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>
53 #include <net/protocol.h>
54 #include <linux/skbuff.h>
56 #include <net/checksum.h>
58 #include <net/tcp_states.h>
59 #include <trace/events/skb.h>
60 #include <net/busy_poll.h>
63 * Is a socket 'connection oriented' ?
65 static inline int connection_based(struct sock *sk)
67 return sk->sk_type == SOCK_SEQPACKET || sk->sk_type == SOCK_STREAM;
70 static int receiver_wake_function(wait_queue_t *wait, unsigned int mode, int sync,
73 unsigned long bits = (unsigned long)key;
76 * Avoid a wakeup if event not interesting for us
78 if (bits && !(bits & (POLLIN | POLLERR)))
80 return autoremove_wake_function(wait, mode, sync, key);
83 * Wait for the last received packet to be different from skb
85 static int wait_for_more_packets(struct sock *sk, int *err, long *timeo_p,
86 const struct sk_buff *skb)
89 DEFINE_WAIT_FUNC(wait, receiver_wake_function);
91 prepare_to_wait_exclusive(sk_sleep(sk), &wait, TASK_INTERRUPTIBLE);
94 error = sock_error(sk);
98 if (sk->sk_receive_queue.prev != skb)
101 /* Socket shut down? */
102 if (sk->sk_shutdown & RCV_SHUTDOWN)
105 /* Sequenced packets can come disconnected.
106 * If so we report the problem
109 if (connection_based(sk) &&
110 !(sk->sk_state == TCP_ESTABLISHED || sk->sk_state == TCP_LISTEN))
114 if (signal_pending(current))
118 *timeo_p = schedule_timeout(*timeo_p);
120 finish_wait(sk_sleep(sk), &wait);
123 error = sock_intr_errno(*timeo_p);
134 * __skb_recv_datagram - Receive a datagram skbuff
137 * @peeked: returns non-zero if this packet has been seen before
138 * @off: an offset in bytes to peek skb from. Returns an offset
139 * within an skb where data actually starts
140 * @err: error code returned
142 * Get a datagram skbuff, understands the peeking, nonblocking wakeups
143 * and possible races. This replaces identical code in packet, raw and
144 * udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
145 * the long standing peek and read race for datagram sockets. If you
146 * alter this routine remember it must be re-entrant.
148 * This function will lock the socket if a skb is returned, so the caller
149 * needs to unlock the socket in that case (usually by calling
152 * * It does not lock socket since today. This function is
153 * * free of race conditions. This measure should/can improve
154 * * significantly datagram socket latencies at high loads,
155 * * when data copying to user space takes lots of time.
156 * * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
160 * The order of the tests when we find no data waiting are specified
161 * quite explicitly by POSIX 1003.1g, don't change them without having
162 * the standard around please.
164 struct sk_buff *__skb_recv_datagram(struct sock *sk, unsigned int flags,
165 int *peeked, int *off, int *err)
167 struct sk_buff *skb, *last;
170 * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
172 int error = sock_error(sk);
177 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
180 /* Again only user level code calls this function, so nothing
181 * interrupt level will suddenly eat the receive_queue.
183 * Look at current nfs client by the way...
184 * However, this function was correct in any case. 8)
186 unsigned long cpu_flags;
187 struct sk_buff_head *queue = &sk->sk_receive_queue;
190 last = (struct sk_buff *)queue;
191 spin_lock_irqsave(&queue->lock, cpu_flags);
192 skb_queue_walk(queue, skb) {
194 *peeked = skb->peeked;
195 if (flags & MSG_PEEK) {
196 if (_off >= skb->len && (skb->len || _off ||
202 atomic_inc(&skb->users);
204 __skb_unlink(skb, queue);
206 spin_unlock_irqrestore(&queue->lock, cpu_flags);
210 spin_unlock_irqrestore(&queue->lock, cpu_flags);
212 if (sk_can_busy_loop(sk) &&
213 sk_busy_loop(sk, flags & MSG_DONTWAIT))
216 /* User doesn't want to wait */
221 } while (!wait_for_more_packets(sk, err, &timeo, last));
229 EXPORT_SYMBOL(__skb_recv_datagram);
231 struct sk_buff *skb_recv_datagram(struct sock *sk, unsigned int flags,
232 int noblock, int *err)
236 return __skb_recv_datagram(sk, flags | (noblock ? MSG_DONTWAIT : 0),
239 EXPORT_SYMBOL(skb_recv_datagram);
241 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
244 sk_mem_reclaim_partial(sk);
246 EXPORT_SYMBOL(skb_free_datagram);
248 void skb_free_datagram_locked(struct sock *sk, struct sk_buff *skb)
252 if (likely(atomic_read(&skb->users) == 1))
254 else if (likely(!atomic_dec_and_test(&skb->users)))
257 slow = lock_sock_fast(sk);
259 sk_mem_reclaim_partial(sk);
260 unlock_sock_fast(sk, slow);
262 /* skb is now orphaned, can be freed outside of locked section */
265 EXPORT_SYMBOL(skb_free_datagram_locked);
268 * skb_kill_datagram - Free a datagram skbuff forcibly
270 * @skb: datagram skbuff
273 * This function frees a datagram skbuff that was received by
274 * skb_recv_datagram. The flags argument must match the one
275 * used for skb_recv_datagram.
277 * If the MSG_PEEK flag is set, and the packet is still on the
278 * receive queue of the socket, it will be taken off the queue
279 * before it is freed.
281 * This function currently only disables BH when acquiring the
282 * sk_receive_queue lock. Therefore it must not be used in a
283 * context where that lock is acquired in an IRQ context.
285 * It returns 0 if the packet was removed by us.
288 int skb_kill_datagram(struct sock *sk, struct sk_buff *skb, unsigned int flags)
292 if (flags & MSG_PEEK) {
294 spin_lock_bh(&sk->sk_receive_queue.lock);
295 if (skb == skb_peek(&sk->sk_receive_queue)) {
296 __skb_unlink(skb, &sk->sk_receive_queue);
297 atomic_dec(&skb->users);
300 spin_unlock_bh(&sk->sk_receive_queue.lock);
304 atomic_inc(&sk->sk_drops);
305 sk_mem_reclaim_partial(sk);
309 EXPORT_SYMBOL(skb_kill_datagram);
312 * skb_copy_datagram_iovec - Copy a datagram to an iovec.
313 * @skb: buffer to copy
314 * @offset: offset in the buffer to start copying from
315 * @to: io vector to copy to
316 * @len: amount of data to copy from buffer to iovec
318 * Note: the iovec is modified during the copy.
320 int skb_copy_datagram_iovec(const struct sk_buff *skb, int offset,
321 struct iovec *to, int len)
323 int start = skb_headlen(skb);
324 int i, copy = start - offset;
325 struct sk_buff *frag_iter;
327 trace_skb_copy_datagram_iovec(skb, len);
333 if (memcpy_toiovec(to, skb->data + offset, copy))
335 if ((len -= copy) == 0)
340 /* Copy paged appendix. Hmm... why does this look so complicated? */
341 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
343 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
345 WARN_ON(start > offset + len);
347 end = start + skb_frag_size(frag);
348 if ((copy = end - offset) > 0) {
351 struct page *page = skb_frag_page(frag);
356 err = memcpy_toiovec(to, vaddr + frag->page_offset +
357 offset - start, copy);
368 skb_walk_frags(skb, frag_iter) {
371 WARN_ON(start > offset + len);
373 end = start + frag_iter->len;
374 if ((copy = end - offset) > 0) {
377 if (skb_copy_datagram_iovec(frag_iter,
381 if ((len -= copy) == 0)
393 EXPORT_SYMBOL(skb_copy_datagram_iovec);
396 * skb_copy_datagram_const_iovec - Copy a datagram to an iovec.
397 * @skb: buffer to copy
398 * @offset: offset in the buffer to start copying from
399 * @to: io vector to copy to
400 * @to_offset: offset in the io vector to start copying to
401 * @len: amount of data to copy from buffer to iovec
403 * Returns 0 or -EFAULT.
404 * Note: the iovec is not modified during the copy.
406 int skb_copy_datagram_const_iovec(const struct sk_buff *skb, int offset,
407 const struct iovec *to, int to_offset,
410 int start = skb_headlen(skb);
411 int i, copy = start - offset;
412 struct sk_buff *frag_iter;
418 if (memcpy_toiovecend(to, skb->data + offset, to_offset, copy))
420 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) {
437 struct page *page = skb_frag_page(frag);
442 err = memcpy_toiovecend(to, vaddr + frag->page_offset +
443 offset - start, to_offset, copy);
455 skb_walk_frags(skb, frag_iter) {
458 WARN_ON(start > offset + len);
460 end = start + frag_iter->len;
461 if ((copy = end - offset) > 0) {
464 if (skb_copy_datagram_const_iovec(frag_iter,
469 if ((len -= copy) == 0)
482 EXPORT_SYMBOL(skb_copy_datagram_const_iovec);
485 * skb_copy_datagram_from_iovec - Copy a datagram from an iovec.
486 * @skb: buffer to copy
487 * @offset: offset in the buffer to start copying to
488 * @from: io vector to copy to
489 * @from_offset: offset in the io vector to start copying from
490 * @len: amount of data to copy to buffer from iovec
492 * Returns 0 or -EFAULT.
493 * Note: the iovec is not modified during the copy.
495 int skb_copy_datagram_from_iovec(struct sk_buff *skb, int offset,
496 const struct iovec *from, int from_offset,
499 int start = skb_headlen(skb);
500 int i, copy = start - offset;
501 struct sk_buff *frag_iter;
507 if (memcpy_fromiovecend(skb->data + offset, from, from_offset,
510 if ((len -= copy) == 0)
516 /* Copy paged appendix. Hmm... why does this look so complicated? */
517 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
519 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
521 WARN_ON(start > offset + len);
523 end = start + skb_frag_size(frag);
524 if ((copy = end - offset) > 0) {
527 struct page *page = skb_frag_page(frag);
532 err = memcpy_fromiovecend(vaddr + frag->page_offset +
534 from, from_offset, copy);
547 skb_walk_frags(skb, frag_iter) {
550 WARN_ON(start > offset + len);
552 end = start + frag_iter->len;
553 if ((copy = end - offset) > 0) {
556 if (skb_copy_datagram_from_iovec(frag_iter,
562 if ((len -= copy) == 0)
575 EXPORT_SYMBOL(skb_copy_datagram_from_iovec);
578 * zerocopy_sg_from_iovec - Build a zerocopy datagram from an iovec
579 * @skb: buffer to copy
580 * @from: io vector to copy from
581 * @offset: offset in the io vector to start copying from
582 * @count: amount of vectors to copy to buffer from
584 * The function will first copy up to headlen, and then pin the userspace
585 * pages and build frags through them.
587 * Returns 0, -EFAULT or -EMSGSIZE.
588 * Note: the iovec is not modified during the copy
590 int zerocopy_sg_from_iovec(struct sk_buff *skb, const struct iovec *from,
591 int offset, size_t count)
593 int len = iov_length(from, count) - offset;
594 int copy = min_t(int, skb_headlen(skb), len);
598 /* copy up to skb headlen */
599 if (skb_copy_datagram_from_iovec(skb, 0, from, offset, copy))
607 struct page *page[MAX_SKB_FRAGS];
610 unsigned long truesize;
612 /* Skip over from offset and copied */
613 if (offset >= from->iov_len) {
614 offset -= from->iov_len;
618 len = from->iov_len - offset;
619 base = (unsigned long)from->iov_base + offset;
620 size = ((base & ~PAGE_MASK) + len + ~PAGE_MASK) >> PAGE_SHIFT;
621 if (i + size > MAX_SKB_FRAGS)
623 num_pages = get_user_pages_fast(base, size, 0, &page[i]);
624 if (num_pages != size) {
625 release_pages(&page[i], num_pages, 0);
628 truesize = size * PAGE_SIZE;
629 skb->data_len += len;
631 skb->truesize += truesize;
632 atomic_add(truesize, &skb->sk->sk_wmem_alloc);
634 int off = base & ~PAGE_MASK;
635 int size = min_t(int, len, PAGE_SIZE - off);
636 skb_fill_page_desc(skb, i, page[i], off, size);
646 EXPORT_SYMBOL(zerocopy_sg_from_iovec);
648 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
649 u8 __user *to, int len,
652 int start = skb_headlen(skb);
653 int i, copy = start - offset;
654 struct sk_buff *frag_iter;
662 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
666 if ((len -= copy) == 0)
673 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
675 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
677 WARN_ON(start > offset + len);
679 end = start + skb_frag_size(frag);
680 if ((copy = end - offset) > 0) {
684 struct page *page = skb_frag_page(frag);
689 csum2 = csum_and_copy_to_user(vaddr +
696 *csump = csum_block_add(*csump, csum2, pos);
706 skb_walk_frags(skb, frag_iter) {
709 WARN_ON(start > offset + len);
711 end = start + frag_iter->len;
712 if ((copy = end - offset) > 0) {
716 if (skb_copy_and_csum_datagram(frag_iter,
721 *csump = csum_block_add(*csump, csum2, pos);
722 if ((len -= copy) == 0)
737 __sum16 __skb_checksum_complete_head(struct sk_buff *skb, int len)
741 sum = csum_fold(skb_checksum(skb, 0, len, skb->csum));
743 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE))
744 netdev_rx_csum_fault(skb->dev);
745 skb->ip_summed = CHECKSUM_UNNECESSARY;
749 EXPORT_SYMBOL(__skb_checksum_complete_head);
751 __sum16 __skb_checksum_complete(struct sk_buff *skb)
753 return __skb_checksum_complete_head(skb, skb->len);
755 EXPORT_SYMBOL(__skb_checksum_complete);
758 * skb_copy_and_csum_datagram_iovec - Copy and checkum skb to user iovec.
760 * @hlen: hardware length
763 * Caller _must_ check that skb will fit to this iovec.
765 * Returns: 0 - success.
766 * -EINVAL - checksum failure.
767 * -EFAULT - fault during copy. Beware, in this case iovec
770 int skb_copy_and_csum_datagram_iovec(struct sk_buff *skb,
771 int hlen, struct iovec *iov)
774 int chunk = skb->len - hlen;
779 /* Skip filled elements.
780 * Pretty silly, look at memcpy_toiovec, though 8)
782 while (!iov->iov_len)
785 if (iov->iov_len < chunk) {
786 if (__skb_checksum_complete(skb))
788 if (skb_copy_datagram_iovec(skb, hlen, iov, chunk))
791 csum = csum_partial(skb->data, hlen, skb->csum);
792 if (skb_copy_and_csum_datagram(skb, hlen, iov->iov_base,
797 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE))
798 netdev_rx_csum_fault(skb->dev);
799 iov->iov_len -= chunk;
800 iov->iov_base += chunk;
808 EXPORT_SYMBOL(skb_copy_and_csum_datagram_iovec);
811 * datagram_poll - generic datagram poll
816 * Datagram poll: Again totally generic. This also handles
817 * sequenced packet sockets providing the socket receive queue
818 * is only ever holding data ready to receive.
820 * Note: when you _don't_ use this routine for this protocol,
821 * and you use a different write policy from sock_writeable()
822 * then please supply your own write_space callback.
824 unsigned int datagram_poll(struct file *file, struct socket *sock,
827 struct sock *sk = sock->sk;
830 sock_poll_wait(file, sk_sleep(sk), wait);
833 /* exceptional events? */
834 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
836 (sock_flag(sk, SOCK_SELECT_ERR_QUEUE) ? POLLPRI : 0);
838 if (sk->sk_shutdown & RCV_SHUTDOWN)
839 mask |= POLLRDHUP | POLLIN | POLLRDNORM;
840 if (sk->sk_shutdown == SHUTDOWN_MASK)
844 if (!skb_queue_empty(&sk->sk_receive_queue))
845 mask |= POLLIN | POLLRDNORM;
847 /* Connection-based need to check for termination and startup */
848 if (connection_based(sk)) {
849 if (sk->sk_state == TCP_CLOSE)
851 /* connection hasn't started yet? */
852 if (sk->sk_state == TCP_SYN_SENT)
857 if (sock_writeable(sk))
858 mask |= POLLOUT | POLLWRNORM | POLLWRBAND;
860 set_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
864 EXPORT_SYMBOL(datagram_poll);