (len <= skb_tailroom(to)) {
4806 if (len)
4807 BUG_ON(skb_copy_bits(from, 0, skb_put(to, len), len));
4808 *delta_truesize = 0;
4809 return true;
4810 }
4811
4812 to_shinfo = skb_shinfo(to);
4813 from_shinfo = skb_shinfo(from);
4814 if (to_shinfo->frag_list || from_shinfo->frag_list)
4815 return false;
4816 if (skb_zcopy(to) || skb_zcopy(from))
4817 return false;
4818
4819 if (skb_headlen(from) != 0) {
4820 struct page *page;
4821 unsigned int offset;
4822
4823 if (to_shinfo->nr_frags +
4824 from_shinfo->nr_frags >= MAX_SKB_FRAGS)
4825 return false;
4826
4827 if (skb_head_is_locked(from))
4828 return false;
4829
4830 delta = from->truesize - SKB_DATA_ALIGN(sizeof(struct sk_buff));
4831
4832 page = virt_to_head_page(from->head);
4833 offset = from->data - (unsigned char *)page_address(page);
4834
4835 skb_fill_page_desc(to, to_shinfo->nr_frags,
4836 page, offset, skb_headlen(from));
4837 *fragstolen = true;
4838 } else {
4839 if (to_shinfo->nr_frags +
4840 from_shinfo->nr_frags > MAX_SKB_FRAGS)
4841 return false;
4842
4843 delta = from->truesize - SKB_TRUESIZE(skb_end_offset(from));
4844 }
4845
4846 WARN_ON_ONCE(delta < len);
4847
4848 memcpy(to_shinfo->frags + to_shinfo->nr_frags,
4849 from_shinfo->frags,
4850 from_shinfo->nr_frags * sizeof(skb_frag_t));
4851 to_shinfo->nr_frags += from_shinfo->nr_frags;
4852
4853 if (!skb_cloned(from))
4854 from_shinfo->nr_frags = 0;
4855
4856 /* if the skb is not cloned this does nothing
4857 * since we set nr_frags to 0.
4858 */
4859 for (i = 0; i < from_shinfo->nr_frags; i++)
4860 __skb_frag_ref(&from_shinfo->frags[i]);
4861
4862 to->truesize += delta;
4863 to->len += len;
4864 to->data_len += len;
4865
4866 *delta_truesize = delta;
4867 return true;
4868}
4869EXPORT_SYMBOL(skb_try_coalesce);
4870
4871/**
4872 * skb_scrub_packet - scrub an skb
4873 *
4874 * @skb: buffer to clean
4875 * @xnet: packet is crossing netns
4876 *
4877 * skb_scrub_packet can be used after encapsulating or decapsulting a packet
4878 * into/from a tunnel. Some information have to be cleared during these
4879 * operations.
4880 * skb_scrub_packet can also be used to clean a skb before injecting it in
4881 * another namespace (@xnet == true). We have to clear all information in the
4882 * skb that could impact namespace isolation.
4883 */
4884void skb_scrub_packet(struct sk_buff *skb, bool xnet)
4885{
4886 skb->tstamp = 0;
4887 skb->pkt_type = PACKET_HOST;
4888 skb->skb_iif = 0;
4889 skb->ignore_df = 0;
4890 skb_dst_drop(skb);
4891 secpath_reset(skb);
4892 nf_reset(skb);
4893 nf_reset_trace(skb);
4894
4895 if (!xnet)
4896 return;
4897
4898 ipvs_reset(skb);
4899 skb_orphan(skb);
4900 skb->mark = 0;
4901}
4902EXPORT_SYMBOL_GPL(skb_scrub_packet);
4903
4904/**
4905 * skb_gso_transport_seglen - Return length of individual segments of a gso packet
4906 *
4907 * @skb: GSO skb
4908 *
4909 * skb_gso_transport_seglen is used to determine the real size of the
4910 * individual segments, including Layer4 headers (TCP/UDP).
4911 *
4912 * The MAC/L2 or network (IP, IPv6) headers are not accounted for.
4913 */
4914static unsigned int skb_gso_transport_seglen(const struct sk_buff *skb)
4915{
4916 const struct skb_shared_info *shinfo = skb_shinfo(skb);
4917 unsigned int thlen = 0;
4918
4919 if (skb->encapsulation) {
4920 thlen = skb_inner_transport_header(skb) -
4921 skb_transport_header(skb);
4922
4923 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
4924 thlen += inner_tcp_hdrlen(skb);
4925 } else if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
4926 thlen = tcp_hdrlen(skb);
4927 } else if (unlikely(skb_is_gso_sctp(skb))) {
4928 thlen = sizeof(struct sctphdr);
4929 }
4930 /* UFO sets gso_size to the size of the fragmentation
4931 * payload, i.e. the size of the L4 (UDP) header is already
4932 * accounted for.
4933 */
4934 return thlen + shinfo->gso_size;
4935}
4936
4937/**
4938 * skb_gso_network_seglen - Return length of individual segments of a gso packet
4939 *
4940 * @skb: GSO skb
4941 *
4942 * skb_gso_network_seglen is used to determine the real size of the
4943 * individual segments, including Layer3 (IP, IPv6) and L4 headers (TCP/UDP).
4944 *
4945 * The MAC/L2 header is not accounted for.
4946 */
4947static unsigned int skb_gso_network_seglen(const struct sk_buff *skb)
4948{
4949 unsigned int hdr_len = skb_transport_header(skb) -
4950 skb_network_header(skb);
4951
4952 return hdr_len + skb_gso_transport_seglen(skb);
4953}
4954
4955/**
4956 * skb_gso_mac_seglen - Return length of individual segments of a gso packet
4957 *
4958 * @skb: GSO skb
4959 *
4960 * skb_gso_mac_seglen is used to determine the real size of the
4961 * individual segments, including MAC/L2, Layer3 (IP, IPv6) and L4
4962 * headers (TCP/UDP).
4963 */
4964static unsigned int skb_gso_mac_seglen(const struct sk_buff *skb)
4965{
4966 unsigned int hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
4967
4968 return hdr_len + skb_gso_transport_seglen(skb);
4969}
4970
4971/**
4972 * skb_gso_size_check - check the skb size, considering GSO_BY_FRAGS
4973 *
4974 * There are a couple of instances where we have a GSO skb, and we
4975 * want to determine what size it would be after it is segmented.
4976 *
4977 * We might want to check:
4978 * - L3+L4+payload size (e.g. IP forwarding)
4979 * - L2+L3+L4+payload size (e.g. sanity check before passing to driver)
4980 *
4981 * This is a helper to do that correctly considering GSO_BY_FRAGS.
4982 *
4983 * @seg_len: The segmented length (from skb_gso_*_seglen). In the
4984 * GSO_BY_FRAGS case this will be [header sizes + GSO_BY_FRAGS].
4985 *
4986 * @max_len: The maximum permissible length.
4987 *
4988 * Returns true if the segmented length <= max length.
4989 */
4990static inline bool skb_gso_size_check(const struct sk_buff *skb,
4991 unsigned int seg_len,
4992 unsigned int max_len) {
4993 const struct skb_shared_info *shinfo = skb_shinfo(skb);
4994 const struct sk_buff *iter;
4995
4996 if (shinfo->gso_size != GSO_BY_FRAGS)
4997 return seg_len <= max_len;
4998
4999 /* Undo this so we can re-use header sizes */
5000 seg_len -= GSO_BY_FRAGS;
5001
5002 skb_walk_frags(skb, iter) {
5003 if (seg_len + skb_headlen(iter) > max_len)
5004 return false;
5005 }
5006
5007 return true;
5008}
5009
5010/**
5011 * skb_gso_validate_network_len - Will a split GSO skb fit into a given MTU?
5012 *
5013 * @skb: GSO skb
5014 * @mtu: MTU to validate against
5015 *
5016 * skb_gso_validate_network_len validates if a given skb will fit a
5017 * wanted MTU once split. It considers L3 headers, L4 headers, and the
5018 * payload.
5019 */
5020bool skb_gso_validate_network_len(const struct sk_buff *skb, unsigned int mtu)
5021{
5022 return skb_gso_size_check(skb, skb_gso_network_seglen(skb), mtu);
5023}
5024EXPORT_SYMBOL_GPL(skb_gso_validate_network_len);
5025
5026/**
5027 * skb_gso_validate_mac_len - Will a split GSO skb fit in a given length?
5028 *
5029 * @skb: GSO skb
5030 * @len: length to validate against
5031 *
5032 * skb_gso_validate_mac_len validates if a given skb will fit a wanted
5033 * length once split, including L2, L3 and L4 headers and the payload.
5034 */
5035bool skb_gso_validate_mac_len(const struct sk_buff *skb, unsigned int len)
5036{
5037 return skb_gso_size_check(skb, skb_gso_mac_seglen(skb), len);
5038}
5039EXPORT_SYMBOL_GPL(skb_gso_validate_mac_len);
5040
5041static struct sk_buff *skb_reorder_vlan_header(struct sk_buff *skb)
5042{
5043 int mac_len;
5044
5045 if (skb_cow(skb, skb_headroom(skb)) < 0) {
5046 kfree_skb(skb);
5047 return NULL;
5048 }
5049
5050 mac_len = skb->data - skb_mac_header(skb);
5051 if (likely(mac_len > VLAN_HLEN + ETH_TLEN)) {
5052 memmove(skb_mac_header(skb) + VLAN_HLEN, skb_mac_header(skb),
5053 mac_len - VLAN_HLEN - ETH_TLEN);
5054 }
5055 skb->mac_header += VLAN_HLEN;
5056 return skb;
5057}
5058
5059struct sk_buff *skb_vlan_untag(struct sk_buff *skb)
5060{
5061 struct vlan_hdr *vhdr;
5062 u16 vlan_tci;
5063
5064 if (unlikely(skb_vlan_tag_present(skb))) {
5065 /* vlan_tci is already set-up so leave this for another time */
5066 return skb;
5067 }
5068
5069 skb = skb_share_check(skb, GFP_ATOMIC);
5070 if (unlikely(!skb))
5071 goto err_free;
5072
5073 if (unlikely(!pskb_may_pull(skb, VLAN_HLEN)))
5074 goto err_free;
5075
5076 vhdr = (struct vlan_hdr *)skb->data;
5077 vlan_tci = ntohs(vhdr->h_vlan_TCI);
5078 __vlan_hwaccel_put_tag(skb, skb->protocol, vlan_tci);
5079
5080 skb_pull_rcsum(skb, VLAN_HLEN);
5081 vlan_set_encap_proto(skb, vhdr);
5082
5083 skb = skb_reorder_vlan_header(skb);
5084 if (unlikely(!skb))
5085 goto err_free;
5086
5087 skb_reset_network_header(skb);
5088 skb_reset_transport_header(skb);
5089 skb_reset_mac_len(skb);
5090
5091 return skb;
5092
5093err_free:
5094 kfree_skb(skb);
5095 return NULL;
5096}
5097EXPORT_SYMBOL(skb_vlan_untag);
5098
5099int skb_ensure_writable(struct sk_buff *skb, int write_len)
5100{
5101 if (!pskb_may_pull(skb, write_len))
5102 return -ENOMEM;
5103
5104 if (!skb_cloned(skb) || skb_clone_writable(skb, write_len))
5105 return 0;
5106
5107 return pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
5108}
5109EXPORT_SYMBOL(skb_ensure_writable);
5110
5111/* remove VLAN header from packet and update csum accordingly.
5112 * expects a non skb_vlan_tag_present skb with a vlan tag payload
5113 */
5114int __skb_vlan_pop(struct sk_buff *skb, u16 *vlan_tci)
5115{
5116 struct vlan_hdr *vhdr;
5117 int offset = skb->data - skb_mac_header(skb);
5118 int err;
5119
5120 if (WARN_ONCE(offset,
5121 "__skb_vlan_pop got skb with skb->data not at mac header (offset %d)\n",
5122 offset)) {
5123 return -EINVAL;
5124 }
5125
5126 err = skb_ensure_writable(skb, VLAN_ETH_HLEN);
5127 if (unlikely(err))
5128 return err;
5129
5130 skb_postpull_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
5131
5132 vhdr = (struct vlan_hdr *)(skb->data + ETH_HLEN);
5133 *vlan_tci = ntohs(vhdr->h_vlan_TCI);
5134
5135 memmove(skb->data + VLAN_HLEN, skb->data, 2 * ETH_ALEN);
5136 __skb_pull(skb, VLAN_HLEN);
5137
5138 vlan_set_encap_proto(skb, vhdr);
5139 skb->mac_header += VLAN_HLEN;
5140
5141 if (skb_network_offset(skb) < ETH_HLEN)
5142 skb_set_network_header(skb, ETH_HLEN);
5143
5144 skb_reset_mac_len(skb);
5145
5146 return err;
5147}
5148EXPORT_SYMBOL(__skb_vlan_pop);
5149
5150/* Pop a vlan tag either from hwaccel or from payload.
5151 * Expects skb->data at mac header.
5152 */
5153int skb_vlan_pop(struct sk_buff *skb)
5154{
5155 u16 vlan_tci;
5156 __be16 vlan_proto;
5157 int err;
5158
5159 if (likely(skb_vlan_tag_present(skb))) {
5160 skb->vlan_tci = 0;
5161 } else {
5162 if (unlikely(!eth_type_vlan(skb->protocol)))
5163 return 0;
5164
5165 err = __skb_vlan_pop(skb, &vlan_tci);
5166 if (err)
5167 return err;
5168 }
5169 /* move next vlan tag to hw accel tag */
5170 if (likely(!eth_type_vlan(skb->protocol)))
5171 return 0;
5172
5173 vlan_proto = skb->protocol;
5174 err = __skb_vlan_pop(skb, &vlan_tci);
5175 if (unlikely(err))
5176 return err;
5177
5178 __vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
5179 return 0;
5180}
5181EXPORT_SYMBOL(skb_vlan_pop);
5182
5183/* Push a vlan tag either into hwaccel or into payload (if hwaccel tag present).
5184 * Expects skb->data at mac header.
5185 */
5186int skb_vlan_push(struct sk_buff *skb, __be16 vlan_proto, u16 vlan_tci)
5187{
5188 if (skb_vlan_tag_present(skb)) {
5189 int offset = skb->data - skb_mac_header(skb);
5190 int err;
5191
5192 if (WARN_ONCE(offset,
5193 "skb_vlan_push got skb with skb->data not at mac header (offset %d)\n",
5194 offset)) {
5195 return -EINVAL;
5196 }
5197
5198 err = __vlan_insert_tag(skb, skb->vlan_proto,
5199 skb_vlan_tag_get(skb));
5200 if (err)
5201 return err;
5202
5203 skb->protocol = skb->vlan_proto;
5204 skb->mac_len += VLAN_HLEN;
5205
5206 skb_postpush_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
5207 }
5208 __vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
5209 return 0;
5210}
5211EXPORT_SYMBOL(skb_vlan_push);
5212
5213/**
5214 * alloc_skb_with_frags - allocate skb with page frags
5215 *
5216 * @header_len: size of linear part
5217 * @data_len: needed length in frags
5218 * @max_page_order: max page order desired.
5219 * @errcode: pointer to error code if any
5220 * @gfp_mask: allocation mask
5221 *
5222 * This can be used to allocate a paged skb, given a maximal order for frags.
5223 */
5224struct sk_buff *alloc_skb_with_frags(unsigned long header_len,
5225 unsigned long data_len,
5226 int max_page_order,
5227 int *errcode,
5228 gfp_t gfp_mask)
5229{
5230 int npages = (data_len + (PAGE_SIZE - 1)) >> PAGE_SHIFT;
5231 unsigned long chunk;
5232 struct sk_buff *skb;
5233 struct page *page;
5234 gfp_t gfp_head;
5235 int i;
5236
5237 *errcode = -EMSGSIZE;
5238 /* Note this test could be relaxed, if we succeed to allocate
5239 * high order pages...
5240 */
5241 if (npages > MAX_SKB_FRAGS)
5242 return NULL;
5243
5244 gfp_head = gfp_mask;
5245 if (gfp_head & __GFP_DIRECT_RECLAIM)
5246 gfp_head |= __GFP_RETRY_MAYFAIL;
5247
5248 *errcode = -ENOBUFS;
5249 skb = alloc_skb(header_len, gfp_head);
5250 if (!skb)
5251 return NULL;
5252
5253 skb->truesize += npages << PAGE_SHIFT;
5254
5255 for (i = 0; npages > 0; i++) {
5256 int order = max_page_order;
5257
5258 while (order) {
5259 if (npages >= 1 << order) {
5260 page = alloc_pages((gfp_mask & ~__GFP_DIRECT_RECLAIM) |
5261 __GFP_COMP |
5262 __GFP_NOWARN |
5263 __GFP_NORETRY,
5264 order);
5265 if (page)
5266 goto fill_page;
5267 /* Do not retry other high order allocations */
5268 order = 1;
5269 max_page_order = 0;
5270 }
5271 order--;
5272 }
5273 page = alloc_page(gfp_mask);
5274 if (!page)
5275 goto failure;
5276fill_page:
5277 chunk = min_t(unsigned long, data_len,
5278 PAGE_SIZE << order);
5279 skb_fill_page_desc(skb, i, page, 0, chunk);
5280 data_len -= chunk;
5281 npages -= 1 << order;
5282 }
5283 return skb;
5284
5285failure:
5286 kfree_skb(skb);
5287 return NULL;
5288}
5289EXPORT_SYMBOL(alloc_skb_with_frags);
5290
5291/* carve out the first off bytes from skb when off < headlen */
5292static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
5293 const int headlen, gfp_t gfp_mask)
5294{
5295 int i;
5296 int size = skb_end_offset(skb);
5297 int new_hlen = headlen - off;
5298 u8 *data;
5299
5300 size = SKB_DATA_ALIGN(size);
5301
5302 if (skb_pfmemalloc(skb))
5303 gfp_mask |= __GFP_MEMALLOC;
5304 data = kmalloc_reserve(size +
5305 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
5306 gfp_mask, NUMA_NO_NODE, NULL);
5307 if (!data)
5308 return -ENOMEM;
5309
5310 size = SKB_WITH_OVERHEAD(ksize(data));
5311
5312 /* Copy real data, and all frags */
5313 skb_copy_from_linear_data_offset(skb, off, data, new_hlen);
5314 skb->len -= off;
5315
5316 memcpy((struct skb_shared_info *)(data + size),
5317 skb_shinfo(skb),
5318 offsetof(struct skb_shared_info,
5319 frags[skb_shinfo(skb)->nr_frags]));
5320 if (skb_cloned(skb)) {
5321 /* drop the old head gracefully */
5322 if (skb_orphan_frags(skb, gfp_mask)) {
5323 kfree(data);
5324 return -ENOMEM;
5325 }
5326 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
5327 skb_frag_ref(skb, i);
5328 if (skb_has_frag_list(skb))
5329 skb_clone_fraglist(skb);
5330 skb_release_data(skb);
5331 } else {
5332 /* we can reuse existing recount- all we did was
5333 * relocate values
5334 */
5335 skb_free_head(skb);
5336 }
5337
5338 skb->head = data;
5339 skb->data = data;
5340 skb->head_frag = 0;
5341#ifdef NET_SKBUFF_DATA_USES_OFFSET
5342 skb->end = size;
5343#else
5344 skb->end = skb->head + size;
5345#endif
5346 skb_set_tail_pointer(skb, skb_headlen(skb));
5347 skb_headers_offset_update(skb, 0);
5348 skb->cloned = 0;
5349 skb->hdr_len = 0;
5350 skb->nohdr = 0;
5351 atomic_set(&skb_shinfo(skb)->dataref, 1);
5352
5353 return 0;
5354}
5355
5356static int pskb_carve(struct sk_buff *skb, const u32 off, gfp_t gfp);
5357
5358/* carve out the first eat bytes from skb's frag_list. May recurse into
5359 * pskb_carve()
5360 */
5361static int pskb_carve_frag_list(struct sk_buff *skb,
5362 struct skb_shared_info *shinfo, int eat,
5363 gfp_t gfp_mask)
5364{
5365 struct sk_buff *list = shinfo->frag_list;
5366 struct sk_buff *clone = NULL;
5367 struct sk_buff *insp = NULL;
5368
5369 do {
5370 if (!list) {
5371 pr_err("Not enough bytes to eat. Want %d\n", eat);
5372 return -EFAULT;
5373 }
5374 if (list->len <= eat) {
5375 /* Eaten as whole. */
5376 eat -= list->len;
5377 list = list->next;
5378 insp = list;
5379 } else {
5380 /* Eaten partially. */
5381 if (skb_shared(list)) {
5382 clone = skb_clone(list, gfp_mask);
5383 if (!clone)
5384 return -ENOMEM;
5385 insp = list->next;
5386 list = clone;
5387 } else {
5388 /* This may be pulled without problems. */
5389 insp = list;
5390 }
5391 if (pskb_carve(list, eat, gfp_mask) < 0) {
5392 kfree_skb(clone);
5393 return -ENOMEM;
5394 }
5395 break;
5396 }
5397 } while (eat);
5398
5399 /* Free pulled out fragments. */
5400 while ((list = shinfo->frag_list) != insp) {
5401 shinfo->frag_list = list->next;
5402 kfree_skb(list);
5403 }
5404 /* And insert new clone at head. */
5405 if (clone) {
5406 clone->next = list;
5407 shinfo->frag_list = clone;
5408 }
5409 return 0;
5410}
5411
5412/* carve off first len bytes from skb. Split line (off) is in the
5413 * non-linear part of skb
5414 */
5415static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
5416 int pos, gfp_t gfp_mask)
5417{
5418 int i, k = 0;
5419 int size = skb_end_offset(skb);
5420 u8 *data;
5421 const int nfrags = skb_shinfo(skb)->nr_frags;
5422 struct skb_shared_info *shinfo;
5423
5424 size = SKB_DATA_ALIGN(size);
5425
5426 if (skb_pfmemalloc(skb))
5427 gfp_mask |= __GFP_MEMALLOC;
5428 data = kmalloc_reserve(size +
5429 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)),
5430 gfp_mask, NUMA_NO_NODE, NULL);
5431 if (!data)
5432 return -ENOMEM;
5433
5434 size = SKB_WITH_OVERHEAD(ksize(data));
5435
5436 memcpy((struct skb_shared_info *)(data + size),
5437 skb_shinfo(skb), offsetof(struct skb_shared_info,
5438 frags[skb_shinfo(skb)->nr_frags]));
5439 if (skb_orphan_frags(skb, gfp_mask)) {
5440 kfree(data);
5441 return -ENOMEM;
5442 }
5443 shinfo = (struct skb_shared_info *)(data + size);
5444 for (i = 0; i < nfrags; i++) {
5445 int fsize = skb_frag_size(&skb_shinfo(skb)->frags[i]);
5446
5447 if (pos + fsize > off) {
5448 shinfo->frags[k] = skb_shinfo(skb)->frags[i];
5449
5450 if (pos < off) {
5451 /* Split frag.
5452 * We have two variants in this case:
5453 * 1. Move all the frag to the second
5454 * part, if it is possible. F.e.
5455 * this approach is mandatory for TUX,
5456 * where splitting is expensive.
5457 * 2. Split is accurately. We make this.
5458 */
5459 shinfo->frags[0].page_offset += off - pos;
5460 skb_frag_size_sub(&shinfo->frags[0], off - pos);
5461 }
5462 skb_frag_ref(skb, i);
5463 k++;
5464 }
5465 pos += fsize;
5466 }
5467 shinfo->nr_frags = k;
5468 if (skb_has_frag_list(skb))
5469 skb_clone_fraglist(skb);
5470
5471 if (k == 0) {
5472 /* split line is in frag list */
5473 pskb_carve_frag_list(skb, shinfo, off - pos, gfp_mask);
5474 }
5475 skb_release_data(skb);
5476
5477 skb->head = data;
5478 skb->head_frag = 0;
5479 skb->data = data;
5480#ifdef NET_SKBUFF_DATA_USES_OFFSET
5481 skb->end = size;
5482#else
5483 skb->end = skb->head + size;
5484#endif
5485 skb_reset_tail_pointer(skb);
5486 skb_headers_offset_update(skb, 0);
5487 skb->cloned = 0;
5488 skb->hdr_len = 0;
5489 skb->nohdr = 0;
5490 skb->len -= off;
5491 skb->data_len = skb->len;
5492 atomic_set(&skb_shinfo(skb)->dataref, 1);
5493 return 0;
5494}
5495
5496/* remove len bytes from the beginning of the skb */
5497static int pskb_carve(struct sk_buff *skb, const u32 len, gfp_t gfp)
5498{
5499 int headlen = skb_headlen(skb);
5500
5501 if (len < headlen)
5502 return pskb_carve_inside_header(skb, len, headlen, gfp);
5503 else
5504 return pskb_carve_inside_nonlinear(skb, len, headlen, gfp);
5505}
5506
5507/* Extract to_copy bytes starting at off from skb, and return this in
5508 * a new skb
5509 */
5510struct sk_buff *pskb_extract(struct sk_buff *skb, int off,
5511 int to_copy, gfp_t gfp)
5512{
5513 struct sk_buff *clone = skb_clone(skb, gfp);
5514
5515 if (!clone)
5516 return NULL;
5517
5518 if (pskb_carve(clone, off, gfp) < 0 ||
5519 pskb_trim(clone, to_copy)) {
5520 kfree_skb(clone);
5521 return NULL;
5522 }
5523 return clone;
5524}
5525EXPORT_SYMBOL(pskb_extract);
5526
5527/**
5528 * skb_condense - try to get rid of fragments/frag_list if possible
5529 * @skb: buffer
5530 *
5531 * Can be used to save memory before skb is added to a busy queue.
5532 * If packet has bytes in frags and enough tail room in skb->head,
5533 * pull all of them, so that we can free the frags right now and adjust
5534 * truesize.
5535 * Notes:
5536 * We do not reallocate skb->head thus can not fail.
5537 * Caller must re-evaluate skb->truesize if needed.
5538 */
5539void skb_condense(struct sk_buff *skb)
5540{
5541 if (skb->data_len) {
5542 if (skb->data_len > skb->end - skb->tail ||
5543 skb_cloned(skb))
5544 return;
5545
5546 /* Nice, we can free page frag(s) right now */
5547 __pskb_pull_tail(skb, skb->data_len);
5548 }
5549 /* At this point, skb->truesize might be over estimated,
5550 * because skb had a fragment, and fragments do not tell
5551 * their truesize.
5552 * When we pulled its content into skb->head, fragment
5553 * was freed, but __pskb_pull_tail() could not possibly
5554 * adjust skb->truesize, not knowing the frag truesize.
5555 */
5556 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5557}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Routines having to do with the 'struct sk_buff' memory handlers.
4 *
5 * Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>
6 * Florian La Roche <rzsfl@rz.uni-sb.de>
7 *
8 * Fixes:
9 * Alan Cox : Fixed the worst of the load
10 * balancer bugs.
11 * Dave Platt : Interrupt stacking fix.
12 * Richard Kooijman : Timestamp fixes.
13 * Alan Cox : Changed buffer format.
14 * Alan Cox : destructor hook for AF_UNIX etc.
15 * Linus Torvalds : Better skb_clone.
16 * Alan Cox : Added skb_copy.
17 * Alan Cox : Added all the changed routines Linus
18 * only put in the headers
19 * Ray VanTassle : Fixed --skb->lock in free
20 * Alan Cox : skb_copy copy arp field
21 * Andi Kleen : slabified it.
22 * Robert Olsson : Removed skb_head_pool
23 *
24 * NOTE:
25 * The __skb_ routines should be called with interrupts
26 * disabled, or you better be *real* sure that the operation is atomic
27 * with respect to whatever list is being frobbed (e.g. via lock_sock()
28 * or via disabling bottom half handlers, etc).
29 */
30
31/*
32 * The functions in this file will not compile correctly with gcc 2.4.x
33 */
34
35#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
36
37#include <linux/module.h>
38#include <linux/types.h>
39#include <linux/kernel.h>
40#include <linux/mm.h>
41#include <linux/interrupt.h>
42#include <linux/in.h>
43#include <linux/inet.h>
44#include <linux/slab.h>
45#include <linux/tcp.h>
46#include <linux/udp.h>
47#include <linux/sctp.h>
48#include <linux/netdevice.h>
49#ifdef CONFIG_NET_CLS_ACT
50#include <net/pkt_sched.h>
51#endif
52#include <linux/string.h>
53#include <linux/skbuff.h>
54#include <linux/splice.h>
55#include <linux/cache.h>
56#include <linux/rtnetlink.h>
57#include <linux/init.h>
58#include <linux/scatterlist.h>
59#include <linux/errqueue.h>
60#include <linux/prefetch.h>
61#include <linux/bitfield.h>
62#include <linux/if_vlan.h>
63#include <linux/mpls.h>
64#include <linux/kcov.h>
65#include <linux/iov_iter.h>
66
67#include <net/protocol.h>
68#include <net/dst.h>
69#include <net/sock.h>
70#include <net/checksum.h>
71#include <net/gso.h>
72#include <net/hotdata.h>
73#include <net/ip6_checksum.h>
74#include <net/xfrm.h>
75#include <net/mpls.h>
76#include <net/mptcp.h>
77#include <net/mctp.h>
78#include <net/page_pool/helpers.h>
79#include <net/dropreason.h>
80
81#include <linux/uaccess.h>
82#include <trace/events/skb.h>
83#include <linux/highmem.h>
84#include <linux/capability.h>
85#include <linux/user_namespace.h>
86#include <linux/indirect_call_wrapper.h>
87#include <linux/textsearch.h>
88
89#include "dev.h"
90#include "sock_destructor.h"
91
92#ifdef CONFIG_SKB_EXTENSIONS
93static struct kmem_cache *skbuff_ext_cache __ro_after_init;
94#endif
95
96#define SKB_SMALL_HEAD_SIZE SKB_HEAD_ALIGN(MAX_TCP_HEADER)
97
98/* We want SKB_SMALL_HEAD_CACHE_SIZE to not be a power of two.
99 * This should ensure that SKB_SMALL_HEAD_HEADROOM is a unique
100 * size, and we can differentiate heads from skb_small_head_cache
101 * vs system slabs by looking at their size (skb_end_offset()).
102 */
103#define SKB_SMALL_HEAD_CACHE_SIZE \
104 (is_power_of_2(SKB_SMALL_HEAD_SIZE) ? \
105 (SKB_SMALL_HEAD_SIZE + L1_CACHE_BYTES) : \
106 SKB_SMALL_HEAD_SIZE)
107
108#define SKB_SMALL_HEAD_HEADROOM \
109 SKB_WITH_OVERHEAD(SKB_SMALL_HEAD_CACHE_SIZE)
110
111int sysctl_max_skb_frags __read_mostly = MAX_SKB_FRAGS;
112EXPORT_SYMBOL(sysctl_max_skb_frags);
113
114/* kcm_write_msgs() relies on casting paged frags to bio_vec to use
115 * iov_iter_bvec(). These static asserts ensure the cast is valid is long as the
116 * netmem is a page.
117 */
118static_assert(offsetof(struct bio_vec, bv_page) ==
119 offsetof(skb_frag_t, netmem));
120static_assert(sizeof_field(struct bio_vec, bv_page) ==
121 sizeof_field(skb_frag_t, netmem));
122
123static_assert(offsetof(struct bio_vec, bv_len) == offsetof(skb_frag_t, len));
124static_assert(sizeof_field(struct bio_vec, bv_len) ==
125 sizeof_field(skb_frag_t, len));
126
127static_assert(offsetof(struct bio_vec, bv_offset) ==
128 offsetof(skb_frag_t, offset));
129static_assert(sizeof_field(struct bio_vec, bv_offset) ==
130 sizeof_field(skb_frag_t, offset));
131
132#undef FN
133#define FN(reason) [SKB_DROP_REASON_##reason] = #reason,
134static const char * const drop_reasons[] = {
135 [SKB_CONSUMED] = "CONSUMED",
136 DEFINE_DROP_REASON(FN, FN)
137};
138
139static const struct drop_reason_list drop_reasons_core = {
140 .reasons = drop_reasons,
141 .n_reasons = ARRAY_SIZE(drop_reasons),
142};
143
144const struct drop_reason_list __rcu *
145drop_reasons_by_subsys[SKB_DROP_REASON_SUBSYS_NUM] = {
146 [SKB_DROP_REASON_SUBSYS_CORE] = RCU_INITIALIZER(&drop_reasons_core),
147};
148EXPORT_SYMBOL(drop_reasons_by_subsys);
149
150/**
151 * drop_reasons_register_subsys - register another drop reason subsystem
152 * @subsys: the subsystem to register, must not be the core
153 * @list: the list of drop reasons within the subsystem, must point to
154 * a statically initialized list
155 */
156void drop_reasons_register_subsys(enum skb_drop_reason_subsys subsys,
157 const struct drop_reason_list *list)
158{
159 if (WARN(subsys <= SKB_DROP_REASON_SUBSYS_CORE ||
160 subsys >= ARRAY_SIZE(drop_reasons_by_subsys),
161 "invalid subsystem %d\n", subsys))
162 return;
163
164 /* must point to statically allocated memory, so INIT is OK */
165 RCU_INIT_POINTER(drop_reasons_by_subsys[subsys], list);
166}
167EXPORT_SYMBOL_GPL(drop_reasons_register_subsys);
168
169/**
170 * drop_reasons_unregister_subsys - unregister a drop reason subsystem
171 * @subsys: the subsystem to remove, must not be the core
172 *
173 * Note: This will synchronize_rcu() to ensure no users when it returns.
174 */
175void drop_reasons_unregister_subsys(enum skb_drop_reason_subsys subsys)
176{
177 if (WARN(subsys <= SKB_DROP_REASON_SUBSYS_CORE ||
178 subsys >= ARRAY_SIZE(drop_reasons_by_subsys),
179 "invalid subsystem %d\n", subsys))
180 return;
181
182 RCU_INIT_POINTER(drop_reasons_by_subsys[subsys], NULL);
183
184 synchronize_rcu();
185}
186EXPORT_SYMBOL_GPL(drop_reasons_unregister_subsys);
187
188/**
189 * skb_panic - private function for out-of-line support
190 * @skb: buffer
191 * @sz: size
192 * @addr: address
193 * @msg: skb_over_panic or skb_under_panic
194 *
195 * Out-of-line support for skb_put() and skb_push().
196 * Called via the wrapper skb_over_panic() or skb_under_panic().
197 * Keep out of line to prevent kernel bloat.
198 * __builtin_return_address is not used because it is not always reliable.
199 */
200static void skb_panic(struct sk_buff *skb, unsigned int sz, void *addr,
201 const char msg[])
202{
203 pr_emerg("%s: text:%px len:%d put:%d head:%px data:%px tail:%#lx end:%#lx dev:%s\n",
204 msg, addr, skb->len, sz, skb->head, skb->data,
205 (unsigned long)skb->tail, (unsigned long)skb->end,
206 skb->dev ? skb->dev->name : "<NULL>");
207 BUG();
208}
209
210static void skb_over_panic(struct sk_buff *skb, unsigned int sz, void *addr)
211{
212 skb_panic(skb, sz, addr, __func__);
213}
214
215static void skb_under_panic(struct sk_buff *skb, unsigned int sz, void *addr)
216{
217 skb_panic(skb, sz, addr, __func__);
218}
219
220#define NAPI_SKB_CACHE_SIZE 64
221#define NAPI_SKB_CACHE_BULK 16
222#define NAPI_SKB_CACHE_HALF (NAPI_SKB_CACHE_SIZE / 2)
223
224#if PAGE_SIZE == SZ_4K
225
226#define NAPI_HAS_SMALL_PAGE_FRAG 1
227#define NAPI_SMALL_PAGE_PFMEMALLOC(nc) ((nc).pfmemalloc)
228
229/* specialized page frag allocator using a single order 0 page
230 * and slicing it into 1K sized fragment. Constrained to systems
231 * with a very limited amount of 1K fragments fitting a single
232 * page - to avoid excessive truesize underestimation
233 */
234
235struct page_frag_1k {
236 void *va;
237 u16 offset;
238 bool pfmemalloc;
239};
240
241static void *page_frag_alloc_1k(struct page_frag_1k *nc, gfp_t gfp)
242{
243 struct page *page;
244 int offset;
245
246 offset = nc->offset - SZ_1K;
247 if (likely(offset >= 0))
248 goto use_frag;
249
250 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
251 if (!page)
252 return NULL;
253
254 nc->va = page_address(page);
255 nc->pfmemalloc = page_is_pfmemalloc(page);
256 offset = PAGE_SIZE - SZ_1K;
257 page_ref_add(page, offset / SZ_1K);
258
259use_frag:
260 nc->offset = offset;
261 return nc->va + offset;
262}
263#else
264
265/* the small page is actually unused in this build; add dummy helpers
266 * to please the compiler and avoid later preprocessor's conditionals
267 */
268#define NAPI_HAS_SMALL_PAGE_FRAG 0
269#define NAPI_SMALL_PAGE_PFMEMALLOC(nc) false
270
271struct page_frag_1k {
272};
273
274static void *page_frag_alloc_1k(struct page_frag_1k *nc, gfp_t gfp_mask)
275{
276 return NULL;
277}
278
279#endif
280
281struct napi_alloc_cache {
282 struct page_frag_cache page;
283 struct page_frag_1k page_small;
284 unsigned int skb_count;
285 void *skb_cache[NAPI_SKB_CACHE_SIZE];
286};
287
288static DEFINE_PER_CPU(struct page_frag_cache, netdev_alloc_cache);
289static DEFINE_PER_CPU(struct napi_alloc_cache, napi_alloc_cache);
290
291/* Double check that napi_get_frags() allocates skbs with
292 * skb->head being backed by slab, not a page fragment.
293 * This is to make sure bug fixed in 3226b158e67c
294 * ("net: avoid 32 x truesize under-estimation for tiny skbs")
295 * does not accidentally come back.
296 */
297void napi_get_frags_check(struct napi_struct *napi)
298{
299 struct sk_buff *skb;
300
301 local_bh_disable();
302 skb = napi_get_frags(napi);
303 WARN_ON_ONCE(!NAPI_HAS_SMALL_PAGE_FRAG && skb && skb->head_frag);
304 napi_free_frags(napi);
305 local_bh_enable();
306}
307
308void *__napi_alloc_frag_align(unsigned int fragsz, unsigned int align_mask)
309{
310 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
311
312 fragsz = SKB_DATA_ALIGN(fragsz);
313
314 return __page_frag_alloc_align(&nc->page, fragsz, GFP_ATOMIC,
315 align_mask);
316}
317EXPORT_SYMBOL(__napi_alloc_frag_align);
318
319void *__netdev_alloc_frag_align(unsigned int fragsz, unsigned int align_mask)
320{
321 void *data;
322
323 fragsz = SKB_DATA_ALIGN(fragsz);
324 if (in_hardirq() || irqs_disabled()) {
325 struct page_frag_cache *nc = this_cpu_ptr(&netdev_alloc_cache);
326
327 data = __page_frag_alloc_align(nc, fragsz, GFP_ATOMIC,
328 align_mask);
329 } else {
330 struct napi_alloc_cache *nc;
331
332 local_bh_disable();
333 nc = this_cpu_ptr(&napi_alloc_cache);
334 data = __page_frag_alloc_align(&nc->page, fragsz, GFP_ATOMIC,
335 align_mask);
336 local_bh_enable();
337 }
338 return data;
339}
340EXPORT_SYMBOL(__netdev_alloc_frag_align);
341
342static struct sk_buff *napi_skb_cache_get(void)
343{
344 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
345 struct sk_buff *skb;
346
347 if (unlikely(!nc->skb_count)) {
348 nc->skb_count = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
349 GFP_ATOMIC,
350 NAPI_SKB_CACHE_BULK,
351 nc->skb_cache);
352 if (unlikely(!nc->skb_count))
353 return NULL;
354 }
355
356 skb = nc->skb_cache[--nc->skb_count];
357 kasan_mempool_unpoison_object(skb, kmem_cache_size(net_hotdata.skbuff_cache));
358
359 return skb;
360}
361
362static inline void __finalize_skb_around(struct sk_buff *skb, void *data,
363 unsigned int size)
364{
365 struct skb_shared_info *shinfo;
366
367 size -= SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
368
369 /* Assumes caller memset cleared SKB */
370 skb->truesize = SKB_TRUESIZE(size);
371 refcount_set(&skb->users, 1);
372 skb->head = data;
373 skb->data = data;
374 skb_reset_tail_pointer(skb);
375 skb_set_end_offset(skb, size);
376 skb->mac_header = (typeof(skb->mac_header))~0U;
377 skb->transport_header = (typeof(skb->transport_header))~0U;
378 skb->alloc_cpu = raw_smp_processor_id();
379 /* make sure we initialize shinfo sequentially */
380 shinfo = skb_shinfo(skb);
381 memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
382 atomic_set(&shinfo->dataref, 1);
383
384 skb_set_kcov_handle(skb, kcov_common_handle());
385}
386
387static inline void *__slab_build_skb(struct sk_buff *skb, void *data,
388 unsigned int *size)
389{
390 void *resized;
391
392 /* Must find the allocation size (and grow it to match). */
393 *size = ksize(data);
394 /* krealloc() will immediately return "data" when
395 * "ksize(data)" is requested: it is the existing upper
396 * bounds. As a result, GFP_ATOMIC will be ignored. Note
397 * that this "new" pointer needs to be passed back to the
398 * caller for use so the __alloc_size hinting will be
399 * tracked correctly.
400 */
401 resized = krealloc(data, *size, GFP_ATOMIC);
402 WARN_ON_ONCE(resized != data);
403 return resized;
404}
405
406/* build_skb() variant which can operate on slab buffers.
407 * Note that this should be used sparingly as slab buffers
408 * cannot be combined efficiently by GRO!
409 */
410struct sk_buff *slab_build_skb(void *data)
411{
412 struct sk_buff *skb;
413 unsigned int size;
414
415 skb = kmem_cache_alloc(net_hotdata.skbuff_cache, GFP_ATOMIC);
416 if (unlikely(!skb))
417 return NULL;
418
419 memset(skb, 0, offsetof(struct sk_buff, tail));
420 data = __slab_build_skb(skb, data, &size);
421 __finalize_skb_around(skb, data, size);
422
423 return skb;
424}
425EXPORT_SYMBOL(slab_build_skb);
426
427/* Caller must provide SKB that is memset cleared */
428static void __build_skb_around(struct sk_buff *skb, void *data,
429 unsigned int frag_size)
430{
431 unsigned int size = frag_size;
432
433 /* frag_size == 0 is considered deprecated now. Callers
434 * using slab buffer should use slab_build_skb() instead.
435 */
436 if (WARN_ONCE(size == 0, "Use slab_build_skb() instead"))
437 data = __slab_build_skb(skb, data, &size);
438
439 __finalize_skb_around(skb, data, size);
440}
441
442/**
443 * __build_skb - build a network buffer
444 * @data: data buffer provided by caller
445 * @frag_size: size of data (must not be 0)
446 *
447 * Allocate a new &sk_buff. Caller provides space holding head and
448 * skb_shared_info. @data must have been allocated from the page
449 * allocator or vmalloc(). (A @frag_size of 0 to indicate a kmalloc()
450 * allocation is deprecated, and callers should use slab_build_skb()
451 * instead.)
452 * The return is the new skb buffer.
453 * On a failure the return is %NULL, and @data is not freed.
454 * Notes :
455 * Before IO, driver allocates only data buffer where NIC put incoming frame
456 * Driver should add room at head (NET_SKB_PAD) and
457 * MUST add room at tail (SKB_DATA_ALIGN(skb_shared_info))
458 * After IO, driver calls build_skb(), to allocate sk_buff and populate it
459 * before giving packet to stack.
460 * RX rings only contains data buffers, not full skbs.
461 */
462struct sk_buff *__build_skb(void *data, unsigned int frag_size)
463{
464 struct sk_buff *skb;
465
466 skb = kmem_cache_alloc(net_hotdata.skbuff_cache, GFP_ATOMIC);
467 if (unlikely(!skb))
468 return NULL;
469
470 memset(skb, 0, offsetof(struct sk_buff, tail));
471 __build_skb_around(skb, data, frag_size);
472
473 return skb;
474}
475
476/* build_skb() is wrapper over __build_skb(), that specifically
477 * takes care of skb->head and skb->pfmemalloc
478 */
479struct sk_buff *build_skb(void *data, unsigned int frag_size)
480{
481 struct sk_buff *skb = __build_skb(data, frag_size);
482
483 if (likely(skb && frag_size)) {
484 skb->head_frag = 1;
485 skb_propagate_pfmemalloc(virt_to_head_page(data), skb);
486 }
487 return skb;
488}
489EXPORT_SYMBOL(build_skb);
490
491/**
492 * build_skb_around - build a network buffer around provided skb
493 * @skb: sk_buff provide by caller, must be memset cleared
494 * @data: data buffer provided by caller
495 * @frag_size: size of data
496 */
497struct sk_buff *build_skb_around(struct sk_buff *skb,
498 void *data, unsigned int frag_size)
499{
500 if (unlikely(!skb))
501 return NULL;
502
503 __build_skb_around(skb, data, frag_size);
504
505 if (frag_size) {
506 skb->head_frag = 1;
507 skb_propagate_pfmemalloc(virt_to_head_page(data), skb);
508 }
509 return skb;
510}
511EXPORT_SYMBOL(build_skb_around);
512
513/**
514 * __napi_build_skb - build a network buffer
515 * @data: data buffer provided by caller
516 * @frag_size: size of data
517 *
518 * Version of __build_skb() that uses NAPI percpu caches to obtain
519 * skbuff_head instead of inplace allocation.
520 *
521 * Returns a new &sk_buff on success, %NULL on allocation failure.
522 */
523static struct sk_buff *__napi_build_skb(void *data, unsigned int frag_size)
524{
525 struct sk_buff *skb;
526
527 skb = napi_skb_cache_get();
528 if (unlikely(!skb))
529 return NULL;
530
531 memset(skb, 0, offsetof(struct sk_buff, tail));
532 __build_skb_around(skb, data, frag_size);
533
534 return skb;
535}
536
537/**
538 * napi_build_skb - build a network buffer
539 * @data: data buffer provided by caller
540 * @frag_size: size of data
541 *
542 * Version of __napi_build_skb() that takes care of skb->head_frag
543 * and skb->pfmemalloc when the data is a page or page fragment.
544 *
545 * Returns a new &sk_buff on success, %NULL on allocation failure.
546 */
547struct sk_buff *napi_build_skb(void *data, unsigned int frag_size)
548{
549 struct sk_buff *skb = __napi_build_skb(data, frag_size);
550
551 if (likely(skb) && frag_size) {
552 skb->head_frag = 1;
553 skb_propagate_pfmemalloc(virt_to_head_page(data), skb);
554 }
555
556 return skb;
557}
558EXPORT_SYMBOL(napi_build_skb);
559
560/*
561 * kmalloc_reserve is a wrapper around kmalloc_node_track_caller that tells
562 * the caller if emergency pfmemalloc reserves are being used. If it is and
563 * the socket is later found to be SOCK_MEMALLOC then PFMEMALLOC reserves
564 * may be used. Otherwise, the packet data may be discarded until enough
565 * memory is free
566 */
567static void *kmalloc_reserve(unsigned int *size, gfp_t flags, int node,
568 bool *pfmemalloc)
569{
570 bool ret_pfmemalloc = false;
571 size_t obj_size;
572 void *obj;
573
574 obj_size = SKB_HEAD_ALIGN(*size);
575 if (obj_size <= SKB_SMALL_HEAD_CACHE_SIZE &&
576 !(flags & KMALLOC_NOT_NORMAL_BITS)) {
577 obj = kmem_cache_alloc_node(net_hotdata.skb_small_head_cache,
578 flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
579 node);
580 *size = SKB_SMALL_HEAD_CACHE_SIZE;
581 if (obj || !(gfp_pfmemalloc_allowed(flags)))
582 goto out;
583 /* Try again but now we are using pfmemalloc reserves */
584 ret_pfmemalloc = true;
585 obj = kmem_cache_alloc_node(net_hotdata.skb_small_head_cache, flags, node);
586 goto out;
587 }
588
589 obj_size = kmalloc_size_roundup(obj_size);
590 /* The following cast might truncate high-order bits of obj_size, this
591 * is harmless because kmalloc(obj_size >= 2^32) will fail anyway.
592 */
593 *size = (unsigned int)obj_size;
594
595 /*
596 * Try a regular allocation, when that fails and we're not entitled
597 * to the reserves, fail.
598 */
599 obj = kmalloc_node_track_caller(obj_size,
600 flags | __GFP_NOMEMALLOC | __GFP_NOWARN,
601 node);
602 if (obj || !(gfp_pfmemalloc_allowed(flags)))
603 goto out;
604
605 /* Try again but now we are using pfmemalloc reserves */
606 ret_pfmemalloc = true;
607 obj = kmalloc_node_track_caller(obj_size, flags, node);
608
609out:
610 if (pfmemalloc)
611 *pfmemalloc = ret_pfmemalloc;
612
613 return obj;
614}
615
616/* Allocate a new skbuff. We do this ourselves so we can fill in a few
617 * 'private' fields and also do memory statistics to find all the
618 * [BEEP] leaks.
619 *
620 */
621
622/**
623 * __alloc_skb - allocate a network buffer
624 * @size: size to allocate
625 * @gfp_mask: allocation mask
626 * @flags: If SKB_ALLOC_FCLONE is set, allocate from fclone cache
627 * instead of head cache and allocate a cloned (child) skb.
628 * If SKB_ALLOC_RX is set, __GFP_MEMALLOC will be used for
629 * allocations in case the data is required for writeback
630 * @node: numa node to allocate memory on
631 *
632 * Allocate a new &sk_buff. The returned buffer has no headroom and a
633 * tail room of at least size bytes. The object has a reference count
634 * of one. The return is the buffer. On a failure the return is %NULL.
635 *
636 * Buffers may only be allocated from interrupts using a @gfp_mask of
637 * %GFP_ATOMIC.
638 */
639struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
640 int flags, int node)
641{
642 struct kmem_cache *cache;
643 struct sk_buff *skb;
644 bool pfmemalloc;
645 u8 *data;
646
647 cache = (flags & SKB_ALLOC_FCLONE)
648 ? net_hotdata.skbuff_fclone_cache : net_hotdata.skbuff_cache;
649
650 if (sk_memalloc_socks() && (flags & SKB_ALLOC_RX))
651 gfp_mask |= __GFP_MEMALLOC;
652
653 /* Get the HEAD */
654 if ((flags & (SKB_ALLOC_FCLONE | SKB_ALLOC_NAPI)) == SKB_ALLOC_NAPI &&
655 likely(node == NUMA_NO_NODE || node == numa_mem_id()))
656 skb = napi_skb_cache_get();
657 else
658 skb = kmem_cache_alloc_node(cache, gfp_mask & ~GFP_DMA, node);
659 if (unlikely(!skb))
660 return NULL;
661 prefetchw(skb);
662
663 /* We do our best to align skb_shared_info on a separate cache
664 * line. It usually works because kmalloc(X > SMP_CACHE_BYTES) gives
665 * aligned memory blocks, unless SLUB/SLAB debug is enabled.
666 * Both skb->head and skb_shared_info are cache line aligned.
667 */
668 data = kmalloc_reserve(&size, gfp_mask, node, &pfmemalloc);
669 if (unlikely(!data))
670 goto nodata;
671 /* kmalloc_size_roundup() might give us more room than requested.
672 * Put skb_shared_info exactly at the end of allocated zone,
673 * to allow max possible filling before reallocation.
674 */
675 prefetchw(data + SKB_WITH_OVERHEAD(size));
676
677 /*
678 * Only clear those fields we need to clear, not those that we will
679 * actually initialise below. Hence, don't put any more fields after
680 * the tail pointer in struct sk_buff!
681 */
682 memset(skb, 0, offsetof(struct sk_buff, tail));
683 __build_skb_around(skb, data, size);
684 skb->pfmemalloc = pfmemalloc;
685
686 if (flags & SKB_ALLOC_FCLONE) {
687 struct sk_buff_fclones *fclones;
688
689 fclones = container_of(skb, struct sk_buff_fclones, skb1);
690
691 skb->fclone = SKB_FCLONE_ORIG;
692 refcount_set(&fclones->fclone_ref, 1);
693 }
694
695 return skb;
696
697nodata:
698 kmem_cache_free(cache, skb);
699 return NULL;
700}
701EXPORT_SYMBOL(__alloc_skb);
702
703/**
704 * __netdev_alloc_skb - allocate an skbuff for rx on a specific device
705 * @dev: network device to receive on
706 * @len: length to allocate
707 * @gfp_mask: get_free_pages mask, passed to alloc_skb
708 *
709 * Allocate a new &sk_buff and assign it a usage count of one. The
710 * buffer has NET_SKB_PAD headroom built in. Users should allocate
711 * the headroom they think they need without accounting for the
712 * built in space. The built in space is used for optimisations.
713 *
714 * %NULL is returned if there is no free memory.
715 */
716struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
717 gfp_t gfp_mask)
718{
719 struct page_frag_cache *nc;
720 struct sk_buff *skb;
721 bool pfmemalloc;
722 void *data;
723
724 len += NET_SKB_PAD;
725
726 /* If requested length is either too small or too big,
727 * we use kmalloc() for skb->head allocation.
728 */
729 if (len <= SKB_WITH_OVERHEAD(1024) ||
730 len > SKB_WITH_OVERHEAD(PAGE_SIZE) ||
731 (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
732 skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX, NUMA_NO_NODE);
733 if (!skb)
734 goto skb_fail;
735 goto skb_success;
736 }
737
738 len = SKB_HEAD_ALIGN(len);
739
740 if (sk_memalloc_socks())
741 gfp_mask |= __GFP_MEMALLOC;
742
743 if (in_hardirq() || irqs_disabled()) {
744 nc = this_cpu_ptr(&netdev_alloc_cache);
745 data = page_frag_alloc(nc, len, gfp_mask);
746 pfmemalloc = nc->pfmemalloc;
747 } else {
748 local_bh_disable();
749 nc = this_cpu_ptr(&napi_alloc_cache.page);
750 data = page_frag_alloc(nc, len, gfp_mask);
751 pfmemalloc = nc->pfmemalloc;
752 local_bh_enable();
753 }
754
755 if (unlikely(!data))
756 return NULL;
757
758 skb = __build_skb(data, len);
759 if (unlikely(!skb)) {
760 skb_free_frag(data);
761 return NULL;
762 }
763
764 if (pfmemalloc)
765 skb->pfmemalloc = 1;
766 skb->head_frag = 1;
767
768skb_success:
769 skb_reserve(skb, NET_SKB_PAD);
770 skb->dev = dev;
771
772skb_fail:
773 return skb;
774}
775EXPORT_SYMBOL(__netdev_alloc_skb);
776
777/**
778 * __napi_alloc_skb - allocate skbuff for rx in a specific NAPI instance
779 * @napi: napi instance this buffer was allocated for
780 * @len: length to allocate
781 * @gfp_mask: get_free_pages mask, passed to alloc_skb and alloc_pages
782 *
783 * Allocate a new sk_buff for use in NAPI receive. This buffer will
784 * attempt to allocate the head from a special reserved region used
785 * only for NAPI Rx allocation. By doing this we can save several
786 * CPU cycles by avoiding having to disable and re-enable IRQs.
787 *
788 * %NULL is returned if there is no free memory.
789 */
790struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
791 gfp_t gfp_mask)
792{
793 struct napi_alloc_cache *nc;
794 struct sk_buff *skb;
795 bool pfmemalloc;
796 void *data;
797
798 DEBUG_NET_WARN_ON_ONCE(!in_softirq());
799 len += NET_SKB_PAD + NET_IP_ALIGN;
800
801 /* If requested length is either too small or too big,
802 * we use kmalloc() for skb->head allocation.
803 * When the small frag allocator is available, prefer it over kmalloc
804 * for small fragments
805 */
806 if ((!NAPI_HAS_SMALL_PAGE_FRAG && len <= SKB_WITH_OVERHEAD(1024)) ||
807 len > SKB_WITH_OVERHEAD(PAGE_SIZE) ||
808 (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
809 skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX | SKB_ALLOC_NAPI,
810 NUMA_NO_NODE);
811 if (!skb)
812 goto skb_fail;
813 goto skb_success;
814 }
815
816 nc = this_cpu_ptr(&napi_alloc_cache);
817
818 if (sk_memalloc_socks())
819 gfp_mask |= __GFP_MEMALLOC;
820
821 if (NAPI_HAS_SMALL_PAGE_FRAG && len <= SKB_WITH_OVERHEAD(1024)) {
822 /* we are artificially inflating the allocation size, but
823 * that is not as bad as it may look like, as:
824 * - 'len' less than GRO_MAX_HEAD makes little sense
825 * - On most systems, larger 'len' values lead to fragment
826 * size above 512 bytes
827 * - kmalloc would use the kmalloc-1k slab for such values
828 * - Builds with smaller GRO_MAX_HEAD will very likely do
829 * little networking, as that implies no WiFi and no
830 * tunnels support, and 32 bits arches.
831 */
832 len = SZ_1K;
833
834 data = page_frag_alloc_1k(&nc->page_small, gfp_mask);
835 pfmemalloc = NAPI_SMALL_PAGE_PFMEMALLOC(nc->page_small);
836 } else {
837 len = SKB_HEAD_ALIGN(len);
838
839 data = page_frag_alloc(&nc->page, len, gfp_mask);
840 pfmemalloc = nc->page.pfmemalloc;
841 }
842
843 if (unlikely(!data))
844 return NULL;
845
846 skb = __napi_build_skb(data, len);
847 if (unlikely(!skb)) {
848 skb_free_frag(data);
849 return NULL;
850 }
851
852 if (pfmemalloc)
853 skb->pfmemalloc = 1;
854 skb->head_frag = 1;
855
856skb_success:
857 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
858 skb->dev = napi->dev;
859
860skb_fail:
861 return skb;
862}
863EXPORT_SYMBOL(__napi_alloc_skb);
864
865void skb_add_rx_frag_netmem(struct sk_buff *skb, int i, netmem_ref netmem,
866 int off, int size, unsigned int truesize)
867{
868 DEBUG_NET_WARN_ON_ONCE(size > truesize);
869
870 skb_fill_netmem_desc(skb, i, netmem, off, size);
871 skb->len += size;
872 skb->data_len += size;
873 skb->truesize += truesize;
874}
875EXPORT_SYMBOL(skb_add_rx_frag_netmem);
876
877void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
878 unsigned int truesize)
879{
880 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
881
882 DEBUG_NET_WARN_ON_ONCE(size > truesize);
883
884 skb_frag_size_add(frag, size);
885 skb->len += size;
886 skb->data_len += size;
887 skb->truesize += truesize;
888}
889EXPORT_SYMBOL(skb_coalesce_rx_frag);
890
891static void skb_drop_list(struct sk_buff **listp)
892{
893 kfree_skb_list(*listp);
894 *listp = NULL;
895}
896
897static inline void skb_drop_fraglist(struct sk_buff *skb)
898{
899 skb_drop_list(&skb_shinfo(skb)->frag_list);
900}
901
902static void skb_clone_fraglist(struct sk_buff *skb)
903{
904 struct sk_buff *list;
905
906 skb_walk_frags(skb, list)
907 skb_get(list);
908}
909
910static bool is_pp_page(struct page *page)
911{
912 return (page->pp_magic & ~0x3UL) == PP_SIGNATURE;
913}
914
915int skb_pp_cow_data(struct page_pool *pool, struct sk_buff **pskb,
916 unsigned int headroom)
917{
918#if IS_ENABLED(CONFIG_PAGE_POOL)
919 u32 size, truesize, len, max_head_size, off;
920 struct sk_buff *skb = *pskb, *nskb;
921 int err, i, head_off;
922 void *data;
923
924 /* XDP does not support fraglist so we need to linearize
925 * the skb.
926 */
927 if (skb_has_frag_list(skb))
928 return -EOPNOTSUPP;
929
930 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE - headroom);
931 if (skb->len > max_head_size + MAX_SKB_FRAGS * PAGE_SIZE)
932 return -ENOMEM;
933
934 size = min_t(u32, skb->len, max_head_size);
935 truesize = SKB_HEAD_ALIGN(size) + headroom;
936 data = page_pool_dev_alloc_va(pool, &truesize);
937 if (!data)
938 return -ENOMEM;
939
940 nskb = napi_build_skb(data, truesize);
941 if (!nskb) {
942 page_pool_free_va(pool, data, true);
943 return -ENOMEM;
944 }
945
946 skb_reserve(nskb, headroom);
947 skb_copy_header(nskb, skb);
948 skb_mark_for_recycle(nskb);
949
950 err = skb_copy_bits(skb, 0, nskb->data, size);
951 if (err) {
952 consume_skb(nskb);
953 return err;
954 }
955 skb_put(nskb, size);
956
957 head_off = skb_headroom(nskb) - skb_headroom(skb);
958 skb_headers_offset_update(nskb, head_off);
959
960 off = size;
961 len = skb->len - off;
962 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
963 struct page *page;
964 u32 page_off;
965
966 size = min_t(u32, len, PAGE_SIZE);
967 truesize = size;
968
969 page = page_pool_dev_alloc(pool, &page_off, &truesize);
970 if (!page) {
971 consume_skb(nskb);
972 return -ENOMEM;
973 }
974
975 skb_add_rx_frag(nskb, i, page, page_off, size, truesize);
976 err = skb_copy_bits(skb, off, page_address(page) + page_off,
977 size);
978 if (err) {
979 consume_skb(nskb);
980 return err;
981 }
982
983 len -= size;
984 off += size;
985 }
986
987 consume_skb(skb);
988 *pskb = nskb;
989
990 return 0;
991#else
992 return -EOPNOTSUPP;
993#endif
994}
995EXPORT_SYMBOL(skb_pp_cow_data);
996
997int skb_cow_data_for_xdp(struct page_pool *pool, struct sk_buff **pskb,
998 struct bpf_prog *prog)
999{
1000 if (!prog->aux->xdp_has_frags)
1001 return -EINVAL;
1002
1003 return skb_pp_cow_data(pool, pskb, XDP_PACKET_HEADROOM);
1004}
1005EXPORT_SYMBOL(skb_cow_data_for_xdp);
1006
1007#if IS_ENABLED(CONFIG_PAGE_POOL)
1008bool napi_pp_put_page(struct page *page, bool napi_safe)
1009{
1010 bool allow_direct = false;
1011 struct page_pool *pp;
1012
1013 page = compound_head(page);
1014
1015 /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
1016 * in order to preserve any existing bits, such as bit 0 for the
1017 * head page of compound page and bit 1 for pfmemalloc page, so
1018 * mask those bits for freeing side when doing below checking,
1019 * and page_is_pfmemalloc() is checked in __page_pool_put_page()
1020 * to avoid recycling the pfmemalloc page.
1021 */
1022 if (unlikely(!is_pp_page(page)))
1023 return false;
1024
1025 pp = page->pp;
1026
1027 /* Allow direct recycle if we have reasons to believe that we are
1028 * in the same context as the consumer would run, so there's
1029 * no possible race.
1030 * __page_pool_put_page() makes sure we're not in hardirq context
1031 * and interrupts are enabled prior to accessing the cache.
1032 */
1033 if (napi_safe || in_softirq()) {
1034 const struct napi_struct *napi = READ_ONCE(pp->p.napi);
1035 unsigned int cpuid = smp_processor_id();
1036
1037 allow_direct = napi && READ_ONCE(napi->list_owner) == cpuid;
1038 allow_direct |= READ_ONCE(pp->cpuid) == cpuid;
1039 }
1040
1041 /* Driver set this to memory recycling info. Reset it on recycle.
1042 * This will *not* work for NIC using a split-page memory model.
1043 * The page will be returned to the pool here regardless of the
1044 * 'flipped' fragment being in use or not.
1045 */
1046 page_pool_put_full_page(pp, page, allow_direct);
1047
1048 return true;
1049}
1050EXPORT_SYMBOL(napi_pp_put_page);
1051#endif
1052
1053static bool skb_pp_recycle(struct sk_buff *skb, void *data, bool napi_safe)
1054{
1055 if (!IS_ENABLED(CONFIG_PAGE_POOL) || !skb->pp_recycle)
1056 return false;
1057 return napi_pp_put_page(virt_to_page(data), napi_safe);
1058}
1059
1060/**
1061 * skb_pp_frag_ref() - Increase fragment references of a page pool aware skb
1062 * @skb: page pool aware skb
1063 *
1064 * Increase the fragment reference count (pp_ref_count) of a skb. This is
1065 * intended to gain fragment references only for page pool aware skbs,
1066 * i.e. when skb->pp_recycle is true, and not for fragments in a
1067 * non-pp-recycling skb. It has a fallback to increase references on normal
1068 * pages, as page pool aware skbs may also have normal page fragments.
1069 */
1070static int skb_pp_frag_ref(struct sk_buff *skb)
1071{
1072 struct skb_shared_info *shinfo;
1073 struct page *head_page;
1074 int i;
1075
1076 if (!skb->pp_recycle)
1077 return -EINVAL;
1078
1079 shinfo = skb_shinfo(skb);
1080
1081 for (i = 0; i < shinfo->nr_frags; i++) {
1082 head_page = compound_head(skb_frag_page(&shinfo->frags[i]));
1083 if (likely(is_pp_page(head_page)))
1084 page_pool_ref_page(head_page);
1085 else
1086 page_ref_inc(head_page);
1087 }
1088 return 0;
1089}
1090
1091static void skb_kfree_head(void *head, unsigned int end_offset)
1092{
1093 if (end_offset == SKB_SMALL_HEAD_HEADROOM)
1094 kmem_cache_free(net_hotdata.skb_small_head_cache, head);
1095 else
1096 kfree(head);
1097}
1098
1099static void skb_free_head(struct sk_buff *skb, bool napi_safe)
1100{
1101 unsigned char *head = skb->head;
1102
1103 if (skb->head_frag) {
1104 if (skb_pp_recycle(skb, head, napi_safe))
1105 return;
1106 skb_free_frag(head);
1107 } else {
1108 skb_kfree_head(head, skb_end_offset(skb));
1109 }
1110}
1111
1112static void skb_release_data(struct sk_buff *skb, enum skb_drop_reason reason,
1113 bool napi_safe)
1114{
1115 struct skb_shared_info *shinfo = skb_shinfo(skb);
1116 int i;
1117
1118 if (!skb_data_unref(skb, shinfo))
1119 goto exit;
1120
1121 if (skb_zcopy(skb)) {
1122 bool skip_unref = shinfo->flags & SKBFL_MANAGED_FRAG_REFS;
1123
1124 skb_zcopy_clear(skb, true);
1125 if (skip_unref)
1126 goto free_head;
1127 }
1128
1129 for (i = 0; i < shinfo->nr_frags; i++)
1130 napi_frag_unref(&shinfo->frags[i], skb->pp_recycle, napi_safe);
1131
1132free_head:
1133 if (shinfo->frag_list)
1134 kfree_skb_list_reason(shinfo->frag_list, reason);
1135
1136 skb_free_head(skb, napi_safe);
1137exit:
1138 /* When we clone an SKB we copy the reycling bit. The pp_recycle
1139 * bit is only set on the head though, so in order to avoid races
1140 * while trying to recycle fragments on __skb_frag_unref() we need
1141 * to make one SKB responsible for triggering the recycle path.
1142 * So disable the recycling bit if an SKB is cloned and we have
1143 * additional references to the fragmented part of the SKB.
1144 * Eventually the last SKB will have the recycling bit set and it's
1145 * dataref set to 0, which will trigger the recycling
1146 */
1147 skb->pp_recycle = 0;
1148}
1149
1150/*
1151 * Free an skbuff by memory without cleaning the state.
1152 */
1153static void kfree_skbmem(struct sk_buff *skb)
1154{
1155 struct sk_buff_fclones *fclones;
1156
1157 switch (skb->fclone) {
1158 case SKB_FCLONE_UNAVAILABLE:
1159 kmem_cache_free(net_hotdata.skbuff_cache, skb);
1160 return;
1161
1162 case SKB_FCLONE_ORIG:
1163 fclones = container_of(skb, struct sk_buff_fclones, skb1);
1164
1165 /* We usually free the clone (TX completion) before original skb
1166 * This test would have no chance to be true for the clone,
1167 * while here, branch prediction will be good.
1168 */
1169 if (refcount_read(&fclones->fclone_ref) == 1)
1170 goto fastpath;
1171 break;
1172
1173 default: /* SKB_FCLONE_CLONE */
1174 fclones = container_of(skb, struct sk_buff_fclones, skb2);
1175 break;
1176 }
1177 if (!refcount_dec_and_test(&fclones->fclone_ref))
1178 return;
1179fastpath:
1180 kmem_cache_free(net_hotdata.skbuff_fclone_cache, fclones);
1181}
1182
1183void skb_release_head_state(struct sk_buff *skb)
1184{
1185 skb_dst_drop(skb);
1186 if (skb->destructor) {
1187 DEBUG_NET_WARN_ON_ONCE(in_hardirq());
1188 skb->destructor(skb);
1189 }
1190#if IS_ENABLED(CONFIG_NF_CONNTRACK)
1191 nf_conntrack_put(skb_nfct(skb));
1192#endif
1193 skb_ext_put(skb);
1194}
1195
1196/* Free everything but the sk_buff shell. */
1197static void skb_release_all(struct sk_buff *skb, enum skb_drop_reason reason,
1198 bool napi_safe)
1199{
1200 skb_release_head_state(skb);
1201 if (likely(skb->head))
1202 skb_release_data(skb, reason, napi_safe);
1203}
1204
1205/**
1206 * __kfree_skb - private function
1207 * @skb: buffer
1208 *
1209 * Free an sk_buff. Release anything attached to the buffer.
1210 * Clean the state. This is an internal helper function. Users should
1211 * always call kfree_skb
1212 */
1213
1214void __kfree_skb(struct sk_buff *skb)
1215{
1216 skb_release_all(skb, SKB_DROP_REASON_NOT_SPECIFIED, false);
1217 kfree_skbmem(skb);
1218}
1219EXPORT_SYMBOL(__kfree_skb);
1220
1221static __always_inline
1222bool __kfree_skb_reason(struct sk_buff *skb, enum skb_drop_reason reason)
1223{
1224 if (unlikely(!skb_unref(skb)))
1225 return false;
1226
1227 DEBUG_NET_WARN_ON_ONCE(reason == SKB_NOT_DROPPED_YET ||
1228 u32_get_bits(reason,
1229 SKB_DROP_REASON_SUBSYS_MASK) >=
1230 SKB_DROP_REASON_SUBSYS_NUM);
1231
1232 if (reason == SKB_CONSUMED)
1233 trace_consume_skb(skb, __builtin_return_address(0));
1234 else
1235 trace_kfree_skb(skb, __builtin_return_address(0), reason);
1236 return true;
1237}
1238
1239/**
1240 * kfree_skb_reason - free an sk_buff with special reason
1241 * @skb: buffer to free
1242 * @reason: reason why this skb is dropped
1243 *
1244 * Drop a reference to the buffer and free it if the usage count has
1245 * hit zero. Meanwhile, pass the drop reason to 'kfree_skb'
1246 * tracepoint.
1247 */
1248void __fix_address
1249kfree_skb_reason(struct sk_buff *skb, enum skb_drop_reason reason)
1250{
1251 if (__kfree_skb_reason(skb, reason))
1252 __kfree_skb(skb);
1253}
1254EXPORT_SYMBOL(kfree_skb_reason);
1255
1256#define KFREE_SKB_BULK_SIZE 16
1257
1258struct skb_free_array {
1259 unsigned int skb_count;
1260 void *skb_array[KFREE_SKB_BULK_SIZE];
1261};
1262
1263static void kfree_skb_add_bulk(struct sk_buff *skb,
1264 struct skb_free_array *sa,
1265 enum skb_drop_reason reason)
1266{
1267 /* if SKB is a clone, don't handle this case */
1268 if (unlikely(skb->fclone != SKB_FCLONE_UNAVAILABLE)) {
1269 __kfree_skb(skb);
1270 return;
1271 }
1272
1273 skb_release_all(skb, reason, false);
1274 sa->skb_array[sa->skb_count++] = skb;
1275
1276 if (unlikely(sa->skb_count == KFREE_SKB_BULK_SIZE)) {
1277 kmem_cache_free_bulk(net_hotdata.skbuff_cache, KFREE_SKB_BULK_SIZE,
1278 sa->skb_array);
1279 sa->skb_count = 0;
1280 }
1281}
1282
1283void __fix_address
1284kfree_skb_list_reason(struct sk_buff *segs, enum skb_drop_reason reason)
1285{
1286 struct skb_free_array sa;
1287
1288 sa.skb_count = 0;
1289
1290 while (segs) {
1291 struct sk_buff *next = segs->next;
1292
1293 if (__kfree_skb_reason(segs, reason)) {
1294 skb_poison_list(segs);
1295 kfree_skb_add_bulk(segs, &sa, reason);
1296 }
1297
1298 segs = next;
1299 }
1300
1301 if (sa.skb_count)
1302 kmem_cache_free_bulk(net_hotdata.skbuff_cache, sa.skb_count, sa.skb_array);
1303}
1304EXPORT_SYMBOL(kfree_skb_list_reason);
1305
1306/* Dump skb information and contents.
1307 *
1308 * Must only be called from net_ratelimit()-ed paths.
1309 *
1310 * Dumps whole packets if full_pkt, only headers otherwise.
1311 */
1312void skb_dump(const char *level, const struct sk_buff *skb, bool full_pkt)
1313{
1314 struct skb_shared_info *sh = skb_shinfo(skb);
1315 struct net_device *dev = skb->dev;
1316 struct sock *sk = skb->sk;
1317 struct sk_buff *list_skb;
1318 bool has_mac, has_trans;
1319 int headroom, tailroom;
1320 int i, len, seg_len;
1321
1322 if (full_pkt)
1323 len = skb->len;
1324 else
1325 len = min_t(int, skb->len, MAX_HEADER + 128);
1326
1327 headroom = skb_headroom(skb);
1328 tailroom = skb_tailroom(skb);
1329
1330 has_mac = skb_mac_header_was_set(skb);
1331 has_trans = skb_transport_header_was_set(skb);
1332
1333 printk("%sskb len=%u headroom=%u headlen=%u tailroom=%u\n"
1334 "mac=(%d,%d) net=(%d,%d) trans=%d\n"
1335 "shinfo(txflags=%u nr_frags=%u gso(size=%hu type=%u segs=%hu))\n"
1336 "csum(0x%x ip_summed=%u complete_sw=%u valid=%u level=%u)\n"
1337 "hash(0x%x sw=%u l4=%u) proto=0x%04x pkttype=%u iif=%d\n",
1338 level, skb->len, headroom, skb_headlen(skb), tailroom,
1339 has_mac ? skb->mac_header : -1,
1340 has_mac ? skb_mac_header_len(skb) : -1,
1341 skb->network_header,
1342 has_trans ? skb_network_header_len(skb) : -1,
1343 has_trans ? skb->transport_header : -1,
1344 sh->tx_flags, sh->nr_frags,
1345 sh->gso_size, sh->gso_type, sh->gso_segs,
1346 skb->csum, skb->ip_summed, skb->csum_complete_sw,
1347 skb->csum_valid, skb->csum_level,
1348 skb->hash, skb->sw_hash, skb->l4_hash,
1349 ntohs(skb->protocol), skb->pkt_type, skb->skb_iif);
1350
1351 if (dev)
1352 printk("%sdev name=%s feat=%pNF\n",
1353 level, dev->name, &dev->features);
1354 if (sk)
1355 printk("%ssk family=%hu type=%u proto=%u\n",
1356 level, sk->sk_family, sk->sk_type, sk->sk_protocol);
1357
1358 if (full_pkt && headroom)
1359 print_hex_dump(level, "skb headroom: ", DUMP_PREFIX_OFFSET,
1360 16, 1, skb->head, headroom, false);
1361
1362 seg_len = min_t(int, skb_headlen(skb), len);
1363 if (seg_len)
1364 print_hex_dump(level, "skb linear: ", DUMP_PREFIX_OFFSET,
1365 16, 1, skb->data, seg_len, false);
1366 len -= seg_len;
1367
1368 if (full_pkt && tailroom)
1369 print_hex_dump(level, "skb tailroom: ", DUMP_PREFIX_OFFSET,
1370 16, 1, skb_tail_pointer(skb), tailroom, false);
1371
1372 for (i = 0; len && i < skb_shinfo(skb)->nr_frags; i++) {
1373 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1374 u32 p_off, p_len, copied;
1375 struct page *p;
1376 u8 *vaddr;
1377
1378 skb_frag_foreach_page(frag, skb_frag_off(frag),
1379 skb_frag_size(frag), p, p_off, p_len,
1380 copied) {
1381 seg_len = min_t(int, p_len, len);
1382 vaddr = kmap_atomic(p);
1383 print_hex_dump(level, "skb frag: ",
1384 DUMP_PREFIX_OFFSET,
1385 16, 1, vaddr + p_off, seg_len, false);
1386 kunmap_atomic(vaddr);
1387 len -= seg_len;
1388 if (!len)
1389 break;
1390 }
1391 }
1392
1393 if (full_pkt && skb_has_frag_list(skb)) {
1394 printk("skb fraglist:\n");
1395 skb_walk_frags(skb, list_skb)
1396 skb_dump(level, list_skb, true);
1397 }
1398}
1399EXPORT_SYMBOL(skb_dump);
1400
1401/**
1402 * skb_tx_error - report an sk_buff xmit error
1403 * @skb: buffer that triggered an error
1404 *
1405 * Report xmit error if a device callback is tracking this skb.
1406 * skb must be freed afterwards.
1407 */
1408void skb_tx_error(struct sk_buff *skb)
1409{
1410 if (skb) {
1411 skb_zcopy_downgrade_managed(skb);
1412 skb_zcopy_clear(skb, true);
1413 }
1414}
1415EXPORT_SYMBOL(skb_tx_error);
1416
1417#ifdef CONFIG_TRACEPOINTS
1418/**
1419 * consume_skb - free an skbuff
1420 * @skb: buffer to free
1421 *
1422 * Drop a ref to the buffer and free it if the usage count has hit zero
1423 * Functions identically to kfree_skb, but kfree_skb assumes that the frame
1424 * is being dropped after a failure and notes that
1425 */
1426void consume_skb(struct sk_buff *skb)
1427{
1428 if (!skb_unref(skb))
1429 return;
1430
1431 trace_consume_skb(skb, __builtin_return_address(0));
1432 __kfree_skb(skb);
1433}
1434EXPORT_SYMBOL(consume_skb);
1435#endif
1436
1437/**
1438 * __consume_stateless_skb - free an skbuff, assuming it is stateless
1439 * @skb: buffer to free
1440 *
1441 * Alike consume_skb(), but this variant assumes that this is the last
1442 * skb reference and all the head states have been already dropped
1443 */
1444void __consume_stateless_skb(struct sk_buff *skb)
1445{
1446 trace_consume_skb(skb, __builtin_return_address(0));
1447 skb_release_data(skb, SKB_CONSUMED, false);
1448 kfree_skbmem(skb);
1449}
1450
1451static void napi_skb_cache_put(struct sk_buff *skb)
1452{
1453 struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
1454 u32 i;
1455
1456 if (!kasan_mempool_poison_object(skb))
1457 return;
1458
1459 nc->skb_cache[nc->skb_count++] = skb;
1460
1461 if (unlikely(nc->skb_count == NAPI_SKB_CACHE_SIZE)) {
1462 for (i = NAPI_SKB_CACHE_HALF; i < NAPI_SKB_CACHE_SIZE; i++)
1463 kasan_mempool_unpoison_object(nc->skb_cache[i],
1464 kmem_cache_size(net_hotdata.skbuff_cache));
1465
1466 kmem_cache_free_bulk(net_hotdata.skbuff_cache, NAPI_SKB_CACHE_HALF,
1467 nc->skb_cache + NAPI_SKB_CACHE_HALF);
1468 nc->skb_count = NAPI_SKB_CACHE_HALF;
1469 }
1470}
1471
1472void __napi_kfree_skb(struct sk_buff *skb, enum skb_drop_reason reason)
1473{
1474 skb_release_all(skb, reason, true);
1475 napi_skb_cache_put(skb);
1476}
1477
1478void napi_skb_free_stolen_head(struct sk_buff *skb)
1479{
1480 if (unlikely(skb->slow_gro)) {
1481 nf_reset_ct(skb);
1482 skb_dst_drop(skb);
1483 skb_ext_put(skb);
1484 skb_orphan(skb);
1485 skb->slow_gro = 0;
1486 }
1487 napi_skb_cache_put(skb);
1488}
1489
1490void napi_consume_skb(struct sk_buff *skb, int budget)
1491{
1492 /* Zero budget indicate non-NAPI context called us, like netpoll */
1493 if (unlikely(!budget)) {
1494 dev_consume_skb_any(skb);
1495 return;
1496 }
1497
1498 DEBUG_NET_WARN_ON_ONCE(!in_softirq());
1499
1500 if (!skb_unref(skb))
1501 return;
1502
1503 /* if reaching here SKB is ready to free */
1504 trace_consume_skb(skb, __builtin_return_address(0));
1505
1506 /* if SKB is a clone, don't handle this case */
1507 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) {
1508 __kfree_skb(skb);
1509 return;
1510 }
1511
1512 skb_release_all(skb, SKB_CONSUMED, !!budget);
1513 napi_skb_cache_put(skb);
1514}
1515EXPORT_SYMBOL(napi_consume_skb);
1516
1517/* Make sure a field is contained by headers group */
1518#define CHECK_SKB_FIELD(field) \
1519 BUILD_BUG_ON(offsetof(struct sk_buff, field) != \
1520 offsetof(struct sk_buff, headers.field)); \
1521
1522static void __copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
1523{
1524 new->tstamp = old->tstamp;
1525 /* We do not copy old->sk */
1526 new->dev = old->dev;
1527 memcpy(new->cb, old->cb, sizeof(old->cb));
1528 skb_dst_copy(new, old);
1529 __skb_ext_copy(new, old);
1530 __nf_copy(new, old, false);
1531
1532 /* Note : this field could be in the headers group.
1533 * It is not yet because we do not want to have a 16 bit hole
1534 */
1535 new->queue_mapping = old->queue_mapping;
1536
1537 memcpy(&new->headers, &old->headers, sizeof(new->headers));
1538 CHECK_SKB_FIELD(protocol);
1539 CHECK_SKB_FIELD(csum);
1540 CHECK_SKB_FIELD(hash);
1541 CHECK_SKB_FIELD(priority);
1542 CHECK_SKB_FIELD(skb_iif);
1543 CHECK_SKB_FIELD(vlan_proto);
1544 CHECK_SKB_FIELD(vlan_tci);
1545 CHECK_SKB_FIELD(transport_header);
1546 CHECK_SKB_FIELD(network_header);
1547 CHECK_SKB_FIELD(mac_header);
1548 CHECK_SKB_FIELD(inner_protocol);
1549 CHECK_SKB_FIELD(inner_transport_header);
1550 CHECK_SKB_FIELD(inner_network_header);
1551 CHECK_SKB_FIELD(inner_mac_header);
1552 CHECK_SKB_FIELD(mark);
1553#ifdef CONFIG_NETWORK_SECMARK
1554 CHECK_SKB_FIELD(secmark);
1555#endif
1556#ifdef CONFIG_NET_RX_BUSY_POLL
1557 CHECK_SKB_FIELD(napi_id);
1558#endif
1559 CHECK_SKB_FIELD(alloc_cpu);
1560#ifdef CONFIG_XPS
1561 CHECK_SKB_FIELD(sender_cpu);
1562#endif
1563#ifdef CONFIG_NET_SCHED
1564 CHECK_SKB_FIELD(tc_index);
1565#endif
1566
1567}
1568
1569/*
1570 * You should not add any new code to this function. Add it to
1571 * __copy_skb_header above instead.
1572 */
1573static struct sk_buff *__skb_clone(struct sk_buff *n, struct sk_buff *skb)
1574{
1575#define C(x) n->x = skb->x
1576
1577 n->next = n->prev = NULL;
1578 n->sk = NULL;
1579 __copy_skb_header(n, skb);
1580
1581 C(len);
1582 C(data_len);
1583 C(mac_len);
1584 n->hdr_len = skb->nohdr ? skb_headroom(skb) : skb->hdr_len;
1585 n->cloned = 1;
1586 n->nohdr = 0;
1587 n->peeked = 0;
1588 C(pfmemalloc);
1589 C(pp_recycle);
1590 n->destructor = NULL;
1591 C(tail);
1592 C(end);
1593 C(head);
1594 C(head_frag);
1595 C(data);
1596 C(truesize);
1597 refcount_set(&n->users, 1);
1598
1599 atomic_inc(&(skb_shinfo(skb)->dataref));
1600 skb->cloned = 1;
1601
1602 return n;
1603#undef C
1604}
1605
1606/**
1607 * alloc_skb_for_msg() - allocate sk_buff to wrap frag list forming a msg
1608 * @first: first sk_buff of the msg
1609 */
1610struct sk_buff *alloc_skb_for_msg(struct sk_buff *first)
1611{
1612 struct sk_buff *n;
1613
1614 n = alloc_skb(0, GFP_ATOMIC);
1615 if (!n)
1616 return NULL;
1617
1618 n->len = first->len;
1619 n->data_len = first->len;
1620 n->truesize = first->truesize;
1621
1622 skb_shinfo(n)->frag_list = first;
1623
1624 __copy_skb_header(n, first);
1625 n->destructor = NULL;
1626
1627 return n;
1628}
1629EXPORT_SYMBOL_GPL(alloc_skb_for_msg);
1630
1631/**
1632 * skb_morph - morph one skb into another
1633 * @dst: the skb to receive the contents
1634 * @src: the skb to supply the contents
1635 *
1636 * This is identical to skb_clone except that the target skb is
1637 * supplied by the user.
1638 *
1639 * The target skb is returned upon exit.
1640 */
1641struct sk_buff *skb_morph(struct sk_buff *dst, struct sk_buff *src)
1642{
1643 skb_release_all(dst, SKB_CONSUMED, false);
1644 return __skb_clone(dst, src);
1645}
1646EXPORT_SYMBOL_GPL(skb_morph);
1647
1648int mm_account_pinned_pages(struct mmpin *mmp, size_t size)
1649{
1650 unsigned long max_pg, num_pg, new_pg, old_pg, rlim;
1651 struct user_struct *user;
1652
1653 if (capable(CAP_IPC_LOCK) || !size)
1654 return 0;
1655
1656 rlim = rlimit(RLIMIT_MEMLOCK);
1657 if (rlim == RLIM_INFINITY)
1658 return 0;
1659
1660 num_pg = (size >> PAGE_SHIFT) + 2; /* worst case */
1661 max_pg = rlim >> PAGE_SHIFT;
1662 user = mmp->user ? : current_user();
1663
1664 old_pg = atomic_long_read(&user->locked_vm);
1665 do {
1666 new_pg = old_pg + num_pg;
1667 if (new_pg > max_pg)
1668 return -ENOBUFS;
1669 } while (!atomic_long_try_cmpxchg(&user->locked_vm, &old_pg, new_pg));
1670
1671 if (!mmp->user) {
1672 mmp->user = get_uid(user);
1673 mmp->num_pg = num_pg;
1674 } else {
1675 mmp->num_pg += num_pg;
1676 }
1677
1678 return 0;
1679}
1680EXPORT_SYMBOL_GPL(mm_account_pinned_pages);
1681
1682void mm_unaccount_pinned_pages(struct mmpin *mmp)
1683{
1684 if (mmp->user) {
1685 atomic_long_sub(mmp->num_pg, &mmp->user->locked_vm);
1686 free_uid(mmp->user);
1687 }
1688}
1689EXPORT_SYMBOL_GPL(mm_unaccount_pinned_pages);
1690
1691static struct ubuf_info *msg_zerocopy_alloc(struct sock *sk, size_t size)
1692{
1693 struct ubuf_info_msgzc *uarg;
1694 struct sk_buff *skb;
1695
1696 WARN_ON_ONCE(!in_task());
1697
1698 skb = sock_omalloc(sk, 0, GFP_KERNEL);
1699 if (!skb)
1700 return NULL;
1701
1702 BUILD_BUG_ON(sizeof(*uarg) > sizeof(skb->cb));
1703 uarg = (void *)skb->cb;
1704 uarg->mmp.user = NULL;
1705
1706 if (mm_account_pinned_pages(&uarg->mmp, size)) {
1707 kfree_skb(skb);
1708 return NULL;
1709 }
1710
1711 uarg->ubuf.callback = msg_zerocopy_callback;
1712 uarg->id = ((u32)atomic_inc_return(&sk->sk_zckey)) - 1;
1713 uarg->len = 1;
1714 uarg->bytelen = size;
1715 uarg->zerocopy = 1;
1716 uarg->ubuf.flags = SKBFL_ZEROCOPY_FRAG | SKBFL_DONT_ORPHAN;
1717 refcount_set(&uarg->ubuf.refcnt, 1);
1718 sock_hold(sk);
1719
1720 return &uarg->ubuf;
1721}
1722
1723static inline struct sk_buff *skb_from_uarg(struct ubuf_info_msgzc *uarg)
1724{
1725 return container_of((void *)uarg, struct sk_buff, cb);
1726}
1727
1728struct ubuf_info *msg_zerocopy_realloc(struct sock *sk, size_t size,
1729 struct ubuf_info *uarg)
1730{
1731 if (uarg) {
1732 struct ubuf_info_msgzc *uarg_zc;
1733 const u32 byte_limit = 1 << 19; /* limit to a few TSO */
1734 u32 bytelen, next;
1735
1736 /* there might be non MSG_ZEROCOPY users */
1737 if (uarg->callback != msg_zerocopy_callback)
1738 return NULL;
1739
1740 /* realloc only when socket is locked (TCP, UDP cork),
1741 * so uarg->len and sk_zckey access is serialized
1742 */
1743 if (!sock_owned_by_user(sk)) {
1744 WARN_ON_ONCE(1);
1745 return NULL;
1746 }
1747
1748 uarg_zc = uarg_to_msgzc(uarg);
1749 bytelen = uarg_zc->bytelen + size;
1750 if (uarg_zc->len == USHRT_MAX - 1 || bytelen > byte_limit) {
1751 /* TCP can create new skb to attach new uarg */
1752 if (sk->sk_type == SOCK_STREAM)
1753 goto new_alloc;
1754 return NULL;
1755 }
1756
1757 next = (u32)atomic_read(&sk->sk_zckey);
1758 if ((u32)(uarg_zc->id + uarg_zc->len) == next) {
1759 if (mm_account_pinned_pages(&uarg_zc->mmp, size))
1760 return NULL;
1761 uarg_zc->len++;
1762 uarg_zc->bytelen = bytelen;
1763 atomic_set(&sk->sk_zckey, ++next);
1764
1765 /* no extra ref when appending to datagram (MSG_MORE) */
1766 if (sk->sk_type == SOCK_STREAM)
1767 net_zcopy_get(uarg);
1768
1769 return uarg;
1770 }
1771 }
1772
1773new_alloc:
1774 return msg_zerocopy_alloc(sk, size);
1775}
1776EXPORT_SYMBOL_GPL(msg_zerocopy_realloc);
1777
1778static bool skb_zerocopy_notify_extend(struct sk_buff *skb, u32 lo, u16 len)
1779{
1780 struct sock_exterr_skb *serr = SKB_EXT_ERR(skb);
1781 u32 old_lo, old_hi;
1782 u64 sum_len;
1783
1784 old_lo = serr->ee.ee_info;
1785 old_hi = serr->ee.ee_data;
1786 sum_len = old_hi - old_lo + 1ULL + len;
1787
1788 if (sum_len >= (1ULL << 32))
1789 return false;
1790
1791 if (lo != old_hi + 1)
1792 return false;
1793
1794 serr->ee.ee_data += len;
1795 return true;
1796}
1797
1798static void __msg_zerocopy_callback(struct ubuf_info_msgzc *uarg)
1799{
1800 struct sk_buff *tail, *skb = skb_from_uarg(uarg);
1801 struct sock_exterr_skb *serr;
1802 struct sock *sk = skb->sk;
1803 struct sk_buff_head *q;
1804 unsigned long flags;
1805 bool is_zerocopy;
1806 u32 lo, hi;
1807 u16 len;
1808
1809 mm_unaccount_pinned_pages(&uarg->mmp);
1810
1811 /* if !len, there was only 1 call, and it was aborted
1812 * so do not queue a completion notification
1813 */
1814 if (!uarg->len || sock_flag(sk, SOCK_DEAD))
1815 goto release;
1816
1817 len = uarg->len;
1818 lo = uarg->id;
1819 hi = uarg->id + len - 1;
1820 is_zerocopy = uarg->zerocopy;
1821
1822 serr = SKB_EXT_ERR(skb);
1823 memset(serr, 0, sizeof(*serr));
1824 serr->ee.ee_errno = 0;
1825 serr->ee.ee_origin = SO_EE_ORIGIN_ZEROCOPY;
1826 serr->ee.ee_data = hi;
1827 serr->ee.ee_info = lo;
1828 if (!is_zerocopy)
1829 serr->ee.ee_code |= SO_EE_CODE_ZEROCOPY_COPIED;
1830
1831 q = &sk->sk_error_queue;
1832 spin_lock_irqsave(&q->lock, flags);
1833 tail = skb_peek_tail(q);
1834 if (!tail || SKB_EXT_ERR(tail)->ee.ee_origin != SO_EE_ORIGIN_ZEROCOPY ||
1835 !skb_zerocopy_notify_extend(tail, lo, len)) {
1836 __skb_queue_tail(q, skb);
1837 skb = NULL;
1838 }
1839 spin_unlock_irqrestore(&q->lock, flags);
1840
1841 sk_error_report(sk);
1842
1843release:
1844 consume_skb(skb);
1845 sock_put(sk);
1846}
1847
1848void msg_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *uarg,
1849 bool success)
1850{
1851 struct ubuf_info_msgzc *uarg_zc = uarg_to_msgzc(uarg);
1852
1853 uarg_zc->zerocopy = uarg_zc->zerocopy & success;
1854
1855 if (refcount_dec_and_test(&uarg->refcnt))
1856 __msg_zerocopy_callback(uarg_zc);
1857}
1858EXPORT_SYMBOL_GPL(msg_zerocopy_callback);
1859
1860void msg_zerocopy_put_abort(struct ubuf_info *uarg, bool have_uref)
1861{
1862 struct sock *sk = skb_from_uarg(uarg_to_msgzc(uarg))->sk;
1863
1864 atomic_dec(&sk->sk_zckey);
1865 uarg_to_msgzc(uarg)->len--;
1866
1867 if (have_uref)
1868 msg_zerocopy_callback(NULL, uarg, true);
1869}
1870EXPORT_SYMBOL_GPL(msg_zerocopy_put_abort);
1871
1872int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
1873 struct msghdr *msg, int len,
1874 struct ubuf_info *uarg)
1875{
1876 struct ubuf_info *orig_uarg = skb_zcopy(skb);
1877 int err, orig_len = skb->len;
1878
1879 /* An skb can only point to one uarg. This edge case happens when
1880 * TCP appends to an skb, but zerocopy_realloc triggered a new alloc.
1881 */
1882 if (orig_uarg && uarg != orig_uarg)
1883 return -EEXIST;
1884
1885 err = __zerocopy_sg_from_iter(msg, sk, skb, &msg->msg_iter, len);
1886 if (err == -EFAULT || (err == -EMSGSIZE && skb->len == orig_len)) {
1887 struct sock *save_sk = skb->sk;
1888
1889 /* Streams do not free skb on error. Reset to prev state. */
1890 iov_iter_revert(&msg->msg_iter, skb->len - orig_len);
1891 skb->sk = sk;
1892 ___pskb_trim(skb, orig_len);
1893 skb->sk = save_sk;
1894 return err;
1895 }
1896
1897 skb_zcopy_set(skb, uarg, NULL);
1898 return skb->len - orig_len;
1899}
1900EXPORT_SYMBOL_GPL(skb_zerocopy_iter_stream);
1901
1902void __skb_zcopy_downgrade_managed(struct sk_buff *skb)
1903{
1904 int i;
1905
1906 skb_shinfo(skb)->flags &= ~SKBFL_MANAGED_FRAG_REFS;
1907 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1908 skb_frag_ref(skb, i);
1909}
1910EXPORT_SYMBOL_GPL(__skb_zcopy_downgrade_managed);
1911
1912static int skb_zerocopy_clone(struct sk_buff *nskb, struct sk_buff *orig,
1913 gfp_t gfp_mask)
1914{
1915 if (skb_zcopy(orig)) {
1916 if (skb_zcopy(nskb)) {
1917 /* !gfp_mask callers are verified to !skb_zcopy(nskb) */
1918 if (!gfp_mask) {
1919 WARN_ON_ONCE(1);
1920 return -ENOMEM;
1921 }
1922 if (skb_uarg(nskb) == skb_uarg(orig))
1923 return 0;
1924 if (skb_copy_ubufs(nskb, GFP_ATOMIC))
1925 return -EIO;
1926 }
1927 skb_zcopy_set(nskb, skb_uarg(orig), NULL);
1928 }
1929 return 0;
1930}
1931
1932/**
1933 * skb_copy_ubufs - copy userspace skb frags buffers to kernel
1934 * @skb: the skb to modify
1935 * @gfp_mask: allocation priority
1936 *
1937 * This must be called on skb with SKBFL_ZEROCOPY_ENABLE.
1938 * It will copy all frags into kernel and drop the reference
1939 * to userspace pages.
1940 *
1941 * If this function is called from an interrupt gfp_mask() must be
1942 * %GFP_ATOMIC.
1943 *
1944 * Returns 0 on success or a negative error code on failure
1945 * to allocate kernel memory to copy to.
1946 */
1947int skb_copy_ubufs(struct sk_buff *skb, gfp_t gfp_mask)
1948{
1949 int num_frags = skb_shinfo(skb)->nr_frags;
1950 struct page *page, *head = NULL;
1951 int i, order, psize, new_frags;
1952 u32 d_off;
1953
1954 if (skb_shared(skb) || skb_unclone(skb, gfp_mask))
1955 return -EINVAL;
1956
1957 if (!num_frags)
1958 goto release;
1959
1960 /* We might have to allocate high order pages, so compute what minimum
1961 * page order is needed.
1962 */
1963 order = 0;
1964 while ((PAGE_SIZE << order) * MAX_SKB_FRAGS < __skb_pagelen(skb))
1965 order++;
1966 psize = (PAGE_SIZE << order);
1967
1968 new_frags = (__skb_pagelen(skb) + psize - 1) >> (PAGE_SHIFT + order);
1969 for (i = 0; i < new_frags; i++) {
1970 page = alloc_pages(gfp_mask | __GFP_COMP, order);
1971 if (!page) {
1972 while (head) {
1973 struct page *next = (struct page *)page_private(head);
1974 put_page(head);
1975 head = next;
1976 }
1977 return -ENOMEM;
1978 }
1979 set_page_private(page, (unsigned long)head);
1980 head = page;
1981 }
1982
1983 page = head;
1984 d_off = 0;
1985 for (i = 0; i < num_frags; i++) {
1986 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
1987 u32 p_off, p_len, copied;
1988 struct page *p;
1989 u8 *vaddr;
1990
1991 skb_frag_foreach_page(f, skb_frag_off(f), skb_frag_size(f),
1992 p, p_off, p_len, copied) {
1993 u32 copy, done = 0;
1994 vaddr = kmap_atomic(p);
1995
1996 while (done < p_len) {
1997 if (d_off == psize) {
1998 d_off = 0;
1999 page = (struct page *)page_private(page);
2000 }
2001 copy = min_t(u32, psize - d_off, p_len - done);
2002 memcpy(page_address(page) + d_off,
2003 vaddr + p_off + done, copy);
2004 done += copy;
2005 d_off += copy;
2006 }
2007 kunmap_atomic(vaddr);
2008 }
2009 }
2010
2011 /* skb frags release userspace buffers */
2012 for (i = 0; i < num_frags; i++)
2013 skb_frag_unref(skb, i);
2014
2015 /* skb frags point to kernel buffers */
2016 for (i = 0; i < new_frags - 1; i++) {
2017 __skb_fill_netmem_desc(skb, i, page_to_netmem(head), 0, psize);
2018 head = (struct page *)page_private(head);
2019 }
2020 __skb_fill_netmem_desc(skb, new_frags - 1, page_to_netmem(head), 0,
2021 d_off);
2022 skb_shinfo(skb)->nr_frags = new_frags;
2023
2024release:
2025 skb_zcopy_clear(skb, false);
2026 return 0;
2027}
2028EXPORT_SYMBOL_GPL(skb_copy_ubufs);
2029
2030/**
2031 * skb_clone - duplicate an sk_buff
2032 * @skb: buffer to clone
2033 * @gfp_mask: allocation priority
2034 *
2035 * Duplicate an &sk_buff. The new one is not owned by a socket. Both
2036 * copies share the same packet data but not structure. The new
2037 * buffer has a reference count of 1. If the allocation fails the
2038 * function returns %NULL otherwise the new buffer is returned.
2039 *
2040 * If this function is called from an interrupt gfp_mask() must be
2041 * %GFP_ATOMIC.
2042 */
2043
2044struct sk_buff *skb_clone(struct sk_buff *skb, gfp_t gfp_mask)
2045{
2046 struct sk_buff_fclones *fclones = container_of(skb,
2047 struct sk_buff_fclones,
2048 skb1);
2049 struct sk_buff *n;
2050
2051 if (skb_orphan_frags(skb, gfp_mask))
2052 return NULL;
2053
2054 if (skb->fclone == SKB_FCLONE_ORIG &&
2055 refcount_read(&fclones->fclone_ref) == 1) {
2056 n = &fclones->skb2;
2057 refcount_set(&fclones->fclone_ref, 2);
2058 n->fclone = SKB_FCLONE_CLONE;
2059 } else {
2060 if (skb_pfmemalloc(skb))
2061 gfp_mask |= __GFP_MEMALLOC;
2062
2063 n = kmem_cache_alloc(net_hotdata.skbuff_cache, gfp_mask);
2064 if (!n)
2065 return NULL;
2066
2067 n->fclone = SKB_FCLONE_UNAVAILABLE;
2068 }
2069
2070 return __skb_clone(n, skb);
2071}
2072EXPORT_SYMBOL(skb_clone);
2073
2074void skb_headers_offset_update(struct sk_buff *skb, int off)
2075{
2076 /* Only adjust this if it actually is csum_start rather than csum */
2077 if (skb->ip_summed == CHECKSUM_PARTIAL)
2078 skb->csum_start += off;
2079 /* {transport,network,mac}_header and tail are relative to skb->head */
2080 skb->transport_header += off;
2081 skb->network_header += off;
2082 if (skb_mac_header_was_set(skb))
2083 skb->mac_header += off;
2084 skb->inner_transport_header += off;
2085 skb->inner_network_header += off;
2086 skb->inner_mac_header += off;
2087}
2088EXPORT_SYMBOL(skb_headers_offset_update);
2089
2090void skb_copy_header(struct sk_buff *new, const struct sk_buff *old)
2091{
2092 __copy_skb_header(new, old);
2093
2094 skb_shinfo(new)->gso_size = skb_shinfo(old)->gso_size;
2095 skb_shinfo(new)->gso_segs = skb_shinfo(old)->gso_segs;
2096 skb_shinfo(new)->gso_type = skb_shinfo(old)->gso_type;
2097}
2098EXPORT_SYMBOL(skb_copy_header);
2099
2100static inline int skb_alloc_rx_flag(const struct sk_buff *skb)
2101{
2102 if (skb_pfmemalloc(skb))
2103 return SKB_ALLOC_RX;
2104 return 0;
2105}
2106
2107/**
2108 * skb_copy - create private copy of an sk_buff
2109 * @skb: buffer to copy
2110 * @gfp_mask: allocation priority
2111 *
2112 * Make a copy of both an &sk_buff and its data. This is used when the
2113 * caller wishes to modify the data and needs a private copy of the
2114 * data to alter. Returns %NULL on failure or the pointer to the buffer
2115 * on success. The returned buffer has a reference count of 1.
2116 *
2117 * As by-product this function converts non-linear &sk_buff to linear
2118 * one, so that &sk_buff becomes completely private and caller is allowed
2119 * to modify all the data of returned buffer. This means that this
2120 * function is not recommended for use in circumstances when only
2121 * header is going to be modified. Use pskb_copy() instead.
2122 */
2123
2124struct sk_buff *skb_copy(const struct sk_buff *skb, gfp_t gfp_mask)
2125{
2126 struct sk_buff *n;
2127 unsigned int size;
2128 int headerlen;
2129
2130 if (WARN_ON_ONCE(skb_shinfo(skb)->gso_type & SKB_GSO_FRAGLIST))
2131 return NULL;
2132
2133 headerlen = skb_headroom(skb);
2134 size = skb_end_offset(skb) + skb->data_len;
2135 n = __alloc_skb(size, gfp_mask,
2136 skb_alloc_rx_flag(skb), NUMA_NO_NODE);
2137 if (!n)
2138 return NULL;
2139
2140 /* Set the data pointer */
2141 skb_reserve(n, headerlen);
2142 /* Set the tail pointer and length */
2143 skb_put(n, skb->len);
2144
2145 BUG_ON(skb_copy_bits(skb, -headerlen, n->head, headerlen + skb->len));
2146
2147 skb_copy_header(n, skb);
2148 return n;
2149}
2150EXPORT_SYMBOL(skb_copy);
2151
2152/**
2153 * __pskb_copy_fclone - create copy of an sk_buff with private head.
2154 * @skb: buffer to copy
2155 * @headroom: headroom of new skb
2156 * @gfp_mask: allocation priority
2157 * @fclone: if true allocate the copy of the skb from the fclone
2158 * cache instead of the head cache; it is recommended to set this
2159 * to true for the cases where the copy will likely be cloned
2160 *
2161 * Make a copy of both an &sk_buff and part of its data, located
2162 * in header. Fragmented data remain shared. This is used when
2163 * the caller wishes to modify only header of &sk_buff and needs
2164 * private copy of the header to alter. Returns %NULL on failure
2165 * or the pointer to the buffer on success.
2166 * The returned buffer has a reference count of 1.
2167 */
2168
2169struct sk_buff *__pskb_copy_fclone(struct sk_buff *skb, int headroom,
2170 gfp_t gfp_mask, bool fclone)
2171{
2172 unsigned int size = skb_headlen(skb) + headroom;
2173 int flags = skb_alloc_rx_flag(skb) | (fclone ? SKB_ALLOC_FCLONE : 0);
2174 struct sk_buff *n = __alloc_skb(size, gfp_mask, flags, NUMA_NO_NODE);
2175
2176 if (!n)
2177 goto out;
2178
2179 /* Set the data pointer */
2180 skb_reserve(n, headroom);
2181 /* Set the tail pointer and length */
2182 skb_put(n, skb_headlen(skb));
2183 /* Copy the bytes */
2184 skb_copy_from_linear_data(skb, n->data, n->len);
2185
2186 n->truesize += skb->data_len;
2187 n->data_len = skb->data_len;
2188 n->len = skb->len;
2189
2190 if (skb_shinfo(skb)->nr_frags) {
2191 int i;
2192
2193 if (skb_orphan_frags(skb, gfp_mask) ||
2194 skb_zerocopy_clone(n, skb, gfp_mask)) {
2195 kfree_skb(n);
2196 n = NULL;
2197 goto out;
2198 }
2199 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2200 skb_shinfo(n)->frags[i] = skb_shinfo(skb)->frags[i];
2201 skb_frag_ref(skb, i);
2202 }
2203 skb_shinfo(n)->nr_frags = i;
2204 }
2205
2206 if (skb_has_frag_list(skb)) {
2207 skb_shinfo(n)->frag_list = skb_shinfo(skb)->frag_list;
2208 skb_clone_fraglist(n);
2209 }
2210
2211 skb_copy_header(n, skb);
2212out:
2213 return n;
2214}
2215EXPORT_SYMBOL(__pskb_copy_fclone);
2216
2217/**
2218 * pskb_expand_head - reallocate header of &sk_buff
2219 * @skb: buffer to reallocate
2220 * @nhead: room to add at head
2221 * @ntail: room to add at tail
2222 * @gfp_mask: allocation priority
2223 *
2224 * Expands (or creates identical copy, if @nhead and @ntail are zero)
2225 * header of @skb. &sk_buff itself is not changed. &sk_buff MUST have
2226 * reference count of 1. Returns zero in the case of success or error,
2227 * if expansion failed. In the last case, &sk_buff is not changed.
2228 *
2229 * All the pointers pointing into skb header may change and must be
2230 * reloaded after call to this function.
2231 */
2232
2233int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,
2234 gfp_t gfp_mask)
2235{
2236 unsigned int osize = skb_end_offset(skb);
2237 unsigned int size = osize + nhead + ntail;
2238 long off;
2239 u8 *data;
2240 int i;
2241
2242 BUG_ON(nhead < 0);
2243
2244 BUG_ON(skb_shared(skb));
2245
2246 skb_zcopy_downgrade_managed(skb);
2247
2248 if (skb_pfmemalloc(skb))
2249 gfp_mask |= __GFP_MEMALLOC;
2250
2251 data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
2252 if (!data)
2253 goto nodata;
2254 size = SKB_WITH_OVERHEAD(size);
2255
2256 /* Copy only real data... and, alas, header. This should be
2257 * optimized for the cases when header is void.
2258 */
2259 memcpy(data + nhead, skb->head, skb_tail_pointer(skb) - skb->head);
2260
2261 memcpy((struct skb_shared_info *)(data + size),
2262 skb_shinfo(skb),
2263 offsetof(struct skb_shared_info, frags[skb_shinfo(skb)->nr_frags]));
2264
2265 /*
2266 * if shinfo is shared we must drop the old head gracefully, but if it
2267 * is not we can just drop the old head and let the existing refcount
2268 * be since all we did is relocate the values
2269 */
2270 if (skb_cloned(skb)) {
2271 if (skb_orphan_frags(skb, gfp_mask))
2272 goto nofrags;
2273 if (skb_zcopy(skb))
2274 refcount_inc(&skb_uarg(skb)->refcnt);
2275 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
2276 skb_frag_ref(skb, i);
2277
2278 if (skb_has_frag_list(skb))
2279 skb_clone_fraglist(skb);
2280
2281 skb_release_data(skb, SKB_CONSUMED, false);
2282 } else {
2283 skb_free_head(skb, false);
2284 }
2285 off = (data + nhead) - skb->head;
2286
2287 skb->head = data;
2288 skb->head_frag = 0;
2289 skb->data += off;
2290
2291 skb_set_end_offset(skb, size);
2292#ifdef NET_SKBUFF_DATA_USES_OFFSET
2293 off = nhead;
2294#endif
2295 skb->tail += off;
2296 skb_headers_offset_update(skb, nhead);
2297 skb->cloned = 0;
2298 skb->hdr_len = 0;
2299 skb->nohdr = 0;
2300 atomic_set(&skb_shinfo(skb)->dataref, 1);
2301
2302 skb_metadata_clear(skb);
2303
2304 /* It is not generally safe to change skb->truesize.
2305 * For the moment, we really care of rx path, or
2306 * when skb is orphaned (not attached to a socket).
2307 */
2308 if (!skb->sk || skb->destructor == sock_edemux)
2309 skb->truesize += size - osize;
2310
2311 return 0;
2312
2313nofrags:
2314 skb_kfree_head(data, size);
2315nodata:
2316 return -ENOMEM;
2317}
2318EXPORT_SYMBOL(pskb_expand_head);
2319
2320/* Make private copy of skb with writable head and some headroom */
2321
2322struct sk_buff *skb_realloc_headroom(struct sk_buff *skb, unsigned int headroom)
2323{
2324 struct sk_buff *skb2;
2325 int delta = headroom - skb_headroom(skb);
2326
2327 if (delta <= 0)
2328 skb2 = pskb_copy(skb, GFP_ATOMIC);
2329 else {
2330 skb2 = skb_clone(skb, GFP_ATOMIC);
2331 if (skb2 && pskb_expand_head(skb2, SKB_DATA_ALIGN(delta), 0,
2332 GFP_ATOMIC)) {
2333 kfree_skb(skb2);
2334 skb2 = NULL;
2335 }
2336 }
2337 return skb2;
2338}
2339EXPORT_SYMBOL(skb_realloc_headroom);
2340
2341/* Note: We plan to rework this in linux-6.4 */
2342int __skb_unclone_keeptruesize(struct sk_buff *skb, gfp_t pri)
2343{
2344 unsigned int saved_end_offset, saved_truesize;
2345 struct skb_shared_info *shinfo;
2346 int res;
2347
2348 saved_end_offset = skb_end_offset(skb);
2349 saved_truesize = skb->truesize;
2350
2351 res = pskb_expand_head(skb, 0, 0, pri);
2352 if (res)
2353 return res;
2354
2355 skb->truesize = saved_truesize;
2356
2357 if (likely(skb_end_offset(skb) == saved_end_offset))
2358 return 0;
2359
2360 /* We can not change skb->end if the original or new value
2361 * is SKB_SMALL_HEAD_HEADROOM, as it might break skb_kfree_head().
2362 */
2363 if (saved_end_offset == SKB_SMALL_HEAD_HEADROOM ||
2364 skb_end_offset(skb) == SKB_SMALL_HEAD_HEADROOM) {
2365 /* We think this path should not be taken.
2366 * Add a temporary trace to warn us just in case.
2367 */
2368 pr_err_once("__skb_unclone_keeptruesize() skb_end_offset() %u -> %u\n",
2369 saved_end_offset, skb_end_offset(skb));
2370 WARN_ON_ONCE(1);
2371 return 0;
2372 }
2373
2374 shinfo = skb_shinfo(skb);
2375
2376 /* We are about to change back skb->end,
2377 * we need to move skb_shinfo() to its new location.
2378 */
2379 memmove(skb->head + saved_end_offset,
2380 shinfo,
2381 offsetof(struct skb_shared_info, frags[shinfo->nr_frags]));
2382
2383 skb_set_end_offset(skb, saved_end_offset);
2384
2385 return 0;
2386}
2387
2388/**
2389 * skb_expand_head - reallocate header of &sk_buff
2390 * @skb: buffer to reallocate
2391 * @headroom: needed headroom
2392 *
2393 * Unlike skb_realloc_headroom, this one does not allocate a new skb
2394 * if possible; copies skb->sk to new skb as needed
2395 * and frees original skb in case of failures.
2396 *
2397 * It expect increased headroom and generates warning otherwise.
2398 */
2399
2400struct sk_buff *skb_expand_head(struct sk_buff *skb, unsigned int headroom)
2401{
2402 int delta = headroom - skb_headroom(skb);
2403 int osize = skb_end_offset(skb);
2404 struct sock *sk = skb->sk;
2405
2406 if (WARN_ONCE(delta <= 0,
2407 "%s is expecting an increase in the headroom", __func__))
2408 return skb;
2409
2410 delta = SKB_DATA_ALIGN(delta);
2411 /* pskb_expand_head() might crash, if skb is shared. */
2412 if (skb_shared(skb) || !is_skb_wmem(skb)) {
2413 struct sk_buff *nskb = skb_clone(skb, GFP_ATOMIC);
2414
2415 if (unlikely(!nskb))
2416 goto fail;
2417
2418 if (sk)
2419 skb_set_owner_w(nskb, sk);
2420 consume_skb(skb);
2421 skb = nskb;
2422 }
2423 if (pskb_expand_head(skb, delta, 0, GFP_ATOMIC))
2424 goto fail;
2425
2426 if (sk && is_skb_wmem(skb)) {
2427 delta = skb_end_offset(skb) - osize;
2428 refcount_add(delta, &sk->sk_wmem_alloc);
2429 skb->truesize += delta;
2430 }
2431 return skb;
2432
2433fail:
2434 kfree_skb(skb);
2435 return NULL;
2436}
2437EXPORT_SYMBOL(skb_expand_head);
2438
2439/**
2440 * skb_copy_expand - copy and expand sk_buff
2441 * @skb: buffer to copy
2442 * @newheadroom: new free bytes at head
2443 * @newtailroom: new free bytes at tail
2444 * @gfp_mask: allocation priority
2445 *
2446 * Make a copy of both an &sk_buff and its data and while doing so
2447 * allocate additional space.
2448 *
2449 * This is used when the caller wishes to modify the data and needs a
2450 * private copy of the data to alter as well as more space for new fields.
2451 * Returns %NULL on failure or the pointer to the buffer
2452 * on success. The returned buffer has a reference count of 1.
2453 *
2454 * You must pass %GFP_ATOMIC as the allocation priority if this function
2455 * is called from an interrupt.
2456 */
2457struct sk_buff *skb_copy_expand(const struct sk_buff *skb,
2458 int newheadroom, int newtailroom,
2459 gfp_t gfp_mask)
2460{
2461 /*
2462 * Allocate the copy buffer
2463 */
2464 int head_copy_len, head_copy_off;
2465 struct sk_buff *n;
2466 int oldheadroom;
2467
2468 if (WARN_ON_ONCE(skb_shinfo(skb)->gso_type & SKB_GSO_FRAGLIST))
2469 return NULL;
2470
2471 oldheadroom = skb_headroom(skb);
2472 n = __alloc_skb(newheadroom + skb->len + newtailroom,
2473 gfp_mask, skb_alloc_rx_flag(skb),
2474 NUMA_NO_NODE);
2475 if (!n)
2476 return NULL;
2477
2478 skb_reserve(n, newheadroom);
2479
2480 /* Set the tail pointer and length */
2481 skb_put(n, skb->len);
2482
2483 head_copy_len = oldheadroom;
2484 head_copy_off = 0;
2485 if (newheadroom <= head_copy_len)
2486 head_copy_len = newheadroom;
2487 else
2488 head_copy_off = newheadroom - head_copy_len;
2489
2490 /* Copy the linear header and data. */
2491 BUG_ON(skb_copy_bits(skb, -head_copy_len, n->head + head_copy_off,
2492 skb->len + head_copy_len));
2493
2494 skb_copy_header(n, skb);
2495
2496 skb_headers_offset_update(n, newheadroom - oldheadroom);
2497
2498 return n;
2499}
2500EXPORT_SYMBOL(skb_copy_expand);
2501
2502/**
2503 * __skb_pad - zero pad the tail of an skb
2504 * @skb: buffer to pad
2505 * @pad: space to pad
2506 * @free_on_error: free buffer on error
2507 *
2508 * Ensure that a buffer is followed by a padding area that is zero
2509 * filled. Used by network drivers which may DMA or transfer data
2510 * beyond the buffer end onto the wire.
2511 *
2512 * May return error in out of memory cases. The skb is freed on error
2513 * if @free_on_error is true.
2514 */
2515
2516int __skb_pad(struct sk_buff *skb, int pad, bool free_on_error)
2517{
2518 int err;
2519 int ntail;
2520
2521 /* If the skbuff is non linear tailroom is always zero.. */
2522 if (!skb_cloned(skb) && skb_tailroom(skb) >= pad) {
2523 memset(skb->data+skb->len, 0, pad);
2524 return 0;
2525 }
2526
2527 ntail = skb->data_len + pad - (skb->end - skb->tail);
2528 if (likely(skb_cloned(skb) || ntail > 0)) {
2529 err = pskb_expand_head(skb, 0, ntail, GFP_ATOMIC);
2530 if (unlikely(err))
2531 goto free_skb;
2532 }
2533
2534 /* FIXME: The use of this function with non-linear skb's really needs
2535 * to be audited.
2536 */
2537 err = skb_linearize(skb);
2538 if (unlikely(err))
2539 goto free_skb;
2540
2541 memset(skb->data + skb->len, 0, pad);
2542 return 0;
2543
2544free_skb:
2545 if (free_on_error)
2546 kfree_skb(skb);
2547 return err;
2548}
2549EXPORT_SYMBOL(__skb_pad);
2550
2551/**
2552 * pskb_put - add data to the tail of a potentially fragmented buffer
2553 * @skb: start of the buffer to use
2554 * @tail: tail fragment of the buffer to use
2555 * @len: amount of data to add
2556 *
2557 * This function extends the used data area of the potentially
2558 * fragmented buffer. @tail must be the last fragment of @skb -- or
2559 * @skb itself. If this would exceed the total buffer size the kernel
2560 * will panic. A pointer to the first byte of the extra data is
2561 * returned.
2562 */
2563
2564void *pskb_put(struct sk_buff *skb, struct sk_buff *tail, int len)
2565{
2566 if (tail != skb) {
2567 skb->data_len += len;
2568 skb->len += len;
2569 }
2570 return skb_put(tail, len);
2571}
2572EXPORT_SYMBOL_GPL(pskb_put);
2573
2574/**
2575 * skb_put - add data to a buffer
2576 * @skb: buffer to use
2577 * @len: amount of data to add
2578 *
2579 * This function extends the used data area of the buffer. If this would
2580 * exceed the total buffer size the kernel will panic. A pointer to the
2581 * first byte of the extra data is returned.
2582 */
2583void *skb_put(struct sk_buff *skb, unsigned int len)
2584{
2585 void *tmp = skb_tail_pointer(skb);
2586 SKB_LINEAR_ASSERT(skb);
2587 skb->tail += len;
2588 skb->len += len;
2589 if (unlikely(skb->tail > skb->end))
2590 skb_over_panic(skb, len, __builtin_return_address(0));
2591 return tmp;
2592}
2593EXPORT_SYMBOL(skb_put);
2594
2595/**
2596 * skb_push - add data to the start of a buffer
2597 * @skb: buffer to use
2598 * @len: amount of data to add
2599 *
2600 * This function extends the used data area of the buffer at the buffer
2601 * start. If this would exceed the total buffer headroom the kernel will
2602 * panic. A pointer to the first byte of the extra data is returned.
2603 */
2604void *skb_push(struct sk_buff *skb, unsigned int len)
2605{
2606 skb->data -= len;
2607 skb->len += len;
2608 if (unlikely(skb->data < skb->head))
2609 skb_under_panic(skb, len, __builtin_return_address(0));
2610 return skb->data;
2611}
2612EXPORT_SYMBOL(skb_push);
2613
2614/**
2615 * skb_pull - remove data from the start of a buffer
2616 * @skb: buffer to use
2617 * @len: amount of data to remove
2618 *
2619 * This function removes data from the start of a buffer, returning
2620 * the memory to the headroom. A pointer to the next data in the buffer
2621 * is returned. Once the data has been pulled future pushes will overwrite
2622 * the old data.
2623 */
2624void *skb_pull(struct sk_buff *skb, unsigned int len)
2625{
2626 return skb_pull_inline(skb, len);
2627}
2628EXPORT_SYMBOL(skb_pull);
2629
2630/**
2631 * skb_pull_data - remove data from the start of a buffer returning its
2632 * original position.
2633 * @skb: buffer to use
2634 * @len: amount of data to remove
2635 *
2636 * This function removes data from the start of a buffer, returning
2637 * the memory to the headroom. A pointer to the original data in the buffer
2638 * is returned after checking if there is enough data to pull. Once the
2639 * data has been pulled future pushes will overwrite the old data.
2640 */
2641void *skb_pull_data(struct sk_buff *skb, size_t len)
2642{
2643 void *data = skb->data;
2644
2645 if (skb->len < len)
2646 return NULL;
2647
2648 skb_pull(skb, len);
2649
2650 return data;
2651}
2652EXPORT_SYMBOL(skb_pull_data);
2653
2654/**
2655 * skb_trim - remove end from a buffer
2656 * @skb: buffer to alter
2657 * @len: new length
2658 *
2659 * Cut the length of a buffer down by removing data from the tail. If
2660 * the buffer is already under the length specified it is not modified.
2661 * The skb must be linear.
2662 */
2663void skb_trim(struct sk_buff *skb, unsigned int len)
2664{
2665 if (skb->len > len)
2666 __skb_trim(skb, len);
2667}
2668EXPORT_SYMBOL(skb_trim);
2669
2670/* Trims skb to length len. It can change skb pointers.
2671 */
2672
2673int ___pskb_trim(struct sk_buff *skb, unsigned int len)
2674{
2675 struct sk_buff **fragp;
2676 struct sk_buff *frag;
2677 int offset = skb_headlen(skb);
2678 int nfrags = skb_shinfo(skb)->nr_frags;
2679 int i;
2680 int err;
2681
2682 if (skb_cloned(skb) &&
2683 unlikely((err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC))))
2684 return err;
2685
2686 i = 0;
2687 if (offset >= len)
2688 goto drop_pages;
2689
2690 for (; i < nfrags; i++) {
2691 int end = offset + skb_frag_size(&skb_shinfo(skb)->frags[i]);
2692
2693 if (end < len) {
2694 offset = end;
2695 continue;
2696 }
2697
2698 skb_frag_size_set(&skb_shinfo(skb)->frags[i++], len - offset);
2699
2700drop_pages:
2701 skb_shinfo(skb)->nr_frags = i;
2702
2703 for (; i < nfrags; i++)
2704 skb_frag_unref(skb, i);
2705
2706 if (skb_has_frag_list(skb))
2707 skb_drop_fraglist(skb);
2708 goto done;
2709 }
2710
2711 for (fragp = &skb_shinfo(skb)->frag_list; (frag = *fragp);
2712 fragp = &frag->next) {
2713 int end = offset + frag->len;
2714
2715 if (skb_shared(frag)) {
2716 struct sk_buff *nfrag;
2717
2718 nfrag = skb_clone(frag, GFP_ATOMIC);
2719 if (unlikely(!nfrag))
2720 return -ENOMEM;
2721
2722 nfrag->next = frag->next;
2723 consume_skb(frag);
2724 frag = nfrag;
2725 *fragp = frag;
2726 }
2727
2728 if (end < len) {
2729 offset = end;
2730 continue;
2731 }
2732
2733 if (end > len &&
2734 unlikely((err = pskb_trim(frag, len - offset))))
2735 return err;
2736
2737 if (frag->next)
2738 skb_drop_list(&frag->next);
2739 break;
2740 }
2741
2742done:
2743 if (len > skb_headlen(skb)) {
2744 skb->data_len -= skb->len - len;
2745 skb->len = len;
2746 } else {
2747 skb->len = len;
2748 skb->data_len = 0;
2749 skb_set_tail_pointer(skb, len);
2750 }
2751
2752 if (!skb->sk || skb->destructor == sock_edemux)
2753 skb_condense(skb);
2754 return 0;
2755}
2756EXPORT_SYMBOL(___pskb_trim);
2757
2758/* Note : use pskb_trim_rcsum() instead of calling this directly
2759 */
2760int pskb_trim_rcsum_slow(struct sk_buff *skb, unsigned int len)
2761{
2762 if (skb->ip_summed == CHECKSUM_COMPLETE) {
2763 int delta = skb->len - len;
2764
2765 skb->csum = csum_block_sub(skb->csum,
2766 skb_checksum(skb, len, delta, 0),
2767 len);
2768 } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
2769 int hdlen = (len > skb_headlen(skb)) ? skb_headlen(skb) : len;
2770 int offset = skb_checksum_start_offset(skb) + skb->csum_offset;
2771
2772 if (offset + sizeof(__sum16) > hdlen)
2773 return -EINVAL;
2774 }
2775 return __pskb_trim(skb, len);
2776}
2777EXPORT_SYMBOL(pskb_trim_rcsum_slow);
2778
2779/**
2780 * __pskb_pull_tail - advance tail of skb header
2781 * @skb: buffer to reallocate
2782 * @delta: number of bytes to advance tail
2783 *
2784 * The function makes a sense only on a fragmented &sk_buff,
2785 * it expands header moving its tail forward and copying necessary
2786 * data from fragmented part.
2787 *
2788 * &sk_buff MUST have reference count of 1.
2789 *
2790 * Returns %NULL (and &sk_buff does not change) if pull failed
2791 * or value of new tail of skb in the case of success.
2792 *
2793 * All the pointers pointing into skb header may change and must be
2794 * reloaded after call to this function.
2795 */
2796
2797/* Moves tail of skb head forward, copying data from fragmented part,
2798 * when it is necessary.
2799 * 1. It may fail due to malloc failure.
2800 * 2. It may change skb pointers.
2801 *
2802 * It is pretty complicated. Luckily, it is called only in exceptional cases.
2803 */
2804void *__pskb_pull_tail(struct sk_buff *skb, int delta)
2805{
2806 /* If skb has not enough free space at tail, get new one
2807 * plus 128 bytes for future expansions. If we have enough
2808 * room at tail, reallocate without expansion only if skb is cloned.
2809 */
2810 int i, k, eat = (skb->tail + delta) - skb->end;
2811
2812 if (eat > 0 || skb_cloned(skb)) {
2813 if (pskb_expand_head(skb, 0, eat > 0 ? eat + 128 : 0,
2814 GFP_ATOMIC))
2815 return NULL;
2816 }
2817
2818 BUG_ON(skb_copy_bits(skb, skb_headlen(skb),
2819 skb_tail_pointer(skb), delta));
2820
2821 /* Optimization: no fragments, no reasons to preestimate
2822 * size of pulled pages. Superb.
2823 */
2824 if (!skb_has_frag_list(skb))
2825 goto pull_pages;
2826
2827 /* Estimate size of pulled pages. */
2828 eat = delta;
2829 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2830 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
2831
2832 if (size >= eat)
2833 goto pull_pages;
2834 eat -= size;
2835 }
2836
2837 /* If we need update frag list, we are in troubles.
2838 * Certainly, it is possible to add an offset to skb data,
2839 * but taking into account that pulling is expected to
2840 * be very rare operation, it is worth to fight against
2841 * further bloating skb head and crucify ourselves here instead.
2842 * Pure masohism, indeed. 8)8)
2843 */
2844 if (eat) {
2845 struct sk_buff *list = skb_shinfo(skb)->frag_list;
2846 struct sk_buff *clone = NULL;
2847 struct sk_buff *insp = NULL;
2848
2849 do {
2850 if (list->len <= eat) {
2851 /* Eaten as whole. */
2852 eat -= list->len;
2853 list = list->next;
2854 insp = list;
2855 } else {
2856 /* Eaten partially. */
2857 if (skb_is_gso(skb) && !list->head_frag &&
2858 skb_headlen(list))
2859 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2860
2861 if (skb_shared(list)) {
2862 /* Sucks! We need to fork list. :-( */
2863 clone = skb_clone(list, GFP_ATOMIC);
2864 if (!clone)
2865 return NULL;
2866 insp = list->next;
2867 list = clone;
2868 } else {
2869 /* This may be pulled without
2870 * problems. */
2871 insp = list;
2872 }
2873 if (!pskb_pull(list, eat)) {
2874 kfree_skb(clone);
2875 return NULL;
2876 }
2877 break;
2878 }
2879 } while (eat);
2880
2881 /* Free pulled out fragments. */
2882 while ((list = skb_shinfo(skb)->frag_list) != insp) {
2883 skb_shinfo(skb)->frag_list = list->next;
2884 consume_skb(list);
2885 }
2886 /* And insert new clone at head. */
2887 if (clone) {
2888 clone->next = list;
2889 skb_shinfo(skb)->frag_list = clone;
2890 }
2891 }
2892 /* Success! Now we may commit changes to skb data. */
2893
2894pull_pages:
2895 eat = delta;
2896 k = 0;
2897 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2898 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
2899
2900 if (size <= eat) {
2901 skb_frag_unref(skb, i);
2902 eat -= size;
2903 } else {
2904 skb_frag_t *frag = &skb_shinfo(skb)->frags[k];
2905
2906 *frag = skb_shinfo(skb)->frags[i];
2907 if (eat) {
2908 skb_frag_off_add(frag, eat);
2909 skb_frag_size_sub(frag, eat);
2910 if (!i)
2911 goto end;
2912 eat = 0;
2913 }
2914 k++;
2915 }
2916 }
2917 skb_shinfo(skb)->nr_frags = k;
2918
2919end:
2920 skb->tail += delta;
2921 skb->data_len -= delta;
2922
2923 if (!skb->data_len)
2924 skb_zcopy_clear(skb, false);
2925
2926 return skb_tail_pointer(skb);
2927}
2928EXPORT_SYMBOL(__pskb_pull_tail);
2929
2930/**
2931 * skb_copy_bits - copy bits from skb to kernel buffer
2932 * @skb: source skb
2933 * @offset: offset in source
2934 * @to: destination buffer
2935 * @len: number of bytes to copy
2936 *
2937 * Copy the specified number of bytes from the source skb to the
2938 * destination buffer.
2939 *
2940 * CAUTION ! :
2941 * If its prototype is ever changed,
2942 * check arch/{*}/net/{*}.S files,
2943 * since it is called from BPF assembly code.
2944 */
2945int skb_copy_bits(const struct sk_buff *skb, int offset, void *to, int len)
2946{
2947 int start = skb_headlen(skb);
2948 struct sk_buff *frag_iter;
2949 int i, copy;
2950
2951 if (offset > (int)skb->len - len)
2952 goto fault;
2953
2954 /* Copy header. */
2955 if ((copy = start - offset) > 0) {
2956 if (copy > len)
2957 copy = len;
2958 skb_copy_from_linear_data_offset(skb, offset, to, copy);
2959 if ((len -= copy) == 0)
2960 return 0;
2961 offset += copy;
2962 to += copy;
2963 }
2964
2965 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2966 int end;
2967 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
2968
2969 WARN_ON(start > offset + len);
2970
2971 end = start + skb_frag_size(f);
2972 if ((copy = end - offset) > 0) {
2973 u32 p_off, p_len, copied;
2974 struct page *p;
2975 u8 *vaddr;
2976
2977 if (copy > len)
2978 copy = len;
2979
2980 skb_frag_foreach_page(f,
2981 skb_frag_off(f) + offset - start,
2982 copy, p, p_off, p_len, copied) {
2983 vaddr = kmap_atomic(p);
2984 memcpy(to + copied, vaddr + p_off, p_len);
2985 kunmap_atomic(vaddr);
2986 }
2987
2988 if ((len -= copy) == 0)
2989 return 0;
2990 offset += copy;
2991 to += copy;
2992 }
2993 start = end;
2994 }
2995
2996 skb_walk_frags(skb, frag_iter) {
2997 int end;
2998
2999 WARN_ON(start > offset + len);
3000
3001 end = start + frag_iter->len;
3002 if ((copy = end - offset) > 0) {
3003 if (copy > len)
3004 copy = len;
3005 if (skb_copy_bits(frag_iter, offset - start, to, copy))
3006 goto fault;
3007 if ((len -= copy) == 0)
3008 return 0;
3009 offset += copy;
3010 to += copy;
3011 }
3012 start = end;
3013 }
3014
3015 if (!len)
3016 return 0;
3017
3018fault:
3019 return -EFAULT;
3020}
3021EXPORT_SYMBOL(skb_copy_bits);
3022
3023/*
3024 * Callback from splice_to_pipe(), if we need to release some pages
3025 * at the end of the spd in case we error'ed out in filling the pipe.
3026 */
3027static void sock_spd_release(struct splice_pipe_desc *spd, unsigned int i)
3028{
3029 put_page(spd->pages[i]);
3030}
3031
3032static struct page *linear_to_page(struct page *page, unsigned int *len,
3033 unsigned int *offset,
3034 struct sock *sk)
3035{
3036 struct page_frag *pfrag = sk_page_frag(sk);
3037
3038 if (!sk_page_frag_refill(sk, pfrag))
3039 return NULL;
3040
3041 *len = min_t(unsigned int, *len, pfrag->size - pfrag->offset);
3042
3043 memcpy(page_address(pfrag->page) + pfrag->offset,
3044 page_address(page) + *offset, *len);
3045 *offset = pfrag->offset;
3046 pfrag->offset += *len;
3047
3048 return pfrag->page;
3049}
3050
3051static bool spd_can_coalesce(const struct splice_pipe_desc *spd,
3052 struct page *page,
3053 unsigned int offset)
3054{
3055 return spd->nr_pages &&
3056 spd->pages[spd->nr_pages - 1] == page &&
3057 (spd->partial[spd->nr_pages - 1].offset +
3058 spd->partial[spd->nr_pages - 1].len == offset);
3059}
3060
3061/*
3062 * Fill page/offset/length into spd, if it can hold more pages.
3063 */
3064static bool spd_fill_page(struct splice_pipe_desc *spd,
3065 struct pipe_inode_info *pipe, struct page *page,
3066 unsigned int *len, unsigned int offset,
3067 bool linear,
3068 struct sock *sk)
3069{
3070 if (unlikely(spd->nr_pages == MAX_SKB_FRAGS))
3071 return true;
3072
3073 if (linear) {
3074 page = linear_to_page(page, len, &offset, sk);
3075 if (!page)
3076 return true;
3077 }
3078 if (spd_can_coalesce(spd, page, offset)) {
3079 spd->partial[spd->nr_pages - 1].len += *len;
3080 return false;
3081 }
3082 get_page(page);
3083 spd->pages[spd->nr_pages] = page;
3084 spd->partial[spd->nr_pages].len = *len;
3085 spd->partial[spd->nr_pages].offset = offset;
3086 spd->nr_pages++;
3087
3088 return false;
3089}
3090
3091static bool __splice_segment(struct page *page, unsigned int poff,
3092 unsigned int plen, unsigned int *off,
3093 unsigned int *len,
3094 struct splice_pipe_desc *spd, bool linear,
3095 struct sock *sk,
3096 struct pipe_inode_info *pipe)
3097{
3098 if (!*len)
3099 return true;
3100
3101 /* skip this segment if already processed */
3102 if (*off >= plen) {
3103 *off -= plen;
3104 return false;
3105 }
3106
3107 /* ignore any bits we already processed */
3108 poff += *off;
3109 plen -= *off;
3110 *off = 0;
3111
3112 do {
3113 unsigned int flen = min(*len, plen);
3114
3115 if (spd_fill_page(spd, pipe, page, &flen, poff,
3116 linear, sk))
3117 return true;
3118 poff += flen;
3119 plen -= flen;
3120 *len -= flen;
3121 } while (*len && plen);
3122
3123 return false;
3124}
3125
3126/*
3127 * Map linear and fragment data from the skb to spd. It reports true if the
3128 * pipe is full or if we already spliced the requested length.
3129 */
3130static bool __skb_splice_bits(struct sk_buff *skb, struct pipe_inode_info *pipe,
3131 unsigned int *offset, unsigned int *len,
3132 struct splice_pipe_desc *spd, struct sock *sk)
3133{
3134 int seg;
3135 struct sk_buff *iter;
3136
3137 /* map the linear part :
3138 * If skb->head_frag is set, this 'linear' part is backed by a
3139 * fragment, and if the head is not shared with any clones then
3140 * we can avoid a copy since we own the head portion of this page.
3141 */
3142 if (__splice_segment(virt_to_page(skb->data),
3143 (unsigned long) skb->data & (PAGE_SIZE - 1),
3144 skb_headlen(skb),
3145 offset, len, spd,
3146 skb_head_is_locked(skb),
3147 sk, pipe))
3148 return true;
3149
3150 /*
3151 * then map the fragments
3152 */
3153 for (seg = 0; seg < skb_shinfo(skb)->nr_frags; seg++) {
3154 const skb_frag_t *f = &skb_shinfo(skb)->frags[seg];
3155
3156 if (__splice_segment(skb_frag_page(f),
3157 skb_frag_off(f), skb_frag_size(f),
3158 offset, len, spd, false, sk, pipe))
3159 return true;
3160 }
3161
3162 skb_walk_frags(skb, iter) {
3163 if (*offset >= iter->len) {
3164 *offset -= iter->len;
3165 continue;
3166 }
3167 /* __skb_splice_bits() only fails if the output has no room
3168 * left, so no point in going over the frag_list for the error
3169 * case.
3170 */
3171 if (__skb_splice_bits(iter, pipe, offset, len, spd, sk))
3172 return true;
3173 }
3174
3175 return false;
3176}
3177
3178/*
3179 * Map data from the skb to a pipe. Should handle both the linear part,
3180 * the fragments, and the frag list.
3181 */
3182int skb_splice_bits(struct sk_buff *skb, struct sock *sk, unsigned int offset,
3183 struct pipe_inode_info *pipe, unsigned int tlen,
3184 unsigned int flags)
3185{
3186 struct partial_page partial[MAX_SKB_FRAGS];
3187 struct page *pages[MAX_SKB_FRAGS];
3188 struct splice_pipe_desc spd = {
3189 .pages = pages,
3190 .partial = partial,
3191 .nr_pages_max = MAX_SKB_FRAGS,
3192 .ops = &nosteal_pipe_buf_ops,
3193 .spd_release = sock_spd_release,
3194 };
3195 int ret = 0;
3196
3197 __skb_splice_bits(skb, pipe, &offset, &tlen, &spd, sk);
3198
3199 if (spd.nr_pages)
3200 ret = splice_to_pipe(pipe, &spd);
3201
3202 return ret;
3203}
3204EXPORT_SYMBOL_GPL(skb_splice_bits);
3205
3206static int sendmsg_locked(struct sock *sk, struct msghdr *msg)
3207{
3208 struct socket *sock = sk->sk_socket;
3209 size_t size = msg_data_left(msg);
3210
3211 if (!sock)
3212 return -EINVAL;
3213
3214 if (!sock->ops->sendmsg_locked)
3215 return sock_no_sendmsg_locked(sk, msg, size);
3216
3217 return sock->ops->sendmsg_locked(sk, msg, size);
3218}
3219
3220static int sendmsg_unlocked(struct sock *sk, struct msghdr *msg)
3221{
3222 struct socket *sock = sk->sk_socket;
3223
3224 if (!sock)
3225 return -EINVAL;
3226 return sock_sendmsg(sock, msg);
3227}
3228
3229typedef int (*sendmsg_func)(struct sock *sk, struct msghdr *msg);
3230static int __skb_send_sock(struct sock *sk, struct sk_buff *skb, int offset,
3231 int len, sendmsg_func sendmsg)
3232{
3233 unsigned int orig_len = len;
3234 struct sk_buff *head = skb;
3235 unsigned short fragidx;
3236 int slen, ret;
3237
3238do_frag_list:
3239
3240 /* Deal with head data */
3241 while (offset < skb_headlen(skb) && len) {
3242 struct kvec kv;
3243 struct msghdr msg;
3244
3245 slen = min_t(int, len, skb_headlen(skb) - offset);
3246 kv.iov_base = skb->data + offset;
3247 kv.iov_len = slen;
3248 memset(&msg, 0, sizeof(msg));
3249 msg.msg_flags = MSG_DONTWAIT;
3250
3251 iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, &kv, 1, slen);
3252 ret = INDIRECT_CALL_2(sendmsg, sendmsg_locked,
3253 sendmsg_unlocked, sk, &msg);
3254 if (ret <= 0)
3255 goto error;
3256
3257 offset += ret;
3258 len -= ret;
3259 }
3260
3261 /* All the data was skb head? */
3262 if (!len)
3263 goto out;
3264
3265 /* Make offset relative to start of frags */
3266 offset -= skb_headlen(skb);
3267
3268 /* Find where we are in frag list */
3269 for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags; fragidx++) {
3270 skb_frag_t *frag = &skb_shinfo(skb)->frags[fragidx];
3271
3272 if (offset < skb_frag_size(frag))
3273 break;
3274
3275 offset -= skb_frag_size(frag);
3276 }
3277
3278 for (; len && fragidx < skb_shinfo(skb)->nr_frags; fragidx++) {
3279 skb_frag_t *frag = &skb_shinfo(skb)->frags[fragidx];
3280
3281 slen = min_t(size_t, len, skb_frag_size(frag) - offset);
3282
3283 while (slen) {
3284 struct bio_vec bvec;
3285 struct msghdr msg = {
3286 .msg_flags = MSG_SPLICE_PAGES | MSG_DONTWAIT,
3287 };
3288
3289 bvec_set_page(&bvec, skb_frag_page(frag), slen,
3290 skb_frag_off(frag) + offset);
3291 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1,
3292 slen);
3293
3294 ret = INDIRECT_CALL_2(sendmsg, sendmsg_locked,
3295 sendmsg_unlocked, sk, &msg);
3296 if (ret <= 0)
3297 goto error;
3298
3299 len -= ret;
3300 offset += ret;
3301 slen -= ret;
3302 }
3303
3304 offset = 0;
3305 }
3306
3307 if (len) {
3308 /* Process any frag lists */
3309
3310 if (skb == head) {
3311 if (skb_has_frag_list(skb)) {
3312 skb = skb_shinfo(skb)->frag_list;
3313 goto do_frag_list;
3314 }
3315 } else if (skb->next) {
3316 skb = skb->next;
3317 goto do_frag_list;
3318 }
3319 }
3320
3321out:
3322 return orig_len - len;
3323
3324error:
3325 return orig_len == len ? ret : orig_len - len;
3326}
3327
3328/* Send skb data on a socket. Socket must be locked. */
3329int skb_send_sock_locked(struct sock *sk, struct sk_buff *skb, int offset,
3330 int len)
3331{
3332 return __skb_send_sock(sk, skb, offset, len, sendmsg_locked);
3333}
3334EXPORT_SYMBOL_GPL(skb_send_sock_locked);
3335
3336/* Send skb data on a socket. Socket must be unlocked. */
3337int skb_send_sock(struct sock *sk, struct sk_buff *skb, int offset, int len)
3338{
3339 return __skb_send_sock(sk, skb, offset, len, sendmsg_unlocked);
3340}
3341
3342/**
3343 * skb_store_bits - store bits from kernel buffer to skb
3344 * @skb: destination buffer
3345 * @offset: offset in destination
3346 * @from: source buffer
3347 * @len: number of bytes to copy
3348 *
3349 * Copy the specified number of bytes from the source buffer to the
3350 * destination skb. This function handles all the messy bits of
3351 * traversing fragment lists and such.
3352 */
3353
3354int skb_store_bits(struct sk_buff *skb, int offset, const void *from, int len)
3355{
3356 int start = skb_headlen(skb);
3357 struct sk_buff *frag_iter;
3358 int i, copy;
3359
3360 if (offset > (int)skb->len - len)
3361 goto fault;
3362
3363 if ((copy = start - offset) > 0) {
3364 if (copy > len)
3365 copy = len;
3366 skb_copy_to_linear_data_offset(skb, offset, from, copy);
3367 if ((len -= copy) == 0)
3368 return 0;
3369 offset += copy;
3370 from += copy;
3371 }
3372
3373 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3374 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3375 int end;
3376
3377 WARN_ON(start > offset + len);
3378
3379 end = start + skb_frag_size(frag);
3380 if ((copy = end - offset) > 0) {
3381 u32 p_off, p_len, copied;
3382 struct page *p;
3383 u8 *vaddr;
3384
3385 if (copy > len)
3386 copy = len;
3387
3388 skb_frag_foreach_page(frag,
3389 skb_frag_off(frag) + offset - start,
3390 copy, p, p_off, p_len, copied) {
3391 vaddr = kmap_atomic(p);
3392 memcpy(vaddr + p_off, from + copied, p_len);
3393 kunmap_atomic(vaddr);
3394 }
3395
3396 if ((len -= copy) == 0)
3397 return 0;
3398 offset += copy;
3399 from += copy;
3400 }
3401 start = end;
3402 }
3403
3404 skb_walk_frags(skb, frag_iter) {
3405 int end;
3406
3407 WARN_ON(start > offset + len);
3408
3409 end = start + frag_iter->len;
3410 if ((copy = end - offset) > 0) {
3411 if (copy > len)
3412 copy = len;
3413 if (skb_store_bits(frag_iter, offset - start,
3414 from, copy))
3415 goto fault;
3416 if ((len -= copy) == 0)
3417 return 0;
3418 offset += copy;
3419 from += copy;
3420 }
3421 start = end;
3422 }
3423 if (!len)
3424 return 0;
3425
3426fault:
3427 return -EFAULT;
3428}
3429EXPORT_SYMBOL(skb_store_bits);
3430
3431/* Checksum skb data. */
3432__wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
3433 __wsum csum, const struct skb_checksum_ops *ops)
3434{
3435 int start = skb_headlen(skb);
3436 int i, copy = start - offset;
3437 struct sk_buff *frag_iter;
3438 int pos = 0;
3439
3440 /* Checksum header. */
3441 if (copy > 0) {
3442 if (copy > len)
3443 copy = len;
3444 csum = INDIRECT_CALL_1(ops->update, csum_partial_ext,
3445 skb->data + offset, copy, csum);
3446 if ((len -= copy) == 0)
3447 return csum;
3448 offset += copy;
3449 pos = copy;
3450 }
3451
3452 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3453 int end;
3454 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3455
3456 WARN_ON(start > offset + len);
3457
3458 end = start + skb_frag_size(frag);
3459 if ((copy = end - offset) > 0) {
3460 u32 p_off, p_len, copied;
3461 struct page *p;
3462 __wsum csum2;
3463 u8 *vaddr;
3464
3465 if (copy > len)
3466 copy = len;
3467
3468 skb_frag_foreach_page(frag,
3469 skb_frag_off(frag) + offset - start,
3470 copy, p, p_off, p_len, copied) {
3471 vaddr = kmap_atomic(p);
3472 csum2 = INDIRECT_CALL_1(ops->update,
3473 csum_partial_ext,
3474 vaddr + p_off, p_len, 0);
3475 kunmap_atomic(vaddr);
3476 csum = INDIRECT_CALL_1(ops->combine,
3477 csum_block_add_ext, csum,
3478 csum2, pos, p_len);
3479 pos += p_len;
3480 }
3481
3482 if (!(len -= copy))
3483 return csum;
3484 offset += copy;
3485 }
3486 start = end;
3487 }
3488
3489 skb_walk_frags(skb, frag_iter) {
3490 int end;
3491
3492 WARN_ON(start > offset + len);
3493
3494 end = start + frag_iter->len;
3495 if ((copy = end - offset) > 0) {
3496 __wsum csum2;
3497 if (copy > len)
3498 copy = len;
3499 csum2 = __skb_checksum(frag_iter, offset - start,
3500 copy, 0, ops);
3501 csum = INDIRECT_CALL_1(ops->combine, csum_block_add_ext,
3502 csum, csum2, pos, copy);
3503 if ((len -= copy) == 0)
3504 return csum;
3505 offset += copy;
3506 pos += copy;
3507 }
3508 start = end;
3509 }
3510 BUG_ON(len);
3511
3512 return csum;
3513}
3514EXPORT_SYMBOL(__skb_checksum);
3515
3516__wsum skb_checksum(const struct sk_buff *skb, int offset,
3517 int len, __wsum csum)
3518{
3519 const struct skb_checksum_ops ops = {
3520 .update = csum_partial_ext,
3521 .combine = csum_block_add_ext,
3522 };
3523
3524 return __skb_checksum(skb, offset, len, csum, &ops);
3525}
3526EXPORT_SYMBOL(skb_checksum);
3527
3528/* Both of above in one bottle. */
3529
3530__wsum skb_copy_and_csum_bits(const struct sk_buff *skb, int offset,
3531 u8 *to, int len)
3532{
3533 int start = skb_headlen(skb);
3534 int i, copy = start - offset;
3535 struct sk_buff *frag_iter;
3536 int pos = 0;
3537 __wsum csum = 0;
3538
3539 /* Copy header. */
3540 if (copy > 0) {
3541 if (copy > len)
3542 copy = len;
3543 csum = csum_partial_copy_nocheck(skb->data + offset, to,
3544 copy);
3545 if ((len -= copy) == 0)
3546 return csum;
3547 offset += copy;
3548 to += copy;
3549 pos = copy;
3550 }
3551
3552 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3553 int end;
3554
3555 WARN_ON(start > offset + len);
3556
3557 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
3558 if ((copy = end - offset) > 0) {
3559 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3560 u32 p_off, p_len, copied;
3561 struct page *p;
3562 __wsum csum2;
3563 u8 *vaddr;
3564
3565 if (copy > len)
3566 copy = len;
3567
3568 skb_frag_foreach_page(frag,
3569 skb_frag_off(frag) + offset - start,
3570 copy, p, p_off, p_len, copied) {
3571 vaddr = kmap_atomic(p);
3572 csum2 = csum_partial_copy_nocheck(vaddr + p_off,
3573 to + copied,
3574 p_len);
3575 kunmap_atomic(vaddr);
3576 csum = csum_block_add(csum, csum2, pos);
3577 pos += p_len;
3578 }
3579
3580 if (!(len -= copy))
3581 return csum;
3582 offset += copy;
3583 to += copy;
3584 }
3585 start = end;
3586 }
3587
3588 skb_walk_frags(skb, frag_iter) {
3589 __wsum csum2;
3590 int end;
3591
3592 WARN_ON(start > offset + len);
3593
3594 end = start + frag_iter->len;
3595 if ((copy = end - offset) > 0) {
3596 if (copy > len)
3597 copy = len;
3598 csum2 = skb_copy_and_csum_bits(frag_iter,
3599 offset - start,
3600 to, copy);
3601 csum = csum_block_add(csum, csum2, pos);
3602 if ((len -= copy) == 0)
3603 return csum;
3604 offset += copy;
3605 to += copy;
3606 pos += copy;
3607 }
3608 start = end;
3609 }
3610 BUG_ON(len);
3611 return csum;
3612}
3613EXPORT_SYMBOL(skb_copy_and_csum_bits);
3614
3615__sum16 __skb_checksum_complete_head(struct sk_buff *skb, int len)
3616{
3617 __sum16 sum;
3618
3619 sum = csum_fold(skb_checksum(skb, 0, len, skb->csum));
3620 /* See comments in __skb_checksum_complete(). */
3621 if (likely(!sum)) {
3622 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
3623 !skb->csum_complete_sw)
3624 netdev_rx_csum_fault(skb->dev, skb);
3625 }
3626 if (!skb_shared(skb))
3627 skb->csum_valid = !sum;
3628 return sum;
3629}
3630EXPORT_SYMBOL(__skb_checksum_complete_head);
3631
3632/* This function assumes skb->csum already holds pseudo header's checksum,
3633 * which has been changed from the hardware checksum, for example, by
3634 * __skb_checksum_validate_complete(). And, the original skb->csum must
3635 * have been validated unsuccessfully for CHECKSUM_COMPLETE case.
3636 *
3637 * It returns non-zero if the recomputed checksum is still invalid, otherwise
3638 * zero. The new checksum is stored back into skb->csum unless the skb is
3639 * shared.
3640 */
3641__sum16 __skb_checksum_complete(struct sk_buff *skb)
3642{
3643 __wsum csum;
3644 __sum16 sum;
3645
3646 csum = skb_checksum(skb, 0, skb->len, 0);
3647
3648 sum = csum_fold(csum_add(skb->csum, csum));
3649 /* This check is inverted, because we already knew the hardware
3650 * checksum is invalid before calling this function. So, if the
3651 * re-computed checksum is valid instead, then we have a mismatch
3652 * between the original skb->csum and skb_checksum(). This means either
3653 * the original hardware checksum is incorrect or we screw up skb->csum
3654 * when moving skb->data around.
3655 */
3656 if (likely(!sum)) {
3657 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
3658 !skb->csum_complete_sw)
3659 netdev_rx_csum_fault(skb->dev, skb);
3660 }
3661
3662 if (!skb_shared(skb)) {
3663 /* Save full packet checksum */
3664 skb->csum = csum;
3665 skb->ip_summed = CHECKSUM_COMPLETE;
3666 skb->csum_complete_sw = 1;
3667 skb->csum_valid = !sum;
3668 }
3669
3670 return sum;
3671}
3672EXPORT_SYMBOL(__skb_checksum_complete);
3673
3674static __wsum warn_crc32c_csum_update(const void *buff, int len, __wsum sum)
3675{
3676 net_warn_ratelimited(
3677 "%s: attempt to compute crc32c without libcrc32c.ko\n",
3678 __func__);
3679 return 0;
3680}
3681
3682static __wsum warn_crc32c_csum_combine(__wsum csum, __wsum csum2,
3683 int offset, int len)
3684{
3685 net_warn_ratelimited(
3686 "%s: attempt to compute crc32c without libcrc32c.ko\n",
3687 __func__);
3688 return 0;
3689}
3690
3691static const struct skb_checksum_ops default_crc32c_ops = {
3692 .update = warn_crc32c_csum_update,
3693 .combine = warn_crc32c_csum_combine,
3694};
3695
3696const struct skb_checksum_ops *crc32c_csum_stub __read_mostly =
3697 &default_crc32c_ops;
3698EXPORT_SYMBOL(crc32c_csum_stub);
3699
3700 /**
3701 * skb_zerocopy_headlen - Calculate headroom needed for skb_zerocopy()
3702 * @from: source buffer
3703 *
3704 * Calculates the amount of linear headroom needed in the 'to' skb passed
3705 * into skb_zerocopy().
3706 */
3707unsigned int
3708skb_zerocopy_headlen(const struct sk_buff *from)
3709{
3710 unsigned int hlen = 0;
3711
3712 if (!from->head_frag ||
3713 skb_headlen(from) < L1_CACHE_BYTES ||
3714 skb_shinfo(from)->nr_frags >= MAX_SKB_FRAGS) {
3715 hlen = skb_headlen(from);
3716 if (!hlen)
3717 hlen = from->len;
3718 }
3719
3720 if (skb_has_frag_list(from))
3721 hlen = from->len;
3722
3723 return hlen;
3724}
3725EXPORT_SYMBOL_GPL(skb_zerocopy_headlen);
3726
3727/**
3728 * skb_zerocopy - Zero copy skb to skb
3729 * @to: destination buffer
3730 * @from: source buffer
3731 * @len: number of bytes to copy from source buffer
3732 * @hlen: size of linear headroom in destination buffer
3733 *
3734 * Copies up to `len` bytes from `from` to `to` by creating references
3735 * to the frags in the source buffer.
3736 *
3737 * The `hlen` as calculated by skb_zerocopy_headlen() specifies the
3738 * headroom in the `to` buffer.
3739 *
3740 * Return value:
3741 * 0: everything is OK
3742 * -ENOMEM: couldn't orphan frags of @from due to lack of memory
3743 * -EFAULT: skb_copy_bits() found some problem with skb geometry
3744 */
3745int
3746skb_zerocopy(struct sk_buff *to, struct sk_buff *from, int len, int hlen)
3747{
3748 int i, j = 0;
3749 int plen = 0; /* length of skb->head fragment */
3750 int ret;
3751 struct page *page;
3752 unsigned int offset;
3753
3754 BUG_ON(!from->head_frag && !hlen);
3755
3756 /* dont bother with small payloads */
3757 if (len <= skb_tailroom(to))
3758 return skb_copy_bits(from, 0, skb_put(to, len), len);
3759
3760 if (hlen) {
3761 ret = skb_copy_bits(from, 0, skb_put(to, hlen), hlen);
3762 if (unlikely(ret))
3763 return ret;
3764 len -= hlen;
3765 } else {
3766 plen = min_t(int, skb_headlen(from), len);
3767 if (plen) {
3768 page = virt_to_head_page(from->head);
3769 offset = from->data - (unsigned char *)page_address(page);
3770 __skb_fill_netmem_desc(to, 0, page_to_netmem(page),
3771 offset, plen);
3772 get_page(page);
3773 j = 1;
3774 len -= plen;
3775 }
3776 }
3777
3778 skb_len_add(to, len + plen);
3779
3780 if (unlikely(skb_orphan_frags(from, GFP_ATOMIC))) {
3781 skb_tx_error(from);
3782 return -ENOMEM;
3783 }
3784 skb_zerocopy_clone(to, from, GFP_ATOMIC);
3785
3786 for (i = 0; i < skb_shinfo(from)->nr_frags; i++) {
3787 int size;
3788
3789 if (!len)
3790 break;
3791 skb_shinfo(to)->frags[j] = skb_shinfo(from)->frags[i];
3792 size = min_t(int, skb_frag_size(&skb_shinfo(to)->frags[j]),
3793 len);
3794 skb_frag_size_set(&skb_shinfo(to)->frags[j], size);
3795 len -= size;
3796 skb_frag_ref(to, j);
3797 j++;
3798 }
3799 skb_shinfo(to)->nr_frags = j;
3800
3801 return 0;
3802}
3803EXPORT_SYMBOL_GPL(skb_zerocopy);
3804
3805void skb_copy_and_csum_dev(const struct sk_buff *skb, u8 *to)
3806{
3807 __wsum csum;
3808 long csstart;
3809
3810 if (skb->ip_summed == CHECKSUM_PARTIAL)
3811 csstart = skb_checksum_start_offset(skb);
3812 else
3813 csstart = skb_headlen(skb);
3814
3815 BUG_ON(csstart > skb_headlen(skb));
3816
3817 skb_copy_from_linear_data(skb, to, csstart);
3818
3819 csum = 0;
3820 if (csstart != skb->len)
3821 csum = skb_copy_and_csum_bits(skb, csstart, to + csstart,
3822 skb->len - csstart);
3823
3824 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3825 long csstuff = csstart + skb->csum_offset;
3826
3827 *((__sum16 *)(to + csstuff)) = csum_fold(csum);
3828 }
3829}
3830EXPORT_SYMBOL(skb_copy_and_csum_dev);
3831
3832/**
3833 * skb_dequeue - remove from the head of the queue
3834 * @list: list to dequeue from
3835 *
3836 * Remove the head of the list. The list lock is taken so the function
3837 * may be used safely with other locking list functions. The head item is
3838 * returned or %NULL if the list is empty.
3839 */
3840
3841struct sk_buff *skb_dequeue(struct sk_buff_head *list)
3842{
3843 unsigned long flags;
3844 struct sk_buff *result;
3845
3846 spin_lock_irqsave(&list->lock, flags);
3847 result = __skb_dequeue(list);
3848 spin_unlock_irqrestore(&list->lock, flags);
3849 return result;
3850}
3851EXPORT_SYMBOL(skb_dequeue);
3852
3853/**
3854 * skb_dequeue_tail - remove from the tail of the queue
3855 * @list: list to dequeue from
3856 *
3857 * Remove the tail of the list. The list lock is taken so the function
3858 * may be used safely with other locking list functions. The tail item is
3859 * returned or %NULL if the list is empty.
3860 */
3861struct sk_buff *skb_dequeue_tail(struct sk_buff_head *list)
3862{
3863 unsigned long flags;
3864 struct sk_buff *result;
3865
3866 spin_lock_irqsave(&list->lock, flags);
3867 result = __skb_dequeue_tail(list);
3868 spin_unlock_irqrestore(&list->lock, flags);
3869 return result;
3870}
3871EXPORT_SYMBOL(skb_dequeue_tail);
3872
3873/**
3874 * skb_queue_purge_reason - empty a list
3875 * @list: list to empty
3876 * @reason: drop reason
3877 *
3878 * Delete all buffers on an &sk_buff list. Each buffer is removed from
3879 * the list and one reference dropped. This function takes the list
3880 * lock and is atomic with respect to other list locking functions.
3881 */
3882void skb_queue_purge_reason(struct sk_buff_head *list,
3883 enum skb_drop_reason reason)
3884{
3885 struct sk_buff_head tmp;
3886 unsigned long flags;
3887
3888 if (skb_queue_empty_lockless(list))
3889 return;
3890
3891 __skb_queue_head_init(&tmp);
3892
3893 spin_lock_irqsave(&list->lock, flags);
3894 skb_queue_splice_init(list, &tmp);
3895 spin_unlock_irqrestore(&list->lock, flags);
3896
3897 __skb_queue_purge_reason(&tmp, reason);
3898}
3899EXPORT_SYMBOL(skb_queue_purge_reason);
3900
3901/**
3902 * skb_rbtree_purge - empty a skb rbtree
3903 * @root: root of the rbtree to empty
3904 * Return value: the sum of truesizes of all purged skbs.
3905 *
3906 * Delete all buffers on an &sk_buff rbtree. Each buffer is removed from
3907 * the list and one reference dropped. This function does not take
3908 * any lock. Synchronization should be handled by the caller (e.g., TCP
3909 * out-of-order queue is protected by the socket lock).
3910 */
3911unsigned int skb_rbtree_purge(struct rb_root *root)
3912{
3913 struct rb_node *p = rb_first(root);
3914 unsigned int sum = 0;
3915
3916 while (p) {
3917 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
3918
3919 p = rb_next(p);
3920 rb_erase(&skb->rbnode, root);
3921 sum += skb->truesize;
3922 kfree_skb(skb);
3923 }
3924 return sum;
3925}
3926
3927void skb_errqueue_purge(struct sk_buff_head *list)
3928{
3929 struct sk_buff *skb, *next;
3930 struct sk_buff_head kill;
3931 unsigned long flags;
3932
3933 __skb_queue_head_init(&kill);
3934
3935 spin_lock_irqsave(&list->lock, flags);
3936 skb_queue_walk_safe(list, skb, next) {
3937 if (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ZEROCOPY ||
3938 SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_TIMESTAMPING)
3939 continue;
3940 __skb_unlink(skb, list);
3941 __skb_queue_tail(&kill, skb);
3942 }
3943 spin_unlock_irqrestore(&list->lock, flags);
3944 __skb_queue_purge(&kill);
3945}
3946EXPORT_SYMBOL(skb_errqueue_purge);
3947
3948/**
3949 * skb_queue_head - queue a buffer at the list head
3950 * @list: list to use
3951 * @newsk: buffer to queue
3952 *
3953 * Queue a buffer at the start of the list. This function takes the
3954 * list lock and can be used safely with other locking &sk_buff functions
3955 * safely.
3956 *
3957 * A buffer cannot be placed on two lists at the same time.
3958 */
3959void skb_queue_head(struct sk_buff_head *list, struct sk_buff *newsk)
3960{
3961 unsigned long flags;
3962
3963 spin_lock_irqsave(&list->lock, flags);
3964 __skb_queue_head(list, newsk);
3965 spin_unlock_irqrestore(&list->lock, flags);
3966}
3967EXPORT_SYMBOL(skb_queue_head);
3968
3969/**
3970 * skb_queue_tail - queue a buffer at the list tail
3971 * @list: list to use
3972 * @newsk: buffer to queue
3973 *
3974 * Queue a buffer at the tail of the list. This function takes the
3975 * list lock and can be used safely with other locking &sk_buff functions
3976 * safely.
3977 *
3978 * A buffer cannot be placed on two lists at the same time.
3979 */
3980void skb_queue_tail(struct sk_buff_head *list, struct sk_buff *newsk)
3981{
3982 unsigned long flags;
3983
3984 spin_lock_irqsave(&list->lock, flags);
3985 __skb_queue_tail(list, newsk);
3986 spin_unlock_irqrestore(&list->lock, flags);
3987}
3988EXPORT_SYMBOL(skb_queue_tail);
3989
3990/**
3991 * skb_unlink - remove a buffer from a list
3992 * @skb: buffer to remove
3993 * @list: list to use
3994 *
3995 * Remove a packet from a list. The list locks are taken and this
3996 * function is atomic with respect to other list locked calls
3997 *
3998 * You must know what list the SKB is on.
3999 */
4000void skb_unlink(struct sk_buff *skb, struct sk_buff_head *list)
4001{
4002 unsigned long flags;
4003
4004 spin_lock_irqsave(&list->lock, flags);
4005 __skb_unlink(skb, list);
4006 spin_unlock_irqrestore(&list->lock, flags);
4007}
4008EXPORT_SYMBOL(skb_unlink);
4009
4010/**
4011 * skb_append - append a buffer
4012 * @old: buffer to insert after
4013 * @newsk: buffer to insert
4014 * @list: list to use
4015 *
4016 * Place a packet after a given packet in a list. The list locks are taken
4017 * and this function is atomic with respect to other list locked calls.
4018 * A buffer cannot be placed on two lists at the same time.
4019 */
4020void skb_append(struct sk_buff *old, struct sk_buff *newsk, struct sk_buff_head *list)
4021{
4022 unsigned long flags;
4023
4024 spin_lock_irqsave(&list->lock, flags);
4025 __skb_queue_after(list, old, newsk);
4026 spin_unlock_irqrestore(&list->lock, flags);
4027}
4028EXPORT_SYMBOL(skb_append);
4029
4030static inline void skb_split_inside_header(struct sk_buff *skb,
4031 struct sk_buff* skb1,
4032 const u32 len, const int pos)
4033{
4034 int i;
4035
4036 skb_copy_from_linear_data_offset(skb, len, skb_put(skb1, pos - len),
4037 pos - len);
4038 /* And move data appendix as is. */
4039 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
4040 skb_shinfo(skb1)->frags[i] = skb_shinfo(skb)->frags[i];
4041
4042 skb_shinfo(skb1)->nr_frags = skb_shinfo(skb)->nr_frags;
4043 skb_shinfo(skb)->nr_frags = 0;
4044 skb1->data_len = skb->data_len;
4045 skb1->len += skb1->data_len;
4046 skb->data_len = 0;
4047 skb->len = len;
4048 skb_set_tail_pointer(skb, len);
4049}
4050
4051static inline void skb_split_no_header(struct sk_buff *skb,
4052 struct sk_buff* skb1,
4053 const u32 len, int pos)
4054{
4055 int i, k = 0;
4056 const int nfrags = skb_shinfo(skb)->nr_frags;
4057
4058 skb_shinfo(skb)->nr_frags = 0;
4059 skb1->len = skb1->data_len = skb->len - len;
4060 skb->len = len;
4061 skb->data_len = len - pos;
4062
4063 for (i = 0; i < nfrags; i++) {
4064 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
4065
4066 if (pos + size > len) {
4067 skb_shinfo(skb1)->frags[k] = skb_shinfo(skb)->frags[i];
4068
4069 if (pos < len) {
4070 /* Split frag.
4071 * We have two variants in this case:
4072 * 1. Move all the frag to the second
4073 * part, if it is possible. F.e.
4074 * this approach is mandatory for TUX,
4075 * where splitting is expensive.
4076 * 2. Split is accurately. We make this.
4077 */
4078 skb_frag_ref(skb, i);
4079 skb_frag_off_add(&skb_shinfo(skb1)->frags[0], len - pos);
4080 skb_frag_size_sub(&skb_shinfo(skb1)->frags[0], len - pos);
4081 skb_frag_size_set(&skb_shinfo(skb)->frags[i], len - pos);
4082 skb_shinfo(skb)->nr_frags++;
4083 }
4084 k++;
4085 } else
4086 skb_shinfo(skb)->nr_frags++;
4087 pos += size;
4088 }
4089 skb_shinfo(skb1)->nr_frags = k;
4090}
4091
4092/**
4093 * skb_split - Split fragmented skb to two parts at length len.
4094 * @skb: the buffer to split
4095 * @skb1: the buffer to receive the second part
4096 * @len: new length for skb
4097 */
4098void skb_split(struct sk_buff *skb, struct sk_buff *skb1, const u32 len)
4099{
4100 int pos = skb_headlen(skb);
4101 const int zc_flags = SKBFL_SHARED_FRAG | SKBFL_PURE_ZEROCOPY;
4102
4103 skb_zcopy_downgrade_managed(skb);
4104
4105 skb_shinfo(skb1)->flags |= skb_shinfo(skb)->flags & zc_flags;
4106 skb_zerocopy_clone(skb1, skb, 0);
4107 if (len < pos) /* Split line is inside header. */
4108 skb_split_inside_header(skb, skb1, len, pos);
4109 else /* Second chunk has no header, nothing to copy. */
4110 skb_split_no_header(skb, skb1, len, pos);
4111}
4112EXPORT_SYMBOL(skb_split);
4113
4114/* Shifting from/to a cloned skb is a no-go.
4115 *
4116 * Caller cannot keep skb_shinfo related pointers past calling here!
4117 */
4118static int skb_prepare_for_shift(struct sk_buff *skb)
4119{
4120 return skb_unclone_keeptruesize(skb, GFP_ATOMIC);
4121}
4122
4123/**
4124 * skb_shift - Shifts paged data partially from skb to another
4125 * @tgt: buffer into which tail data gets added
4126 * @skb: buffer from which the paged data comes from
4127 * @shiftlen: shift up to this many bytes
4128 *
4129 * Attempts to shift up to shiftlen worth of bytes, which may be less than
4130 * the length of the skb, from skb to tgt. Returns number bytes shifted.
4131 * It's up to caller to free skb if everything was shifted.
4132 *
4133 * If @tgt runs out of frags, the whole operation is aborted.
4134 *
4135 * Skb cannot include anything else but paged data while tgt is allowed
4136 * to have non-paged data as well.
4137 *
4138 * TODO: full sized shift could be optimized but that would need
4139 * specialized skb free'er to handle frags without up-to-date nr_frags.
4140 */
4141int skb_shift(struct sk_buff *tgt, struct sk_buff *skb, int shiftlen)
4142{
4143 int from, to, merge, todo;
4144 skb_frag_t *fragfrom, *fragto;
4145
4146 BUG_ON(shiftlen > skb->len);
4147
4148 if (skb_headlen(skb))
4149 return 0;
4150 if (skb_zcopy(tgt) || skb_zcopy(skb))
4151 return 0;
4152
4153 todo = shiftlen;
4154 from = 0;
4155 to = skb_shinfo(tgt)->nr_frags;
4156 fragfrom = &skb_shinfo(skb)->frags[from];
4157
4158 /* Actual merge is delayed until the point when we know we can
4159 * commit all, so that we don't have to undo partial changes
4160 */
4161 if (!to ||
4162 !skb_can_coalesce(tgt, to, skb_frag_page(fragfrom),
4163 skb_frag_off(fragfrom))) {
4164 merge = -1;
4165 } else {
4166 merge = to - 1;
4167
4168 todo -= skb_frag_size(fragfrom);
4169 if (todo < 0) {
4170 if (skb_prepare_for_shift(skb) ||
4171 skb_prepare_for_shift(tgt))
4172 return 0;
4173
4174 /* All previous frag pointers might be stale! */
4175 fragfrom = &skb_shinfo(skb)->frags[from];
4176 fragto = &skb_shinfo(tgt)->frags[merge];
4177
4178 skb_frag_size_add(fragto, shiftlen);
4179 skb_frag_size_sub(fragfrom, shiftlen);
4180 skb_frag_off_add(fragfrom, shiftlen);
4181
4182 goto onlymerged;
4183 }
4184
4185 from++;
4186 }
4187
4188 /* Skip full, not-fitting skb to avoid expensive operations */
4189 if ((shiftlen == skb->len) &&
4190 (skb_shinfo(skb)->nr_frags - from) > (MAX_SKB_FRAGS - to))
4191 return 0;
4192
4193 if (skb_prepare_for_shift(skb) || skb_prepare_for_shift(tgt))
4194 return 0;
4195
4196 while ((todo > 0) && (from < skb_shinfo(skb)->nr_frags)) {
4197 if (to == MAX_SKB_FRAGS)
4198 return 0;
4199
4200 fragfrom = &skb_shinfo(skb)->frags[from];
4201 fragto = &skb_shinfo(tgt)->frags[to];
4202
4203 if (todo >= skb_frag_size(fragfrom)) {
4204 *fragto = *fragfrom;
4205 todo -= skb_frag_size(fragfrom);
4206 from++;
4207 to++;
4208
4209 } else {
4210 __skb_frag_ref(fragfrom);
4211 skb_frag_page_copy(fragto, fragfrom);
4212 skb_frag_off_copy(fragto, fragfrom);
4213 skb_frag_size_set(fragto, todo);
4214
4215 skb_frag_off_add(fragfrom, todo);
4216 skb_frag_size_sub(fragfrom, todo);
4217 todo = 0;
4218
4219 to++;
4220 break;
4221 }
4222 }
4223
4224 /* Ready to "commit" this state change to tgt */
4225 skb_shinfo(tgt)->nr_frags = to;
4226
4227 if (merge >= 0) {
4228 fragfrom = &skb_shinfo(skb)->frags[0];
4229 fragto = &skb_shinfo(tgt)->frags[merge];
4230
4231 skb_frag_size_add(fragto, skb_frag_size(fragfrom));
4232 __skb_frag_unref(fragfrom, skb->pp_recycle);
4233 }
4234
4235 /* Reposition in the original skb */
4236 to = 0;
4237 while (from < skb_shinfo(skb)->nr_frags)
4238 skb_shinfo(skb)->frags[to++] = skb_shinfo(skb)->frags[from++];
4239 skb_shinfo(skb)->nr_frags = to;
4240
4241 BUG_ON(todo > 0 && !skb_shinfo(skb)->nr_frags);
4242
4243onlymerged:
4244 /* Most likely the tgt won't ever need its checksum anymore, skb on
4245 * the other hand might need it if it needs to be resent
4246 */
4247 tgt->ip_summed = CHECKSUM_PARTIAL;
4248 skb->ip_summed = CHECKSUM_PARTIAL;
4249
4250 skb_len_add(skb, -shiftlen);
4251 skb_len_add(tgt, shiftlen);
4252
4253 return shiftlen;
4254}
4255
4256/**
4257 * skb_prepare_seq_read - Prepare a sequential read of skb data
4258 * @skb: the buffer to read
4259 * @from: lower offset of data to be read
4260 * @to: upper offset of data to be read
4261 * @st: state variable
4262 *
4263 * Initializes the specified state variable. Must be called before
4264 * invoking skb_seq_read() for the first time.
4265 */
4266void skb_prepare_seq_read(struct sk_buff *skb, unsigned int from,
4267 unsigned int to, struct skb_seq_state *st)
4268{
4269 st->lower_offset = from;
4270 st->upper_offset = to;
4271 st->root_skb = st->cur_skb = skb;
4272 st->frag_idx = st->stepped_offset = 0;
4273 st->frag_data = NULL;
4274 st->frag_off = 0;
4275}
4276EXPORT_SYMBOL(skb_prepare_seq_read);
4277
4278/**
4279 * skb_seq_read - Sequentially read skb data
4280 * @consumed: number of bytes consumed by the caller so far
4281 * @data: destination pointer for data to be returned
4282 * @st: state variable
4283 *
4284 * Reads a block of skb data at @consumed relative to the
4285 * lower offset specified to skb_prepare_seq_read(). Assigns
4286 * the head of the data block to @data and returns the length
4287 * of the block or 0 if the end of the skb data or the upper
4288 * offset has been reached.
4289 *
4290 * The caller is not required to consume all of the data
4291 * returned, i.e. @consumed is typically set to the number
4292 * of bytes already consumed and the next call to
4293 * skb_seq_read() will return the remaining part of the block.
4294 *
4295 * Note 1: The size of each block of data returned can be arbitrary,
4296 * this limitation is the cost for zerocopy sequential
4297 * reads of potentially non linear data.
4298 *
4299 * Note 2: Fragment lists within fragments are not implemented
4300 * at the moment, state->root_skb could be replaced with
4301 * a stack for this purpose.
4302 */
4303unsigned int skb_seq_read(unsigned int consumed, const u8 **data,
4304 struct skb_seq_state *st)
4305{
4306 unsigned int block_limit, abs_offset = consumed + st->lower_offset;
4307 skb_frag_t *frag;
4308
4309 if (unlikely(abs_offset >= st->upper_offset)) {
4310 if (st->frag_data) {
4311 kunmap_atomic(st->frag_data);
4312 st->frag_data = NULL;
4313 }
4314 return 0;
4315 }
4316
4317next_skb:
4318 block_limit = skb_headlen(st->cur_skb) + st->stepped_offset;
4319
4320 if (abs_offset < block_limit && !st->frag_data) {
4321 *data = st->cur_skb->data + (abs_offset - st->stepped_offset);
4322 return block_limit - abs_offset;
4323 }
4324
4325 if (st->frag_idx == 0 && !st->frag_data)
4326 st->stepped_offset += skb_headlen(st->cur_skb);
4327
4328 while (st->frag_idx < skb_shinfo(st->cur_skb)->nr_frags) {
4329 unsigned int pg_idx, pg_off, pg_sz;
4330
4331 frag = &skb_shinfo(st->cur_skb)->frags[st->frag_idx];
4332
4333 pg_idx = 0;
4334 pg_off = skb_frag_off(frag);
4335 pg_sz = skb_frag_size(frag);
4336
4337 if (skb_frag_must_loop(skb_frag_page(frag))) {
4338 pg_idx = (pg_off + st->frag_off) >> PAGE_SHIFT;
4339 pg_off = offset_in_page(pg_off + st->frag_off);
4340 pg_sz = min_t(unsigned int, pg_sz - st->frag_off,
4341 PAGE_SIZE - pg_off);
4342 }
4343
4344 block_limit = pg_sz + st->stepped_offset;
4345 if (abs_offset < block_limit) {
4346 if (!st->frag_data)
4347 st->frag_data = kmap_atomic(skb_frag_page(frag) + pg_idx);
4348
4349 *data = (u8 *)st->frag_data + pg_off +
4350 (abs_offset - st->stepped_offset);
4351
4352 return block_limit - abs_offset;
4353 }
4354
4355 if (st->frag_data) {
4356 kunmap_atomic(st->frag_data);
4357 st->frag_data = NULL;
4358 }
4359
4360 st->stepped_offset += pg_sz;
4361 st->frag_off += pg_sz;
4362 if (st->frag_off == skb_frag_size(frag)) {
4363 st->frag_off = 0;
4364 st->frag_idx++;
4365 }
4366 }
4367
4368 if (st->frag_data) {
4369 kunmap_atomic(st->frag_data);
4370 st->frag_data = NULL;
4371 }
4372
4373 if (st->root_skb == st->cur_skb && skb_has_frag_list(st->root_skb)) {
4374 st->cur_skb = skb_shinfo(st->root_skb)->frag_list;
4375 st->frag_idx = 0;
4376 goto next_skb;
4377 } else if (st->cur_skb->next) {
4378 st->cur_skb = st->cur_skb->next;
4379 st->frag_idx = 0;
4380 goto next_skb;
4381 }
4382
4383 return 0;
4384}
4385EXPORT_SYMBOL(skb_seq_read);
4386
4387/**
4388 * skb_abort_seq_read - Abort a sequential read of skb data
4389 * @st: state variable
4390 *
4391 * Must be called if skb_seq_read() was not called until it
4392 * returned 0.
4393 */
4394void skb_abort_seq_read(struct skb_seq_state *st)
4395{
4396 if (st->frag_data)
4397 kunmap_atomic(st->frag_data);
4398}
4399EXPORT_SYMBOL(skb_abort_seq_read);
4400
4401#define TS_SKB_CB(state) ((struct skb_seq_state *) &((state)->cb))
4402
4403static unsigned int skb_ts_get_next_block(unsigned int offset, const u8 **text,
4404 struct ts_config *conf,
4405 struct ts_state *state)
4406{
4407 return skb_seq_read(offset, text, TS_SKB_CB(state));
4408}
4409
4410static void skb_ts_finish(struct ts_config *conf, struct ts_state *state)
4411{
4412 skb_abort_seq_read(TS_SKB_CB(state));
4413}
4414
4415/**
4416 * skb_find_text - Find a text pattern in skb data
4417 * @skb: the buffer to look in
4418 * @from: search offset
4419 * @to: search limit
4420 * @config: textsearch configuration
4421 *
4422 * Finds a pattern in the skb data according to the specified
4423 * textsearch configuration. Use textsearch_next() to retrieve
4424 * subsequent occurrences of the pattern. Returns the offset
4425 * to the first occurrence or UINT_MAX if no match was found.
4426 */
4427unsigned int skb_find_text(struct sk_buff *skb, unsigned int from,
4428 unsigned int to, struct ts_config *config)
4429{
4430 unsigned int patlen = config->ops->get_pattern_len(config);
4431 struct ts_state state;
4432 unsigned int ret;
4433
4434 BUILD_BUG_ON(sizeof(struct skb_seq_state) > sizeof(state.cb));
4435
4436 config->get_next_block = skb_ts_get_next_block;
4437 config->finish = skb_ts_finish;
4438
4439 skb_prepare_seq_read(skb, from, to, TS_SKB_CB(&state));
4440
4441 ret = textsearch_find(config, &state);
4442 return (ret + patlen <= to - from ? ret : UINT_MAX);
4443}
4444EXPORT_SYMBOL(skb_find_text);
4445
4446int skb_append_pagefrags(struct sk_buff *skb, struct page *page,
4447 int offset, size_t size, size_t max_frags)
4448{
4449 int i = skb_shinfo(skb)->nr_frags;
4450
4451 if (skb_can_coalesce(skb, i, page, offset)) {
4452 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
4453 } else if (i < max_frags) {
4454 skb_zcopy_downgrade_managed(skb);
4455 get_page(page);
4456 skb_fill_page_desc_noacc(skb, i, page, offset, size);
4457 } else {
4458 return -EMSGSIZE;
4459 }
4460
4461 return 0;
4462}
4463EXPORT_SYMBOL_GPL(skb_append_pagefrags);
4464
4465/**
4466 * skb_pull_rcsum - pull skb and update receive checksum
4467 * @skb: buffer to update
4468 * @len: length of data pulled
4469 *
4470 * This function performs an skb_pull on the packet and updates
4471 * the CHECKSUM_COMPLETE checksum. It should be used on
4472 * receive path processing instead of skb_pull unless you know
4473 * that the checksum difference is zero (e.g., a valid IP header)
4474 * or you are setting ip_summed to CHECKSUM_NONE.
4475 */
4476void *skb_pull_rcsum(struct sk_buff *skb, unsigned int len)
4477{
4478 unsigned char *data = skb->data;
4479
4480 BUG_ON(len > skb->len);
4481 __skb_pull(skb, len);
4482 skb_postpull_rcsum(skb, data, len);
4483 return skb->data;
4484}
4485EXPORT_SYMBOL_GPL(skb_pull_rcsum);
4486
4487static inline skb_frag_t skb_head_frag_to_page_desc(struct sk_buff *frag_skb)
4488{
4489 skb_frag_t head_frag;
4490 struct page *page;
4491
4492 page = virt_to_head_page(frag_skb->head);
4493 skb_frag_fill_page_desc(&head_frag, page, frag_skb->data -
4494 (unsigned char *)page_address(page),
4495 skb_headlen(frag_skb));
4496 return head_frag;
4497}
4498
4499struct sk_buff *skb_segment_list(struct sk_buff *skb,
4500 netdev_features_t features,
4501 unsigned int offset)
4502{
4503 struct sk_buff *list_skb = skb_shinfo(skb)->frag_list;
4504 unsigned int tnl_hlen = skb_tnl_header_len(skb);
4505 unsigned int delta_truesize = 0;
4506 unsigned int delta_len = 0;
4507 struct sk_buff *tail = NULL;
4508 struct sk_buff *nskb, *tmp;
4509 int len_diff, err;
4510
4511 skb_push(skb, -skb_network_offset(skb) + offset);
4512
4513 /* Ensure the head is writeable before touching the shared info */
4514 err = skb_unclone(skb, GFP_ATOMIC);
4515 if (err)
4516 goto err_linearize;
4517
4518 skb_shinfo(skb)->frag_list = NULL;
4519
4520 while (list_skb) {
4521 nskb = list_skb;
4522 list_skb = list_skb->next;
4523
4524 err = 0;
4525 delta_truesize += nskb->truesize;
4526 if (skb_shared(nskb)) {
4527 tmp = skb_clone(nskb, GFP_ATOMIC);
4528 if (tmp) {
4529 consume_skb(nskb);
4530 nskb = tmp;
4531 err = skb_unclone(nskb, GFP_ATOMIC);
4532 } else {
4533 err = -ENOMEM;
4534 }
4535 }
4536
4537 if (!tail)
4538 skb->next = nskb;
4539 else
4540 tail->next = nskb;
4541
4542 if (unlikely(err)) {
4543 nskb->next = list_skb;
4544 goto err_linearize;
4545 }
4546
4547 tail = nskb;
4548
4549 delta_len += nskb->len;
4550
4551 skb_push(nskb, -skb_network_offset(nskb) + offset);
4552
4553 skb_release_head_state(nskb);
4554 len_diff = skb_network_header_len(nskb) - skb_network_header_len(skb);
4555 __copy_skb_header(nskb, skb);
4556
4557 skb_headers_offset_update(nskb, skb_headroom(nskb) - skb_headroom(skb));
4558 nskb->transport_header += len_diff;
4559 skb_copy_from_linear_data_offset(skb, -tnl_hlen,
4560 nskb->data - tnl_hlen,
4561 offset + tnl_hlen);
4562
4563 if (skb_needs_linearize(nskb, features) &&
4564 __skb_linearize(nskb))
4565 goto err_linearize;
4566 }
4567
4568 skb->truesize = skb->truesize - delta_truesize;
4569 skb->data_len = skb->data_len - delta_len;
4570 skb->len = skb->len - delta_len;
4571
4572 skb_gso_reset(skb);
4573
4574 skb->prev = tail;
4575
4576 if (skb_needs_linearize(skb, features) &&
4577 __skb_linearize(skb))
4578 goto err_linearize;
4579
4580 skb_get(skb);
4581
4582 return skb;
4583
4584err_linearize:
4585 kfree_skb_list(skb->next);
4586 skb->next = NULL;
4587 return ERR_PTR(-ENOMEM);
4588}
4589EXPORT_SYMBOL_GPL(skb_segment_list);
4590
4591/**
4592 * skb_segment - Perform protocol segmentation on skb.
4593 * @head_skb: buffer to segment
4594 * @features: features for the output path (see dev->features)
4595 *
4596 * This function performs segmentation on the given skb. It returns
4597 * a pointer to the first in a list of new skbs for the segments.
4598 * In case of error it returns ERR_PTR(err).
4599 */
4600struct sk_buff *skb_segment(struct sk_buff *head_skb,
4601 netdev_features_t features)
4602{
4603 struct sk_buff *segs = NULL;
4604 struct sk_buff *tail = NULL;
4605 struct sk_buff *list_skb = skb_shinfo(head_skb)->frag_list;
4606 unsigned int mss = skb_shinfo(head_skb)->gso_size;
4607 unsigned int doffset = head_skb->data - skb_mac_header(head_skb);
4608 unsigned int offset = doffset;
4609 unsigned int tnl_hlen = skb_tnl_header_len(head_skb);
4610 unsigned int partial_segs = 0;
4611 unsigned int headroom;
4612 unsigned int len = head_skb->len;
4613 struct sk_buff *frag_skb;
4614 skb_frag_t *frag;
4615 __be16 proto;
4616 bool csum, sg;
4617 int err = -ENOMEM;
4618 int i = 0;
4619 int nfrags, pos;
4620
4621 if ((skb_shinfo(head_skb)->gso_type & SKB_GSO_DODGY) &&
4622 mss != GSO_BY_FRAGS && mss != skb_headlen(head_skb)) {
4623 struct sk_buff *check_skb;
4624
4625 for (check_skb = list_skb; check_skb; check_skb = check_skb->next) {
4626 if (skb_headlen(check_skb) && !check_skb->head_frag) {
4627 /* gso_size is untrusted, and we have a frag_list with
4628 * a linear non head_frag item.
4629 *
4630 * If head_skb's headlen does not fit requested gso_size,
4631 * it means that the frag_list members do NOT terminate
4632 * on exact gso_size boundaries. Hence we cannot perform
4633 * skb_frag_t page sharing. Therefore we must fallback to
4634 * copying the frag_list skbs; we do so by disabling SG.
4635 */
4636 features &= ~NETIF_F_SG;
4637 break;
4638 }
4639 }
4640 }
4641
4642 __skb_push(head_skb, doffset);
4643 proto = skb_network_protocol(head_skb, NULL);
4644 if (unlikely(!proto))
4645 return ERR_PTR(-EINVAL);
4646
4647 sg = !!(features & NETIF_F_SG);
4648 csum = !!can_checksum_protocol(features, proto);
4649
4650 if (sg && csum && (mss != GSO_BY_FRAGS)) {
4651 if (!(features & NETIF_F_GSO_PARTIAL)) {
4652 struct sk_buff *iter;
4653 unsigned int frag_len;
4654
4655 if (!list_skb ||
4656 !net_gso_ok(features, skb_shinfo(head_skb)->gso_type))
4657 goto normal;
4658
4659 /* If we get here then all the required
4660 * GSO features except frag_list are supported.
4661 * Try to split the SKB to multiple GSO SKBs
4662 * with no frag_list.
4663 * Currently we can do that only when the buffers don't
4664 * have a linear part and all the buffers except
4665 * the last are of the same length.
4666 */
4667 frag_len = list_skb->len;
4668 skb_walk_frags(head_skb, iter) {
4669 if (frag_len != iter->len && iter->next)
4670 goto normal;
4671 if (skb_headlen(iter) && !iter->head_frag)
4672 goto normal;
4673
4674 len -= iter->len;
4675 }
4676
4677 if (len != frag_len)
4678 goto normal;
4679 }
4680
4681 /* GSO partial only requires that we trim off any excess that
4682 * doesn't fit into an MSS sized block, so take care of that
4683 * now.
4684 * Cap len to not accidentally hit GSO_BY_FRAGS.
4685 */
4686 partial_segs = min(len, GSO_BY_FRAGS - 1) / mss;
4687 if (partial_segs > 1)
4688 mss *= partial_segs;
4689 else
4690 partial_segs = 0;
4691 }
4692
4693normal:
4694 headroom = skb_headroom(head_skb);
4695 pos = skb_headlen(head_skb);
4696
4697 if (skb_orphan_frags(head_skb, GFP_ATOMIC))
4698 return ERR_PTR(-ENOMEM);
4699
4700 nfrags = skb_shinfo(head_skb)->nr_frags;
4701 frag = skb_shinfo(head_skb)->frags;
4702 frag_skb = head_skb;
4703
4704 do {
4705 struct sk_buff *nskb;
4706 skb_frag_t *nskb_frag;
4707 int hsize;
4708 int size;
4709
4710 if (unlikely(mss == GSO_BY_FRAGS)) {
4711 len = list_skb->len;
4712 } else {
4713 len = head_skb->len - offset;
4714 if (len > mss)
4715 len = mss;
4716 }
4717
4718 hsize = skb_headlen(head_skb) - offset;
4719
4720 if (hsize <= 0 && i >= nfrags && skb_headlen(list_skb) &&
4721 (skb_headlen(list_skb) == len || sg)) {
4722 BUG_ON(skb_headlen(list_skb) > len);
4723
4724 nskb = skb_clone(list_skb, GFP_ATOMIC);
4725 if (unlikely(!nskb))
4726 goto err;
4727
4728 i = 0;
4729 nfrags = skb_shinfo(list_skb)->nr_frags;
4730 frag = skb_shinfo(list_skb)->frags;
4731 frag_skb = list_skb;
4732 pos += skb_headlen(list_skb);
4733
4734 while (pos < offset + len) {
4735 BUG_ON(i >= nfrags);
4736
4737 size = skb_frag_size(frag);
4738 if (pos + size > offset + len)
4739 break;
4740
4741 i++;
4742 pos += size;
4743 frag++;
4744 }
4745
4746 list_skb = list_skb->next;
4747
4748 if (unlikely(pskb_trim(nskb, len))) {
4749 kfree_skb(nskb);
4750 goto err;
4751 }
4752
4753 hsize = skb_end_offset(nskb);
4754 if (skb_cow_head(nskb, doffset + headroom)) {
4755 kfree_skb(nskb);
4756 goto err;
4757 }
4758
4759 nskb->truesize += skb_end_offset(nskb) - hsize;
4760 skb_release_head_state(nskb);
4761 __skb_push(nskb, doffset);
4762 } else {
4763 if (hsize < 0)
4764 hsize = 0;
4765 if (hsize > len || !sg)
4766 hsize = len;
4767
4768 nskb = __alloc_skb(hsize + doffset + headroom,
4769 GFP_ATOMIC, skb_alloc_rx_flag(head_skb),
4770 NUMA_NO_NODE);
4771
4772 if (unlikely(!nskb))
4773 goto err;
4774
4775 skb_reserve(nskb, headroom);
4776 __skb_put(nskb, doffset);
4777 }
4778
4779 if (segs)
4780 tail->next = nskb;
4781 else
4782 segs = nskb;
4783 tail = nskb;
4784
4785 __copy_skb_header(nskb, head_skb);
4786
4787 skb_headers_offset_update(nskb, skb_headroom(nskb) - headroom);
4788 skb_reset_mac_len(nskb);
4789
4790 skb_copy_from_linear_data_offset(head_skb, -tnl_hlen,
4791 nskb->data - tnl_hlen,
4792 doffset + tnl_hlen);
4793
4794 if (nskb->len == len + doffset)
4795 goto perform_csum_check;
4796
4797 if (!sg) {
4798 if (!csum) {
4799 if (!nskb->remcsum_offload)
4800 nskb->ip_summed = CHECKSUM_NONE;
4801 SKB_GSO_CB(nskb)->csum =
4802 skb_copy_and_csum_bits(head_skb, offset,
4803 skb_put(nskb,
4804 len),
4805 len);
4806 SKB_GSO_CB(nskb)->csum_start =
4807 skb_headroom(nskb) + doffset;
4808 } else {
4809 if (skb_copy_bits(head_skb, offset, skb_put(nskb, len), len))
4810 goto err;
4811 }
4812 continue;
4813 }
4814
4815 nskb_frag = skb_shinfo(nskb)->frags;
4816
4817 skb_copy_from_linear_data_offset(head_skb, offset,
4818 skb_put(nskb, hsize), hsize);
4819
4820 skb_shinfo(nskb)->flags |= skb_shinfo(head_skb)->flags &
4821 SKBFL_SHARED_FRAG;
4822
4823 if (skb_zerocopy_clone(nskb, frag_skb, GFP_ATOMIC))
4824 goto err;
4825
4826 while (pos < offset + len) {
4827 if (i >= nfrags) {
4828 if (skb_orphan_frags(list_skb, GFP_ATOMIC) ||
4829 skb_zerocopy_clone(nskb, list_skb,
4830 GFP_ATOMIC))
4831 goto err;
4832
4833 i = 0;
4834 nfrags = skb_shinfo(list_skb)->nr_frags;
4835 frag = skb_shinfo(list_skb)->frags;
4836 frag_skb = list_skb;
4837 if (!skb_headlen(list_skb)) {
4838 BUG_ON(!nfrags);
4839 } else {
4840 BUG_ON(!list_skb->head_frag);
4841
4842 /* to make room for head_frag. */
4843 i--;
4844 frag--;
4845 }
4846
4847 list_skb = list_skb->next;
4848 }
4849
4850 if (unlikely(skb_shinfo(nskb)->nr_frags >=
4851 MAX_SKB_FRAGS)) {
4852 net_warn_ratelimited(
4853 "skb_segment: too many frags: %u %u\n",
4854 pos, mss);
4855 err = -EINVAL;
4856 goto err;
4857 }
4858
4859 *nskb_frag = (i < 0) ? skb_head_frag_to_page_desc(frag_skb) : *frag;
4860 __skb_frag_ref(nskb_frag);
4861 size = skb_frag_size(nskb_frag);
4862
4863 if (pos < offset) {
4864 skb_frag_off_add(nskb_frag, offset - pos);
4865 skb_frag_size_sub(nskb_frag, offset - pos);
4866 }
4867
4868 skb_shinfo(nskb)->nr_frags++;
4869
4870 if (pos + size <= offset + len) {
4871 i++;
4872 frag++;
4873 pos += size;
4874 } else {
4875 skb_frag_size_sub(nskb_frag, pos + size - (offset + len));
4876 goto skip_fraglist;
4877 }
4878
4879 nskb_frag++;
4880 }
4881
4882skip_fraglist:
4883 nskb->data_len = len - hsize;
4884 nskb->len += nskb->data_len;
4885 nskb->truesize += nskb->data_len;
4886
4887perform_csum_check:
4888 if (!csum) {
4889 if (skb_has_shared_frag(nskb) &&
4890 __skb_linearize(nskb))
4891 goto err;
4892
4893 if (!nskb->remcsum_offload)
4894 nskb->ip_summed = CHECKSUM_NONE;
4895 SKB_GSO_CB(nskb)->csum =
4896 skb_checksum(nskb, doffset,
4897 nskb->len - doffset, 0);
4898 SKB_GSO_CB(nskb)->csum_start =
4899 skb_headroom(nskb) + doffset;
4900 }
4901 } while ((offset += len) < head_skb->len);
4902
4903 /* Some callers want to get the end of the list.
4904 * Put it in segs->prev to avoid walking the list.
4905 * (see validate_xmit_skb_list() for example)
4906 */
4907 segs->prev = tail;
4908
4909 if (partial_segs) {
4910 struct sk_buff *iter;
4911 int type = skb_shinfo(head_skb)->gso_type;
4912 unsigned short gso_size = skb_shinfo(head_skb)->gso_size;
4913
4914 /* Update type to add partial and then remove dodgy if set */
4915 type |= (features & NETIF_F_GSO_PARTIAL) / NETIF_F_GSO_PARTIAL * SKB_GSO_PARTIAL;
4916 type &= ~SKB_GSO_DODGY;
4917
4918 /* Update GSO info and prepare to start updating headers on
4919 * our way back down the stack of protocols.
4920 */
4921 for (iter = segs; iter; iter = iter->next) {
4922 skb_shinfo(iter)->gso_size = gso_size;
4923 skb_shinfo(iter)->gso_segs = partial_segs;
4924 skb_shinfo(iter)->gso_type = type;
4925 SKB_GSO_CB(iter)->data_offset = skb_headroom(iter) + doffset;
4926 }
4927
4928 if (tail->len - doffset <= gso_size)
4929 skb_shinfo(tail)->gso_size = 0;
4930 else if (tail != segs)
4931 skb_shinfo(tail)->gso_segs = DIV_ROUND_UP(tail->len - doffset, gso_size);
4932 }
4933
4934 /* Following permits correct backpressure, for protocols
4935 * using skb_set_owner_w().
4936 * Idea is to tranfert ownership from head_skb to last segment.
4937 */
4938 if (head_skb->destructor == sock_wfree) {
4939 swap(tail->truesize, head_skb->truesize);
4940 swap(tail->destructor, head_skb->destructor);
4941 swap(tail->sk, head_skb->sk);
4942 }
4943 return segs;
4944
4945err:
4946 kfree_skb_list(segs);
4947 return ERR_PTR(err);
4948}
4949EXPORT_SYMBOL_GPL(skb_segment);
4950
4951#ifdef CONFIG_SKB_EXTENSIONS
4952#define SKB_EXT_ALIGN_VALUE 8
4953#define SKB_EXT_CHUNKSIZEOF(x) (ALIGN((sizeof(x)), SKB_EXT_ALIGN_VALUE) / SKB_EXT_ALIGN_VALUE)
4954
4955static const u8 skb_ext_type_len[] = {
4956#if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
4957 [SKB_EXT_BRIDGE_NF] = SKB_EXT_CHUNKSIZEOF(struct nf_bridge_info),
4958#endif
4959#ifdef CONFIG_XFRM
4960 [SKB_EXT_SEC_PATH] = SKB_EXT_CHUNKSIZEOF(struct sec_path),
4961#endif
4962#if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
4963 [TC_SKB_EXT] = SKB_EXT_CHUNKSIZEOF(struct tc_skb_ext),
4964#endif
4965#if IS_ENABLED(CONFIG_MPTCP)
4966 [SKB_EXT_MPTCP] = SKB_EXT_CHUNKSIZEOF(struct mptcp_ext),
4967#endif
4968#if IS_ENABLED(CONFIG_MCTP_FLOWS)
4969 [SKB_EXT_MCTP] = SKB_EXT_CHUNKSIZEOF(struct mctp_flow),
4970#endif
4971};
4972
4973static __always_inline unsigned int skb_ext_total_length(void)
4974{
4975 unsigned int l = SKB_EXT_CHUNKSIZEOF(struct skb_ext);
4976 int i;
4977
4978 for (i = 0; i < ARRAY_SIZE(skb_ext_type_len); i++)
4979 l += skb_ext_type_len[i];
4980
4981 return l;
4982}
4983
4984static void skb_extensions_init(void)
4985{
4986 BUILD_BUG_ON(SKB_EXT_NUM >= 8);
4987#if !IS_ENABLED(CONFIG_KCOV_INSTRUMENT_ALL)
4988 BUILD_BUG_ON(skb_ext_total_length() > 255);
4989#endif
4990
4991 skbuff_ext_cache = kmem_cache_create("skbuff_ext_cache",
4992 SKB_EXT_ALIGN_VALUE * skb_ext_total_length(),
4993 0,
4994 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
4995 NULL);
4996}
4997#else
4998static void skb_extensions_init(void) {}
4999#endif
5000
5001/* The SKB kmem_cache slab is critical for network performance. Never
5002 * merge/alias the slab with similar sized objects. This avoids fragmentation
5003 * that hurts performance of kmem_cache_{alloc,free}_bulk APIs.
5004 */
5005#ifndef CONFIG_SLUB_TINY
5006#define FLAG_SKB_NO_MERGE SLAB_NO_MERGE
5007#else /* CONFIG_SLUB_TINY - simple loop in kmem_cache_alloc_bulk */
5008#define FLAG_SKB_NO_MERGE 0
5009#endif
5010
5011void __init skb_init(void)
5012{
5013 net_hotdata.skbuff_cache = kmem_cache_create_usercopy("skbuff_head_cache",
5014 sizeof(struct sk_buff),
5015 0,
5016 SLAB_HWCACHE_ALIGN|SLAB_PANIC|
5017 FLAG_SKB_NO_MERGE,
5018 offsetof(struct sk_buff, cb),
5019 sizeof_field(struct sk_buff, cb),
5020 NULL);
5021 net_hotdata.skbuff_fclone_cache = kmem_cache_create("skbuff_fclone_cache",
5022 sizeof(struct sk_buff_fclones),
5023 0,
5024 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
5025 NULL);
5026 /* usercopy should only access first SKB_SMALL_HEAD_HEADROOM bytes.
5027 * struct skb_shared_info is located at the end of skb->head,
5028 * and should not be copied to/from user.
5029 */
5030 net_hotdata.skb_small_head_cache = kmem_cache_create_usercopy("skbuff_small_head",
5031 SKB_SMALL_HEAD_CACHE_SIZE,
5032 0,
5033 SLAB_HWCACHE_ALIGN | SLAB_PANIC,
5034 0,
5035 SKB_SMALL_HEAD_HEADROOM,
5036 NULL);
5037 skb_extensions_init();
5038}
5039
5040static int
5041__skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len,
5042 unsigned int recursion_level)
5043{
5044 int start = skb_headlen(skb);
5045 int i, copy = start - offset;
5046 struct sk_buff *frag_iter;
5047 int elt = 0;
5048
5049 if (unlikely(recursion_level >= 24))
5050 return -EMSGSIZE;
5051
5052 if (copy > 0) {
5053 if (copy > len)
5054 copy = len;
5055 sg_set_buf(sg, skb->data + offset, copy);
5056 elt++;
5057 if ((len -= copy) == 0)
5058 return elt;
5059 offset += copy;
5060 }
5061
5062 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
5063 int end;
5064
5065 WARN_ON(start > offset + len);
5066
5067 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
5068 if ((copy = end - offset) > 0) {
5069 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
5070 if (unlikely(elt && sg_is_last(&sg[elt - 1])))
5071 return -EMSGSIZE;
5072
5073 if (copy > len)
5074 copy = len;
5075 sg_set_page(&sg[elt], skb_frag_page(frag), copy,
5076 skb_frag_off(frag) + offset - start);
5077 elt++;
5078 if (!(len -= copy))
5079 return elt;
5080 offset += copy;
5081 }
5082 start = end;
5083 }
5084
5085 skb_walk_frags(skb, frag_iter) {
5086 int end, ret;
5087
5088 WARN_ON(start > offset + len);
5089
5090 end = start + frag_iter->len;
5091 if ((copy = end - offset) > 0) {
5092 if (unlikely(elt && sg_is_last(&sg[elt - 1])))
5093 return -EMSGSIZE;
5094
5095 if (copy > len)
5096 copy = len;
5097 ret = __skb_to_sgvec(frag_iter, sg+elt, offset - start,
5098 copy, recursion_level + 1);
5099 if (unlikely(ret < 0))
5100 return ret;
5101 elt += ret;
5102 if ((len -= copy) == 0)
5103 return elt;
5104 offset += copy;
5105 }
5106 start = end;
5107 }
5108 BUG_ON(len);
5109 return elt;
5110}
5111
5112/**
5113 * skb_to_sgvec - Fill a scatter-gather list from a socket buffer
5114 * @skb: Socket buffer containing the buffers to be mapped
5115 * @sg: The scatter-gather list to map into
5116 * @offset: The offset into the buffer's contents to start mapping
5117 * @len: Length of buffer space to be mapped
5118 *
5119 * Fill the specified scatter-gather list with mappings/pointers into a
5120 * region of the buffer space attached to a socket buffer. Returns either
5121 * the number of scatterlist items used, or -EMSGSIZE if the contents
5122 * could not fit.
5123 */
5124int skb_to_sgvec(struct sk_buff *skb, struct scatterlist *sg, int offset, int len)
5125{
5126 int nsg = __skb_to_sgvec(skb, sg, offset, len, 0);
5127
5128 if (nsg <= 0)
5129 return nsg;
5130
5131 sg_mark_end(&sg[nsg - 1]);
5132
5133 return nsg;
5134}
5135EXPORT_SYMBOL_GPL(skb_to_sgvec);
5136
5137/* As compared with skb_to_sgvec, skb_to_sgvec_nomark only map skb to given
5138 * sglist without mark the sg which contain last skb data as the end.
5139 * So the caller can mannipulate sg list as will when padding new data after
5140 * the first call without calling sg_unmark_end to expend sg list.
5141 *
5142 * Scenario to use skb_to_sgvec_nomark:
5143 * 1. sg_init_table
5144 * 2. skb_to_sgvec_nomark(payload1)
5145 * 3. skb_to_sgvec_nomark(payload2)
5146 *
5147 * This is equivalent to:
5148 * 1. sg_init_table
5149 * 2. skb_to_sgvec(payload1)
5150 * 3. sg_unmark_end
5151 * 4. skb_to_sgvec(payload2)
5152 *
5153 * When mapping mutilple payload conditionally, skb_to_sgvec_nomark
5154 * is more preferable.
5155 */
5156int skb_to_sgvec_nomark(struct sk_buff *skb, struct scatterlist *sg,
5157 int offset, int len)
5158{
5159 return __skb_to_sgvec(skb, sg, offset, len, 0);
5160}
5161EXPORT_SYMBOL_GPL(skb_to_sgvec_nomark);
5162
5163
5164
5165/**
5166 * skb_cow_data - Check that a socket buffer's data buffers are writable
5167 * @skb: The socket buffer to check.
5168 * @tailbits: Amount of trailing space to be added
5169 * @trailer: Returned pointer to the skb where the @tailbits space begins
5170 *
5171 * Make sure that the data buffers attached to a socket buffer are
5172 * writable. If they are not, private copies are made of the data buffers
5173 * and the socket buffer is set to use these instead.
5174 *
5175 * If @tailbits is given, make sure that there is space to write @tailbits
5176 * bytes of data beyond current end of socket buffer. @trailer will be
5177 * set to point to the skb in which this space begins.
5178 *
5179 * The number of scatterlist elements required to completely map the
5180 * COW'd and extended socket buffer will be returned.
5181 */
5182int skb_cow_data(struct sk_buff *skb, int tailbits, struct sk_buff **trailer)
5183{
5184 int copyflag;
5185 int elt;
5186 struct sk_buff *skb1, **skb_p;
5187
5188 /* If skb is cloned or its head is paged, reallocate
5189 * head pulling out all the pages (pages are considered not writable
5190 * at the moment even if they are anonymous).
5191 */
5192 if ((skb_cloned(skb) || skb_shinfo(skb)->nr_frags) &&
5193 !__pskb_pull_tail(skb, __skb_pagelen(skb)))
5194 return -ENOMEM;
5195
5196 /* Easy case. Most of packets will go this way. */
5197 if (!skb_has_frag_list(skb)) {
5198 /* A little of trouble, not enough of space for trailer.
5199 * This should not happen, when stack is tuned to generate
5200 * good frames. OK, on miss we reallocate and reserve even more
5201 * space, 128 bytes is fair. */
5202
5203 if (skb_tailroom(skb) < tailbits &&
5204 pskb_expand_head(skb, 0, tailbits-skb_tailroom(skb)+128, GFP_ATOMIC))
5205 return -ENOMEM;
5206
5207 /* Voila! */
5208 *trailer = skb;
5209 return 1;
5210 }
5211
5212 /* Misery. We are in troubles, going to mincer fragments... */
5213
5214 elt = 1;
5215 skb_p = &skb_shinfo(skb)->frag_list;
5216 copyflag = 0;
5217
5218 while ((skb1 = *skb_p) != NULL) {
5219 int ntail = 0;
5220
5221 /* The fragment is partially pulled by someone,
5222 * this can happen on input. Copy it and everything
5223 * after it. */
5224
5225 if (skb_shared(skb1))
5226 copyflag = 1;
5227
5228 /* If the skb is the last, worry about trailer. */
5229
5230 if (skb1->next == NULL && tailbits) {
5231 if (skb_shinfo(skb1)->nr_frags ||
5232 skb_has_frag_list(skb1) ||
5233 skb_tailroom(skb1) < tailbits)
5234 ntail = tailbits + 128;
5235 }
5236
5237 if (copyflag ||
5238 skb_cloned(skb1) ||
5239 ntail ||
5240 skb_shinfo(skb1)->nr_frags ||
5241 skb_has_frag_list(skb1)) {
5242 struct sk_buff *skb2;
5243
5244 /* Fuck, we are miserable poor guys... */
5245 if (ntail == 0)
5246 skb2 = skb_copy(skb1, GFP_ATOMIC);
5247 else
5248 skb2 = skb_copy_expand(skb1,
5249 skb_headroom(skb1),
5250 ntail,
5251 GFP_ATOMIC);
5252 if (unlikely(skb2 == NULL))
5253 return -ENOMEM;
5254
5255 if (skb1->sk)
5256 skb_set_owner_w(skb2, skb1->sk);
5257
5258 /* Looking around. Are we still alive?
5259 * OK, link new skb, drop old one */
5260
5261 skb2->next = skb1->next;
5262 *skb_p = skb2;
5263 kfree_skb(skb1);
5264 skb1 = skb2;
5265 }
5266 elt++;
5267 *trailer = skb1;
5268 skb_p = &skb1->next;
5269 }
5270
5271 return elt;
5272}
5273EXPORT_SYMBOL_GPL(skb_cow_data);
5274
5275static void sock_rmem_free(struct sk_buff *skb)
5276{
5277 struct sock *sk = skb->sk;
5278
5279 atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
5280}
5281
5282static void skb_set_err_queue(struct sk_buff *skb)
5283{
5284 /* pkt_type of skbs received on local sockets is never PACKET_OUTGOING.
5285 * So, it is safe to (mis)use it to mark skbs on the error queue.
5286 */
5287 skb->pkt_type = PACKET_OUTGOING;
5288 BUILD_BUG_ON(PACKET_OUTGOING == 0);
5289}
5290
5291/*
5292 * Note: We dont mem charge error packets (no sk_forward_alloc changes)
5293 */
5294int sock_queue_err_skb(struct sock *sk, struct sk_buff *skb)
5295{
5296 if (atomic_read(&sk->sk_rmem_alloc) + skb->truesize >=
5297 (unsigned int)READ_ONCE(sk->sk_rcvbuf))
5298 return -ENOMEM;
5299
5300 skb_orphan(skb);
5301 skb->sk = sk;
5302 skb->destructor = sock_rmem_free;
5303 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
5304 skb_set_err_queue(skb);
5305
5306 /* before exiting rcu section, make sure dst is refcounted */
5307 skb_dst_force(skb);
5308
5309 skb_queue_tail(&sk->sk_error_queue, skb);
5310 if (!sock_flag(sk, SOCK_DEAD))
5311 sk_error_report(sk);
5312 return 0;
5313}
5314EXPORT_SYMBOL(sock_queue_err_skb);
5315
5316static bool is_icmp_err_skb(const struct sk_buff *skb)
5317{
5318 return skb && (SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP ||
5319 SKB_EXT_ERR(skb)->ee.ee_origin == SO_EE_ORIGIN_ICMP6);
5320}
5321
5322struct sk_buff *sock_dequeue_err_skb(struct sock *sk)
5323{
5324 struct sk_buff_head *q = &sk->sk_error_queue;
5325 struct sk_buff *skb, *skb_next = NULL;
5326 bool icmp_next = false;
5327 unsigned long flags;
5328
5329 if (skb_queue_empty_lockless(q))
5330 return NULL;
5331
5332 spin_lock_irqsave(&q->lock, flags);
5333 skb = __skb_dequeue(q);
5334 if (skb && (skb_next = skb_peek(q))) {
5335 icmp_next = is_icmp_err_skb(skb_next);
5336 if (icmp_next)
5337 sk->sk_err = SKB_EXT_ERR(skb_next)->ee.ee_errno;
5338 }
5339 spin_unlock_irqrestore(&q->lock, flags);
5340
5341 if (is_icmp_err_skb(skb) && !icmp_next)
5342 sk->sk_err = 0;
5343
5344 if (skb_next)
5345 sk_error_report(sk);
5346
5347 return skb;
5348}
5349EXPORT_SYMBOL(sock_dequeue_err_skb);
5350
5351/**
5352 * skb_clone_sk - create clone of skb, and take reference to socket
5353 * @skb: the skb to clone
5354 *
5355 * This function creates a clone of a buffer that holds a reference on
5356 * sk_refcnt. Buffers created via this function are meant to be
5357 * returned using sock_queue_err_skb, or free via kfree_skb.
5358 *
5359 * When passing buffers allocated with this function to sock_queue_err_skb
5360 * it is necessary to wrap the call with sock_hold/sock_put in order to
5361 * prevent the socket from being released prior to being enqueued on
5362 * the sk_error_queue.
5363 */
5364struct sk_buff *skb_clone_sk(struct sk_buff *skb)
5365{
5366 struct sock *sk = skb->sk;
5367 struct sk_buff *clone;
5368
5369 if (!sk || !refcount_inc_not_zero(&sk->sk_refcnt))
5370 return NULL;
5371
5372 clone = skb_clone(skb, GFP_ATOMIC);
5373 if (!clone) {
5374 sock_put(sk);
5375 return NULL;
5376 }
5377
5378 clone->sk = sk;
5379 clone->destructor = sock_efree;
5380
5381 return clone;
5382}
5383EXPORT_SYMBOL(skb_clone_sk);
5384
5385static void __skb_complete_tx_timestamp(struct sk_buff *skb,
5386 struct sock *sk,
5387 int tstype,
5388 bool opt_stats)
5389{
5390 struct sock_exterr_skb *serr;
5391 int err;
5392
5393 BUILD_BUG_ON(sizeof(struct sock_exterr_skb) > sizeof(skb->cb));
5394
5395 serr = SKB_EXT_ERR(skb);
5396 memset(serr, 0, sizeof(*serr));
5397 serr->ee.ee_errno = ENOMSG;
5398 serr->ee.ee_origin = SO_EE_ORIGIN_TIMESTAMPING;
5399 serr->ee.ee_info = tstype;
5400 serr->opt_stats = opt_stats;
5401 serr->header.h4.iif = skb->dev ? skb->dev->ifindex : 0;
5402 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_OPT_ID) {
5403 serr->ee.ee_data = skb_shinfo(skb)->tskey;
5404 if (sk_is_tcp(sk))
5405 serr->ee.ee_data -= atomic_read(&sk->sk_tskey);
5406 }
5407
5408 err = sock_queue_err_skb(sk, skb);
5409
5410 if (err)
5411 kfree_skb(skb);
5412}
5413
5414static bool skb_may_tx_timestamp(struct sock *sk, bool tsonly)
5415{
5416 bool ret;
5417
5418 if (likely(READ_ONCE(sysctl_tstamp_allow_data) || tsonly))
5419 return true;
5420
5421 read_lock_bh(&sk->sk_callback_lock);
5422 ret = sk->sk_socket && sk->sk_socket->file &&
5423 file_ns_capable(sk->sk_socket->file, &init_user_ns, CAP_NET_RAW);
5424 read_unlock_bh(&sk->sk_callback_lock);
5425 return ret;
5426}
5427
5428void skb_complete_tx_timestamp(struct sk_buff *skb,
5429 struct skb_shared_hwtstamps *hwtstamps)
5430{
5431 struct sock *sk = skb->sk;
5432
5433 if (!skb_may_tx_timestamp(sk, false))
5434 goto err;
5435
5436 /* Take a reference to prevent skb_orphan() from freeing the socket,
5437 * but only if the socket refcount is not zero.
5438 */
5439 if (likely(refcount_inc_not_zero(&sk->sk_refcnt))) {
5440 *skb_hwtstamps(skb) = *hwtstamps;
5441 __skb_complete_tx_timestamp(skb, sk, SCM_TSTAMP_SND, false);
5442 sock_put(sk);
5443 return;
5444 }
5445
5446err:
5447 kfree_skb(skb);
5448}
5449EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp);
5450
5451void __skb_tstamp_tx(struct sk_buff *orig_skb,
5452 const struct sk_buff *ack_skb,
5453 struct skb_shared_hwtstamps *hwtstamps,
5454 struct sock *sk, int tstype)
5455{
5456 struct sk_buff *skb;
5457 bool tsonly, opt_stats = false;
5458 u32 tsflags;
5459
5460 if (!sk)
5461 return;
5462
5463 tsflags = READ_ONCE(sk->sk_tsflags);
5464 if (!hwtstamps && !(tsflags & SOF_TIMESTAMPING_OPT_TX_SWHW) &&
5465 skb_shinfo(orig_skb)->tx_flags & SKBTX_IN_PROGRESS)
5466 return;
5467
5468 tsonly = tsflags & SOF_TIMESTAMPING_OPT_TSONLY;
5469 if (!skb_may_tx_timestamp(sk, tsonly))
5470 return;
5471
5472 if (tsonly) {
5473#ifdef CONFIG_INET
5474 if ((tsflags & SOF_TIMESTAMPING_OPT_STATS) &&
5475 sk_is_tcp(sk)) {
5476 skb = tcp_get_timestamping_opt_stats(sk, orig_skb,
5477 ack_skb);
5478 opt_stats = true;
5479 } else
5480#endif
5481 skb = alloc_skb(0, GFP_ATOMIC);
5482 } else {
5483 skb = skb_clone(orig_skb, GFP_ATOMIC);
5484
5485 if (skb_orphan_frags_rx(skb, GFP_ATOMIC)) {
5486 kfree_skb(skb);
5487 return;
5488 }
5489 }
5490 if (!skb)
5491 return;
5492
5493 if (tsonly) {
5494 skb_shinfo(skb)->tx_flags |= skb_shinfo(orig_skb)->tx_flags &
5495 SKBTX_ANY_TSTAMP;
5496 skb_shinfo(skb)->tskey = skb_shinfo(orig_skb)->tskey;
5497 }
5498
5499 if (hwtstamps)
5500 *skb_hwtstamps(skb) = *hwtstamps;
5501 else
5502 __net_timestamp(skb);
5503
5504 __skb_complete_tx_timestamp(skb, sk, tstype, opt_stats);
5505}
5506EXPORT_SYMBOL_GPL(__skb_tstamp_tx);
5507
5508void skb_tstamp_tx(struct sk_buff *orig_skb,
5509 struct skb_shared_hwtstamps *hwtstamps)
5510{
5511 return __skb_tstamp_tx(orig_skb, NULL, hwtstamps, orig_skb->sk,
5512 SCM_TSTAMP_SND);
5513}
5514EXPORT_SYMBOL_GPL(skb_tstamp_tx);
5515
5516#ifdef CONFIG_WIRELESS
5517void skb_complete_wifi_ack(struct sk_buff *skb, bool acked)
5518{
5519 struct sock *sk = skb->sk;
5520 struct sock_exterr_skb *serr;
5521 int err = 1;
5522
5523 skb->wifi_acked_valid = 1;
5524 skb->wifi_acked = acked;
5525
5526 serr = SKB_EXT_ERR(skb);
5527 memset(serr, 0, sizeof(*serr));
5528 serr->ee.ee_errno = ENOMSG;
5529 serr->ee.ee_origin = SO_EE_ORIGIN_TXSTATUS;
5530
5531 /* Take a reference to prevent skb_orphan() from freeing the socket,
5532 * but only if the socket refcount is not zero.
5533 */
5534 if (likely(refcount_inc_not_zero(&sk->sk_refcnt))) {
5535 err = sock_queue_err_skb(sk, skb);
5536 sock_put(sk);
5537 }
5538 if (err)
5539 kfree_skb(skb);
5540}
5541EXPORT_SYMBOL_GPL(skb_complete_wifi_ack);
5542#endif /* CONFIG_WIRELESS */
5543
5544/**
5545 * skb_partial_csum_set - set up and verify partial csum values for packet
5546 * @skb: the skb to set
5547 * @start: the number of bytes after skb->data to start checksumming.
5548 * @off: the offset from start to place the checksum.
5549 *
5550 * For untrusted partially-checksummed packets, we need to make sure the values
5551 * for skb->csum_start and skb->csum_offset are valid so we don't oops.
5552 *
5553 * This function checks and sets those values and skb->ip_summed: if this
5554 * returns false you should drop the packet.
5555 */
5556bool skb_partial_csum_set(struct sk_buff *skb, u16 start, u16 off)
5557{
5558 u32 csum_end = (u32)start + (u32)off + sizeof(__sum16);
5559 u32 csum_start = skb_headroom(skb) + (u32)start;
5560
5561 if (unlikely(csum_start >= U16_MAX || csum_end > skb_headlen(skb))) {
5562 net_warn_ratelimited("bad partial csum: csum=%u/%u headroom=%u headlen=%u\n",
5563 start, off, skb_headroom(skb), skb_headlen(skb));
5564 return false;
5565 }
5566 skb->ip_summed = CHECKSUM_PARTIAL;
5567 skb->csum_start = csum_start;
5568 skb->csum_offset = off;
5569 skb->transport_header = csum_start;
5570 return true;
5571}
5572EXPORT_SYMBOL_GPL(skb_partial_csum_set);
5573
5574static int skb_maybe_pull_tail(struct sk_buff *skb, unsigned int len,
5575 unsigned int max)
5576{
5577 if (skb_headlen(skb) >= len)
5578 return 0;
5579
5580 /* If we need to pullup then pullup to the max, so we
5581 * won't need to do it again.
5582 */
5583 if (max > skb->len)
5584 max = skb->len;
5585
5586 if (__pskb_pull_tail(skb, max - skb_headlen(skb)) == NULL)
5587 return -ENOMEM;
5588
5589 if (skb_headlen(skb) < len)
5590 return -EPROTO;
5591
5592 return 0;
5593}
5594
5595#define MAX_TCP_HDR_LEN (15 * 4)
5596
5597static __sum16 *skb_checksum_setup_ip(struct sk_buff *skb,
5598 typeof(IPPROTO_IP) proto,
5599 unsigned int off)
5600{
5601 int err;
5602
5603 switch (proto) {
5604 case IPPROTO_TCP:
5605 err = skb_maybe_pull_tail(skb, off + sizeof(struct tcphdr),
5606 off + MAX_TCP_HDR_LEN);
5607 if (!err && !skb_partial_csum_set(skb, off,
5608 offsetof(struct tcphdr,
5609 check)))
5610 err = -EPROTO;
5611 return err ? ERR_PTR(err) : &tcp_hdr(skb)->check;
5612
5613 case IPPROTO_UDP:
5614 err = skb_maybe_pull_tail(skb, off + sizeof(struct udphdr),
5615 off + sizeof(struct udphdr));
5616 if (!err && !skb_partial_csum_set(skb, off,
5617 offsetof(struct udphdr,
5618 check)))
5619 err = -EPROTO;
5620 return err ? ERR_PTR(err) : &udp_hdr(skb)->check;
5621 }
5622
5623 return ERR_PTR(-EPROTO);
5624}
5625
5626/* This value should be large enough to cover a tagged ethernet header plus
5627 * maximally sized IP and TCP or UDP headers.
5628 */
5629#define MAX_IP_HDR_LEN 128
5630
5631static int skb_checksum_setup_ipv4(struct sk_buff *skb, bool recalculate)
5632{
5633 unsigned int off;
5634 bool fragment;
5635 __sum16 *csum;
5636 int err;
5637
5638 fragment = false;
5639
5640 err = skb_maybe_pull_tail(skb,
5641 sizeof(struct iphdr),
5642 MAX_IP_HDR_LEN);
5643 if (err < 0)
5644 goto out;
5645
5646 if (ip_is_fragment(ip_hdr(skb)))
5647 fragment = true;
5648
5649 off = ip_hdrlen(skb);
5650
5651 err = -EPROTO;
5652
5653 if (fragment)
5654 goto out;
5655
5656 csum = skb_checksum_setup_ip(skb, ip_hdr(skb)->protocol, off);
5657 if (IS_ERR(csum))
5658 return PTR_ERR(csum);
5659
5660 if (recalculate)
5661 *csum = ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
5662 ip_hdr(skb)->daddr,
5663 skb->len - off,
5664 ip_hdr(skb)->protocol, 0);
5665 err = 0;
5666
5667out:
5668 return err;
5669}
5670
5671/* This value should be large enough to cover a tagged ethernet header plus
5672 * an IPv6 header, all options, and a maximal TCP or UDP header.
5673 */
5674#define MAX_IPV6_HDR_LEN 256
5675
5676#define OPT_HDR(type, skb, off) \
5677 (type *)(skb_network_header(skb) + (off))
5678
5679static int skb_checksum_setup_ipv6(struct sk_buff *skb, bool recalculate)
5680{
5681 int err;
5682 u8 nexthdr;
5683 unsigned int off;
5684 unsigned int len;
5685 bool fragment;
5686 bool done;
5687 __sum16 *csum;
5688
5689 fragment = false;
5690 done = false;
5691
5692 off = sizeof(struct ipv6hdr);
5693
5694 err = skb_maybe_pull_tail(skb, off, MAX_IPV6_HDR_LEN);
5695 if (err < 0)
5696 goto out;
5697
5698 nexthdr = ipv6_hdr(skb)->nexthdr;
5699
5700 len = sizeof(struct ipv6hdr) + ntohs(ipv6_hdr(skb)->payload_len);
5701 while (off <= len && !done) {
5702 switch (nexthdr) {
5703 case IPPROTO_DSTOPTS:
5704 case IPPROTO_HOPOPTS:
5705 case IPPROTO_ROUTING: {
5706 struct ipv6_opt_hdr *hp;
5707
5708 err = skb_maybe_pull_tail(skb,
5709 off +
5710 sizeof(struct ipv6_opt_hdr),
5711 MAX_IPV6_HDR_LEN);
5712 if (err < 0)
5713 goto out;
5714
5715 hp = OPT_HDR(struct ipv6_opt_hdr, skb, off);
5716 nexthdr = hp->nexthdr;
5717 off += ipv6_optlen(hp);
5718 break;
5719 }
5720 case IPPROTO_AH: {
5721 struct ip_auth_hdr *hp;
5722
5723 err = skb_maybe_pull_tail(skb,
5724 off +
5725 sizeof(struct ip_auth_hdr),
5726 MAX_IPV6_HDR_LEN);
5727 if (err < 0)
5728 goto out;
5729
5730 hp = OPT_HDR(struct ip_auth_hdr, skb, off);
5731 nexthdr = hp->nexthdr;
5732 off += ipv6_authlen(hp);
5733 break;
5734 }
5735 case IPPROTO_FRAGMENT: {
5736 struct frag_hdr *hp;
5737
5738 err = skb_maybe_pull_tail(skb,
5739 off +
5740 sizeof(struct frag_hdr),
5741 MAX_IPV6_HDR_LEN);
5742 if (err < 0)
5743 goto out;
5744
5745 hp = OPT_HDR(struct frag_hdr, skb, off);
5746
5747 if (hp->frag_off & htons(IP6_OFFSET | IP6_MF))
5748 fragment = true;
5749
5750 nexthdr = hp->nexthdr;
5751 off += sizeof(struct frag_hdr);
5752 break;
5753 }
5754 default:
5755 done = true;
5756 break;
5757 }
5758 }
5759
5760 err = -EPROTO;
5761
5762 if (!done || fragment)
5763 goto out;
5764
5765 csum = skb_checksum_setup_ip(skb, nexthdr, off);
5766 if (IS_ERR(csum))
5767 return PTR_ERR(csum);
5768
5769 if (recalculate)
5770 *csum = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
5771 &ipv6_hdr(skb)->daddr,
5772 skb->len - off, nexthdr, 0);
5773 err = 0;
5774
5775out:
5776 return err;
5777}
5778
5779/**
5780 * skb_checksum_setup - set up partial checksum offset
5781 * @skb: the skb to set up
5782 * @recalculate: if true the pseudo-header checksum will be recalculated
5783 */
5784int skb_checksum_setup(struct sk_buff *skb, bool recalculate)
5785{
5786 int err;
5787
5788 switch (skb->protocol) {
5789 case htons(ETH_P_IP):
5790 err = skb_checksum_setup_ipv4(skb, recalculate);
5791 break;
5792
5793 case htons(ETH_P_IPV6):
5794 err = skb_checksum_setup_ipv6(skb, recalculate);
5795 break;
5796
5797 default:
5798 err = -EPROTO;
5799 break;
5800 }
5801
5802 return err;
5803}
5804EXPORT_SYMBOL(skb_checksum_setup);
5805
5806/**
5807 * skb_checksum_maybe_trim - maybe trims the given skb
5808 * @skb: the skb to check
5809 * @transport_len: the data length beyond the network header
5810 *
5811 * Checks whether the given skb has data beyond the given transport length.
5812 * If so, returns a cloned skb trimmed to this transport length.
5813 * Otherwise returns the provided skb. Returns NULL in error cases
5814 * (e.g. transport_len exceeds skb length or out-of-memory).
5815 *
5816 * Caller needs to set the skb transport header and free any returned skb if it
5817 * differs from the provided skb.
5818 */
5819static struct sk_buff *skb_checksum_maybe_trim(struct sk_buff *skb,
5820 unsigned int transport_len)
5821{
5822 struct sk_buff *skb_chk;
5823 unsigned int len = skb_transport_offset(skb) + transport_len;
5824 int ret;
5825
5826 if (skb->len < len)
5827 return NULL;
5828 else if (skb->len == len)
5829 return skb;
5830
5831 skb_chk = skb_clone(skb, GFP_ATOMIC);
5832 if (!skb_chk)
5833 return NULL;
5834
5835 ret = pskb_trim_rcsum(skb_chk, len);
5836 if (ret) {
5837 kfree_skb(skb_chk);
5838 return NULL;
5839 }
5840
5841 return skb_chk;
5842}
5843
5844/**
5845 * skb_checksum_trimmed - validate checksum of an skb
5846 * @skb: the skb to check
5847 * @transport_len: the data length beyond the network header
5848 * @skb_chkf: checksum function to use
5849 *
5850 * Applies the given checksum function skb_chkf to the provided skb.
5851 * Returns a checked and maybe trimmed skb. Returns NULL on error.
5852 *
5853 * If the skb has data beyond the given transport length, then a
5854 * trimmed & cloned skb is checked and returned.
5855 *
5856 * Caller needs to set the skb transport header and free any returned skb if it
5857 * differs from the provided skb.
5858 */
5859struct sk_buff *skb_checksum_trimmed(struct sk_buff *skb,
5860 unsigned int transport_len,
5861 __sum16(*skb_chkf)(struct sk_buff *skb))
5862{
5863 struct sk_buff *skb_chk;
5864 unsigned int offset = skb_transport_offset(skb);
5865 __sum16 ret;
5866
5867 skb_chk = skb_checksum_maybe_trim(skb, transport_len);
5868 if (!skb_chk)
5869 goto err;
5870
5871 if (!pskb_may_pull(skb_chk, offset))
5872 goto err;
5873
5874 skb_pull_rcsum(skb_chk, offset);
5875 ret = skb_chkf(skb_chk);
5876 skb_push_rcsum(skb_chk, offset);
5877
5878 if (ret)
5879 goto err;
5880
5881 return skb_chk;
5882
5883err:
5884 if (skb_chk && skb_chk != skb)
5885 kfree_skb(skb_chk);
5886
5887 return NULL;
5888
5889}
5890EXPORT_SYMBOL(skb_checksum_trimmed);
5891
5892void __skb_warn_lro_forwarding(const struct sk_buff *skb)
5893{
5894 net_warn_ratelimited("%s: received packets cannot be forwarded while LRO is enabled\n",
5895 skb->dev->name);
5896}
5897EXPORT_SYMBOL(__skb_warn_lro_forwarding);
5898
5899void kfree_skb_partial(struct sk_buff *skb, bool head_stolen)
5900{
5901 if (head_stolen) {
5902 skb_release_head_state(skb);
5903 kmem_cache_free(net_hotdata.skbuff_cache, skb);
5904 } else {
5905 __kfree_skb(skb);
5906 }
5907}
5908EXPORT_SYMBOL(kfree_skb_partial);
5909
5910/**
5911 * skb_try_coalesce - try to merge skb to prior one
5912 * @to: prior buffer
5913 * @from: buffer to add
5914 * @fragstolen: pointer to boolean
5915 * @delta_truesize: how much more was allocated than was requested
5916 */
5917bool skb_try_coalesce(struct sk_buff *to, struct sk_buff *from,
5918 bool *fragstolen, int *delta_truesize)
5919{
5920 struct skb_shared_info *to_shinfo, *from_shinfo;
5921 int i, delta, len = from->len;
5922
5923 *fragstolen = false;
5924
5925 if (skb_cloned(to))
5926 return false;
5927
5928 /* In general, avoid mixing page_pool and non-page_pool allocated
5929 * pages within the same SKB. In theory we could take full
5930 * references if @from is cloned and !@to->pp_recycle but its
5931 * tricky (due to potential race with the clone disappearing) and
5932 * rare, so not worth dealing with.
5933 */
5934 if (to->pp_recycle != from->pp_recycle)
5935 return false;
5936
5937 if (len <= skb_tailroom(to)) {
5938 if (len)
5939 BUG_ON(skb_copy_bits(from, 0, skb_put(to, len), len));
5940 *delta_truesize = 0;
5941 return true;
5942 }
5943
5944 to_shinfo = skb_shinfo(to);
5945 from_shinfo = skb_shinfo(from);
5946 if (to_shinfo->frag_list || from_shinfo->frag_list)
5947 return false;
5948 if (skb_zcopy(to) || skb_zcopy(from))
5949 return false;
5950
5951 if (skb_headlen(from) != 0) {
5952 struct page *page;
5953 unsigned int offset;
5954
5955 if (to_shinfo->nr_frags +
5956 from_shinfo->nr_frags >= MAX_SKB_FRAGS)
5957 return false;
5958
5959 if (skb_head_is_locked(from))
5960 return false;
5961
5962 delta = from->truesize - SKB_DATA_ALIGN(sizeof(struct sk_buff));
5963
5964 page = virt_to_head_page(from->head);
5965 offset = from->data - (unsigned char *)page_address(page);
5966
5967 skb_fill_page_desc(to, to_shinfo->nr_frags,
5968 page, offset, skb_headlen(from));
5969 *fragstolen = true;
5970 } else {
5971 if (to_shinfo->nr_frags +
5972 from_shinfo->nr_frags > MAX_SKB_FRAGS)
5973 return false;
5974
5975 delta = from->truesize - SKB_TRUESIZE(skb_end_offset(from));
5976 }
5977
5978 WARN_ON_ONCE(delta < len);
5979
5980 memcpy(to_shinfo->frags + to_shinfo->nr_frags,
5981 from_shinfo->frags,
5982 from_shinfo->nr_frags * sizeof(skb_frag_t));
5983 to_shinfo->nr_frags += from_shinfo->nr_frags;
5984
5985 if (!skb_cloned(from))
5986 from_shinfo->nr_frags = 0;
5987
5988 /* if the skb is not cloned this does nothing
5989 * since we set nr_frags to 0.
5990 */
5991 if (skb_pp_frag_ref(from)) {
5992 for (i = 0; i < from_shinfo->nr_frags; i++)
5993 __skb_frag_ref(&from_shinfo->frags[i]);
5994 }
5995
5996 to->truesize += delta;
5997 to->len += len;
5998 to->data_len += len;
5999
6000 *delta_truesize = delta;
6001 return true;
6002}
6003EXPORT_SYMBOL(skb_try_coalesce);
6004
6005/**
6006 * skb_scrub_packet - scrub an skb
6007 *
6008 * @skb: buffer to clean
6009 * @xnet: packet is crossing netns
6010 *
6011 * skb_scrub_packet can be used after encapsulating or decapsulting a packet
6012 * into/from a tunnel. Some information have to be cleared during these
6013 * operations.
6014 * skb_scrub_packet can also be used to clean a skb before injecting it in
6015 * another namespace (@xnet == true). We have to clear all information in the
6016 * skb that could impact namespace isolation.
6017 */
6018void skb_scrub_packet(struct sk_buff *skb, bool xnet)
6019{
6020 skb->pkt_type = PACKET_HOST;
6021 skb->skb_iif = 0;
6022 skb->ignore_df = 0;
6023 skb_dst_drop(skb);
6024 skb_ext_reset(skb);
6025 nf_reset_ct(skb);
6026 nf_reset_trace(skb);
6027
6028#ifdef CONFIG_NET_SWITCHDEV
6029 skb->offload_fwd_mark = 0;
6030 skb->offload_l3_fwd_mark = 0;
6031#endif
6032
6033 if (!xnet)
6034 return;
6035
6036 ipvs_reset(skb);
6037 skb->mark = 0;
6038 skb_clear_tstamp(skb);
6039}
6040EXPORT_SYMBOL_GPL(skb_scrub_packet);
6041
6042static struct sk_buff *skb_reorder_vlan_header(struct sk_buff *skb)
6043{
6044 int mac_len, meta_len;
6045 void *meta;
6046
6047 if (skb_cow(skb, skb_headroom(skb)) < 0) {
6048 kfree_skb(skb);
6049 return NULL;
6050 }
6051
6052 mac_len = skb->data - skb_mac_header(skb);
6053 if (likely(mac_len > VLAN_HLEN + ETH_TLEN)) {
6054 memmove(skb_mac_header(skb) + VLAN_HLEN, skb_mac_header(skb),
6055 mac_len - VLAN_HLEN - ETH_TLEN);
6056 }
6057
6058 meta_len = skb_metadata_len(skb);
6059 if (meta_len) {
6060 meta = skb_metadata_end(skb) - meta_len;
6061 memmove(meta + VLAN_HLEN, meta, meta_len);
6062 }
6063
6064 skb->mac_header += VLAN_HLEN;
6065 return skb;
6066}
6067
6068struct sk_buff *skb_vlan_untag(struct sk_buff *skb)
6069{
6070 struct vlan_hdr *vhdr;
6071 u16 vlan_tci;
6072
6073 if (unlikely(skb_vlan_tag_present(skb))) {
6074 /* vlan_tci is already set-up so leave this for another time */
6075 return skb;
6076 }
6077
6078 skb = skb_share_check(skb, GFP_ATOMIC);
6079 if (unlikely(!skb))
6080 goto err_free;
6081 /* We may access the two bytes after vlan_hdr in vlan_set_encap_proto(). */
6082 if (unlikely(!pskb_may_pull(skb, VLAN_HLEN + sizeof(unsigned short))))
6083 goto err_free;
6084
6085 vhdr = (struct vlan_hdr *)skb->data;
6086 vlan_tci = ntohs(vhdr->h_vlan_TCI);
6087 __vlan_hwaccel_put_tag(skb, skb->protocol, vlan_tci);
6088
6089 skb_pull_rcsum(skb, VLAN_HLEN);
6090 vlan_set_encap_proto(skb, vhdr);
6091
6092 skb = skb_reorder_vlan_header(skb);
6093 if (unlikely(!skb))
6094 goto err_free;
6095
6096 skb_reset_network_header(skb);
6097 if (!skb_transport_header_was_set(skb))
6098 skb_reset_transport_header(skb);
6099 skb_reset_mac_len(skb);
6100
6101 return skb;
6102
6103err_free:
6104 kfree_skb(skb);
6105 return NULL;
6106}
6107EXPORT_SYMBOL(skb_vlan_untag);
6108
6109int skb_ensure_writable(struct sk_buff *skb, unsigned int write_len)
6110{
6111 if (!pskb_may_pull(skb, write_len))
6112 return -ENOMEM;
6113
6114 if (!skb_cloned(skb) || skb_clone_writable(skb, write_len))
6115 return 0;
6116
6117 return pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
6118}
6119EXPORT_SYMBOL(skb_ensure_writable);
6120
6121int skb_ensure_writable_head_tail(struct sk_buff *skb, struct net_device *dev)
6122{
6123 int needed_headroom = dev->needed_headroom;
6124 int needed_tailroom = dev->needed_tailroom;
6125
6126 /* For tail taggers, we need to pad short frames ourselves, to ensure
6127 * that the tail tag does not fail at its role of being at the end of
6128 * the packet, once the conduit interface pads the frame. Account for
6129 * that pad length here, and pad later.
6130 */
6131 if (unlikely(needed_tailroom && skb->len < ETH_ZLEN))
6132 needed_tailroom += ETH_ZLEN - skb->len;
6133 /* skb_headroom() returns unsigned int... */
6134 needed_headroom = max_t(int, needed_headroom - skb_headroom(skb), 0);
6135 needed_tailroom = max_t(int, needed_tailroom - skb_tailroom(skb), 0);
6136
6137 if (likely(!needed_headroom && !needed_tailroom && !skb_cloned(skb)))
6138 /* No reallocation needed, yay! */
6139 return 0;
6140
6141 return pskb_expand_head(skb, needed_headroom, needed_tailroom,
6142 GFP_ATOMIC);
6143}
6144EXPORT_SYMBOL(skb_ensure_writable_head_tail);
6145
6146/* remove VLAN header from packet and update csum accordingly.
6147 * expects a non skb_vlan_tag_present skb with a vlan tag payload
6148 */
6149int __skb_vlan_pop(struct sk_buff *skb, u16 *vlan_tci)
6150{
6151 int offset = skb->data - skb_mac_header(skb);
6152 int err;
6153
6154 if (WARN_ONCE(offset,
6155 "__skb_vlan_pop got skb with skb->data not at mac header (offset %d)\n",
6156 offset)) {
6157 return -EINVAL;
6158 }
6159
6160 err = skb_ensure_writable(skb, VLAN_ETH_HLEN);
6161 if (unlikely(err))
6162 return err;
6163
6164 skb_postpull_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
6165
6166 vlan_remove_tag(skb, vlan_tci);
6167
6168 skb->mac_header += VLAN_HLEN;
6169
6170 if (skb_network_offset(skb) < ETH_HLEN)
6171 skb_set_network_header(skb, ETH_HLEN);
6172
6173 skb_reset_mac_len(skb);
6174
6175 return err;
6176}
6177EXPORT_SYMBOL(__skb_vlan_pop);
6178
6179/* Pop a vlan tag either from hwaccel or from payload.
6180 * Expects skb->data at mac header.
6181 */
6182int skb_vlan_pop(struct sk_buff *skb)
6183{
6184 u16 vlan_tci;
6185 __be16 vlan_proto;
6186 int err;
6187
6188 if (likely(skb_vlan_tag_present(skb))) {
6189 __vlan_hwaccel_clear_tag(skb);
6190 } else {
6191 if (unlikely(!eth_type_vlan(skb->protocol)))
6192 return 0;
6193
6194 err = __skb_vlan_pop(skb, &vlan_tci);
6195 if (err)
6196 return err;
6197 }
6198 /* move next vlan tag to hw accel tag */
6199 if (likely(!eth_type_vlan(skb->protocol)))
6200 return 0;
6201
6202 vlan_proto = skb->protocol;
6203 err = __skb_vlan_pop(skb, &vlan_tci);
6204 if (unlikely(err))
6205 return err;
6206
6207 __vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
6208 return 0;
6209}
6210EXPORT_SYMBOL(skb_vlan_pop);
6211
6212/* Push a vlan tag either into hwaccel or into payload (if hwaccel tag present).
6213 * Expects skb->data at mac header.
6214 */
6215int skb_vlan_push(struct sk_buff *skb, __be16 vlan_proto, u16 vlan_tci)
6216{
6217 if (skb_vlan_tag_present(skb)) {
6218 int offset = skb->data - skb_mac_header(skb);
6219 int err;
6220
6221 if (WARN_ONCE(offset,
6222 "skb_vlan_push got skb with skb->data not at mac header (offset %d)\n",
6223 offset)) {
6224 return -EINVAL;
6225 }
6226
6227 err = __vlan_insert_tag(skb, skb->vlan_proto,
6228 skb_vlan_tag_get(skb));
6229 if (err)
6230 return err;
6231
6232 skb->protocol = skb->vlan_proto;
6233 skb->mac_len += VLAN_HLEN;
6234
6235 skb_postpush_rcsum(skb, skb->data + (2 * ETH_ALEN), VLAN_HLEN);
6236 }
6237 __vlan_hwaccel_put_tag(skb, vlan_proto, vlan_tci);
6238 return 0;
6239}
6240EXPORT_SYMBOL(skb_vlan_push);
6241
6242/**
6243 * skb_eth_pop() - Drop the Ethernet header at the head of a packet
6244 *
6245 * @skb: Socket buffer to modify
6246 *
6247 * Drop the Ethernet header of @skb.
6248 *
6249 * Expects that skb->data points to the mac header and that no VLAN tags are
6250 * present.
6251 *
6252 * Returns 0 on success, -errno otherwise.
6253 */
6254int skb_eth_pop(struct sk_buff *skb)
6255{
6256 if (!pskb_may_pull(skb, ETH_HLEN) || skb_vlan_tagged(skb) ||
6257 skb_network_offset(skb) < ETH_HLEN)
6258 return -EPROTO;
6259
6260 skb_pull_rcsum(skb, ETH_HLEN);
6261 skb_reset_mac_header(skb);
6262 skb_reset_mac_len(skb);
6263
6264 return 0;
6265}
6266EXPORT_SYMBOL(skb_eth_pop);
6267
6268/**
6269 * skb_eth_push() - Add a new Ethernet header at the head of a packet
6270 *
6271 * @skb: Socket buffer to modify
6272 * @dst: Destination MAC address of the new header
6273 * @src: Source MAC address of the new header
6274 *
6275 * Prepend @skb with a new Ethernet header.
6276 *
6277 * Expects that skb->data points to the mac header, which must be empty.
6278 *
6279 * Returns 0 on success, -errno otherwise.
6280 */
6281int skb_eth_push(struct sk_buff *skb, const unsigned char *dst,
6282 const unsigned char *src)
6283{
6284 struct ethhdr *eth;
6285 int err;
6286
6287 if (skb_network_offset(skb) || skb_vlan_tag_present(skb))
6288 return -EPROTO;
6289
6290 err = skb_cow_head(skb, sizeof(*eth));
6291 if (err < 0)
6292 return err;
6293
6294 skb_push(skb, sizeof(*eth));
6295 skb_reset_mac_header(skb);
6296 skb_reset_mac_len(skb);
6297
6298 eth = eth_hdr(skb);
6299 ether_addr_copy(eth->h_dest, dst);
6300 ether_addr_copy(eth->h_source, src);
6301 eth->h_proto = skb->protocol;
6302
6303 skb_postpush_rcsum(skb, eth, sizeof(*eth));
6304
6305 return 0;
6306}
6307EXPORT_SYMBOL(skb_eth_push);
6308
6309/* Update the ethertype of hdr and the skb csum value if required. */
6310static void skb_mod_eth_type(struct sk_buff *skb, struct ethhdr *hdr,
6311 __be16 ethertype)
6312{
6313 if (skb->ip_summed == CHECKSUM_COMPLETE) {
6314 __be16 diff[] = { ~hdr->h_proto, ethertype };
6315
6316 skb->csum = csum_partial((char *)diff, sizeof(diff), skb->csum);
6317 }
6318
6319 hdr->h_proto = ethertype;
6320}
6321
6322/**
6323 * skb_mpls_push() - push a new MPLS header after mac_len bytes from start of
6324 * the packet
6325 *
6326 * @skb: buffer
6327 * @mpls_lse: MPLS label stack entry to push
6328 * @mpls_proto: ethertype of the new MPLS header (expects 0x8847 or 0x8848)
6329 * @mac_len: length of the MAC header
6330 * @ethernet: flag to indicate if the resulting packet after skb_mpls_push is
6331 * ethernet
6332 *
6333 * Expects skb->data at mac header.
6334 *
6335 * Returns 0 on success, -errno otherwise.
6336 */
6337int skb_mpls_push(struct sk_buff *skb, __be32 mpls_lse, __be16 mpls_proto,
6338 int mac_len, bool ethernet)
6339{
6340 struct mpls_shim_hdr *lse;
6341 int err;
6342
6343 if (unlikely(!eth_p_mpls(mpls_proto)))
6344 return -EINVAL;
6345
6346 /* Networking stack does not allow simultaneous Tunnel and MPLS GSO. */
6347 if (skb->encapsulation)
6348 return -EINVAL;
6349
6350 err = skb_cow_head(skb, MPLS_HLEN);
6351 if (unlikely(err))
6352 return err;
6353
6354 if (!skb->inner_protocol) {
6355 skb_set_inner_network_header(skb, skb_network_offset(skb));
6356 skb_set_inner_protocol(skb, skb->protocol);
6357 }
6358
6359 skb_push(skb, MPLS_HLEN);
6360 memmove(skb_mac_header(skb) - MPLS_HLEN, skb_mac_header(skb),
6361 mac_len);
6362 skb_reset_mac_header(skb);
6363 skb_set_network_header(skb, mac_len);
6364 skb_reset_mac_len(skb);
6365
6366 lse = mpls_hdr(skb);
6367 lse->label_stack_entry = mpls_lse;
6368 skb_postpush_rcsum(skb, lse, MPLS_HLEN);
6369
6370 if (ethernet && mac_len >= ETH_HLEN)
6371 skb_mod_eth_type(skb, eth_hdr(skb), mpls_proto);
6372 skb->protocol = mpls_proto;
6373
6374 return 0;
6375}
6376EXPORT_SYMBOL_GPL(skb_mpls_push);
6377
6378/**
6379 * skb_mpls_pop() - pop the outermost MPLS header
6380 *
6381 * @skb: buffer
6382 * @next_proto: ethertype of header after popped MPLS header
6383 * @mac_len: length of the MAC header
6384 * @ethernet: flag to indicate if the packet is ethernet
6385 *
6386 * Expects skb->data at mac header.
6387 *
6388 * Returns 0 on success, -errno otherwise.
6389 */
6390int skb_mpls_pop(struct sk_buff *skb, __be16 next_proto, int mac_len,
6391 bool ethernet)
6392{
6393 int err;
6394
6395 if (unlikely(!eth_p_mpls(skb->protocol)))
6396 return 0;
6397
6398 err = skb_ensure_writable(skb, mac_len + MPLS_HLEN);
6399 if (unlikely(err))
6400 return err;
6401
6402 skb_postpull_rcsum(skb, mpls_hdr(skb), MPLS_HLEN);
6403 memmove(skb_mac_header(skb) + MPLS_HLEN, skb_mac_header(skb),
6404 mac_len);
6405
6406 __skb_pull(skb, MPLS_HLEN);
6407 skb_reset_mac_header(skb);
6408 skb_set_network_header(skb, mac_len);
6409
6410 if (ethernet && mac_len >= ETH_HLEN) {
6411 struct ethhdr *hdr;
6412
6413 /* use mpls_hdr() to get ethertype to account for VLANs. */
6414 hdr = (struct ethhdr *)((void *)mpls_hdr(skb) - ETH_HLEN);
6415 skb_mod_eth_type(skb, hdr, next_proto);
6416 }
6417 skb->protocol = next_proto;
6418
6419 return 0;
6420}
6421EXPORT_SYMBOL_GPL(skb_mpls_pop);
6422
6423/**
6424 * skb_mpls_update_lse() - modify outermost MPLS header and update csum
6425 *
6426 * @skb: buffer
6427 * @mpls_lse: new MPLS label stack entry to update to
6428 *
6429 * Expects skb->data at mac header.
6430 *
6431 * Returns 0 on success, -errno otherwise.
6432 */
6433int skb_mpls_update_lse(struct sk_buff *skb, __be32 mpls_lse)
6434{
6435 int err;
6436
6437 if (unlikely(!eth_p_mpls(skb->protocol)))
6438 return -EINVAL;
6439
6440 err = skb_ensure_writable(skb, skb->mac_len + MPLS_HLEN);
6441 if (unlikely(err))
6442 return err;
6443
6444 if (skb->ip_summed == CHECKSUM_COMPLETE) {
6445 __be32 diff[] = { ~mpls_hdr(skb)->label_stack_entry, mpls_lse };
6446
6447 skb->csum = csum_partial((char *)diff, sizeof(diff), skb->csum);
6448 }
6449
6450 mpls_hdr(skb)->label_stack_entry = mpls_lse;
6451
6452 return 0;
6453}
6454EXPORT_SYMBOL_GPL(skb_mpls_update_lse);
6455
6456/**
6457 * skb_mpls_dec_ttl() - decrement the TTL of the outermost MPLS header
6458 *
6459 * @skb: buffer
6460 *
6461 * Expects skb->data at mac header.
6462 *
6463 * Returns 0 on success, -errno otherwise.
6464 */
6465int skb_mpls_dec_ttl(struct sk_buff *skb)
6466{
6467 u32 lse;
6468 u8 ttl;
6469
6470 if (unlikely(!eth_p_mpls(skb->protocol)))
6471 return -EINVAL;
6472
6473 if (!pskb_may_pull(skb, skb_network_offset(skb) + MPLS_HLEN))
6474 return -ENOMEM;
6475
6476 lse = be32_to_cpu(mpls_hdr(skb)->label_stack_entry);
6477 ttl = (lse & MPLS_LS_TTL_MASK) >> MPLS_LS_TTL_SHIFT;
6478 if (!--ttl)
6479 return -EINVAL;
6480
6481 lse &= ~MPLS_LS_TTL_MASK;
6482 lse |= ttl << MPLS_LS_TTL_SHIFT;
6483
6484 return skb_mpls_update_lse(skb, cpu_to_be32(lse));
6485}
6486EXPORT_SYMBOL_GPL(skb_mpls_dec_ttl);
6487
6488/**
6489 * alloc_skb_with_frags - allocate skb with page frags
6490 *
6491 * @header_len: size of linear part
6492 * @data_len: needed length in frags
6493 * @order: max page order desired.
6494 * @errcode: pointer to error code if any
6495 * @gfp_mask: allocation mask
6496 *
6497 * This can be used to allocate a paged skb, given a maximal order for frags.
6498 */
6499struct sk_buff *alloc_skb_with_frags(unsigned long header_len,
6500 unsigned long data_len,
6501 int order,
6502 int *errcode,
6503 gfp_t gfp_mask)
6504{
6505 unsigned long chunk;
6506 struct sk_buff *skb;
6507 struct page *page;
6508 int nr_frags = 0;
6509
6510 *errcode = -EMSGSIZE;
6511 if (unlikely(data_len > MAX_SKB_FRAGS * (PAGE_SIZE << order)))
6512 return NULL;
6513
6514 *errcode = -ENOBUFS;
6515 skb = alloc_skb(header_len, gfp_mask);
6516 if (!skb)
6517 return NULL;
6518
6519 while (data_len) {
6520 if (nr_frags == MAX_SKB_FRAGS - 1)
6521 goto failure;
6522 while (order && PAGE_ALIGN(data_len) < (PAGE_SIZE << order))
6523 order--;
6524
6525 if (order) {
6526 page = alloc_pages((gfp_mask & ~__GFP_DIRECT_RECLAIM) |
6527 __GFP_COMP |
6528 __GFP_NOWARN,
6529 order);
6530 if (!page) {
6531 order--;
6532 continue;
6533 }
6534 } else {
6535 page = alloc_page(gfp_mask);
6536 if (!page)
6537 goto failure;
6538 }
6539 chunk = min_t(unsigned long, data_len,
6540 PAGE_SIZE << order);
6541 skb_fill_page_desc(skb, nr_frags, page, 0, chunk);
6542 nr_frags++;
6543 skb->truesize += (PAGE_SIZE << order);
6544 data_len -= chunk;
6545 }
6546 return skb;
6547
6548failure:
6549 kfree_skb(skb);
6550 return NULL;
6551}
6552EXPORT_SYMBOL(alloc_skb_with_frags);
6553
6554/* carve out the first off bytes from skb when off < headlen */
6555static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
6556 const int headlen, gfp_t gfp_mask)
6557{
6558 int i;
6559 unsigned int size = skb_end_offset(skb);
6560 int new_hlen = headlen - off;
6561 u8 *data;
6562
6563 if (skb_pfmemalloc(skb))
6564 gfp_mask |= __GFP_MEMALLOC;
6565
6566 data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
6567 if (!data)
6568 return -ENOMEM;
6569 size = SKB_WITH_OVERHEAD(size);
6570
6571 /* Copy real data, and all frags */
6572 skb_copy_from_linear_data_offset(skb, off, data, new_hlen);
6573 skb->len -= off;
6574
6575 memcpy((struct skb_shared_info *)(data + size),
6576 skb_shinfo(skb),
6577 offsetof(struct skb_shared_info,
6578 frags[skb_shinfo(skb)->nr_frags]));
6579 if (skb_cloned(skb)) {
6580 /* drop the old head gracefully */
6581 if (skb_orphan_frags(skb, gfp_mask)) {
6582 skb_kfree_head(data, size);
6583 return -ENOMEM;
6584 }
6585 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
6586 skb_frag_ref(skb, i);
6587 if (skb_has_frag_list(skb))
6588 skb_clone_fraglist(skb);
6589 skb_release_data(skb, SKB_CONSUMED, false);
6590 } else {
6591 /* we can reuse existing recount- all we did was
6592 * relocate values
6593 */
6594 skb_free_head(skb, false);
6595 }
6596
6597 skb->head = data;
6598 skb->data = data;
6599 skb->head_frag = 0;
6600 skb_set_end_offset(skb, size);
6601 skb_set_tail_pointer(skb, skb_headlen(skb));
6602 skb_headers_offset_update(skb, 0);
6603 skb->cloned = 0;
6604 skb->hdr_len = 0;
6605 skb->nohdr = 0;
6606 atomic_set(&skb_shinfo(skb)->dataref, 1);
6607
6608 return 0;
6609}
6610
6611static int pskb_carve(struct sk_buff *skb, const u32 off, gfp_t gfp);
6612
6613/* carve out the first eat bytes from skb's frag_list. May recurse into
6614 * pskb_carve()
6615 */
6616static int pskb_carve_frag_list(struct sk_buff *skb,
6617 struct skb_shared_info *shinfo, int eat,
6618 gfp_t gfp_mask)
6619{
6620 struct sk_buff *list = shinfo->frag_list;
6621 struct sk_buff *clone = NULL;
6622 struct sk_buff *insp = NULL;
6623
6624 do {
6625 if (!list) {
6626 pr_err("Not enough bytes to eat. Want %d\n", eat);
6627 return -EFAULT;
6628 }
6629 if (list->len <= eat) {
6630 /* Eaten as whole. */
6631 eat -= list->len;
6632 list = list->next;
6633 insp = list;
6634 } else {
6635 /* Eaten partially. */
6636 if (skb_shared(list)) {
6637 clone = skb_clone(list, gfp_mask);
6638 if (!clone)
6639 return -ENOMEM;
6640 insp = list->next;
6641 list = clone;
6642 } else {
6643 /* This may be pulled without problems. */
6644 insp = list;
6645 }
6646 if (pskb_carve(list, eat, gfp_mask) < 0) {
6647 kfree_skb(clone);
6648 return -ENOMEM;
6649 }
6650 break;
6651 }
6652 } while (eat);
6653
6654 /* Free pulled out fragments. */
6655 while ((list = shinfo->frag_list) != insp) {
6656 shinfo->frag_list = list->next;
6657 consume_skb(list);
6658 }
6659 /* And insert new clone at head. */
6660 if (clone) {
6661 clone->next = list;
6662 shinfo->frag_list = clone;
6663 }
6664 return 0;
6665}
6666
6667/* carve off first len bytes from skb. Split line (off) is in the
6668 * non-linear part of skb
6669 */
6670static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
6671 int pos, gfp_t gfp_mask)
6672{
6673 int i, k = 0;
6674 unsigned int size = skb_end_offset(skb);
6675 u8 *data;
6676 const int nfrags = skb_shinfo(skb)->nr_frags;
6677 struct skb_shared_info *shinfo;
6678
6679 if (skb_pfmemalloc(skb))
6680 gfp_mask |= __GFP_MEMALLOC;
6681
6682 data = kmalloc_reserve(&size, gfp_mask, NUMA_NO_NODE, NULL);
6683 if (!data)
6684 return -ENOMEM;
6685 size = SKB_WITH_OVERHEAD(size);
6686
6687 memcpy((struct skb_shared_info *)(data + size),
6688 skb_shinfo(skb), offsetof(struct skb_shared_info, frags[0]));
6689 if (skb_orphan_frags(skb, gfp_mask)) {
6690 skb_kfree_head(data, size);
6691 return -ENOMEM;
6692 }
6693 shinfo = (struct skb_shared_info *)(data + size);
6694 for (i = 0; i < nfrags; i++) {
6695 int fsize = skb_frag_size(&skb_shinfo(skb)->frags[i]);
6696
6697 if (pos + fsize > off) {
6698 shinfo->frags[k] = skb_shinfo(skb)->frags[i];
6699
6700 if (pos < off) {
6701 /* Split frag.
6702 * We have two variants in this case:
6703 * 1. Move all the frag to the second
6704 * part, if it is possible. F.e.
6705 * this approach is mandatory for TUX,
6706 * where splitting is expensive.
6707 * 2. Split is accurately. We make this.
6708 */
6709 skb_frag_off_add(&shinfo->frags[0], off - pos);
6710 skb_frag_size_sub(&shinfo->frags[0], off - pos);
6711 }
6712 skb_frag_ref(skb, i);
6713 k++;
6714 }
6715 pos += fsize;
6716 }
6717 shinfo->nr_frags = k;
6718 if (skb_has_frag_list(skb))
6719 skb_clone_fraglist(skb);
6720
6721 /* split line is in frag list */
6722 if (k == 0 && pskb_carve_frag_list(skb, shinfo, off - pos, gfp_mask)) {
6723 /* skb_frag_unref() is not needed here as shinfo->nr_frags = 0. */
6724 if (skb_has_frag_list(skb))
6725 kfree_skb_list(skb_shinfo(skb)->frag_list);
6726 skb_kfree_head(data, size);
6727 return -ENOMEM;
6728 }
6729 skb_release_data(skb, SKB_CONSUMED, false);
6730
6731 skb->head = data;
6732 skb->head_frag = 0;
6733 skb->data = data;
6734 skb_set_end_offset(skb, size);
6735 skb_reset_tail_pointer(skb);
6736 skb_headers_offset_update(skb, 0);
6737 skb->cloned = 0;
6738 skb->hdr_len = 0;
6739 skb->nohdr = 0;
6740 skb->len -= off;
6741 skb->data_len = skb->len;
6742 atomic_set(&skb_shinfo(skb)->dataref, 1);
6743 return 0;
6744}
6745
6746/* remove len bytes from the beginning of the skb */
6747static int pskb_carve(struct sk_buff *skb, const u32 len, gfp_t gfp)
6748{
6749 int headlen = skb_headlen(skb);
6750
6751 if (len < headlen)
6752 return pskb_carve_inside_header(skb, len, headlen, gfp);
6753 else
6754 return pskb_carve_inside_nonlinear(skb, len, headlen, gfp);
6755}
6756
6757/* Extract to_copy bytes starting at off from skb, and return this in
6758 * a new skb
6759 */
6760struct sk_buff *pskb_extract(struct sk_buff *skb, int off,
6761 int to_copy, gfp_t gfp)
6762{
6763 struct sk_buff *clone = skb_clone(skb, gfp);
6764
6765 if (!clone)
6766 return NULL;
6767
6768 if (pskb_carve(clone, off, gfp) < 0 ||
6769 pskb_trim(clone, to_copy)) {
6770 kfree_skb(clone);
6771 return NULL;
6772 }
6773 return clone;
6774}
6775EXPORT_SYMBOL(pskb_extract);
6776
6777/**
6778 * skb_condense - try to get rid of fragments/frag_list if possible
6779 * @skb: buffer
6780 *
6781 * Can be used to save memory before skb is added to a busy queue.
6782 * If packet has bytes in frags and enough tail room in skb->head,
6783 * pull all of them, so that we can free the frags right now and adjust
6784 * truesize.
6785 * Notes:
6786 * We do not reallocate skb->head thus can not fail.
6787 * Caller must re-evaluate skb->truesize if needed.
6788 */
6789void skb_condense(struct sk_buff *skb)
6790{
6791 if (skb->data_len) {
6792 if (skb->data_len > skb->end - skb->tail ||
6793 skb_cloned(skb))
6794 return;
6795
6796 /* Nice, we can free page frag(s) right now */
6797 __pskb_pull_tail(skb, skb->data_len);
6798 }
6799 /* At this point, skb->truesize might be over estimated,
6800 * because skb had a fragment, and fragments do not tell
6801 * their truesize.
6802 * When we pulled its content into skb->head, fragment
6803 * was freed, but __pskb_pull_tail() could not possibly
6804 * adjust skb->truesize, not knowing the frag truesize.
6805 */
6806 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
6807}
6808EXPORT_SYMBOL(skb_condense);
6809
6810#ifdef CONFIG_SKB_EXTENSIONS
6811static void *skb_ext_get_ptr(struct skb_ext *ext, enum skb_ext_id id)
6812{
6813 return (void *)ext + (ext->offset[id] * SKB_EXT_ALIGN_VALUE);
6814}
6815
6816/**
6817 * __skb_ext_alloc - allocate a new skb extensions storage
6818 *
6819 * @flags: See kmalloc().
6820 *
6821 * Returns the newly allocated pointer. The pointer can later attached to a
6822 * skb via __skb_ext_set().
6823 * Note: caller must handle the skb_ext as an opaque data.
6824 */
6825struct skb_ext *__skb_ext_alloc(gfp_t flags)
6826{
6827 struct skb_ext *new = kmem_cache_alloc(skbuff_ext_cache, flags);
6828
6829 if (new) {
6830 memset(new->offset, 0, sizeof(new->offset));
6831 refcount_set(&new->refcnt, 1);
6832 }
6833
6834 return new;
6835}
6836
6837static struct skb_ext *skb_ext_maybe_cow(struct skb_ext *old,
6838 unsigned int old_active)
6839{
6840 struct skb_ext *new;
6841
6842 if (refcount_read(&old->refcnt) == 1)
6843 return old;
6844
6845 new = kmem_cache_alloc(skbuff_ext_cache, GFP_ATOMIC);
6846 if (!new)
6847 return NULL;
6848
6849 memcpy(new, old, old->chunks * SKB_EXT_ALIGN_VALUE);
6850 refcount_set(&new->refcnt, 1);
6851
6852#ifdef CONFIG_XFRM
6853 if (old_active & (1 << SKB_EXT_SEC_PATH)) {
6854 struct sec_path *sp = skb_ext_get_ptr(old, SKB_EXT_SEC_PATH);
6855 unsigned int i;
6856
6857 for (i = 0; i < sp->len; i++)
6858 xfrm_state_hold(sp->xvec[i]);
6859 }
6860#endif
6861#ifdef CONFIG_MCTP_FLOWS
6862 if (old_active & (1 << SKB_EXT_MCTP)) {
6863 struct mctp_flow *flow = skb_ext_get_ptr(old, SKB_EXT_MCTP);
6864
6865 if (flow->key)
6866 refcount_inc(&flow->key->refs);
6867 }
6868#endif
6869 __skb_ext_put(old);
6870 return new;
6871}
6872
6873/**
6874 * __skb_ext_set - attach the specified extension storage to this skb
6875 * @skb: buffer
6876 * @id: extension id
6877 * @ext: extension storage previously allocated via __skb_ext_alloc()
6878 *
6879 * Existing extensions, if any, are cleared.
6880 *
6881 * Returns the pointer to the extension.
6882 */
6883void *__skb_ext_set(struct sk_buff *skb, enum skb_ext_id id,
6884 struct skb_ext *ext)
6885{
6886 unsigned int newlen, newoff = SKB_EXT_CHUNKSIZEOF(*ext);
6887
6888 skb_ext_put(skb);
6889 newlen = newoff + skb_ext_type_len[id];
6890 ext->chunks = newlen;
6891 ext->offset[id] = newoff;
6892 skb->extensions = ext;
6893 skb->active_extensions = 1 << id;
6894 return skb_ext_get_ptr(ext, id);
6895}
6896
6897/**
6898 * skb_ext_add - allocate space for given extension, COW if needed
6899 * @skb: buffer
6900 * @id: extension to allocate space for
6901 *
6902 * Allocates enough space for the given extension.
6903 * If the extension is already present, a pointer to that extension
6904 * is returned.
6905 *
6906 * If the skb was cloned, COW applies and the returned memory can be
6907 * modified without changing the extension space of clones buffers.
6908 *
6909 * Returns pointer to the extension or NULL on allocation failure.
6910 */
6911void *skb_ext_add(struct sk_buff *skb, enum skb_ext_id id)
6912{
6913 struct skb_ext *new, *old = NULL;
6914 unsigned int newlen, newoff;
6915
6916 if (skb->active_extensions) {
6917 old = skb->extensions;
6918
6919 new = skb_ext_maybe_cow(old, skb->active_extensions);
6920 if (!new)
6921 return NULL;
6922
6923 if (__skb_ext_exist(new, id))
6924 goto set_active;
6925
6926 newoff = new->chunks;
6927 } else {
6928 newoff = SKB_EXT_CHUNKSIZEOF(*new);
6929
6930 new = __skb_ext_alloc(GFP_ATOMIC);
6931 if (!new)
6932 return NULL;
6933 }
6934
6935 newlen = newoff + skb_ext_type_len[id];
6936 new->chunks = newlen;
6937 new->offset[id] = newoff;
6938set_active:
6939 skb->slow_gro = 1;
6940 skb->extensions = new;
6941 skb->active_extensions |= 1 << id;
6942 return skb_ext_get_ptr(new, id);
6943}
6944EXPORT_SYMBOL(skb_ext_add);
6945
6946#ifdef CONFIG_XFRM
6947static void skb_ext_put_sp(struct sec_path *sp)
6948{
6949 unsigned int i;
6950
6951 for (i = 0; i < sp->len; i++)
6952 xfrm_state_put(sp->xvec[i]);
6953}
6954#endif
6955
6956#ifdef CONFIG_MCTP_FLOWS
6957static void skb_ext_put_mctp(struct mctp_flow *flow)
6958{
6959 if (flow->key)
6960 mctp_key_unref(flow->key);
6961}
6962#endif
6963
6964void __skb_ext_del(struct sk_buff *skb, enum skb_ext_id id)
6965{
6966 struct skb_ext *ext = skb->extensions;
6967
6968 skb->active_extensions &= ~(1 << id);
6969 if (skb->active_extensions == 0) {
6970 skb->extensions = NULL;
6971 __skb_ext_put(ext);
6972#ifdef CONFIG_XFRM
6973 } else if (id == SKB_EXT_SEC_PATH &&
6974 refcount_read(&ext->refcnt) == 1) {
6975 struct sec_path *sp = skb_ext_get_ptr(ext, SKB_EXT_SEC_PATH);
6976
6977 skb_ext_put_sp(sp);
6978 sp->len = 0;
6979#endif
6980 }
6981}
6982EXPORT_SYMBOL(__skb_ext_del);
6983
6984void __skb_ext_put(struct skb_ext *ext)
6985{
6986 /* If this is last clone, nothing can increment
6987 * it after check passes. Avoids one atomic op.
6988 */
6989 if (refcount_read(&ext->refcnt) == 1)
6990 goto free_now;
6991
6992 if (!refcount_dec_and_test(&ext->refcnt))
6993 return;
6994free_now:
6995#ifdef CONFIG_XFRM
6996 if (__skb_ext_exist(ext, SKB_EXT_SEC_PATH))
6997 skb_ext_put_sp(skb_ext_get_ptr(ext, SKB_EXT_SEC_PATH));
6998#endif
6999#ifdef CONFIG_MCTP_FLOWS
7000 if (__skb_ext_exist(ext, SKB_EXT_MCTP))
7001 skb_ext_put_mctp(skb_ext_get_ptr(ext, SKB_EXT_MCTP));
7002#endif
7003
7004 kmem_cache_free(skbuff_ext_cache, ext);
7005}
7006EXPORT_SYMBOL(__skb_ext_put);
7007#endif /* CONFIG_SKB_EXTENSIONS */
7008
7009/**
7010 * skb_attempt_defer_free - queue skb for remote freeing
7011 * @skb: buffer
7012 *
7013 * Put @skb in a per-cpu list, using the cpu which
7014 * allocated the skb/pages to reduce false sharing
7015 * and memory zone spinlock contention.
7016 */
7017void skb_attempt_defer_free(struct sk_buff *skb)
7018{
7019 int cpu = skb->alloc_cpu;
7020 struct softnet_data *sd;
7021 unsigned int defer_max;
7022 bool kick;
7023
7024 if (WARN_ON_ONCE(cpu >= nr_cpu_ids) ||
7025 !cpu_online(cpu) ||
7026 cpu == raw_smp_processor_id()) {
7027nodefer: __kfree_skb(skb);
7028 return;
7029 }
7030
7031 DEBUG_NET_WARN_ON_ONCE(skb_dst(skb));
7032 DEBUG_NET_WARN_ON_ONCE(skb->destructor);
7033
7034 sd = &per_cpu(softnet_data, cpu);
7035 defer_max = READ_ONCE(sysctl_skb_defer_max);
7036 if (READ_ONCE(sd->defer_count) >= defer_max)
7037 goto nodefer;
7038
7039 spin_lock_bh(&sd->defer_lock);
7040 /* Send an IPI every time queue reaches half capacity. */
7041 kick = sd->defer_count == (defer_max >> 1);
7042 /* Paired with the READ_ONCE() few lines above */
7043 WRITE_ONCE(sd->defer_count, sd->defer_count + 1);
7044
7045 skb->next = sd->defer_list;
7046 /* Paired with READ_ONCE() in skb_defer_free_flush() */
7047 WRITE_ONCE(sd->defer_list, skb);
7048 spin_unlock_bh(&sd->defer_lock);
7049
7050 /* Make sure to trigger NET_RX_SOFTIRQ on the remote CPU
7051 * if we are unlucky enough (this seems very unlikely).
7052 */
7053 if (unlikely(kick) && !cmpxchg(&sd->defer_ipi_scheduled, 0, 1))
7054 smp_call_function_single_async(cpu, &sd->defer_csd);
7055}
7056
7057static void skb_splice_csum_page(struct sk_buff *skb, struct page *page,
7058 size_t offset, size_t len)
7059{
7060 const char *kaddr;
7061 __wsum csum;
7062
7063 kaddr = kmap_local_page(page);
7064 csum = csum_partial(kaddr + offset, len, 0);
7065 kunmap_local(kaddr);
7066 skb->csum = csum_block_add(skb->csum, csum, skb->len);
7067}
7068
7069/**
7070 * skb_splice_from_iter - Splice (or copy) pages to skbuff
7071 * @skb: The buffer to add pages to
7072 * @iter: Iterator representing the pages to be added
7073 * @maxsize: Maximum amount of pages to be added
7074 * @gfp: Allocation flags
7075 *
7076 * This is a common helper function for supporting MSG_SPLICE_PAGES. It
7077 * extracts pages from an iterator and adds them to the socket buffer if
7078 * possible, copying them to fragments if not possible (such as if they're slab
7079 * pages).
7080 *
7081 * Returns the amount of data spliced/copied or -EMSGSIZE if there's
7082 * insufficient space in the buffer to transfer anything.
7083 */
7084ssize_t skb_splice_from_iter(struct sk_buff *skb, struct iov_iter *iter,
7085 ssize_t maxsize, gfp_t gfp)
7086{
7087 size_t frag_limit = READ_ONCE(sysctl_max_skb_frags);
7088 struct page *pages[8], **ppages = pages;
7089 ssize_t spliced = 0, ret = 0;
7090 unsigned int i;
7091
7092 while (iter->count > 0) {
7093 ssize_t space, nr, len;
7094 size_t off;
7095
7096 ret = -EMSGSIZE;
7097 space = frag_limit - skb_shinfo(skb)->nr_frags;
7098 if (space < 0)
7099 break;
7100
7101 /* We might be able to coalesce without increasing nr_frags */
7102 nr = clamp_t(size_t, space, 1, ARRAY_SIZE(pages));
7103
7104 len = iov_iter_extract_pages(iter, &ppages, maxsize, nr, 0, &off);
7105 if (len <= 0) {
7106 ret = len ?: -EIO;
7107 break;
7108 }
7109
7110 i = 0;
7111 do {
7112 struct page *page = pages[i++];
7113 size_t part = min_t(size_t, PAGE_SIZE - off, len);
7114
7115 ret = -EIO;
7116 if (WARN_ON_ONCE(!sendpage_ok(page)))
7117 goto out;
7118
7119 ret = skb_append_pagefrags(skb, page, off, part,
7120 frag_limit);
7121 if (ret < 0) {
7122 iov_iter_revert(iter, len);
7123 goto out;
7124 }
7125
7126 if (skb->ip_summed == CHECKSUM_NONE)
7127 skb_splice_csum_page(skb, page, off, part);
7128
7129 off = 0;
7130 spliced += part;
7131 maxsize -= part;
7132 len -= part;
7133 } while (len > 0);
7134
7135 if (maxsize <= 0)
7136 break;
7137 }
7138
7139out:
7140 skb_len_add(skb, spliced);
7141 return spliced ?: ret;
7142}
7143EXPORT_SYMBOL(skb_splice_from_iter);
7144
7145static __always_inline
7146size_t memcpy_from_iter_csum(void *iter_from, size_t progress,
7147 size_t len, void *to, void *priv2)
7148{
7149 __wsum *csum = priv2;
7150 __wsum next = csum_partial_copy_nocheck(iter_from, to + progress, len);
7151
7152 *csum = csum_block_add(*csum, next, progress);
7153 return 0;
7154}
7155
7156static __always_inline
7157size_t copy_from_user_iter_csum(void __user *iter_from, size_t progress,
7158 size_t len, void *to, void *priv2)
7159{
7160 __wsum next, *csum = priv2;
7161
7162 next = csum_and_copy_from_user(iter_from, to + progress, len);
7163 *csum = csum_block_add(*csum, next, progress);
7164 return next ? 0 : len;
7165}
7166
7167bool csum_and_copy_from_iter_full(void *addr, size_t bytes,
7168 __wsum *csum, struct iov_iter *i)
7169{
7170 size_t copied;
7171
7172 if (WARN_ON_ONCE(!i->data_source))
7173 return false;
7174 copied = iterate_and_advance2(i, bytes, addr, csum,
7175 copy_from_user_iter_csum,
7176 memcpy_from_iter_csum);
7177 if (likely(copied == bytes))
7178 return true;
7179 iov_iter_revert(i, copied);
7180 return false;
7181}
7182EXPORT_SYMBOL(csum_and_copy_from_iter_full);