xref: /linux/net/kcm/kcmsock.c (revision 561add0da6d3d07c9bccb0832fb6ed5619167d26)
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 (copied && sock->type == SOCK_SEQPACKET) {
934 		/* Wrote some bytes before encountering an
935 		 * error, return partial success.
936 		 */
937 		goto partial_message;
938 	}
939 
940 	if (head != kcm->seq_skb)
941 		kfree_skb(head);
942 
943 	err = sk_stream_error(sk, msg->msg_flags, err);
944 
945 	/* make sure we wake any epoll edge trigger waiter */
946 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
947 		sk->sk_write_space(sk);
948 
949 	release_sock(sk);
950 	return err;
951 }
952 
953 static void kcm_splice_eof(struct socket *sock)
954 {
955 	struct sock *sk = sock->sk;
956 	struct kcm_sock *kcm = kcm_sk(sk);
957 
958 	if (skb_queue_empty_lockless(&sk->sk_write_queue))
959 		return;
960 
961 	lock_sock(sk);
962 	kcm_write_msgs(kcm);
963 	release_sock(sk);
964 }
965 
966 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
967 		       size_t len, int flags)
968 {
969 	struct sock *sk = sock->sk;
970 	struct kcm_sock *kcm = kcm_sk(sk);
971 	int err = 0;
972 	struct strp_msg *stm;
973 	int copied = 0;
974 	struct sk_buff *skb;
975 
976 	skb = skb_recv_datagram(sk, flags, &err);
977 	if (!skb)
978 		goto out;
979 
980 	/* Okay, have a message on the receive queue */
981 
982 	stm = strp_msg(skb);
983 
984 	if (len > stm->full_len)
985 		len = stm->full_len;
986 
987 	err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
988 	if (err < 0)
989 		goto out;
990 
991 	copied = len;
992 	if (likely(!(flags & MSG_PEEK))) {
993 		KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
994 		if (copied < stm->full_len) {
995 			if (sock->type == SOCK_DGRAM) {
996 				/* Truncated message */
997 				msg->msg_flags |= MSG_TRUNC;
998 				goto msg_finished;
999 			}
1000 			stm->offset += copied;
1001 			stm->full_len -= copied;
1002 		} else {
1003 msg_finished:
1004 			/* Finished with message */
1005 			msg->msg_flags |= MSG_EOR;
1006 			KCM_STATS_INCR(kcm->stats.rx_msgs);
1007 		}
1008 	}
1009 
1010 out:
1011 	skb_free_datagram(sk, skb);
1012 	return copied ? : err;
1013 }
1014 
1015 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1016 			       struct pipe_inode_info *pipe, size_t len,
1017 			       unsigned int flags)
1018 {
1019 	struct sock *sk = sock->sk;
1020 	struct kcm_sock *kcm = kcm_sk(sk);
1021 	struct strp_msg *stm;
1022 	int err = 0;
1023 	ssize_t copied;
1024 	struct sk_buff *skb;
1025 
1026 	/* Only support splice for SOCKSEQPACKET */
1027 
1028 	skb = skb_recv_datagram(sk, flags, &err);
1029 	if (!skb)
1030 		goto err_out;
1031 
1032 	/* Okay, have a message on the receive queue */
1033 
1034 	stm = strp_msg(skb);
1035 
1036 	if (len > stm->full_len)
1037 		len = stm->full_len;
1038 
1039 	copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1040 	if (copied < 0) {
1041 		err = copied;
1042 		goto err_out;
1043 	}
1044 
1045 	KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1046 
1047 	stm->offset += copied;
1048 	stm->full_len -= copied;
1049 
1050 	/* We have no way to return MSG_EOR. If all the bytes have been
1051 	 * read we still leave the message in the receive socket buffer.
1052 	 * A subsequent recvmsg needs to be done to return MSG_EOR and
1053 	 * finish reading the message.
1054 	 */
1055 
1056 	skb_free_datagram(sk, skb);
1057 	return copied;
1058 
1059 err_out:
1060 	skb_free_datagram(sk, skb);
1061 	return err;
1062 }
1063 
1064 /* kcm sock lock held */
1065 static void kcm_recv_disable(struct kcm_sock *kcm)
1066 {
1067 	struct kcm_mux *mux = kcm->mux;
1068 
1069 	if (kcm->rx_disabled)
1070 		return;
1071 
1072 	spin_lock_bh(&mux->rx_lock);
1073 
1074 	kcm->rx_disabled = 1;
1075 
1076 	/* If a psock is reserved we'll do cleanup in unreserve */
1077 	if (!kcm->rx_psock) {
1078 		if (kcm->rx_wait) {
1079 			list_del(&kcm->wait_rx_list);
1080 			/* paired with lockless reads in kcm_rfree() */
1081 			WRITE_ONCE(kcm->rx_wait, false);
1082 		}
1083 
1084 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1085 	}
1086 
1087 	spin_unlock_bh(&mux->rx_lock);
1088 }
1089 
1090 /* kcm sock lock held */
1091 static void kcm_recv_enable(struct kcm_sock *kcm)
1092 {
1093 	struct kcm_mux *mux = kcm->mux;
1094 
1095 	if (!kcm->rx_disabled)
1096 		return;
1097 
1098 	spin_lock_bh(&mux->rx_lock);
1099 
1100 	kcm->rx_disabled = 0;
1101 	kcm_rcv_ready(kcm);
1102 
1103 	spin_unlock_bh(&mux->rx_lock);
1104 }
1105 
1106 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1107 			  sockptr_t optval, unsigned int optlen)
1108 {
1109 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1110 	int val, valbool;
1111 	int err = 0;
1112 
1113 	if (level != SOL_KCM)
1114 		return -ENOPROTOOPT;
1115 
1116 	if (optlen < sizeof(int))
1117 		return -EINVAL;
1118 
1119 	if (copy_from_sockptr(&val, optval, sizeof(int)))
1120 		return -EFAULT;
1121 
1122 	valbool = val ? 1 : 0;
1123 
1124 	switch (optname) {
1125 	case KCM_RECV_DISABLE:
1126 		lock_sock(&kcm->sk);
1127 		if (valbool)
1128 			kcm_recv_disable(kcm);
1129 		else
1130 			kcm_recv_enable(kcm);
1131 		release_sock(&kcm->sk);
1132 		break;
1133 	default:
1134 		err = -ENOPROTOOPT;
1135 	}
1136 
1137 	return err;
1138 }
1139 
1140 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1141 			  char __user *optval, int __user *optlen)
1142 {
1143 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1144 	int val, len;
1145 
1146 	if (level != SOL_KCM)
1147 		return -ENOPROTOOPT;
1148 
1149 	if (get_user(len, optlen))
1150 		return -EFAULT;
1151 
1152 	len = min_t(unsigned int, len, sizeof(int));
1153 	if (len < 0)
1154 		return -EINVAL;
1155 
1156 	switch (optname) {
1157 	case KCM_RECV_DISABLE:
1158 		val = kcm->rx_disabled;
1159 		break;
1160 	default:
1161 		return -ENOPROTOOPT;
1162 	}
1163 
1164 	if (put_user(len, optlen))
1165 		return -EFAULT;
1166 	if (copy_to_user(optval, &val, len))
1167 		return -EFAULT;
1168 	return 0;
1169 }
1170 
1171 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1172 {
1173 	struct kcm_sock *tkcm;
1174 	struct list_head *head;
1175 	int index = 0;
1176 
1177 	/* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1178 	 * we set sk_state, otherwise epoll_wait always returns right away with
1179 	 * EPOLLHUP
1180 	 */
1181 	kcm->sk.sk_state = TCP_ESTABLISHED;
1182 
1183 	/* Add to mux's kcm sockets list */
1184 	kcm->mux = mux;
1185 	spin_lock_bh(&mux->lock);
1186 
1187 	head = &mux->kcm_socks;
1188 	list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1189 		if (tkcm->index != index)
1190 			break;
1191 		head = &tkcm->kcm_sock_list;
1192 		index++;
1193 	}
1194 
1195 	list_add(&kcm->kcm_sock_list, head);
1196 	kcm->index = index;
1197 
1198 	mux->kcm_socks_cnt++;
1199 	spin_unlock_bh(&mux->lock);
1200 
1201 	INIT_WORK(&kcm->tx_work, kcm_tx_work);
1202 
1203 	spin_lock_bh(&mux->rx_lock);
1204 	kcm_rcv_ready(kcm);
1205 	spin_unlock_bh(&mux->rx_lock);
1206 }
1207 
1208 static int kcm_attach(struct socket *sock, struct socket *csock,
1209 		      struct bpf_prog *prog)
1210 {
1211 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1212 	struct kcm_mux *mux = kcm->mux;
1213 	struct sock *csk;
1214 	struct kcm_psock *psock = NULL, *tpsock;
1215 	struct list_head *head;
1216 	int index = 0;
1217 	static const struct strp_callbacks cb = {
1218 		.rcv_msg = kcm_rcv_strparser,
1219 		.parse_msg = kcm_parse_func_strparser,
1220 		.read_sock_done = kcm_read_sock_done,
1221 	};
1222 	int err = 0;
1223 
1224 	csk = csock->sk;
1225 	if (!csk)
1226 		return -EINVAL;
1227 
1228 	lock_sock(csk);
1229 
1230 	/* Only allow TCP sockets to be attached for now */
1231 	if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1232 	    csk->sk_protocol != IPPROTO_TCP) {
1233 		err = -EOPNOTSUPP;
1234 		goto out;
1235 	}
1236 
1237 	/* Don't allow listeners or closed sockets */
1238 	if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1239 		err = -EOPNOTSUPP;
1240 		goto out;
1241 	}
1242 
1243 	psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1244 	if (!psock) {
1245 		err = -ENOMEM;
1246 		goto out;
1247 	}
1248 
1249 	psock->mux = mux;
1250 	psock->sk = csk;
1251 	psock->bpf_prog = prog;
1252 
1253 	write_lock_bh(&csk->sk_callback_lock);
1254 
1255 	/* Check if sk_user_data is already by KCM or someone else.
1256 	 * Must be done under lock to prevent race conditions.
1257 	 */
1258 	if (csk->sk_user_data) {
1259 		write_unlock_bh(&csk->sk_callback_lock);
1260 		kmem_cache_free(kcm_psockp, psock);
1261 		err = -EALREADY;
1262 		goto out;
1263 	}
1264 
1265 	err = strp_init(&psock->strp, csk, &cb);
1266 	if (err) {
1267 		write_unlock_bh(&csk->sk_callback_lock);
1268 		kmem_cache_free(kcm_psockp, psock);
1269 		goto out;
1270 	}
1271 
1272 	psock->save_data_ready = csk->sk_data_ready;
1273 	psock->save_write_space = csk->sk_write_space;
1274 	psock->save_state_change = csk->sk_state_change;
1275 	csk->sk_user_data = psock;
1276 	csk->sk_data_ready = psock_data_ready;
1277 	csk->sk_write_space = psock_write_space;
1278 	csk->sk_state_change = psock_state_change;
1279 
1280 	write_unlock_bh(&csk->sk_callback_lock);
1281 
1282 	sock_hold(csk);
1283 
1284 	/* Finished initialization, now add the psock to the MUX. */
1285 	spin_lock_bh(&mux->lock);
1286 	head = &mux->psocks;
1287 	list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1288 		if (tpsock->index != index)
1289 			break;
1290 		head = &tpsock->psock_list;
1291 		index++;
1292 	}
1293 
1294 	list_add(&psock->psock_list, head);
1295 	psock->index = index;
1296 
1297 	KCM_STATS_INCR(mux->stats.psock_attach);
1298 	mux->psocks_cnt++;
1299 	psock_now_avail(psock);
1300 	spin_unlock_bh(&mux->lock);
1301 
1302 	/* Schedule RX work in case there are already bytes queued */
1303 	strp_check_rcv(&psock->strp);
1304 
1305 out:
1306 	release_sock(csk);
1307 
1308 	return err;
1309 }
1310 
1311 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1312 {
1313 	struct socket *csock;
1314 	struct bpf_prog *prog;
1315 	int err;
1316 
1317 	csock = sockfd_lookup(info->fd, &err);
1318 	if (!csock)
1319 		return -ENOENT;
1320 
1321 	prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1322 	if (IS_ERR(prog)) {
1323 		err = PTR_ERR(prog);
1324 		goto out;
1325 	}
1326 
1327 	err = kcm_attach(sock, csock, prog);
1328 	if (err) {
1329 		bpf_prog_put(prog);
1330 		goto out;
1331 	}
1332 
1333 	/* Keep reference on file also */
1334 
1335 	return 0;
1336 out:
1337 	sockfd_put(csock);
1338 	return err;
1339 }
1340 
1341 static void kcm_unattach(struct kcm_psock *psock)
1342 {
1343 	struct sock *csk = psock->sk;
1344 	struct kcm_mux *mux = psock->mux;
1345 
1346 	lock_sock(csk);
1347 
1348 	/* Stop getting callbacks from TCP socket. After this there should
1349 	 * be no way to reserve a kcm for this psock.
1350 	 */
1351 	write_lock_bh(&csk->sk_callback_lock);
1352 	csk->sk_user_data = NULL;
1353 	csk->sk_data_ready = psock->save_data_ready;
1354 	csk->sk_write_space = psock->save_write_space;
1355 	csk->sk_state_change = psock->save_state_change;
1356 	strp_stop(&psock->strp);
1357 
1358 	if (WARN_ON(psock->rx_kcm)) {
1359 		write_unlock_bh(&csk->sk_callback_lock);
1360 		release_sock(csk);
1361 		return;
1362 	}
1363 
1364 	spin_lock_bh(&mux->rx_lock);
1365 
1366 	/* Stop receiver activities. After this point psock should not be
1367 	 * able to get onto ready list either through callbacks or work.
1368 	 */
1369 	if (psock->ready_rx_msg) {
1370 		list_del(&psock->psock_ready_list);
1371 		kfree_skb(psock->ready_rx_msg);
1372 		psock->ready_rx_msg = NULL;
1373 		KCM_STATS_INCR(mux->stats.rx_ready_drops);
1374 	}
1375 
1376 	spin_unlock_bh(&mux->rx_lock);
1377 
1378 	write_unlock_bh(&csk->sk_callback_lock);
1379 
1380 	/* Call strp_done without sock lock */
1381 	release_sock(csk);
1382 	strp_done(&psock->strp);
1383 	lock_sock(csk);
1384 
1385 	bpf_prog_put(psock->bpf_prog);
1386 
1387 	spin_lock_bh(&mux->lock);
1388 
1389 	aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1390 	save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1391 
1392 	KCM_STATS_INCR(mux->stats.psock_unattach);
1393 
1394 	if (psock->tx_kcm) {
1395 		/* psock was reserved.  Just mark it finished and we will clean
1396 		 * up in the kcm paths, we need kcm lock which can not be
1397 		 * acquired here.
1398 		 */
1399 		KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1400 		spin_unlock_bh(&mux->lock);
1401 
1402 		/* We are unattaching a socket that is reserved. Abort the
1403 		 * socket since we may be out of sync in sending on it. We need
1404 		 * to do this without the mux lock.
1405 		 */
1406 		kcm_abort_tx_psock(psock, EPIPE, false);
1407 
1408 		spin_lock_bh(&mux->lock);
1409 		if (!psock->tx_kcm) {
1410 			/* psock now unreserved in window mux was unlocked */
1411 			goto no_reserved;
1412 		}
1413 		psock->done = 1;
1414 
1415 		/* Commit done before queuing work to process it */
1416 		smp_mb();
1417 
1418 		/* Queue tx work to make sure psock->done is handled */
1419 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1420 		spin_unlock_bh(&mux->lock);
1421 	} else {
1422 no_reserved:
1423 		if (!psock->tx_stopped)
1424 			list_del(&psock->psock_avail_list);
1425 		list_del(&psock->psock_list);
1426 		mux->psocks_cnt--;
1427 		spin_unlock_bh(&mux->lock);
1428 
1429 		sock_put(csk);
1430 		fput(csk->sk_socket->file);
1431 		kmem_cache_free(kcm_psockp, psock);
1432 	}
1433 
1434 	release_sock(csk);
1435 }
1436 
1437 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1438 {
1439 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1440 	struct kcm_mux *mux = kcm->mux;
1441 	struct kcm_psock *psock;
1442 	struct socket *csock;
1443 	struct sock *csk;
1444 	int err;
1445 
1446 	csock = sockfd_lookup(info->fd, &err);
1447 	if (!csock)
1448 		return -ENOENT;
1449 
1450 	csk = csock->sk;
1451 	if (!csk) {
1452 		err = -EINVAL;
1453 		goto out;
1454 	}
1455 
1456 	err = -ENOENT;
1457 
1458 	spin_lock_bh(&mux->lock);
1459 
1460 	list_for_each_entry(psock, &mux->psocks, psock_list) {
1461 		if (psock->sk != csk)
1462 			continue;
1463 
1464 		/* Found the matching psock */
1465 
1466 		if (psock->unattaching || WARN_ON(psock->done)) {
1467 			err = -EALREADY;
1468 			break;
1469 		}
1470 
1471 		psock->unattaching = 1;
1472 
1473 		spin_unlock_bh(&mux->lock);
1474 
1475 		/* Lower socket lock should already be held */
1476 		kcm_unattach(psock);
1477 
1478 		err = 0;
1479 		goto out;
1480 	}
1481 
1482 	spin_unlock_bh(&mux->lock);
1483 
1484 out:
1485 	sockfd_put(csock);
1486 	return err;
1487 }
1488 
1489 static struct proto kcm_proto = {
1490 	.name	= "KCM",
1491 	.owner	= THIS_MODULE,
1492 	.obj_size = sizeof(struct kcm_sock),
1493 };
1494 
1495 /* Clone a kcm socket. */
1496 static struct file *kcm_clone(struct socket *osock)
1497 {
1498 	struct socket *newsock;
1499 	struct sock *newsk;
1500 
1501 	newsock = sock_alloc();
1502 	if (!newsock)
1503 		return ERR_PTR(-ENFILE);
1504 
1505 	newsock->type = osock->type;
1506 	newsock->ops = osock->ops;
1507 
1508 	__module_get(newsock->ops->owner);
1509 
1510 	newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1511 			 &kcm_proto, false);
1512 	if (!newsk) {
1513 		sock_release(newsock);
1514 		return ERR_PTR(-ENOMEM);
1515 	}
1516 	sock_init_data(newsock, newsk);
1517 	init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1518 
1519 	return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1520 }
1521 
1522 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1523 {
1524 	int err;
1525 
1526 	switch (cmd) {
1527 	case SIOCKCMATTACH: {
1528 		struct kcm_attach info;
1529 
1530 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1531 			return -EFAULT;
1532 
1533 		err = kcm_attach_ioctl(sock, &info);
1534 
1535 		break;
1536 	}
1537 	case SIOCKCMUNATTACH: {
1538 		struct kcm_unattach info;
1539 
1540 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1541 			return -EFAULT;
1542 
1543 		err = kcm_unattach_ioctl(sock, &info);
1544 
1545 		break;
1546 	}
1547 	case SIOCKCMCLONE: {
1548 		struct kcm_clone info;
1549 		struct file *file;
1550 
1551 		info.fd = get_unused_fd_flags(0);
1552 		if (unlikely(info.fd < 0))
1553 			return info.fd;
1554 
1555 		file = kcm_clone(sock);
1556 		if (IS_ERR(file)) {
1557 			put_unused_fd(info.fd);
1558 			return PTR_ERR(file);
1559 		}
1560 		if (copy_to_user((void __user *)arg, &info,
1561 				 sizeof(info))) {
1562 			put_unused_fd(info.fd);
1563 			fput(file);
1564 			return -EFAULT;
1565 		}
1566 		fd_install(info.fd, file);
1567 		err = 0;
1568 		break;
1569 	}
1570 	default:
1571 		err = -ENOIOCTLCMD;
1572 		break;
1573 	}
1574 
1575 	return err;
1576 }
1577 
1578 static void free_mux(struct rcu_head *rcu)
1579 {
1580 	struct kcm_mux *mux = container_of(rcu,
1581 	    struct kcm_mux, rcu);
1582 
1583 	kmem_cache_free(kcm_muxp, mux);
1584 }
1585 
1586 static void release_mux(struct kcm_mux *mux)
1587 {
1588 	struct kcm_net *knet = mux->knet;
1589 	struct kcm_psock *psock, *tmp_psock;
1590 
1591 	/* Release psocks */
1592 	list_for_each_entry_safe(psock, tmp_psock,
1593 				 &mux->psocks, psock_list) {
1594 		if (!WARN_ON(psock->unattaching))
1595 			kcm_unattach(psock);
1596 	}
1597 
1598 	if (WARN_ON(mux->psocks_cnt))
1599 		return;
1600 
1601 	__skb_queue_purge(&mux->rx_hold_queue);
1602 
1603 	mutex_lock(&knet->mutex);
1604 	aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1605 	aggregate_psock_stats(&mux->aggregate_psock_stats,
1606 			      &knet->aggregate_psock_stats);
1607 	aggregate_strp_stats(&mux->aggregate_strp_stats,
1608 			     &knet->aggregate_strp_stats);
1609 	list_del_rcu(&mux->kcm_mux_list);
1610 	knet->count--;
1611 	mutex_unlock(&knet->mutex);
1612 
1613 	call_rcu(&mux->rcu, free_mux);
1614 }
1615 
1616 static void kcm_done(struct kcm_sock *kcm)
1617 {
1618 	struct kcm_mux *mux = kcm->mux;
1619 	struct sock *sk = &kcm->sk;
1620 	int socks_cnt;
1621 
1622 	spin_lock_bh(&mux->rx_lock);
1623 	if (kcm->rx_psock) {
1624 		/* Cleanup in unreserve_rx_kcm */
1625 		WARN_ON(kcm->done);
1626 		kcm->rx_disabled = 1;
1627 		kcm->done = 1;
1628 		spin_unlock_bh(&mux->rx_lock);
1629 		return;
1630 	}
1631 
1632 	if (kcm->rx_wait) {
1633 		list_del(&kcm->wait_rx_list);
1634 		/* paired with lockless reads in kcm_rfree() */
1635 		WRITE_ONCE(kcm->rx_wait, false);
1636 	}
1637 	/* Move any pending receive messages to other kcm sockets */
1638 	requeue_rx_msgs(mux, &sk->sk_receive_queue);
1639 
1640 	spin_unlock_bh(&mux->rx_lock);
1641 
1642 	if (WARN_ON(sk_rmem_alloc_get(sk)))
1643 		return;
1644 
1645 	/* Detach from MUX */
1646 	spin_lock_bh(&mux->lock);
1647 
1648 	list_del(&kcm->kcm_sock_list);
1649 	mux->kcm_socks_cnt--;
1650 	socks_cnt = mux->kcm_socks_cnt;
1651 
1652 	spin_unlock_bh(&mux->lock);
1653 
1654 	if (!socks_cnt) {
1655 		/* We are done with the mux now. */
1656 		release_mux(mux);
1657 	}
1658 
1659 	WARN_ON(kcm->rx_wait);
1660 
1661 	sock_put(&kcm->sk);
1662 }
1663 
1664 /* Called by kcm_release to close a KCM socket.
1665  * If this is the last KCM socket on the MUX, destroy the MUX.
1666  */
1667 static int kcm_release(struct socket *sock)
1668 {
1669 	struct sock *sk = sock->sk;
1670 	struct kcm_sock *kcm;
1671 	struct kcm_mux *mux;
1672 	struct kcm_psock *psock;
1673 
1674 	if (!sk)
1675 		return 0;
1676 
1677 	kcm = kcm_sk(sk);
1678 	mux = kcm->mux;
1679 
1680 	lock_sock(sk);
1681 	sock_orphan(sk);
1682 	kfree_skb(kcm->seq_skb);
1683 
1684 	/* Purge queue under lock to avoid race condition with tx_work trying
1685 	 * to act when queue is nonempty. If tx_work runs after this point
1686 	 * it will just return.
1687 	 */
1688 	__skb_queue_purge(&sk->sk_write_queue);
1689 
1690 	/* Set tx_stopped. This is checked when psock is bound to a kcm and we
1691 	 * get a writespace callback. This prevents further work being queued
1692 	 * from the callback (unbinding the psock occurs after canceling work.
1693 	 */
1694 	kcm->tx_stopped = 1;
1695 
1696 	release_sock(sk);
1697 
1698 	spin_lock_bh(&mux->lock);
1699 	if (kcm->tx_wait) {
1700 		/* Take of tx_wait list, after this point there should be no way
1701 		 * that a psock will be assigned to this kcm.
1702 		 */
1703 		list_del(&kcm->wait_psock_list);
1704 		kcm->tx_wait = false;
1705 	}
1706 	spin_unlock_bh(&mux->lock);
1707 
1708 	/* Cancel work. After this point there should be no outside references
1709 	 * to the kcm socket.
1710 	 */
1711 	cancel_work_sync(&kcm->tx_work);
1712 
1713 	lock_sock(sk);
1714 	psock = kcm->tx_psock;
1715 	if (psock) {
1716 		/* A psock was reserved, so we need to kill it since it
1717 		 * may already have some bytes queued from a message. We
1718 		 * need to do this after removing kcm from tx_wait list.
1719 		 */
1720 		kcm_abort_tx_psock(psock, EPIPE, false);
1721 		unreserve_psock(kcm);
1722 	}
1723 	release_sock(sk);
1724 
1725 	WARN_ON(kcm->tx_wait);
1726 	WARN_ON(kcm->tx_psock);
1727 
1728 	sock->sk = NULL;
1729 
1730 	kcm_done(kcm);
1731 
1732 	return 0;
1733 }
1734 
1735 static const struct proto_ops kcm_dgram_ops = {
1736 	.family =	PF_KCM,
1737 	.owner =	THIS_MODULE,
1738 	.release =	kcm_release,
1739 	.bind =		sock_no_bind,
1740 	.connect =	sock_no_connect,
1741 	.socketpair =	sock_no_socketpair,
1742 	.accept =	sock_no_accept,
1743 	.getname =	sock_no_getname,
1744 	.poll =		datagram_poll,
1745 	.ioctl =	kcm_ioctl,
1746 	.listen =	sock_no_listen,
1747 	.shutdown =	sock_no_shutdown,
1748 	.setsockopt =	kcm_setsockopt,
1749 	.getsockopt =	kcm_getsockopt,
1750 	.sendmsg =	kcm_sendmsg,
1751 	.recvmsg =	kcm_recvmsg,
1752 	.mmap =		sock_no_mmap,
1753 	.splice_eof =	kcm_splice_eof,
1754 };
1755 
1756 static const struct proto_ops kcm_seqpacket_ops = {
1757 	.family =	PF_KCM,
1758 	.owner =	THIS_MODULE,
1759 	.release =	kcm_release,
1760 	.bind =		sock_no_bind,
1761 	.connect =	sock_no_connect,
1762 	.socketpair =	sock_no_socketpair,
1763 	.accept =	sock_no_accept,
1764 	.getname =	sock_no_getname,
1765 	.poll =		datagram_poll,
1766 	.ioctl =	kcm_ioctl,
1767 	.listen =	sock_no_listen,
1768 	.shutdown =	sock_no_shutdown,
1769 	.setsockopt =	kcm_setsockopt,
1770 	.getsockopt =	kcm_getsockopt,
1771 	.sendmsg =	kcm_sendmsg,
1772 	.recvmsg =	kcm_recvmsg,
1773 	.mmap =		sock_no_mmap,
1774 	.splice_eof =	kcm_splice_eof,
1775 	.splice_read =	kcm_splice_read,
1776 };
1777 
1778 /* Create proto operation for kcm sockets */
1779 static int kcm_create(struct net *net, struct socket *sock,
1780 		      int protocol, int kern)
1781 {
1782 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1783 	struct sock *sk;
1784 	struct kcm_mux *mux;
1785 
1786 	switch (sock->type) {
1787 	case SOCK_DGRAM:
1788 		sock->ops = &kcm_dgram_ops;
1789 		break;
1790 	case SOCK_SEQPACKET:
1791 		sock->ops = &kcm_seqpacket_ops;
1792 		break;
1793 	default:
1794 		return -ESOCKTNOSUPPORT;
1795 	}
1796 
1797 	if (protocol != KCMPROTO_CONNECTED)
1798 		return -EPROTONOSUPPORT;
1799 
1800 	sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1801 	if (!sk)
1802 		return -ENOMEM;
1803 
1804 	/* Allocate a kcm mux, shared between KCM sockets */
1805 	mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1806 	if (!mux) {
1807 		sk_free(sk);
1808 		return -ENOMEM;
1809 	}
1810 
1811 	spin_lock_init(&mux->lock);
1812 	spin_lock_init(&mux->rx_lock);
1813 	INIT_LIST_HEAD(&mux->kcm_socks);
1814 	INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1815 	INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1816 
1817 	INIT_LIST_HEAD(&mux->psocks);
1818 	INIT_LIST_HEAD(&mux->psocks_ready);
1819 	INIT_LIST_HEAD(&mux->psocks_avail);
1820 
1821 	mux->knet = knet;
1822 
1823 	/* Add new MUX to list */
1824 	mutex_lock(&knet->mutex);
1825 	list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1826 	knet->count++;
1827 	mutex_unlock(&knet->mutex);
1828 
1829 	skb_queue_head_init(&mux->rx_hold_queue);
1830 
1831 	/* Init KCM socket */
1832 	sock_init_data(sock, sk);
1833 	init_kcm_sock(kcm_sk(sk), mux);
1834 
1835 	return 0;
1836 }
1837 
1838 static const struct net_proto_family kcm_family_ops = {
1839 	.family = PF_KCM,
1840 	.create = kcm_create,
1841 	.owner  = THIS_MODULE,
1842 };
1843 
1844 static __net_init int kcm_init_net(struct net *net)
1845 {
1846 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1847 
1848 	INIT_LIST_HEAD_RCU(&knet->mux_list);
1849 	mutex_init(&knet->mutex);
1850 
1851 	return 0;
1852 }
1853 
1854 static __net_exit void kcm_exit_net(struct net *net)
1855 {
1856 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1857 
1858 	/* All KCM sockets should be closed at this point, which should mean
1859 	 * that all multiplexors and psocks have been destroyed.
1860 	 */
1861 	WARN_ON(!list_empty(&knet->mux_list));
1862 
1863 	mutex_destroy(&knet->mutex);
1864 }
1865 
1866 static struct pernet_operations kcm_net_ops = {
1867 	.init = kcm_init_net,
1868 	.exit = kcm_exit_net,
1869 	.id   = &kcm_net_id,
1870 	.size = sizeof(struct kcm_net),
1871 };
1872 
1873 static int __init kcm_init(void)
1874 {
1875 	int err = -ENOMEM;
1876 
1877 	kcm_muxp = kmem_cache_create("kcm_mux_cache",
1878 				     sizeof(struct kcm_mux), 0,
1879 				     SLAB_HWCACHE_ALIGN, NULL);
1880 	if (!kcm_muxp)
1881 		goto fail;
1882 
1883 	kcm_psockp = kmem_cache_create("kcm_psock_cache",
1884 				       sizeof(struct kcm_psock), 0,
1885 					SLAB_HWCACHE_ALIGN, NULL);
1886 	if (!kcm_psockp)
1887 		goto fail;
1888 
1889 	kcm_wq = create_singlethread_workqueue("kkcmd");
1890 	if (!kcm_wq)
1891 		goto fail;
1892 
1893 	err = proto_register(&kcm_proto, 1);
1894 	if (err)
1895 		goto fail;
1896 
1897 	err = register_pernet_device(&kcm_net_ops);
1898 	if (err)
1899 		goto net_ops_fail;
1900 
1901 	err = sock_register(&kcm_family_ops);
1902 	if (err)
1903 		goto sock_register_fail;
1904 
1905 	err = kcm_proc_init();
1906 	if (err)
1907 		goto proc_init_fail;
1908 
1909 	return 0;
1910 
1911 proc_init_fail:
1912 	sock_unregister(PF_KCM);
1913 
1914 sock_register_fail:
1915 	unregister_pernet_device(&kcm_net_ops);
1916 
1917 net_ops_fail:
1918 	proto_unregister(&kcm_proto);
1919 
1920 fail:
1921 	kmem_cache_destroy(kcm_muxp);
1922 	kmem_cache_destroy(kcm_psockp);
1923 
1924 	if (kcm_wq)
1925 		destroy_workqueue(kcm_wq);
1926 
1927 	return err;
1928 }
1929 
1930 static void __exit kcm_exit(void)
1931 {
1932 	kcm_proc_exit();
1933 	sock_unregister(PF_KCM);
1934 	unregister_pernet_device(&kcm_net_ops);
1935 	proto_unregister(&kcm_proto);
1936 	destroy_workqueue(kcm_wq);
1937 
1938 	kmem_cache_destroy(kcm_muxp);
1939 	kmem_cache_destroy(kcm_psockp);
1940 }
1941 
1942 module_init(kcm_init);
1943 module_exit(kcm_exit);
1944 
1945 MODULE_LICENSE("GPL");
1946 MODULE_ALIAS_NETPROTO(PF_KCM);
1947