xref: /openbmc/linux/net/kcm/kcmsock.c (revision babbdf5b)
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/in.h>
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/net.h>
16 #include <linux/netdevice.h>
17 #include <linux/poll.h>
18 #include <linux/rculist.h>
19 #include <linux/skbuff.h>
20 #include <linux/socket.h>
21 #include <linux/uaccess.h>
22 #include <linux/workqueue.h>
23 #include <linux/syscalls.h>
24 #include <linux/sched/signal.h>
25 
26 #include <net/kcm.h>
27 #include <net/netns/generic.h>
28 #include <net/sock.h>
29 #include <uapi/linux/kcm.h>
30 
31 unsigned int kcm_net_id;
32 
33 static struct kmem_cache *kcm_psockp __read_mostly;
34 static struct kmem_cache *kcm_muxp __read_mostly;
35 static struct workqueue_struct *kcm_wq;
36 
37 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
38 {
39 	return (struct kcm_sock *)sk;
40 }
41 
42 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
43 {
44 	return (struct kcm_tx_msg *)skb->cb;
45 }
46 
47 static void report_csk_error(struct sock *csk, int err)
48 {
49 	csk->sk_err = EPIPE;
50 	csk->sk_error_report(csk);
51 }
52 
53 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
54 			       bool wakeup_kcm)
55 {
56 	struct sock *csk = psock->sk;
57 	struct kcm_mux *mux = psock->mux;
58 
59 	/* Unrecoverable error in transmit */
60 
61 	spin_lock_bh(&mux->lock);
62 
63 	if (psock->tx_stopped) {
64 		spin_unlock_bh(&mux->lock);
65 		return;
66 	}
67 
68 	psock->tx_stopped = 1;
69 	KCM_STATS_INCR(psock->stats.tx_aborts);
70 
71 	if (!psock->tx_kcm) {
72 		/* Take off psocks_avail list */
73 		list_del(&psock->psock_avail_list);
74 	} else if (wakeup_kcm) {
75 		/* In this case psock is being aborted while outside of
76 		 * write_msgs and psock is reserved. Schedule tx_work
77 		 * to handle the failure there. Need to commit tx_stopped
78 		 * before queuing work.
79 		 */
80 		smp_mb();
81 
82 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
83 	}
84 
85 	spin_unlock_bh(&mux->lock);
86 
87 	/* Report error on lower socket */
88 	report_csk_error(csk, err);
89 }
90 
91 /* RX mux lock held. */
92 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
93 				    struct kcm_psock *psock)
94 {
95 	STRP_STATS_ADD(mux->stats.rx_bytes,
96 		       psock->strp.stats.bytes -
97 		       psock->saved_rx_bytes);
98 	mux->stats.rx_msgs +=
99 		psock->strp.stats.msgs - psock->saved_rx_msgs;
100 	psock->saved_rx_msgs = psock->strp.stats.msgs;
101 	psock->saved_rx_bytes = psock->strp.stats.bytes;
102 }
103 
104 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
105 				    struct kcm_psock *psock)
106 {
107 	KCM_STATS_ADD(mux->stats.tx_bytes,
108 		      psock->stats.tx_bytes - psock->saved_tx_bytes);
109 	mux->stats.tx_msgs +=
110 		psock->stats.tx_msgs - psock->saved_tx_msgs;
111 	psock->saved_tx_msgs = psock->stats.tx_msgs;
112 	psock->saved_tx_bytes = psock->stats.tx_bytes;
113 }
114 
115 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
116 
117 /* KCM is ready to receive messages on its queue-- either the KCM is new or
118  * has become unblocked after being blocked on full socket buffer. Queue any
119  * pending ready messages on a psock. RX mux lock held.
120  */
121 static void kcm_rcv_ready(struct kcm_sock *kcm)
122 {
123 	struct kcm_mux *mux = kcm->mux;
124 	struct kcm_psock *psock;
125 	struct sk_buff *skb;
126 
127 	if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
128 		return;
129 
130 	while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
131 		if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
132 			/* Assuming buffer limit has been reached */
133 			skb_queue_head(&mux->rx_hold_queue, skb);
134 			WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
135 			return;
136 		}
137 	}
138 
139 	while (!list_empty(&mux->psocks_ready)) {
140 		psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
141 					 psock_ready_list);
142 
143 		if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
144 			/* Assuming buffer limit has been reached */
145 			WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
146 			return;
147 		}
148 
149 		/* Consumed the ready message on the psock. Schedule rx_work to
150 		 * get more messages.
151 		 */
152 		list_del(&psock->psock_ready_list);
153 		psock->ready_rx_msg = NULL;
154 		/* Commit clearing of ready_rx_msg for queuing work */
155 		smp_mb();
156 
157 		strp_unpause(&psock->strp);
158 		strp_check_rcv(&psock->strp);
159 	}
160 
161 	/* Buffer limit is okay now, add to ready list */
162 	list_add_tail(&kcm->wait_rx_list,
163 		      &kcm->mux->kcm_rx_waiters);
164 	kcm->rx_wait = true;
165 }
166 
167 static void kcm_rfree(struct sk_buff *skb)
168 {
169 	struct sock *sk = skb->sk;
170 	struct kcm_sock *kcm = kcm_sk(sk);
171 	struct kcm_mux *mux = kcm->mux;
172 	unsigned int len = skb->truesize;
173 
174 	sk_mem_uncharge(sk, len);
175 	atomic_sub(len, &sk->sk_rmem_alloc);
176 
177 	/* For reading rx_wait and rx_psock without holding lock */
178 	smp_mb__after_atomic();
179 
180 	if (!kcm->rx_wait && !kcm->rx_psock &&
181 	    sk_rmem_alloc_get(sk) < sk->sk_rcvlowat) {
182 		spin_lock_bh(&mux->rx_lock);
183 		kcm_rcv_ready(kcm);
184 		spin_unlock_bh(&mux->rx_lock);
185 	}
186 }
187 
188 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
189 {
190 	struct sk_buff_head *list = &sk->sk_receive_queue;
191 
192 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
193 		return -ENOMEM;
194 
195 	if (!sk_rmem_schedule(sk, skb, skb->truesize))
196 		return -ENOBUFS;
197 
198 	skb->dev = NULL;
199 
200 	skb_orphan(skb);
201 	skb->sk = sk;
202 	skb->destructor = kcm_rfree;
203 	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
204 	sk_mem_charge(sk, skb->truesize);
205 
206 	skb_queue_tail(list, skb);
207 
208 	if (!sock_flag(sk, SOCK_DEAD))
209 		sk->sk_data_ready(sk);
210 
211 	return 0;
212 }
213 
214 /* Requeue received messages for a kcm socket to other kcm sockets. This is
215  * called with a kcm socket is receive disabled.
216  * RX mux lock held.
217  */
218 static void requeue_rx_msgs(struct kcm_mux *mux, struct sk_buff_head *head)
219 {
220 	struct sk_buff *skb;
221 	struct kcm_sock *kcm;
222 
223 	while ((skb = __skb_dequeue(head))) {
224 		/* Reset destructor to avoid calling kcm_rcv_ready */
225 		skb->destructor = sock_rfree;
226 		skb_orphan(skb);
227 try_again:
228 		if (list_empty(&mux->kcm_rx_waiters)) {
229 			skb_queue_tail(&mux->rx_hold_queue, skb);
230 			continue;
231 		}
232 
233 		kcm = list_first_entry(&mux->kcm_rx_waiters,
234 				       struct kcm_sock, wait_rx_list);
235 
236 		if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
237 			/* Should mean socket buffer full */
238 			list_del(&kcm->wait_rx_list);
239 			kcm->rx_wait = false;
240 
241 			/* Commit rx_wait to read in kcm_free */
242 			smp_wmb();
243 
244 			goto try_again;
245 		}
246 	}
247 }
248 
249 /* Lower sock lock held */
250 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
251 				       struct sk_buff *head)
252 {
253 	struct kcm_mux *mux = psock->mux;
254 	struct kcm_sock *kcm;
255 
256 	WARN_ON(psock->ready_rx_msg);
257 
258 	if (psock->rx_kcm)
259 		return psock->rx_kcm;
260 
261 	spin_lock_bh(&mux->rx_lock);
262 
263 	if (psock->rx_kcm) {
264 		spin_unlock_bh(&mux->rx_lock);
265 		return psock->rx_kcm;
266 	}
267 
268 	kcm_update_rx_mux_stats(mux, psock);
269 
270 	if (list_empty(&mux->kcm_rx_waiters)) {
271 		psock->ready_rx_msg = head;
272 		strp_pause(&psock->strp);
273 		list_add_tail(&psock->psock_ready_list,
274 			      &mux->psocks_ready);
275 		spin_unlock_bh(&mux->rx_lock);
276 		return NULL;
277 	}
278 
279 	kcm = list_first_entry(&mux->kcm_rx_waiters,
280 			       struct kcm_sock, wait_rx_list);
281 	list_del(&kcm->wait_rx_list);
282 	kcm->rx_wait = false;
283 
284 	psock->rx_kcm = kcm;
285 	kcm->rx_psock = psock;
286 
287 	spin_unlock_bh(&mux->rx_lock);
288 
289 	return kcm;
290 }
291 
292 static void kcm_done(struct kcm_sock *kcm);
293 
294 static void kcm_done_work(struct work_struct *w)
295 {
296 	kcm_done(container_of(w, struct kcm_sock, done_work));
297 }
298 
299 /* Lower sock held */
300 static void unreserve_rx_kcm(struct kcm_psock *psock,
301 			     bool rcv_ready)
302 {
303 	struct kcm_sock *kcm = psock->rx_kcm;
304 	struct kcm_mux *mux = psock->mux;
305 
306 	if (!kcm)
307 		return;
308 
309 	spin_lock_bh(&mux->rx_lock);
310 
311 	psock->rx_kcm = NULL;
312 	kcm->rx_psock = NULL;
313 
314 	/* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
315 	 * kcm_rfree
316 	 */
317 	smp_mb();
318 
319 	if (unlikely(kcm->done)) {
320 		spin_unlock_bh(&mux->rx_lock);
321 
322 		/* Need to run kcm_done in a task since we need to qcquire
323 		 * callback locks which may already be held here.
324 		 */
325 		INIT_WORK(&kcm->done_work, kcm_done_work);
326 		schedule_work(&kcm->done_work);
327 		return;
328 	}
329 
330 	if (unlikely(kcm->rx_disabled)) {
331 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
332 	} else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
333 		/* Check for degenerative race with rx_wait that all
334 		 * data was dequeued (accounted for in kcm_rfree).
335 		 */
336 		kcm_rcv_ready(kcm);
337 	}
338 	spin_unlock_bh(&mux->rx_lock);
339 }
340 
341 /* Lower sock lock held */
342 static void psock_data_ready(struct sock *sk)
343 {
344 	struct kcm_psock *psock;
345 
346 	read_lock_bh(&sk->sk_callback_lock);
347 
348 	psock = (struct kcm_psock *)sk->sk_user_data;
349 	if (likely(psock))
350 		strp_data_ready(&psock->strp);
351 
352 	read_unlock_bh(&sk->sk_callback_lock);
353 }
354 
355 /* Called with lower sock held */
356 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
357 {
358 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
359 	struct kcm_sock *kcm;
360 
361 try_queue:
362 	kcm = reserve_rx_kcm(psock, skb);
363 	if (!kcm) {
364 		 /* Unable to reserve a KCM, message is held in psock and strp
365 		  * is paused.
366 		  */
367 		return;
368 	}
369 
370 	if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
371 		/* Should mean socket buffer full */
372 		unreserve_rx_kcm(psock, false);
373 		goto try_queue;
374 	}
375 }
376 
377 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
378 {
379 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
380 	struct bpf_prog *prog = psock->bpf_prog;
381 	int res;
382 
383 	res = bpf_prog_run_pin_on_cpu(prog, skb);
384 	return res;
385 }
386 
387 static int kcm_read_sock_done(struct strparser *strp, int err)
388 {
389 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
390 
391 	unreserve_rx_kcm(psock, true);
392 
393 	return err;
394 }
395 
396 static void psock_state_change(struct sock *sk)
397 {
398 	/* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
399 	 * since application will normally not poll with EPOLLIN
400 	 * on the TCP sockets.
401 	 */
402 
403 	report_csk_error(sk, EPIPE);
404 }
405 
406 static void psock_write_space(struct sock *sk)
407 {
408 	struct kcm_psock *psock;
409 	struct kcm_mux *mux;
410 	struct kcm_sock *kcm;
411 
412 	read_lock_bh(&sk->sk_callback_lock);
413 
414 	psock = (struct kcm_psock *)sk->sk_user_data;
415 	if (unlikely(!psock))
416 		goto out;
417 	mux = psock->mux;
418 
419 	spin_lock_bh(&mux->lock);
420 
421 	/* Check if the socket is reserved so someone is waiting for sending. */
422 	kcm = psock->tx_kcm;
423 	if (kcm && !unlikely(kcm->tx_stopped))
424 		queue_work(kcm_wq, &kcm->tx_work);
425 
426 	spin_unlock_bh(&mux->lock);
427 out:
428 	read_unlock_bh(&sk->sk_callback_lock);
429 }
430 
431 static void unreserve_psock(struct kcm_sock *kcm);
432 
433 /* kcm sock is locked. */
434 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
435 {
436 	struct kcm_mux *mux = kcm->mux;
437 	struct kcm_psock *psock;
438 
439 	psock = kcm->tx_psock;
440 
441 	smp_rmb(); /* Must read tx_psock before tx_wait */
442 
443 	if (psock) {
444 		WARN_ON(kcm->tx_wait);
445 		if (unlikely(psock->tx_stopped))
446 			unreserve_psock(kcm);
447 		else
448 			return kcm->tx_psock;
449 	}
450 
451 	spin_lock_bh(&mux->lock);
452 
453 	/* Check again under lock to see if psock was reserved for this
454 	 * psock via psock_unreserve.
455 	 */
456 	psock = kcm->tx_psock;
457 	if (unlikely(psock)) {
458 		WARN_ON(kcm->tx_wait);
459 		spin_unlock_bh(&mux->lock);
460 		return kcm->tx_psock;
461 	}
462 
463 	if (!list_empty(&mux->psocks_avail)) {
464 		psock = list_first_entry(&mux->psocks_avail,
465 					 struct kcm_psock,
466 					 psock_avail_list);
467 		list_del(&psock->psock_avail_list);
468 		if (kcm->tx_wait) {
469 			list_del(&kcm->wait_psock_list);
470 			kcm->tx_wait = false;
471 		}
472 		kcm->tx_psock = psock;
473 		psock->tx_kcm = kcm;
474 		KCM_STATS_INCR(psock->stats.reserved);
475 	} else if (!kcm->tx_wait) {
476 		list_add_tail(&kcm->wait_psock_list,
477 			      &mux->kcm_tx_waiters);
478 		kcm->tx_wait = true;
479 	}
480 
481 	spin_unlock_bh(&mux->lock);
482 
483 	return psock;
484 }
485 
486 /* mux lock held */
487 static void psock_now_avail(struct kcm_psock *psock)
488 {
489 	struct kcm_mux *mux = psock->mux;
490 	struct kcm_sock *kcm;
491 
492 	if (list_empty(&mux->kcm_tx_waiters)) {
493 		list_add_tail(&psock->psock_avail_list,
494 			      &mux->psocks_avail);
495 	} else {
496 		kcm = list_first_entry(&mux->kcm_tx_waiters,
497 				       struct kcm_sock,
498 				       wait_psock_list);
499 		list_del(&kcm->wait_psock_list);
500 		kcm->tx_wait = false;
501 		psock->tx_kcm = kcm;
502 
503 		/* Commit before changing tx_psock since that is read in
504 		 * reserve_psock before queuing work.
505 		 */
506 		smp_mb();
507 
508 		kcm->tx_psock = psock;
509 		KCM_STATS_INCR(psock->stats.reserved);
510 		queue_work(kcm_wq, &kcm->tx_work);
511 	}
512 }
513 
514 /* kcm sock is locked. */
515 static void unreserve_psock(struct kcm_sock *kcm)
516 {
517 	struct kcm_psock *psock;
518 	struct kcm_mux *mux = kcm->mux;
519 
520 	spin_lock_bh(&mux->lock);
521 
522 	psock = kcm->tx_psock;
523 
524 	if (WARN_ON(!psock)) {
525 		spin_unlock_bh(&mux->lock);
526 		return;
527 	}
528 
529 	smp_rmb(); /* Read tx_psock before tx_wait */
530 
531 	kcm_update_tx_mux_stats(mux, psock);
532 
533 	WARN_ON(kcm->tx_wait);
534 
535 	kcm->tx_psock = NULL;
536 	psock->tx_kcm = NULL;
537 	KCM_STATS_INCR(psock->stats.unreserved);
538 
539 	if (unlikely(psock->tx_stopped)) {
540 		if (psock->done) {
541 			/* Deferred free */
542 			list_del(&psock->psock_list);
543 			mux->psocks_cnt--;
544 			sock_put(psock->sk);
545 			fput(psock->sk->sk_socket->file);
546 			kmem_cache_free(kcm_psockp, psock);
547 		}
548 
549 		/* Don't put back on available list */
550 
551 		spin_unlock_bh(&mux->lock);
552 
553 		return;
554 	}
555 
556 	psock_now_avail(psock);
557 
558 	spin_unlock_bh(&mux->lock);
559 }
560 
561 static void kcm_report_tx_retry(struct kcm_sock *kcm)
562 {
563 	struct kcm_mux *mux = kcm->mux;
564 
565 	spin_lock_bh(&mux->lock);
566 	KCM_STATS_INCR(mux->stats.tx_retries);
567 	spin_unlock_bh(&mux->lock);
568 }
569 
570 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
571  * held.  Return bytes actually sent or error.
572  */
573 static int kcm_write_msgs(struct kcm_sock *kcm)
574 {
575 	struct sock *sk = &kcm->sk;
576 	struct kcm_psock *psock;
577 	struct sk_buff *skb, *head;
578 	struct kcm_tx_msg *txm;
579 	unsigned short fragidx, frag_offset;
580 	unsigned int sent, total_sent = 0;
581 	int ret = 0;
582 
583 	kcm->tx_wait_more = false;
584 	psock = kcm->tx_psock;
585 	if (unlikely(psock && psock->tx_stopped)) {
586 		/* A reserved psock was aborted asynchronously. Unreserve
587 		 * it and we'll retry the message.
588 		 */
589 		unreserve_psock(kcm);
590 		kcm_report_tx_retry(kcm);
591 		if (skb_queue_empty(&sk->sk_write_queue))
592 			return 0;
593 
594 		kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
595 
596 	} else if (skb_queue_empty(&sk->sk_write_queue)) {
597 		return 0;
598 	}
599 
600 	head = skb_peek(&sk->sk_write_queue);
601 	txm = kcm_tx_msg(head);
602 
603 	if (txm->sent) {
604 		/* Send of first skbuff in queue already in progress */
605 		if (WARN_ON(!psock)) {
606 			ret = -EINVAL;
607 			goto out;
608 		}
609 		sent = txm->sent;
610 		frag_offset = txm->frag_offset;
611 		fragidx = txm->fragidx;
612 		skb = txm->frag_skb;
613 
614 		goto do_frag;
615 	}
616 
617 try_again:
618 	psock = reserve_psock(kcm);
619 	if (!psock)
620 		goto out;
621 
622 	do {
623 		skb = head;
624 		txm = kcm_tx_msg(head);
625 		sent = 0;
626 
627 do_frag_list:
628 		if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
629 			ret = -EINVAL;
630 			goto out;
631 		}
632 
633 		for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
634 		     fragidx++) {
635 			skb_frag_t *frag;
636 
637 			frag_offset = 0;
638 do_frag:
639 			frag = &skb_shinfo(skb)->frags[fragidx];
640 			if (WARN_ON(!skb_frag_size(frag))) {
641 				ret = -EINVAL;
642 				goto out;
643 			}
644 
645 			ret = kernel_sendpage(psock->sk->sk_socket,
646 					      skb_frag_page(frag),
647 					      skb_frag_off(frag) + frag_offset,
648 					      skb_frag_size(frag) - frag_offset,
649 					      MSG_DONTWAIT);
650 			if (ret <= 0) {
651 				if (ret == -EAGAIN) {
652 					/* Save state to try again when there's
653 					 * write space on the socket
654 					 */
655 					txm->sent = sent;
656 					txm->frag_offset = frag_offset;
657 					txm->fragidx = fragidx;
658 					txm->frag_skb = skb;
659 
660 					ret = 0;
661 					goto out;
662 				}
663 
664 				/* Hard failure in sending message, abort this
665 				 * psock since it has lost framing
666 				 * synchronization and retry sending the
667 				 * message from the beginning.
668 				 */
669 				kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
670 						   true);
671 				unreserve_psock(kcm);
672 
673 				txm->sent = 0;
674 				kcm_report_tx_retry(kcm);
675 				ret = 0;
676 
677 				goto try_again;
678 			}
679 
680 			sent += ret;
681 			frag_offset += ret;
682 			KCM_STATS_ADD(psock->stats.tx_bytes, ret);
683 			if (frag_offset < skb_frag_size(frag)) {
684 				/* Not finished with this frag */
685 				goto do_frag;
686 			}
687 		}
688 
689 		if (skb == head) {
690 			if (skb_has_frag_list(skb)) {
691 				skb = skb_shinfo(skb)->frag_list;
692 				goto do_frag_list;
693 			}
694 		} else if (skb->next) {
695 			skb = skb->next;
696 			goto do_frag_list;
697 		}
698 
699 		/* Successfully sent the whole packet, account for it. */
700 		skb_dequeue(&sk->sk_write_queue);
701 		kfree_skb(head);
702 		sk->sk_wmem_queued -= sent;
703 		total_sent += sent;
704 		KCM_STATS_INCR(psock->stats.tx_msgs);
705 	} while ((head = skb_peek(&sk->sk_write_queue)));
706 out:
707 	if (!head) {
708 		/* Done with all queued messages. */
709 		WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
710 		unreserve_psock(kcm);
711 	}
712 
713 	/* Check if write space is available */
714 	sk->sk_write_space(sk);
715 
716 	return total_sent ? : ret;
717 }
718 
719 static void kcm_tx_work(struct work_struct *w)
720 {
721 	struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
722 	struct sock *sk = &kcm->sk;
723 	int err;
724 
725 	lock_sock(sk);
726 
727 	/* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
728 	 * aborts
729 	 */
730 	err = kcm_write_msgs(kcm);
731 	if (err < 0) {
732 		/* Hard failure in write, report error on KCM socket */
733 		pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
734 		report_csk_error(&kcm->sk, -err);
735 		goto out;
736 	}
737 
738 	/* Primarily for SOCK_SEQPACKET sockets */
739 	if (likely(sk->sk_socket) &&
740 	    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
741 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
742 		sk->sk_write_space(sk);
743 	}
744 
745 out:
746 	release_sock(sk);
747 }
748 
749 static void kcm_push(struct kcm_sock *kcm)
750 {
751 	if (kcm->tx_wait_more)
752 		kcm_write_msgs(kcm);
753 }
754 
755 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
756 			    int offset, size_t size, int flags)
757 
758 {
759 	struct sock *sk = sock->sk;
760 	struct kcm_sock *kcm = kcm_sk(sk);
761 	struct sk_buff *skb = NULL, *head = NULL;
762 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
763 	bool eor;
764 	int err = 0;
765 	int i;
766 
767 	if (flags & MSG_SENDPAGE_NOTLAST)
768 		flags |= MSG_MORE;
769 
770 	/* No MSG_EOR from splice, only look at MSG_MORE */
771 	eor = !(flags & MSG_MORE);
772 
773 	lock_sock(sk);
774 
775 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
776 
777 	err = -EPIPE;
778 	if (sk->sk_err)
779 		goto out_error;
780 
781 	if (kcm->seq_skb) {
782 		/* Previously opened message */
783 		head = kcm->seq_skb;
784 		skb = kcm_tx_msg(head)->last_skb;
785 		i = skb_shinfo(skb)->nr_frags;
786 
787 		if (skb_can_coalesce(skb, i, page, offset)) {
788 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
789 			skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
790 			goto coalesced;
791 		}
792 
793 		if (i >= MAX_SKB_FRAGS) {
794 			struct sk_buff *tskb;
795 
796 			tskb = alloc_skb(0, sk->sk_allocation);
797 			while (!tskb) {
798 				kcm_push(kcm);
799 				err = sk_stream_wait_memory(sk, &timeo);
800 				if (err)
801 					goto out_error;
802 			}
803 
804 			if (head == skb)
805 				skb_shinfo(head)->frag_list = tskb;
806 			else
807 				skb->next = tskb;
808 
809 			skb = tskb;
810 			skb->ip_summed = CHECKSUM_UNNECESSARY;
811 			i = 0;
812 		}
813 	} else {
814 		/* Call the sk_stream functions to manage the sndbuf mem. */
815 		if (!sk_stream_memory_free(sk)) {
816 			kcm_push(kcm);
817 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
818 			err = sk_stream_wait_memory(sk, &timeo);
819 			if (err)
820 				goto out_error;
821 		}
822 
823 		head = alloc_skb(0, sk->sk_allocation);
824 		while (!head) {
825 			kcm_push(kcm);
826 			err = sk_stream_wait_memory(sk, &timeo);
827 			if (err)
828 				goto out_error;
829 		}
830 
831 		skb = head;
832 		i = 0;
833 	}
834 
835 	get_page(page);
836 	skb_fill_page_desc(skb, i, page, offset, size);
837 	skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
838 
839 coalesced:
840 	skb->len += size;
841 	skb->data_len += size;
842 	skb->truesize += size;
843 	sk->sk_wmem_queued += size;
844 	sk_mem_charge(sk, size);
845 
846 	if (head != skb) {
847 		head->len += size;
848 		head->data_len += size;
849 		head->truesize += size;
850 	}
851 
852 	if (eor) {
853 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
854 
855 		/* Message complete, queue it on send buffer */
856 		__skb_queue_tail(&sk->sk_write_queue, head);
857 		kcm->seq_skb = NULL;
858 		KCM_STATS_INCR(kcm->stats.tx_msgs);
859 
860 		if (flags & MSG_BATCH) {
861 			kcm->tx_wait_more = true;
862 		} else if (kcm->tx_wait_more || not_busy) {
863 			err = kcm_write_msgs(kcm);
864 			if (err < 0) {
865 				/* We got a hard error in write_msgs but have
866 				 * already queued this message. Report an error
867 				 * in the socket, but don't affect return value
868 				 * from sendmsg
869 				 */
870 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
871 				report_csk_error(&kcm->sk, -err);
872 			}
873 		}
874 	} else {
875 		/* Message not complete, save state */
876 		kcm->seq_skb = head;
877 		kcm_tx_msg(head)->last_skb = skb;
878 	}
879 
880 	KCM_STATS_ADD(kcm->stats.tx_bytes, size);
881 
882 	release_sock(sk);
883 	return size;
884 
885 out_error:
886 	kcm_push(kcm);
887 
888 	err = sk_stream_error(sk, flags, err);
889 
890 	/* make sure we wake any epoll edge trigger waiter */
891 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
892 		sk->sk_write_space(sk);
893 
894 	release_sock(sk);
895 	return err;
896 }
897 
898 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
899 {
900 	struct sock *sk = sock->sk;
901 	struct kcm_sock *kcm = kcm_sk(sk);
902 	struct sk_buff *skb = NULL, *head = NULL;
903 	size_t copy, copied = 0;
904 	long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
905 	int eor = (sock->type == SOCK_DGRAM) ?
906 		  !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
907 	int err = -EPIPE;
908 
909 	lock_sock(sk);
910 
911 	/* Per tcp_sendmsg this should be in poll */
912 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
913 
914 	if (sk->sk_err)
915 		goto out_error;
916 
917 	if (kcm->seq_skb) {
918 		/* Previously opened message */
919 		head = kcm->seq_skb;
920 		skb = kcm_tx_msg(head)->last_skb;
921 		goto start;
922 	}
923 
924 	/* Call the sk_stream functions to manage the sndbuf mem. */
925 	if (!sk_stream_memory_free(sk)) {
926 		kcm_push(kcm);
927 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
928 		err = sk_stream_wait_memory(sk, &timeo);
929 		if (err)
930 			goto out_error;
931 	}
932 
933 	if (msg_data_left(msg)) {
934 		/* New message, alloc head skb */
935 		head = alloc_skb(0, sk->sk_allocation);
936 		while (!head) {
937 			kcm_push(kcm);
938 			err = sk_stream_wait_memory(sk, &timeo);
939 			if (err)
940 				goto out_error;
941 
942 			head = alloc_skb(0, sk->sk_allocation);
943 		}
944 
945 		skb = head;
946 
947 		/* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
948 		 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
949 		 */
950 		skb->ip_summed = CHECKSUM_UNNECESSARY;
951 	}
952 
953 start:
954 	while (msg_data_left(msg)) {
955 		bool merge = true;
956 		int i = skb_shinfo(skb)->nr_frags;
957 		struct page_frag *pfrag = sk_page_frag(sk);
958 
959 		if (!sk_page_frag_refill(sk, pfrag))
960 			goto wait_for_memory;
961 
962 		if (!skb_can_coalesce(skb, i, pfrag->page,
963 				      pfrag->offset)) {
964 			if (i == MAX_SKB_FRAGS) {
965 				struct sk_buff *tskb;
966 
967 				tskb = alloc_skb(0, sk->sk_allocation);
968 				if (!tskb)
969 					goto wait_for_memory;
970 
971 				if (head == skb)
972 					skb_shinfo(head)->frag_list = tskb;
973 				else
974 					skb->next = tskb;
975 
976 				skb = tskb;
977 				skb->ip_summed = CHECKSUM_UNNECESSARY;
978 				continue;
979 			}
980 			merge = false;
981 		}
982 
983 		copy = min_t(int, msg_data_left(msg),
984 			     pfrag->size - pfrag->offset);
985 
986 		if (!sk_wmem_schedule(sk, copy))
987 			goto wait_for_memory;
988 
989 		err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
990 					       pfrag->page,
991 					       pfrag->offset,
992 					       copy);
993 		if (err)
994 			goto out_error;
995 
996 		/* Update the skb. */
997 		if (merge) {
998 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
999 		} else {
1000 			skb_fill_page_desc(skb, i, pfrag->page,
1001 					   pfrag->offset, copy);
1002 			get_page(pfrag->page);
1003 		}
1004 
1005 		pfrag->offset += copy;
1006 		copied += copy;
1007 		if (head != skb) {
1008 			head->len += copy;
1009 			head->data_len += copy;
1010 		}
1011 
1012 		continue;
1013 
1014 wait_for_memory:
1015 		kcm_push(kcm);
1016 		err = sk_stream_wait_memory(sk, &timeo);
1017 		if (err)
1018 			goto out_error;
1019 	}
1020 
1021 	if (eor) {
1022 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1023 
1024 		if (head) {
1025 			/* Message complete, queue it on send buffer */
1026 			__skb_queue_tail(&sk->sk_write_queue, head);
1027 			kcm->seq_skb = NULL;
1028 			KCM_STATS_INCR(kcm->stats.tx_msgs);
1029 		}
1030 
1031 		if (msg->msg_flags & MSG_BATCH) {
1032 			kcm->tx_wait_more = true;
1033 		} else if (kcm->tx_wait_more || not_busy) {
1034 			err = kcm_write_msgs(kcm);
1035 			if (err < 0) {
1036 				/* We got a hard error in write_msgs but have
1037 				 * already queued this message. Report an error
1038 				 * in the socket, but don't affect return value
1039 				 * from sendmsg
1040 				 */
1041 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1042 				report_csk_error(&kcm->sk, -err);
1043 			}
1044 		}
1045 	} else {
1046 		/* Message not complete, save state */
1047 partial_message:
1048 		if (head) {
1049 			kcm->seq_skb = head;
1050 			kcm_tx_msg(head)->last_skb = skb;
1051 		}
1052 	}
1053 
1054 	KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1055 
1056 	release_sock(sk);
1057 	return copied;
1058 
1059 out_error:
1060 	kcm_push(kcm);
1061 
1062 	if (copied && sock->type == SOCK_SEQPACKET) {
1063 		/* Wrote some bytes before encountering an
1064 		 * error, return partial success.
1065 		 */
1066 		goto partial_message;
1067 	}
1068 
1069 	if (head != kcm->seq_skb)
1070 		kfree_skb(head);
1071 
1072 	err = sk_stream_error(sk, msg->msg_flags, err);
1073 
1074 	/* make sure we wake any epoll edge trigger waiter */
1075 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1076 		sk->sk_write_space(sk);
1077 
1078 	release_sock(sk);
1079 	return err;
1080 }
1081 
1082 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1083 				     long timeo, int *err)
1084 {
1085 	struct sk_buff *skb;
1086 
1087 	while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1088 		if (sk->sk_err) {
1089 			*err = sock_error(sk);
1090 			return NULL;
1091 		}
1092 
1093 		if (sock_flag(sk, SOCK_DONE))
1094 			return NULL;
1095 
1096 		if ((flags & MSG_DONTWAIT) || !timeo) {
1097 			*err = -EAGAIN;
1098 			return NULL;
1099 		}
1100 
1101 		sk_wait_data(sk, &timeo, NULL);
1102 
1103 		/* Handle signals */
1104 		if (signal_pending(current)) {
1105 			*err = sock_intr_errno(timeo);
1106 			return NULL;
1107 		}
1108 	}
1109 
1110 	return skb;
1111 }
1112 
1113 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1114 		       size_t len, int flags)
1115 {
1116 	struct sock *sk = sock->sk;
1117 	struct kcm_sock *kcm = kcm_sk(sk);
1118 	int err = 0;
1119 	long timeo;
1120 	struct strp_msg *stm;
1121 	int copied = 0;
1122 	struct sk_buff *skb;
1123 
1124 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1125 
1126 	lock_sock(sk);
1127 
1128 	skb = kcm_wait_data(sk, flags, timeo, &err);
1129 	if (!skb)
1130 		goto out;
1131 
1132 	/* Okay, have a message on the receive queue */
1133 
1134 	stm = strp_msg(skb);
1135 
1136 	if (len > stm->full_len)
1137 		len = stm->full_len;
1138 
1139 	err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1140 	if (err < 0)
1141 		goto out;
1142 
1143 	copied = len;
1144 	if (likely(!(flags & MSG_PEEK))) {
1145 		KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1146 		if (copied < stm->full_len) {
1147 			if (sock->type == SOCK_DGRAM) {
1148 				/* Truncated message */
1149 				msg->msg_flags |= MSG_TRUNC;
1150 				goto msg_finished;
1151 			}
1152 			stm->offset += copied;
1153 			stm->full_len -= copied;
1154 		} else {
1155 msg_finished:
1156 			/* Finished with message */
1157 			msg->msg_flags |= MSG_EOR;
1158 			KCM_STATS_INCR(kcm->stats.rx_msgs);
1159 			skb_unlink(skb, &sk->sk_receive_queue);
1160 			kfree_skb(skb);
1161 		}
1162 	}
1163 
1164 out:
1165 	release_sock(sk);
1166 
1167 	return copied ? : err;
1168 }
1169 
1170 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1171 			       struct pipe_inode_info *pipe, size_t len,
1172 			       unsigned int flags)
1173 {
1174 	struct sock *sk = sock->sk;
1175 	struct kcm_sock *kcm = kcm_sk(sk);
1176 	long timeo;
1177 	struct strp_msg *stm;
1178 	int err = 0;
1179 	ssize_t copied;
1180 	struct sk_buff *skb;
1181 
1182 	/* Only support splice for SOCKSEQPACKET */
1183 
1184 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1185 
1186 	lock_sock(sk);
1187 
1188 	skb = kcm_wait_data(sk, flags, timeo, &err);
1189 	if (!skb)
1190 		goto err_out;
1191 
1192 	/* Okay, have a message on the receive queue */
1193 
1194 	stm = strp_msg(skb);
1195 
1196 	if (len > stm->full_len)
1197 		len = stm->full_len;
1198 
1199 	copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1200 	if (copied < 0) {
1201 		err = copied;
1202 		goto err_out;
1203 	}
1204 
1205 	KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1206 
1207 	stm->offset += copied;
1208 	stm->full_len -= copied;
1209 
1210 	/* We have no way to return MSG_EOR. If all the bytes have been
1211 	 * read we still leave the message in the receive socket buffer.
1212 	 * A subsequent recvmsg needs to be done to return MSG_EOR and
1213 	 * finish reading the message.
1214 	 */
1215 
1216 	release_sock(sk);
1217 
1218 	return copied;
1219 
1220 err_out:
1221 	release_sock(sk);
1222 
1223 	return err;
1224 }
1225 
1226 /* kcm sock lock held */
1227 static void kcm_recv_disable(struct kcm_sock *kcm)
1228 {
1229 	struct kcm_mux *mux = kcm->mux;
1230 
1231 	if (kcm->rx_disabled)
1232 		return;
1233 
1234 	spin_lock_bh(&mux->rx_lock);
1235 
1236 	kcm->rx_disabled = 1;
1237 
1238 	/* If a psock is reserved we'll do cleanup in unreserve */
1239 	if (!kcm->rx_psock) {
1240 		if (kcm->rx_wait) {
1241 			list_del(&kcm->wait_rx_list);
1242 			kcm->rx_wait = false;
1243 		}
1244 
1245 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1246 	}
1247 
1248 	spin_unlock_bh(&mux->rx_lock);
1249 }
1250 
1251 /* kcm sock lock held */
1252 static void kcm_recv_enable(struct kcm_sock *kcm)
1253 {
1254 	struct kcm_mux *mux = kcm->mux;
1255 
1256 	if (!kcm->rx_disabled)
1257 		return;
1258 
1259 	spin_lock_bh(&mux->rx_lock);
1260 
1261 	kcm->rx_disabled = 0;
1262 	kcm_rcv_ready(kcm);
1263 
1264 	spin_unlock_bh(&mux->rx_lock);
1265 }
1266 
1267 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1268 			  sockptr_t optval, unsigned int optlen)
1269 {
1270 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1271 	int val, valbool;
1272 	int err = 0;
1273 
1274 	if (level != SOL_KCM)
1275 		return -ENOPROTOOPT;
1276 
1277 	if (optlen < sizeof(int))
1278 		return -EINVAL;
1279 
1280 	if (copy_from_sockptr(&val, optval, sizeof(int)))
1281 		return -EFAULT;
1282 
1283 	valbool = val ? 1 : 0;
1284 
1285 	switch (optname) {
1286 	case KCM_RECV_DISABLE:
1287 		lock_sock(&kcm->sk);
1288 		if (valbool)
1289 			kcm_recv_disable(kcm);
1290 		else
1291 			kcm_recv_enable(kcm);
1292 		release_sock(&kcm->sk);
1293 		break;
1294 	default:
1295 		err = -ENOPROTOOPT;
1296 	}
1297 
1298 	return err;
1299 }
1300 
1301 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1302 			  char __user *optval, int __user *optlen)
1303 {
1304 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1305 	int val, len;
1306 
1307 	if (level != SOL_KCM)
1308 		return -ENOPROTOOPT;
1309 
1310 	if (get_user(len, optlen))
1311 		return -EFAULT;
1312 
1313 	len = min_t(unsigned int, len, sizeof(int));
1314 	if (len < 0)
1315 		return -EINVAL;
1316 
1317 	switch (optname) {
1318 	case KCM_RECV_DISABLE:
1319 		val = kcm->rx_disabled;
1320 		break;
1321 	default:
1322 		return -ENOPROTOOPT;
1323 	}
1324 
1325 	if (put_user(len, optlen))
1326 		return -EFAULT;
1327 	if (copy_to_user(optval, &val, len))
1328 		return -EFAULT;
1329 	return 0;
1330 }
1331 
1332 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1333 {
1334 	struct kcm_sock *tkcm;
1335 	struct list_head *head;
1336 	int index = 0;
1337 
1338 	/* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1339 	 * we set sk_state, otherwise epoll_wait always returns right away with
1340 	 * EPOLLHUP
1341 	 */
1342 	kcm->sk.sk_state = TCP_ESTABLISHED;
1343 
1344 	/* Add to mux's kcm sockets list */
1345 	kcm->mux = mux;
1346 	spin_lock_bh(&mux->lock);
1347 
1348 	head = &mux->kcm_socks;
1349 	list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1350 		if (tkcm->index != index)
1351 			break;
1352 		head = &tkcm->kcm_sock_list;
1353 		index++;
1354 	}
1355 
1356 	list_add(&kcm->kcm_sock_list, head);
1357 	kcm->index = index;
1358 
1359 	mux->kcm_socks_cnt++;
1360 	spin_unlock_bh(&mux->lock);
1361 
1362 	INIT_WORK(&kcm->tx_work, kcm_tx_work);
1363 
1364 	spin_lock_bh(&mux->rx_lock);
1365 	kcm_rcv_ready(kcm);
1366 	spin_unlock_bh(&mux->rx_lock);
1367 }
1368 
1369 static int kcm_attach(struct socket *sock, struct socket *csock,
1370 		      struct bpf_prog *prog)
1371 {
1372 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1373 	struct kcm_mux *mux = kcm->mux;
1374 	struct sock *csk;
1375 	struct kcm_psock *psock = NULL, *tpsock;
1376 	struct list_head *head;
1377 	int index = 0;
1378 	static const struct strp_callbacks cb = {
1379 		.rcv_msg = kcm_rcv_strparser,
1380 		.parse_msg = kcm_parse_func_strparser,
1381 		.read_sock_done = kcm_read_sock_done,
1382 	};
1383 	int err = 0;
1384 
1385 	csk = csock->sk;
1386 	if (!csk)
1387 		return -EINVAL;
1388 
1389 	lock_sock(csk);
1390 
1391 	/* Only allow TCP sockets to be attached for now */
1392 	if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1393 	    csk->sk_protocol != IPPROTO_TCP) {
1394 		err = -EOPNOTSUPP;
1395 		goto out;
1396 	}
1397 
1398 	/* Don't allow listeners or closed sockets */
1399 	if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1400 		err = -EOPNOTSUPP;
1401 		goto out;
1402 	}
1403 
1404 	psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1405 	if (!psock) {
1406 		err = -ENOMEM;
1407 		goto out;
1408 	}
1409 
1410 	psock->mux = mux;
1411 	psock->sk = csk;
1412 	psock->bpf_prog = prog;
1413 
1414 	err = strp_init(&psock->strp, csk, &cb);
1415 	if (err) {
1416 		kmem_cache_free(kcm_psockp, psock);
1417 		goto out;
1418 	}
1419 
1420 	write_lock_bh(&csk->sk_callback_lock);
1421 
1422 	/* Check if sk_user_data is already by KCM or someone else.
1423 	 * Must be done under lock to prevent race conditions.
1424 	 */
1425 	if (csk->sk_user_data) {
1426 		write_unlock_bh(&csk->sk_callback_lock);
1427 		strp_stop(&psock->strp);
1428 		strp_done(&psock->strp);
1429 		kmem_cache_free(kcm_psockp, psock);
1430 		err = -EALREADY;
1431 		goto out;
1432 	}
1433 
1434 	psock->save_data_ready = csk->sk_data_ready;
1435 	psock->save_write_space = csk->sk_write_space;
1436 	psock->save_state_change = csk->sk_state_change;
1437 	csk->sk_user_data = psock;
1438 	csk->sk_data_ready = psock_data_ready;
1439 	csk->sk_write_space = psock_write_space;
1440 	csk->sk_state_change = psock_state_change;
1441 
1442 	write_unlock_bh(&csk->sk_callback_lock);
1443 
1444 	sock_hold(csk);
1445 
1446 	/* Finished initialization, now add the psock to the MUX. */
1447 	spin_lock_bh(&mux->lock);
1448 	head = &mux->psocks;
1449 	list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1450 		if (tpsock->index != index)
1451 			break;
1452 		head = &tpsock->psock_list;
1453 		index++;
1454 	}
1455 
1456 	list_add(&psock->psock_list, head);
1457 	psock->index = index;
1458 
1459 	KCM_STATS_INCR(mux->stats.psock_attach);
1460 	mux->psocks_cnt++;
1461 	psock_now_avail(psock);
1462 	spin_unlock_bh(&mux->lock);
1463 
1464 	/* Schedule RX work in case there are already bytes queued */
1465 	strp_check_rcv(&psock->strp);
1466 
1467 out:
1468 	release_sock(csk);
1469 
1470 	return err;
1471 }
1472 
1473 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1474 {
1475 	struct socket *csock;
1476 	struct bpf_prog *prog;
1477 	int err;
1478 
1479 	csock = sockfd_lookup(info->fd, &err);
1480 	if (!csock)
1481 		return -ENOENT;
1482 
1483 	prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1484 	if (IS_ERR(prog)) {
1485 		err = PTR_ERR(prog);
1486 		goto out;
1487 	}
1488 
1489 	err = kcm_attach(sock, csock, prog);
1490 	if (err) {
1491 		bpf_prog_put(prog);
1492 		goto out;
1493 	}
1494 
1495 	/* Keep reference on file also */
1496 
1497 	return 0;
1498 out:
1499 	sockfd_put(csock);
1500 	return err;
1501 }
1502 
1503 static void kcm_unattach(struct kcm_psock *psock)
1504 {
1505 	struct sock *csk = psock->sk;
1506 	struct kcm_mux *mux = psock->mux;
1507 
1508 	lock_sock(csk);
1509 
1510 	/* Stop getting callbacks from TCP socket. After this there should
1511 	 * be no way to reserve a kcm for this psock.
1512 	 */
1513 	write_lock_bh(&csk->sk_callback_lock);
1514 	csk->sk_user_data = NULL;
1515 	csk->sk_data_ready = psock->save_data_ready;
1516 	csk->sk_write_space = psock->save_write_space;
1517 	csk->sk_state_change = psock->save_state_change;
1518 	strp_stop(&psock->strp);
1519 
1520 	if (WARN_ON(psock->rx_kcm)) {
1521 		write_unlock_bh(&csk->sk_callback_lock);
1522 		release_sock(csk);
1523 		return;
1524 	}
1525 
1526 	spin_lock_bh(&mux->rx_lock);
1527 
1528 	/* Stop receiver activities. After this point psock should not be
1529 	 * able to get onto ready list either through callbacks or work.
1530 	 */
1531 	if (psock->ready_rx_msg) {
1532 		list_del(&psock->psock_ready_list);
1533 		kfree_skb(psock->ready_rx_msg);
1534 		psock->ready_rx_msg = NULL;
1535 		KCM_STATS_INCR(mux->stats.rx_ready_drops);
1536 	}
1537 
1538 	spin_unlock_bh(&mux->rx_lock);
1539 
1540 	write_unlock_bh(&csk->sk_callback_lock);
1541 
1542 	/* Call strp_done without sock lock */
1543 	release_sock(csk);
1544 	strp_done(&psock->strp);
1545 	lock_sock(csk);
1546 
1547 	bpf_prog_put(psock->bpf_prog);
1548 
1549 	spin_lock_bh(&mux->lock);
1550 
1551 	aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1552 	save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1553 
1554 	KCM_STATS_INCR(mux->stats.psock_unattach);
1555 
1556 	if (psock->tx_kcm) {
1557 		/* psock was reserved.  Just mark it finished and we will clean
1558 		 * up in the kcm paths, we need kcm lock which can not be
1559 		 * acquired here.
1560 		 */
1561 		KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1562 		spin_unlock_bh(&mux->lock);
1563 
1564 		/* We are unattaching a socket that is reserved. Abort the
1565 		 * socket since we may be out of sync in sending on it. We need
1566 		 * to do this without the mux lock.
1567 		 */
1568 		kcm_abort_tx_psock(psock, EPIPE, false);
1569 
1570 		spin_lock_bh(&mux->lock);
1571 		if (!psock->tx_kcm) {
1572 			/* psock now unreserved in window mux was unlocked */
1573 			goto no_reserved;
1574 		}
1575 		psock->done = 1;
1576 
1577 		/* Commit done before queuing work to process it */
1578 		smp_mb();
1579 
1580 		/* Queue tx work to make sure psock->done is handled */
1581 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1582 		spin_unlock_bh(&mux->lock);
1583 	} else {
1584 no_reserved:
1585 		if (!psock->tx_stopped)
1586 			list_del(&psock->psock_avail_list);
1587 		list_del(&psock->psock_list);
1588 		mux->psocks_cnt--;
1589 		spin_unlock_bh(&mux->lock);
1590 
1591 		sock_put(csk);
1592 		fput(csk->sk_socket->file);
1593 		kmem_cache_free(kcm_psockp, psock);
1594 	}
1595 
1596 	release_sock(csk);
1597 }
1598 
1599 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1600 {
1601 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1602 	struct kcm_mux *mux = kcm->mux;
1603 	struct kcm_psock *psock;
1604 	struct socket *csock;
1605 	struct sock *csk;
1606 	int err;
1607 
1608 	csock = sockfd_lookup(info->fd, &err);
1609 	if (!csock)
1610 		return -ENOENT;
1611 
1612 	csk = csock->sk;
1613 	if (!csk) {
1614 		err = -EINVAL;
1615 		goto out;
1616 	}
1617 
1618 	err = -ENOENT;
1619 
1620 	spin_lock_bh(&mux->lock);
1621 
1622 	list_for_each_entry(psock, &mux->psocks, psock_list) {
1623 		if (psock->sk != csk)
1624 			continue;
1625 
1626 		/* Found the matching psock */
1627 
1628 		if (psock->unattaching || WARN_ON(psock->done)) {
1629 			err = -EALREADY;
1630 			break;
1631 		}
1632 
1633 		psock->unattaching = 1;
1634 
1635 		spin_unlock_bh(&mux->lock);
1636 
1637 		/* Lower socket lock should already be held */
1638 		kcm_unattach(psock);
1639 
1640 		err = 0;
1641 		goto out;
1642 	}
1643 
1644 	spin_unlock_bh(&mux->lock);
1645 
1646 out:
1647 	sockfd_put(csock);
1648 	return err;
1649 }
1650 
1651 static struct proto kcm_proto = {
1652 	.name	= "KCM",
1653 	.owner	= THIS_MODULE,
1654 	.obj_size = sizeof(struct kcm_sock),
1655 };
1656 
1657 /* Clone a kcm socket. */
1658 static struct file *kcm_clone(struct socket *osock)
1659 {
1660 	struct socket *newsock;
1661 	struct sock *newsk;
1662 
1663 	newsock = sock_alloc();
1664 	if (!newsock)
1665 		return ERR_PTR(-ENFILE);
1666 
1667 	newsock->type = osock->type;
1668 	newsock->ops = osock->ops;
1669 
1670 	__module_get(newsock->ops->owner);
1671 
1672 	newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1673 			 &kcm_proto, false);
1674 	if (!newsk) {
1675 		sock_release(newsock);
1676 		return ERR_PTR(-ENOMEM);
1677 	}
1678 	sock_init_data(newsock, newsk);
1679 	init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1680 
1681 	return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1682 }
1683 
1684 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1685 {
1686 	int err;
1687 
1688 	switch (cmd) {
1689 	case SIOCKCMATTACH: {
1690 		struct kcm_attach info;
1691 
1692 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1693 			return -EFAULT;
1694 
1695 		err = kcm_attach_ioctl(sock, &info);
1696 
1697 		break;
1698 	}
1699 	case SIOCKCMUNATTACH: {
1700 		struct kcm_unattach info;
1701 
1702 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1703 			return -EFAULT;
1704 
1705 		err = kcm_unattach_ioctl(sock, &info);
1706 
1707 		break;
1708 	}
1709 	case SIOCKCMCLONE: {
1710 		struct kcm_clone info;
1711 		struct file *file;
1712 
1713 		info.fd = get_unused_fd_flags(0);
1714 		if (unlikely(info.fd < 0))
1715 			return info.fd;
1716 
1717 		file = kcm_clone(sock);
1718 		if (IS_ERR(file)) {
1719 			put_unused_fd(info.fd);
1720 			return PTR_ERR(file);
1721 		}
1722 		if (copy_to_user((void __user *)arg, &info,
1723 				 sizeof(info))) {
1724 			put_unused_fd(info.fd);
1725 			fput(file);
1726 			return -EFAULT;
1727 		}
1728 		fd_install(info.fd, file);
1729 		err = 0;
1730 		break;
1731 	}
1732 	default:
1733 		err = -ENOIOCTLCMD;
1734 		break;
1735 	}
1736 
1737 	return err;
1738 }
1739 
1740 static void free_mux(struct rcu_head *rcu)
1741 {
1742 	struct kcm_mux *mux = container_of(rcu,
1743 	    struct kcm_mux, rcu);
1744 
1745 	kmem_cache_free(kcm_muxp, mux);
1746 }
1747 
1748 static void release_mux(struct kcm_mux *mux)
1749 {
1750 	struct kcm_net *knet = mux->knet;
1751 	struct kcm_psock *psock, *tmp_psock;
1752 
1753 	/* Release psocks */
1754 	list_for_each_entry_safe(psock, tmp_psock,
1755 				 &mux->psocks, psock_list) {
1756 		if (!WARN_ON(psock->unattaching))
1757 			kcm_unattach(psock);
1758 	}
1759 
1760 	if (WARN_ON(mux->psocks_cnt))
1761 		return;
1762 
1763 	__skb_queue_purge(&mux->rx_hold_queue);
1764 
1765 	mutex_lock(&knet->mutex);
1766 	aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1767 	aggregate_psock_stats(&mux->aggregate_psock_stats,
1768 			      &knet->aggregate_psock_stats);
1769 	aggregate_strp_stats(&mux->aggregate_strp_stats,
1770 			     &knet->aggregate_strp_stats);
1771 	list_del_rcu(&mux->kcm_mux_list);
1772 	knet->count--;
1773 	mutex_unlock(&knet->mutex);
1774 
1775 	call_rcu(&mux->rcu, free_mux);
1776 }
1777 
1778 static void kcm_done(struct kcm_sock *kcm)
1779 {
1780 	struct kcm_mux *mux = kcm->mux;
1781 	struct sock *sk = &kcm->sk;
1782 	int socks_cnt;
1783 
1784 	spin_lock_bh(&mux->rx_lock);
1785 	if (kcm->rx_psock) {
1786 		/* Cleanup in unreserve_rx_kcm */
1787 		WARN_ON(kcm->done);
1788 		kcm->rx_disabled = 1;
1789 		kcm->done = 1;
1790 		spin_unlock_bh(&mux->rx_lock);
1791 		return;
1792 	}
1793 
1794 	if (kcm->rx_wait) {
1795 		list_del(&kcm->wait_rx_list);
1796 		kcm->rx_wait = false;
1797 	}
1798 	/* Move any pending receive messages to other kcm sockets */
1799 	requeue_rx_msgs(mux, &sk->sk_receive_queue);
1800 
1801 	spin_unlock_bh(&mux->rx_lock);
1802 
1803 	if (WARN_ON(sk_rmem_alloc_get(sk)))
1804 		return;
1805 
1806 	/* Detach from MUX */
1807 	spin_lock_bh(&mux->lock);
1808 
1809 	list_del(&kcm->kcm_sock_list);
1810 	mux->kcm_socks_cnt--;
1811 	socks_cnt = mux->kcm_socks_cnt;
1812 
1813 	spin_unlock_bh(&mux->lock);
1814 
1815 	if (!socks_cnt) {
1816 		/* We are done with the mux now. */
1817 		release_mux(mux);
1818 	}
1819 
1820 	WARN_ON(kcm->rx_wait);
1821 
1822 	sock_put(&kcm->sk);
1823 }
1824 
1825 /* Called by kcm_release to close a KCM socket.
1826  * If this is the last KCM socket on the MUX, destroy the MUX.
1827  */
1828 static int kcm_release(struct socket *sock)
1829 {
1830 	struct sock *sk = sock->sk;
1831 	struct kcm_sock *kcm;
1832 	struct kcm_mux *mux;
1833 	struct kcm_psock *psock;
1834 
1835 	if (!sk)
1836 		return 0;
1837 
1838 	kcm = kcm_sk(sk);
1839 	mux = kcm->mux;
1840 
1841 	sock_orphan(sk);
1842 	kfree_skb(kcm->seq_skb);
1843 
1844 	lock_sock(sk);
1845 	/* Purge queue under lock to avoid race condition with tx_work trying
1846 	 * to act when queue is nonempty. If tx_work runs after this point
1847 	 * it will just return.
1848 	 */
1849 	__skb_queue_purge(&sk->sk_write_queue);
1850 
1851 	/* Set tx_stopped. This is checked when psock is bound to a kcm and we
1852 	 * get a writespace callback. This prevents further work being queued
1853 	 * from the callback (unbinding the psock occurs after canceling work.
1854 	 */
1855 	kcm->tx_stopped = 1;
1856 
1857 	release_sock(sk);
1858 
1859 	spin_lock_bh(&mux->lock);
1860 	if (kcm->tx_wait) {
1861 		/* Take of tx_wait list, after this point there should be no way
1862 		 * that a psock will be assigned to this kcm.
1863 		 */
1864 		list_del(&kcm->wait_psock_list);
1865 		kcm->tx_wait = false;
1866 	}
1867 	spin_unlock_bh(&mux->lock);
1868 
1869 	/* Cancel work. After this point there should be no outside references
1870 	 * to the kcm socket.
1871 	 */
1872 	cancel_work_sync(&kcm->tx_work);
1873 
1874 	lock_sock(sk);
1875 	psock = kcm->tx_psock;
1876 	if (psock) {
1877 		/* A psock was reserved, so we need to kill it since it
1878 		 * may already have some bytes queued from a message. We
1879 		 * need to do this after removing kcm from tx_wait list.
1880 		 */
1881 		kcm_abort_tx_psock(psock, EPIPE, false);
1882 		unreserve_psock(kcm);
1883 	}
1884 	release_sock(sk);
1885 
1886 	WARN_ON(kcm->tx_wait);
1887 	WARN_ON(kcm->tx_psock);
1888 
1889 	sock->sk = NULL;
1890 
1891 	kcm_done(kcm);
1892 
1893 	return 0;
1894 }
1895 
1896 static const struct proto_ops kcm_dgram_ops = {
1897 	.family =	PF_KCM,
1898 	.owner =	THIS_MODULE,
1899 	.release =	kcm_release,
1900 	.bind =		sock_no_bind,
1901 	.connect =	sock_no_connect,
1902 	.socketpair =	sock_no_socketpair,
1903 	.accept =	sock_no_accept,
1904 	.getname =	sock_no_getname,
1905 	.poll =		datagram_poll,
1906 	.ioctl =	kcm_ioctl,
1907 	.listen =	sock_no_listen,
1908 	.shutdown =	sock_no_shutdown,
1909 	.setsockopt =	kcm_setsockopt,
1910 	.getsockopt =	kcm_getsockopt,
1911 	.sendmsg =	kcm_sendmsg,
1912 	.recvmsg =	kcm_recvmsg,
1913 	.mmap =		sock_no_mmap,
1914 	.sendpage =	kcm_sendpage,
1915 };
1916 
1917 static const struct proto_ops kcm_seqpacket_ops = {
1918 	.family =	PF_KCM,
1919 	.owner =	THIS_MODULE,
1920 	.release =	kcm_release,
1921 	.bind =		sock_no_bind,
1922 	.connect =	sock_no_connect,
1923 	.socketpair =	sock_no_socketpair,
1924 	.accept =	sock_no_accept,
1925 	.getname =	sock_no_getname,
1926 	.poll =		datagram_poll,
1927 	.ioctl =	kcm_ioctl,
1928 	.listen =	sock_no_listen,
1929 	.shutdown =	sock_no_shutdown,
1930 	.setsockopt =	kcm_setsockopt,
1931 	.getsockopt =	kcm_getsockopt,
1932 	.sendmsg =	kcm_sendmsg,
1933 	.recvmsg =	kcm_recvmsg,
1934 	.mmap =		sock_no_mmap,
1935 	.sendpage =	kcm_sendpage,
1936 	.splice_read =	kcm_splice_read,
1937 };
1938 
1939 /* Create proto operation for kcm sockets */
1940 static int kcm_create(struct net *net, struct socket *sock,
1941 		      int protocol, int kern)
1942 {
1943 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1944 	struct sock *sk;
1945 	struct kcm_mux *mux;
1946 
1947 	switch (sock->type) {
1948 	case SOCK_DGRAM:
1949 		sock->ops = &kcm_dgram_ops;
1950 		break;
1951 	case SOCK_SEQPACKET:
1952 		sock->ops = &kcm_seqpacket_ops;
1953 		break;
1954 	default:
1955 		return -ESOCKTNOSUPPORT;
1956 	}
1957 
1958 	if (protocol != KCMPROTO_CONNECTED)
1959 		return -EPROTONOSUPPORT;
1960 
1961 	sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1962 	if (!sk)
1963 		return -ENOMEM;
1964 
1965 	/* Allocate a kcm mux, shared between KCM sockets */
1966 	mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1967 	if (!mux) {
1968 		sk_free(sk);
1969 		return -ENOMEM;
1970 	}
1971 
1972 	spin_lock_init(&mux->lock);
1973 	spin_lock_init(&mux->rx_lock);
1974 	INIT_LIST_HEAD(&mux->kcm_socks);
1975 	INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1976 	INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1977 
1978 	INIT_LIST_HEAD(&mux->psocks);
1979 	INIT_LIST_HEAD(&mux->psocks_ready);
1980 	INIT_LIST_HEAD(&mux->psocks_avail);
1981 
1982 	mux->knet = knet;
1983 
1984 	/* Add new MUX to list */
1985 	mutex_lock(&knet->mutex);
1986 	list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1987 	knet->count++;
1988 	mutex_unlock(&knet->mutex);
1989 
1990 	skb_queue_head_init(&mux->rx_hold_queue);
1991 
1992 	/* Init KCM socket */
1993 	sock_init_data(sock, sk);
1994 	init_kcm_sock(kcm_sk(sk), mux);
1995 
1996 	return 0;
1997 }
1998 
1999 static const struct net_proto_family kcm_family_ops = {
2000 	.family = PF_KCM,
2001 	.create = kcm_create,
2002 	.owner  = THIS_MODULE,
2003 };
2004 
2005 static __net_init int kcm_init_net(struct net *net)
2006 {
2007 	struct kcm_net *knet = net_generic(net, kcm_net_id);
2008 
2009 	INIT_LIST_HEAD_RCU(&knet->mux_list);
2010 	mutex_init(&knet->mutex);
2011 
2012 	return 0;
2013 }
2014 
2015 static __net_exit void kcm_exit_net(struct net *net)
2016 {
2017 	struct kcm_net *knet = net_generic(net, kcm_net_id);
2018 
2019 	/* All KCM sockets should be closed at this point, which should mean
2020 	 * that all multiplexors and psocks have been destroyed.
2021 	 */
2022 	WARN_ON(!list_empty(&knet->mux_list));
2023 }
2024 
2025 static struct pernet_operations kcm_net_ops = {
2026 	.init = kcm_init_net,
2027 	.exit = kcm_exit_net,
2028 	.id   = &kcm_net_id,
2029 	.size = sizeof(struct kcm_net),
2030 };
2031 
2032 static int __init kcm_init(void)
2033 {
2034 	int err = -ENOMEM;
2035 
2036 	kcm_muxp = kmem_cache_create("kcm_mux_cache",
2037 				     sizeof(struct kcm_mux), 0,
2038 				     SLAB_HWCACHE_ALIGN, NULL);
2039 	if (!kcm_muxp)
2040 		goto fail;
2041 
2042 	kcm_psockp = kmem_cache_create("kcm_psock_cache",
2043 				       sizeof(struct kcm_psock), 0,
2044 					SLAB_HWCACHE_ALIGN, NULL);
2045 	if (!kcm_psockp)
2046 		goto fail;
2047 
2048 	kcm_wq = create_singlethread_workqueue("kkcmd");
2049 	if (!kcm_wq)
2050 		goto fail;
2051 
2052 	err = proto_register(&kcm_proto, 1);
2053 	if (err)
2054 		goto fail;
2055 
2056 	err = register_pernet_device(&kcm_net_ops);
2057 	if (err)
2058 		goto net_ops_fail;
2059 
2060 	err = sock_register(&kcm_family_ops);
2061 	if (err)
2062 		goto sock_register_fail;
2063 
2064 	err = kcm_proc_init();
2065 	if (err)
2066 		goto proc_init_fail;
2067 
2068 	return 0;
2069 
2070 proc_init_fail:
2071 	sock_unregister(PF_KCM);
2072 
2073 sock_register_fail:
2074 	unregister_pernet_device(&kcm_net_ops);
2075 
2076 net_ops_fail:
2077 	proto_unregister(&kcm_proto);
2078 
2079 fail:
2080 	kmem_cache_destroy(kcm_muxp);
2081 	kmem_cache_destroy(kcm_psockp);
2082 
2083 	if (kcm_wq)
2084 		destroy_workqueue(kcm_wq);
2085 
2086 	return err;
2087 }
2088 
2089 static void __exit kcm_exit(void)
2090 {
2091 	kcm_proc_exit();
2092 	sock_unregister(PF_KCM);
2093 	unregister_pernet_device(&kcm_net_ops);
2094 	proto_unregister(&kcm_proto);
2095 	destroy_workqueue(kcm_wq);
2096 
2097 	kmem_cache_destroy(kcm_muxp);
2098 	kmem_cache_destroy(kcm_psockp);
2099 }
2100 
2101 module_init(kcm_init);
2102 module_exit(kcm_exit);
2103 
2104 MODULE_LICENSE("GPL");
2105 MODULE_ALIAS_NETPROTO(PF_KCM);
2106