xref: /openbmc/linux/io_uring/net.c (revision eac52bfe2333acb382cf03db5186587ee96eda4a)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/file.h>
5 #include <linux/slab.h>
6 #include <linux/net.h>
7 #include <linux/compat.h>
8 #include <net/compat.h>
9 #include <linux/io_uring.h>
10 
11 #include <uapi/linux/io_uring.h>
12 
13 #include "io_uring.h"
14 #include "kbuf.h"
15 #include "alloc_cache.h"
16 #include "net.h"
17 #include "notif.h"
18 #include "rsrc.h"
19 
20 #if defined(CONFIG_NET)
21 struct io_shutdown {
22 	struct file			*file;
23 	int				how;
24 };
25 
26 struct io_accept {
27 	struct file			*file;
28 	struct sockaddr __user		*addr;
29 	int __user			*addr_len;
30 	int				flags;
31 	u32				file_slot;
32 	unsigned long			nofile;
33 };
34 
35 struct io_socket {
36 	struct file			*file;
37 	int				domain;
38 	int				type;
39 	int				protocol;
40 	int				flags;
41 	u32				file_slot;
42 	unsigned long			nofile;
43 };
44 
45 struct io_connect {
46 	struct file			*file;
47 	struct sockaddr __user		*addr;
48 	int				addr_len;
49 	bool				in_progress;
50 	bool				seen_econnaborted;
51 };
52 
53 struct io_sr_msg {
54 	struct file			*file;
55 	union {
56 		struct compat_msghdr __user	*umsg_compat;
57 		struct user_msghdr __user	*umsg;
58 		void __user			*buf;
59 	};
60 	unsigned			len;
61 	unsigned			done_io;
62 	unsigned			msg_flags;
63 	u16				flags;
64 	/* initialised and used only by !msg send variants */
65 	u16				addr_len;
66 	u16				buf_group;
67 	void __user			*addr;
68 	void __user			*msg_control;
69 	/* used only for send zerocopy */
70 	struct io_kiocb 		*notif;
71 };
72 
73 static inline bool io_check_multishot(struct io_kiocb *req,
74 				      unsigned int issue_flags)
75 {
76 	/*
77 	 * When ->locked_cq is set we only allow to post CQEs from the original
78 	 * task context. Usual request completions will be handled in other
79 	 * generic paths but multipoll may decide to post extra cqes.
80 	 */
81 	return !(issue_flags & IO_URING_F_IOWQ) ||
82 		!(issue_flags & IO_URING_F_MULTISHOT) ||
83 		!req->ctx->task_complete;
84 }
85 
86 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
87 {
88 	struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
89 
90 	if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
91 		     sqe->buf_index || sqe->splice_fd_in))
92 		return -EINVAL;
93 
94 	shutdown->how = READ_ONCE(sqe->len);
95 	req->flags |= REQ_F_FORCE_ASYNC;
96 	return 0;
97 }
98 
99 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
100 {
101 	struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
102 	struct socket *sock;
103 	int ret;
104 
105 	WARN_ON_ONCE(issue_flags & IO_URING_F_NONBLOCK);
106 
107 	sock = sock_from_file(req->file);
108 	if (unlikely(!sock))
109 		return -ENOTSOCK;
110 
111 	ret = __sys_shutdown_sock(sock, shutdown->how);
112 	io_req_set_res(req, ret, 0);
113 	return IOU_OK;
114 }
115 
116 static bool io_net_retry(struct socket *sock, int flags)
117 {
118 	if (!(flags & MSG_WAITALL))
119 		return false;
120 	return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
121 }
122 
123 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
124 {
125 	struct io_async_msghdr *hdr = req->async_data;
126 
127 	if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
128 		return;
129 
130 	/* Let normal cleanup path reap it if we fail adding to the cache */
131 	if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
132 		req->async_data = NULL;
133 		req->flags &= ~REQ_F_ASYNC_DATA;
134 	}
135 }
136 
137 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
138 						  unsigned int issue_flags)
139 {
140 	struct io_ring_ctx *ctx = req->ctx;
141 	struct io_cache_entry *entry;
142 	struct io_async_msghdr *hdr;
143 
144 	if (!(issue_flags & IO_URING_F_UNLOCKED)) {
145 		entry = io_alloc_cache_get(&ctx->netmsg_cache);
146 		if (entry) {
147 			hdr = container_of(entry, struct io_async_msghdr, cache);
148 			hdr->free_iov = NULL;
149 			req->flags |= REQ_F_ASYNC_DATA;
150 			req->async_data = hdr;
151 			return hdr;
152 		}
153 	}
154 
155 	if (!io_alloc_async_data(req)) {
156 		hdr = req->async_data;
157 		hdr->free_iov = NULL;
158 		return hdr;
159 	}
160 	return NULL;
161 }
162 
163 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
164 {
165 	/* ->prep_async is always called from the submission context */
166 	return io_msg_alloc_async(req, 0);
167 }
168 
169 static int io_setup_async_msg(struct io_kiocb *req,
170 			      struct io_async_msghdr *kmsg,
171 			      unsigned int issue_flags)
172 {
173 	struct io_async_msghdr *async_msg;
174 
175 	if (req_has_async_data(req))
176 		return -EAGAIN;
177 	async_msg = io_msg_alloc_async(req, issue_flags);
178 	if (!async_msg) {
179 		kfree(kmsg->free_iov);
180 		return -ENOMEM;
181 	}
182 	req->flags |= REQ_F_NEED_CLEANUP;
183 	memcpy(async_msg, kmsg, sizeof(*kmsg));
184 	if (async_msg->msg.msg_name)
185 		async_msg->msg.msg_name = &async_msg->addr;
186 
187 	if ((req->flags & REQ_F_BUFFER_SELECT) && !async_msg->msg.msg_iter.nr_segs)
188 		return -EAGAIN;
189 
190 	/* if were using fast_iov, set it to the new one */
191 	if (iter_is_iovec(&kmsg->msg.msg_iter) && !kmsg->free_iov) {
192 		size_t fast_idx = iter_iov(&kmsg->msg.msg_iter) - kmsg->fast_iov;
193 		async_msg->msg.msg_iter.__iov = &async_msg->fast_iov[fast_idx];
194 	}
195 
196 	return -EAGAIN;
197 }
198 
199 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
200 			       struct io_async_msghdr *iomsg)
201 {
202 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
203 	int ret;
204 
205 	iomsg->msg.msg_name = &iomsg->addr;
206 	iomsg->free_iov = iomsg->fast_iov;
207 	ret = sendmsg_copy_msghdr(&iomsg->msg, sr->umsg, sr->msg_flags,
208 					&iomsg->free_iov);
209 	/* save msg_control as sys_sendmsg() overwrites it */
210 	sr->msg_control = iomsg->msg.msg_control_user;
211 	return ret;
212 }
213 
214 int io_send_prep_async(struct io_kiocb *req)
215 {
216 	struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
217 	struct io_async_msghdr *io;
218 	int ret;
219 
220 	if (!zc->addr || req_has_async_data(req))
221 		return 0;
222 	io = io_msg_alloc_async_prep(req);
223 	if (!io)
224 		return -ENOMEM;
225 	ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
226 	return ret;
227 }
228 
229 static int io_setup_async_addr(struct io_kiocb *req,
230 			      struct sockaddr_storage *addr_storage,
231 			      unsigned int issue_flags)
232 {
233 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
234 	struct io_async_msghdr *io;
235 
236 	if (!sr->addr || req_has_async_data(req))
237 		return -EAGAIN;
238 	io = io_msg_alloc_async(req, issue_flags);
239 	if (!io)
240 		return -ENOMEM;
241 	memcpy(&io->addr, addr_storage, sizeof(io->addr));
242 	return -EAGAIN;
243 }
244 
245 int io_sendmsg_prep_async(struct io_kiocb *req)
246 {
247 	int ret;
248 
249 	if (!io_msg_alloc_async_prep(req))
250 		return -ENOMEM;
251 	ret = io_sendmsg_copy_hdr(req, req->async_data);
252 	if (!ret)
253 		req->flags |= REQ_F_NEED_CLEANUP;
254 	return ret;
255 }
256 
257 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
258 {
259 	struct io_async_msghdr *io = req->async_data;
260 
261 	kfree(io->free_iov);
262 }
263 
264 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
265 {
266 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
267 
268 	if (req->opcode == IORING_OP_SEND) {
269 		if (READ_ONCE(sqe->__pad3[0]))
270 			return -EINVAL;
271 		sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
272 		sr->addr_len = READ_ONCE(sqe->addr_len);
273 	} else if (sqe->addr2 || sqe->file_index) {
274 		return -EINVAL;
275 	}
276 
277 	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
278 	sr->len = READ_ONCE(sqe->len);
279 	sr->flags = READ_ONCE(sqe->ioprio);
280 	if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
281 		return -EINVAL;
282 	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
283 	if (sr->msg_flags & MSG_DONTWAIT)
284 		req->flags |= REQ_F_NOWAIT;
285 
286 #ifdef CONFIG_COMPAT
287 	if (req->ctx->compat)
288 		sr->msg_flags |= MSG_CMSG_COMPAT;
289 #endif
290 	sr->done_io = 0;
291 	return 0;
292 }
293 
294 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
295 {
296 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
297 	struct io_async_msghdr iomsg, *kmsg;
298 	struct socket *sock;
299 	unsigned flags;
300 	int min_ret = 0;
301 	int ret;
302 
303 	sock = sock_from_file(req->file);
304 	if (unlikely(!sock))
305 		return -ENOTSOCK;
306 
307 	if (req_has_async_data(req)) {
308 		kmsg = req->async_data;
309 		kmsg->msg.msg_control_user = sr->msg_control;
310 	} else {
311 		ret = io_sendmsg_copy_hdr(req, &iomsg);
312 		if (ret)
313 			return ret;
314 		kmsg = &iomsg;
315 	}
316 
317 	if (!(req->flags & REQ_F_POLLED) &&
318 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
319 		return io_setup_async_msg(req, kmsg, issue_flags);
320 
321 	flags = sr->msg_flags;
322 	if (issue_flags & IO_URING_F_NONBLOCK)
323 		flags |= MSG_DONTWAIT;
324 	if (flags & MSG_WAITALL)
325 		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
326 
327 	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
328 
329 	if (ret < min_ret) {
330 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
331 			return io_setup_async_msg(req, kmsg, issue_flags);
332 		if (ret > 0 && io_net_retry(sock, flags)) {
333 			kmsg->msg.msg_controllen = 0;
334 			kmsg->msg.msg_control = NULL;
335 			sr->done_io += ret;
336 			req->flags |= REQ_F_PARTIAL_IO;
337 			return io_setup_async_msg(req, kmsg, issue_flags);
338 		}
339 		if (ret == -ERESTARTSYS)
340 			ret = -EINTR;
341 		req_set_fail(req);
342 	}
343 	/* fast path, check for non-NULL to avoid function call */
344 	if (kmsg->free_iov)
345 		kfree(kmsg->free_iov);
346 	req->flags &= ~REQ_F_NEED_CLEANUP;
347 	io_netmsg_recycle(req, issue_flags);
348 	if (ret >= 0)
349 		ret += sr->done_io;
350 	else if (sr->done_io)
351 		ret = sr->done_io;
352 	io_req_set_res(req, ret, 0);
353 	return IOU_OK;
354 }
355 
356 int io_send(struct io_kiocb *req, unsigned int issue_flags)
357 {
358 	struct sockaddr_storage __address;
359 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
360 	struct msghdr msg;
361 	struct socket *sock;
362 	unsigned flags;
363 	int min_ret = 0;
364 	int ret;
365 
366 	msg.msg_name = NULL;
367 	msg.msg_control = NULL;
368 	msg.msg_controllen = 0;
369 	msg.msg_namelen = 0;
370 	msg.msg_ubuf = NULL;
371 
372 	if (sr->addr) {
373 		if (req_has_async_data(req)) {
374 			struct io_async_msghdr *io = req->async_data;
375 
376 			msg.msg_name = &io->addr;
377 		} else {
378 			ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
379 			if (unlikely(ret < 0))
380 				return ret;
381 			msg.msg_name = (struct sockaddr *)&__address;
382 		}
383 		msg.msg_namelen = sr->addr_len;
384 	}
385 
386 	if (!(req->flags & REQ_F_POLLED) &&
387 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
388 		return io_setup_async_addr(req, &__address, issue_flags);
389 
390 	sock = sock_from_file(req->file);
391 	if (unlikely(!sock))
392 		return -ENOTSOCK;
393 
394 	ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len, &msg.msg_iter);
395 	if (unlikely(ret))
396 		return ret;
397 
398 	flags = sr->msg_flags;
399 	if (issue_flags & IO_URING_F_NONBLOCK)
400 		flags |= MSG_DONTWAIT;
401 	if (flags & MSG_WAITALL)
402 		min_ret = iov_iter_count(&msg.msg_iter);
403 
404 	flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
405 	msg.msg_flags = flags;
406 	ret = sock_sendmsg(sock, &msg);
407 	if (ret < min_ret) {
408 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
409 			return io_setup_async_addr(req, &__address, issue_flags);
410 
411 		if (ret > 0 && io_net_retry(sock, flags)) {
412 			sr->len -= ret;
413 			sr->buf += ret;
414 			sr->done_io += ret;
415 			req->flags |= REQ_F_PARTIAL_IO;
416 			return io_setup_async_addr(req, &__address, issue_flags);
417 		}
418 		if (ret == -ERESTARTSYS)
419 			ret = -EINTR;
420 		req_set_fail(req);
421 	}
422 	if (ret >= 0)
423 		ret += sr->done_io;
424 	else if (sr->done_io)
425 		ret = sr->done_io;
426 	io_req_set_res(req, ret, 0);
427 	return IOU_OK;
428 }
429 
430 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
431 {
432 	int hdr;
433 
434 	if (iomsg->namelen < 0)
435 		return true;
436 	if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
437 			       iomsg->namelen, &hdr))
438 		return true;
439 	if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
440 		return true;
441 
442 	return false;
443 }
444 
445 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
446 				 struct io_async_msghdr *iomsg)
447 {
448 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
449 	struct user_msghdr msg;
450 	int ret;
451 
452 	if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
453 		return -EFAULT;
454 
455 	ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
456 	if (ret)
457 		return ret;
458 
459 	if (req->flags & REQ_F_BUFFER_SELECT) {
460 		if (msg.msg_iovlen == 0) {
461 			sr->len = iomsg->fast_iov[0].iov_len = 0;
462 			iomsg->fast_iov[0].iov_base = NULL;
463 			iomsg->free_iov = NULL;
464 		} else if (msg.msg_iovlen > 1) {
465 			return -EINVAL;
466 		} else {
467 			if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
468 				return -EFAULT;
469 			sr->len = iomsg->fast_iov[0].iov_len;
470 			iomsg->free_iov = NULL;
471 		}
472 
473 		if (req->flags & REQ_F_APOLL_MULTISHOT) {
474 			iomsg->namelen = msg.msg_namelen;
475 			iomsg->controllen = msg.msg_controllen;
476 			if (io_recvmsg_multishot_overflow(iomsg))
477 				return -EOVERFLOW;
478 		}
479 	} else {
480 		iomsg->free_iov = iomsg->fast_iov;
481 		ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
482 				     &iomsg->free_iov, &iomsg->msg.msg_iter,
483 				     false);
484 		if (ret > 0)
485 			ret = 0;
486 	}
487 
488 	return ret;
489 }
490 
491 #ifdef CONFIG_COMPAT
492 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
493 					struct io_async_msghdr *iomsg)
494 {
495 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
496 	struct compat_msghdr msg;
497 	struct compat_iovec __user *uiov;
498 	int ret;
499 
500 	if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
501 		return -EFAULT;
502 
503 	ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
504 	if (ret)
505 		return ret;
506 
507 	uiov = compat_ptr(msg.msg_iov);
508 	if (req->flags & REQ_F_BUFFER_SELECT) {
509 		compat_ssize_t clen;
510 
511 		iomsg->free_iov = NULL;
512 		if (msg.msg_iovlen == 0) {
513 			sr->len = 0;
514 		} else if (msg.msg_iovlen > 1) {
515 			return -EINVAL;
516 		} else {
517 			if (!access_ok(uiov, sizeof(*uiov)))
518 				return -EFAULT;
519 			if (__get_user(clen, &uiov->iov_len))
520 				return -EFAULT;
521 			if (clen < 0)
522 				return -EINVAL;
523 			sr->len = clen;
524 		}
525 
526 		if (req->flags & REQ_F_APOLL_MULTISHOT) {
527 			iomsg->namelen = msg.msg_namelen;
528 			iomsg->controllen = msg.msg_controllen;
529 			if (io_recvmsg_multishot_overflow(iomsg))
530 				return -EOVERFLOW;
531 		}
532 	} else {
533 		iomsg->free_iov = iomsg->fast_iov;
534 		ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
535 				   UIO_FASTIOV, &iomsg->free_iov,
536 				   &iomsg->msg.msg_iter, true);
537 		if (ret < 0)
538 			return ret;
539 	}
540 
541 	return 0;
542 }
543 #endif
544 
545 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
546 			       struct io_async_msghdr *iomsg)
547 {
548 	iomsg->msg.msg_name = &iomsg->addr;
549 	iomsg->msg.msg_iter.nr_segs = 0;
550 
551 #ifdef CONFIG_COMPAT
552 	if (req->ctx->compat)
553 		return __io_compat_recvmsg_copy_hdr(req, iomsg);
554 #endif
555 
556 	return __io_recvmsg_copy_hdr(req, iomsg);
557 }
558 
559 int io_recvmsg_prep_async(struct io_kiocb *req)
560 {
561 	int ret;
562 
563 	if (!io_msg_alloc_async_prep(req))
564 		return -ENOMEM;
565 	ret = io_recvmsg_copy_hdr(req, req->async_data);
566 	if (!ret)
567 		req->flags |= REQ_F_NEED_CLEANUP;
568 	return ret;
569 }
570 
571 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
572 
573 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
574 {
575 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
576 
577 	if (unlikely(sqe->file_index || sqe->addr2))
578 		return -EINVAL;
579 
580 	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
581 	sr->len = READ_ONCE(sqe->len);
582 	sr->flags = READ_ONCE(sqe->ioprio);
583 	if (sr->flags & ~(RECVMSG_FLAGS))
584 		return -EINVAL;
585 	sr->msg_flags = READ_ONCE(sqe->msg_flags);
586 	if (sr->msg_flags & MSG_DONTWAIT)
587 		req->flags |= REQ_F_NOWAIT;
588 	if (sr->msg_flags & MSG_ERRQUEUE)
589 		req->flags |= REQ_F_CLEAR_POLLIN;
590 	if (sr->flags & IORING_RECV_MULTISHOT) {
591 		if (!(req->flags & REQ_F_BUFFER_SELECT))
592 			return -EINVAL;
593 		if (sr->msg_flags & MSG_WAITALL)
594 			return -EINVAL;
595 		if (req->opcode == IORING_OP_RECV && sr->len)
596 			return -EINVAL;
597 		req->flags |= REQ_F_APOLL_MULTISHOT;
598 		/*
599 		 * Store the buffer group for this multishot receive separately,
600 		 * as if we end up doing an io-wq based issue that selects a
601 		 * buffer, it has to be committed immediately and that will
602 		 * clear ->buf_list. This means we lose the link to the buffer
603 		 * list, and the eventual buffer put on completion then cannot
604 		 * restore it.
605 		 */
606 		sr->buf_group = req->buf_index;
607 	}
608 
609 #ifdef CONFIG_COMPAT
610 	if (req->ctx->compat)
611 		sr->msg_flags |= MSG_CMSG_COMPAT;
612 #endif
613 	sr->done_io = 0;
614 	return 0;
615 }
616 
617 static inline void io_recv_prep_retry(struct io_kiocb *req)
618 {
619 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
620 
621 	sr->done_io = 0;
622 	sr->len = 0; /* get from the provided buffer */
623 	req->buf_index = sr->buf_group;
624 }
625 
626 /*
627  * Finishes io_recv and io_recvmsg.
628  *
629  * Returns true if it is actually finished, or false if it should run
630  * again (for multishot).
631  */
632 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
633 				  struct msghdr *msg, bool mshot_finished,
634 				  unsigned issue_flags)
635 {
636 	unsigned int cflags;
637 
638 	cflags = io_put_kbuf(req, issue_flags);
639 	if (msg->msg_inq && msg->msg_inq != -1)
640 		cflags |= IORING_CQE_F_SOCK_NONEMPTY;
641 
642 	if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
643 		io_req_set_res(req, *ret, cflags);
644 		*ret = IOU_OK;
645 		return true;
646 	}
647 
648 	if (mshot_finished)
649 		goto finish;
650 
651 	/*
652 	 * Fill CQE for this receive and see if we should keep trying to
653 	 * receive from this socket.
654 	 */
655 	if (io_fill_cqe_req_aux(req, issue_flags & IO_URING_F_COMPLETE_DEFER,
656 				*ret, cflags | IORING_CQE_F_MORE)) {
657 		io_recv_prep_retry(req);
658 		/* Known not-empty or unknown state, retry */
659 		if (cflags & IORING_CQE_F_SOCK_NONEMPTY || msg->msg_inq == -1)
660 			return false;
661 		if (issue_flags & IO_URING_F_MULTISHOT)
662 			*ret = IOU_ISSUE_SKIP_COMPLETE;
663 		else
664 			*ret = -EAGAIN;
665 		return true;
666 	}
667 	/* Otherwise stop multishot but use the current result. */
668 finish:
669 	io_req_set_res(req, *ret, cflags);
670 
671 	if (issue_flags & IO_URING_F_MULTISHOT)
672 		*ret = IOU_STOP_MULTISHOT;
673 	else
674 		*ret = IOU_OK;
675 	return true;
676 }
677 
678 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
679 				     struct io_sr_msg *sr, void __user **buf,
680 				     size_t *len)
681 {
682 	unsigned long ubuf = (unsigned long) *buf;
683 	unsigned long hdr;
684 
685 	hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
686 		kmsg->controllen;
687 	if (*len < hdr)
688 		return -EFAULT;
689 
690 	if (kmsg->controllen) {
691 		unsigned long control = ubuf + hdr - kmsg->controllen;
692 
693 		kmsg->msg.msg_control_user = (void __user *) control;
694 		kmsg->msg.msg_controllen = kmsg->controllen;
695 	}
696 
697 	sr->buf = *buf; /* stash for later copy */
698 	*buf = (void __user *) (ubuf + hdr);
699 	kmsg->payloadlen = *len = *len - hdr;
700 	return 0;
701 }
702 
703 struct io_recvmsg_multishot_hdr {
704 	struct io_uring_recvmsg_out msg;
705 	struct sockaddr_storage addr;
706 };
707 
708 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
709 				struct io_async_msghdr *kmsg,
710 				unsigned int flags, bool *finished)
711 {
712 	int err;
713 	int copy_len;
714 	struct io_recvmsg_multishot_hdr hdr;
715 
716 	if (kmsg->namelen)
717 		kmsg->msg.msg_name = &hdr.addr;
718 	kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
719 	kmsg->msg.msg_namelen = 0;
720 
721 	if (sock->file->f_flags & O_NONBLOCK)
722 		flags |= MSG_DONTWAIT;
723 
724 	err = sock_recvmsg(sock, &kmsg->msg, flags);
725 	*finished = err <= 0;
726 	if (err < 0)
727 		return err;
728 
729 	hdr.msg = (struct io_uring_recvmsg_out) {
730 		.controllen = kmsg->controllen - kmsg->msg.msg_controllen,
731 		.flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
732 	};
733 
734 	hdr.msg.payloadlen = err;
735 	if (err > kmsg->payloadlen)
736 		err = kmsg->payloadlen;
737 
738 	copy_len = sizeof(struct io_uring_recvmsg_out);
739 	if (kmsg->msg.msg_namelen > kmsg->namelen)
740 		copy_len += kmsg->namelen;
741 	else
742 		copy_len += kmsg->msg.msg_namelen;
743 
744 	/*
745 	 *      "fromlen shall refer to the value before truncation.."
746 	 *                      1003.1g
747 	 */
748 	hdr.msg.namelen = kmsg->msg.msg_namelen;
749 
750 	/* ensure that there is no gap between hdr and sockaddr_storage */
751 	BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
752 		     sizeof(struct io_uring_recvmsg_out));
753 	if (copy_to_user(io->buf, &hdr, copy_len)) {
754 		*finished = true;
755 		return -EFAULT;
756 	}
757 
758 	return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
759 			kmsg->controllen + err;
760 }
761 
762 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
763 {
764 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
765 	struct io_async_msghdr iomsg, *kmsg;
766 	struct socket *sock;
767 	unsigned flags;
768 	int ret, min_ret = 0;
769 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
770 	bool mshot_finished = true;
771 
772 	sock = sock_from_file(req->file);
773 	if (unlikely(!sock))
774 		return -ENOTSOCK;
775 
776 	if (req_has_async_data(req)) {
777 		kmsg = req->async_data;
778 	} else {
779 		ret = io_recvmsg_copy_hdr(req, &iomsg);
780 		if (ret)
781 			return ret;
782 		kmsg = &iomsg;
783 	}
784 
785 	if (!(req->flags & REQ_F_POLLED) &&
786 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
787 		return io_setup_async_msg(req, kmsg, issue_flags);
788 
789 	if (!io_check_multishot(req, issue_flags))
790 		return io_setup_async_msg(req, kmsg, issue_flags);
791 
792 retry_multishot:
793 	if (io_do_buffer_select(req)) {
794 		void __user *buf;
795 		size_t len = sr->len;
796 
797 		buf = io_buffer_select(req, &len, issue_flags);
798 		if (!buf)
799 			return -ENOBUFS;
800 
801 		if (req->flags & REQ_F_APOLL_MULTISHOT) {
802 			ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
803 			if (ret) {
804 				io_kbuf_recycle(req, issue_flags);
805 				return ret;
806 			}
807 		}
808 
809 		iov_iter_ubuf(&kmsg->msg.msg_iter, ITER_DEST, buf, len);
810 	}
811 
812 	flags = sr->msg_flags;
813 	if (force_nonblock)
814 		flags |= MSG_DONTWAIT;
815 
816 	kmsg->msg.msg_get_inq = 1;
817 	kmsg->msg.msg_inq = -1;
818 	if (req->flags & REQ_F_APOLL_MULTISHOT) {
819 		ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
820 					   &mshot_finished);
821 	} else {
822 		/* disable partial retry for recvmsg with cmsg attached */
823 		if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
824 			min_ret = iov_iter_count(&kmsg->msg.msg_iter);
825 
826 		ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
827 					 kmsg->uaddr, flags);
828 	}
829 
830 	if (ret < min_ret) {
831 		if (ret == -EAGAIN && force_nonblock) {
832 			ret = io_setup_async_msg(req, kmsg, issue_flags);
833 			if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
834 				io_kbuf_recycle(req, issue_flags);
835 				return IOU_ISSUE_SKIP_COMPLETE;
836 			}
837 			return ret;
838 		}
839 		if (ret > 0 && io_net_retry(sock, flags)) {
840 			sr->done_io += ret;
841 			req->flags |= REQ_F_PARTIAL_IO;
842 			return io_setup_async_msg(req, kmsg, issue_flags);
843 		}
844 		if (ret == -ERESTARTSYS)
845 			ret = -EINTR;
846 		req_set_fail(req);
847 	} else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
848 		req_set_fail(req);
849 	}
850 
851 	if (ret > 0)
852 		ret += sr->done_io;
853 	else if (sr->done_io)
854 		ret = sr->done_io;
855 	else
856 		io_kbuf_recycle(req, issue_flags);
857 
858 	if (!io_recv_finish(req, &ret, &kmsg->msg, mshot_finished, issue_flags))
859 		goto retry_multishot;
860 
861 	if (mshot_finished) {
862 		/* fast path, check for non-NULL to avoid function call */
863 		if (kmsg->free_iov)
864 			kfree(kmsg->free_iov);
865 		io_netmsg_recycle(req, issue_flags);
866 		req->flags &= ~REQ_F_NEED_CLEANUP;
867 	}
868 
869 	return ret;
870 }
871 
872 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
873 {
874 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
875 	struct msghdr msg;
876 	struct socket *sock;
877 	unsigned flags;
878 	int ret, min_ret = 0;
879 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
880 	size_t len = sr->len;
881 
882 	if (!(req->flags & REQ_F_POLLED) &&
883 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
884 		return -EAGAIN;
885 
886 	if (!io_check_multishot(req, issue_flags))
887 		return -EAGAIN;
888 
889 	sock = sock_from_file(req->file);
890 	if (unlikely(!sock))
891 		return -ENOTSOCK;
892 
893 	msg.msg_name = NULL;
894 	msg.msg_namelen = 0;
895 	msg.msg_control = NULL;
896 	msg.msg_get_inq = 1;
897 	msg.msg_controllen = 0;
898 	msg.msg_iocb = NULL;
899 	msg.msg_ubuf = NULL;
900 
901 retry_multishot:
902 	if (io_do_buffer_select(req)) {
903 		void __user *buf;
904 
905 		buf = io_buffer_select(req, &len, issue_flags);
906 		if (!buf)
907 			return -ENOBUFS;
908 		sr->buf = buf;
909 		sr->len = len;
910 	}
911 
912 	ret = import_ubuf(ITER_DEST, sr->buf, len, &msg.msg_iter);
913 	if (unlikely(ret))
914 		goto out_free;
915 
916 	msg.msg_inq = -1;
917 	msg.msg_flags = 0;
918 
919 	flags = sr->msg_flags;
920 	if (force_nonblock)
921 		flags |= MSG_DONTWAIT;
922 	if (flags & MSG_WAITALL)
923 		min_ret = iov_iter_count(&msg.msg_iter);
924 
925 	ret = sock_recvmsg(sock, &msg, flags);
926 	if (ret < min_ret) {
927 		if (ret == -EAGAIN && force_nonblock) {
928 			if (issue_flags & IO_URING_F_MULTISHOT) {
929 				io_kbuf_recycle(req, issue_flags);
930 				return IOU_ISSUE_SKIP_COMPLETE;
931 			}
932 
933 			return -EAGAIN;
934 		}
935 		if (ret > 0 && io_net_retry(sock, flags)) {
936 			sr->len -= ret;
937 			sr->buf += ret;
938 			sr->done_io += ret;
939 			req->flags |= REQ_F_PARTIAL_IO;
940 			return -EAGAIN;
941 		}
942 		if (ret == -ERESTARTSYS)
943 			ret = -EINTR;
944 		req_set_fail(req);
945 	} else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
946 out_free:
947 		req_set_fail(req);
948 	}
949 
950 	if (ret > 0)
951 		ret += sr->done_io;
952 	else if (sr->done_io)
953 		ret = sr->done_io;
954 	else
955 		io_kbuf_recycle(req, issue_flags);
956 
957 	if (!io_recv_finish(req, &ret, &msg, ret <= 0, issue_flags))
958 		goto retry_multishot;
959 
960 	return ret;
961 }
962 
963 void io_send_zc_cleanup(struct io_kiocb *req)
964 {
965 	struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
966 	struct io_async_msghdr *io;
967 
968 	if (req_has_async_data(req)) {
969 		io = req->async_data;
970 		/* might be ->fast_iov if *msg_copy_hdr failed */
971 		if (io->free_iov != io->fast_iov)
972 			kfree(io->free_iov);
973 	}
974 	if (zc->notif) {
975 		io_notif_flush(zc->notif);
976 		zc->notif = NULL;
977 	}
978 }
979 
980 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
981 #define IO_ZC_FLAGS_VALID  (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
982 
983 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
984 {
985 	struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
986 	struct io_ring_ctx *ctx = req->ctx;
987 	struct io_kiocb *notif;
988 
989 	if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
990 		return -EINVAL;
991 	/* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
992 	if (req->flags & REQ_F_CQE_SKIP)
993 		return -EINVAL;
994 
995 	notif = zc->notif = io_alloc_notif(ctx);
996 	if (!notif)
997 		return -ENOMEM;
998 	notif->cqe.user_data = req->cqe.user_data;
999 	notif->cqe.res = 0;
1000 	notif->cqe.flags = IORING_CQE_F_NOTIF;
1001 	req->flags |= REQ_F_NEED_CLEANUP;
1002 
1003 	zc->flags = READ_ONCE(sqe->ioprio);
1004 	if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
1005 		if (zc->flags & ~IO_ZC_FLAGS_VALID)
1006 			return -EINVAL;
1007 		if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
1008 			io_notif_set_extended(notif);
1009 			io_notif_to_data(notif)->zc_report = true;
1010 		}
1011 	}
1012 
1013 	if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1014 		unsigned idx = READ_ONCE(sqe->buf_index);
1015 
1016 		if (unlikely(idx >= ctx->nr_user_bufs))
1017 			return -EFAULT;
1018 		idx = array_index_nospec(idx, ctx->nr_user_bufs);
1019 		req->imu = READ_ONCE(ctx->user_bufs[idx]);
1020 		io_req_set_rsrc_node(notif, ctx, 0);
1021 	}
1022 
1023 	if (req->opcode == IORING_OP_SEND_ZC) {
1024 		if (READ_ONCE(sqe->__pad3[0]))
1025 			return -EINVAL;
1026 		zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1027 		zc->addr_len = READ_ONCE(sqe->addr_len);
1028 	} else {
1029 		if (unlikely(sqe->addr2 || sqe->file_index))
1030 			return -EINVAL;
1031 		if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1032 			return -EINVAL;
1033 	}
1034 
1035 	zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1036 	zc->len = READ_ONCE(sqe->len);
1037 	zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1038 	if (zc->msg_flags & MSG_DONTWAIT)
1039 		req->flags |= REQ_F_NOWAIT;
1040 
1041 	zc->done_io = 0;
1042 
1043 #ifdef CONFIG_COMPAT
1044 	if (req->ctx->compat)
1045 		zc->msg_flags |= MSG_CMSG_COMPAT;
1046 #endif
1047 	return 0;
1048 }
1049 
1050 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1051 				 struct iov_iter *from, size_t length)
1052 {
1053 	skb_zcopy_downgrade_managed(skb);
1054 	return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1055 }
1056 
1057 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1058 			   struct iov_iter *from, size_t length)
1059 {
1060 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1061 	int frag = shinfo->nr_frags;
1062 	int ret = 0;
1063 	struct bvec_iter bi;
1064 	ssize_t copied = 0;
1065 	unsigned long truesize = 0;
1066 
1067 	if (!frag)
1068 		shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1069 	else if (unlikely(!skb_zcopy_managed(skb)))
1070 		return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1071 
1072 	bi.bi_size = min(from->count, length);
1073 	bi.bi_bvec_done = from->iov_offset;
1074 	bi.bi_idx = 0;
1075 
1076 	while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1077 		struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1078 
1079 		copied += v.bv_len;
1080 		truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1081 		__skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1082 					   v.bv_offset, v.bv_len);
1083 		bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1084 	}
1085 	if (bi.bi_size)
1086 		ret = -EMSGSIZE;
1087 
1088 	shinfo->nr_frags = frag;
1089 	from->bvec += bi.bi_idx;
1090 	from->nr_segs -= bi.bi_idx;
1091 	from->count -= copied;
1092 	from->iov_offset = bi.bi_bvec_done;
1093 
1094 	skb->data_len += copied;
1095 	skb->len += copied;
1096 	skb->truesize += truesize;
1097 
1098 	if (sk && sk->sk_type == SOCK_STREAM) {
1099 		sk_wmem_queued_add(sk, truesize);
1100 		if (!skb_zcopy_pure(skb))
1101 			sk_mem_charge(sk, truesize);
1102 	} else {
1103 		refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1104 	}
1105 	return ret;
1106 }
1107 
1108 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1109 {
1110 	struct sockaddr_storage __address;
1111 	struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1112 	struct msghdr msg;
1113 	struct socket *sock;
1114 	unsigned msg_flags;
1115 	int ret, min_ret = 0;
1116 
1117 	sock = sock_from_file(req->file);
1118 	if (unlikely(!sock))
1119 		return -ENOTSOCK;
1120 	if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1121 		return -EOPNOTSUPP;
1122 
1123 	msg.msg_name = NULL;
1124 	msg.msg_control = NULL;
1125 	msg.msg_controllen = 0;
1126 	msg.msg_namelen = 0;
1127 
1128 	if (zc->addr) {
1129 		if (req_has_async_data(req)) {
1130 			struct io_async_msghdr *io = req->async_data;
1131 
1132 			msg.msg_name = &io->addr;
1133 		} else {
1134 			ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1135 			if (unlikely(ret < 0))
1136 				return ret;
1137 			msg.msg_name = (struct sockaddr *)&__address;
1138 		}
1139 		msg.msg_namelen = zc->addr_len;
1140 	}
1141 
1142 	if (!(req->flags & REQ_F_POLLED) &&
1143 	    (zc->flags & IORING_RECVSEND_POLL_FIRST))
1144 		return io_setup_async_addr(req, &__address, issue_flags);
1145 
1146 	if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1147 		ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1148 					(u64)(uintptr_t)zc->buf, zc->len);
1149 		if (unlikely(ret))
1150 			return ret;
1151 		msg.sg_from_iter = io_sg_from_iter;
1152 	} else {
1153 		io_notif_set_extended(zc->notif);
1154 		ret = import_ubuf(ITER_SOURCE, zc->buf, zc->len, &msg.msg_iter);
1155 		if (unlikely(ret))
1156 			return ret;
1157 		ret = io_notif_account_mem(zc->notif, zc->len);
1158 		if (unlikely(ret))
1159 			return ret;
1160 		msg.sg_from_iter = io_sg_from_iter_iovec;
1161 	}
1162 
1163 	msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1164 	if (issue_flags & IO_URING_F_NONBLOCK)
1165 		msg_flags |= MSG_DONTWAIT;
1166 	if (msg_flags & MSG_WAITALL)
1167 		min_ret = iov_iter_count(&msg.msg_iter);
1168 	msg_flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
1169 
1170 	msg.msg_flags = msg_flags;
1171 	msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1172 	ret = sock_sendmsg(sock, &msg);
1173 
1174 	if (unlikely(ret < min_ret)) {
1175 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1176 			return io_setup_async_addr(req, &__address, issue_flags);
1177 
1178 		if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1179 			zc->len -= ret;
1180 			zc->buf += ret;
1181 			zc->done_io += ret;
1182 			req->flags |= REQ_F_PARTIAL_IO;
1183 			return io_setup_async_addr(req, &__address, issue_flags);
1184 		}
1185 		if (ret == -ERESTARTSYS)
1186 			ret = -EINTR;
1187 		req_set_fail(req);
1188 	}
1189 
1190 	if (ret >= 0)
1191 		ret += zc->done_io;
1192 	else if (zc->done_io)
1193 		ret = zc->done_io;
1194 
1195 	/*
1196 	 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1197 	 * flushing notif to io_send_zc_cleanup()
1198 	 */
1199 	if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1200 		io_notif_flush(zc->notif);
1201 		req->flags &= ~REQ_F_NEED_CLEANUP;
1202 	}
1203 	io_req_set_res(req, ret, IORING_CQE_F_MORE);
1204 	return IOU_OK;
1205 }
1206 
1207 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1208 {
1209 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1210 	struct io_async_msghdr iomsg, *kmsg;
1211 	struct socket *sock;
1212 	unsigned flags;
1213 	int ret, min_ret = 0;
1214 
1215 	io_notif_set_extended(sr->notif);
1216 
1217 	sock = sock_from_file(req->file);
1218 	if (unlikely(!sock))
1219 		return -ENOTSOCK;
1220 	if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1221 		return -EOPNOTSUPP;
1222 
1223 	if (req_has_async_data(req)) {
1224 		kmsg = req->async_data;
1225 	} else {
1226 		ret = io_sendmsg_copy_hdr(req, &iomsg);
1227 		if (ret)
1228 			return ret;
1229 		kmsg = &iomsg;
1230 	}
1231 
1232 	if (!(req->flags & REQ_F_POLLED) &&
1233 	    (sr->flags & IORING_RECVSEND_POLL_FIRST))
1234 		return io_setup_async_msg(req, kmsg, issue_flags);
1235 
1236 	flags = sr->msg_flags | MSG_ZEROCOPY;
1237 	if (issue_flags & IO_URING_F_NONBLOCK)
1238 		flags |= MSG_DONTWAIT;
1239 	if (flags & MSG_WAITALL)
1240 		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1241 
1242 	kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1243 	kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1244 	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1245 
1246 	if (unlikely(ret < min_ret)) {
1247 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1248 			return io_setup_async_msg(req, kmsg, issue_flags);
1249 
1250 		if (ret > 0 && io_net_retry(sock, flags)) {
1251 			sr->done_io += ret;
1252 			req->flags |= REQ_F_PARTIAL_IO;
1253 			return io_setup_async_msg(req, kmsg, issue_flags);
1254 		}
1255 		if (ret == -ERESTARTSYS)
1256 			ret = -EINTR;
1257 		req_set_fail(req);
1258 	}
1259 	/* fast path, check for non-NULL to avoid function call */
1260 	if (kmsg->free_iov) {
1261 		kfree(kmsg->free_iov);
1262 		kmsg->free_iov = NULL;
1263 	}
1264 
1265 	io_netmsg_recycle(req, issue_flags);
1266 	if (ret >= 0)
1267 		ret += sr->done_io;
1268 	else if (sr->done_io)
1269 		ret = sr->done_io;
1270 
1271 	/*
1272 	 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1273 	 * flushing notif to io_send_zc_cleanup()
1274 	 */
1275 	if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1276 		io_notif_flush(sr->notif);
1277 		req->flags &= ~REQ_F_NEED_CLEANUP;
1278 	}
1279 	io_req_set_res(req, ret, IORING_CQE_F_MORE);
1280 	return IOU_OK;
1281 }
1282 
1283 void io_sendrecv_fail(struct io_kiocb *req)
1284 {
1285 	struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1286 
1287 	if (req->flags & REQ_F_PARTIAL_IO)
1288 		req->cqe.res = sr->done_io;
1289 
1290 	if ((req->flags & REQ_F_NEED_CLEANUP) &&
1291 	    (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1292 		req->cqe.flags |= IORING_CQE_F_MORE;
1293 }
1294 
1295 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1296 {
1297 	struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1298 	unsigned flags;
1299 
1300 	if (sqe->len || sqe->buf_index)
1301 		return -EINVAL;
1302 
1303 	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1304 	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1305 	accept->flags = READ_ONCE(sqe->accept_flags);
1306 	accept->nofile = rlimit(RLIMIT_NOFILE);
1307 	flags = READ_ONCE(sqe->ioprio);
1308 	if (flags & ~IORING_ACCEPT_MULTISHOT)
1309 		return -EINVAL;
1310 
1311 	accept->file_slot = READ_ONCE(sqe->file_index);
1312 	if (accept->file_slot) {
1313 		if (accept->flags & SOCK_CLOEXEC)
1314 			return -EINVAL;
1315 		if (flags & IORING_ACCEPT_MULTISHOT &&
1316 		    accept->file_slot != IORING_FILE_INDEX_ALLOC)
1317 			return -EINVAL;
1318 	}
1319 	if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1320 		return -EINVAL;
1321 	if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1322 		accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1323 	if (flags & IORING_ACCEPT_MULTISHOT)
1324 		req->flags |= REQ_F_APOLL_MULTISHOT;
1325 	return 0;
1326 }
1327 
1328 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1329 {
1330 	struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1331 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1332 	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1333 	bool fixed = !!accept->file_slot;
1334 	struct file *file;
1335 	int ret, fd;
1336 
1337 	if (!io_check_multishot(req, issue_flags))
1338 		return -EAGAIN;
1339 retry:
1340 	if (!fixed) {
1341 		fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1342 		if (unlikely(fd < 0))
1343 			return fd;
1344 	}
1345 	file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1346 			 accept->flags);
1347 	if (IS_ERR(file)) {
1348 		if (!fixed)
1349 			put_unused_fd(fd);
1350 		ret = PTR_ERR(file);
1351 		if (ret == -EAGAIN && force_nonblock) {
1352 			/*
1353 			 * if it's multishot and polled, we don't need to
1354 			 * return EAGAIN to arm the poll infra since it
1355 			 * has already been done
1356 			 */
1357 			if (issue_flags & IO_URING_F_MULTISHOT)
1358 				ret = IOU_ISSUE_SKIP_COMPLETE;
1359 			return ret;
1360 		}
1361 		if (ret == -ERESTARTSYS)
1362 			ret = -EINTR;
1363 		req_set_fail(req);
1364 	} else if (!fixed) {
1365 		fd_install(fd, file);
1366 		ret = fd;
1367 	} else {
1368 		ret = io_fixed_fd_install(req, issue_flags, file,
1369 						accept->file_slot);
1370 	}
1371 
1372 	if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1373 		io_req_set_res(req, ret, 0);
1374 		return IOU_OK;
1375 	}
1376 
1377 	if (ret < 0)
1378 		return ret;
1379 	if (io_fill_cqe_req_aux(req, issue_flags & IO_URING_F_COMPLETE_DEFER,
1380 				ret, IORING_CQE_F_MORE))
1381 		goto retry;
1382 
1383 	return -ECANCELED;
1384 }
1385 
1386 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1387 {
1388 	struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1389 
1390 	if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1391 		return -EINVAL;
1392 
1393 	sock->domain = READ_ONCE(sqe->fd);
1394 	sock->type = READ_ONCE(sqe->off);
1395 	sock->protocol = READ_ONCE(sqe->len);
1396 	sock->file_slot = READ_ONCE(sqe->file_index);
1397 	sock->nofile = rlimit(RLIMIT_NOFILE);
1398 
1399 	sock->flags = sock->type & ~SOCK_TYPE_MASK;
1400 	if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1401 		return -EINVAL;
1402 	if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1403 		return -EINVAL;
1404 	return 0;
1405 }
1406 
1407 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1408 {
1409 	struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1410 	bool fixed = !!sock->file_slot;
1411 	struct file *file;
1412 	int ret, fd;
1413 
1414 	if (!fixed) {
1415 		fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1416 		if (unlikely(fd < 0))
1417 			return fd;
1418 	}
1419 	file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1420 	if (IS_ERR(file)) {
1421 		if (!fixed)
1422 			put_unused_fd(fd);
1423 		ret = PTR_ERR(file);
1424 		if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1425 			return -EAGAIN;
1426 		if (ret == -ERESTARTSYS)
1427 			ret = -EINTR;
1428 		req_set_fail(req);
1429 	} else if (!fixed) {
1430 		fd_install(fd, file);
1431 		ret = fd;
1432 	} else {
1433 		ret = io_fixed_fd_install(req, issue_flags, file,
1434 					    sock->file_slot);
1435 	}
1436 	io_req_set_res(req, ret, 0);
1437 	return IOU_OK;
1438 }
1439 
1440 int io_connect_prep_async(struct io_kiocb *req)
1441 {
1442 	struct io_async_connect *io = req->async_data;
1443 	struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1444 
1445 	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1446 }
1447 
1448 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1449 {
1450 	struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1451 
1452 	if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1453 		return -EINVAL;
1454 
1455 	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1456 	conn->addr_len =  READ_ONCE(sqe->addr2);
1457 	conn->in_progress = conn->seen_econnaborted = false;
1458 	return 0;
1459 }
1460 
1461 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1462 {
1463 	struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1464 	struct io_async_connect __io, *io;
1465 	unsigned file_flags;
1466 	int ret;
1467 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1468 
1469 	if (req_has_async_data(req)) {
1470 		io = req->async_data;
1471 	} else {
1472 		ret = move_addr_to_kernel(connect->addr,
1473 						connect->addr_len,
1474 						&__io.address);
1475 		if (ret)
1476 			goto out;
1477 		io = &__io;
1478 	}
1479 
1480 	file_flags = force_nonblock ? O_NONBLOCK : 0;
1481 
1482 	ret = __sys_connect_file(req->file, &io->address,
1483 					connect->addr_len, file_flags);
1484 	if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1485 	    && force_nonblock) {
1486 		if (ret == -EINPROGRESS) {
1487 			connect->in_progress = true;
1488 		} else if (ret == -ECONNABORTED) {
1489 			if (connect->seen_econnaborted)
1490 				goto out;
1491 			connect->seen_econnaborted = true;
1492 		}
1493 		if (req_has_async_data(req))
1494 			return -EAGAIN;
1495 		if (io_alloc_async_data(req)) {
1496 			ret = -ENOMEM;
1497 			goto out;
1498 		}
1499 		memcpy(req->async_data, &__io, sizeof(__io));
1500 		return -EAGAIN;
1501 	}
1502 	if (connect->in_progress) {
1503 		/*
1504 		 * At least bluetooth will return -EBADFD on a re-connect
1505 		 * attempt, and it's (supposedly) also valid to get -EISCONN
1506 		 * which means the previous result is good. For both of these,
1507 		 * grab the sock_error() and use that for the completion.
1508 		 */
1509 		if (ret == -EBADFD || ret == -EISCONN)
1510 			ret = sock_error(sock_from_file(req->file)->sk);
1511 	}
1512 	if (ret == -ERESTARTSYS)
1513 		ret = -EINTR;
1514 out:
1515 	if (ret < 0)
1516 		req_set_fail(req);
1517 	io_req_set_res(req, ret, 0);
1518 	return IOU_OK;
1519 }
1520 
1521 void io_netmsg_cache_free(struct io_cache_entry *entry)
1522 {
1523 	kfree(container_of(entry, struct io_async_msghdr, cache));
1524 }
1525 #endif
1526