xref: /linux/net/kcm/kcmsock.c (revision 5d6ce399d986dcf40d331d825fb2e4f8d0329865)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel Connection Multiplexor
4  *
5  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
6  */
7 
8 #include <linux/bpf.h>
9 #include <linux/errno.h>
10 #include <linux/errqueue.h>
11 #include <linux/file.h>
12 #include <linux/filter.h>
13 #include <linux/in.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>
26 
27 #include <net/kcm.h>
28 #include <net/netns/generic.h>
29 #include <net/sock.h>
30 #include <uapi/linux/kcm.h>
31 #include <trace/events/sock.h>
32 
33 unsigned int kcm_net_id;
34 
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;
38 
39 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
40 {
41 	return (struct kcm_sock *)sk;
42 }
43 
44 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
45 {
46 	return (struct kcm_tx_msg *)skb->cb;
47 }
48 
49 static void report_csk_error(struct sock *csk, int err)
50 {
51 	csk->sk_err = EPIPE;
52 	sk_error_report(csk);
53 }
54 
55 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
56 			       bool wakeup_kcm)
57 {
58 	struct sock *csk = psock->sk;
59 	struct kcm_mux *mux = psock->mux;
60 
61 	/* Unrecoverable error in transmit */
62 
63 	spin_lock_bh(&mux->lock);
64 
65 	if (psock->tx_stopped) {
66 		spin_unlock_bh(&mux->lock);
67 		return;
68 	}
69 
70 	psock->tx_stopped = 1;
71 	KCM_STATS_INCR(psock->stats.tx_aborts);
72 
73 	if (!psock->tx_kcm) {
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.
81 		 */
82 		smp_mb();
83 
84 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
85 	}
86 
87 	spin_unlock_bh(&mux->lock);
88 
89 	/* Report error on lower socket */
90 	report_csk_error(csk, err);
91 }
92 
93 /* RX mux lock held. */
94 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
95 				    struct kcm_psock *psock)
96 {
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;
104 }
105 
106 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
107 				    struct kcm_psock *psock)
108 {
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;
115 }
116 
117 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
118 
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.
122  */
123 static void kcm_rcv_ready(struct kcm_sock *kcm)
124 {
125 	struct kcm_mux *mux = kcm->mux;
126 	struct kcm_psock *psock;
127 	struct sk_buff *skb;
128 
129 	if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
130 		return;
131 
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));
137 			return;
138 		}
139 	}
140 
141 	while (!list_empty(&mux->psocks_ready)) {
142 		psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
143 					 psock_ready_list);
144 
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));
148 			return;
149 		}
150 
151 		/* Consumed the ready message on the psock. Schedule rx_work to
152 		 * get more messages.
153 		 */
154 		list_del(&psock->psock_ready_list);
155 		psock->ready_rx_msg = NULL;
156 		/* Commit clearing of ready_rx_msg for queuing work */
157 		smp_mb();
158 
159 		strp_unpause(&psock->strp);
160 		strp_check_rcv(&psock->strp);
161 	}
162 
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);
168 }
169 
170 static void kcm_rfree(struct sk_buff *skb)
171 {
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;
176 
177 	sk_mem_uncharge(sk, len);
178 	atomic_sub(len, &sk->sk_rmem_alloc);
179 
180 	/* For reading rx_wait and rx_psock without holding lock */
181 	smp_mb__after_atomic();
182 
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);
186 		kcm_rcv_ready(kcm);
187 		spin_unlock_bh(&mux->rx_lock);
188 	}
189 }
190 
191 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
192 {
193 	struct sk_buff_head *list = &sk->sk_receive_queue;
194 
195 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
196 		return -ENOMEM;
197 
198 	if (!sk_rmem_schedule(sk, skb, skb->truesize))
199 		return -ENOBUFS;
200 
201 	skb->dev = NULL;
202 
203 	skb_orphan(skb);
204 	skb->sk = sk;
205 	skb->destructor = kcm_rfree;
206 	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
207 	sk_mem_charge(sk, skb->truesize);
208 
209 	skb_queue_tail(list, skb);
210 
211 	if (!sock_flag(sk, SOCK_DEAD))
212 		sk->sk_data_ready(sk);
213 
214 	return 0;
215 }
216 
217 /* Requeue received messages for a kcm socket to other kcm sockets. This is
218  * called with a kcm socket is receive disabled.
219  * RX mux lock held.
220  */
221 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
222 {
223 	struct sk_buff *skb;
224 	struct kcm_sock *kcm;
225 
226 	while ((skb = skb_dequeue(head))) {
227 		/* Reset destructor to avoid calling kcm_rcv_ready */
228 		skb->destructor = sock_rfree;
229 		skb_orphan(skb);
230 try_again:
231 		if (list_empty(&mux->kcm_rx_waiters)) {
232 			skb_queue_tail(&mux->rx_hold_queue, skb);
233 			continue;
234 		}
235 
236 		kcm = list_first_entry(&mux->kcm_rx_waiters,
237 				       struct kcm_sock, wait_rx_list);
238 
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);
244 
245 			/* Commit rx_wait to read in kcm_free */
246 			smp_wmb();
247 
248 			goto try_again;
249 		}
250 	}
251 }
252 
253 /* Lower sock lock held */
254 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
255 				       struct sk_buff *head)
256 {
257 	struct kcm_mux *mux = psock->mux;
258 	struct kcm_sock *kcm;
259 
260 	WARN_ON(psock->ready_rx_msg);
261 
262 	if (psock->rx_kcm)
263 		return psock->rx_kcm;
264 
265 	spin_lock_bh(&mux->rx_lock);
266 
267 	if (psock->rx_kcm) {
268 		spin_unlock_bh(&mux->rx_lock);
269 		return psock->rx_kcm;
270 	}
271 
272 	kcm_update_rx_mux_stats(mux, psock);
273 
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,
278 			      &mux->psocks_ready);
279 		spin_unlock_bh(&mux->rx_lock);
280 		return NULL;
281 	}
282 
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);
288 
289 	psock->rx_kcm = kcm;
290 	/* paired with lockless reads in kcm_rfree() */
291 	WRITE_ONCE(kcm->rx_psock, psock);
292 
293 	spin_unlock_bh(&mux->rx_lock);
294 
295 	return kcm;
296 }
297 
298 static void kcm_done(struct kcm_sock *kcm);
299 
300 static void kcm_done_work(struct work_struct *w)
301 {
302 	kcm_done(container_of(w, struct kcm_sock, done_work));
303 }
304 
305 /* Lower sock held */
306 static void unreserve_rx_kcm(struct kcm_psock *psock,
307 			     bool rcv_ready)
308 {
309 	struct kcm_sock *kcm = psock->rx_kcm;
310 	struct kcm_mux *mux = psock->mux;
311 
312 	if (!kcm)
313 		return;
314 
315 	spin_lock_bh(&mux->rx_lock);
316 
317 	psock->rx_kcm = NULL;
318 	/* paired with lockless reads in kcm_rfree() */
319 	WRITE_ONCE(kcm->rx_psock, NULL);
320 
321 	/* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
322 	 * kcm_rfree
323 	 */
324 	smp_mb();
325 
326 	if (unlikely(kcm->done)) {
327 		spin_unlock_bh(&mux->rx_lock);
328 
329 		/* Need to run kcm_done in a task since we need to qcquire
330 		 * callback locks which may already be held here.
331 		 */
332 		INIT_WORK(&kcm->done_work, kcm_done_work);
333 		schedule_work(&kcm->done_work);
334 		return;
335 	}
336 
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).
342 		 */
343 		kcm_rcv_ready(kcm);
344 	}
345 	spin_unlock_bh(&mux->rx_lock);
346 }
347 
348 /* Lower sock lock held */
349 static void psock_data_ready(struct sock *sk)
350 {
351 	struct kcm_psock *psock;
352 
353 	trace_sk_data_ready(sk);
354 
355 	read_lock_bh(&sk->sk_callback_lock);
356 
357 	psock = (struct kcm_psock *)sk->sk_user_data;
358 	if (likely(psock))
359 		strp_data_ready(&psock->strp);
360 
361 	read_unlock_bh(&sk->sk_callback_lock);
362 }
363 
364 /* Called with lower sock held */
365 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
366 {
367 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
368 	struct kcm_sock *kcm;
369 
370 try_queue:
371 	kcm = reserve_rx_kcm(psock, skb);
372 	if (!kcm) {
373 		 /* Unable to reserve a KCM, message is held in psock and strp
374 		  * is paused.
375 		  */
376 		return;
377 	}
378 
379 	if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
380 		/* Should mean socket buffer full */
381 		unreserve_rx_kcm(psock, false);
382 		goto try_queue;
383 	}
384 }
385 
386 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
387 {
388 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
389 	struct bpf_prog *prog = psock->bpf_prog;
390 	int res;
391 
392 	res = bpf_prog_run_pin_on_cpu(prog, skb);
393 	return res;
394 }
395 
396 static int kcm_read_sock_done(struct strparser *strp, int err)
397 {
398 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
399 
400 	unreserve_rx_kcm(psock, true);
401 
402 	return err;
403 }
404 
405 static void psock_state_change(struct sock *sk)
406 {
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.
410 	 */
411 
412 	report_csk_error(sk, EPIPE);
413 }
414 
415 static void psock_write_space(struct sock *sk)
416 {
417 	struct kcm_psock *psock;
418 	struct kcm_mux *mux;
419 	struct kcm_sock *kcm;
420 
421 	read_lock_bh(&sk->sk_callback_lock);
422 
423 	psock = (struct kcm_psock *)sk->sk_user_data;
424 	if (unlikely(!psock))
425 		goto out;
426 	mux = psock->mux;
427 
428 	spin_lock_bh(&mux->lock);
429 
430 	/* Check if the socket is reserved so someone is waiting for sending. */
431 	kcm = psock->tx_kcm;
432 	if (kcm && !unlikely(kcm->tx_stopped))
433 		queue_work(kcm_wq, &kcm->tx_work);
434 
435 	spin_unlock_bh(&mux->lock);
436 out:
437 	read_unlock_bh(&sk->sk_callback_lock);
438 }
439 
440 static void unreserve_psock(struct kcm_sock *kcm);
441 
442 /* kcm sock is locked. */
443 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
444 {
445 	struct kcm_mux *mux = kcm->mux;
446 	struct kcm_psock *psock;
447 
448 	psock = kcm->tx_psock;
449 
450 	smp_rmb(); /* Must read tx_psock before tx_wait */
451 
452 	if (psock) {
453 		WARN_ON(kcm->tx_wait);
454 		if (unlikely(psock->tx_stopped))
455 			unreserve_psock(kcm);
456 		else
457 			return kcm->tx_psock;
458 	}
459 
460 	spin_lock_bh(&mux->lock);
461 
462 	/* Check again under lock to see if psock was reserved for this
463 	 * psock via psock_unreserve.
464 	 */
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;
470 	}
471 
472 	if (!list_empty(&mux->psocks_avail)) {
473 		psock = list_first_entry(&mux->psocks_avail,
474 					 struct kcm_psock,
475 					 psock_avail_list);
476 		list_del(&psock->psock_avail_list);
477 		if (kcm->tx_wait) {
478 			list_del(&kcm->wait_psock_list);
479 			kcm->tx_wait = false;
480 		}
481 		kcm->tx_psock = psock;
482 		psock->tx_kcm = kcm;
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);
487 		kcm->tx_wait = true;
488 	}
489 
490 	spin_unlock_bh(&mux->lock);
491 
492 	return psock;
493 }
494 
495 /* mux lock held */
496 static void psock_now_avail(struct kcm_psock *psock)
497 {
498 	struct kcm_mux *mux = psock->mux;
499 	struct kcm_sock *kcm;
500 
501 	if (list_empty(&mux->kcm_tx_waiters)) {
502 		list_add_tail(&psock->psock_avail_list,
503 			      &mux->psocks_avail);
504 	} else {
505 		kcm = list_first_entry(&mux->kcm_tx_waiters,
506 				       struct kcm_sock,
507 				       wait_psock_list);
508 		list_del(&kcm->wait_psock_list);
509 		kcm->tx_wait = false;
510 		psock->tx_kcm = kcm;
511 
512 		/* Commit before changing tx_psock since that is read in
513 		 * reserve_psock before queuing work.
514 		 */
515 		smp_mb();
516 
517 		kcm->tx_psock = psock;
518 		KCM_STATS_INCR(psock->stats.reserved);
519 		queue_work(kcm_wq, &kcm->tx_work);
520 	}
521 }
522 
523 /* kcm sock is locked. */
524 static void unreserve_psock(struct kcm_sock *kcm)
525 {
526 	struct kcm_psock *psock;
527 	struct kcm_mux *mux = kcm->mux;
528 
529 	spin_lock_bh(&mux->lock);
530 
531 	psock = kcm->tx_psock;
532 
533 	if (WARN_ON(!psock)) {
534 		spin_unlock_bh(&mux->lock);
535 		return;
536 	}
537 
538 	smp_rmb(); /* Read tx_psock before tx_wait */
539 
540 	kcm_update_tx_mux_stats(mux, psock);
541 
542 	WARN_ON(kcm->tx_wait);
543 
544 	kcm->tx_psock = NULL;
545 	psock->tx_kcm = NULL;
546 	KCM_STATS_INCR(psock->stats.unreserved);
547 
548 	if (unlikely(psock->tx_stopped)) {
549 		if (psock->done) {
550 			/* Deferred free */
551 			list_del(&psock->psock_list);
552 			mux->psocks_cnt--;
553 			sock_put(psock->sk);
554 			fput(psock->sk->sk_socket->file);
555 			kmem_cache_free(kcm_psockp, psock);
556 		}
557 
558 		/* Don't put back on available list */
559 
560 		spin_unlock_bh(&mux->lock);
561 
562 		return;
563 	}
564 
565 	psock_now_avail(psock);
566 
567 	spin_unlock_bh(&mux->lock);
568 }
569 
570 static void kcm_report_tx_retry(struct kcm_sock *kcm)
571 {
572 	struct kcm_mux *mux = kcm->mux;
573 
574 	spin_lock_bh(&mux->lock);
575 	KCM_STATS_INCR(mux->stats.tx_retries);
576 	spin_unlock_bh(&mux->lock);
577 }
578 
579 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
580  * held.  Return bytes actually sent or error.
581  */
582 static int kcm_write_msgs(struct kcm_sock *kcm)
583 {
584 	unsigned int total_sent = 0;
585 	struct sock *sk = &kcm->sk;
586 	struct kcm_psock *psock;
587 	struct sk_buff *head;
588 	int ret = 0;
589 
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.
595 		 */
596 		unreserve_psock(kcm);
597 		kcm_report_tx_retry(kcm);
598 		if (skb_queue_empty(&sk->sk_write_queue))
599 			return 0;
600 
601 		kcm_tx_msg(skb_peek(&sk->sk_write_queue))->started_tx = false;
602 	}
603 
604 retry:
605 	while ((head = skb_peek(&sk->sk_write_queue))) {
606 		struct msghdr msg = {
607 			.msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
608 		};
609 		struct kcm_tx_msg *txm = kcm_tx_msg(head);
610 		struct sk_buff *skb;
611 		unsigned int msize;
612 		int i;
613 
614 		if (!txm->started_tx) {
615 			psock = reserve_psock(kcm);
616 			if (!psock)
617 				goto out;
618 			skb = head;
619 			txm->frag_offset = 0;
620 			txm->sent = 0;
621 			txm->started_tx = true;
622 		} else {
623 			if (WARN_ON(!psock)) {
624 				ret = -EINVAL;
625 				goto out;
626 			}
627 			skb = txm->frag_skb;
628 		}
629 
630 		if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
631 			ret = -EINVAL;
632 			goto out;
633 		}
634 
635 		msize = 0;
636 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
637 			msize += skb_shinfo(skb)->frags[i].bv_len;
638 
639 		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE,
640 			      skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags,
641 			      msize);
642 		iov_iter_advance(&msg.msg_iter, txm->frag_offset);
643 
644 		do {
645 			ret = sock_sendmsg(psock->sk->sk_socket, &msg);
646 			if (ret <= 0) {
647 				if (ret == -EAGAIN) {
648 					/* Save state to try again when there's
649 					 * write space on the socket
650 					 */
651 					txm->frag_skb = skb;
652 					ret = 0;
653 					goto out;
654 				}
655 
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.
660 				 */
661 				kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
662 						   true);
663 				unreserve_psock(kcm);
664 				psock = NULL;
665 
666 				txm->started_tx = false;
667 				kcm_report_tx_retry(kcm);
668 				ret = 0;
669 				goto retry;
670 			}
671 
672 			txm->sent += ret;
673 			txm->frag_offset += ret;
674 			KCM_STATS_ADD(psock->stats.tx_bytes, ret);
675 		} while (msg.msg_iter.count > 0);
676 
677 		if (skb == head) {
678 			if (skb_has_frag_list(skb)) {
679 				txm->frag_skb = skb_shinfo(skb)->frag_list;
680 				txm->frag_offset = 0;
681 				continue;
682 			}
683 		} else if (skb->next) {
684 			txm->frag_skb = skb->next;
685 			txm->frag_offset = 0;
686 			continue;
687 		}
688 
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);
693 		kfree_skb(head);
694 		KCM_STATS_INCR(psock->stats.tx_msgs);
695 	}
696 out:
697 	if (!head) {
698 		/* Done with all queued messages. */
699 		WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
700 		if (psock)
701 			unreserve_psock(kcm);
702 	}
703 
704 	/* Check if write space is available */
705 	sk->sk_write_space(sk);
706 
707 	return total_sent ? : ret;
708 }
709 
710 static void kcm_tx_work(struct work_struct *w)
711 {
712 	struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
713 	struct sock *sk = &kcm->sk;
714 	int err;
715 
716 	lock_sock(sk);
717 
718 	/* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
719 	 * aborts
720 	 */
721 	err = kcm_write_msgs(kcm);
722 	if (err < 0) {
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);
726 		goto out;
727 	}
728 
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);
734 	}
735 
736 out:
737 	release_sock(sk);
738 }
739 
740 static void kcm_push(struct kcm_sock *kcm)
741 {
742 	if (kcm->tx_wait_more)
743 		kcm_write_msgs(kcm);
744 }
745 
746 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
747 {
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);
755 	int err = -EPIPE;
756 
757 	lock_sock(sk);
758 
759 	/* Per tcp_sendmsg this should be in poll */
760 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
761 
762 	if (sk->sk_err)
763 		goto out_error;
764 
765 	if (kcm->seq_skb) {
766 		/* Previously opened message */
767 		head = kcm->seq_skb;
768 		skb = kcm_tx_msg(head)->last_skb;
769 		goto start;
770 	}
771 
772 	/* Call the sk_stream functions to manage the sndbuf mem. */
773 	if (!sk_stream_memory_free(sk)) {
774 		kcm_push(kcm);
775 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
776 		err = sk_stream_wait_memory(sk, &timeo);
777 		if (err)
778 			goto out_error;
779 	}
780 
781 	if (msg_data_left(msg)) {
782 		/* New message, alloc head skb */
783 		head = alloc_skb(0, sk->sk_allocation);
784 		while (!head) {
785 			kcm_push(kcm);
786 			err = sk_stream_wait_memory(sk, &timeo);
787 			if (err)
788 				goto out_error;
789 
790 			head = alloc_skb(0, sk->sk_allocation);
791 		}
792 
793 		skb = head;
794 
795 		/* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
796 		 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
797 		 */
798 		skb->ip_summed = CHECKSUM_UNNECESSARY;
799 	}
800 
801 start:
802 	while (msg_data_left(msg)) {
803 		bool merge = true;
804 		int i = skb_shinfo(skb)->nr_frags;
805 		struct page_frag *pfrag = sk_page_frag(sk);
806 
807 		if (!sk_page_frag_refill(sk, pfrag))
808 			goto wait_for_memory;
809 
810 		if (!skb_can_coalesce(skb, i, pfrag->page,
811 				      pfrag->offset)) {
812 			if (i == MAX_SKB_FRAGS) {
813 				struct sk_buff *tskb;
814 
815 				tskb = alloc_skb(0, sk->sk_allocation);
816 				if (!tskb)
817 					goto wait_for_memory;
818 
819 				if (head == skb)
820 					skb_shinfo(head)->frag_list = tskb;
821 				else
822 					skb->next = tskb;
823 
824 				skb = tskb;
825 				skb->ip_summed = CHECKSUM_UNNECESSARY;
826 				continue;
827 			}
828 			merge = false;
829 		}
830 
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;
835 
836 			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
837 						   sk->sk_allocation);
838 			if (err < 0) {
839 				if (err == -EMSGSIZE)
840 					goto wait_for_memory;
841 				goto out_error;
842 			}
843 
844 			copy = err;
845 			skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
846 			sk_wmem_queued_add(sk, copy);
847 			sk_mem_charge(sk, copy);
848 
849 			if (head != skb)
850 				head->truesize += copy;
851 		} else {
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;
856 
857 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
858 						       pfrag->page,
859 						       pfrag->offset,
860 						       copy);
861 			if (err)
862 				goto out_error;
863 
864 			/* Update the skb. */
865 			if (merge) {
866 				skb_frag_size_add(
867 					&skb_shinfo(skb)->frags[i - 1], copy);
868 			} else {
869 				skb_fill_page_desc(skb, i, pfrag->page,
870 						   pfrag->offset, copy);
871 				get_page(pfrag->page);
872 			}
873 
874 			pfrag->offset += copy;
875 		}
876 
877 		copied += copy;
878 		if (head != skb) {
879 			head->len += copy;
880 			head->data_len += copy;
881 		}
882 
883 		continue;
884 
885 wait_for_memory:
886 		kcm_push(kcm);
887 		err = sk_stream_wait_memory(sk, &timeo);
888 		if (err)
889 			goto out_error;
890 	}
891 
892 	if (eor) {
893 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
894 
895 		if (head) {
896 			/* Message complete, queue it on send buffer */
897 			__skb_queue_tail(&sk->sk_write_queue, head);
898 			kcm->seq_skb = NULL;
899 			KCM_STATS_INCR(kcm->stats.tx_msgs);
900 		}
901 
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);
906 			if (err < 0) {
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
910 				 * from sendmsg
911 				 */
912 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
913 				report_csk_error(&kcm->sk, -err);
914 			}
915 		}
916 	} else {
917 		/* Message not complete, save state */
918 partial_message:
919 		if (head) {
920 			kcm->seq_skb = head;
921 			kcm_tx_msg(head)->last_skb = skb;
922 		}
923 	}
924 
925 	KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
926 
927 	release_sock(sk);
928 	return copied;
929 
930 out_error:
931 	kcm_push(kcm);
932 
933 	if (sock->type == SOCK_SEQPACKET) {
934 		/* Wrote some bytes before encountering an
935 		 * error, return partial success.
936 		 */
937 		if (copied)
938 			goto partial_message;
939 		if (head != kcm->seq_skb)
940 			kfree_skb(head);
941 	} else {
942 		kfree_skb(head);
943 		kcm->seq_skb = NULL;
944 	}
945 
946 	err = sk_stream_error(sk, msg->msg_flags, err);
947 
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);
951 
952 	release_sock(sk);
953 	return err;
954 }
955 
956 static void kcm_splice_eof(struct socket *sock)
957 {
958 	struct sock *sk = sock->sk;
959 	struct kcm_sock *kcm = kcm_sk(sk);
960 
961 	if (skb_queue_empty_lockless(&sk->sk_write_queue))
962 		return;
963 
964 	lock_sock(sk);
965 	kcm_write_msgs(kcm);
966 	release_sock(sk);
967 }
968 
969 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
970 		       size_t len, int flags)
971 {
972 	struct sock *sk = sock->sk;
973 	struct kcm_sock *kcm = kcm_sk(sk);
974 	int err = 0;
975 	struct strp_msg *stm;
976 	int copied = 0;
977 	struct sk_buff *skb;
978 
979 	skb = skb_recv_datagram(sk, flags, &err);
980 	if (!skb)
981 		goto out;
982 
983 	/* Okay, have a message on the receive queue */
984 
985 	stm = strp_msg(skb);
986 
987 	if (len > stm->full_len)
988 		len = stm->full_len;
989 
990 	err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
991 	if (err < 0)
992 		goto out;
993 
994 	copied = 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;
1001 				goto msg_finished;
1002 			}
1003 			stm->offset += copied;
1004 			stm->full_len -= copied;
1005 		} else {
1006 msg_finished:
1007 			/* Finished with message */
1008 			msg->msg_flags |= MSG_EOR;
1009 			KCM_STATS_INCR(kcm->stats.rx_msgs);
1010 		}
1011 	}
1012 
1013 out:
1014 	skb_free_datagram(sk, skb);
1015 	return copied ? : err;
1016 }
1017 
1018 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1019 			       struct pipe_inode_info *pipe, size_t len,
1020 			       unsigned int flags)
1021 {
1022 	struct sock *sk = sock->sk;
1023 	struct kcm_sock *kcm = kcm_sk(sk);
1024 	struct strp_msg *stm;
1025 	int err = 0;
1026 	ssize_t copied;
1027 	struct sk_buff *skb;
1028 
1029 	/* Only support splice for SOCKSEQPACKET */
1030 
1031 	skb = skb_recv_datagram(sk, flags, &err);
1032 	if (!skb)
1033 		goto err_out;
1034 
1035 	/* Okay, have a message on the receive queue */
1036 
1037 	stm = strp_msg(skb);
1038 
1039 	if (len > stm->full_len)
1040 		len = stm->full_len;
1041 
1042 	copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1043 	if (copied < 0) {
1044 		err = copied;
1045 		goto err_out;
1046 	}
1047 
1048 	KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1049 
1050 	stm->offset += copied;
1051 	stm->full_len -= copied;
1052 
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.
1057 	 */
1058 
1059 	skb_free_datagram(sk, skb);
1060 	return copied;
1061 
1062 err_out:
1063 	skb_free_datagram(sk, skb);
1064 	return err;
1065 }
1066 
1067 /* kcm sock lock held */
1068 static void kcm_recv_disable(struct kcm_sock *kcm)
1069 {
1070 	struct kcm_mux *mux = kcm->mux;
1071 
1072 	if (kcm->rx_disabled)
1073 		return;
1074 
1075 	spin_lock_bh(&mux->rx_lock);
1076 
1077 	kcm->rx_disabled = 1;
1078 
1079 	/* If a psock is reserved we'll do cleanup in unreserve */
1080 	if (!kcm->rx_psock) {
1081 		if (kcm->rx_wait) {
1082 			list_del(&kcm->wait_rx_list);
1083 			/* paired with lockless reads in kcm_rfree() */
1084 			WRITE_ONCE(kcm->rx_wait, false);
1085 		}
1086 
1087 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1088 	}
1089 
1090 	spin_unlock_bh(&mux->rx_lock);
1091 }
1092 
1093 /* kcm sock lock held */
1094 static void kcm_recv_enable(struct kcm_sock *kcm)
1095 {
1096 	struct kcm_mux *mux = kcm->mux;
1097 
1098 	if (!kcm->rx_disabled)
1099 		return;
1100 
1101 	spin_lock_bh(&mux->rx_lock);
1102 
1103 	kcm->rx_disabled = 0;
1104 	kcm_rcv_ready(kcm);
1105 
1106 	spin_unlock_bh(&mux->rx_lock);
1107 }
1108 
1109 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1110 			  sockptr_t optval, unsigned int optlen)
1111 {
1112 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1113 	int val, valbool;
1114 	int err = 0;
1115 
1116 	if (level != SOL_KCM)
1117 		return -ENOPROTOOPT;
1118 
1119 	if (optlen < sizeof(int))
1120 		return -EINVAL;
1121 
1122 	if (copy_from_sockptr(&val, optval, sizeof(int)))
1123 		return -EFAULT;
1124 
1125 	valbool = val ? 1 : 0;
1126 
1127 	switch (optname) {
1128 	case KCM_RECV_DISABLE:
1129 		lock_sock(&kcm->sk);
1130 		if (valbool)
1131 			kcm_recv_disable(kcm);
1132 		else
1133 			kcm_recv_enable(kcm);
1134 		release_sock(&kcm->sk);
1135 		break;
1136 	default:
1137 		err = -ENOPROTOOPT;
1138 	}
1139 
1140 	return err;
1141 }
1142 
1143 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1144 			  char __user *optval, int __user *optlen)
1145 {
1146 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1147 	int val, len;
1148 
1149 	if (level != SOL_KCM)
1150 		return -ENOPROTOOPT;
1151 
1152 	if (get_user(len, optlen))
1153 		return -EFAULT;
1154 
1155 	len = min_t(unsigned int, len, sizeof(int));
1156 	if (len < 0)
1157 		return -EINVAL;
1158 
1159 	switch (optname) {
1160 	case KCM_RECV_DISABLE:
1161 		val = kcm->rx_disabled;
1162 		break;
1163 	default:
1164 		return -ENOPROTOOPT;
1165 	}
1166 
1167 	if (put_user(len, optlen))
1168 		return -EFAULT;
1169 	if (copy_to_user(optval, &val, len))
1170 		return -EFAULT;
1171 	return 0;
1172 }
1173 
1174 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1175 {
1176 	struct kcm_sock *tkcm;
1177 	struct list_head *head;
1178 	int index = 0;
1179 
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
1182 	 * EPOLLHUP
1183 	 */
1184 	kcm->sk.sk_state = TCP_ESTABLISHED;
1185 
1186 	/* Add to mux's kcm sockets list */
1187 	kcm->mux = mux;
1188 	spin_lock_bh(&mux->lock);
1189 
1190 	head = &mux->kcm_socks;
1191 	list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1192 		if (tkcm->index != index)
1193 			break;
1194 		head = &tkcm->kcm_sock_list;
1195 		index++;
1196 	}
1197 
1198 	list_add(&kcm->kcm_sock_list, head);
1199 	kcm->index = index;
1200 
1201 	mux->kcm_socks_cnt++;
1202 	spin_unlock_bh(&mux->lock);
1203 
1204 	INIT_WORK(&kcm->tx_work, kcm_tx_work);
1205 
1206 	spin_lock_bh(&mux->rx_lock);
1207 	kcm_rcv_ready(kcm);
1208 	spin_unlock_bh(&mux->rx_lock);
1209 }
1210 
1211 static int kcm_attach(struct socket *sock, struct socket *csock,
1212 		      struct bpf_prog *prog)
1213 {
1214 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1215 	struct kcm_mux *mux = kcm->mux;
1216 	struct sock *csk;
1217 	struct kcm_psock *psock = NULL, *tpsock;
1218 	struct list_head *head;
1219 	int index = 0;
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,
1224 	};
1225 	int err = 0;
1226 
1227 	csk = csock->sk;
1228 	if (!csk)
1229 		return -EINVAL;
1230 
1231 	lock_sock(csk);
1232 
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) {
1236 		err = -EOPNOTSUPP;
1237 		goto out;
1238 	}
1239 
1240 	/* Don't allow listeners or closed sockets */
1241 	if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1242 		err = -EOPNOTSUPP;
1243 		goto out;
1244 	}
1245 
1246 	psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1247 	if (!psock) {
1248 		err = -ENOMEM;
1249 		goto out;
1250 	}
1251 
1252 	psock->mux = mux;
1253 	psock->sk = csk;
1254 	psock->bpf_prog = prog;
1255 
1256 	write_lock_bh(&csk->sk_callback_lock);
1257 
1258 	/* Check if sk_user_data is already by KCM or someone else.
1259 	 * Must be done under lock to prevent race conditions.
1260 	 */
1261 	if (csk->sk_user_data) {
1262 		write_unlock_bh(&csk->sk_callback_lock);
1263 		kmem_cache_free(kcm_psockp, psock);
1264 		err = -EALREADY;
1265 		goto out;
1266 	}
1267 
1268 	err = strp_init(&psock->strp, csk, &cb);
1269 	if (err) {
1270 		write_unlock_bh(&csk->sk_callback_lock);
1271 		kmem_cache_free(kcm_psockp, psock);
1272 		goto out;
1273 	}
1274 
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;
1282 
1283 	write_unlock_bh(&csk->sk_callback_lock);
1284 
1285 	sock_hold(csk);
1286 
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)
1292 			break;
1293 		head = &tpsock->psock_list;
1294 		index++;
1295 	}
1296 
1297 	list_add(&psock->psock_list, head);
1298 	psock->index = index;
1299 
1300 	KCM_STATS_INCR(mux->stats.psock_attach);
1301 	mux->psocks_cnt++;
1302 	psock_now_avail(psock);
1303 	spin_unlock_bh(&mux->lock);
1304 
1305 	/* Schedule RX work in case there are already bytes queued */
1306 	strp_check_rcv(&psock->strp);
1307 
1308 out:
1309 	release_sock(csk);
1310 
1311 	return err;
1312 }
1313 
1314 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1315 {
1316 	struct socket *csock;
1317 	struct bpf_prog *prog;
1318 	int err;
1319 
1320 	csock = sockfd_lookup(info->fd, &err);
1321 	if (!csock)
1322 		return -ENOENT;
1323 
1324 	prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1325 	if (IS_ERR(prog)) {
1326 		err = PTR_ERR(prog);
1327 		goto out;
1328 	}
1329 
1330 	err = kcm_attach(sock, csock, prog);
1331 	if (err) {
1332 		bpf_prog_put(prog);
1333 		goto out;
1334 	}
1335 
1336 	/* Keep reference on file also */
1337 
1338 	return 0;
1339 out:
1340 	sockfd_put(csock);
1341 	return err;
1342 }
1343 
1344 static void kcm_unattach(struct kcm_psock *psock)
1345 {
1346 	struct sock *csk = psock->sk;
1347 	struct kcm_mux *mux = psock->mux;
1348 
1349 	lock_sock(csk);
1350 
1351 	/* Stop getting callbacks from TCP socket. After this there should
1352 	 * be no way to reserve a kcm for this psock.
1353 	 */
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);
1360 
1361 	if (WARN_ON(psock->rx_kcm)) {
1362 		write_unlock_bh(&csk->sk_callback_lock);
1363 		release_sock(csk);
1364 		return;
1365 	}
1366 
1367 	spin_lock_bh(&mux->rx_lock);
1368 
1369 	/* Stop receiver activities. After this point psock should not be
1370 	 * able to get onto ready list either through callbacks or work.
1371 	 */
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);
1377 	}
1378 
1379 	spin_unlock_bh(&mux->rx_lock);
1380 
1381 	write_unlock_bh(&csk->sk_callback_lock);
1382 
1383 	/* Call strp_done without sock lock */
1384 	release_sock(csk);
1385 	strp_done(&psock->strp);
1386 	lock_sock(csk);
1387 
1388 	bpf_prog_put(psock->bpf_prog);
1389 
1390 	spin_lock_bh(&mux->lock);
1391 
1392 	aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1393 	save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1394 
1395 	KCM_STATS_INCR(mux->stats.psock_unattach);
1396 
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
1400 		 * acquired here.
1401 		 */
1402 		KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1403 		spin_unlock_bh(&mux->lock);
1404 
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.
1408 		 */
1409 		kcm_abort_tx_psock(psock, EPIPE, false);
1410 
1411 		spin_lock_bh(&mux->lock);
1412 		if (!psock->tx_kcm) {
1413 			/* psock now unreserved in window mux was unlocked */
1414 			goto no_reserved;
1415 		}
1416 		psock->done = 1;
1417 
1418 		/* Commit done before queuing work to process it */
1419 		smp_mb();
1420 
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);
1424 	} else {
1425 no_reserved:
1426 		if (!psock->tx_stopped)
1427 			list_del(&psock->psock_avail_list);
1428 		list_del(&psock->psock_list);
1429 		mux->psocks_cnt--;
1430 		spin_unlock_bh(&mux->lock);
1431 
1432 		sock_put(csk);
1433 		fput(csk->sk_socket->file);
1434 		kmem_cache_free(kcm_psockp, psock);
1435 	}
1436 
1437 	release_sock(csk);
1438 }
1439 
1440 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1441 {
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;
1446 	struct sock *csk;
1447 	int err;
1448 
1449 	csock = sockfd_lookup(info->fd, &err);
1450 	if (!csock)
1451 		return -ENOENT;
1452 
1453 	csk = csock->sk;
1454 	if (!csk) {
1455 		err = -EINVAL;
1456 		goto out;
1457 	}
1458 
1459 	err = -ENOENT;
1460 
1461 	spin_lock_bh(&mux->lock);
1462 
1463 	list_for_each_entry(psock, &mux->psocks, psock_list) {
1464 		if (psock->sk != csk)
1465 			continue;
1466 
1467 		/* Found the matching psock */
1468 
1469 		if (psock->unattaching || WARN_ON(psock->done)) {
1470 			err = -EALREADY;
1471 			break;
1472 		}
1473 
1474 		psock->unattaching = 1;
1475 
1476 		spin_unlock_bh(&mux->lock);
1477 
1478 		/* Lower socket lock should already be held */
1479 		kcm_unattach(psock);
1480 
1481 		err = 0;
1482 		goto out;
1483 	}
1484 
1485 	spin_unlock_bh(&mux->lock);
1486 
1487 out:
1488 	sockfd_put(csock);
1489 	return err;
1490 }
1491 
1492 static struct proto kcm_proto = {
1493 	.name	= "KCM",
1494 	.owner	= THIS_MODULE,
1495 	.obj_size = sizeof(struct kcm_sock),
1496 };
1497 
1498 /* Clone a kcm socket. */
1499 static struct file *kcm_clone(struct socket *osock)
1500 {
1501 	struct socket *newsock;
1502 	struct sock *newsk;
1503 
1504 	newsock = sock_alloc();
1505 	if (!newsock)
1506 		return ERR_PTR(-ENFILE);
1507 
1508 	newsock->type = osock->type;
1509 	newsock->ops = osock->ops;
1510 
1511 	__module_get(newsock->ops->owner);
1512 
1513 	newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1514 			 &kcm_proto, false);
1515 	if (!newsk) {
1516 		sock_release(newsock);
1517 		return ERR_PTR(-ENOMEM);
1518 	}
1519 	sock_init_data(newsock, newsk);
1520 	init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1521 
1522 	return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1523 }
1524 
1525 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1526 {
1527 	int err;
1528 
1529 	switch (cmd) {
1530 	case SIOCKCMATTACH: {
1531 		struct kcm_attach info;
1532 
1533 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1534 			return -EFAULT;
1535 
1536 		err = kcm_attach_ioctl(sock, &info);
1537 
1538 		break;
1539 	}
1540 	case SIOCKCMUNATTACH: {
1541 		struct kcm_unattach info;
1542 
1543 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1544 			return -EFAULT;
1545 
1546 		err = kcm_unattach_ioctl(sock, &info);
1547 
1548 		break;
1549 	}
1550 	case SIOCKCMCLONE: {
1551 		struct kcm_clone info;
1552 		struct file *file;
1553 
1554 		info.fd = get_unused_fd_flags(0);
1555 		if (unlikely(info.fd < 0))
1556 			return info.fd;
1557 
1558 		file = kcm_clone(sock);
1559 		if (IS_ERR(file)) {
1560 			put_unused_fd(info.fd);
1561 			return PTR_ERR(file);
1562 		}
1563 		if (copy_to_user((void __user *)arg, &info,
1564 				 sizeof(info))) {
1565 			put_unused_fd(info.fd);
1566 			fput(file);
1567 			return -EFAULT;
1568 		}
1569 		fd_install(info.fd, file);
1570 		err = 0;
1571 		break;
1572 	}
1573 	default:
1574 		err = -ENOIOCTLCMD;
1575 		break;
1576 	}
1577 
1578 	return err;
1579 }
1580 
1581 static void free_mux(struct rcu_head *rcu)
1582 {
1583 	struct kcm_mux *mux = container_of(rcu,
1584 	    struct kcm_mux, rcu);
1585 
1586 	kmem_cache_free(kcm_muxp, mux);
1587 }
1588 
1589 static void release_mux(struct kcm_mux *mux)
1590 {
1591 	struct kcm_net *knet = mux->knet;
1592 	struct kcm_psock *psock, *tmp_psock;
1593 
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);
1599 	}
1600 
1601 	if (WARN_ON(mux->psocks_cnt))
1602 		return;
1603 
1604 	__skb_queue_purge(&mux->rx_hold_queue);
1605 
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);
1613 	knet->count--;
1614 	mutex_unlock(&knet->mutex);
1615 
1616 	call_rcu(&mux->rcu, free_mux);
1617 }
1618 
1619 static void kcm_done(struct kcm_sock *kcm)
1620 {
1621 	struct kcm_mux *mux = kcm->mux;
1622 	struct sock *sk = &kcm->sk;
1623 	int socks_cnt;
1624 
1625 	spin_lock_bh(&mux->rx_lock);
1626 	if (kcm->rx_psock) {
1627 		/* Cleanup in unreserve_rx_kcm */
1628 		WARN_ON(kcm->done);
1629 		kcm->rx_disabled = 1;
1630 		kcm->done = 1;
1631 		spin_unlock_bh(&mux->rx_lock);
1632 		return;
1633 	}
1634 
1635 	if (kcm->rx_wait) {
1636 		list_del(&kcm->wait_rx_list);
1637 		/* paired with lockless reads in kcm_rfree() */
1638 		WRITE_ONCE(kcm->rx_wait, false);
1639 	}
1640 	/* Move any pending receive messages to other kcm sockets */
1641 	requeue_rx_msgs(mux, &sk->sk_receive_queue);
1642 
1643 	spin_unlock_bh(&mux->rx_lock);
1644 
1645 	if (WARN_ON(sk_rmem_alloc_get(sk)))
1646 		return;
1647 
1648 	/* Detach from MUX */
1649 	spin_lock_bh(&mux->lock);
1650 
1651 	list_del(&kcm->kcm_sock_list);
1652 	mux->kcm_socks_cnt--;
1653 	socks_cnt = mux->kcm_socks_cnt;
1654 
1655 	spin_unlock_bh(&mux->lock);
1656 
1657 	if (!socks_cnt) {
1658 		/* We are done with the mux now. */
1659 		release_mux(mux);
1660 	}
1661 
1662 	WARN_ON(kcm->rx_wait);
1663 
1664 	sock_put(&kcm->sk);
1665 }
1666 
1667 /* Called by kcm_release to close a KCM socket.
1668  * If this is the last KCM socket on the MUX, destroy the MUX.
1669  */
1670 static int kcm_release(struct socket *sock)
1671 {
1672 	struct sock *sk = sock->sk;
1673 	struct kcm_sock *kcm;
1674 	struct kcm_mux *mux;
1675 	struct kcm_psock *psock;
1676 
1677 	if (!sk)
1678 		return 0;
1679 
1680 	kcm = kcm_sk(sk);
1681 	mux = kcm->mux;
1682 
1683 	lock_sock(sk);
1684 	sock_orphan(sk);
1685 	kfree_skb(kcm->seq_skb);
1686 
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.
1690 	 */
1691 	__skb_queue_purge(&sk->sk_write_queue);
1692 
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.
1696 	 */
1697 	kcm->tx_stopped = 1;
1698 
1699 	release_sock(sk);
1700 
1701 	spin_lock_bh(&mux->lock);
1702 	if (kcm->tx_wait) {
1703 		/* Take of tx_wait list, after this point there should be no way
1704 		 * that a psock will be assigned to this kcm.
1705 		 */
1706 		list_del(&kcm->wait_psock_list);
1707 		kcm->tx_wait = false;
1708 	}
1709 	spin_unlock_bh(&mux->lock);
1710 
1711 	/* Cancel work. After this point there should be no outside references
1712 	 * to the kcm socket.
1713 	 */
1714 	cancel_work_sync(&kcm->tx_work);
1715 
1716 	lock_sock(sk);
1717 	psock = kcm->tx_psock;
1718 	if (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.
1722 		 */
1723 		kcm_abort_tx_psock(psock, EPIPE, false);
1724 		unreserve_psock(kcm);
1725 	}
1726 	release_sock(sk);
1727 
1728 	WARN_ON(kcm->tx_wait);
1729 	WARN_ON(kcm->tx_psock);
1730 
1731 	sock->sk = NULL;
1732 
1733 	kcm_done(kcm);
1734 
1735 	return 0;
1736 }
1737 
1738 static const struct proto_ops kcm_dgram_ops = {
1739 	.family =	PF_KCM,
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,
1748 	.ioctl =	kcm_ioctl,
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,
1757 };
1758 
1759 static const struct proto_ops kcm_seqpacket_ops = {
1760 	.family =	PF_KCM,
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,
1769 	.ioctl =	kcm_ioctl,
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,
1779 };
1780 
1781 /* Create proto operation for kcm sockets */
1782 static int kcm_create(struct net *net, struct socket *sock,
1783 		      int protocol, int kern)
1784 {
1785 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1786 	struct sock *sk;
1787 	struct kcm_mux *mux;
1788 
1789 	switch (sock->type) {
1790 	case SOCK_DGRAM:
1791 		sock->ops = &kcm_dgram_ops;
1792 		break;
1793 	case SOCK_SEQPACKET:
1794 		sock->ops = &kcm_seqpacket_ops;
1795 		break;
1796 	default:
1797 		return -ESOCKTNOSUPPORT;
1798 	}
1799 
1800 	if (protocol != KCMPROTO_CONNECTED)
1801 		return -EPROTONOSUPPORT;
1802 
1803 	sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1804 	if (!sk)
1805 		return -ENOMEM;
1806 
1807 	/* Allocate a kcm mux, shared between KCM sockets */
1808 	mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1809 	if (!mux) {
1810 		sk_free(sk);
1811 		return -ENOMEM;
1812 	}
1813 
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);
1819 
1820 	INIT_LIST_HEAD(&mux->psocks);
1821 	INIT_LIST_HEAD(&mux->psocks_ready);
1822 	INIT_LIST_HEAD(&mux->psocks_avail);
1823 
1824 	mux->knet = knet;
1825 
1826 	/* Add new MUX to list */
1827 	mutex_lock(&knet->mutex);
1828 	list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1829 	knet->count++;
1830 	mutex_unlock(&knet->mutex);
1831 
1832 	skb_queue_head_init(&mux->rx_hold_queue);
1833 
1834 	/* Init KCM socket */
1835 	sock_init_data(sock, sk);
1836 	init_kcm_sock(kcm_sk(sk), mux);
1837 
1838 	return 0;
1839 }
1840 
1841 static const struct net_proto_family kcm_family_ops = {
1842 	.family = PF_KCM,
1843 	.create = kcm_create,
1844 	.owner  = THIS_MODULE,
1845 };
1846 
1847 static __net_init int kcm_init_net(struct net *net)
1848 {
1849 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1850 
1851 	INIT_LIST_HEAD_RCU(&knet->mux_list);
1852 	mutex_init(&knet->mutex);
1853 
1854 	return 0;
1855 }
1856 
1857 static __net_exit void kcm_exit_net(struct net *net)
1858 {
1859 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1860 
1861 	/* All KCM sockets should be closed at this point, which should mean
1862 	 * that all multiplexors and psocks have been destroyed.
1863 	 */
1864 	WARN_ON(!list_empty(&knet->mux_list));
1865 
1866 	mutex_destroy(&knet->mutex);
1867 }
1868 
1869 static struct pernet_operations kcm_net_ops = {
1870 	.init = kcm_init_net,
1871 	.exit = kcm_exit_net,
1872 	.id   = &kcm_net_id,
1873 	.size = sizeof(struct kcm_net),
1874 };
1875 
1876 static int __init kcm_init(void)
1877 {
1878 	int err = -ENOMEM;
1879 
1880 	kcm_muxp = kmem_cache_create("kcm_mux_cache",
1881 				     sizeof(struct kcm_mux), 0,
1882 				     SLAB_HWCACHE_ALIGN, NULL);
1883 	if (!kcm_muxp)
1884 		goto fail;
1885 
1886 	kcm_psockp = kmem_cache_create("kcm_psock_cache",
1887 				       sizeof(struct kcm_psock), 0,
1888 					SLAB_HWCACHE_ALIGN, NULL);
1889 	if (!kcm_psockp)
1890 		goto fail;
1891 
1892 	kcm_wq = create_singlethread_workqueue("kkcmd");
1893 	if (!kcm_wq)
1894 		goto fail;
1895 
1896 	err = proto_register(&kcm_proto, 1);
1897 	if (err)
1898 		goto fail;
1899 
1900 	err = register_pernet_device(&kcm_net_ops);
1901 	if (err)
1902 		goto net_ops_fail;
1903 
1904 	err = sock_register(&kcm_family_ops);
1905 	if (err)
1906 		goto sock_register_fail;
1907 
1908 	err = kcm_proc_init();
1909 	if (err)
1910 		goto proc_init_fail;
1911 
1912 	return 0;
1913 
1914 proc_init_fail:
1915 	sock_unregister(PF_KCM);
1916 
1917 sock_register_fail:
1918 	unregister_pernet_device(&kcm_net_ops);
1919 
1920 net_ops_fail:
1921 	proto_unregister(&kcm_proto);
1922 
1923 fail:
1924 	kmem_cache_destroy(kcm_muxp);
1925 	kmem_cache_destroy(kcm_psockp);
1926 
1927 	if (kcm_wq)
1928 		destroy_workqueue(kcm_wq);
1929 
1930 	return err;
1931 }
1932 
1933 static void __exit kcm_exit(void)
1934 {
1935 	kcm_proc_exit();
1936 	sock_unregister(PF_KCM);
1937 	unregister_pernet_device(&kcm_net_ops);
1938 	proto_unregister(&kcm_proto);
1939 	destroy_workqueue(kcm_wq);
1940 
1941 	kmem_cache_destroy(kcm_muxp);
1942 	kmem_cache_destroy(kcm_psockp);
1943 }
1944 
1945 module_init(kcm_init);
1946 module_exit(kcm_exit);
1947 
1948 MODULE_LICENSE("GPL");
1949 MODULE_DESCRIPTION("KCM (Kernel Connection Multiplexor) sockets");
1950 MODULE_ALIAS_NETPROTO(PF_KCM);
1951