xref: /linux/net/kcm/kcmsock.c (revision 8bc7c5e525584903ea83332e18a2118ed3b1985e)
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 		    WARN_ON_ONCE(!skb_frag_page(&skb_shinfo(skb)->frags[0]))) {
632 			ret = -EINVAL;
633 			goto out;
634 		}
635 
636 		msize = 0;
637 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
638 			msize += skb_frag_size(&skb_shinfo(skb)->frags[i]);
639 
640 		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE,
641 			      (const struct bio_vec *)skb_shinfo(skb)->frags,
642 			      skb_shinfo(skb)->nr_frags, msize);
643 		iov_iter_advance(&msg.msg_iter, txm->frag_offset);
644 
645 		do {
646 			ret = sock_sendmsg(psock->sk->sk_socket, &msg);
647 			if (ret <= 0) {
648 				if (ret == -EAGAIN) {
649 					/* Save state to try again when there's
650 					 * write space on the socket
651 					 */
652 					txm->frag_skb = skb;
653 					ret = 0;
654 					goto out;
655 				}
656 
657 				/* Hard failure in sending message, abort this
658 				 * psock since it has lost framing
659 				 * synchronization and retry sending the
660 				 * message from the beginning.
661 				 */
662 				kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
663 						   true);
664 				unreserve_psock(kcm);
665 				psock = NULL;
666 
667 				txm->started_tx = false;
668 				kcm_report_tx_retry(kcm);
669 				ret = 0;
670 				goto retry;
671 			}
672 
673 			txm->sent += ret;
674 			txm->frag_offset += ret;
675 			KCM_STATS_ADD(psock->stats.tx_bytes, ret);
676 		} while (msg.msg_iter.count > 0);
677 
678 		if (skb == head) {
679 			if (skb_has_frag_list(skb)) {
680 				txm->frag_skb = skb_shinfo(skb)->frag_list;
681 				txm->frag_offset = 0;
682 				continue;
683 			}
684 		} else if (skb->next) {
685 			txm->frag_skb = skb->next;
686 			txm->frag_offset = 0;
687 			continue;
688 		}
689 
690 		/* Successfully sent the whole packet, account for it. */
691 		sk->sk_wmem_queued -= txm->sent;
692 		total_sent += txm->sent;
693 		skb_dequeue(&sk->sk_write_queue);
694 		kfree_skb(head);
695 		KCM_STATS_INCR(psock->stats.tx_msgs);
696 	}
697 out:
698 	if (!head) {
699 		/* Done with all queued messages. */
700 		WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
701 		if (psock)
702 			unreserve_psock(kcm);
703 	}
704 
705 	/* Check if write space is available */
706 	sk->sk_write_space(sk);
707 
708 	return total_sent ? : ret;
709 }
710 
711 static void kcm_tx_work(struct work_struct *w)
712 {
713 	struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
714 	struct sock *sk = &kcm->sk;
715 	int err;
716 
717 	lock_sock(sk);
718 
719 	/* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
720 	 * aborts
721 	 */
722 	err = kcm_write_msgs(kcm);
723 	if (err < 0) {
724 		/* Hard failure in write, report error on KCM socket */
725 		pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
726 		report_csk_error(&kcm->sk, -err);
727 		goto out;
728 	}
729 
730 	/* Primarily for SOCK_SEQPACKET sockets */
731 	if (likely(sk->sk_socket) &&
732 	    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
733 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
734 		sk->sk_write_space(sk);
735 	}
736 
737 out:
738 	release_sock(sk);
739 }
740 
741 static void kcm_push(struct kcm_sock *kcm)
742 {
743 	if (kcm->tx_wait_more)
744 		kcm_write_msgs(kcm);
745 }
746 
747 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
748 {
749 	struct sock *sk = sock->sk;
750 	struct kcm_sock *kcm = kcm_sk(sk);
751 	struct sk_buff *skb = NULL, *head = NULL;
752 	size_t copy, copied = 0;
753 	long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
754 	int eor = (sock->type == SOCK_DGRAM) ?
755 		  !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
756 	int err = -EPIPE;
757 
758 	mutex_lock(&kcm->tx_mutex);
759 	lock_sock(sk);
760 
761 	/* Per tcp_sendmsg this should be in poll */
762 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
763 
764 	if (sk->sk_err)
765 		goto out_error;
766 
767 	if (kcm->seq_skb) {
768 		/* Previously opened message */
769 		head = kcm->seq_skb;
770 		skb = kcm_tx_msg(head)->last_skb;
771 		goto start;
772 	}
773 
774 	/* Call the sk_stream functions to manage the sndbuf mem. */
775 	if (!sk_stream_memory_free(sk)) {
776 		kcm_push(kcm);
777 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
778 		err = sk_stream_wait_memory(sk, &timeo);
779 		if (err)
780 			goto out_error;
781 	}
782 
783 	if (msg_data_left(msg)) {
784 		/* New message, alloc head skb */
785 		head = alloc_skb(0, sk->sk_allocation);
786 		while (!head) {
787 			kcm_push(kcm);
788 			err = sk_stream_wait_memory(sk, &timeo);
789 			if (err)
790 				goto out_error;
791 
792 			head = alloc_skb(0, sk->sk_allocation);
793 		}
794 
795 		skb = head;
796 
797 		/* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
798 		 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
799 		 */
800 		skb->ip_summed = CHECKSUM_UNNECESSARY;
801 	}
802 
803 start:
804 	while (msg_data_left(msg)) {
805 		bool merge = true;
806 		int i = skb_shinfo(skb)->nr_frags;
807 		struct page_frag *pfrag = sk_page_frag(sk);
808 
809 		if (!sk_page_frag_refill(sk, pfrag))
810 			goto wait_for_memory;
811 
812 		if (!skb_can_coalesce(skb, i, pfrag->page,
813 				      pfrag->offset)) {
814 			if (i == MAX_SKB_FRAGS) {
815 				struct sk_buff *tskb;
816 
817 				tskb = alloc_skb(0, sk->sk_allocation);
818 				if (!tskb)
819 					goto wait_for_memory;
820 
821 				if (head == skb)
822 					skb_shinfo(head)->frag_list = tskb;
823 				else
824 					skb->next = tskb;
825 
826 				skb = tskb;
827 				skb->ip_summed = CHECKSUM_UNNECESSARY;
828 				continue;
829 			}
830 			merge = false;
831 		}
832 
833 		if (msg->msg_flags & MSG_SPLICE_PAGES) {
834 			copy = msg_data_left(msg);
835 			if (!sk_wmem_schedule(sk, copy))
836 				goto wait_for_memory;
837 
838 			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
839 						   sk->sk_allocation);
840 			if (err < 0) {
841 				if (err == -EMSGSIZE)
842 					goto wait_for_memory;
843 				goto out_error;
844 			}
845 
846 			copy = err;
847 			skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
848 			sk_wmem_queued_add(sk, copy);
849 			sk_mem_charge(sk, copy);
850 
851 			if (head != skb)
852 				head->truesize += copy;
853 		} else {
854 			copy = min_t(int, msg_data_left(msg),
855 				     pfrag->size - pfrag->offset);
856 			if (!sk_wmem_schedule(sk, copy))
857 				goto wait_for_memory;
858 
859 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
860 						       pfrag->page,
861 						       pfrag->offset,
862 						       copy);
863 			if (err)
864 				goto out_error;
865 
866 			/* Update the skb. */
867 			if (merge) {
868 				skb_frag_size_add(
869 					&skb_shinfo(skb)->frags[i - 1], copy);
870 			} else {
871 				skb_fill_page_desc(skb, i, pfrag->page,
872 						   pfrag->offset, copy);
873 				get_page(pfrag->page);
874 			}
875 
876 			pfrag->offset += copy;
877 		}
878 
879 		copied += copy;
880 		if (head != skb) {
881 			head->len += copy;
882 			head->data_len += copy;
883 		}
884 
885 		continue;
886 
887 wait_for_memory:
888 		kcm_push(kcm);
889 		err = sk_stream_wait_memory(sk, &timeo);
890 		if (err)
891 			goto out_error;
892 	}
893 
894 	if (eor) {
895 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
896 
897 		if (head) {
898 			/* Message complete, queue it on send buffer */
899 			__skb_queue_tail(&sk->sk_write_queue, head);
900 			kcm->seq_skb = NULL;
901 			KCM_STATS_INCR(kcm->stats.tx_msgs);
902 		}
903 
904 		if (msg->msg_flags & MSG_BATCH) {
905 			kcm->tx_wait_more = true;
906 		} else if (kcm->tx_wait_more || not_busy) {
907 			err = kcm_write_msgs(kcm);
908 			if (err < 0) {
909 				/* We got a hard error in write_msgs but have
910 				 * already queued this message. Report an error
911 				 * in the socket, but don't affect return value
912 				 * from sendmsg
913 				 */
914 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
915 				report_csk_error(&kcm->sk, -err);
916 			}
917 		}
918 	} else {
919 		/* Message not complete, save state */
920 partial_message:
921 		if (head) {
922 			kcm->seq_skb = head;
923 			kcm_tx_msg(head)->last_skb = skb;
924 		}
925 	}
926 
927 	KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
928 
929 	release_sock(sk);
930 	mutex_unlock(&kcm->tx_mutex);
931 	return copied;
932 
933 out_error:
934 	kcm_push(kcm);
935 
936 	if (sock->type == SOCK_SEQPACKET) {
937 		/* Wrote some bytes before encountering an
938 		 * error, return partial success.
939 		 */
940 		if (copied)
941 			goto partial_message;
942 		if (head != kcm->seq_skb)
943 			kfree_skb(head);
944 	} else {
945 		kfree_skb(head);
946 		kcm->seq_skb = NULL;
947 	}
948 
949 	err = sk_stream_error(sk, msg->msg_flags, err);
950 
951 	/* make sure we wake any epoll edge trigger waiter */
952 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
953 		sk->sk_write_space(sk);
954 
955 	release_sock(sk);
956 	mutex_unlock(&kcm->tx_mutex);
957 	return err;
958 }
959 
960 static void kcm_splice_eof(struct socket *sock)
961 {
962 	struct sock *sk = sock->sk;
963 	struct kcm_sock *kcm = kcm_sk(sk);
964 
965 	if (skb_queue_empty_lockless(&sk->sk_write_queue))
966 		return;
967 
968 	lock_sock(sk);
969 	kcm_write_msgs(kcm);
970 	release_sock(sk);
971 }
972 
973 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
974 		       size_t len, int flags)
975 {
976 	struct sock *sk = sock->sk;
977 	struct kcm_sock *kcm = kcm_sk(sk);
978 	int err = 0;
979 	struct strp_msg *stm;
980 	int copied = 0;
981 	struct sk_buff *skb;
982 
983 	skb = skb_recv_datagram(sk, flags, &err);
984 	if (!skb)
985 		goto out;
986 
987 	/* Okay, have a message on the receive queue */
988 
989 	stm = strp_msg(skb);
990 
991 	if (len > stm->full_len)
992 		len = stm->full_len;
993 
994 	err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
995 	if (err < 0)
996 		goto out;
997 
998 	copied = len;
999 	if (likely(!(flags & MSG_PEEK))) {
1000 		KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1001 		if (copied < stm->full_len) {
1002 			if (sock->type == SOCK_DGRAM) {
1003 				/* Truncated message */
1004 				msg->msg_flags |= MSG_TRUNC;
1005 				goto msg_finished;
1006 			}
1007 			stm->offset += copied;
1008 			stm->full_len -= copied;
1009 		} else {
1010 msg_finished:
1011 			/* Finished with message */
1012 			msg->msg_flags |= MSG_EOR;
1013 			KCM_STATS_INCR(kcm->stats.rx_msgs);
1014 		}
1015 	}
1016 
1017 out:
1018 	skb_free_datagram(sk, skb);
1019 	return copied ? : err;
1020 }
1021 
1022 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1023 			       struct pipe_inode_info *pipe, size_t len,
1024 			       unsigned int flags)
1025 {
1026 	struct sock *sk = sock->sk;
1027 	struct kcm_sock *kcm = kcm_sk(sk);
1028 	struct strp_msg *stm;
1029 	int err = 0;
1030 	ssize_t copied;
1031 	struct sk_buff *skb;
1032 
1033 	/* Only support splice for SOCKSEQPACKET */
1034 
1035 	skb = skb_recv_datagram(sk, flags, &err);
1036 	if (!skb)
1037 		goto err_out;
1038 
1039 	/* Okay, have a message on the receive queue */
1040 
1041 	stm = strp_msg(skb);
1042 
1043 	if (len > stm->full_len)
1044 		len = stm->full_len;
1045 
1046 	copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1047 	if (copied < 0) {
1048 		err = copied;
1049 		goto err_out;
1050 	}
1051 
1052 	KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1053 
1054 	stm->offset += copied;
1055 	stm->full_len -= copied;
1056 
1057 	/* We have no way to return MSG_EOR. If all the bytes have been
1058 	 * read we still leave the message in the receive socket buffer.
1059 	 * A subsequent recvmsg needs to be done to return MSG_EOR and
1060 	 * finish reading the message.
1061 	 */
1062 
1063 	skb_free_datagram(sk, skb);
1064 	return copied;
1065 
1066 err_out:
1067 	skb_free_datagram(sk, skb);
1068 	return err;
1069 }
1070 
1071 /* kcm sock lock held */
1072 static void kcm_recv_disable(struct kcm_sock *kcm)
1073 {
1074 	struct kcm_mux *mux = kcm->mux;
1075 
1076 	if (kcm->rx_disabled)
1077 		return;
1078 
1079 	spin_lock_bh(&mux->rx_lock);
1080 
1081 	kcm->rx_disabled = 1;
1082 
1083 	/* If a psock is reserved we'll do cleanup in unreserve */
1084 	if (!kcm->rx_psock) {
1085 		if (kcm->rx_wait) {
1086 			list_del(&kcm->wait_rx_list);
1087 			/* paired with lockless reads in kcm_rfree() */
1088 			WRITE_ONCE(kcm->rx_wait, false);
1089 		}
1090 
1091 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1092 	}
1093 
1094 	spin_unlock_bh(&mux->rx_lock);
1095 }
1096 
1097 /* kcm sock lock held */
1098 static void kcm_recv_enable(struct kcm_sock *kcm)
1099 {
1100 	struct kcm_mux *mux = kcm->mux;
1101 
1102 	if (!kcm->rx_disabled)
1103 		return;
1104 
1105 	spin_lock_bh(&mux->rx_lock);
1106 
1107 	kcm->rx_disabled = 0;
1108 	kcm_rcv_ready(kcm);
1109 
1110 	spin_unlock_bh(&mux->rx_lock);
1111 }
1112 
1113 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1114 			  sockptr_t optval, unsigned int optlen)
1115 {
1116 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1117 	int val, valbool;
1118 	int err = 0;
1119 
1120 	if (level != SOL_KCM)
1121 		return -ENOPROTOOPT;
1122 
1123 	if (optlen < sizeof(int))
1124 		return -EINVAL;
1125 
1126 	if (copy_from_sockptr(&val, optval, sizeof(int)))
1127 		return -EFAULT;
1128 
1129 	valbool = val ? 1 : 0;
1130 
1131 	switch (optname) {
1132 	case KCM_RECV_DISABLE:
1133 		lock_sock(&kcm->sk);
1134 		if (valbool)
1135 			kcm_recv_disable(kcm);
1136 		else
1137 			kcm_recv_enable(kcm);
1138 		release_sock(&kcm->sk);
1139 		break;
1140 	default:
1141 		err = -ENOPROTOOPT;
1142 	}
1143 
1144 	return err;
1145 }
1146 
1147 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1148 			  char __user *optval, int __user *optlen)
1149 {
1150 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1151 	int val, len;
1152 
1153 	if (level != SOL_KCM)
1154 		return -ENOPROTOOPT;
1155 
1156 	if (get_user(len, optlen))
1157 		return -EFAULT;
1158 
1159 	if (len < 0)
1160 		return -EINVAL;
1161 
1162 	len = min_t(unsigned int, len, sizeof(int));
1163 
1164 	switch (optname) {
1165 	case KCM_RECV_DISABLE:
1166 		val = kcm->rx_disabled;
1167 		break;
1168 	default:
1169 		return -ENOPROTOOPT;
1170 	}
1171 
1172 	if (put_user(len, optlen))
1173 		return -EFAULT;
1174 	if (copy_to_user(optval, &val, len))
1175 		return -EFAULT;
1176 	return 0;
1177 }
1178 
1179 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1180 {
1181 	struct kcm_sock *tkcm;
1182 	struct list_head *head;
1183 	int index = 0;
1184 
1185 	/* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1186 	 * we set sk_state, otherwise epoll_wait always returns right away with
1187 	 * EPOLLHUP
1188 	 */
1189 	kcm->sk.sk_state = TCP_ESTABLISHED;
1190 
1191 	/* Add to mux's kcm sockets list */
1192 	kcm->mux = mux;
1193 	spin_lock_bh(&mux->lock);
1194 
1195 	head = &mux->kcm_socks;
1196 	list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1197 		if (tkcm->index != index)
1198 			break;
1199 		head = &tkcm->kcm_sock_list;
1200 		index++;
1201 	}
1202 
1203 	list_add(&kcm->kcm_sock_list, head);
1204 	kcm->index = index;
1205 
1206 	mux->kcm_socks_cnt++;
1207 	spin_unlock_bh(&mux->lock);
1208 
1209 	INIT_WORK(&kcm->tx_work, kcm_tx_work);
1210 	mutex_init(&kcm->tx_mutex);
1211 
1212 	spin_lock_bh(&mux->rx_lock);
1213 	kcm_rcv_ready(kcm);
1214 	spin_unlock_bh(&mux->rx_lock);
1215 }
1216 
1217 static int kcm_attach(struct socket *sock, struct socket *csock,
1218 		      struct bpf_prog *prog)
1219 {
1220 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1221 	struct kcm_mux *mux = kcm->mux;
1222 	struct sock *csk;
1223 	struct kcm_psock *psock = NULL, *tpsock;
1224 	struct list_head *head;
1225 	int index = 0;
1226 	static const struct strp_callbacks cb = {
1227 		.rcv_msg = kcm_rcv_strparser,
1228 		.parse_msg = kcm_parse_func_strparser,
1229 		.read_sock_done = kcm_read_sock_done,
1230 	};
1231 	int err = 0;
1232 
1233 	csk = csock->sk;
1234 	if (!csk)
1235 		return -EINVAL;
1236 
1237 	lock_sock(csk);
1238 
1239 	/* Only allow TCP sockets to be attached for now */
1240 	if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1241 	    csk->sk_protocol != IPPROTO_TCP) {
1242 		err = -EOPNOTSUPP;
1243 		goto out;
1244 	}
1245 
1246 	/* Don't allow listeners or closed sockets */
1247 	if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1248 		err = -EOPNOTSUPP;
1249 		goto out;
1250 	}
1251 
1252 	psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1253 	if (!psock) {
1254 		err = -ENOMEM;
1255 		goto out;
1256 	}
1257 
1258 	psock->mux = mux;
1259 	psock->sk = csk;
1260 	psock->bpf_prog = prog;
1261 
1262 	write_lock_bh(&csk->sk_callback_lock);
1263 
1264 	/* Check if sk_user_data is already by KCM or someone else.
1265 	 * Must be done under lock to prevent race conditions.
1266 	 */
1267 	if (csk->sk_user_data) {
1268 		write_unlock_bh(&csk->sk_callback_lock);
1269 		kmem_cache_free(kcm_psockp, psock);
1270 		err = -EALREADY;
1271 		goto out;
1272 	}
1273 
1274 	err = strp_init(&psock->strp, csk, &cb);
1275 	if (err) {
1276 		write_unlock_bh(&csk->sk_callback_lock);
1277 		kmem_cache_free(kcm_psockp, psock);
1278 		goto out;
1279 	}
1280 
1281 	psock->save_data_ready = csk->sk_data_ready;
1282 	psock->save_write_space = csk->sk_write_space;
1283 	psock->save_state_change = csk->sk_state_change;
1284 	csk->sk_user_data = psock;
1285 	csk->sk_data_ready = psock_data_ready;
1286 	csk->sk_write_space = psock_write_space;
1287 	csk->sk_state_change = psock_state_change;
1288 
1289 	write_unlock_bh(&csk->sk_callback_lock);
1290 
1291 	sock_hold(csk);
1292 
1293 	/* Finished initialization, now add the psock to the MUX. */
1294 	spin_lock_bh(&mux->lock);
1295 	head = &mux->psocks;
1296 	list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1297 		if (tpsock->index != index)
1298 			break;
1299 		head = &tpsock->psock_list;
1300 		index++;
1301 	}
1302 
1303 	list_add(&psock->psock_list, head);
1304 	psock->index = index;
1305 
1306 	KCM_STATS_INCR(mux->stats.psock_attach);
1307 	mux->psocks_cnt++;
1308 	psock_now_avail(psock);
1309 	spin_unlock_bh(&mux->lock);
1310 
1311 	/* Schedule RX work in case there are already bytes queued */
1312 	strp_check_rcv(&psock->strp);
1313 
1314 out:
1315 	release_sock(csk);
1316 
1317 	return err;
1318 }
1319 
1320 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1321 {
1322 	struct socket *csock;
1323 	struct bpf_prog *prog;
1324 	int err;
1325 
1326 	csock = sockfd_lookup(info->fd, &err);
1327 	if (!csock)
1328 		return -ENOENT;
1329 
1330 	prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1331 	if (IS_ERR(prog)) {
1332 		err = PTR_ERR(prog);
1333 		goto out;
1334 	}
1335 
1336 	err = kcm_attach(sock, csock, prog);
1337 	if (err) {
1338 		bpf_prog_put(prog);
1339 		goto out;
1340 	}
1341 
1342 	/* Keep reference on file also */
1343 
1344 	return 0;
1345 out:
1346 	sockfd_put(csock);
1347 	return err;
1348 }
1349 
1350 static void kcm_unattach(struct kcm_psock *psock)
1351 {
1352 	struct sock *csk = psock->sk;
1353 	struct kcm_mux *mux = psock->mux;
1354 
1355 	lock_sock(csk);
1356 
1357 	/* Stop getting callbacks from TCP socket. After this there should
1358 	 * be no way to reserve a kcm for this psock.
1359 	 */
1360 	write_lock_bh(&csk->sk_callback_lock);
1361 	csk->sk_user_data = NULL;
1362 	csk->sk_data_ready = psock->save_data_ready;
1363 	csk->sk_write_space = psock->save_write_space;
1364 	csk->sk_state_change = psock->save_state_change;
1365 	strp_stop(&psock->strp);
1366 
1367 	if (WARN_ON(psock->rx_kcm)) {
1368 		write_unlock_bh(&csk->sk_callback_lock);
1369 		release_sock(csk);
1370 		return;
1371 	}
1372 
1373 	spin_lock_bh(&mux->rx_lock);
1374 
1375 	/* Stop receiver activities. After this point psock should not be
1376 	 * able to get onto ready list either through callbacks or work.
1377 	 */
1378 	if (psock->ready_rx_msg) {
1379 		list_del(&psock->psock_ready_list);
1380 		kfree_skb(psock->ready_rx_msg);
1381 		psock->ready_rx_msg = NULL;
1382 		KCM_STATS_INCR(mux->stats.rx_ready_drops);
1383 	}
1384 
1385 	spin_unlock_bh(&mux->rx_lock);
1386 
1387 	write_unlock_bh(&csk->sk_callback_lock);
1388 
1389 	/* Call strp_done without sock lock */
1390 	release_sock(csk);
1391 	strp_done(&psock->strp);
1392 	lock_sock(csk);
1393 
1394 	bpf_prog_put(psock->bpf_prog);
1395 
1396 	spin_lock_bh(&mux->lock);
1397 
1398 	aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1399 	save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1400 
1401 	KCM_STATS_INCR(mux->stats.psock_unattach);
1402 
1403 	if (psock->tx_kcm) {
1404 		/* psock was reserved.  Just mark it finished and we will clean
1405 		 * up in the kcm paths, we need kcm lock which can not be
1406 		 * acquired here.
1407 		 */
1408 		KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1409 		spin_unlock_bh(&mux->lock);
1410 
1411 		/* We are unattaching a socket that is reserved. Abort the
1412 		 * socket since we may be out of sync in sending on it. We need
1413 		 * to do this without the mux lock.
1414 		 */
1415 		kcm_abort_tx_psock(psock, EPIPE, false);
1416 
1417 		spin_lock_bh(&mux->lock);
1418 		if (!psock->tx_kcm) {
1419 			/* psock now unreserved in window mux was unlocked */
1420 			goto no_reserved;
1421 		}
1422 		psock->done = 1;
1423 
1424 		/* Commit done before queuing work to process it */
1425 		smp_mb();
1426 
1427 		/* Queue tx work to make sure psock->done is handled */
1428 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1429 		spin_unlock_bh(&mux->lock);
1430 	} else {
1431 no_reserved:
1432 		if (!psock->tx_stopped)
1433 			list_del(&psock->psock_avail_list);
1434 		list_del(&psock->psock_list);
1435 		mux->psocks_cnt--;
1436 		spin_unlock_bh(&mux->lock);
1437 
1438 		sock_put(csk);
1439 		fput(csk->sk_socket->file);
1440 		kmem_cache_free(kcm_psockp, psock);
1441 	}
1442 
1443 	release_sock(csk);
1444 }
1445 
1446 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1447 {
1448 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1449 	struct kcm_mux *mux = kcm->mux;
1450 	struct kcm_psock *psock;
1451 	struct socket *csock;
1452 	struct sock *csk;
1453 	int err;
1454 
1455 	csock = sockfd_lookup(info->fd, &err);
1456 	if (!csock)
1457 		return -ENOENT;
1458 
1459 	csk = csock->sk;
1460 	if (!csk) {
1461 		err = -EINVAL;
1462 		goto out;
1463 	}
1464 
1465 	err = -ENOENT;
1466 
1467 	spin_lock_bh(&mux->lock);
1468 
1469 	list_for_each_entry(psock, &mux->psocks, psock_list) {
1470 		if (psock->sk != csk)
1471 			continue;
1472 
1473 		/* Found the matching psock */
1474 
1475 		if (psock->unattaching || WARN_ON(psock->done)) {
1476 			err = -EALREADY;
1477 			break;
1478 		}
1479 
1480 		psock->unattaching = 1;
1481 
1482 		spin_unlock_bh(&mux->lock);
1483 
1484 		/* Lower socket lock should already be held */
1485 		kcm_unattach(psock);
1486 
1487 		err = 0;
1488 		goto out;
1489 	}
1490 
1491 	spin_unlock_bh(&mux->lock);
1492 
1493 out:
1494 	sockfd_put(csock);
1495 	return err;
1496 }
1497 
1498 static struct proto kcm_proto = {
1499 	.name	= "KCM",
1500 	.owner	= THIS_MODULE,
1501 	.obj_size = sizeof(struct kcm_sock),
1502 };
1503 
1504 /* Clone a kcm socket. */
1505 static struct file *kcm_clone(struct socket *osock)
1506 {
1507 	struct socket *newsock;
1508 	struct sock *newsk;
1509 
1510 	newsock = sock_alloc();
1511 	if (!newsock)
1512 		return ERR_PTR(-ENFILE);
1513 
1514 	newsock->type = osock->type;
1515 	newsock->ops = osock->ops;
1516 
1517 	__module_get(newsock->ops->owner);
1518 
1519 	newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1520 			 &kcm_proto, false);
1521 	if (!newsk) {
1522 		sock_release(newsock);
1523 		return ERR_PTR(-ENOMEM);
1524 	}
1525 	sock_init_data(newsock, newsk);
1526 	init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1527 
1528 	return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1529 }
1530 
1531 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1532 {
1533 	int err;
1534 
1535 	switch (cmd) {
1536 	case SIOCKCMATTACH: {
1537 		struct kcm_attach info;
1538 
1539 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1540 			return -EFAULT;
1541 
1542 		err = kcm_attach_ioctl(sock, &info);
1543 
1544 		break;
1545 	}
1546 	case SIOCKCMUNATTACH: {
1547 		struct kcm_unattach info;
1548 
1549 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1550 			return -EFAULT;
1551 
1552 		err = kcm_unattach_ioctl(sock, &info);
1553 
1554 		break;
1555 	}
1556 	case SIOCKCMCLONE: {
1557 		struct kcm_clone info;
1558 		struct file *file;
1559 
1560 		info.fd = get_unused_fd_flags(0);
1561 		if (unlikely(info.fd < 0))
1562 			return info.fd;
1563 
1564 		file = kcm_clone(sock);
1565 		if (IS_ERR(file)) {
1566 			put_unused_fd(info.fd);
1567 			return PTR_ERR(file);
1568 		}
1569 		if (copy_to_user((void __user *)arg, &info,
1570 				 sizeof(info))) {
1571 			put_unused_fd(info.fd);
1572 			fput(file);
1573 			return -EFAULT;
1574 		}
1575 		fd_install(info.fd, file);
1576 		err = 0;
1577 		break;
1578 	}
1579 	default:
1580 		err = -ENOIOCTLCMD;
1581 		break;
1582 	}
1583 
1584 	return err;
1585 }
1586 
1587 static void free_mux(struct rcu_head *rcu)
1588 {
1589 	struct kcm_mux *mux = container_of(rcu,
1590 	    struct kcm_mux, rcu);
1591 
1592 	kmem_cache_free(kcm_muxp, mux);
1593 }
1594 
1595 static void release_mux(struct kcm_mux *mux)
1596 {
1597 	struct kcm_net *knet = mux->knet;
1598 	struct kcm_psock *psock, *tmp_psock;
1599 
1600 	/* Release psocks */
1601 	list_for_each_entry_safe(psock, tmp_psock,
1602 				 &mux->psocks, psock_list) {
1603 		if (!WARN_ON(psock->unattaching))
1604 			kcm_unattach(psock);
1605 	}
1606 
1607 	if (WARN_ON(mux->psocks_cnt))
1608 		return;
1609 
1610 	__skb_queue_purge(&mux->rx_hold_queue);
1611 
1612 	mutex_lock(&knet->mutex);
1613 	aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1614 	aggregate_psock_stats(&mux->aggregate_psock_stats,
1615 			      &knet->aggregate_psock_stats);
1616 	aggregate_strp_stats(&mux->aggregate_strp_stats,
1617 			     &knet->aggregate_strp_stats);
1618 	list_del_rcu(&mux->kcm_mux_list);
1619 	knet->count--;
1620 	mutex_unlock(&knet->mutex);
1621 
1622 	call_rcu(&mux->rcu, free_mux);
1623 }
1624 
1625 static void kcm_done(struct kcm_sock *kcm)
1626 {
1627 	struct kcm_mux *mux = kcm->mux;
1628 	struct sock *sk = &kcm->sk;
1629 	int socks_cnt;
1630 
1631 	spin_lock_bh(&mux->rx_lock);
1632 	if (kcm->rx_psock) {
1633 		/* Cleanup in unreserve_rx_kcm */
1634 		WARN_ON(kcm->done);
1635 		kcm->rx_disabled = 1;
1636 		kcm->done = 1;
1637 		spin_unlock_bh(&mux->rx_lock);
1638 		return;
1639 	}
1640 
1641 	if (kcm->rx_wait) {
1642 		list_del(&kcm->wait_rx_list);
1643 		/* paired with lockless reads in kcm_rfree() */
1644 		WRITE_ONCE(kcm->rx_wait, false);
1645 	}
1646 	/* Move any pending receive messages to other kcm sockets */
1647 	requeue_rx_msgs(mux, &sk->sk_receive_queue);
1648 
1649 	spin_unlock_bh(&mux->rx_lock);
1650 
1651 	if (WARN_ON(sk_rmem_alloc_get(sk)))
1652 		return;
1653 
1654 	/* Detach from MUX */
1655 	spin_lock_bh(&mux->lock);
1656 
1657 	list_del(&kcm->kcm_sock_list);
1658 	mux->kcm_socks_cnt--;
1659 	socks_cnt = mux->kcm_socks_cnt;
1660 
1661 	spin_unlock_bh(&mux->lock);
1662 
1663 	if (!socks_cnt) {
1664 		/* We are done with the mux now. */
1665 		release_mux(mux);
1666 	}
1667 
1668 	WARN_ON(kcm->rx_wait);
1669 
1670 	sock_put(&kcm->sk);
1671 }
1672 
1673 /* Called by kcm_release to close a KCM socket.
1674  * If this is the last KCM socket on the MUX, destroy the MUX.
1675  */
1676 static int kcm_release(struct socket *sock)
1677 {
1678 	struct sock *sk = sock->sk;
1679 	struct kcm_sock *kcm;
1680 	struct kcm_mux *mux;
1681 	struct kcm_psock *psock;
1682 
1683 	if (!sk)
1684 		return 0;
1685 
1686 	kcm = kcm_sk(sk);
1687 	mux = kcm->mux;
1688 
1689 	lock_sock(sk);
1690 	sock_orphan(sk);
1691 	kfree_skb(kcm->seq_skb);
1692 
1693 	/* Purge queue under lock to avoid race condition with tx_work trying
1694 	 * to act when queue is nonempty. If tx_work runs after this point
1695 	 * it will just return.
1696 	 */
1697 	__skb_queue_purge(&sk->sk_write_queue);
1698 
1699 	/* Set tx_stopped. This is checked when psock is bound to a kcm and we
1700 	 * get a writespace callback. This prevents further work being queued
1701 	 * from the callback (unbinding the psock occurs after canceling work.
1702 	 */
1703 	kcm->tx_stopped = 1;
1704 
1705 	release_sock(sk);
1706 
1707 	spin_lock_bh(&mux->lock);
1708 	if (kcm->tx_wait) {
1709 		/* Take of tx_wait list, after this point there should be no way
1710 		 * that a psock will be assigned to this kcm.
1711 		 */
1712 		list_del(&kcm->wait_psock_list);
1713 		kcm->tx_wait = false;
1714 	}
1715 	spin_unlock_bh(&mux->lock);
1716 
1717 	/* Cancel work. After this point there should be no outside references
1718 	 * to the kcm socket.
1719 	 */
1720 	cancel_work_sync(&kcm->tx_work);
1721 
1722 	lock_sock(sk);
1723 	psock = kcm->tx_psock;
1724 	if (psock) {
1725 		/* A psock was reserved, so we need to kill it since it
1726 		 * may already have some bytes queued from a message. We
1727 		 * need to do this after removing kcm from tx_wait list.
1728 		 */
1729 		kcm_abort_tx_psock(psock, EPIPE, false);
1730 		unreserve_psock(kcm);
1731 	}
1732 	release_sock(sk);
1733 
1734 	WARN_ON(kcm->tx_wait);
1735 	WARN_ON(kcm->tx_psock);
1736 
1737 	sock->sk = NULL;
1738 
1739 	kcm_done(kcm);
1740 
1741 	return 0;
1742 }
1743 
1744 static const struct proto_ops kcm_dgram_ops = {
1745 	.family =	PF_KCM,
1746 	.owner =	THIS_MODULE,
1747 	.release =	kcm_release,
1748 	.bind =		sock_no_bind,
1749 	.connect =	sock_no_connect,
1750 	.socketpair =	sock_no_socketpair,
1751 	.accept =	sock_no_accept,
1752 	.getname =	sock_no_getname,
1753 	.poll =		datagram_poll,
1754 	.ioctl =	kcm_ioctl,
1755 	.listen =	sock_no_listen,
1756 	.shutdown =	sock_no_shutdown,
1757 	.setsockopt =	kcm_setsockopt,
1758 	.getsockopt =	kcm_getsockopt,
1759 	.sendmsg =	kcm_sendmsg,
1760 	.recvmsg =	kcm_recvmsg,
1761 	.mmap =		sock_no_mmap,
1762 	.splice_eof =	kcm_splice_eof,
1763 };
1764 
1765 static const struct proto_ops kcm_seqpacket_ops = {
1766 	.family =	PF_KCM,
1767 	.owner =	THIS_MODULE,
1768 	.release =	kcm_release,
1769 	.bind =		sock_no_bind,
1770 	.connect =	sock_no_connect,
1771 	.socketpair =	sock_no_socketpair,
1772 	.accept =	sock_no_accept,
1773 	.getname =	sock_no_getname,
1774 	.poll =		datagram_poll,
1775 	.ioctl =	kcm_ioctl,
1776 	.listen =	sock_no_listen,
1777 	.shutdown =	sock_no_shutdown,
1778 	.setsockopt =	kcm_setsockopt,
1779 	.getsockopt =	kcm_getsockopt,
1780 	.sendmsg =	kcm_sendmsg,
1781 	.recvmsg =	kcm_recvmsg,
1782 	.mmap =		sock_no_mmap,
1783 	.splice_eof =	kcm_splice_eof,
1784 	.splice_read =	kcm_splice_read,
1785 };
1786 
1787 /* Create proto operation for kcm sockets */
1788 static int kcm_create(struct net *net, struct socket *sock,
1789 		      int protocol, int kern)
1790 {
1791 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1792 	struct sock *sk;
1793 	struct kcm_mux *mux;
1794 
1795 	switch (sock->type) {
1796 	case SOCK_DGRAM:
1797 		sock->ops = &kcm_dgram_ops;
1798 		break;
1799 	case SOCK_SEQPACKET:
1800 		sock->ops = &kcm_seqpacket_ops;
1801 		break;
1802 	default:
1803 		return -ESOCKTNOSUPPORT;
1804 	}
1805 
1806 	if (protocol != KCMPROTO_CONNECTED)
1807 		return -EPROTONOSUPPORT;
1808 
1809 	sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1810 	if (!sk)
1811 		return -ENOMEM;
1812 
1813 	/* Allocate a kcm mux, shared between KCM sockets */
1814 	mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1815 	if (!mux) {
1816 		sk_free(sk);
1817 		return -ENOMEM;
1818 	}
1819 
1820 	spin_lock_init(&mux->lock);
1821 	spin_lock_init(&mux->rx_lock);
1822 	INIT_LIST_HEAD(&mux->kcm_socks);
1823 	INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1824 	INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1825 
1826 	INIT_LIST_HEAD(&mux->psocks);
1827 	INIT_LIST_HEAD(&mux->psocks_ready);
1828 	INIT_LIST_HEAD(&mux->psocks_avail);
1829 
1830 	mux->knet = knet;
1831 
1832 	/* Add new MUX to list */
1833 	mutex_lock(&knet->mutex);
1834 	list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1835 	knet->count++;
1836 	mutex_unlock(&knet->mutex);
1837 
1838 	skb_queue_head_init(&mux->rx_hold_queue);
1839 
1840 	/* Init KCM socket */
1841 	sock_init_data(sock, sk);
1842 	init_kcm_sock(kcm_sk(sk), mux);
1843 
1844 	return 0;
1845 }
1846 
1847 static const struct net_proto_family kcm_family_ops = {
1848 	.family = PF_KCM,
1849 	.create = kcm_create,
1850 	.owner  = THIS_MODULE,
1851 };
1852 
1853 static __net_init int kcm_init_net(struct net *net)
1854 {
1855 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1856 
1857 	INIT_LIST_HEAD_RCU(&knet->mux_list);
1858 	mutex_init(&knet->mutex);
1859 
1860 	return 0;
1861 }
1862 
1863 static __net_exit void kcm_exit_net(struct net *net)
1864 {
1865 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1866 
1867 	/* All KCM sockets should be closed at this point, which should mean
1868 	 * that all multiplexors and psocks have been destroyed.
1869 	 */
1870 	WARN_ON(!list_empty(&knet->mux_list));
1871 
1872 	mutex_destroy(&knet->mutex);
1873 }
1874 
1875 static struct pernet_operations kcm_net_ops = {
1876 	.init = kcm_init_net,
1877 	.exit = kcm_exit_net,
1878 	.id   = &kcm_net_id,
1879 	.size = sizeof(struct kcm_net),
1880 };
1881 
1882 static int __init kcm_init(void)
1883 {
1884 	int err = -ENOMEM;
1885 
1886 	kcm_muxp = KMEM_CACHE(kcm_mux, SLAB_HWCACHE_ALIGN);
1887 	if (!kcm_muxp)
1888 		goto fail;
1889 
1890 	kcm_psockp = KMEM_CACHE(kcm_psock, SLAB_HWCACHE_ALIGN);
1891 	if (!kcm_psockp)
1892 		goto fail;
1893 
1894 	kcm_wq = create_singlethread_workqueue("kkcmd");
1895 	if (!kcm_wq)
1896 		goto fail;
1897 
1898 	err = proto_register(&kcm_proto, 1);
1899 	if (err)
1900 		goto fail;
1901 
1902 	err = register_pernet_device(&kcm_net_ops);
1903 	if (err)
1904 		goto net_ops_fail;
1905 
1906 	err = sock_register(&kcm_family_ops);
1907 	if (err)
1908 		goto sock_register_fail;
1909 
1910 	err = kcm_proc_init();
1911 	if (err)
1912 		goto proc_init_fail;
1913 
1914 	return 0;
1915 
1916 proc_init_fail:
1917 	sock_unregister(PF_KCM);
1918 
1919 sock_register_fail:
1920 	unregister_pernet_device(&kcm_net_ops);
1921 
1922 net_ops_fail:
1923 	proto_unregister(&kcm_proto);
1924 
1925 fail:
1926 	kmem_cache_destroy(kcm_muxp);
1927 	kmem_cache_destroy(kcm_psockp);
1928 
1929 	if (kcm_wq)
1930 		destroy_workqueue(kcm_wq);
1931 
1932 	return err;
1933 }
1934 
1935 static void __exit kcm_exit(void)
1936 {
1937 	kcm_proc_exit();
1938 	sock_unregister(PF_KCM);
1939 	unregister_pernet_device(&kcm_net_ops);
1940 	proto_unregister(&kcm_proto);
1941 	destroy_workqueue(kcm_wq);
1942 
1943 	kmem_cache_destroy(kcm_muxp);
1944 	kmem_cache_destroy(kcm_psockp);
1945 }
1946 
1947 module_init(kcm_init);
1948 module_exit(kcm_exit);
1949 
1950 MODULE_LICENSE("GPL");
1951 MODULE_DESCRIPTION("KCM (Kernel Connection Multiplexor) sockets");
1952 MODULE_ALIAS_NETPROTO(PF_KCM);
1953