1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel Connection Multiplexor
5 * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
9 #include <linux/errno.h>
10 #include <linux/errqueue.h>
11 #include <linux/file.h>
12 #include <linux/filter.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/net.h>
17 #include <linux/netdevice.h>
18 #include <linux/poll.h>
19 #include <linux/rculist.h>
20 #include <linux/skbuff.h>
21 #include <linux/socket.h>
22 #include <linux/uaccess.h>
23 #include <linux/workqueue.h>
24 #include <linux/syscalls.h>
25 #include <linux/sched/signal.h>
28 #include <net/netns/generic.h>
30 #include <uapi/linux/kcm.h>
31 #include <trace/events/sock.h>
33 unsigned int kcm_net_id;
35 static struct kmem_cache *kcm_psockp __read_mostly;
36 static struct kmem_cache *kcm_muxp __read_mostly;
37 static struct workqueue_struct *kcm_wq;
39 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
41 return (struct kcm_sock *)sk;
44 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
46 return (struct kcm_tx_msg *)skb->cb;
49 static void report_csk_error(struct sock *csk, int err)
55 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
58 struct sock *csk = psock->sk;
59 struct kcm_mux *mux = psock->mux;
61 /* Unrecoverable error in transmit */
63 spin_lock_bh(&mux->lock);
65 if (psock->tx_stopped) {
66 spin_unlock_bh(&mux->lock);
70 psock->tx_stopped = 1;
71 KCM_STATS_INCR(psock->stats.tx_aborts);
74 /* Take off psocks_avail list */
75 list_del(&psock->psock_avail_list);
76 } else if (wakeup_kcm) {
77 /* In this case psock is being aborted while outside of
78 * write_msgs and psock is reserved. Schedule tx_work
79 * to handle the failure there. Need to commit tx_stopped
80 * before queuing work.
84 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
87 spin_unlock_bh(&mux->lock);
89 /* Report error on lower socket */
90 report_csk_error(csk, err);
93 /* RX mux lock held. */
94 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
95 struct kcm_psock *psock)
97 STRP_STATS_ADD(mux->stats.rx_bytes,
98 psock->strp.stats.bytes -
99 psock->saved_rx_bytes);
100 mux->stats.rx_msgs +=
101 psock->strp.stats.msgs - psock->saved_rx_msgs;
102 psock->saved_rx_msgs = psock->strp.stats.msgs;
103 psock->saved_rx_bytes = psock->strp.stats.bytes;
106 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
107 struct kcm_psock *psock)
109 KCM_STATS_ADD(mux->stats.tx_bytes,
110 psock->stats.tx_bytes - psock->saved_tx_bytes);
111 mux->stats.tx_msgs +=
112 psock->stats.tx_msgs - psock->saved_tx_msgs;
113 psock->saved_tx_msgs = psock->stats.tx_msgs;
114 psock->saved_tx_bytes = psock->stats.tx_bytes;
117 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
119 /* KCM is ready to receive messages on its queue-- either the KCM is new or
120 * has become unblocked after being blocked on full socket buffer. Queue any
121 * pending ready messages on a psock. RX mux lock held.
123 static void kcm_rcv_ready(struct kcm_sock *kcm)
125 struct kcm_mux *mux = kcm->mux;
126 struct kcm_psock *psock;
129 if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
132 while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
133 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
134 /* Assuming buffer limit has been reached */
135 skb_queue_head(&mux->rx_hold_queue, skb);
136 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
141 while (!list_empty(&mux->psocks_ready)) {
142 psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
145 if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
146 /* Assuming buffer limit has been reached */
147 WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
151 /* Consumed the ready message on the psock. Schedule rx_work to
154 list_del(&psock->psock_ready_list);
155 psock->ready_rx_msg = NULL;
156 /* Commit clearing of ready_rx_msg for queuing work */
159 strp_unpause(&psock->strp);
160 strp_check_rcv(&psock->strp);
163 /* Buffer limit is okay now, add to ready list */
164 list_add_tail(&kcm->wait_rx_list,
165 &kcm->mux->kcm_rx_waiters);
166 /* paired with lockless reads in kcm_rfree() */
167 WRITE_ONCE(kcm->rx_wait, true);
170 static void kcm_rfree(struct sk_buff *skb)
172 struct sock *sk = skb->sk;
173 struct kcm_sock *kcm = kcm_sk(sk);
174 struct kcm_mux *mux = kcm->mux;
175 unsigned int len = skb->truesize;
177 sk_mem_uncharge(sk, len);
178 atomic_sub(len, &sk->sk_rmem_alloc);
180 /* For reading rx_wait and rx_psock without holding lock */
181 smp_mb__after_atomic();
183 if (!READ_ONCE(kcm->rx_wait) && !READ_ONCE(kcm->rx_psock) &&
184 sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
185 spin_lock_bh(&mux->rx_lock);
187 spin_unlock_bh(&mux->rx_lock);
191 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
193 struct sk_buff_head *list = &sk->sk_receive_queue;
195 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
198 if (!sk_rmem_schedule(sk, skb, skb->truesize))
205 skb->destructor = kcm_rfree;
206 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
207 sk_mem_charge(sk, skb->truesize);
209 skb_queue_tail(list, skb);
211 if (!sock_flag(sk, SOCK_DEAD))
212 sk->sk_data_ready(sk);
217 /* Requeue received messages for a kcm socket to other kcm sockets. This is
218 * called with a kcm socket is receive disabled.
221 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
224 struct kcm_sock *kcm;
226 while ((skb = skb_dequeue(head))) {
227 /* Reset destructor to avoid calling kcm_rcv_ready */
228 skb->destructor = sock_rfree;
231 if (list_empty(&mux->kcm_rx_waiters)) {
232 skb_queue_tail(&mux->rx_hold_queue, skb);
236 kcm = list_first_entry(&mux->kcm_rx_waiters,
237 struct kcm_sock, wait_rx_list);
239 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
240 /* Should mean socket buffer full */
241 list_del(&kcm->wait_rx_list);
242 /* paired with lockless reads in kcm_rfree() */
243 WRITE_ONCE(kcm->rx_wait, false);
245 /* Commit rx_wait to read in kcm_free */
253 /* Lower sock lock held */
254 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
255 struct sk_buff *head)
257 struct kcm_mux *mux = psock->mux;
258 struct kcm_sock *kcm;
260 WARN_ON(psock->ready_rx_msg);
263 return psock->rx_kcm;
265 spin_lock_bh(&mux->rx_lock);
268 spin_unlock_bh(&mux->rx_lock);
269 return psock->rx_kcm;
272 kcm_update_rx_mux_stats(mux, psock);
274 if (list_empty(&mux->kcm_rx_waiters)) {
275 psock->ready_rx_msg = head;
276 strp_pause(&psock->strp);
277 list_add_tail(&psock->psock_ready_list,
279 spin_unlock_bh(&mux->rx_lock);
283 kcm = list_first_entry(&mux->kcm_rx_waiters,
284 struct kcm_sock, wait_rx_list);
285 list_del(&kcm->wait_rx_list);
286 /* paired with lockless reads in kcm_rfree() */
287 WRITE_ONCE(kcm->rx_wait, false);
290 /* paired with lockless reads in kcm_rfree() */
291 WRITE_ONCE(kcm->rx_psock, psock);
293 spin_unlock_bh(&mux->rx_lock);
298 static void kcm_done(struct kcm_sock *kcm);
300 static void kcm_done_work(struct work_struct *w)
302 kcm_done(container_of(w, struct kcm_sock, done_work));
305 /* Lower sock held */
306 static void unreserve_rx_kcm(struct kcm_psock *psock,
309 struct kcm_sock *kcm = psock->rx_kcm;
310 struct kcm_mux *mux = psock->mux;
315 spin_lock_bh(&mux->rx_lock);
317 psock->rx_kcm = NULL;
318 /* paired with lockless reads in kcm_rfree() */
319 WRITE_ONCE(kcm->rx_psock, NULL);
321 /* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
326 if (unlikely(kcm->done)) {
327 spin_unlock_bh(&mux->rx_lock);
329 /* Need to run kcm_done in a task since we need to qcquire
330 * callback locks which may already be held here.
332 INIT_WORK(&kcm->done_work, kcm_done_work);
333 schedule_work(&kcm->done_work);
337 if (unlikely(kcm->rx_disabled)) {
338 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
339 } else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
340 /* Check for degenerative race with rx_wait that all
341 * data was dequeued (accounted for in kcm_rfree).
345 spin_unlock_bh(&mux->rx_lock);
348 /* Lower sock lock held */
349 static void psock_data_ready(struct sock *sk)
351 struct kcm_psock *psock;
353 trace_sk_data_ready(sk);
355 read_lock_bh(&sk->sk_callback_lock);
357 psock = (struct kcm_psock *)sk->sk_user_data;
359 strp_data_ready(&psock->strp);
361 read_unlock_bh(&sk->sk_callback_lock);
364 /* Called with lower sock held */
365 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
367 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
368 struct kcm_sock *kcm;
371 kcm = reserve_rx_kcm(psock, skb);
373 /* Unable to reserve a KCM, message is held in psock and strp
379 if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
380 /* Should mean socket buffer full */
381 unreserve_rx_kcm(psock, false);
386 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
388 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
389 struct bpf_prog *prog = psock->bpf_prog;
392 res = bpf_prog_run_pin_on_cpu(prog, skb);
396 static int kcm_read_sock_done(struct strparser *strp, int err)
398 struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
400 unreserve_rx_kcm(psock, true);
405 static void psock_state_change(struct sock *sk)
407 /* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
408 * since application will normally not poll with EPOLLIN
409 * on the TCP sockets.
412 report_csk_error(sk, EPIPE);
415 static void psock_write_space(struct sock *sk)
417 struct kcm_psock *psock;
419 struct kcm_sock *kcm;
421 read_lock_bh(&sk->sk_callback_lock);
423 psock = (struct kcm_psock *)sk->sk_user_data;
424 if (unlikely(!psock))
428 spin_lock_bh(&mux->lock);
430 /* Check if the socket is reserved so someone is waiting for sending. */
432 if (kcm && !unlikely(kcm->tx_stopped))
433 queue_work(kcm_wq, &kcm->tx_work);
435 spin_unlock_bh(&mux->lock);
437 read_unlock_bh(&sk->sk_callback_lock);
440 static void unreserve_psock(struct kcm_sock *kcm);
442 /* kcm sock is locked. */
443 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
445 struct kcm_mux *mux = kcm->mux;
446 struct kcm_psock *psock;
448 psock = kcm->tx_psock;
450 smp_rmb(); /* Must read tx_psock before tx_wait */
453 WARN_ON(kcm->tx_wait);
454 if (unlikely(psock->tx_stopped))
455 unreserve_psock(kcm);
457 return kcm->tx_psock;
460 spin_lock_bh(&mux->lock);
462 /* Check again under lock to see if psock was reserved for this
463 * psock via psock_unreserve.
465 psock = kcm->tx_psock;
466 if (unlikely(psock)) {
467 WARN_ON(kcm->tx_wait);
468 spin_unlock_bh(&mux->lock);
469 return kcm->tx_psock;
472 if (!list_empty(&mux->psocks_avail)) {
473 psock = list_first_entry(&mux->psocks_avail,
476 list_del(&psock->psock_avail_list);
478 list_del(&kcm->wait_psock_list);
479 kcm->tx_wait = false;
481 kcm->tx_psock = psock;
483 KCM_STATS_INCR(psock->stats.reserved);
484 } else if (!kcm->tx_wait) {
485 list_add_tail(&kcm->wait_psock_list,
486 &mux->kcm_tx_waiters);
490 spin_unlock_bh(&mux->lock);
496 static void psock_now_avail(struct kcm_psock *psock)
498 struct kcm_mux *mux = psock->mux;
499 struct kcm_sock *kcm;
501 if (list_empty(&mux->kcm_tx_waiters)) {
502 list_add_tail(&psock->psock_avail_list,
505 kcm = list_first_entry(&mux->kcm_tx_waiters,
508 list_del(&kcm->wait_psock_list);
509 kcm->tx_wait = false;
512 /* Commit before changing tx_psock since that is read in
513 * reserve_psock before queuing work.
517 kcm->tx_psock = psock;
518 KCM_STATS_INCR(psock->stats.reserved);
519 queue_work(kcm_wq, &kcm->tx_work);
523 /* kcm sock is locked. */
524 static void unreserve_psock(struct kcm_sock *kcm)
526 struct kcm_psock *psock;
527 struct kcm_mux *mux = kcm->mux;
529 spin_lock_bh(&mux->lock);
531 psock = kcm->tx_psock;
533 if (WARN_ON(!psock)) {
534 spin_unlock_bh(&mux->lock);
538 smp_rmb(); /* Read tx_psock before tx_wait */
540 kcm_update_tx_mux_stats(mux, psock);
542 WARN_ON(kcm->tx_wait);
544 kcm->tx_psock = NULL;
545 psock->tx_kcm = NULL;
546 KCM_STATS_INCR(psock->stats.unreserved);
548 if (unlikely(psock->tx_stopped)) {
551 list_del(&psock->psock_list);
554 fput(psock->sk->sk_socket->file);
555 kmem_cache_free(kcm_psockp, psock);
558 /* Don't put back on available list */
560 spin_unlock_bh(&mux->lock);
565 psock_now_avail(psock);
567 spin_unlock_bh(&mux->lock);
570 static void kcm_report_tx_retry(struct kcm_sock *kcm)
572 struct kcm_mux *mux = kcm->mux;
574 spin_lock_bh(&mux->lock);
575 KCM_STATS_INCR(mux->stats.tx_retries);
576 spin_unlock_bh(&mux->lock);
579 /* Write any messages ready on the kcm socket. Called with kcm sock lock
580 * held. Return bytes actually sent or error.
582 static int kcm_write_msgs(struct kcm_sock *kcm)
584 unsigned int total_sent = 0;
585 struct sock *sk = &kcm->sk;
586 struct kcm_psock *psock;
587 struct sk_buff *head;
590 kcm->tx_wait_more = false;
591 psock = kcm->tx_psock;
592 if (unlikely(psock && psock->tx_stopped)) {
593 /* A reserved psock was aborted asynchronously. Unreserve
594 * it and we'll retry the message.
596 unreserve_psock(kcm);
597 kcm_report_tx_retry(kcm);
598 if (skb_queue_empty(&sk->sk_write_queue))
601 kcm_tx_msg(skb_peek(&sk->sk_write_queue))->started_tx = false;
605 while ((head = skb_peek(&sk->sk_write_queue))) {
606 struct msghdr msg = {
607 .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
609 struct kcm_tx_msg *txm = kcm_tx_msg(head);
614 if (!txm->started_tx) {
615 psock = reserve_psock(kcm);
619 txm->frag_offset = 0;
621 txm->started_tx = true;
623 if (WARN_ON(!psock)) {
630 if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
636 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
637 msize += skb_shinfo(skb)->frags[i].bv_len;
639 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE,
640 skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags,
642 iov_iter_advance(&msg.msg_iter, txm->frag_offset);
645 ret = sock_sendmsg(psock->sk->sk_socket, &msg);
647 if (ret == -EAGAIN) {
648 /* Save state to try again when there's
649 * write space on the socket
656 /* Hard failure in sending message, abort this
657 * psock since it has lost framing
658 * synchronization and retry sending the
659 * message from the beginning.
661 kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
663 unreserve_psock(kcm);
666 txm->started_tx = false;
667 kcm_report_tx_retry(kcm);
673 txm->frag_offset += ret;
674 KCM_STATS_ADD(psock->stats.tx_bytes, ret);
675 } while (msg.msg_iter.count > 0);
678 if (skb_has_frag_list(skb)) {
679 txm->frag_skb = skb_shinfo(skb)->frag_list;
680 txm->frag_offset = 0;
683 } else if (skb->next) {
684 txm->frag_skb = skb->next;
685 txm->frag_offset = 0;
689 /* Successfully sent the whole packet, account for it. */
690 sk->sk_wmem_queued -= txm->sent;
691 total_sent += txm->sent;
692 skb_dequeue(&sk->sk_write_queue);
694 KCM_STATS_INCR(psock->stats.tx_msgs);
698 /* Done with all queued messages. */
699 WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
701 unreserve_psock(kcm);
704 /* Check if write space is available */
705 sk->sk_write_space(sk);
707 return total_sent ? : ret;
710 static void kcm_tx_work(struct work_struct *w)
712 struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
713 struct sock *sk = &kcm->sk;
718 /* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
721 err = kcm_write_msgs(kcm);
723 /* Hard failure in write, report error on KCM socket */
724 pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
725 report_csk_error(&kcm->sk, -err);
729 /* Primarily for SOCK_SEQPACKET sockets */
730 if (likely(sk->sk_socket) &&
731 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
732 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
733 sk->sk_write_space(sk);
740 static void kcm_push(struct kcm_sock *kcm)
742 if (kcm->tx_wait_more)
746 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
748 struct sock *sk = sock->sk;
749 struct kcm_sock *kcm = kcm_sk(sk);
750 struct sk_buff *skb = NULL, *head = NULL;
751 size_t copy, copied = 0;
752 long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
753 int eor = (sock->type == SOCK_DGRAM) ?
754 !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
759 /* Per tcp_sendmsg this should be in poll */
760 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
766 /* Previously opened message */
768 skb = kcm_tx_msg(head)->last_skb;
772 /* Call the sk_stream functions to manage the sndbuf mem. */
773 if (!sk_stream_memory_free(sk)) {
775 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
776 err = sk_stream_wait_memory(sk, &timeo);
781 if (msg_data_left(msg)) {
782 /* New message, alloc head skb */
783 head = alloc_skb(0, sk->sk_allocation);
786 err = sk_stream_wait_memory(sk, &timeo);
790 head = alloc_skb(0, sk->sk_allocation);
795 /* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
796 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
798 skb->ip_summed = CHECKSUM_UNNECESSARY;
802 while (msg_data_left(msg)) {
804 int i = skb_shinfo(skb)->nr_frags;
805 struct page_frag *pfrag = sk_page_frag(sk);
807 if (!sk_page_frag_refill(sk, pfrag))
808 goto wait_for_memory;
810 if (!skb_can_coalesce(skb, i, pfrag->page,
812 if (i == MAX_SKB_FRAGS) {
813 struct sk_buff *tskb;
815 tskb = alloc_skb(0, sk->sk_allocation);
817 goto wait_for_memory;
820 skb_shinfo(head)->frag_list = tskb;
825 skb->ip_summed = CHECKSUM_UNNECESSARY;
831 if (msg->msg_flags & MSG_SPLICE_PAGES) {
832 copy = msg_data_left(msg);
833 if (!sk_wmem_schedule(sk, copy))
834 goto wait_for_memory;
836 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
839 if (err == -EMSGSIZE)
840 goto wait_for_memory;
845 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
846 sk_wmem_queued_add(sk, copy);
847 sk_mem_charge(sk, copy);
850 head->truesize += copy;
852 copy = min_t(int, msg_data_left(msg),
853 pfrag->size - pfrag->offset);
854 if (!sk_wmem_schedule(sk, copy))
855 goto wait_for_memory;
857 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
864 /* Update the skb. */
867 &skb_shinfo(skb)->frags[i - 1], copy);
869 skb_fill_page_desc(skb, i, pfrag->page,
870 pfrag->offset, copy);
871 get_page(pfrag->page);
874 pfrag->offset += copy;
880 head->data_len += copy;
887 err = sk_stream_wait_memory(sk, &timeo);
893 bool not_busy = skb_queue_empty(&sk->sk_write_queue);
896 /* Message complete, queue it on send buffer */
897 __skb_queue_tail(&sk->sk_write_queue, head);
899 KCM_STATS_INCR(kcm->stats.tx_msgs);
902 if (msg->msg_flags & MSG_BATCH) {
903 kcm->tx_wait_more = true;
904 } else if (kcm->tx_wait_more || not_busy) {
905 err = kcm_write_msgs(kcm);
907 /* We got a hard error in write_msgs but have
908 * already queued this message. Report an error
909 * in the socket, but don't affect return value
912 pr_warn("KCM: Hard failure on kcm_write_msgs\n");
913 report_csk_error(&kcm->sk, -err);
917 /* Message not complete, save state */
921 kcm_tx_msg(head)->last_skb = skb;
925 KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
933 if (sock->type == SOCK_SEQPACKET) {
934 /* Wrote some bytes before encountering an
935 * error, return partial success.
938 goto partial_message;
939 if (head != kcm->seq_skb)
946 err = sk_stream_error(sk, msg->msg_flags, err);
948 /* make sure we wake any epoll edge trigger waiter */
949 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
950 sk->sk_write_space(sk);
956 static void kcm_splice_eof(struct socket *sock)
958 struct sock *sk = sock->sk;
959 struct kcm_sock *kcm = kcm_sk(sk);
961 if (skb_queue_empty_lockless(&sk->sk_write_queue))
969 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
970 size_t len, int flags)
972 struct sock *sk = sock->sk;
973 struct kcm_sock *kcm = kcm_sk(sk);
975 struct strp_msg *stm;
979 skb = skb_recv_datagram(sk, flags, &err);
983 /* Okay, have a message on the receive queue */
987 if (len > stm->full_len)
990 err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
995 if (likely(!(flags & MSG_PEEK))) {
996 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
997 if (copied < stm->full_len) {
998 if (sock->type == SOCK_DGRAM) {
999 /* Truncated message */
1000 msg->msg_flags |= MSG_TRUNC;
1003 stm->offset += copied;
1004 stm->full_len -= copied;
1007 /* Finished with message */
1008 msg->msg_flags |= MSG_EOR;
1009 KCM_STATS_INCR(kcm->stats.rx_msgs);
1014 skb_free_datagram(sk, skb);
1015 return copied ? : err;
1018 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1019 struct pipe_inode_info *pipe, size_t len,
1022 struct sock *sk = sock->sk;
1023 struct kcm_sock *kcm = kcm_sk(sk);
1024 struct strp_msg *stm;
1027 struct sk_buff *skb;
1029 /* Only support splice for SOCKSEQPACKET */
1031 skb = skb_recv_datagram(sk, flags, &err);
1035 /* Okay, have a message on the receive queue */
1037 stm = strp_msg(skb);
1039 if (len > stm->full_len)
1040 len = stm->full_len;
1042 copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1048 KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1050 stm->offset += copied;
1051 stm->full_len -= copied;
1053 /* We have no way to return MSG_EOR. If all the bytes have been
1054 * read we still leave the message in the receive socket buffer.
1055 * A subsequent recvmsg needs to be done to return MSG_EOR and
1056 * finish reading the message.
1059 skb_free_datagram(sk, skb);
1063 skb_free_datagram(sk, skb);
1067 /* kcm sock lock held */
1068 static void kcm_recv_disable(struct kcm_sock *kcm)
1070 struct kcm_mux *mux = kcm->mux;
1072 if (kcm->rx_disabled)
1075 spin_lock_bh(&mux->rx_lock);
1077 kcm->rx_disabled = 1;
1079 /* If a psock is reserved we'll do cleanup in unreserve */
1080 if (!kcm->rx_psock) {
1082 list_del(&kcm->wait_rx_list);
1083 /* paired with lockless reads in kcm_rfree() */
1084 WRITE_ONCE(kcm->rx_wait, false);
1087 requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1090 spin_unlock_bh(&mux->rx_lock);
1093 /* kcm sock lock held */
1094 static void kcm_recv_enable(struct kcm_sock *kcm)
1096 struct kcm_mux *mux = kcm->mux;
1098 if (!kcm->rx_disabled)
1101 spin_lock_bh(&mux->rx_lock);
1103 kcm->rx_disabled = 0;
1106 spin_unlock_bh(&mux->rx_lock);
1109 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1110 sockptr_t optval, unsigned int optlen)
1112 struct kcm_sock *kcm = kcm_sk(sock->sk);
1116 if (level != SOL_KCM)
1117 return -ENOPROTOOPT;
1119 if (optlen < sizeof(int))
1122 if (copy_from_sockptr(&val, optval, sizeof(int)))
1125 valbool = val ? 1 : 0;
1128 case KCM_RECV_DISABLE:
1129 lock_sock(&kcm->sk);
1131 kcm_recv_disable(kcm);
1133 kcm_recv_enable(kcm);
1134 release_sock(&kcm->sk);
1143 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1144 char __user *optval, int __user *optlen)
1146 struct kcm_sock *kcm = kcm_sk(sock->sk);
1149 if (level != SOL_KCM)
1150 return -ENOPROTOOPT;
1152 if (get_user(len, optlen))
1155 len = min_t(unsigned int, len, sizeof(int));
1160 case KCM_RECV_DISABLE:
1161 val = kcm->rx_disabled;
1164 return -ENOPROTOOPT;
1167 if (put_user(len, optlen))
1169 if (copy_to_user(optval, &val, len))
1174 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1176 struct kcm_sock *tkcm;
1177 struct list_head *head;
1180 /* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1181 * we set sk_state, otherwise epoll_wait always returns right away with
1184 kcm->sk.sk_state = TCP_ESTABLISHED;
1186 /* Add to mux's kcm sockets list */
1188 spin_lock_bh(&mux->lock);
1190 head = &mux->kcm_socks;
1191 list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1192 if (tkcm->index != index)
1194 head = &tkcm->kcm_sock_list;
1198 list_add(&kcm->kcm_sock_list, head);
1201 mux->kcm_socks_cnt++;
1202 spin_unlock_bh(&mux->lock);
1204 INIT_WORK(&kcm->tx_work, kcm_tx_work);
1206 spin_lock_bh(&mux->rx_lock);
1208 spin_unlock_bh(&mux->rx_lock);
1211 static int kcm_attach(struct socket *sock, struct socket *csock,
1212 struct bpf_prog *prog)
1214 struct kcm_sock *kcm = kcm_sk(sock->sk);
1215 struct kcm_mux *mux = kcm->mux;
1217 struct kcm_psock *psock = NULL, *tpsock;
1218 struct list_head *head;
1220 static const struct strp_callbacks cb = {
1221 .rcv_msg = kcm_rcv_strparser,
1222 .parse_msg = kcm_parse_func_strparser,
1223 .read_sock_done = kcm_read_sock_done,
1233 /* Only allow TCP sockets to be attached for now */
1234 if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1235 csk->sk_protocol != IPPROTO_TCP) {
1240 /* Don't allow listeners or closed sockets */
1241 if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1246 psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1254 psock->bpf_prog = prog;
1256 write_lock_bh(&csk->sk_callback_lock);
1258 /* Check if sk_user_data is already by KCM or someone else.
1259 * Must be done under lock to prevent race conditions.
1261 if (csk->sk_user_data) {
1262 write_unlock_bh(&csk->sk_callback_lock);
1263 kmem_cache_free(kcm_psockp, psock);
1268 err = strp_init(&psock->strp, csk, &cb);
1270 write_unlock_bh(&csk->sk_callback_lock);
1271 kmem_cache_free(kcm_psockp, psock);
1275 psock->save_data_ready = csk->sk_data_ready;
1276 psock->save_write_space = csk->sk_write_space;
1277 psock->save_state_change = csk->sk_state_change;
1278 csk->sk_user_data = psock;
1279 csk->sk_data_ready = psock_data_ready;
1280 csk->sk_write_space = psock_write_space;
1281 csk->sk_state_change = psock_state_change;
1283 write_unlock_bh(&csk->sk_callback_lock);
1287 /* Finished initialization, now add the psock to the MUX. */
1288 spin_lock_bh(&mux->lock);
1289 head = &mux->psocks;
1290 list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1291 if (tpsock->index != index)
1293 head = &tpsock->psock_list;
1297 list_add(&psock->psock_list, head);
1298 psock->index = index;
1300 KCM_STATS_INCR(mux->stats.psock_attach);
1302 psock_now_avail(psock);
1303 spin_unlock_bh(&mux->lock);
1305 /* Schedule RX work in case there are already bytes queued */
1306 strp_check_rcv(&psock->strp);
1314 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1316 struct socket *csock;
1317 struct bpf_prog *prog;
1320 csock = sockfd_lookup(info->fd, &err);
1324 prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1326 err = PTR_ERR(prog);
1330 err = kcm_attach(sock, csock, prog);
1336 /* Keep reference on file also */
1344 static void kcm_unattach(struct kcm_psock *psock)
1346 struct sock *csk = psock->sk;
1347 struct kcm_mux *mux = psock->mux;
1351 /* Stop getting callbacks from TCP socket. After this there should
1352 * be no way to reserve a kcm for this psock.
1354 write_lock_bh(&csk->sk_callback_lock);
1355 csk->sk_user_data = NULL;
1356 csk->sk_data_ready = psock->save_data_ready;
1357 csk->sk_write_space = psock->save_write_space;
1358 csk->sk_state_change = psock->save_state_change;
1359 strp_stop(&psock->strp);
1361 if (WARN_ON(psock->rx_kcm)) {
1362 write_unlock_bh(&csk->sk_callback_lock);
1367 spin_lock_bh(&mux->rx_lock);
1369 /* Stop receiver activities. After this point psock should not be
1370 * able to get onto ready list either through callbacks or work.
1372 if (psock->ready_rx_msg) {
1373 list_del(&psock->psock_ready_list);
1374 kfree_skb(psock->ready_rx_msg);
1375 psock->ready_rx_msg = NULL;
1376 KCM_STATS_INCR(mux->stats.rx_ready_drops);
1379 spin_unlock_bh(&mux->rx_lock);
1381 write_unlock_bh(&csk->sk_callback_lock);
1383 /* Call strp_done without sock lock */
1385 strp_done(&psock->strp);
1388 bpf_prog_put(psock->bpf_prog);
1390 spin_lock_bh(&mux->lock);
1392 aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1393 save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1395 KCM_STATS_INCR(mux->stats.psock_unattach);
1397 if (psock->tx_kcm) {
1398 /* psock was reserved. Just mark it finished and we will clean
1399 * up in the kcm paths, we need kcm lock which can not be
1402 KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1403 spin_unlock_bh(&mux->lock);
1405 /* We are unattaching a socket that is reserved. Abort the
1406 * socket since we may be out of sync in sending on it. We need
1407 * to do this without the mux lock.
1409 kcm_abort_tx_psock(psock, EPIPE, false);
1411 spin_lock_bh(&mux->lock);
1412 if (!psock->tx_kcm) {
1413 /* psock now unreserved in window mux was unlocked */
1418 /* Commit done before queuing work to process it */
1421 /* Queue tx work to make sure psock->done is handled */
1422 queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1423 spin_unlock_bh(&mux->lock);
1426 if (!psock->tx_stopped)
1427 list_del(&psock->psock_avail_list);
1428 list_del(&psock->psock_list);
1430 spin_unlock_bh(&mux->lock);
1433 fput(csk->sk_socket->file);
1434 kmem_cache_free(kcm_psockp, psock);
1440 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1442 struct kcm_sock *kcm = kcm_sk(sock->sk);
1443 struct kcm_mux *mux = kcm->mux;
1444 struct kcm_psock *psock;
1445 struct socket *csock;
1449 csock = sockfd_lookup(info->fd, &err);
1461 spin_lock_bh(&mux->lock);
1463 list_for_each_entry(psock, &mux->psocks, psock_list) {
1464 if (psock->sk != csk)
1467 /* Found the matching psock */
1469 if (psock->unattaching || WARN_ON(psock->done)) {
1474 psock->unattaching = 1;
1476 spin_unlock_bh(&mux->lock);
1478 /* Lower socket lock should already be held */
1479 kcm_unattach(psock);
1485 spin_unlock_bh(&mux->lock);
1492 static struct proto kcm_proto = {
1494 .owner = THIS_MODULE,
1495 .obj_size = sizeof(struct kcm_sock),
1498 /* Clone a kcm socket. */
1499 static struct file *kcm_clone(struct socket *osock)
1501 struct socket *newsock;
1504 newsock = sock_alloc();
1506 return ERR_PTR(-ENFILE);
1508 newsock->type = osock->type;
1509 newsock->ops = osock->ops;
1511 __module_get(newsock->ops->owner);
1513 newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1516 sock_release(newsock);
1517 return ERR_PTR(-ENOMEM);
1519 sock_init_data(newsock, newsk);
1520 init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1522 return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1525 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1530 case SIOCKCMATTACH: {
1531 struct kcm_attach info;
1533 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1536 err = kcm_attach_ioctl(sock, &info);
1540 case SIOCKCMUNATTACH: {
1541 struct kcm_unattach info;
1543 if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1546 err = kcm_unattach_ioctl(sock, &info);
1550 case SIOCKCMCLONE: {
1551 struct kcm_clone info;
1554 info.fd = get_unused_fd_flags(0);
1555 if (unlikely(info.fd < 0))
1558 file = kcm_clone(sock);
1560 put_unused_fd(info.fd);
1561 return PTR_ERR(file);
1563 if (copy_to_user((void __user *)arg, &info,
1565 put_unused_fd(info.fd);
1569 fd_install(info.fd, file);
1581 static void free_mux(struct rcu_head *rcu)
1583 struct kcm_mux *mux = container_of(rcu,
1584 struct kcm_mux, rcu);
1586 kmem_cache_free(kcm_muxp, mux);
1589 static void release_mux(struct kcm_mux *mux)
1591 struct kcm_net *knet = mux->knet;
1592 struct kcm_psock *psock, *tmp_psock;
1594 /* Release psocks */
1595 list_for_each_entry_safe(psock, tmp_psock,
1596 &mux->psocks, psock_list) {
1597 if (!WARN_ON(psock->unattaching))
1598 kcm_unattach(psock);
1601 if (WARN_ON(mux->psocks_cnt))
1604 __skb_queue_purge(&mux->rx_hold_queue);
1606 mutex_lock(&knet->mutex);
1607 aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1608 aggregate_psock_stats(&mux->aggregate_psock_stats,
1609 &knet->aggregate_psock_stats);
1610 aggregate_strp_stats(&mux->aggregate_strp_stats,
1611 &knet->aggregate_strp_stats);
1612 list_del_rcu(&mux->kcm_mux_list);
1614 mutex_unlock(&knet->mutex);
1616 call_rcu(&mux->rcu, free_mux);
1619 static void kcm_done(struct kcm_sock *kcm)
1621 struct kcm_mux *mux = kcm->mux;
1622 struct sock *sk = &kcm->sk;
1625 spin_lock_bh(&mux->rx_lock);
1626 if (kcm->rx_psock) {
1627 /* Cleanup in unreserve_rx_kcm */
1629 kcm->rx_disabled = 1;
1631 spin_unlock_bh(&mux->rx_lock);
1636 list_del(&kcm->wait_rx_list);
1637 /* paired with lockless reads in kcm_rfree() */
1638 WRITE_ONCE(kcm->rx_wait, false);
1640 /* Move any pending receive messages to other kcm sockets */
1641 requeue_rx_msgs(mux, &sk->sk_receive_queue);
1643 spin_unlock_bh(&mux->rx_lock);
1645 if (WARN_ON(sk_rmem_alloc_get(sk)))
1648 /* Detach from MUX */
1649 spin_lock_bh(&mux->lock);
1651 list_del(&kcm->kcm_sock_list);
1652 mux->kcm_socks_cnt--;
1653 socks_cnt = mux->kcm_socks_cnt;
1655 spin_unlock_bh(&mux->lock);
1658 /* We are done with the mux now. */
1662 WARN_ON(kcm->rx_wait);
1667 /* Called by kcm_release to close a KCM socket.
1668 * If this is the last KCM socket on the MUX, destroy the MUX.
1670 static int kcm_release(struct socket *sock)
1672 struct sock *sk = sock->sk;
1673 struct kcm_sock *kcm;
1674 struct kcm_mux *mux;
1675 struct kcm_psock *psock;
1685 kfree_skb(kcm->seq_skb);
1687 /* Purge queue under lock to avoid race condition with tx_work trying
1688 * to act when queue is nonempty. If tx_work runs after this point
1689 * it will just return.
1691 __skb_queue_purge(&sk->sk_write_queue);
1693 /* Set tx_stopped. This is checked when psock is bound to a kcm and we
1694 * get a writespace callback. This prevents further work being queued
1695 * from the callback (unbinding the psock occurs after canceling work.
1697 kcm->tx_stopped = 1;
1701 spin_lock_bh(&mux->lock);
1703 /* Take of tx_wait list, after this point there should be no way
1704 * that a psock will be assigned to this kcm.
1706 list_del(&kcm->wait_psock_list);
1707 kcm->tx_wait = false;
1709 spin_unlock_bh(&mux->lock);
1711 /* Cancel work. After this point there should be no outside references
1712 * to the kcm socket.
1714 cancel_work_sync(&kcm->tx_work);
1717 psock = kcm->tx_psock;
1719 /* A psock was reserved, so we need to kill it since it
1720 * may already have some bytes queued from a message. We
1721 * need to do this after removing kcm from tx_wait list.
1723 kcm_abort_tx_psock(psock, EPIPE, false);
1724 unreserve_psock(kcm);
1728 WARN_ON(kcm->tx_wait);
1729 WARN_ON(kcm->tx_psock);
1738 static const struct proto_ops kcm_dgram_ops = {
1740 .owner = THIS_MODULE,
1741 .release = kcm_release,
1742 .bind = sock_no_bind,
1743 .connect = sock_no_connect,
1744 .socketpair = sock_no_socketpair,
1745 .accept = sock_no_accept,
1746 .getname = sock_no_getname,
1747 .poll = datagram_poll,
1749 .listen = sock_no_listen,
1750 .shutdown = sock_no_shutdown,
1751 .setsockopt = kcm_setsockopt,
1752 .getsockopt = kcm_getsockopt,
1753 .sendmsg = kcm_sendmsg,
1754 .recvmsg = kcm_recvmsg,
1755 .mmap = sock_no_mmap,
1756 .splice_eof = kcm_splice_eof,
1759 static const struct proto_ops kcm_seqpacket_ops = {
1761 .owner = THIS_MODULE,
1762 .release = kcm_release,
1763 .bind = sock_no_bind,
1764 .connect = sock_no_connect,
1765 .socketpair = sock_no_socketpair,
1766 .accept = sock_no_accept,
1767 .getname = sock_no_getname,
1768 .poll = datagram_poll,
1770 .listen = sock_no_listen,
1771 .shutdown = sock_no_shutdown,
1772 .setsockopt = kcm_setsockopt,
1773 .getsockopt = kcm_getsockopt,
1774 .sendmsg = kcm_sendmsg,
1775 .recvmsg = kcm_recvmsg,
1776 .mmap = sock_no_mmap,
1777 .splice_eof = kcm_splice_eof,
1778 .splice_read = kcm_splice_read,
1781 /* Create proto operation for kcm sockets */
1782 static int kcm_create(struct net *net, struct socket *sock,
1783 int protocol, int kern)
1785 struct kcm_net *knet = net_generic(net, kcm_net_id);
1787 struct kcm_mux *mux;
1789 switch (sock->type) {
1791 sock->ops = &kcm_dgram_ops;
1793 case SOCK_SEQPACKET:
1794 sock->ops = &kcm_seqpacket_ops;
1797 return -ESOCKTNOSUPPORT;
1800 if (protocol != KCMPROTO_CONNECTED)
1801 return -EPROTONOSUPPORT;
1803 sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1807 /* Allocate a kcm mux, shared between KCM sockets */
1808 mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1814 spin_lock_init(&mux->lock);
1815 spin_lock_init(&mux->rx_lock);
1816 INIT_LIST_HEAD(&mux->kcm_socks);
1817 INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1818 INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1820 INIT_LIST_HEAD(&mux->psocks);
1821 INIT_LIST_HEAD(&mux->psocks_ready);
1822 INIT_LIST_HEAD(&mux->psocks_avail);
1826 /* Add new MUX to list */
1827 mutex_lock(&knet->mutex);
1828 list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1830 mutex_unlock(&knet->mutex);
1832 skb_queue_head_init(&mux->rx_hold_queue);
1834 /* Init KCM socket */
1835 sock_init_data(sock, sk);
1836 init_kcm_sock(kcm_sk(sk), mux);
1841 static const struct net_proto_family kcm_family_ops = {
1843 .create = kcm_create,
1844 .owner = THIS_MODULE,
1847 static __net_init int kcm_init_net(struct net *net)
1849 struct kcm_net *knet = net_generic(net, kcm_net_id);
1851 INIT_LIST_HEAD_RCU(&knet->mux_list);
1852 mutex_init(&knet->mutex);
1857 static __net_exit void kcm_exit_net(struct net *net)
1859 struct kcm_net *knet = net_generic(net, kcm_net_id);
1861 /* All KCM sockets should be closed at this point, which should mean
1862 * that all multiplexors and psocks have been destroyed.
1864 WARN_ON(!list_empty(&knet->mux_list));
1866 mutex_destroy(&knet->mutex);
1869 static struct pernet_operations kcm_net_ops = {
1870 .init = kcm_init_net,
1871 .exit = kcm_exit_net,
1873 .size = sizeof(struct kcm_net),
1876 static int __init kcm_init(void)
1880 kcm_muxp = kmem_cache_create("kcm_mux_cache",
1881 sizeof(struct kcm_mux), 0,
1882 SLAB_HWCACHE_ALIGN, NULL);
1886 kcm_psockp = kmem_cache_create("kcm_psock_cache",
1887 sizeof(struct kcm_psock), 0,
1888 SLAB_HWCACHE_ALIGN, NULL);
1892 kcm_wq = create_singlethread_workqueue("kkcmd");
1896 err = proto_register(&kcm_proto, 1);
1900 err = register_pernet_device(&kcm_net_ops);
1904 err = sock_register(&kcm_family_ops);
1906 goto sock_register_fail;
1908 err = kcm_proc_init();
1910 goto proc_init_fail;
1915 sock_unregister(PF_KCM);
1918 unregister_pernet_device(&kcm_net_ops);
1921 proto_unregister(&kcm_proto);
1924 kmem_cache_destroy(kcm_muxp);
1925 kmem_cache_destroy(kcm_psockp);
1928 destroy_workqueue(kcm_wq);
1933 static void __exit kcm_exit(void)
1936 sock_unregister(PF_KCM);
1937 unregister_pernet_device(&kcm_net_ops);
1938 proto_unregister(&kcm_proto);
1939 destroy_workqueue(kcm_wq);
1941 kmem_cache_destroy(kcm_muxp);
1942 kmem_cache_destroy(kcm_psockp);
1945 module_init(kcm_init);
1946 module_exit(kcm_exit);
1948 MODULE_LICENSE("GPL");
1949 MODULE_ALIAS_NETPROTO(PF_KCM);