xref: /openbmc/linux/net/core/skmsg.c (revision 09c434b8)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
3 
4 #include <linux/skmsg.h>
5 #include <linux/skbuff.h>
6 #include <linux/scatterlist.h>
7 
8 #include <net/sock.h>
9 #include <net/tcp.h>
10 
11 static bool sk_msg_try_coalesce_ok(struct sk_msg *msg, int elem_first_coalesce)
12 {
13 	if (msg->sg.end > msg->sg.start &&
14 	    elem_first_coalesce < msg->sg.end)
15 		return true;
16 
17 	if (msg->sg.end < msg->sg.start &&
18 	    (elem_first_coalesce > msg->sg.start ||
19 	     elem_first_coalesce < msg->sg.end))
20 		return true;
21 
22 	return false;
23 }
24 
25 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
26 		 int elem_first_coalesce)
27 {
28 	struct page_frag *pfrag = sk_page_frag(sk);
29 	int ret = 0;
30 
31 	len -= msg->sg.size;
32 	while (len > 0) {
33 		struct scatterlist *sge;
34 		u32 orig_offset;
35 		int use, i;
36 
37 		if (!sk_page_frag_refill(sk, pfrag))
38 			return -ENOMEM;
39 
40 		orig_offset = pfrag->offset;
41 		use = min_t(int, len, pfrag->size - orig_offset);
42 		if (!sk_wmem_schedule(sk, use))
43 			return -ENOMEM;
44 
45 		i = msg->sg.end;
46 		sk_msg_iter_var_prev(i);
47 		sge = &msg->sg.data[i];
48 
49 		if (sk_msg_try_coalesce_ok(msg, elem_first_coalesce) &&
50 		    sg_page(sge) == pfrag->page &&
51 		    sge->offset + sge->length == orig_offset) {
52 			sge->length += use;
53 		} else {
54 			if (sk_msg_full(msg)) {
55 				ret = -ENOSPC;
56 				break;
57 			}
58 
59 			sge = &msg->sg.data[msg->sg.end];
60 			sg_unmark_end(sge);
61 			sg_set_page(sge, pfrag->page, use, orig_offset);
62 			get_page(pfrag->page);
63 			sk_msg_iter_next(msg, end);
64 		}
65 
66 		sk_mem_charge(sk, use);
67 		msg->sg.size += use;
68 		pfrag->offset += use;
69 		len -= use;
70 	}
71 
72 	return ret;
73 }
74 EXPORT_SYMBOL_GPL(sk_msg_alloc);
75 
76 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
77 		 u32 off, u32 len)
78 {
79 	int i = src->sg.start;
80 	struct scatterlist *sge = sk_msg_elem(src, i);
81 	struct scatterlist *sgd = NULL;
82 	u32 sge_len, sge_off;
83 
84 	while (off) {
85 		if (sge->length > off)
86 			break;
87 		off -= sge->length;
88 		sk_msg_iter_var_next(i);
89 		if (i == src->sg.end && off)
90 			return -ENOSPC;
91 		sge = sk_msg_elem(src, i);
92 	}
93 
94 	while (len) {
95 		sge_len = sge->length - off;
96 		if (sge_len > len)
97 			sge_len = len;
98 
99 		if (dst->sg.end)
100 			sgd = sk_msg_elem(dst, dst->sg.end - 1);
101 
102 		if (sgd &&
103 		    (sg_page(sge) == sg_page(sgd)) &&
104 		    (sg_virt(sge) + off == sg_virt(sgd) + sgd->length)) {
105 			sgd->length += sge_len;
106 			dst->sg.size += sge_len;
107 		} else if (!sk_msg_full(dst)) {
108 			sge_off = sge->offset + off;
109 			sk_msg_page_add(dst, sg_page(sge), sge_len, sge_off);
110 		} else {
111 			return -ENOSPC;
112 		}
113 
114 		off = 0;
115 		len -= sge_len;
116 		sk_mem_charge(sk, sge_len);
117 		sk_msg_iter_var_next(i);
118 		if (i == src->sg.end && len)
119 			return -ENOSPC;
120 		sge = sk_msg_elem(src, i);
121 	}
122 
123 	return 0;
124 }
125 EXPORT_SYMBOL_GPL(sk_msg_clone);
126 
127 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes)
128 {
129 	int i = msg->sg.start;
130 
131 	do {
132 		struct scatterlist *sge = sk_msg_elem(msg, i);
133 
134 		if (bytes < sge->length) {
135 			sge->length -= bytes;
136 			sge->offset += bytes;
137 			sk_mem_uncharge(sk, bytes);
138 			break;
139 		}
140 
141 		sk_mem_uncharge(sk, sge->length);
142 		bytes -= sge->length;
143 		sge->length = 0;
144 		sge->offset = 0;
145 		sk_msg_iter_var_next(i);
146 	} while (bytes && i != msg->sg.end);
147 	msg->sg.start = i;
148 }
149 EXPORT_SYMBOL_GPL(sk_msg_return_zero);
150 
151 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes)
152 {
153 	int i = msg->sg.start;
154 
155 	do {
156 		struct scatterlist *sge = &msg->sg.data[i];
157 		int uncharge = (bytes < sge->length) ? bytes : sge->length;
158 
159 		sk_mem_uncharge(sk, uncharge);
160 		bytes -= uncharge;
161 		sk_msg_iter_var_next(i);
162 	} while (i != msg->sg.end);
163 }
164 EXPORT_SYMBOL_GPL(sk_msg_return);
165 
166 static int sk_msg_free_elem(struct sock *sk, struct sk_msg *msg, u32 i,
167 			    bool charge)
168 {
169 	struct scatterlist *sge = sk_msg_elem(msg, i);
170 	u32 len = sge->length;
171 
172 	if (charge)
173 		sk_mem_uncharge(sk, len);
174 	if (!msg->skb)
175 		put_page(sg_page(sge));
176 	memset(sge, 0, sizeof(*sge));
177 	return len;
178 }
179 
180 static int __sk_msg_free(struct sock *sk, struct sk_msg *msg, u32 i,
181 			 bool charge)
182 {
183 	struct scatterlist *sge = sk_msg_elem(msg, i);
184 	int freed = 0;
185 
186 	while (msg->sg.size) {
187 		msg->sg.size -= sge->length;
188 		freed += sk_msg_free_elem(sk, msg, i, charge);
189 		sk_msg_iter_var_next(i);
190 		sk_msg_check_to_free(msg, i, msg->sg.size);
191 		sge = sk_msg_elem(msg, i);
192 	}
193 	if (msg->skb)
194 		consume_skb(msg->skb);
195 	sk_msg_init(msg);
196 	return freed;
197 }
198 
199 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg)
200 {
201 	return __sk_msg_free(sk, msg, msg->sg.start, false);
202 }
203 EXPORT_SYMBOL_GPL(sk_msg_free_nocharge);
204 
205 int sk_msg_free(struct sock *sk, struct sk_msg *msg)
206 {
207 	return __sk_msg_free(sk, msg, msg->sg.start, true);
208 }
209 EXPORT_SYMBOL_GPL(sk_msg_free);
210 
211 static void __sk_msg_free_partial(struct sock *sk, struct sk_msg *msg,
212 				  u32 bytes, bool charge)
213 {
214 	struct scatterlist *sge;
215 	u32 i = msg->sg.start;
216 
217 	while (bytes) {
218 		sge = sk_msg_elem(msg, i);
219 		if (!sge->length)
220 			break;
221 		if (bytes < sge->length) {
222 			if (charge)
223 				sk_mem_uncharge(sk, bytes);
224 			sge->length -= bytes;
225 			sge->offset += bytes;
226 			msg->sg.size -= bytes;
227 			break;
228 		}
229 
230 		msg->sg.size -= sge->length;
231 		bytes -= sge->length;
232 		sk_msg_free_elem(sk, msg, i, charge);
233 		sk_msg_iter_var_next(i);
234 		sk_msg_check_to_free(msg, i, bytes);
235 	}
236 	msg->sg.start = i;
237 }
238 
239 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes)
240 {
241 	__sk_msg_free_partial(sk, msg, bytes, true);
242 }
243 EXPORT_SYMBOL_GPL(sk_msg_free_partial);
244 
245 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
246 				  u32 bytes)
247 {
248 	__sk_msg_free_partial(sk, msg, bytes, false);
249 }
250 
251 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len)
252 {
253 	int trim = msg->sg.size - len;
254 	u32 i = msg->sg.end;
255 
256 	if (trim <= 0) {
257 		WARN_ON(trim < 0);
258 		return;
259 	}
260 
261 	sk_msg_iter_var_prev(i);
262 	msg->sg.size = len;
263 	while (msg->sg.data[i].length &&
264 	       trim >= msg->sg.data[i].length) {
265 		trim -= msg->sg.data[i].length;
266 		sk_msg_free_elem(sk, msg, i, true);
267 		sk_msg_iter_var_prev(i);
268 		if (!trim)
269 			goto out;
270 	}
271 
272 	msg->sg.data[i].length -= trim;
273 	sk_mem_uncharge(sk, trim);
274 out:
275 	/* If we trim data before curr pointer update copybreak and current
276 	 * so that any future copy operations start at new copy location.
277 	 * However trimed data that has not yet been used in a copy op
278 	 * does not require an update.
279 	 */
280 	if (msg->sg.curr >= i) {
281 		msg->sg.curr = i;
282 		msg->sg.copybreak = msg->sg.data[i].length;
283 	}
284 	sk_msg_iter_var_next(i);
285 	msg->sg.end = i;
286 }
287 EXPORT_SYMBOL_GPL(sk_msg_trim);
288 
289 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
290 			      struct sk_msg *msg, u32 bytes)
291 {
292 	int i, maxpages, ret = 0, num_elems = sk_msg_elem_used(msg);
293 	const int to_max_pages = MAX_MSG_FRAGS;
294 	struct page *pages[MAX_MSG_FRAGS];
295 	ssize_t orig, copied, use, offset;
296 
297 	orig = msg->sg.size;
298 	while (bytes > 0) {
299 		i = 0;
300 		maxpages = to_max_pages - num_elems;
301 		if (maxpages == 0) {
302 			ret = -EFAULT;
303 			goto out;
304 		}
305 
306 		copied = iov_iter_get_pages(from, pages, bytes, maxpages,
307 					    &offset);
308 		if (copied <= 0) {
309 			ret = -EFAULT;
310 			goto out;
311 		}
312 
313 		iov_iter_advance(from, copied);
314 		bytes -= copied;
315 		msg->sg.size += copied;
316 
317 		while (copied) {
318 			use = min_t(int, copied, PAGE_SIZE - offset);
319 			sg_set_page(&msg->sg.data[msg->sg.end],
320 				    pages[i], use, offset);
321 			sg_unmark_end(&msg->sg.data[msg->sg.end]);
322 			sk_mem_charge(sk, use);
323 
324 			offset = 0;
325 			copied -= use;
326 			sk_msg_iter_next(msg, end);
327 			num_elems++;
328 			i++;
329 		}
330 		/* When zerocopy is mixed with sk_msg_*copy* operations we
331 		 * may have a copybreak set in this case clear and prefer
332 		 * zerocopy remainder when possible.
333 		 */
334 		msg->sg.copybreak = 0;
335 		msg->sg.curr = msg->sg.end;
336 	}
337 out:
338 	/* Revert iov_iter updates, msg will need to use 'trim' later if it
339 	 * also needs to be cleared.
340 	 */
341 	if (ret)
342 		iov_iter_revert(from, msg->sg.size - orig);
343 	return ret;
344 }
345 EXPORT_SYMBOL_GPL(sk_msg_zerocopy_from_iter);
346 
347 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
348 			     struct sk_msg *msg, u32 bytes)
349 {
350 	int ret = -ENOSPC, i = msg->sg.curr;
351 	struct scatterlist *sge;
352 	u32 copy, buf_size;
353 	void *to;
354 
355 	do {
356 		sge = sk_msg_elem(msg, i);
357 		/* This is possible if a trim operation shrunk the buffer */
358 		if (msg->sg.copybreak >= sge->length) {
359 			msg->sg.copybreak = 0;
360 			sk_msg_iter_var_next(i);
361 			if (i == msg->sg.end)
362 				break;
363 			sge = sk_msg_elem(msg, i);
364 		}
365 
366 		buf_size = sge->length - msg->sg.copybreak;
367 		copy = (buf_size > bytes) ? bytes : buf_size;
368 		to = sg_virt(sge) + msg->sg.copybreak;
369 		msg->sg.copybreak += copy;
370 		if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
371 			ret = copy_from_iter_nocache(to, copy, from);
372 		else
373 			ret = copy_from_iter(to, copy, from);
374 		if (ret != copy) {
375 			ret = -EFAULT;
376 			goto out;
377 		}
378 		bytes -= copy;
379 		if (!bytes)
380 			break;
381 		msg->sg.copybreak = 0;
382 		sk_msg_iter_var_next(i);
383 	} while (i != msg->sg.end);
384 out:
385 	msg->sg.curr = i;
386 	return ret;
387 }
388 EXPORT_SYMBOL_GPL(sk_msg_memcopy_from_iter);
389 
390 static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb)
391 {
392 	struct sock *sk = psock->sk;
393 	int copied = 0, num_sge;
394 	struct sk_msg *msg;
395 
396 	msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC);
397 	if (unlikely(!msg))
398 		return -EAGAIN;
399 	if (!sk_rmem_schedule(sk, skb, skb->len)) {
400 		kfree(msg);
401 		return -EAGAIN;
402 	}
403 
404 	sk_msg_init(msg);
405 	num_sge = skb_to_sgvec(skb, msg->sg.data, 0, skb->len);
406 	if (unlikely(num_sge < 0)) {
407 		kfree(msg);
408 		return num_sge;
409 	}
410 
411 	sk_mem_charge(sk, skb->len);
412 	copied = skb->len;
413 	msg->sg.start = 0;
414 	msg->sg.end = num_sge == MAX_MSG_FRAGS ? 0 : num_sge;
415 	msg->skb = skb;
416 
417 	sk_psock_queue_msg(psock, msg);
418 	sk_psock_data_ready(sk, psock);
419 	return copied;
420 }
421 
422 static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
423 			       u32 off, u32 len, bool ingress)
424 {
425 	if (ingress)
426 		return sk_psock_skb_ingress(psock, skb);
427 	else
428 		return skb_send_sock_locked(psock->sk, skb, off, len);
429 }
430 
431 static void sk_psock_backlog(struct work_struct *work)
432 {
433 	struct sk_psock *psock = container_of(work, struct sk_psock, work);
434 	struct sk_psock_work_state *state = &psock->work_state;
435 	struct sk_buff *skb;
436 	bool ingress;
437 	u32 len, off;
438 	int ret;
439 
440 	/* Lock sock to avoid losing sk_socket during loop. */
441 	lock_sock(psock->sk);
442 	if (state->skb) {
443 		skb = state->skb;
444 		len = state->len;
445 		off = state->off;
446 		state->skb = NULL;
447 		goto start;
448 	}
449 
450 	while ((skb = skb_dequeue(&psock->ingress_skb))) {
451 		len = skb->len;
452 		off = 0;
453 start:
454 		ingress = tcp_skb_bpf_ingress(skb);
455 		do {
456 			ret = -EIO;
457 			if (likely(psock->sk->sk_socket))
458 				ret = sk_psock_handle_skb(psock, skb, off,
459 							  len, ingress);
460 			if (ret <= 0) {
461 				if (ret == -EAGAIN) {
462 					state->skb = skb;
463 					state->len = len;
464 					state->off = off;
465 					goto end;
466 				}
467 				/* Hard errors break pipe and stop xmit. */
468 				sk_psock_report_error(psock, ret ? -ret : EPIPE);
469 				sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
470 				kfree_skb(skb);
471 				goto end;
472 			}
473 			off += ret;
474 			len -= ret;
475 		} while (len);
476 
477 		if (!ingress)
478 			kfree_skb(skb);
479 	}
480 end:
481 	release_sock(psock->sk);
482 }
483 
484 struct sk_psock *sk_psock_init(struct sock *sk, int node)
485 {
486 	struct sk_psock *psock = kzalloc_node(sizeof(*psock),
487 					      GFP_ATOMIC | __GFP_NOWARN,
488 					      node);
489 	if (!psock)
490 		return NULL;
491 
492 	psock->sk = sk;
493 	psock->eval =  __SK_NONE;
494 
495 	INIT_LIST_HEAD(&psock->link);
496 	spin_lock_init(&psock->link_lock);
497 
498 	INIT_WORK(&psock->work, sk_psock_backlog);
499 	INIT_LIST_HEAD(&psock->ingress_msg);
500 	skb_queue_head_init(&psock->ingress_skb);
501 
502 	sk_psock_set_state(psock, SK_PSOCK_TX_ENABLED);
503 	refcount_set(&psock->refcnt, 1);
504 
505 	rcu_assign_sk_user_data(sk, psock);
506 	sock_hold(sk);
507 
508 	return psock;
509 }
510 EXPORT_SYMBOL_GPL(sk_psock_init);
511 
512 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock)
513 {
514 	struct sk_psock_link *link;
515 
516 	spin_lock_bh(&psock->link_lock);
517 	link = list_first_entry_or_null(&psock->link, struct sk_psock_link,
518 					list);
519 	if (link)
520 		list_del(&link->list);
521 	spin_unlock_bh(&psock->link_lock);
522 	return link;
523 }
524 
525 void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
526 {
527 	struct sk_msg *msg, *tmp;
528 
529 	list_for_each_entry_safe(msg, tmp, &psock->ingress_msg, list) {
530 		list_del(&msg->list);
531 		sk_msg_free(psock->sk, msg);
532 		kfree(msg);
533 	}
534 }
535 
536 static void sk_psock_zap_ingress(struct sk_psock *psock)
537 {
538 	__skb_queue_purge(&psock->ingress_skb);
539 	__sk_psock_purge_ingress_msg(psock);
540 }
541 
542 static void sk_psock_link_destroy(struct sk_psock *psock)
543 {
544 	struct sk_psock_link *link, *tmp;
545 
546 	list_for_each_entry_safe(link, tmp, &psock->link, list) {
547 		list_del(&link->list);
548 		sk_psock_free_link(link);
549 	}
550 }
551 
552 static void sk_psock_destroy_deferred(struct work_struct *gc)
553 {
554 	struct sk_psock *psock = container_of(gc, struct sk_psock, gc);
555 
556 	/* No sk_callback_lock since already detached. */
557 	strp_stop(&psock->parser.strp);
558 	strp_done(&psock->parser.strp);
559 
560 	cancel_work_sync(&psock->work);
561 
562 	psock_progs_drop(&psock->progs);
563 
564 	sk_psock_link_destroy(psock);
565 	sk_psock_cork_free(psock);
566 	sk_psock_zap_ingress(psock);
567 
568 	if (psock->sk_redir)
569 		sock_put(psock->sk_redir);
570 	sock_put(psock->sk);
571 	kfree(psock);
572 }
573 
574 void sk_psock_destroy(struct rcu_head *rcu)
575 {
576 	struct sk_psock *psock = container_of(rcu, struct sk_psock, rcu);
577 
578 	INIT_WORK(&psock->gc, sk_psock_destroy_deferred);
579 	schedule_work(&psock->gc);
580 }
581 EXPORT_SYMBOL_GPL(sk_psock_destroy);
582 
583 void sk_psock_drop(struct sock *sk, struct sk_psock *psock)
584 {
585 	rcu_assign_sk_user_data(sk, NULL);
586 	sk_psock_cork_free(psock);
587 	sk_psock_zap_ingress(psock);
588 	sk_psock_restore_proto(sk, psock);
589 
590 	write_lock_bh(&sk->sk_callback_lock);
591 	if (psock->progs.skb_parser)
592 		sk_psock_stop_strp(sk, psock);
593 	write_unlock_bh(&sk->sk_callback_lock);
594 	sk_psock_clear_state(psock, SK_PSOCK_TX_ENABLED);
595 
596 	call_rcu(&psock->rcu, sk_psock_destroy);
597 }
598 EXPORT_SYMBOL_GPL(sk_psock_drop);
599 
600 static int sk_psock_map_verd(int verdict, bool redir)
601 {
602 	switch (verdict) {
603 	case SK_PASS:
604 		return redir ? __SK_REDIRECT : __SK_PASS;
605 	case SK_DROP:
606 	default:
607 		break;
608 	}
609 
610 	return __SK_DROP;
611 }
612 
613 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
614 			 struct sk_msg *msg)
615 {
616 	struct bpf_prog *prog;
617 	int ret;
618 
619 	preempt_disable();
620 	rcu_read_lock();
621 	prog = READ_ONCE(psock->progs.msg_parser);
622 	if (unlikely(!prog)) {
623 		ret = __SK_PASS;
624 		goto out;
625 	}
626 
627 	sk_msg_compute_data_pointers(msg);
628 	msg->sk = sk;
629 	ret = BPF_PROG_RUN(prog, msg);
630 	ret = sk_psock_map_verd(ret, msg->sk_redir);
631 	psock->apply_bytes = msg->apply_bytes;
632 	if (ret == __SK_REDIRECT) {
633 		if (psock->sk_redir)
634 			sock_put(psock->sk_redir);
635 		psock->sk_redir = msg->sk_redir;
636 		if (!psock->sk_redir) {
637 			ret = __SK_DROP;
638 			goto out;
639 		}
640 		sock_hold(psock->sk_redir);
641 	}
642 out:
643 	rcu_read_unlock();
644 	preempt_enable();
645 	return ret;
646 }
647 EXPORT_SYMBOL_GPL(sk_psock_msg_verdict);
648 
649 static int sk_psock_bpf_run(struct sk_psock *psock, struct bpf_prog *prog,
650 			    struct sk_buff *skb)
651 {
652 	int ret;
653 
654 	skb->sk = psock->sk;
655 	bpf_compute_data_end_sk_skb(skb);
656 	preempt_disable();
657 	ret = BPF_PROG_RUN(prog, skb);
658 	preempt_enable();
659 	/* strparser clones the skb before handing it to a upper layer,
660 	 * meaning skb_orphan has been called. We NULL sk on the way out
661 	 * to ensure we don't trigger a BUG_ON() in skb/sk operations
662 	 * later and because we are not charging the memory of this skb
663 	 * to any socket yet.
664 	 */
665 	skb->sk = NULL;
666 	return ret;
667 }
668 
669 static struct sk_psock *sk_psock_from_strp(struct strparser *strp)
670 {
671 	struct sk_psock_parser *parser;
672 
673 	parser = container_of(strp, struct sk_psock_parser, strp);
674 	return container_of(parser, struct sk_psock, parser);
675 }
676 
677 static void sk_psock_verdict_apply(struct sk_psock *psock,
678 				   struct sk_buff *skb, int verdict)
679 {
680 	struct sk_psock *psock_other;
681 	struct sock *sk_other;
682 	bool ingress;
683 
684 	switch (verdict) {
685 	case __SK_PASS:
686 		sk_other = psock->sk;
687 		if (sock_flag(sk_other, SOCK_DEAD) ||
688 		    !sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)) {
689 			goto out_free;
690 		}
691 		if (atomic_read(&sk_other->sk_rmem_alloc) <=
692 		    sk_other->sk_rcvbuf) {
693 			struct tcp_skb_cb *tcp = TCP_SKB_CB(skb);
694 
695 			tcp->bpf.flags |= BPF_F_INGRESS;
696 			skb_queue_tail(&psock->ingress_skb, skb);
697 			schedule_work(&psock->work);
698 			break;
699 		}
700 		goto out_free;
701 	case __SK_REDIRECT:
702 		sk_other = tcp_skb_bpf_redirect_fetch(skb);
703 		if (unlikely(!sk_other))
704 			goto out_free;
705 		psock_other = sk_psock(sk_other);
706 		if (!psock_other || sock_flag(sk_other, SOCK_DEAD) ||
707 		    !sk_psock_test_state(psock_other, SK_PSOCK_TX_ENABLED))
708 			goto out_free;
709 		ingress = tcp_skb_bpf_ingress(skb);
710 		if ((!ingress && sock_writeable(sk_other)) ||
711 		    (ingress &&
712 		     atomic_read(&sk_other->sk_rmem_alloc) <=
713 		     sk_other->sk_rcvbuf)) {
714 			if (!ingress)
715 				skb_set_owner_w(skb, sk_other);
716 			skb_queue_tail(&psock_other->ingress_skb, skb);
717 			schedule_work(&psock_other->work);
718 			break;
719 		}
720 		/* fall-through */
721 	case __SK_DROP:
722 		/* fall-through */
723 	default:
724 out_free:
725 		kfree_skb(skb);
726 	}
727 }
728 
729 static void sk_psock_strp_read(struct strparser *strp, struct sk_buff *skb)
730 {
731 	struct sk_psock *psock = sk_psock_from_strp(strp);
732 	struct bpf_prog *prog;
733 	int ret = __SK_DROP;
734 
735 	rcu_read_lock();
736 	prog = READ_ONCE(psock->progs.skb_verdict);
737 	if (likely(prog)) {
738 		skb_orphan(skb);
739 		tcp_skb_bpf_redirect_clear(skb);
740 		ret = sk_psock_bpf_run(psock, prog, skb);
741 		ret = sk_psock_map_verd(ret, tcp_skb_bpf_redirect_fetch(skb));
742 	}
743 	rcu_read_unlock();
744 	sk_psock_verdict_apply(psock, skb, ret);
745 }
746 
747 static int sk_psock_strp_read_done(struct strparser *strp, int err)
748 {
749 	return err;
750 }
751 
752 static int sk_psock_strp_parse(struct strparser *strp, struct sk_buff *skb)
753 {
754 	struct sk_psock *psock = sk_psock_from_strp(strp);
755 	struct bpf_prog *prog;
756 	int ret = skb->len;
757 
758 	rcu_read_lock();
759 	prog = READ_ONCE(psock->progs.skb_parser);
760 	if (likely(prog))
761 		ret = sk_psock_bpf_run(psock, prog, skb);
762 	rcu_read_unlock();
763 	return ret;
764 }
765 
766 /* Called with socket lock held. */
767 static void sk_psock_strp_data_ready(struct sock *sk)
768 {
769 	struct sk_psock *psock;
770 
771 	rcu_read_lock();
772 	psock = sk_psock(sk);
773 	if (likely(psock)) {
774 		write_lock_bh(&sk->sk_callback_lock);
775 		strp_data_ready(&psock->parser.strp);
776 		write_unlock_bh(&sk->sk_callback_lock);
777 	}
778 	rcu_read_unlock();
779 }
780 
781 static void sk_psock_write_space(struct sock *sk)
782 {
783 	struct sk_psock *psock;
784 	void (*write_space)(struct sock *sk);
785 
786 	rcu_read_lock();
787 	psock = sk_psock(sk);
788 	if (likely(psock && sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED)))
789 		schedule_work(&psock->work);
790 	write_space = psock->saved_write_space;
791 	rcu_read_unlock();
792 	write_space(sk);
793 }
794 
795 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
796 {
797 	static const struct strp_callbacks cb = {
798 		.rcv_msg	= sk_psock_strp_read,
799 		.read_sock_done	= sk_psock_strp_read_done,
800 		.parse_msg	= sk_psock_strp_parse,
801 	};
802 
803 	psock->parser.enabled = false;
804 	return strp_init(&psock->parser.strp, sk, &cb);
805 }
806 
807 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
808 {
809 	struct sk_psock_parser *parser = &psock->parser;
810 
811 	if (parser->enabled)
812 		return;
813 
814 	parser->saved_data_ready = sk->sk_data_ready;
815 	sk->sk_data_ready = sk_psock_strp_data_ready;
816 	sk->sk_write_space = sk_psock_write_space;
817 	parser->enabled = true;
818 }
819 
820 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
821 {
822 	struct sk_psock_parser *parser = &psock->parser;
823 
824 	if (!parser->enabled)
825 		return;
826 
827 	sk->sk_data_ready = parser->saved_data_ready;
828 	parser->saved_data_ready = NULL;
829 	strp_stop(&parser->strp);
830 	parser->enabled = false;
831 }
832