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_entry_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
;
164 struct sk_buff
*__skb_try_recv_from_queue(struct sock
*sk
,
165 struct sk_buff_head
*queue
,
167 void (*destructor
)(struct sock
*sk
,
168 struct sk_buff
*skb
),
169 int *peeked
, int *off
, int *err
,
170 struct sk_buff
**last
)
172 bool peek_at_off
= false;
176 if (unlikely(flags
& MSG_PEEK
&& *off
>= 0)) {
182 skb_queue_walk(queue
, skb
) {
183 if (flags
& MSG_PEEK
) {
184 if (peek_at_off
&& _off
>= skb
->len
&&
185 (_off
|| skb
->peeked
)) {
190 skb
= skb_set_peeked(skb
);
191 if (unlikely(IS_ERR(skb
))) {
197 refcount_inc(&skb
->users
);
199 __skb_unlink(skb
, queue
);
210 * __skb_try_recv_datagram - Receive a datagram skbuff
212 * @flags: MSG\_ flags
213 * @destructor: invoked under the receive lock on successful dequeue
214 * @peeked: returns non-zero if this packet has been seen before
215 * @off: an offset in bytes to peek skb from. Returns an offset
216 * within an skb where data actually starts
217 * @err: error code returned
218 * @last: set to last peeked message to inform the wait function
219 * what to look for when peeking
221 * Get a datagram skbuff, understands the peeking, nonblocking wakeups
222 * and possible races. This replaces identical code in packet, raw and
223 * udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
224 * the long standing peek and read race for datagram sockets. If you
225 * alter this routine remember it must be re-entrant.
227 * This function will lock the socket if a skb is returned, so
228 * the caller needs to unlock the socket in that case (usually by
229 * calling skb_free_datagram). Returns NULL with @err set to
230 * -EAGAIN if no data was available or to some other value if an
231 * error was detected.
233 * * It does not lock socket since today. This function is
234 * * free of race conditions. This measure should/can improve
235 * * significantly datagram socket latencies at high loads,
236 * * when data copying to user space takes lots of time.
237 * * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
241 * The order of the tests when we find no data waiting are specified
242 * quite explicitly by POSIX 1003.1g, don't change them without having
243 * the standard around please.
245 struct sk_buff
*__skb_try_recv_datagram(struct sock
*sk
, unsigned int flags
,
246 void (*destructor
)(struct sock
*sk
,
247 struct sk_buff
*skb
),
248 int *peeked
, int *off
, int *err
,
249 struct sk_buff
**last
)
251 struct sk_buff_head
*queue
= &sk
->sk_receive_queue
;
253 unsigned long cpu_flags
;
255 * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
257 int error
= sock_error(sk
);
264 /* Again only user level code calls this function, so nothing
265 * interrupt level will suddenly eat the receive_queue.
267 * Look at current nfs client by the way...
268 * However, this function was correct in any case. 8)
270 spin_lock_irqsave(&queue
->lock
, cpu_flags
);
271 skb
= __skb_try_recv_from_queue(sk
, queue
, flags
, destructor
,
272 peeked
, off
, &error
, last
);
273 spin_unlock_irqrestore(&queue
->lock
, cpu_flags
);
279 if (!sk_can_busy_loop(sk
))
282 sk_busy_loop(sk
, flags
& MSG_DONTWAIT
);
283 } while (!skb_queue_empty(&sk
->sk_receive_queue
));
291 EXPORT_SYMBOL(__skb_try_recv_datagram
);
293 struct sk_buff
*__skb_recv_datagram(struct sock
*sk
, unsigned int flags
,
294 void (*destructor
)(struct sock
*sk
,
295 struct sk_buff
*skb
),
296 int *peeked
, int *off
, int *err
)
298 struct sk_buff
*skb
, *last
;
301 timeo
= sock_rcvtimeo(sk
, flags
& MSG_DONTWAIT
);
304 skb
= __skb_try_recv_datagram(sk
, flags
, destructor
, peeked
,
312 !__skb_wait_for_more_packets(sk
, err
, &timeo
, last
));
316 EXPORT_SYMBOL(__skb_recv_datagram
);
318 struct sk_buff
*skb_recv_datagram(struct sock
*sk
, unsigned int flags
,
319 int noblock
, int *err
)
323 return __skb_recv_datagram(sk
, flags
| (noblock
? MSG_DONTWAIT
: 0),
324 NULL
, &peeked
, &off
, err
);
326 EXPORT_SYMBOL(skb_recv_datagram
);
328 void skb_free_datagram(struct sock
*sk
, struct sk_buff
*skb
)
331 sk_mem_reclaim_partial(sk
);
333 EXPORT_SYMBOL(skb_free_datagram
);
335 void __skb_free_datagram_locked(struct sock
*sk
, struct sk_buff
*skb
, int len
)
339 if (!skb_unref(skb
)) {
340 sk_peek_offset_bwd(sk
, len
);
344 slow
= lock_sock_fast(sk
);
345 sk_peek_offset_bwd(sk
, len
);
347 sk_mem_reclaim_partial(sk
);
348 unlock_sock_fast(sk
, slow
);
350 /* skb is now orphaned, can be freed outside of locked section */
353 EXPORT_SYMBOL(__skb_free_datagram_locked
);
355 int __sk_queue_drop_skb(struct sock
*sk
, struct sk_buff_head
*sk_queue
,
356 struct sk_buff
*skb
, unsigned int flags
,
357 void (*destructor
)(struct sock
*sk
,
358 struct sk_buff
*skb
))
362 if (flags
& MSG_PEEK
) {
364 spin_lock_bh(&sk_queue
->lock
);
366 __skb_unlink(skb
, sk_queue
);
367 refcount_dec(&skb
->users
);
372 spin_unlock_bh(&sk_queue
->lock
);
375 atomic_inc(&sk
->sk_drops
);
378 EXPORT_SYMBOL(__sk_queue_drop_skb
);
381 * skb_kill_datagram - Free a datagram skbuff forcibly
383 * @skb: datagram skbuff
384 * @flags: MSG\_ flags
386 * This function frees a datagram skbuff that was received by
387 * skb_recv_datagram. The flags argument must match the one
388 * used for skb_recv_datagram.
390 * If the MSG_PEEK flag is set, and the packet is still on the
391 * receive queue of the socket, it will be taken off the queue
392 * before it is freed.
394 * This function currently only disables BH when acquiring the
395 * sk_receive_queue lock. Therefore it must not be used in a
396 * context where that lock is acquired in an IRQ context.
398 * It returns 0 if the packet was removed by us.
401 int skb_kill_datagram(struct sock
*sk
, struct sk_buff
*skb
, unsigned int flags
)
403 int err
= __sk_queue_drop_skb(sk
, &sk
->sk_receive_queue
, skb
, flags
,
407 sk_mem_reclaim_partial(sk
);
410 EXPORT_SYMBOL(skb_kill_datagram
);
413 * skb_copy_datagram_iter - Copy a datagram to an iovec iterator.
414 * @skb: buffer to copy
415 * @offset: offset in the buffer to start copying from
416 * @to: iovec iterator to copy to
417 * @len: amount of data to copy from buffer to iovec
419 int skb_copy_datagram_iter(const struct sk_buff
*skb
, int offset
,
420 struct iov_iter
*to
, int len
)
422 int start
= skb_headlen(skb
);
423 int i
, copy
= start
- offset
, start_off
= offset
, n
;
424 struct sk_buff
*frag_iter
;
426 trace_skb_copy_datagram_iovec(skb
, len
);
432 n
= copy_to_iter(skb
->data
+ offset
, copy
, to
);
436 if ((len
-= copy
) == 0)
440 /* Copy paged appendix. Hmm... why does this look so complicated? */
441 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
443 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
445 WARN_ON(start
> offset
+ len
);
447 end
= start
+ skb_frag_size(frag
);
448 if ((copy
= end
- offset
) > 0) {
451 n
= copy_page_to_iter(skb_frag_page(frag
),
452 frag
->page_offset
+ offset
-
463 skb_walk_frags(skb
, frag_iter
) {
466 WARN_ON(start
> offset
+ len
);
468 end
= start
+ frag_iter
->len
;
469 if ((copy
= end
- offset
) > 0) {
472 if (skb_copy_datagram_iter(frag_iter
, offset
- start
,
475 if ((len
-= copy
) == 0)
484 /* This is not really a user copy fault, but rather someone
485 * gave us a bogus length on the skb. We should probably
486 * print a warning here as it may indicate a kernel bug.
490 iov_iter_revert(to
, offset
- start_off
);
494 if (iov_iter_count(to
))
499 EXPORT_SYMBOL(skb_copy_datagram_iter
);
502 * skb_copy_datagram_from_iter - Copy a datagram from an iov_iter.
503 * @skb: buffer to copy
504 * @offset: offset in the buffer to start copying to
505 * @from: the copy source
506 * @len: amount of data to copy to buffer from iovec
508 * Returns 0 or -EFAULT.
510 int skb_copy_datagram_from_iter(struct sk_buff
*skb
, int offset
,
511 struct iov_iter
*from
,
514 int start
= skb_headlen(skb
);
515 int i
, copy
= start
- offset
;
516 struct sk_buff
*frag_iter
;
522 if (copy_from_iter(skb
->data
+ offset
, copy
, from
) != copy
)
524 if ((len
-= copy
) == 0)
529 /* Copy paged appendix. Hmm... why does this look so complicated? */
530 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
532 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
534 WARN_ON(start
> offset
+ len
);
536 end
= start
+ skb_frag_size(frag
);
537 if ((copy
= end
- offset
) > 0) {
542 copied
= copy_page_from_iter(skb_frag_page(frag
),
543 frag
->page_offset
+ offset
- start
,
555 skb_walk_frags(skb
, frag_iter
) {
558 WARN_ON(start
> offset
+ len
);
560 end
= start
+ frag_iter
->len
;
561 if ((copy
= end
- offset
) > 0) {
564 if (skb_copy_datagram_from_iter(frag_iter
,
568 if ((len
-= copy
) == 0)
580 EXPORT_SYMBOL(skb_copy_datagram_from_iter
);
583 * zerocopy_sg_from_iter - Build a zerocopy datagram from an iov_iter
584 * @skb: buffer to copy
585 * @from: the source to copy from
587 * The function will first copy up to headlen, and then pin the userspace
588 * pages and build frags through them.
590 * Returns 0, -EFAULT or -EMSGSIZE.
592 int zerocopy_sg_from_iter(struct sk_buff
*skb
, struct iov_iter
*from
)
594 int len
= iov_iter_count(from
);
595 int copy
= min_t(int, skb_headlen(skb
), len
);
598 /* copy up to skb headlen */
599 if (skb_copy_datagram_from_iter(skb
, 0, from
, copy
))
602 while (iov_iter_count(from
)) {
603 struct page
*pages
[MAX_SKB_FRAGS
];
606 unsigned long truesize
;
609 if (frag
== MAX_SKB_FRAGS
)
612 copied
= iov_iter_get_pages(from
, pages
, ~0U,
613 MAX_SKB_FRAGS
- frag
, &start
);
617 iov_iter_advance(from
, copied
);
619 truesize
= PAGE_ALIGN(copied
+ start
);
620 skb
->data_len
+= copied
;
622 skb
->truesize
+= truesize
;
623 refcount_add(truesize
, &skb
->sk
->sk_wmem_alloc
);
625 int size
= min_t(int, copied
, PAGE_SIZE
- start
);
626 skb_fill_page_desc(skb
, frag
++, pages
[n
], start
, size
);
634 EXPORT_SYMBOL(zerocopy_sg_from_iter
);
636 static int skb_copy_and_csum_datagram(const struct sk_buff
*skb
, int offset
,
637 struct iov_iter
*to
, int len
,
640 int start
= skb_headlen(skb
);
641 int i
, copy
= start
- offset
, start_off
= offset
;
642 struct sk_buff
*frag_iter
;
650 n
= csum_and_copy_to_iter(skb
->data
+ offset
, copy
, csump
, to
);
654 if ((len
-= copy
) == 0)
659 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
661 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
663 WARN_ON(start
> offset
+ len
);
665 end
= start
+ skb_frag_size(frag
);
666 if ((copy
= end
- offset
) > 0) {
668 struct page
*page
= skb_frag_page(frag
);
669 u8
*vaddr
= kmap(page
);
673 n
= csum_and_copy_to_iter(vaddr
+ frag
->page_offset
+
674 offset
- start
, copy
,
680 *csump
= csum_block_add(*csump
, csum2
, pos
);
688 skb_walk_frags(skb
, frag_iter
) {
691 WARN_ON(start
> offset
+ len
);
693 end
= start
+ frag_iter
->len
;
694 if ((copy
= end
- offset
) > 0) {
698 if (skb_copy_and_csum_datagram(frag_iter
,
703 *csump
= csum_block_add(*csump
, csum2
, pos
);
704 if ((len
-= copy
) == 0)
715 iov_iter_revert(to
, offset
- start_off
);
719 __sum16
__skb_checksum_complete_head(struct sk_buff
*skb
, int len
)
723 sum
= csum_fold(skb_checksum(skb
, 0, len
, skb
->csum
));
725 if (unlikely(skb
->ip_summed
== CHECKSUM_COMPLETE
) &&
726 !skb
->csum_complete_sw
)
727 netdev_rx_csum_fault(skb
->dev
);
729 if (!skb_shared(skb
))
730 skb
->csum_valid
= !sum
;
733 EXPORT_SYMBOL(__skb_checksum_complete_head
);
735 __sum16
__skb_checksum_complete(struct sk_buff
*skb
)
740 csum
= skb_checksum(skb
, 0, skb
->len
, 0);
742 /* skb->csum holds pseudo checksum */
743 sum
= csum_fold(csum_add(skb
->csum
, csum
));
745 if (unlikely(skb
->ip_summed
== CHECKSUM_COMPLETE
) &&
746 !skb
->csum_complete_sw
)
747 netdev_rx_csum_fault(skb
->dev
);
750 if (!skb_shared(skb
)) {
751 /* Save full packet checksum */
753 skb
->ip_summed
= CHECKSUM_COMPLETE
;
754 skb
->csum_complete_sw
= 1;
755 skb
->csum_valid
= !sum
;
760 EXPORT_SYMBOL(__skb_checksum_complete
);
763 * skb_copy_and_csum_datagram_msg - Copy and checksum skb to user iovec.
765 * @hlen: hardware length
768 * Caller _must_ check that skb will fit to this iovec.
770 * Returns: 0 - success.
771 * -EINVAL - checksum failure.
772 * -EFAULT - fault during copy.
774 int skb_copy_and_csum_datagram_msg(struct sk_buff
*skb
,
775 int hlen
, struct msghdr
*msg
)
778 int chunk
= skb
->len
- hlen
;
783 if (msg_data_left(msg
) < chunk
) {
784 if (__skb_checksum_complete(skb
))
786 if (skb_copy_datagram_msg(skb
, hlen
, msg
, chunk
))
789 csum
= csum_partial(skb
->data
, hlen
, skb
->csum
);
790 if (skb_copy_and_csum_datagram(skb
, hlen
, &msg
->msg_iter
,
794 if (csum_fold(csum
)) {
795 iov_iter_revert(&msg
->msg_iter
, chunk
);
799 if (unlikely(skb
->ip_summed
== CHECKSUM_COMPLETE
))
800 netdev_rx_csum_fault(skb
->dev
);
806 EXPORT_SYMBOL(skb_copy_and_csum_datagram_msg
);
809 * datagram_poll - generic datagram poll
814 * Datagram poll: Again totally generic. This also handles
815 * sequenced packet sockets providing the socket receive queue
816 * is only ever holding data ready to receive.
818 * Note: when you *don't* use this routine for this protocol,
819 * and you use a different write policy from sock_writeable()
820 * then please supply your own write_space callback.
822 unsigned int datagram_poll(struct file
*file
, struct socket
*sock
,
825 struct sock
*sk
= sock
->sk
;
828 sock_poll_wait(file
, sk_sleep(sk
), wait
);
831 /* exceptional events? */
832 if (sk
->sk_err
|| !skb_queue_empty(&sk
->sk_error_queue
))
834 (sock_flag(sk
, SOCK_SELECT_ERR_QUEUE
) ? POLLPRI
: 0);
836 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
837 mask
|= POLLRDHUP
| POLLIN
| POLLRDNORM
;
838 if (sk
->sk_shutdown
== SHUTDOWN_MASK
)
842 if (!skb_queue_empty(&sk
->sk_receive_queue
))
843 mask
|= POLLIN
| POLLRDNORM
;
845 /* Connection-based need to check for termination and startup */
846 if (connection_based(sk
)) {
847 if (sk
->sk_state
== TCP_CLOSE
)
849 /* connection hasn't started yet? */
850 if (sk
->sk_state
== TCP_SYN_SENT
)
855 if (sock_writeable(sk
))
856 mask
|= POLLOUT
| POLLWRNORM
| POLLWRBAND
;
858 sk_set_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
862 EXPORT_SYMBOL(datagram_poll
);