4 * Generic datagram handling routines. These are generic for all
5 * protocols. Possibly a generic IP version on top of these would
6 * make sense. Not tonight however 8-).
7 * This is used because UDP, RAW, PACKET, DDP, IPX, AX.25 and
8 * NetROM layer all have identical poll code and mostly
9 * identical recvmsg() code. So we share it here. The poll was
10 * shared before but buried in udp.c so I moved it.
12 * Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>. (datagram_poll() from old
16 * Alan Cox : NULL return from skb_peek_copy()
18 * Alan Cox : Rewrote skb_read_datagram to avoid the
19 * skb_peek_copy stuff.
20 * Alan Cox : Added support for SOCK_SEQPACKET.
21 * IPX can no longer use the SO_TYPE hack
22 * but AX.25 now works right, and SPX is
24 * Alan Cox : Fixed write poll of non IP protocol
26 * Florian La Roche: Changed for my new skbuff handling.
27 * Darryl Miles : Fixed non-blocking SOCK_SEQPACKET.
28 * Linus Torvalds : BSD semantic fixes.
29 * Alan Cox : Datagram iovec handling
30 * Darryl Miles : Fixed non-blocking SOCK_STREAM.
31 * Alan Cox : POSIXisms
32 * Pete Wyckoff : Unconnected accept() fix.
36 #include <linux/module.h>
37 #include <linux/types.h>
38 #include <linux/kernel.h>
39 #include <asm/uaccess.h>
41 #include <linux/interrupt.h>
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/inet.h>
45 #include <linux/netdevice.h>
46 #include <linux/rtnetlink.h>
47 #include <linux/poll.h>
48 #include <linux/highmem.h>
49 #include <linux/spinlock.h>
50 #include <linux/slab.h>
51 #include <linux/pagemap.h>
52 #include <linux/uio.h>
54 #include <net/protocol.h>
55 #include <linux/skbuff.h>
57 #include <net/checksum.h>
59 #include <net/tcp_states.h>
60 #include <trace/events/skb.h>
61 #include <net/busy_poll.h>
64 * Is a socket 'connection oriented' ?
66 static inline int connection_based(struct sock *sk)
68 return sk->sk_type == SOCK_SEQPACKET || sk->sk_type == SOCK_STREAM;
71 static int receiver_wake_function(wait_queue_t *wait, unsigned int mode, int sync,
74 unsigned long bits = (unsigned long)key;
77 * Avoid a wakeup if event not interesting for us
79 if (bits && !(bits & (POLLIN | POLLERR)))
81 return autoremove_wake_function(wait, mode, sync, key);
84 * Wait for the last received packet to be different from skb
86 static int wait_for_more_packets(struct sock *sk, int *err, long *timeo_p,
87 const struct sk_buff *skb)
90 DEFINE_WAIT_FUNC(wait, receiver_wake_function);
92 prepare_to_wait_exclusive(sk_sleep(sk), &wait, TASK_INTERRUPTIBLE);
95 error = sock_error(sk);
99 if (sk->sk_receive_queue.prev != skb)
102 /* Socket shut down? */
103 if (sk->sk_shutdown & RCV_SHUTDOWN)
106 /* Sequenced packets can come disconnected.
107 * If so we report the problem
110 if (connection_based(sk) &&
111 !(sk->sk_state == TCP_ESTABLISHED || sk->sk_state == TCP_LISTEN))
115 if (signal_pending(current))
119 *timeo_p = schedule_timeout(*timeo_p);
121 finish_wait(sk_sleep(sk), &wait);
124 error = sock_intr_errno(*timeo_p);
135 * __skb_recv_datagram - Receive a datagram skbuff
138 * @peeked: returns non-zero if this packet has been seen before
139 * @off: an offset in bytes to peek skb from. Returns an offset
140 * within an skb where data actually starts
141 * @err: error code returned
143 * Get a datagram skbuff, understands the peeking, nonblocking wakeups
144 * and possible races. This replaces identical code in packet, raw and
145 * udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
146 * the long standing peek and read race for datagram sockets. If you
147 * alter this routine remember it must be re-entrant.
149 * This function will lock the socket if a skb is returned, so the caller
150 * needs to unlock the socket in that case (usually by calling
153 * * It does not lock socket since today. This function is
154 * * free of race conditions. This measure should/can improve
155 * * significantly datagram socket latencies at high loads,
156 * * when data copying to user space takes lots of time.
157 * * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
161 * The order of the tests when we find no data waiting are specified
162 * quite explicitly by POSIX 1003.1g, don't change them without having
163 * the standard around please.
165 struct sk_buff *__skb_recv_datagram(struct sock *sk, unsigned int flags,
166 int *peeked, int *off, int *err)
168 struct sk_buff *skb, *last;
171 * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
173 int error = sock_error(sk);
178 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
181 /* Again only user level code calls this function, so nothing
182 * interrupt level will suddenly eat the receive_queue.
184 * Look at current nfs client by the way...
185 * However, this function was correct in any case. 8)
187 unsigned long cpu_flags;
188 struct sk_buff_head *queue = &sk->sk_receive_queue;
191 last = (struct sk_buff *)queue;
192 spin_lock_irqsave(&queue->lock, cpu_flags);
193 skb_queue_walk(queue, skb) {
195 *peeked = skb->peeked;
196 if (flags & MSG_PEEK) {
197 if (_off >= skb->len && (skb->len || _off ||
203 atomic_inc(&skb->users);
205 __skb_unlink(skb, queue);
207 spin_unlock_irqrestore(&queue->lock, cpu_flags);
211 spin_unlock_irqrestore(&queue->lock, cpu_flags);
213 if (sk_can_busy_loop(sk) &&
214 sk_busy_loop(sk, flags & MSG_DONTWAIT))
217 /* User doesn't want to wait */
222 } while (!wait_for_more_packets(sk, err, &timeo, last));
230 EXPORT_SYMBOL(__skb_recv_datagram);
232 struct sk_buff *skb_recv_datagram(struct sock *sk, unsigned int flags,
233 int noblock, int *err)
237 return __skb_recv_datagram(sk, flags | (noblock ? MSG_DONTWAIT : 0),
240 EXPORT_SYMBOL(skb_recv_datagram);
242 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
245 sk_mem_reclaim_partial(sk);
247 EXPORT_SYMBOL(skb_free_datagram);
249 void skb_free_datagram_locked(struct sock *sk, struct sk_buff *skb)
253 if (likely(atomic_read(&skb->users) == 1))
255 else if (likely(!atomic_dec_and_test(&skb->users)))
258 slow = lock_sock_fast(sk);
260 sk_mem_reclaim_partial(sk);
261 unlock_sock_fast(sk, slow);
263 /* skb is now orphaned, can be freed outside of locked section */
266 EXPORT_SYMBOL(skb_free_datagram_locked);
269 * skb_kill_datagram - Free a datagram skbuff forcibly
271 * @skb: datagram skbuff
274 * This function frees a datagram skbuff that was received by
275 * skb_recv_datagram. The flags argument must match the one
276 * used for skb_recv_datagram.
278 * If the MSG_PEEK flag is set, and the packet is still on the
279 * receive queue of the socket, it will be taken off the queue
280 * before it is freed.
282 * This function currently only disables BH when acquiring the
283 * sk_receive_queue lock. Therefore it must not be used in a
284 * context where that lock is acquired in an IRQ context.
286 * It returns 0 if the packet was removed by us.
289 int skb_kill_datagram(struct sock *sk, struct sk_buff *skb, unsigned int flags)
293 if (flags & MSG_PEEK) {
295 spin_lock_bh(&sk->sk_receive_queue.lock);
296 if (skb == skb_peek(&sk->sk_receive_queue)) {
297 __skb_unlink(skb, &sk->sk_receive_queue);
298 atomic_dec(&skb->users);
301 spin_unlock_bh(&sk->sk_receive_queue.lock);
305 atomic_inc(&sk->sk_drops);
306 sk_mem_reclaim_partial(sk);
310 EXPORT_SYMBOL(skb_kill_datagram);
313 * skb_copy_datagram_iter - Copy a datagram to an iovec iterator.
314 * @skb: buffer to copy
315 * @offset: offset in the buffer to start copying from
316 * @to: iovec iterator to copy to
317 * @len: amount of data to copy from buffer to iovec
319 int skb_copy_datagram_iter(const struct sk_buff *skb, int offset,
320 struct iov_iter *to, int len)
322 int start = skb_headlen(skb);
323 int i, copy = start - offset;
324 struct sk_buff *frag_iter;
326 trace_skb_copy_datagram_iovec(skb, len);
332 if (copy_to_iter(skb->data + offset, copy, to) != copy)
334 if ((len -= copy) == 0)
339 /* Copy paged appendix. Hmm... why does this look so complicated? */
340 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
342 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
344 WARN_ON(start > offset + len);
346 end = start + skb_frag_size(frag);
347 if ((copy = end - offset) > 0) {
350 if (copy_page_to_iter(skb_frag_page(frag),
351 frag->page_offset + offset -
352 start, copy, to) != copy)
361 skb_walk_frags(skb, frag_iter) {
364 WARN_ON(start > offset + len);
366 end = start + frag_iter->len;
367 if ((copy = end - offset) > 0) {
370 if (skb_copy_datagram_iter(frag_iter, offset - start,
373 if ((len -= copy) == 0)
382 /* This is not really a user copy fault, but rather someone
383 * gave us a bogus length on the skb. We should probably
384 * print a warning here as it may indicate a kernel bug.
391 if (iov_iter_count(to))
396 EXPORT_SYMBOL(skb_copy_datagram_iter);
399 * skb_copy_datagram_from_iter - Copy a datagram from an iov_iter.
400 * @skb: buffer to copy
401 * @offset: offset in the buffer to start copying to
402 * @from: the copy source
403 * @len: amount of data to copy to buffer from iovec
405 * Returns 0 or -EFAULT.
407 int skb_copy_datagram_from_iter(struct sk_buff *skb, int offset,
408 struct iov_iter *from,
411 int start = skb_headlen(skb);
412 int i, copy = start - offset;
413 struct sk_buff *frag_iter;
419 if (copy_from_iter(skb->data + offset, copy, from) != copy)
421 if ((len -= copy) == 0)
426 /* Copy paged appendix. Hmm... why does this look so complicated? */
427 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
429 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
431 WARN_ON(start > offset + len);
433 end = start + skb_frag_size(frag);
434 if ((copy = end - offset) > 0) {
439 copied = copy_page_from_iter(skb_frag_page(frag),
440 frag->page_offset + offset - start,
452 skb_walk_frags(skb, frag_iter) {
455 WARN_ON(start > offset + len);
457 end = start + frag_iter->len;
458 if ((copy = end - offset) > 0) {
461 if (skb_copy_datagram_from_iter(frag_iter,
465 if ((len -= copy) == 0)
477 EXPORT_SYMBOL(skb_copy_datagram_from_iter);
480 * zerocopy_sg_from_iter - Build a zerocopy datagram from an iov_iter
481 * @skb: buffer to copy
482 * @from: the source to copy from
484 * The function will first copy up to headlen, and then pin the userspace
485 * pages and build frags through them.
487 * Returns 0, -EFAULT or -EMSGSIZE.
489 int zerocopy_sg_from_iter(struct sk_buff *skb, struct iov_iter *from)
491 int len = iov_iter_count(from);
492 int copy = min_t(int, skb_headlen(skb), len);
495 /* copy up to skb headlen */
496 if (skb_copy_datagram_from_iter(skb, 0, from, copy))
499 while (iov_iter_count(from)) {
500 struct page *pages[MAX_SKB_FRAGS];
503 unsigned long truesize;
506 if (frag == MAX_SKB_FRAGS)
509 copied = iov_iter_get_pages(from, pages, ~0U,
510 MAX_SKB_FRAGS - frag, &start);
514 iov_iter_advance(from, copied);
516 truesize = PAGE_ALIGN(copied + start);
517 skb->data_len += copied;
519 skb->truesize += truesize;
520 atomic_add(truesize, &skb->sk->sk_wmem_alloc);
522 int size = min_t(int, copied, PAGE_SIZE - start);
523 skb_fill_page_desc(skb, frag++, pages[n], start, size);
531 EXPORT_SYMBOL(zerocopy_sg_from_iter);
533 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
534 struct iov_iter *to, int len,
537 int start = skb_headlen(skb);
538 int i, copy = start - offset;
539 struct sk_buff *frag_iter;
547 n = csum_and_copy_to_iter(skb->data + offset, copy, csump, to);
550 if ((len -= copy) == 0)
556 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
558 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
560 WARN_ON(start > offset + len);
562 end = start + skb_frag_size(frag);
563 if ((copy = end - offset) > 0) {
565 struct page *page = skb_frag_page(frag);
566 u8 *vaddr = kmap(page);
570 n = csum_and_copy_to_iter(vaddr + frag->page_offset +
571 offset - start, copy,
576 *csump = csum_block_add(*csump, csum2, pos);
585 skb_walk_frags(skb, frag_iter) {
588 WARN_ON(start > offset + len);
590 end = start + frag_iter->len;
591 if ((copy = end - offset) > 0) {
595 if (skb_copy_and_csum_datagram(frag_iter,
600 *csump = csum_block_add(*csump, csum2, pos);
601 if ((len -= copy) == 0)
615 __sum16 __skb_checksum_complete_head(struct sk_buff *skb, int len)
619 sum = csum_fold(skb_checksum(skb, 0, len, skb->csum));
621 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
622 !skb->csum_complete_sw)
623 netdev_rx_csum_fault(skb->dev);
625 skb->csum_valid = !sum;
628 EXPORT_SYMBOL(__skb_checksum_complete_head);
630 __sum16 __skb_checksum_complete(struct sk_buff *skb)
635 csum = skb_checksum(skb, 0, skb->len, 0);
637 /* skb->csum holds pseudo checksum */
638 sum = csum_fold(csum_add(skb->csum, csum));
640 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
641 !skb->csum_complete_sw)
642 netdev_rx_csum_fault(skb->dev);
645 /* Save full packet checksum */
647 skb->ip_summed = CHECKSUM_COMPLETE;
648 skb->csum_complete_sw = 1;
649 skb->csum_valid = !sum;
653 EXPORT_SYMBOL(__skb_checksum_complete);
656 * skb_copy_and_csum_datagram_msg - Copy and checksum skb to user iovec.
658 * @hlen: hardware length
661 * Caller _must_ check that skb will fit to this iovec.
663 * Returns: 0 - success.
664 * -EINVAL - checksum failure.
665 * -EFAULT - fault during copy.
667 int skb_copy_and_csum_datagram_msg(struct sk_buff *skb,
668 int hlen, struct msghdr *msg)
671 int chunk = skb->len - hlen;
676 if (msg_data_left(msg) < chunk) {
677 if (__skb_checksum_complete(skb))
679 if (skb_copy_datagram_msg(skb, hlen, msg, chunk))
682 csum = csum_partial(skb->data, hlen, skb->csum);
683 if (skb_copy_and_csum_datagram(skb, hlen, &msg->msg_iter,
688 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE))
689 netdev_rx_csum_fault(skb->dev);
697 EXPORT_SYMBOL(skb_copy_and_csum_datagram_msg);
700 * datagram_poll - generic datagram poll
705 * Datagram poll: Again totally generic. This also handles
706 * sequenced packet sockets providing the socket receive queue
707 * is only ever holding data ready to receive.
709 * Note: when you _don't_ use this routine for this protocol,
710 * and you use a different write policy from sock_writeable()
711 * then please supply your own write_space callback.
713 unsigned int datagram_poll(struct file *file, struct socket *sock,
716 struct sock *sk = sock->sk;
719 sock_poll_wait(file, sk_sleep(sk), wait);
722 /* exceptional events? */
723 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
725 (sock_flag(sk, SOCK_SELECT_ERR_QUEUE) ? POLLPRI : 0);
727 if (sk->sk_shutdown & RCV_SHUTDOWN)
728 mask |= POLLRDHUP | POLLIN | POLLRDNORM;
729 if (sk->sk_shutdown == SHUTDOWN_MASK)
733 if (!skb_queue_empty(&sk->sk_receive_queue))
734 mask |= POLLIN | POLLRDNORM;
736 /* Connection-based need to check for termination and startup */
737 if (connection_based(sk)) {
738 if (sk->sk_state == TCP_CLOSE)
740 /* connection hasn't started yet? */
741 if (sk->sk_state == TCP_SYN_SENT)
746 if (sock_writeable(sk))
747 mask |= POLLOUT | POLLWRNORM | POLLWRBAND;
749 set_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
753 EXPORT_SYMBOL(datagram_poll);