1 // SPDX-License-Identifier: GPL-2.0
5 * Generic datagram handling routines. These are generic for all
6 * protocols. Possibly a generic IP version on top of these would
7 * make sense. Not tonight however 8-).
8 * This is used because UDP, RAW, PACKET, DDP, IPX, AX.25 and
9 * NetROM layer all have identical poll code and mostly
10 * identical recvmsg() code. So we share it here. The poll was
11 * shared before but buried in udp.c so I moved it.
13 * Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>. (datagram_poll() from old
17 * Alan Cox : NULL return from skb_peek_copy()
19 * Alan Cox : Rewrote skb_read_datagram to avoid the
20 * skb_peek_copy stuff.
21 * Alan Cox : Added support for SOCK_SEQPACKET.
22 * IPX can no longer use the SO_TYPE hack
23 * but AX.25 now works right, and SPX is
25 * Alan Cox : Fixed write poll of non IP protocol
27 * Florian La Roche: Changed for my new skbuff handling.
28 * Darryl Miles : Fixed non-blocking SOCK_SEQPACKET.
29 * Linus Torvalds : BSD semantic fixes.
30 * Alan Cox : Datagram iovec handling
31 * Darryl Miles : Fixed non-blocking SOCK_STREAM.
32 * Alan Cox : POSIXisms
33 * Pete Wyckoff : Unconnected accept() fix.
37 #include <linux/module.h>
38 #include <linux/types.h>
39 #include <linux/kernel.h>
40 #include <linux/uaccess.h>
42 #include <linux/interrupt.h>
43 #include <linux/errno.h>
44 #include <linux/sched.h>
45 #include <linux/inet.h>
46 #include <linux/netdevice.h>
47 #include <linux/rtnetlink.h>
48 #include <linux/poll.h>
49 #include <linux/highmem.h>
50 #include <linux/spinlock.h>
51 #include <linux/slab.h>
52 #include <linux/pagemap.h>
53 #include <linux/uio.h>
54 #include <linux/indirect_call_wrapper.h>
56 #include <net/protocol.h>
57 #include <linux/skbuff.h>
59 #include <net/checksum.h>
61 #include <net/tcp_states.h>
62 #include <trace/events/skb.h>
63 #include <net/busy_poll.h>
68 * Is a socket 'connection oriented' ?
70 static inline int connection_based(struct sock *sk)
72 return sk->sk_type == SOCK_SEQPACKET || sk->sk_type == SOCK_STREAM;
75 static int receiver_wake_function(wait_queue_entry_t *wait, unsigned int mode, int sync,
79 * Avoid a wakeup if event not interesting for us
81 if (key && !(key_to_poll(key) & (EPOLLIN | EPOLLERR)))
83 return autoremove_wake_function(wait, mode, sync, key);
86 * Wait for the last received packet to be different from skb
88 int __skb_wait_for_more_packets(struct sock *sk, struct sk_buff_head *queue,
89 int *err, long *timeo_p,
90 const struct sk_buff *skb)
93 DEFINE_WAIT_FUNC(wait, receiver_wake_function);
95 prepare_to_wait_exclusive(sk_sleep(sk), &wait, TASK_INTERRUPTIBLE);
98 error = sock_error(sk);
102 if (READ_ONCE(queue->prev) != skb)
105 /* Socket shut down? */
106 if (sk->sk_shutdown & RCV_SHUTDOWN)
109 /* Sequenced packets can come disconnected.
110 * If so we report the problem
113 if (connection_based(sk) &&
114 !(sk->sk_state == TCP_ESTABLISHED || sk->sk_state == TCP_LISTEN))
118 if (signal_pending(current))
122 *timeo_p = schedule_timeout(*timeo_p);
124 finish_wait(sk_sleep(sk), &wait);
127 error = sock_intr_errno(*timeo_p);
136 EXPORT_SYMBOL(__skb_wait_for_more_packets);
138 static struct sk_buff *skb_set_peeked(struct sk_buff *skb)
140 struct sk_buff *nskb;
145 /* We have to unshare an skb before modifying it. */
146 if (!skb_shared(skb))
149 nskb = skb_clone(skb, GFP_ATOMIC);
151 return ERR_PTR(-ENOMEM);
153 skb->prev->next = nskb;
154 skb->next->prev = nskb;
155 nskb->prev = skb->prev;
156 nskb->next = skb->next;
167 struct sk_buff *__skb_try_recv_from_queue(struct sock *sk,
168 struct sk_buff_head *queue,
171 struct sk_buff **last)
173 bool peek_at_off = false;
177 if (unlikely(flags & MSG_PEEK && *off >= 0)) {
183 skb_queue_walk(queue, skb) {
184 if (flags & MSG_PEEK) {
185 if (peek_at_off && _off >= skb->len &&
186 (_off || skb->peeked)) {
191 skb = skb_set_peeked(skb);
197 refcount_inc(&skb->users);
199 __skb_unlink(skb, queue);
208 * __skb_try_recv_datagram - Receive a datagram skbuff
210 * @queue: socket queue from which to receive
211 * @flags: MSG\_ flags
212 * @off: an offset in bytes to peek skb from. Returns an offset
213 * within an skb where data actually starts
214 * @err: error code returned
215 * @last: set to last peeked message to inform the wait function
216 * what to look for when peeking
218 * Get a datagram skbuff, understands the peeking, nonblocking wakeups
219 * and possible races. This replaces identical code in packet, raw and
220 * udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
221 * the long standing peek and read race for datagram sockets. If you
222 * alter this routine remember it must be re-entrant.
224 * This function will lock the socket if a skb is returned, so
225 * the caller needs to unlock the socket in that case (usually by
226 * calling skb_free_datagram). Returns NULL with @err set to
227 * -EAGAIN if no data was available or to some other value if an
228 * error was detected.
230 * * It does not lock socket since today. This function is
231 * * free of race conditions. This measure should/can improve
232 * * significantly datagram socket latencies at high loads,
233 * * when data copying to user space takes lots of time.
234 * * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
238 * The order of the tests when we find no data waiting are specified
239 * quite explicitly by POSIX 1003.1g, don't change them without having
240 * the standard around please.
242 struct sk_buff *__skb_try_recv_datagram(struct sock *sk,
243 struct sk_buff_head *queue,
244 unsigned int flags, int *off, int *err,
245 struct sk_buff **last)
248 unsigned long cpu_flags;
250 * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
252 int error = sock_error(sk);
258 /* Again only user level code calls this function, so nothing
259 * interrupt level will suddenly eat the receive_queue.
261 * Look at current nfs client by the way...
262 * However, this function was correct in any case. 8)
264 spin_lock_irqsave(&queue->lock, cpu_flags);
265 skb = __skb_try_recv_from_queue(sk, queue, flags, off, &error,
267 spin_unlock_irqrestore(&queue->lock, cpu_flags);
273 if (!sk_can_busy_loop(sk))
276 sk_busy_loop(sk, flags & MSG_DONTWAIT);
277 } while (READ_ONCE(queue->prev) != *last);
285 EXPORT_SYMBOL(__skb_try_recv_datagram);
287 struct sk_buff *__skb_recv_datagram(struct sock *sk,
288 struct sk_buff_head *sk_queue,
289 unsigned int flags, int *off, int *err)
291 struct sk_buff *skb, *last;
294 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
297 skb = __skb_try_recv_datagram(sk, sk_queue, flags, off, err,
305 !__skb_wait_for_more_packets(sk, sk_queue, err,
310 EXPORT_SYMBOL(__skb_recv_datagram);
312 struct sk_buff *skb_recv_datagram(struct sock *sk, unsigned int flags,
313 int noblock, int *err)
317 return __skb_recv_datagram(sk, &sk->sk_receive_queue,
318 flags | (noblock ? MSG_DONTWAIT : 0),
321 EXPORT_SYMBOL(skb_recv_datagram);
323 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
326 sk_mem_reclaim_partial(sk);
328 EXPORT_SYMBOL(skb_free_datagram);
330 void __skb_free_datagram_locked(struct sock *sk, struct sk_buff *skb, int len)
334 if (!skb_unref(skb)) {
335 sk_peek_offset_bwd(sk, len);
339 slow = lock_sock_fast(sk);
340 sk_peek_offset_bwd(sk, len);
342 sk_mem_reclaim_partial(sk);
343 unlock_sock_fast(sk, slow);
345 /* skb is now orphaned, can be freed outside of locked section */
348 EXPORT_SYMBOL(__skb_free_datagram_locked);
350 int __sk_queue_drop_skb(struct sock *sk, struct sk_buff_head *sk_queue,
351 struct sk_buff *skb, unsigned int flags,
352 void (*destructor)(struct sock *sk,
353 struct sk_buff *skb))
357 if (flags & MSG_PEEK) {
359 spin_lock_bh(&sk_queue->lock);
361 __skb_unlink(skb, sk_queue);
362 refcount_dec(&skb->users);
367 spin_unlock_bh(&sk_queue->lock);
370 atomic_inc(&sk->sk_drops);
373 EXPORT_SYMBOL(__sk_queue_drop_skb);
376 * skb_kill_datagram - Free a datagram skbuff forcibly
378 * @skb: datagram skbuff
379 * @flags: MSG\_ flags
381 * This function frees a datagram skbuff that was received by
382 * skb_recv_datagram. The flags argument must match the one
383 * used for skb_recv_datagram.
385 * If the MSG_PEEK flag is set, and the packet is still on the
386 * receive queue of the socket, it will be taken off the queue
387 * before it is freed.
389 * This function currently only disables BH when acquiring the
390 * sk_receive_queue lock. Therefore it must not be used in a
391 * context where that lock is acquired in an IRQ context.
393 * It returns 0 if the packet was removed by us.
396 int skb_kill_datagram(struct sock *sk, struct sk_buff *skb, unsigned int flags)
398 int err = __sk_queue_drop_skb(sk, &sk->sk_receive_queue, skb, flags,
402 sk_mem_reclaim_partial(sk);
405 EXPORT_SYMBOL(skb_kill_datagram);
407 INDIRECT_CALLABLE_DECLARE(static size_t simple_copy_to_iter(const void *addr,
409 void *data __always_unused,
410 struct iov_iter *i));
412 static int __skb_datagram_iter(const struct sk_buff *skb, int offset,
413 struct iov_iter *to, int len, bool fault_short,
414 size_t (*cb)(const void *, size_t, void *,
415 struct iov_iter *), void *data)
417 int start = skb_headlen(skb);
418 int i, copy = start - offset, start_off = offset, n;
419 struct sk_buff *frag_iter;
425 n = INDIRECT_CALL_1(cb, simple_copy_to_iter,
426 skb->data + offset, copy, data, to);
430 if ((len -= copy) == 0)
434 /* Copy paged appendix. Hmm... why does this look so complicated? */
435 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
437 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
439 WARN_ON(start > offset + len);
441 end = start + skb_frag_size(frag);
442 if ((copy = end - offset) > 0) {
443 struct page *page = skb_frag_page(frag);
444 u8 *vaddr = kmap(page);
448 n = INDIRECT_CALL_1(cb, simple_copy_to_iter,
449 vaddr + skb_frag_off(frag) + offset - start,
461 skb_walk_frags(skb, frag_iter) {
464 WARN_ON(start > offset + len);
466 end = start + frag_iter->len;
467 if ((copy = end - offset) > 0) {
470 if (__skb_datagram_iter(frag_iter, offset - start,
471 to, copy, fault_short, cb, data))
473 if ((len -= copy) == 0)
482 /* This is not really a user copy fault, but rather someone
483 * gave us a bogus length on the skb. We should probably
484 * print a warning here as it may indicate a kernel bug.
488 iov_iter_revert(to, offset - start_off);
492 if (fault_short || iov_iter_count(to))
499 * skb_copy_and_hash_datagram_iter - Copy datagram to an iovec iterator
501 * @skb: buffer to copy
502 * @offset: offset in the buffer to start copying from
503 * @to: iovec iterator to copy to
504 * @len: amount of data to copy from buffer to iovec
505 * @hash: hash request to update
507 int skb_copy_and_hash_datagram_iter(const struct sk_buff *skb, int offset,
508 struct iov_iter *to, int len,
509 struct ahash_request *hash)
511 return __skb_datagram_iter(skb, offset, to, len, true,
512 hash_and_copy_to_iter, hash);
514 EXPORT_SYMBOL(skb_copy_and_hash_datagram_iter);
516 static size_t simple_copy_to_iter(const void *addr, size_t bytes,
517 void *data __always_unused, struct iov_iter *i)
519 return copy_to_iter(addr, bytes, i);
523 * skb_copy_datagram_iter - Copy a datagram to an iovec iterator.
524 * @skb: buffer to copy
525 * @offset: offset in the buffer to start copying from
526 * @to: iovec iterator to copy to
527 * @len: amount of data to copy from buffer to iovec
529 int skb_copy_datagram_iter(const struct sk_buff *skb, int offset,
530 struct iov_iter *to, int len)
532 trace_skb_copy_datagram_iovec(skb, len);
533 return __skb_datagram_iter(skb, offset, to, len, false,
534 simple_copy_to_iter, NULL);
536 EXPORT_SYMBOL(skb_copy_datagram_iter);
539 * skb_copy_datagram_from_iter - Copy a datagram from an iov_iter.
540 * @skb: buffer to copy
541 * @offset: offset in the buffer to start copying to
542 * @from: the copy source
543 * @len: amount of data to copy to buffer from iovec
545 * Returns 0 or -EFAULT.
547 int skb_copy_datagram_from_iter(struct sk_buff *skb, int offset,
548 struct iov_iter *from,
551 int start = skb_headlen(skb);
552 int i, copy = start - offset;
553 struct sk_buff *frag_iter;
559 if (copy_from_iter(skb->data + offset, copy, from) != copy)
561 if ((len -= copy) == 0)
566 /* Copy paged appendix. Hmm... why does this look so complicated? */
567 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
569 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
571 WARN_ON(start > offset + len);
573 end = start + skb_frag_size(frag);
574 if ((copy = end - offset) > 0) {
579 copied = copy_page_from_iter(skb_frag_page(frag),
580 skb_frag_off(frag) + offset - start,
592 skb_walk_frags(skb, frag_iter) {
595 WARN_ON(start > offset + len);
597 end = start + frag_iter->len;
598 if ((copy = end - offset) > 0) {
601 if (skb_copy_datagram_from_iter(frag_iter,
605 if ((len -= copy) == 0)
617 EXPORT_SYMBOL(skb_copy_datagram_from_iter);
619 int __zerocopy_sg_from_iter(struct sock *sk, struct sk_buff *skb,
620 struct iov_iter *from, size_t length)
622 int frag = skb_shinfo(skb)->nr_frags;
624 while (length && iov_iter_count(from)) {
625 struct page *pages[MAX_SKB_FRAGS];
626 struct page *last_head = NULL;
629 unsigned long truesize;
632 if (frag == MAX_SKB_FRAGS)
635 copied = iov_iter_get_pages(from, pages, length,
636 MAX_SKB_FRAGS - frag, &start);
640 iov_iter_advance(from, copied);
643 truesize = PAGE_ALIGN(copied + start);
644 skb->data_len += copied;
646 skb->truesize += truesize;
647 if (sk && sk->sk_type == SOCK_STREAM) {
648 sk_wmem_queued_add(sk, truesize);
649 sk_mem_charge(sk, truesize);
651 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
653 for (refs = 0; copied != 0; start = 0) {
654 int size = min_t(int, copied, PAGE_SIZE - start);
655 struct page *head = compound_head(pages[n]);
657 start += (pages[n] - head) << PAGE_SHIFT;
661 skb_frag_t *last = &skb_shinfo(skb)->frags[frag - 1];
663 if (head == skb_frag_page(last) &&
664 start == skb_frag_off(last) + skb_frag_size(last)) {
665 skb_frag_size_add(last, size);
666 /* We combined this page, we need to release
667 * a reference. Since compound pages refcount
668 * is shared among many pages, batch the refcount
669 * adjustments to limit false sharing.
677 page_ref_sub(last_head, refs);
680 skb_fill_page_desc(skb, frag++, head, start, size);
683 page_ref_sub(last_head, refs);
687 EXPORT_SYMBOL(__zerocopy_sg_from_iter);
690 * zerocopy_sg_from_iter - Build a zerocopy datagram from an iov_iter
691 * @skb: buffer to copy
692 * @from: the source to copy from
694 * The function will first copy up to headlen, and then pin the userspace
695 * pages and build frags through them.
697 * Returns 0, -EFAULT or -EMSGSIZE.
699 int zerocopy_sg_from_iter(struct sk_buff *skb, struct iov_iter *from)
701 int copy = min_t(int, skb_headlen(skb), iov_iter_count(from));
703 /* copy up to skb headlen */
704 if (skb_copy_datagram_from_iter(skb, 0, from, copy))
707 return __zerocopy_sg_from_iter(NULL, skb, from, ~0U);
709 EXPORT_SYMBOL(zerocopy_sg_from_iter);
712 * skb_copy_and_csum_datagram - Copy datagram to an iovec iterator
713 * and update a checksum.
714 * @skb: buffer to copy
715 * @offset: offset in the buffer to start copying from
716 * @to: iovec iterator to copy to
717 * @len: amount of data to copy from buffer to iovec
718 * @csump: checksum pointer
720 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
721 struct iov_iter *to, int len,
724 struct csum_state csdata = { .csum = *csump };
727 ret = __skb_datagram_iter(skb, offset, to, len, true,
728 csum_and_copy_to_iter, &csdata);
732 *csump = csdata.csum;
737 * skb_copy_and_csum_datagram_msg - Copy and checksum skb to user iovec.
739 * @hlen: hardware length
742 * Caller _must_ check that skb will fit to this iovec.
744 * Returns: 0 - success.
745 * -EINVAL - checksum failure.
746 * -EFAULT - fault during copy.
748 int skb_copy_and_csum_datagram_msg(struct sk_buff *skb,
749 int hlen, struct msghdr *msg)
752 int chunk = skb->len - hlen;
757 if (msg_data_left(msg) < chunk) {
758 if (__skb_checksum_complete(skb))
760 if (skb_copy_datagram_msg(skb, hlen, msg, chunk))
763 csum = csum_partial(skb->data, hlen, skb->csum);
764 if (skb_copy_and_csum_datagram(skb, hlen, &msg->msg_iter,
768 if (csum_fold(csum)) {
769 iov_iter_revert(&msg->msg_iter, chunk);
773 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
774 !skb->csum_complete_sw)
775 netdev_rx_csum_fault(NULL, skb);
781 EXPORT_SYMBOL(skb_copy_and_csum_datagram_msg);
784 * datagram_poll - generic datagram poll
789 * Datagram poll: Again totally generic. This also handles
790 * sequenced packet sockets providing the socket receive queue
791 * is only ever holding data ready to receive.
793 * Note: when you *don't* use this routine for this protocol,
794 * and you use a different write policy from sock_writeable()
795 * then please supply your own write_space callback.
797 __poll_t datagram_poll(struct file *file, struct socket *sock,
800 struct sock *sk = sock->sk;
803 sock_poll_wait(file, sock, wait);
806 /* exceptional events? */
807 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
809 (sock_flag(sk, SOCK_SELECT_ERR_QUEUE) ? EPOLLPRI : 0);
811 if (sk->sk_shutdown & RCV_SHUTDOWN)
812 mask |= EPOLLRDHUP | EPOLLIN | EPOLLRDNORM;
813 if (sk->sk_shutdown == SHUTDOWN_MASK)
817 if (!skb_queue_empty_lockless(&sk->sk_receive_queue))
818 mask |= EPOLLIN | EPOLLRDNORM;
820 /* Connection-based need to check for termination and startup */
821 if (connection_based(sk)) {
822 if (sk->sk_state == TCP_CLOSE)
824 /* connection hasn't started yet? */
825 if (sk->sk_state == TCP_SYN_SENT)
830 if (sock_writeable(sk))
831 mask |= EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND;
833 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
837 EXPORT_SYMBOL(datagram_poll);