xref: /openbmc/linux/net/kcm/kcmsock.c (revision 8bd1369b)
1 /*
2  * Kernel Connection Multiplexor
3  *
4  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2
8  * as published by the Free Software Foundation.
9  */
10 
11 #include <linux/bpf.h>
12 #include <linux/errno.h>
13 #include <linux/errqueue.h>
14 #include <linux/file.h>
15 #include <linux/in.h>
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/net.h>
19 #include <linux/netdevice.h>
20 #include <linux/poll.h>
21 #include <linux/rculist.h>
22 #include <linux/skbuff.h>
23 #include <linux/socket.h>
24 #include <linux/uaccess.h>
25 #include <linux/workqueue.h>
26 #include <linux/syscalls.h>
27 #include <linux/sched/signal.h>
28 
29 #include <net/kcm.h>
30 #include <net/netns/generic.h>
31 #include <net/sock.h>
32 #include <uapi/linux/kcm.h>
33 
34 unsigned int kcm_net_id;
35 
36 static struct kmem_cache *kcm_psockp __read_mostly;
37 static struct kmem_cache *kcm_muxp __read_mostly;
38 static struct workqueue_struct *kcm_wq;
39 
40 static inline struct kcm_sock *kcm_sk(const struct sock *sk)
41 {
42 	return (struct kcm_sock *)sk;
43 }
44 
45 static inline struct kcm_tx_msg *kcm_tx_msg(struct sk_buff *skb)
46 {
47 	return (struct kcm_tx_msg *)skb->cb;
48 }
49 
50 static void report_csk_error(struct sock *csk, int err)
51 {
52 	csk->sk_err = EPIPE;
53 	csk->sk_error_report(csk);
54 }
55 
56 static void kcm_abort_tx_psock(struct kcm_psock *psock, int err,
57 			       bool wakeup_kcm)
58 {
59 	struct sock *csk = psock->sk;
60 	struct kcm_mux *mux = psock->mux;
61 
62 	/* Unrecoverable error in transmit */
63 
64 	spin_lock_bh(&mux->lock);
65 
66 	if (psock->tx_stopped) {
67 		spin_unlock_bh(&mux->lock);
68 		return;
69 	}
70 
71 	psock->tx_stopped = 1;
72 	KCM_STATS_INCR(psock->stats.tx_aborts);
73 
74 	if (!psock->tx_kcm) {
75 		/* Take off psocks_avail list */
76 		list_del(&psock->psock_avail_list);
77 	} else if (wakeup_kcm) {
78 		/* In this case psock is being aborted while outside of
79 		 * write_msgs and psock is reserved. Schedule tx_work
80 		 * to handle the failure there. Need to commit tx_stopped
81 		 * before queuing work.
82 		 */
83 		smp_mb();
84 
85 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
86 	}
87 
88 	spin_unlock_bh(&mux->lock);
89 
90 	/* Report error on lower socket */
91 	report_csk_error(csk, err);
92 }
93 
94 /* RX mux lock held. */
95 static void kcm_update_rx_mux_stats(struct kcm_mux *mux,
96 				    struct kcm_psock *psock)
97 {
98 	STRP_STATS_ADD(mux->stats.rx_bytes,
99 		       psock->strp.stats.bytes -
100 		       psock->saved_rx_bytes);
101 	mux->stats.rx_msgs +=
102 		psock->strp.stats.msgs - psock->saved_rx_msgs;
103 	psock->saved_rx_msgs = psock->strp.stats.msgs;
104 	psock->saved_rx_bytes = psock->strp.stats.bytes;
105 }
106 
107 static void kcm_update_tx_mux_stats(struct kcm_mux *mux,
108 				    struct kcm_psock *psock)
109 {
110 	KCM_STATS_ADD(mux->stats.tx_bytes,
111 		      psock->stats.tx_bytes - psock->saved_tx_bytes);
112 	mux->stats.tx_msgs +=
113 		psock->stats.tx_msgs - psock->saved_tx_msgs;
114 	psock->saved_tx_msgs = psock->stats.tx_msgs;
115 	psock->saved_tx_bytes = psock->stats.tx_bytes;
116 }
117 
118 static int kcm_queue_rcv_skb(struct sock *sk, struct sk_buff *skb);
119 
120 /* KCM is ready to receive messages on its queue-- either the KCM is new or
121  * has become unblocked after being blocked on full socket buffer. Queue any
122  * pending ready messages on a psock. RX mux lock held.
123  */
124 static void kcm_rcv_ready(struct kcm_sock *kcm)
125 {
126 	struct kcm_mux *mux = kcm->mux;
127 	struct kcm_psock *psock;
128 	struct sk_buff *skb;
129 
130 	if (unlikely(kcm->rx_wait || kcm->rx_psock || kcm->rx_disabled))
131 		return;
132 
133 	while (unlikely((skb = __skb_dequeue(&mux->rx_hold_queue)))) {
134 		if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
135 			/* Assuming buffer limit has been reached */
136 			skb_queue_head(&mux->rx_hold_queue, skb);
137 			WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
138 			return;
139 		}
140 	}
141 
142 	while (!list_empty(&mux->psocks_ready)) {
143 		psock = list_first_entry(&mux->psocks_ready, struct kcm_psock,
144 					 psock_ready_list);
145 
146 		if (kcm_queue_rcv_skb(&kcm->sk, psock->ready_rx_msg)) {
147 			/* Assuming buffer limit has been reached */
148 			WARN_ON(!sk_rmem_alloc_get(&kcm->sk));
149 			return;
150 		}
151 
152 		/* Consumed the ready message on the psock. Schedule rx_work to
153 		 * get more messages.
154 		 */
155 		list_del(&psock->psock_ready_list);
156 		psock->ready_rx_msg = NULL;
157 		/* Commit clearing of ready_rx_msg for queuing work */
158 		smp_mb();
159 
160 		strp_unpause(&psock->strp);
161 		strp_check_rcv(&psock->strp);
162 	}
163 
164 	/* Buffer limit is okay now, add to ready list */
165 	list_add_tail(&kcm->wait_rx_list,
166 		      &kcm->mux->kcm_rx_waiters);
167 	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 (!kcm->rx_wait && !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 			kcm->rx_wait = false;
243 
244 			/* Commit rx_wait to read in kcm_free */
245 			smp_wmb();
246 
247 			goto try_again;
248 		}
249 	}
250 }
251 
252 /* Lower sock lock held */
253 static struct kcm_sock *reserve_rx_kcm(struct kcm_psock *psock,
254 				       struct sk_buff *head)
255 {
256 	struct kcm_mux *mux = psock->mux;
257 	struct kcm_sock *kcm;
258 
259 	WARN_ON(psock->ready_rx_msg);
260 
261 	if (psock->rx_kcm)
262 		return psock->rx_kcm;
263 
264 	spin_lock_bh(&mux->rx_lock);
265 
266 	if (psock->rx_kcm) {
267 		spin_unlock_bh(&mux->rx_lock);
268 		return psock->rx_kcm;
269 	}
270 
271 	kcm_update_rx_mux_stats(mux, psock);
272 
273 	if (list_empty(&mux->kcm_rx_waiters)) {
274 		psock->ready_rx_msg = head;
275 		strp_pause(&psock->strp);
276 		list_add_tail(&psock->psock_ready_list,
277 			      &mux->psocks_ready);
278 		spin_unlock_bh(&mux->rx_lock);
279 		return NULL;
280 	}
281 
282 	kcm = list_first_entry(&mux->kcm_rx_waiters,
283 			       struct kcm_sock, wait_rx_list);
284 	list_del(&kcm->wait_rx_list);
285 	kcm->rx_wait = false;
286 
287 	psock->rx_kcm = kcm;
288 	kcm->rx_psock = psock;
289 
290 	spin_unlock_bh(&mux->rx_lock);
291 
292 	return kcm;
293 }
294 
295 static void kcm_done(struct kcm_sock *kcm);
296 
297 static void kcm_done_work(struct work_struct *w)
298 {
299 	kcm_done(container_of(w, struct kcm_sock, done_work));
300 }
301 
302 /* Lower sock held */
303 static void unreserve_rx_kcm(struct kcm_psock *psock,
304 			     bool rcv_ready)
305 {
306 	struct kcm_sock *kcm = psock->rx_kcm;
307 	struct kcm_mux *mux = psock->mux;
308 
309 	if (!kcm)
310 		return;
311 
312 	spin_lock_bh(&mux->rx_lock);
313 
314 	psock->rx_kcm = NULL;
315 	kcm->rx_psock = NULL;
316 
317 	/* Commit kcm->rx_psock before sk_rmem_alloc_get to sync with
318 	 * kcm_rfree
319 	 */
320 	smp_mb();
321 
322 	if (unlikely(kcm->done)) {
323 		spin_unlock_bh(&mux->rx_lock);
324 
325 		/* Need to run kcm_done in a task since we need to qcquire
326 		 * callback locks which may already be held here.
327 		 */
328 		INIT_WORK(&kcm->done_work, kcm_done_work);
329 		schedule_work(&kcm->done_work);
330 		return;
331 	}
332 
333 	if (unlikely(kcm->rx_disabled)) {
334 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
335 	} else if (rcv_ready || unlikely(!sk_rmem_alloc_get(&kcm->sk))) {
336 		/* Check for degenerative race with rx_wait that all
337 		 * data was dequeued (accounted for in kcm_rfree).
338 		 */
339 		kcm_rcv_ready(kcm);
340 	}
341 	spin_unlock_bh(&mux->rx_lock);
342 }
343 
344 /* Lower sock lock held */
345 static void psock_data_ready(struct sock *sk)
346 {
347 	struct kcm_psock *psock;
348 
349 	read_lock_bh(&sk->sk_callback_lock);
350 
351 	psock = (struct kcm_psock *)sk->sk_user_data;
352 	if (likely(psock))
353 		strp_data_ready(&psock->strp);
354 
355 	read_unlock_bh(&sk->sk_callback_lock);
356 }
357 
358 /* Called with lower sock held */
359 static void kcm_rcv_strparser(struct strparser *strp, struct sk_buff *skb)
360 {
361 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
362 	struct kcm_sock *kcm;
363 
364 try_queue:
365 	kcm = reserve_rx_kcm(psock, skb);
366 	if (!kcm) {
367 		 /* Unable to reserve a KCM, message is held in psock and strp
368 		  * is paused.
369 		  */
370 		return;
371 	}
372 
373 	if (kcm_queue_rcv_skb(&kcm->sk, skb)) {
374 		/* Should mean socket buffer full */
375 		unreserve_rx_kcm(psock, false);
376 		goto try_queue;
377 	}
378 }
379 
380 static int kcm_parse_func_strparser(struct strparser *strp, struct sk_buff *skb)
381 {
382 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
383 	struct bpf_prog *prog = psock->bpf_prog;
384 
385 	return (*prog->bpf_func)(skb, prog->insnsi);
386 }
387 
388 static int kcm_read_sock_done(struct strparser *strp, int err)
389 {
390 	struct kcm_psock *psock = container_of(strp, struct kcm_psock, strp);
391 
392 	unreserve_rx_kcm(psock, true);
393 
394 	return err;
395 }
396 
397 static void psock_state_change(struct sock *sk)
398 {
399 	/* TCP only does a EPOLLIN for a half close. Do a EPOLLHUP here
400 	 * since application will normally not poll with EPOLLIN
401 	 * on the TCP sockets.
402 	 */
403 
404 	report_csk_error(sk, EPIPE);
405 }
406 
407 static void psock_write_space(struct sock *sk)
408 {
409 	struct kcm_psock *psock;
410 	struct kcm_mux *mux;
411 	struct kcm_sock *kcm;
412 
413 	read_lock_bh(&sk->sk_callback_lock);
414 
415 	psock = (struct kcm_psock *)sk->sk_user_data;
416 	if (unlikely(!psock))
417 		goto out;
418 	mux = psock->mux;
419 
420 	spin_lock_bh(&mux->lock);
421 
422 	/* Check if the socket is reserved so someone is waiting for sending. */
423 	kcm = psock->tx_kcm;
424 	if (kcm && !unlikely(kcm->tx_stopped))
425 		queue_work(kcm_wq, &kcm->tx_work);
426 
427 	spin_unlock_bh(&mux->lock);
428 out:
429 	read_unlock_bh(&sk->sk_callback_lock);
430 }
431 
432 static void unreserve_psock(struct kcm_sock *kcm);
433 
434 /* kcm sock is locked. */
435 static struct kcm_psock *reserve_psock(struct kcm_sock *kcm)
436 {
437 	struct kcm_mux *mux = kcm->mux;
438 	struct kcm_psock *psock;
439 
440 	psock = kcm->tx_psock;
441 
442 	smp_rmb(); /* Must read tx_psock before tx_wait */
443 
444 	if (psock) {
445 		WARN_ON(kcm->tx_wait);
446 		if (unlikely(psock->tx_stopped))
447 			unreserve_psock(kcm);
448 		else
449 			return kcm->tx_psock;
450 	}
451 
452 	spin_lock_bh(&mux->lock);
453 
454 	/* Check again under lock to see if psock was reserved for this
455 	 * psock via psock_unreserve.
456 	 */
457 	psock = kcm->tx_psock;
458 	if (unlikely(psock)) {
459 		WARN_ON(kcm->tx_wait);
460 		spin_unlock_bh(&mux->lock);
461 		return kcm->tx_psock;
462 	}
463 
464 	if (!list_empty(&mux->psocks_avail)) {
465 		psock = list_first_entry(&mux->psocks_avail,
466 					 struct kcm_psock,
467 					 psock_avail_list);
468 		list_del(&psock->psock_avail_list);
469 		if (kcm->tx_wait) {
470 			list_del(&kcm->wait_psock_list);
471 			kcm->tx_wait = false;
472 		}
473 		kcm->tx_psock = psock;
474 		psock->tx_kcm = kcm;
475 		KCM_STATS_INCR(psock->stats.reserved);
476 	} else if (!kcm->tx_wait) {
477 		list_add_tail(&kcm->wait_psock_list,
478 			      &mux->kcm_tx_waiters);
479 		kcm->tx_wait = true;
480 	}
481 
482 	spin_unlock_bh(&mux->lock);
483 
484 	return psock;
485 }
486 
487 /* mux lock held */
488 static void psock_now_avail(struct kcm_psock *psock)
489 {
490 	struct kcm_mux *mux = psock->mux;
491 	struct kcm_sock *kcm;
492 
493 	if (list_empty(&mux->kcm_tx_waiters)) {
494 		list_add_tail(&psock->psock_avail_list,
495 			      &mux->psocks_avail);
496 	} else {
497 		kcm = list_first_entry(&mux->kcm_tx_waiters,
498 				       struct kcm_sock,
499 				       wait_psock_list);
500 		list_del(&kcm->wait_psock_list);
501 		kcm->tx_wait = false;
502 		psock->tx_kcm = kcm;
503 
504 		/* Commit before changing tx_psock since that is read in
505 		 * reserve_psock before queuing work.
506 		 */
507 		smp_mb();
508 
509 		kcm->tx_psock = psock;
510 		KCM_STATS_INCR(psock->stats.reserved);
511 		queue_work(kcm_wq, &kcm->tx_work);
512 	}
513 }
514 
515 /* kcm sock is locked. */
516 static void unreserve_psock(struct kcm_sock *kcm)
517 {
518 	struct kcm_psock *psock;
519 	struct kcm_mux *mux = kcm->mux;
520 
521 	spin_lock_bh(&mux->lock);
522 
523 	psock = kcm->tx_psock;
524 
525 	if (WARN_ON(!psock)) {
526 		spin_unlock_bh(&mux->lock);
527 		return;
528 	}
529 
530 	smp_rmb(); /* Read tx_psock before tx_wait */
531 
532 	kcm_update_tx_mux_stats(mux, psock);
533 
534 	WARN_ON(kcm->tx_wait);
535 
536 	kcm->tx_psock = NULL;
537 	psock->tx_kcm = NULL;
538 	KCM_STATS_INCR(psock->stats.unreserved);
539 
540 	if (unlikely(psock->tx_stopped)) {
541 		if (psock->done) {
542 			/* Deferred free */
543 			list_del(&psock->psock_list);
544 			mux->psocks_cnt--;
545 			sock_put(psock->sk);
546 			fput(psock->sk->sk_socket->file);
547 			kmem_cache_free(kcm_psockp, psock);
548 		}
549 
550 		/* Don't put back on available list */
551 
552 		spin_unlock_bh(&mux->lock);
553 
554 		return;
555 	}
556 
557 	psock_now_avail(psock);
558 
559 	spin_unlock_bh(&mux->lock);
560 }
561 
562 static void kcm_report_tx_retry(struct kcm_sock *kcm)
563 {
564 	struct kcm_mux *mux = kcm->mux;
565 
566 	spin_lock_bh(&mux->lock);
567 	KCM_STATS_INCR(mux->stats.tx_retries);
568 	spin_unlock_bh(&mux->lock);
569 }
570 
571 /* Write any messages ready on the kcm socket.  Called with kcm sock lock
572  * held.  Return bytes actually sent or error.
573  */
574 static int kcm_write_msgs(struct kcm_sock *kcm)
575 {
576 	struct sock *sk = &kcm->sk;
577 	struct kcm_psock *psock;
578 	struct sk_buff *skb, *head;
579 	struct kcm_tx_msg *txm;
580 	unsigned short fragidx, frag_offset;
581 	unsigned int sent, total_sent = 0;
582 	int ret = 0;
583 
584 	kcm->tx_wait_more = false;
585 	psock = kcm->tx_psock;
586 	if (unlikely(psock && psock->tx_stopped)) {
587 		/* A reserved psock was aborted asynchronously. Unreserve
588 		 * it and we'll retry the message.
589 		 */
590 		unreserve_psock(kcm);
591 		kcm_report_tx_retry(kcm);
592 		if (skb_queue_empty(&sk->sk_write_queue))
593 			return 0;
594 
595 		kcm_tx_msg(skb_peek(&sk->sk_write_queue))->sent = 0;
596 
597 	} else if (skb_queue_empty(&sk->sk_write_queue)) {
598 		return 0;
599 	}
600 
601 	head = skb_peek(&sk->sk_write_queue);
602 	txm = kcm_tx_msg(head);
603 
604 	if (txm->sent) {
605 		/* Send of first skbuff in queue already in progress */
606 		if (WARN_ON(!psock)) {
607 			ret = -EINVAL;
608 			goto out;
609 		}
610 		sent = txm->sent;
611 		frag_offset = txm->frag_offset;
612 		fragidx = txm->fragidx;
613 		skb = txm->frag_skb;
614 
615 		goto do_frag;
616 	}
617 
618 try_again:
619 	psock = reserve_psock(kcm);
620 	if (!psock)
621 		goto out;
622 
623 	do {
624 		skb = head;
625 		txm = kcm_tx_msg(head);
626 		sent = 0;
627 
628 do_frag_list:
629 		if (WARN_ON(!skb_shinfo(skb)->nr_frags)) {
630 			ret = -EINVAL;
631 			goto out;
632 		}
633 
634 		for (fragidx = 0; fragidx < skb_shinfo(skb)->nr_frags;
635 		     fragidx++) {
636 			skb_frag_t *frag;
637 
638 			frag_offset = 0;
639 do_frag:
640 			frag = &skb_shinfo(skb)->frags[fragidx];
641 			if (WARN_ON(!frag->size)) {
642 				ret = -EINVAL;
643 				goto out;
644 			}
645 
646 			ret = kernel_sendpage(psock->sk->sk_socket,
647 					      frag->page.p,
648 					      frag->page_offset + frag_offset,
649 					      frag->size - frag_offset,
650 					      MSG_DONTWAIT);
651 			if (ret <= 0) {
652 				if (ret == -EAGAIN) {
653 					/* Save state to try again when there's
654 					 * write space on the socket
655 					 */
656 					txm->sent = sent;
657 					txm->frag_offset = frag_offset;
658 					txm->fragidx = fragidx;
659 					txm->frag_skb = skb;
660 
661 					ret = 0;
662 					goto out;
663 				}
664 
665 				/* Hard failure in sending message, abort this
666 				 * psock since it has lost framing
667 				 * synchonization and retry sending the
668 				 * message from the beginning.
669 				 */
670 				kcm_abort_tx_psock(psock, ret ? -ret : EPIPE,
671 						   true);
672 				unreserve_psock(kcm);
673 
674 				txm->sent = 0;
675 				kcm_report_tx_retry(kcm);
676 				ret = 0;
677 
678 				goto try_again;
679 			}
680 
681 			sent += ret;
682 			frag_offset += ret;
683 			KCM_STATS_ADD(psock->stats.tx_bytes, ret);
684 			if (frag_offset < frag->size) {
685 				/* Not finished with this frag */
686 				goto do_frag;
687 			}
688 		}
689 
690 		if (skb == head) {
691 			if (skb_has_frag_list(skb)) {
692 				skb = skb_shinfo(skb)->frag_list;
693 				goto do_frag_list;
694 			}
695 		} else if (skb->next) {
696 			skb = skb->next;
697 			goto do_frag_list;
698 		}
699 
700 		/* Successfully sent the whole packet, account for it. */
701 		skb_dequeue(&sk->sk_write_queue);
702 		kfree_skb(head);
703 		sk->sk_wmem_queued -= sent;
704 		total_sent += sent;
705 		KCM_STATS_INCR(psock->stats.tx_msgs);
706 	} while ((head = skb_peek(&sk->sk_write_queue)));
707 out:
708 	if (!head) {
709 		/* Done with all queued messages. */
710 		WARN_ON(!skb_queue_empty(&sk->sk_write_queue));
711 		unreserve_psock(kcm);
712 	}
713 
714 	/* Check if write space is available */
715 	sk->sk_write_space(sk);
716 
717 	return total_sent ? : ret;
718 }
719 
720 static void kcm_tx_work(struct work_struct *w)
721 {
722 	struct kcm_sock *kcm = container_of(w, struct kcm_sock, tx_work);
723 	struct sock *sk = &kcm->sk;
724 	int err;
725 
726 	lock_sock(sk);
727 
728 	/* Primarily for SOCK_DGRAM sockets, also handle asynchronous tx
729 	 * aborts
730 	 */
731 	err = kcm_write_msgs(kcm);
732 	if (err < 0) {
733 		/* Hard failure in write, report error on KCM socket */
734 		pr_warn("KCM: Hard failure on kcm_write_msgs %d\n", err);
735 		report_csk_error(&kcm->sk, -err);
736 		goto out;
737 	}
738 
739 	/* Primarily for SOCK_SEQPACKET sockets */
740 	if (likely(sk->sk_socket) &&
741 	    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
742 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
743 		sk->sk_write_space(sk);
744 	}
745 
746 out:
747 	release_sock(sk);
748 }
749 
750 static void kcm_push(struct kcm_sock *kcm)
751 {
752 	if (kcm->tx_wait_more)
753 		kcm_write_msgs(kcm);
754 }
755 
756 static ssize_t kcm_sendpage(struct socket *sock, struct page *page,
757 			    int offset, size_t size, int flags)
758 
759 {
760 	struct sock *sk = sock->sk;
761 	struct kcm_sock *kcm = kcm_sk(sk);
762 	struct sk_buff *skb = NULL, *head = NULL;
763 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
764 	bool eor;
765 	int err = 0;
766 	int i;
767 
768 	if (flags & MSG_SENDPAGE_NOTLAST)
769 		flags |= MSG_MORE;
770 
771 	/* No MSG_EOR from splice, only look at MSG_MORE */
772 	eor = !(flags & MSG_MORE);
773 
774 	lock_sock(sk);
775 
776 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
777 
778 	err = -EPIPE;
779 	if (sk->sk_err)
780 		goto out_error;
781 
782 	if (kcm->seq_skb) {
783 		/* Previously opened message */
784 		head = kcm->seq_skb;
785 		skb = kcm_tx_msg(head)->last_skb;
786 		i = skb_shinfo(skb)->nr_frags;
787 
788 		if (skb_can_coalesce(skb, i, page, offset)) {
789 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], size);
790 			skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
791 			goto coalesced;
792 		}
793 
794 		if (i >= MAX_SKB_FRAGS) {
795 			struct sk_buff *tskb;
796 
797 			tskb = alloc_skb(0, sk->sk_allocation);
798 			while (!tskb) {
799 				kcm_push(kcm);
800 				err = sk_stream_wait_memory(sk, &timeo);
801 				if (err)
802 					goto out_error;
803 			}
804 
805 			if (head == skb)
806 				skb_shinfo(head)->frag_list = tskb;
807 			else
808 				skb->next = tskb;
809 
810 			skb = tskb;
811 			skb->ip_summed = CHECKSUM_UNNECESSARY;
812 			i = 0;
813 		}
814 	} else {
815 		/* Call the sk_stream functions to manage the sndbuf mem. */
816 		if (!sk_stream_memory_free(sk)) {
817 			kcm_push(kcm);
818 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
819 			err = sk_stream_wait_memory(sk, &timeo);
820 			if (err)
821 				goto out_error;
822 		}
823 
824 		head = alloc_skb(0, sk->sk_allocation);
825 		while (!head) {
826 			kcm_push(kcm);
827 			err = sk_stream_wait_memory(sk, &timeo);
828 			if (err)
829 				goto out_error;
830 		}
831 
832 		skb = head;
833 		i = 0;
834 	}
835 
836 	get_page(page);
837 	skb_fill_page_desc(skb, i, page, offset, size);
838 	skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
839 
840 coalesced:
841 	skb->len += size;
842 	skb->data_len += size;
843 	skb->truesize += size;
844 	sk->sk_wmem_queued += size;
845 	sk_mem_charge(sk, size);
846 
847 	if (head != skb) {
848 		head->len += size;
849 		head->data_len += size;
850 		head->truesize += size;
851 	}
852 
853 	if (eor) {
854 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
855 
856 		/* Message complete, queue it on send buffer */
857 		__skb_queue_tail(&sk->sk_write_queue, head);
858 		kcm->seq_skb = NULL;
859 		KCM_STATS_INCR(kcm->stats.tx_msgs);
860 
861 		if (flags & MSG_BATCH) {
862 			kcm->tx_wait_more = true;
863 		} else if (kcm->tx_wait_more || not_busy) {
864 			err = kcm_write_msgs(kcm);
865 			if (err < 0) {
866 				/* We got a hard error in write_msgs but have
867 				 * already queued this message. Report an error
868 				 * in the socket, but don't affect return value
869 				 * from sendmsg
870 				 */
871 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
872 				report_csk_error(&kcm->sk, -err);
873 			}
874 		}
875 	} else {
876 		/* Message not complete, save state */
877 		kcm->seq_skb = head;
878 		kcm_tx_msg(head)->last_skb = skb;
879 	}
880 
881 	KCM_STATS_ADD(kcm->stats.tx_bytes, size);
882 
883 	release_sock(sk);
884 	return size;
885 
886 out_error:
887 	kcm_push(kcm);
888 
889 	err = sk_stream_error(sk, flags, err);
890 
891 	/* make sure we wake any epoll edge trigger waiter */
892 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
893 		sk->sk_write_space(sk);
894 
895 	release_sock(sk);
896 	return err;
897 }
898 
899 static int kcm_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
900 {
901 	struct sock *sk = sock->sk;
902 	struct kcm_sock *kcm = kcm_sk(sk);
903 	struct sk_buff *skb = NULL, *head = NULL;
904 	size_t copy, copied = 0;
905 	long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
906 	int eor = (sock->type == SOCK_DGRAM) ?
907 		  !(msg->msg_flags & MSG_MORE) : !!(msg->msg_flags & MSG_EOR);
908 	int err = -EPIPE;
909 
910 	lock_sock(sk);
911 
912 	/* Per tcp_sendmsg this should be in poll */
913 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
914 
915 	if (sk->sk_err)
916 		goto out_error;
917 
918 	if (kcm->seq_skb) {
919 		/* Previously opened message */
920 		head = kcm->seq_skb;
921 		skb = kcm_tx_msg(head)->last_skb;
922 		goto start;
923 	}
924 
925 	/* Call the sk_stream functions to manage the sndbuf mem. */
926 	if (!sk_stream_memory_free(sk)) {
927 		kcm_push(kcm);
928 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
929 		err = sk_stream_wait_memory(sk, &timeo);
930 		if (err)
931 			goto out_error;
932 	}
933 
934 	if (msg_data_left(msg)) {
935 		/* New message, alloc head skb */
936 		head = alloc_skb(0, sk->sk_allocation);
937 		while (!head) {
938 			kcm_push(kcm);
939 			err = sk_stream_wait_memory(sk, &timeo);
940 			if (err)
941 				goto out_error;
942 
943 			head = alloc_skb(0, sk->sk_allocation);
944 		}
945 
946 		skb = head;
947 
948 		/* Set ip_summed to CHECKSUM_UNNECESSARY to avoid calling
949 		 * csum_and_copy_from_iter from skb_do_copy_data_nocache.
950 		 */
951 		skb->ip_summed = CHECKSUM_UNNECESSARY;
952 	}
953 
954 start:
955 	while (msg_data_left(msg)) {
956 		bool merge = true;
957 		int i = skb_shinfo(skb)->nr_frags;
958 		struct page_frag *pfrag = sk_page_frag(sk);
959 
960 		if (!sk_page_frag_refill(sk, pfrag))
961 			goto wait_for_memory;
962 
963 		if (!skb_can_coalesce(skb, i, pfrag->page,
964 				      pfrag->offset)) {
965 			if (i == MAX_SKB_FRAGS) {
966 				struct sk_buff *tskb;
967 
968 				tskb = alloc_skb(0, sk->sk_allocation);
969 				if (!tskb)
970 					goto wait_for_memory;
971 
972 				if (head == skb)
973 					skb_shinfo(head)->frag_list = tskb;
974 				else
975 					skb->next = tskb;
976 
977 				skb = tskb;
978 				skb->ip_summed = CHECKSUM_UNNECESSARY;
979 				continue;
980 			}
981 			merge = false;
982 		}
983 
984 		copy = min_t(int, msg_data_left(msg),
985 			     pfrag->size - pfrag->offset);
986 
987 		if (!sk_wmem_schedule(sk, copy))
988 			goto wait_for_memory;
989 
990 		err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
991 					       pfrag->page,
992 					       pfrag->offset,
993 					       copy);
994 		if (err)
995 			goto out_error;
996 
997 		/* Update the skb. */
998 		if (merge) {
999 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1000 		} else {
1001 			skb_fill_page_desc(skb, i, pfrag->page,
1002 					   pfrag->offset, copy);
1003 			get_page(pfrag->page);
1004 		}
1005 
1006 		pfrag->offset += copy;
1007 		copied += copy;
1008 		if (head != skb) {
1009 			head->len += copy;
1010 			head->data_len += copy;
1011 		}
1012 
1013 		continue;
1014 
1015 wait_for_memory:
1016 		kcm_push(kcm);
1017 		err = sk_stream_wait_memory(sk, &timeo);
1018 		if (err)
1019 			goto out_error;
1020 	}
1021 
1022 	if (eor) {
1023 		bool not_busy = skb_queue_empty(&sk->sk_write_queue);
1024 
1025 		if (head) {
1026 			/* Message complete, queue it on send buffer */
1027 			__skb_queue_tail(&sk->sk_write_queue, head);
1028 			kcm->seq_skb = NULL;
1029 			KCM_STATS_INCR(kcm->stats.tx_msgs);
1030 		}
1031 
1032 		if (msg->msg_flags & MSG_BATCH) {
1033 			kcm->tx_wait_more = true;
1034 		} else if (kcm->tx_wait_more || not_busy) {
1035 			err = kcm_write_msgs(kcm);
1036 			if (err < 0) {
1037 				/* We got a hard error in write_msgs but have
1038 				 * already queued this message. Report an error
1039 				 * in the socket, but don't affect return value
1040 				 * from sendmsg
1041 				 */
1042 				pr_warn("KCM: Hard failure on kcm_write_msgs\n");
1043 				report_csk_error(&kcm->sk, -err);
1044 			}
1045 		}
1046 	} else {
1047 		/* Message not complete, save state */
1048 partial_message:
1049 		if (head) {
1050 			kcm->seq_skb = head;
1051 			kcm_tx_msg(head)->last_skb = skb;
1052 		}
1053 	}
1054 
1055 	KCM_STATS_ADD(kcm->stats.tx_bytes, copied);
1056 
1057 	release_sock(sk);
1058 	return copied;
1059 
1060 out_error:
1061 	kcm_push(kcm);
1062 
1063 	if (copied && sock->type == SOCK_SEQPACKET) {
1064 		/* Wrote some bytes before encountering an
1065 		 * error, return partial success.
1066 		 */
1067 		goto partial_message;
1068 	}
1069 
1070 	if (head != kcm->seq_skb)
1071 		kfree_skb(head);
1072 
1073 	err = sk_stream_error(sk, msg->msg_flags, err);
1074 
1075 	/* make sure we wake any epoll edge trigger waiter */
1076 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1077 		sk->sk_write_space(sk);
1078 
1079 	release_sock(sk);
1080 	return err;
1081 }
1082 
1083 static struct sk_buff *kcm_wait_data(struct sock *sk, int flags,
1084 				     long timeo, int *err)
1085 {
1086 	struct sk_buff *skb;
1087 
1088 	while (!(skb = skb_peek(&sk->sk_receive_queue))) {
1089 		if (sk->sk_err) {
1090 			*err = sock_error(sk);
1091 			return NULL;
1092 		}
1093 
1094 		if (sock_flag(sk, SOCK_DONE))
1095 			return NULL;
1096 
1097 		if ((flags & MSG_DONTWAIT) || !timeo) {
1098 			*err = -EAGAIN;
1099 			return NULL;
1100 		}
1101 
1102 		sk_wait_data(sk, &timeo, NULL);
1103 
1104 		/* Handle signals */
1105 		if (signal_pending(current)) {
1106 			*err = sock_intr_errno(timeo);
1107 			return NULL;
1108 		}
1109 	}
1110 
1111 	return skb;
1112 }
1113 
1114 static int kcm_recvmsg(struct socket *sock, struct msghdr *msg,
1115 		       size_t len, int flags)
1116 {
1117 	struct sock *sk = sock->sk;
1118 	struct kcm_sock *kcm = kcm_sk(sk);
1119 	int err = 0;
1120 	long timeo;
1121 	struct strp_msg *stm;
1122 	int copied = 0;
1123 	struct sk_buff *skb;
1124 
1125 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1126 
1127 	lock_sock(sk);
1128 
1129 	skb = kcm_wait_data(sk, flags, timeo, &err);
1130 	if (!skb)
1131 		goto out;
1132 
1133 	/* Okay, have a message on the receive queue */
1134 
1135 	stm = strp_msg(skb);
1136 
1137 	if (len > stm->full_len)
1138 		len = stm->full_len;
1139 
1140 	err = skb_copy_datagram_msg(skb, stm->offset, msg, len);
1141 	if (err < 0)
1142 		goto out;
1143 
1144 	copied = len;
1145 	if (likely(!(flags & MSG_PEEK))) {
1146 		KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1147 		if (copied < stm->full_len) {
1148 			if (sock->type == SOCK_DGRAM) {
1149 				/* Truncated message */
1150 				msg->msg_flags |= MSG_TRUNC;
1151 				goto msg_finished;
1152 			}
1153 			stm->offset += copied;
1154 			stm->full_len -= copied;
1155 		} else {
1156 msg_finished:
1157 			/* Finished with message */
1158 			msg->msg_flags |= MSG_EOR;
1159 			KCM_STATS_INCR(kcm->stats.rx_msgs);
1160 			skb_unlink(skb, &sk->sk_receive_queue);
1161 			kfree_skb(skb);
1162 		}
1163 	}
1164 
1165 out:
1166 	release_sock(sk);
1167 
1168 	return copied ? : err;
1169 }
1170 
1171 static ssize_t kcm_splice_read(struct socket *sock, loff_t *ppos,
1172 			       struct pipe_inode_info *pipe, size_t len,
1173 			       unsigned int flags)
1174 {
1175 	struct sock *sk = sock->sk;
1176 	struct kcm_sock *kcm = kcm_sk(sk);
1177 	long timeo;
1178 	struct strp_msg *stm;
1179 	int err = 0;
1180 	ssize_t copied;
1181 	struct sk_buff *skb;
1182 
1183 	/* Only support splice for SOCKSEQPACKET */
1184 
1185 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1186 
1187 	lock_sock(sk);
1188 
1189 	skb = kcm_wait_data(sk, flags, timeo, &err);
1190 	if (!skb)
1191 		goto err_out;
1192 
1193 	/* Okay, have a message on the receive queue */
1194 
1195 	stm = strp_msg(skb);
1196 
1197 	if (len > stm->full_len)
1198 		len = stm->full_len;
1199 
1200 	copied = skb_splice_bits(skb, sk, stm->offset, pipe, len, flags);
1201 	if (copied < 0) {
1202 		err = copied;
1203 		goto err_out;
1204 	}
1205 
1206 	KCM_STATS_ADD(kcm->stats.rx_bytes, copied);
1207 
1208 	stm->offset += copied;
1209 	stm->full_len -= copied;
1210 
1211 	/* We have no way to return MSG_EOR. If all the bytes have been
1212 	 * read we still leave the message in the receive socket buffer.
1213 	 * A subsequent recvmsg needs to be done to return MSG_EOR and
1214 	 * finish reading the message.
1215 	 */
1216 
1217 	release_sock(sk);
1218 
1219 	return copied;
1220 
1221 err_out:
1222 	release_sock(sk);
1223 
1224 	return err;
1225 }
1226 
1227 /* kcm sock lock held */
1228 static void kcm_recv_disable(struct kcm_sock *kcm)
1229 {
1230 	struct kcm_mux *mux = kcm->mux;
1231 
1232 	if (kcm->rx_disabled)
1233 		return;
1234 
1235 	spin_lock_bh(&mux->rx_lock);
1236 
1237 	kcm->rx_disabled = 1;
1238 
1239 	/* If a psock is reserved we'll do cleanup in unreserve */
1240 	if (!kcm->rx_psock) {
1241 		if (kcm->rx_wait) {
1242 			list_del(&kcm->wait_rx_list);
1243 			kcm->rx_wait = false;
1244 		}
1245 
1246 		requeue_rx_msgs(mux, &kcm->sk.sk_receive_queue);
1247 	}
1248 
1249 	spin_unlock_bh(&mux->rx_lock);
1250 }
1251 
1252 /* kcm sock lock held */
1253 static void kcm_recv_enable(struct kcm_sock *kcm)
1254 {
1255 	struct kcm_mux *mux = kcm->mux;
1256 
1257 	if (!kcm->rx_disabled)
1258 		return;
1259 
1260 	spin_lock_bh(&mux->rx_lock);
1261 
1262 	kcm->rx_disabled = 0;
1263 	kcm_rcv_ready(kcm);
1264 
1265 	spin_unlock_bh(&mux->rx_lock);
1266 }
1267 
1268 static int kcm_setsockopt(struct socket *sock, int level, int optname,
1269 			  char __user *optval, unsigned int optlen)
1270 {
1271 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1272 	int val, valbool;
1273 	int err = 0;
1274 
1275 	if (level != SOL_KCM)
1276 		return -ENOPROTOOPT;
1277 
1278 	if (optlen < sizeof(int))
1279 		return -EINVAL;
1280 
1281 	if (get_user(val, (int __user *)optval))
1282 		return -EINVAL;
1283 
1284 	valbool = val ? 1 : 0;
1285 
1286 	switch (optname) {
1287 	case KCM_RECV_DISABLE:
1288 		lock_sock(&kcm->sk);
1289 		if (valbool)
1290 			kcm_recv_disable(kcm);
1291 		else
1292 			kcm_recv_enable(kcm);
1293 		release_sock(&kcm->sk);
1294 		break;
1295 	default:
1296 		err = -ENOPROTOOPT;
1297 	}
1298 
1299 	return err;
1300 }
1301 
1302 static int kcm_getsockopt(struct socket *sock, int level, int optname,
1303 			  char __user *optval, int __user *optlen)
1304 {
1305 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1306 	int val, len;
1307 
1308 	if (level != SOL_KCM)
1309 		return -ENOPROTOOPT;
1310 
1311 	if (get_user(len, optlen))
1312 		return -EFAULT;
1313 
1314 	len = min_t(unsigned int, len, sizeof(int));
1315 	if (len < 0)
1316 		return -EINVAL;
1317 
1318 	switch (optname) {
1319 	case KCM_RECV_DISABLE:
1320 		val = kcm->rx_disabled;
1321 		break;
1322 	default:
1323 		return -ENOPROTOOPT;
1324 	}
1325 
1326 	if (put_user(len, optlen))
1327 		return -EFAULT;
1328 	if (copy_to_user(optval, &val, len))
1329 		return -EFAULT;
1330 	return 0;
1331 }
1332 
1333 static void init_kcm_sock(struct kcm_sock *kcm, struct kcm_mux *mux)
1334 {
1335 	struct kcm_sock *tkcm;
1336 	struct list_head *head;
1337 	int index = 0;
1338 
1339 	/* For SOCK_SEQPACKET sock type, datagram_poll checks the sk_state, so
1340 	 * we set sk_state, otherwise epoll_wait always returns right away with
1341 	 * EPOLLHUP
1342 	 */
1343 	kcm->sk.sk_state = TCP_ESTABLISHED;
1344 
1345 	/* Add to mux's kcm sockets list */
1346 	kcm->mux = mux;
1347 	spin_lock_bh(&mux->lock);
1348 
1349 	head = &mux->kcm_socks;
1350 	list_for_each_entry(tkcm, &mux->kcm_socks, kcm_sock_list) {
1351 		if (tkcm->index != index)
1352 			break;
1353 		head = &tkcm->kcm_sock_list;
1354 		index++;
1355 	}
1356 
1357 	list_add(&kcm->kcm_sock_list, head);
1358 	kcm->index = index;
1359 
1360 	mux->kcm_socks_cnt++;
1361 	spin_unlock_bh(&mux->lock);
1362 
1363 	INIT_WORK(&kcm->tx_work, kcm_tx_work);
1364 
1365 	spin_lock_bh(&mux->rx_lock);
1366 	kcm_rcv_ready(kcm);
1367 	spin_unlock_bh(&mux->rx_lock);
1368 }
1369 
1370 static int kcm_attach(struct socket *sock, struct socket *csock,
1371 		      struct bpf_prog *prog)
1372 {
1373 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1374 	struct kcm_mux *mux = kcm->mux;
1375 	struct sock *csk;
1376 	struct kcm_psock *psock = NULL, *tpsock;
1377 	struct list_head *head;
1378 	int index = 0;
1379 	static const struct strp_callbacks cb = {
1380 		.rcv_msg = kcm_rcv_strparser,
1381 		.parse_msg = kcm_parse_func_strparser,
1382 		.read_sock_done = kcm_read_sock_done,
1383 	};
1384 	int err = 0;
1385 
1386 	csk = csock->sk;
1387 	if (!csk)
1388 		return -EINVAL;
1389 
1390 	lock_sock(csk);
1391 
1392 	/* Only allow TCP sockets to be attached for now */
1393 	if ((csk->sk_family != AF_INET && csk->sk_family != AF_INET6) ||
1394 	    csk->sk_protocol != IPPROTO_TCP) {
1395 		err = -EOPNOTSUPP;
1396 		goto out;
1397 	}
1398 
1399 	/* Don't allow listeners or closed sockets */
1400 	if (csk->sk_state == TCP_LISTEN || csk->sk_state == TCP_CLOSE) {
1401 		err = -EOPNOTSUPP;
1402 		goto out;
1403 	}
1404 
1405 	psock = kmem_cache_zalloc(kcm_psockp, GFP_KERNEL);
1406 	if (!psock) {
1407 		err = -ENOMEM;
1408 		goto out;
1409 	}
1410 
1411 	psock->mux = mux;
1412 	psock->sk = csk;
1413 	psock->bpf_prog = prog;
1414 
1415 	err = strp_init(&psock->strp, csk, &cb);
1416 	if (err) {
1417 		kmem_cache_free(kcm_psockp, psock);
1418 		goto out;
1419 	}
1420 
1421 	write_lock_bh(&csk->sk_callback_lock);
1422 
1423 	/* Check if sk_user_data is aready by KCM or someone else.
1424 	 * Must be done under lock to prevent race conditions.
1425 	 */
1426 	if (csk->sk_user_data) {
1427 		write_unlock_bh(&csk->sk_callback_lock);
1428 		strp_stop(&psock->strp);
1429 		strp_done(&psock->strp);
1430 		kmem_cache_free(kcm_psockp, psock);
1431 		err = -EALREADY;
1432 		goto out;
1433 	}
1434 
1435 	psock->save_data_ready = csk->sk_data_ready;
1436 	psock->save_write_space = csk->sk_write_space;
1437 	psock->save_state_change = csk->sk_state_change;
1438 	csk->sk_user_data = psock;
1439 	csk->sk_data_ready = psock_data_ready;
1440 	csk->sk_write_space = psock_write_space;
1441 	csk->sk_state_change = psock_state_change;
1442 
1443 	write_unlock_bh(&csk->sk_callback_lock);
1444 
1445 	sock_hold(csk);
1446 
1447 	/* Finished initialization, now add the psock to the MUX. */
1448 	spin_lock_bh(&mux->lock);
1449 	head = &mux->psocks;
1450 	list_for_each_entry(tpsock, &mux->psocks, psock_list) {
1451 		if (tpsock->index != index)
1452 			break;
1453 		head = &tpsock->psock_list;
1454 		index++;
1455 	}
1456 
1457 	list_add(&psock->psock_list, head);
1458 	psock->index = index;
1459 
1460 	KCM_STATS_INCR(mux->stats.psock_attach);
1461 	mux->psocks_cnt++;
1462 	psock_now_avail(psock);
1463 	spin_unlock_bh(&mux->lock);
1464 
1465 	/* Schedule RX work in case there are already bytes queued */
1466 	strp_check_rcv(&psock->strp);
1467 
1468 out:
1469 	release_sock(csk);
1470 
1471 	return err;
1472 }
1473 
1474 static int kcm_attach_ioctl(struct socket *sock, struct kcm_attach *info)
1475 {
1476 	struct socket *csock;
1477 	struct bpf_prog *prog;
1478 	int err;
1479 
1480 	csock = sockfd_lookup(info->fd, &err);
1481 	if (!csock)
1482 		return -ENOENT;
1483 
1484 	prog = bpf_prog_get_type(info->bpf_fd, BPF_PROG_TYPE_SOCKET_FILTER);
1485 	if (IS_ERR(prog)) {
1486 		err = PTR_ERR(prog);
1487 		goto out;
1488 	}
1489 
1490 	err = kcm_attach(sock, csock, prog);
1491 	if (err) {
1492 		bpf_prog_put(prog);
1493 		goto out;
1494 	}
1495 
1496 	/* Keep reference on file also */
1497 
1498 	return 0;
1499 out:
1500 	fput(csock->file);
1501 	return err;
1502 }
1503 
1504 static void kcm_unattach(struct kcm_psock *psock)
1505 {
1506 	struct sock *csk = psock->sk;
1507 	struct kcm_mux *mux = psock->mux;
1508 
1509 	lock_sock(csk);
1510 
1511 	/* Stop getting callbacks from TCP socket. After this there should
1512 	 * be no way to reserve a kcm for this psock.
1513 	 */
1514 	write_lock_bh(&csk->sk_callback_lock);
1515 	csk->sk_user_data = NULL;
1516 	csk->sk_data_ready = psock->save_data_ready;
1517 	csk->sk_write_space = psock->save_write_space;
1518 	csk->sk_state_change = psock->save_state_change;
1519 	strp_stop(&psock->strp);
1520 
1521 	if (WARN_ON(psock->rx_kcm)) {
1522 		write_unlock_bh(&csk->sk_callback_lock);
1523 		release_sock(csk);
1524 		return;
1525 	}
1526 
1527 	spin_lock_bh(&mux->rx_lock);
1528 
1529 	/* Stop receiver activities. After this point psock should not be
1530 	 * able to get onto ready list either through callbacks or work.
1531 	 */
1532 	if (psock->ready_rx_msg) {
1533 		list_del(&psock->psock_ready_list);
1534 		kfree_skb(psock->ready_rx_msg);
1535 		psock->ready_rx_msg = NULL;
1536 		KCM_STATS_INCR(mux->stats.rx_ready_drops);
1537 	}
1538 
1539 	spin_unlock_bh(&mux->rx_lock);
1540 
1541 	write_unlock_bh(&csk->sk_callback_lock);
1542 
1543 	/* Call strp_done without sock lock */
1544 	release_sock(csk);
1545 	strp_done(&psock->strp);
1546 	lock_sock(csk);
1547 
1548 	bpf_prog_put(psock->bpf_prog);
1549 
1550 	spin_lock_bh(&mux->lock);
1551 
1552 	aggregate_psock_stats(&psock->stats, &mux->aggregate_psock_stats);
1553 	save_strp_stats(&psock->strp, &mux->aggregate_strp_stats);
1554 
1555 	KCM_STATS_INCR(mux->stats.psock_unattach);
1556 
1557 	if (psock->tx_kcm) {
1558 		/* psock was reserved.  Just mark it finished and we will clean
1559 		 * up in the kcm paths, we need kcm lock which can not be
1560 		 * acquired here.
1561 		 */
1562 		KCM_STATS_INCR(mux->stats.psock_unattach_rsvd);
1563 		spin_unlock_bh(&mux->lock);
1564 
1565 		/* We are unattaching a socket that is reserved. Abort the
1566 		 * socket since we may be out of sync in sending on it. We need
1567 		 * to do this without the mux lock.
1568 		 */
1569 		kcm_abort_tx_psock(psock, EPIPE, false);
1570 
1571 		spin_lock_bh(&mux->lock);
1572 		if (!psock->tx_kcm) {
1573 			/* psock now unreserved in window mux was unlocked */
1574 			goto no_reserved;
1575 		}
1576 		psock->done = 1;
1577 
1578 		/* Commit done before queuing work to process it */
1579 		smp_mb();
1580 
1581 		/* Queue tx work to make sure psock->done is handled */
1582 		queue_work(kcm_wq, &psock->tx_kcm->tx_work);
1583 		spin_unlock_bh(&mux->lock);
1584 	} else {
1585 no_reserved:
1586 		if (!psock->tx_stopped)
1587 			list_del(&psock->psock_avail_list);
1588 		list_del(&psock->psock_list);
1589 		mux->psocks_cnt--;
1590 		spin_unlock_bh(&mux->lock);
1591 
1592 		sock_put(csk);
1593 		fput(csk->sk_socket->file);
1594 		kmem_cache_free(kcm_psockp, psock);
1595 	}
1596 
1597 	release_sock(csk);
1598 }
1599 
1600 static int kcm_unattach_ioctl(struct socket *sock, struct kcm_unattach *info)
1601 {
1602 	struct kcm_sock *kcm = kcm_sk(sock->sk);
1603 	struct kcm_mux *mux = kcm->mux;
1604 	struct kcm_psock *psock;
1605 	struct socket *csock;
1606 	struct sock *csk;
1607 	int err;
1608 
1609 	csock = sockfd_lookup(info->fd, &err);
1610 	if (!csock)
1611 		return -ENOENT;
1612 
1613 	csk = csock->sk;
1614 	if (!csk) {
1615 		err = -EINVAL;
1616 		goto out;
1617 	}
1618 
1619 	err = -ENOENT;
1620 
1621 	spin_lock_bh(&mux->lock);
1622 
1623 	list_for_each_entry(psock, &mux->psocks, psock_list) {
1624 		if (psock->sk != csk)
1625 			continue;
1626 
1627 		/* Found the matching psock */
1628 
1629 		if (psock->unattaching || WARN_ON(psock->done)) {
1630 			err = -EALREADY;
1631 			break;
1632 		}
1633 
1634 		psock->unattaching = 1;
1635 
1636 		spin_unlock_bh(&mux->lock);
1637 
1638 		/* Lower socket lock should already be held */
1639 		kcm_unattach(psock);
1640 
1641 		err = 0;
1642 		goto out;
1643 	}
1644 
1645 	spin_unlock_bh(&mux->lock);
1646 
1647 out:
1648 	fput(csock->file);
1649 	return err;
1650 }
1651 
1652 static struct proto kcm_proto = {
1653 	.name	= "KCM",
1654 	.owner	= THIS_MODULE,
1655 	.obj_size = sizeof(struct kcm_sock),
1656 };
1657 
1658 /* Clone a kcm socket. */
1659 static struct file *kcm_clone(struct socket *osock)
1660 {
1661 	struct socket *newsock;
1662 	struct sock *newsk;
1663 
1664 	newsock = sock_alloc();
1665 	if (!newsock)
1666 		return ERR_PTR(-ENFILE);
1667 
1668 	newsock->type = osock->type;
1669 	newsock->ops = osock->ops;
1670 
1671 	__module_get(newsock->ops->owner);
1672 
1673 	newsk = sk_alloc(sock_net(osock->sk), PF_KCM, GFP_KERNEL,
1674 			 &kcm_proto, false);
1675 	if (!newsk) {
1676 		sock_release(newsock);
1677 		return ERR_PTR(-ENOMEM);
1678 	}
1679 	sock_init_data(newsock, newsk);
1680 	init_kcm_sock(kcm_sk(newsk), kcm_sk(osock->sk)->mux);
1681 
1682 	return sock_alloc_file(newsock, 0, osock->sk->sk_prot_creator->name);
1683 }
1684 
1685 static int kcm_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
1686 {
1687 	int err;
1688 
1689 	switch (cmd) {
1690 	case SIOCKCMATTACH: {
1691 		struct kcm_attach info;
1692 
1693 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1694 			return -EFAULT;
1695 
1696 		err = kcm_attach_ioctl(sock, &info);
1697 
1698 		break;
1699 	}
1700 	case SIOCKCMUNATTACH: {
1701 		struct kcm_unattach info;
1702 
1703 		if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
1704 			return -EFAULT;
1705 
1706 		err = kcm_unattach_ioctl(sock, &info);
1707 
1708 		break;
1709 	}
1710 	case SIOCKCMCLONE: {
1711 		struct kcm_clone info;
1712 		struct file *file;
1713 
1714 		info.fd = get_unused_fd_flags(0);
1715 		if (unlikely(info.fd < 0))
1716 			return info.fd;
1717 
1718 		file = kcm_clone(sock);
1719 		if (IS_ERR(file)) {
1720 			put_unused_fd(info.fd);
1721 			return PTR_ERR(file);
1722 		}
1723 		if (copy_to_user((void __user *)arg, &info,
1724 				 sizeof(info))) {
1725 			put_unused_fd(info.fd);
1726 			fput(file);
1727 			return -EFAULT;
1728 		}
1729 		fd_install(info.fd, file);
1730 		err = 0;
1731 		break;
1732 	}
1733 	default:
1734 		err = -ENOIOCTLCMD;
1735 		break;
1736 	}
1737 
1738 	return err;
1739 }
1740 
1741 static void free_mux(struct rcu_head *rcu)
1742 {
1743 	struct kcm_mux *mux = container_of(rcu,
1744 	    struct kcm_mux, rcu);
1745 
1746 	kmem_cache_free(kcm_muxp, mux);
1747 }
1748 
1749 static void release_mux(struct kcm_mux *mux)
1750 {
1751 	struct kcm_net *knet = mux->knet;
1752 	struct kcm_psock *psock, *tmp_psock;
1753 
1754 	/* Release psocks */
1755 	list_for_each_entry_safe(psock, tmp_psock,
1756 				 &mux->psocks, psock_list) {
1757 		if (!WARN_ON(psock->unattaching))
1758 			kcm_unattach(psock);
1759 	}
1760 
1761 	if (WARN_ON(mux->psocks_cnt))
1762 		return;
1763 
1764 	__skb_queue_purge(&mux->rx_hold_queue);
1765 
1766 	mutex_lock(&knet->mutex);
1767 	aggregate_mux_stats(&mux->stats, &knet->aggregate_mux_stats);
1768 	aggregate_psock_stats(&mux->aggregate_psock_stats,
1769 			      &knet->aggregate_psock_stats);
1770 	aggregate_strp_stats(&mux->aggregate_strp_stats,
1771 			     &knet->aggregate_strp_stats);
1772 	list_del_rcu(&mux->kcm_mux_list);
1773 	knet->count--;
1774 	mutex_unlock(&knet->mutex);
1775 
1776 	call_rcu(&mux->rcu, free_mux);
1777 }
1778 
1779 static void kcm_done(struct kcm_sock *kcm)
1780 {
1781 	struct kcm_mux *mux = kcm->mux;
1782 	struct sock *sk = &kcm->sk;
1783 	int socks_cnt;
1784 
1785 	spin_lock_bh(&mux->rx_lock);
1786 	if (kcm->rx_psock) {
1787 		/* Cleanup in unreserve_rx_kcm */
1788 		WARN_ON(kcm->done);
1789 		kcm->rx_disabled = 1;
1790 		kcm->done = 1;
1791 		spin_unlock_bh(&mux->rx_lock);
1792 		return;
1793 	}
1794 
1795 	if (kcm->rx_wait) {
1796 		list_del(&kcm->wait_rx_list);
1797 		kcm->rx_wait = false;
1798 	}
1799 	/* Move any pending receive messages to other kcm sockets */
1800 	requeue_rx_msgs(mux, &sk->sk_receive_queue);
1801 
1802 	spin_unlock_bh(&mux->rx_lock);
1803 
1804 	if (WARN_ON(sk_rmem_alloc_get(sk)))
1805 		return;
1806 
1807 	/* Detach from MUX */
1808 	spin_lock_bh(&mux->lock);
1809 
1810 	list_del(&kcm->kcm_sock_list);
1811 	mux->kcm_socks_cnt--;
1812 	socks_cnt = mux->kcm_socks_cnt;
1813 
1814 	spin_unlock_bh(&mux->lock);
1815 
1816 	if (!socks_cnt) {
1817 		/* We are done with the mux now. */
1818 		release_mux(mux);
1819 	}
1820 
1821 	WARN_ON(kcm->rx_wait);
1822 
1823 	sock_put(&kcm->sk);
1824 }
1825 
1826 /* Called by kcm_release to close a KCM socket.
1827  * If this is the last KCM socket on the MUX, destroy the MUX.
1828  */
1829 static int kcm_release(struct socket *sock)
1830 {
1831 	struct sock *sk = sock->sk;
1832 	struct kcm_sock *kcm;
1833 	struct kcm_mux *mux;
1834 	struct kcm_psock *psock;
1835 
1836 	if (!sk)
1837 		return 0;
1838 
1839 	kcm = kcm_sk(sk);
1840 	mux = kcm->mux;
1841 
1842 	sock_orphan(sk);
1843 	kfree_skb(kcm->seq_skb);
1844 
1845 	lock_sock(sk);
1846 	/* Purge queue under lock to avoid race condition with tx_work trying
1847 	 * to act when queue is nonempty. If tx_work runs after this point
1848 	 * it will just return.
1849 	 */
1850 	__skb_queue_purge(&sk->sk_write_queue);
1851 
1852 	/* Set tx_stopped. This is checked when psock is bound to a kcm and we
1853 	 * get a writespace callback. This prevents further work being queued
1854 	 * from the callback (unbinding the psock occurs after canceling work.
1855 	 */
1856 	kcm->tx_stopped = 1;
1857 
1858 	release_sock(sk);
1859 
1860 	spin_lock_bh(&mux->lock);
1861 	if (kcm->tx_wait) {
1862 		/* Take of tx_wait list, after this point there should be no way
1863 		 * that a psock will be assigned to this kcm.
1864 		 */
1865 		list_del(&kcm->wait_psock_list);
1866 		kcm->tx_wait = false;
1867 	}
1868 	spin_unlock_bh(&mux->lock);
1869 
1870 	/* Cancel work. After this point there should be no outside references
1871 	 * to the kcm socket.
1872 	 */
1873 	cancel_work_sync(&kcm->tx_work);
1874 
1875 	lock_sock(sk);
1876 	psock = kcm->tx_psock;
1877 	if (psock) {
1878 		/* A psock was reserved, so we need to kill it since it
1879 		 * may already have some bytes queued from a message. We
1880 		 * need to do this after removing kcm from tx_wait list.
1881 		 */
1882 		kcm_abort_tx_psock(psock, EPIPE, false);
1883 		unreserve_psock(kcm);
1884 	}
1885 	release_sock(sk);
1886 
1887 	WARN_ON(kcm->tx_wait);
1888 	WARN_ON(kcm->tx_psock);
1889 
1890 	sock->sk = NULL;
1891 
1892 	kcm_done(kcm);
1893 
1894 	return 0;
1895 }
1896 
1897 static const struct proto_ops kcm_dgram_ops = {
1898 	.family =	PF_KCM,
1899 	.owner =	THIS_MODULE,
1900 	.release =	kcm_release,
1901 	.bind =		sock_no_bind,
1902 	.connect =	sock_no_connect,
1903 	.socketpair =	sock_no_socketpair,
1904 	.accept =	sock_no_accept,
1905 	.getname =	sock_no_getname,
1906 	.poll =		datagram_poll,
1907 	.ioctl =	kcm_ioctl,
1908 	.listen =	sock_no_listen,
1909 	.shutdown =	sock_no_shutdown,
1910 	.setsockopt =	kcm_setsockopt,
1911 	.getsockopt =	kcm_getsockopt,
1912 	.sendmsg =	kcm_sendmsg,
1913 	.recvmsg =	kcm_recvmsg,
1914 	.mmap =		sock_no_mmap,
1915 	.sendpage =	kcm_sendpage,
1916 };
1917 
1918 static const struct proto_ops kcm_seqpacket_ops = {
1919 	.family =	PF_KCM,
1920 	.owner =	THIS_MODULE,
1921 	.release =	kcm_release,
1922 	.bind =		sock_no_bind,
1923 	.connect =	sock_no_connect,
1924 	.socketpair =	sock_no_socketpair,
1925 	.accept =	sock_no_accept,
1926 	.getname =	sock_no_getname,
1927 	.poll =		datagram_poll,
1928 	.ioctl =	kcm_ioctl,
1929 	.listen =	sock_no_listen,
1930 	.shutdown =	sock_no_shutdown,
1931 	.setsockopt =	kcm_setsockopt,
1932 	.getsockopt =	kcm_getsockopt,
1933 	.sendmsg =	kcm_sendmsg,
1934 	.recvmsg =	kcm_recvmsg,
1935 	.mmap =		sock_no_mmap,
1936 	.sendpage =	kcm_sendpage,
1937 	.splice_read =	kcm_splice_read,
1938 };
1939 
1940 /* Create proto operation for kcm sockets */
1941 static int kcm_create(struct net *net, struct socket *sock,
1942 		      int protocol, int kern)
1943 {
1944 	struct kcm_net *knet = net_generic(net, kcm_net_id);
1945 	struct sock *sk;
1946 	struct kcm_mux *mux;
1947 
1948 	switch (sock->type) {
1949 	case SOCK_DGRAM:
1950 		sock->ops = &kcm_dgram_ops;
1951 		break;
1952 	case SOCK_SEQPACKET:
1953 		sock->ops = &kcm_seqpacket_ops;
1954 		break;
1955 	default:
1956 		return -ESOCKTNOSUPPORT;
1957 	}
1958 
1959 	if (protocol != KCMPROTO_CONNECTED)
1960 		return -EPROTONOSUPPORT;
1961 
1962 	sk = sk_alloc(net, PF_KCM, GFP_KERNEL, &kcm_proto, kern);
1963 	if (!sk)
1964 		return -ENOMEM;
1965 
1966 	/* Allocate a kcm mux, shared between KCM sockets */
1967 	mux = kmem_cache_zalloc(kcm_muxp, GFP_KERNEL);
1968 	if (!mux) {
1969 		sk_free(sk);
1970 		return -ENOMEM;
1971 	}
1972 
1973 	spin_lock_init(&mux->lock);
1974 	spin_lock_init(&mux->rx_lock);
1975 	INIT_LIST_HEAD(&mux->kcm_socks);
1976 	INIT_LIST_HEAD(&mux->kcm_rx_waiters);
1977 	INIT_LIST_HEAD(&mux->kcm_tx_waiters);
1978 
1979 	INIT_LIST_HEAD(&mux->psocks);
1980 	INIT_LIST_HEAD(&mux->psocks_ready);
1981 	INIT_LIST_HEAD(&mux->psocks_avail);
1982 
1983 	mux->knet = knet;
1984 
1985 	/* Add new MUX to list */
1986 	mutex_lock(&knet->mutex);
1987 	list_add_rcu(&mux->kcm_mux_list, &knet->mux_list);
1988 	knet->count++;
1989 	mutex_unlock(&knet->mutex);
1990 
1991 	skb_queue_head_init(&mux->rx_hold_queue);
1992 
1993 	/* Init KCM socket */
1994 	sock_init_data(sock, sk);
1995 	init_kcm_sock(kcm_sk(sk), mux);
1996 
1997 	return 0;
1998 }
1999 
2000 static const struct net_proto_family kcm_family_ops = {
2001 	.family = PF_KCM,
2002 	.create = kcm_create,
2003 	.owner  = THIS_MODULE,
2004 };
2005 
2006 static __net_init int kcm_init_net(struct net *net)
2007 {
2008 	struct kcm_net *knet = net_generic(net, kcm_net_id);
2009 
2010 	INIT_LIST_HEAD_RCU(&knet->mux_list);
2011 	mutex_init(&knet->mutex);
2012 
2013 	return 0;
2014 }
2015 
2016 static __net_exit void kcm_exit_net(struct net *net)
2017 {
2018 	struct kcm_net *knet = net_generic(net, kcm_net_id);
2019 
2020 	/* All KCM sockets should be closed at this point, which should mean
2021 	 * that all multiplexors and psocks have been destroyed.
2022 	 */
2023 	WARN_ON(!list_empty(&knet->mux_list));
2024 }
2025 
2026 static struct pernet_operations kcm_net_ops = {
2027 	.init = kcm_init_net,
2028 	.exit = kcm_exit_net,
2029 	.id   = &kcm_net_id,
2030 	.size = sizeof(struct kcm_net),
2031 };
2032 
2033 static int __init kcm_init(void)
2034 {
2035 	int err = -ENOMEM;
2036 
2037 	kcm_muxp = kmem_cache_create("kcm_mux_cache",
2038 				     sizeof(struct kcm_mux), 0,
2039 				     SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2040 	if (!kcm_muxp)
2041 		goto fail;
2042 
2043 	kcm_psockp = kmem_cache_create("kcm_psock_cache",
2044 				       sizeof(struct kcm_psock), 0,
2045 					SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2046 	if (!kcm_psockp)
2047 		goto fail;
2048 
2049 	kcm_wq = create_singlethread_workqueue("kkcmd");
2050 	if (!kcm_wq)
2051 		goto fail;
2052 
2053 	err = proto_register(&kcm_proto, 1);
2054 	if (err)
2055 		goto fail;
2056 
2057 	err = sock_register(&kcm_family_ops);
2058 	if (err)
2059 		goto sock_register_fail;
2060 
2061 	err = register_pernet_device(&kcm_net_ops);
2062 	if (err)
2063 		goto net_ops_fail;
2064 
2065 	err = kcm_proc_init();
2066 	if (err)
2067 		goto proc_init_fail;
2068 
2069 	return 0;
2070 
2071 proc_init_fail:
2072 	unregister_pernet_device(&kcm_net_ops);
2073 
2074 net_ops_fail:
2075 	sock_unregister(PF_KCM);
2076 
2077 sock_register_fail:
2078 	proto_unregister(&kcm_proto);
2079 
2080 fail:
2081 	kmem_cache_destroy(kcm_muxp);
2082 	kmem_cache_destroy(kcm_psockp);
2083 
2084 	if (kcm_wq)
2085 		destroy_workqueue(kcm_wq);
2086 
2087 	return err;
2088 }
2089 
2090 static void __exit kcm_exit(void)
2091 {
2092 	kcm_proc_exit();
2093 	unregister_pernet_device(&kcm_net_ops);
2094 	sock_unregister(PF_KCM);
2095 	proto_unregister(&kcm_proto);
2096 	destroy_workqueue(kcm_wq);
2097 
2098 	kmem_cache_destroy(kcm_muxp);
2099 	kmem_cache_destroy(kcm_psockp);
2100 }
2101 
2102 module_init(kcm_init);
2103 module_exit(kcm_exit);
2104 
2105 MODULE_LICENSE("GPL");
2106 MODULE_ALIAS_NETPROTO(PF_KCM);
2107 
2108