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 <linux/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 int __skb_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);
133 EXPORT_SYMBOL(__skb_wait_for_more_packets);
135 static struct sk_buff *skb_set_peeked(struct sk_buff *skb)
137 struct sk_buff *nskb;
142 /* We have to unshare an skb before modifying it. */
143 if (!skb_shared(skb))
146 nskb = skb_clone(skb, GFP_ATOMIC);
148 return ERR_PTR(-ENOMEM);
150 skb->prev->next = nskb;
151 skb->next->prev = nskb;
152 nskb->prev = skb->prev;
153 nskb->next = skb->next;
165 * __skb_try_recv_datagram - Receive a datagram skbuff
168 * @destructor: invoked under the receive lock on successful dequeue
169 * @peeked: returns non-zero if this packet has been seen before
170 * @off: an offset in bytes to peek skb from. Returns an offset
171 * within an skb where data actually starts
172 * @err: error code returned
173 * @last: set to last peeked message to inform the wait function
174 * what to look for when peeking
176 * Get a datagram skbuff, understands the peeking, nonblocking wakeups
177 * and possible races. This replaces identical code in packet, raw and
178 * udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
179 * the long standing peek and read race for datagram sockets. If you
180 * alter this routine remember it must be re-entrant.
182 * This function will lock the socket if a skb is returned, so
183 * the caller needs to unlock the socket in that case (usually by
184 * calling skb_free_datagram). Returns NULL with *err set to
185 * -EAGAIN if no data was available or to some other value if an
186 * error was detected.
188 * * It does not lock socket since today. This function is
189 * * free of race conditions. This measure should/can improve
190 * * significantly datagram socket latencies at high loads,
191 * * when data copying to user space takes lots of time.
192 * * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
196 * The order of the tests when we find no data waiting are specified
197 * quite explicitly by POSIX 1003.1g, don't change them without having
198 * the standard around please.
200 struct sk_buff *__skb_try_recv_datagram(struct sock *sk, unsigned int flags,
201 void (*destructor)(struct sock *sk,
202 struct sk_buff *skb),
203 int *peeked, int *off, int *err,
204 struct sk_buff **last)
206 struct sk_buff_head *queue = &sk->sk_receive_queue;
208 unsigned long cpu_flags;
210 * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
212 int error = sock_error(sk);
219 /* Again only user level code calls this function, so nothing
220 * interrupt level will suddenly eat the receive_queue.
222 * Look at current nfs client by the way...
223 * However, this function was correct in any case. 8)
227 *last = (struct sk_buff *)queue;
228 spin_lock_irqsave(&queue->lock, cpu_flags);
229 skb_queue_walk(queue, skb) {
231 if (flags & MSG_PEEK) {
232 if (_off >= skb->len && (skb->len || _off ||
238 skb = skb_set_peeked(skb);
240 error = PTR_ERR(skb);
241 spin_unlock_irqrestore(&queue->lock,
247 atomic_inc(&skb->users);
249 __skb_unlink(skb, queue);
253 spin_unlock_irqrestore(&queue->lock, cpu_flags);
258 spin_unlock_irqrestore(&queue->lock, cpu_flags);
260 if (!sk_can_busy_loop(sk))
263 sk_busy_loop(sk, flags & MSG_DONTWAIT);
264 } while (!skb_queue_empty(&sk->sk_receive_queue));
272 EXPORT_SYMBOL(__skb_try_recv_datagram);
274 struct sk_buff *__skb_recv_datagram(struct sock *sk, unsigned int flags,
275 void (*destructor)(struct sock *sk,
276 struct sk_buff *skb),
277 int *peeked, int *off, int *err)
279 struct sk_buff *skb, *last;
282 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
285 skb = __skb_try_recv_datagram(sk, flags, destructor, peeked,
293 !__skb_wait_for_more_packets(sk, err, &timeo, last));
297 EXPORT_SYMBOL(__skb_recv_datagram);
299 struct sk_buff *skb_recv_datagram(struct sock *sk, unsigned int flags,
300 int noblock, int *err)
304 return __skb_recv_datagram(sk, flags | (noblock ? MSG_DONTWAIT : 0),
305 NULL, &peeked, &off, err);
307 EXPORT_SYMBOL(skb_recv_datagram);
309 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
312 sk_mem_reclaim_partial(sk);
314 EXPORT_SYMBOL(skb_free_datagram);
316 void __skb_free_datagram_locked(struct sock *sk, struct sk_buff *skb, int len)
320 if (likely(atomic_read(&skb->users) == 1))
322 else if (likely(!atomic_dec_and_test(&skb->users))) {
323 sk_peek_offset_bwd(sk, len);
327 slow = lock_sock_fast(sk);
328 sk_peek_offset_bwd(sk, len);
330 sk_mem_reclaim_partial(sk);
331 unlock_sock_fast(sk, slow);
333 /* skb is now orphaned, can be freed outside of locked section */
336 EXPORT_SYMBOL(__skb_free_datagram_locked);
338 int __sk_queue_drop_skb(struct sock *sk, struct sk_buff *skb,
340 void (*destructor)(struct sock *sk,
341 struct sk_buff *skb))
345 if (flags & MSG_PEEK) {
347 spin_lock_bh(&sk->sk_receive_queue.lock);
348 if (skb == skb_peek(&sk->sk_receive_queue)) {
349 __skb_unlink(skb, &sk->sk_receive_queue);
350 atomic_dec(&skb->users);
355 spin_unlock_bh(&sk->sk_receive_queue.lock);
358 atomic_inc(&sk->sk_drops);
361 EXPORT_SYMBOL(__sk_queue_drop_skb);
364 * skb_kill_datagram - Free a datagram skbuff forcibly
366 * @skb: datagram skbuff
369 * This function frees a datagram skbuff that was received by
370 * skb_recv_datagram. The flags argument must match the one
371 * used for skb_recv_datagram.
373 * If the MSG_PEEK flag is set, and the packet is still on the
374 * receive queue of the socket, it will be taken off the queue
375 * before it is freed.
377 * This function currently only disables BH when acquiring the
378 * sk_receive_queue lock. Therefore it must not be used in a
379 * context where that lock is acquired in an IRQ context.
381 * It returns 0 if the packet was removed by us.
384 int skb_kill_datagram(struct sock *sk, struct sk_buff *skb, unsigned int flags)
386 int err = __sk_queue_drop_skb(sk, skb, flags, NULL);
389 sk_mem_reclaim_partial(sk);
392 EXPORT_SYMBOL(skb_kill_datagram);
395 * skb_copy_datagram_iter - Copy a datagram to an iovec iterator.
396 * @skb: buffer to copy
397 * @offset: offset in the buffer to start copying from
398 * @to: iovec iterator to copy to
399 * @len: amount of data to copy from buffer to iovec
401 int skb_copy_datagram_iter(const struct sk_buff *skb, int offset,
402 struct iov_iter *to, int len)
404 int start = skb_headlen(skb);
405 int i, copy = start - offset, start_off = offset, n;
406 struct sk_buff *frag_iter;
408 trace_skb_copy_datagram_iovec(skb, len);
414 n = copy_to_iter(skb->data + offset, copy, to);
418 if ((len -= copy) == 0)
422 /* Copy paged appendix. Hmm... why does this look so complicated? */
423 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
425 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
427 WARN_ON(start > offset + len);
429 end = start + skb_frag_size(frag);
430 if ((copy = end - offset) > 0) {
433 n = copy_page_to_iter(skb_frag_page(frag),
434 frag->page_offset + offset -
445 skb_walk_frags(skb, frag_iter) {
448 WARN_ON(start > offset + len);
450 end = start + frag_iter->len;
451 if ((copy = end - offset) > 0) {
454 if (skb_copy_datagram_iter(frag_iter, offset - start,
457 if ((len -= copy) == 0)
466 /* This is not really a user copy fault, but rather someone
467 * gave us a bogus length on the skb. We should probably
468 * print a warning here as it may indicate a kernel bug.
472 iov_iter_revert(to, offset - start_off);
476 if (iov_iter_count(to))
481 EXPORT_SYMBOL(skb_copy_datagram_iter);
484 * skb_copy_datagram_from_iter - Copy a datagram from an iov_iter.
485 * @skb: buffer to copy
486 * @offset: offset in the buffer to start copying to
487 * @from: the copy source
488 * @len: amount of data to copy to buffer from iovec
490 * Returns 0 or -EFAULT.
492 int skb_copy_datagram_from_iter(struct sk_buff *skb, int offset,
493 struct iov_iter *from,
496 int start = skb_headlen(skb);
497 int i, copy = start - offset;
498 struct sk_buff *frag_iter;
504 if (copy_from_iter(skb->data + offset, copy, from) != copy)
506 if ((len -= copy) == 0)
511 /* Copy paged appendix. Hmm... why does this look so complicated? */
512 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
514 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
516 WARN_ON(start > offset + len);
518 end = start + skb_frag_size(frag);
519 if ((copy = end - offset) > 0) {
524 copied = copy_page_from_iter(skb_frag_page(frag),
525 frag->page_offset + offset - start,
537 skb_walk_frags(skb, frag_iter) {
540 WARN_ON(start > offset + len);
542 end = start + frag_iter->len;
543 if ((copy = end - offset) > 0) {
546 if (skb_copy_datagram_from_iter(frag_iter,
550 if ((len -= copy) == 0)
562 EXPORT_SYMBOL(skb_copy_datagram_from_iter);
565 * zerocopy_sg_from_iter - Build a zerocopy datagram from an iov_iter
566 * @skb: buffer to copy
567 * @from: the source to copy from
569 * The function will first copy up to headlen, and then pin the userspace
570 * pages and build frags through them.
572 * Returns 0, -EFAULT or -EMSGSIZE.
574 int zerocopy_sg_from_iter(struct sk_buff *skb, struct iov_iter *from)
576 int len = iov_iter_count(from);
577 int copy = min_t(int, skb_headlen(skb), len);
580 /* copy up to skb headlen */
581 if (skb_copy_datagram_from_iter(skb, 0, from, copy))
584 while (iov_iter_count(from)) {
585 struct page *pages[MAX_SKB_FRAGS];
588 unsigned long truesize;
591 if (frag == MAX_SKB_FRAGS)
594 copied = iov_iter_get_pages(from, pages, ~0U,
595 MAX_SKB_FRAGS - frag, &start);
599 iov_iter_advance(from, copied);
601 truesize = PAGE_ALIGN(copied + start);
602 skb->data_len += copied;
604 skb->truesize += truesize;
605 atomic_add(truesize, &skb->sk->sk_wmem_alloc);
607 int size = min_t(int, copied, PAGE_SIZE - start);
608 skb_fill_page_desc(skb, frag++, pages[n], start, size);
616 EXPORT_SYMBOL(zerocopy_sg_from_iter);
618 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
619 struct iov_iter *to, int len,
622 int start = skb_headlen(skb);
623 int i, copy = start - offset, start_off = offset;
624 struct sk_buff *frag_iter;
632 n = csum_and_copy_to_iter(skb->data + offset, copy, csump, to);
636 if ((len -= copy) == 0)
641 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
643 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
645 WARN_ON(start > offset + len);
647 end = start + skb_frag_size(frag);
648 if ((copy = end - offset) > 0) {
650 struct page *page = skb_frag_page(frag);
651 u8 *vaddr = kmap(page);
655 n = csum_and_copy_to_iter(vaddr + frag->page_offset +
656 offset - start, copy,
662 *csump = csum_block_add(*csump, csum2, pos);
670 skb_walk_frags(skb, frag_iter) {
673 WARN_ON(start > offset + len);
675 end = start + frag_iter->len;
676 if ((copy = end - offset) > 0) {
680 if (skb_copy_and_csum_datagram(frag_iter,
685 *csump = csum_block_add(*csump, csum2, pos);
686 if ((len -= copy) == 0)
697 iov_iter_revert(to, offset - start_off);
701 __sum16 __skb_checksum_complete_head(struct sk_buff *skb, int len)
705 sum = csum_fold(skb_checksum(skb, 0, len, skb->csum));
707 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
708 !skb->csum_complete_sw)
709 netdev_rx_csum_fault(skb->dev);
711 if (!skb_shared(skb))
712 skb->csum_valid = !sum;
715 EXPORT_SYMBOL(__skb_checksum_complete_head);
717 __sum16 __skb_checksum_complete(struct sk_buff *skb)
722 csum = skb_checksum(skb, 0, skb->len, 0);
724 /* skb->csum holds pseudo checksum */
725 sum = csum_fold(csum_add(skb->csum, csum));
727 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
728 !skb->csum_complete_sw)
729 netdev_rx_csum_fault(skb->dev);
732 if (!skb_shared(skb)) {
733 /* Save full packet checksum */
735 skb->ip_summed = CHECKSUM_COMPLETE;
736 skb->csum_complete_sw = 1;
737 skb->csum_valid = !sum;
742 EXPORT_SYMBOL(__skb_checksum_complete);
745 * skb_copy_and_csum_datagram_msg - Copy and checksum skb to user iovec.
747 * @hlen: hardware length
750 * Caller _must_ check that skb will fit to this iovec.
752 * Returns: 0 - success.
753 * -EINVAL - checksum failure.
754 * -EFAULT - fault during copy.
756 int skb_copy_and_csum_datagram_msg(struct sk_buff *skb,
757 int hlen, struct msghdr *msg)
760 int chunk = skb->len - hlen;
765 if (msg_data_left(msg) < chunk) {
766 if (__skb_checksum_complete(skb))
768 if (skb_copy_datagram_msg(skb, hlen, msg, chunk))
771 csum = csum_partial(skb->data, hlen, skb->csum);
772 if (skb_copy_and_csum_datagram(skb, hlen, &msg->msg_iter,
776 if (csum_fold(csum)) {
777 iov_iter_revert(&msg->msg_iter, chunk);
781 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE))
782 netdev_rx_csum_fault(skb->dev);
788 EXPORT_SYMBOL(skb_copy_and_csum_datagram_msg);
791 * datagram_poll - generic datagram poll
796 * Datagram poll: Again totally generic. This also handles
797 * sequenced packet sockets providing the socket receive queue
798 * is only ever holding data ready to receive.
800 * Note: when you _don't_ use this routine for this protocol,
801 * and you use a different write policy from sock_writeable()
802 * then please supply your own write_space callback.
804 unsigned int datagram_poll(struct file *file, struct socket *sock,
807 struct sock *sk = sock->sk;
810 sock_poll_wait(file, sk_sleep(sk), wait);
813 /* exceptional events? */
814 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
816 (sock_flag(sk, SOCK_SELECT_ERR_QUEUE) ? POLLPRI : 0);
818 if (sk->sk_shutdown & RCV_SHUTDOWN)
819 mask |= POLLRDHUP | POLLIN | POLLRDNORM;
820 if (sk->sk_shutdown == SHUTDOWN_MASK)
824 if (!skb_queue_empty(&sk->sk_receive_queue))
825 mask |= POLLIN | POLLRDNORM;
827 /* Connection-based need to check for termination and startup */
828 if (connection_based(sk)) {
829 if (sk->sk_state == TCP_CLOSE)
831 /* connection hasn't started yet? */
832 if (sk->sk_state == TCP_SYN_SENT)
837 if (sock_writeable(sk))
838 mask |= POLLOUT | POLLWRNORM | POLLWRBAND;
840 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
844 EXPORT_SYMBOL(datagram_poll);