Loading...
Note: File does not exist in v5.9.
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)
21struct io_shutdown {
22 struct file *file;
23 int how;
24};
25
26struct io_accept {
27 struct file *file;
28 struct sockaddr __user *addr;
29 int __user *addr_len;
30 int flags;
31 int iou_flags;
32 u32 file_slot;
33 unsigned long nofile;
34};
35
36struct io_socket {
37 struct file *file;
38 int domain;
39 int type;
40 int protocol;
41 int flags;
42 u32 file_slot;
43 unsigned long nofile;
44};
45
46struct io_connect {
47 struct file *file;
48 struct sockaddr __user *addr;
49 int addr_len;
50 bool in_progress;
51 bool seen_econnaborted;
52};
53
54struct io_bind {
55 struct file *file;
56 int addr_len;
57};
58
59struct io_listen {
60 struct file *file;
61 int backlog;
62};
63
64struct io_sr_msg {
65 struct file *file;
66 union {
67 struct compat_msghdr __user *umsg_compat;
68 struct user_msghdr __user *umsg;
69 void __user *buf;
70 };
71 int len;
72 unsigned done_io;
73 unsigned msg_flags;
74 unsigned nr_multishot_loops;
75 u16 flags;
76 /* initialised and used only by !msg send variants */
77 u16 buf_group;
78 u16 buf_index;
79 void __user *msg_control;
80 /* used only for send zerocopy */
81 struct io_kiocb *notif;
82};
83
84/*
85 * Number of times we'll try and do receives if there's more data. If we
86 * exceed this limit, then add us to the back of the queue and retry from
87 * there. This helps fairness between flooding clients.
88 */
89#define MULTISHOT_MAX_RETRY 32
90
91int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
92{
93 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
94
95 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
96 sqe->buf_index || sqe->splice_fd_in))
97 return -EINVAL;
98
99 shutdown->how = READ_ONCE(sqe->len);
100 req->flags |= REQ_F_FORCE_ASYNC;
101 return 0;
102}
103
104int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
105{
106 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
107 struct socket *sock;
108 int ret;
109
110 WARN_ON_ONCE(issue_flags & IO_URING_F_NONBLOCK);
111
112 sock = sock_from_file(req->file);
113 if (unlikely(!sock))
114 return -ENOTSOCK;
115
116 ret = __sys_shutdown_sock(sock, shutdown->how);
117 io_req_set_res(req, ret, 0);
118 return IOU_OK;
119}
120
121static bool io_net_retry(struct socket *sock, int flags)
122{
123 if (!(flags & MSG_WAITALL))
124 return false;
125 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
126}
127
128static void io_netmsg_iovec_free(struct io_async_msghdr *kmsg)
129{
130 if (kmsg->free_iov) {
131 kfree(kmsg->free_iov);
132 kmsg->free_iov_nr = 0;
133 kmsg->free_iov = NULL;
134 }
135}
136
137static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
138{
139 struct io_async_msghdr *hdr = req->async_data;
140 struct iovec *iov;
141
142 /* can't recycle, ensure we free the iovec if we have one */
143 if (unlikely(issue_flags & IO_URING_F_UNLOCKED)) {
144 io_netmsg_iovec_free(hdr);
145 return;
146 }
147
148 /* Let normal cleanup path reap it if we fail adding to the cache */
149 iov = hdr->free_iov;
150 if (io_alloc_cache_put(&req->ctx->netmsg_cache, hdr)) {
151 if (iov)
152 kasan_mempool_poison_object(iov);
153 req->async_data = NULL;
154 req->flags &= ~REQ_F_ASYNC_DATA;
155 }
156}
157
158static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req)
159{
160 struct io_ring_ctx *ctx = req->ctx;
161 struct io_async_msghdr *hdr;
162
163 hdr = io_alloc_cache_get(&ctx->netmsg_cache);
164 if (hdr) {
165 if (hdr->free_iov) {
166 kasan_mempool_unpoison_object(hdr->free_iov,
167 hdr->free_iov_nr * sizeof(struct iovec));
168 req->flags |= REQ_F_NEED_CLEANUP;
169 }
170 req->flags |= REQ_F_ASYNC_DATA;
171 req->async_data = hdr;
172 return hdr;
173 }
174
175 if (!io_alloc_async_data(req)) {
176 hdr = req->async_data;
177 hdr->free_iov_nr = 0;
178 hdr->free_iov = NULL;
179 return hdr;
180 }
181 return NULL;
182}
183
184/* assign new iovec to kmsg, if we need to */
185static int io_net_vec_assign(struct io_kiocb *req, struct io_async_msghdr *kmsg,
186 struct iovec *iov)
187{
188 if (iov) {
189 req->flags |= REQ_F_NEED_CLEANUP;
190 kmsg->free_iov_nr = kmsg->msg.msg_iter.nr_segs;
191 if (kmsg->free_iov)
192 kfree(kmsg->free_iov);
193 kmsg->free_iov = iov;
194 }
195 return 0;
196}
197
198static inline void io_mshot_prep_retry(struct io_kiocb *req,
199 struct io_async_msghdr *kmsg)
200{
201 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
202
203 req->flags &= ~REQ_F_BL_EMPTY;
204 sr->done_io = 0;
205 sr->len = 0; /* get from the provided buffer */
206 req->buf_index = sr->buf_group;
207}
208
209#ifdef CONFIG_COMPAT
210static int io_compat_msg_copy_hdr(struct io_kiocb *req,
211 struct io_async_msghdr *iomsg,
212 struct compat_msghdr *msg, int ddir)
213{
214 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
215 struct compat_iovec __user *uiov;
216 struct iovec *iov;
217 int ret, nr_segs;
218
219 if (iomsg->free_iov) {
220 nr_segs = iomsg->free_iov_nr;
221 iov = iomsg->free_iov;
222 } else {
223 iov = &iomsg->fast_iov;
224 nr_segs = 1;
225 }
226
227 if (copy_from_user(msg, sr->umsg_compat, sizeof(*msg)))
228 return -EFAULT;
229
230 uiov = compat_ptr(msg->msg_iov);
231 if (req->flags & REQ_F_BUFFER_SELECT) {
232 compat_ssize_t clen;
233
234 if (msg->msg_iovlen == 0) {
235 sr->len = iov->iov_len = 0;
236 iov->iov_base = NULL;
237 } else if (msg->msg_iovlen > 1) {
238 return -EINVAL;
239 } else {
240 if (!access_ok(uiov, sizeof(*uiov)))
241 return -EFAULT;
242 if (__get_user(clen, &uiov->iov_len))
243 return -EFAULT;
244 if (clen < 0)
245 return -EINVAL;
246 sr->len = clen;
247 }
248
249 return 0;
250 }
251
252 ret = __import_iovec(ddir, (struct iovec __user *)uiov, msg->msg_iovlen,
253 nr_segs, &iov, &iomsg->msg.msg_iter, true);
254 if (unlikely(ret < 0))
255 return ret;
256
257 return io_net_vec_assign(req, iomsg, iov);
258}
259#endif
260
261static int io_msg_copy_hdr(struct io_kiocb *req, struct io_async_msghdr *iomsg,
262 struct user_msghdr *msg, int ddir)
263{
264 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
265 struct user_msghdr __user *umsg = sr->umsg;
266 struct iovec *iov;
267 int ret, nr_segs;
268
269 if (iomsg->free_iov) {
270 nr_segs = iomsg->free_iov_nr;
271 iov = iomsg->free_iov;
272 } else {
273 iov = &iomsg->fast_iov;
274 nr_segs = 1;
275 }
276
277 if (!user_access_begin(umsg, sizeof(*umsg)))
278 return -EFAULT;
279
280 ret = -EFAULT;
281 unsafe_get_user(msg->msg_name, &umsg->msg_name, ua_end);
282 unsafe_get_user(msg->msg_namelen, &umsg->msg_namelen, ua_end);
283 unsafe_get_user(msg->msg_iov, &umsg->msg_iov, ua_end);
284 unsafe_get_user(msg->msg_iovlen, &umsg->msg_iovlen, ua_end);
285 unsafe_get_user(msg->msg_control, &umsg->msg_control, ua_end);
286 unsafe_get_user(msg->msg_controllen, &umsg->msg_controllen, ua_end);
287 msg->msg_flags = 0;
288
289 if (req->flags & REQ_F_BUFFER_SELECT) {
290 if (msg->msg_iovlen == 0) {
291 sr->len = iov->iov_len = 0;
292 iov->iov_base = NULL;
293 } else if (msg->msg_iovlen > 1) {
294 ret = -EINVAL;
295 goto ua_end;
296 } else {
297 /* we only need the length for provided buffers */
298 if (!access_ok(&msg->msg_iov[0].iov_len, sizeof(__kernel_size_t)))
299 goto ua_end;
300 unsafe_get_user(iov->iov_len, &msg->msg_iov[0].iov_len,
301 ua_end);
302 sr->len = iov->iov_len;
303 }
304 ret = 0;
305ua_end:
306 user_access_end();
307 return ret;
308 }
309
310 user_access_end();
311 ret = __import_iovec(ddir, msg->msg_iov, msg->msg_iovlen, nr_segs,
312 &iov, &iomsg->msg.msg_iter, false);
313 if (unlikely(ret < 0))
314 return ret;
315
316 return io_net_vec_assign(req, iomsg, iov);
317}
318
319static int io_sendmsg_copy_hdr(struct io_kiocb *req,
320 struct io_async_msghdr *iomsg)
321{
322 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
323 struct user_msghdr msg;
324 int ret;
325
326 iomsg->msg.msg_name = &iomsg->addr;
327 iomsg->msg.msg_iter.nr_segs = 0;
328
329#ifdef CONFIG_COMPAT
330 if (unlikely(req->ctx->compat)) {
331 struct compat_msghdr cmsg;
332
333 ret = io_compat_msg_copy_hdr(req, iomsg, &cmsg, ITER_SOURCE);
334 if (unlikely(ret))
335 return ret;
336
337 ret = __get_compat_msghdr(&iomsg->msg, &cmsg, NULL);
338 sr->msg_control = iomsg->msg.msg_control_user;
339 return ret;
340 }
341#endif
342
343 ret = io_msg_copy_hdr(req, iomsg, &msg, ITER_SOURCE);
344 if (unlikely(ret))
345 return ret;
346
347 ret = __copy_msghdr(&iomsg->msg, &msg, NULL);
348
349 /* save msg_control as sys_sendmsg() overwrites it */
350 sr->msg_control = iomsg->msg.msg_control_user;
351 return ret;
352}
353
354void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
355{
356 struct io_async_msghdr *io = req->async_data;
357
358 io_netmsg_iovec_free(io);
359}
360
361static int io_send_setup(struct io_kiocb *req, const struct io_uring_sqe *sqe)
362{
363 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
364 struct io_async_msghdr *kmsg = req->async_data;
365 void __user *addr;
366 u16 addr_len;
367 int ret;
368
369 sr->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
370
371 if (READ_ONCE(sqe->__pad3[0]))
372 return -EINVAL;
373
374 kmsg->msg.msg_name = NULL;
375 kmsg->msg.msg_namelen = 0;
376 kmsg->msg.msg_control = NULL;
377 kmsg->msg.msg_controllen = 0;
378 kmsg->msg.msg_ubuf = NULL;
379
380 addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
381 addr_len = READ_ONCE(sqe->addr_len);
382 if (addr) {
383 ret = move_addr_to_kernel(addr, addr_len, &kmsg->addr);
384 if (unlikely(ret < 0))
385 return ret;
386 kmsg->msg.msg_name = &kmsg->addr;
387 kmsg->msg.msg_namelen = addr_len;
388 }
389 if (!io_do_buffer_select(req)) {
390 ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len,
391 &kmsg->msg.msg_iter);
392 if (unlikely(ret < 0))
393 return ret;
394 }
395 return 0;
396}
397
398static int io_sendmsg_setup(struct io_kiocb *req, const struct io_uring_sqe *sqe)
399{
400 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
401 struct io_async_msghdr *kmsg = req->async_data;
402 int ret;
403
404 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
405
406 ret = io_sendmsg_copy_hdr(req, kmsg);
407 if (!ret)
408 req->flags |= REQ_F_NEED_CLEANUP;
409 return ret;
410}
411
412#define SENDMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_BUNDLE)
413
414int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
415{
416 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
417
418 sr->done_io = 0;
419
420 if (req->opcode != IORING_OP_SEND) {
421 if (sqe->addr2 || sqe->file_index)
422 return -EINVAL;
423 }
424
425 sr->len = READ_ONCE(sqe->len);
426 sr->flags = READ_ONCE(sqe->ioprio);
427 if (sr->flags & ~SENDMSG_FLAGS)
428 return -EINVAL;
429 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
430 if (sr->msg_flags & MSG_DONTWAIT)
431 req->flags |= REQ_F_NOWAIT;
432 if (sr->flags & IORING_RECVSEND_BUNDLE) {
433 if (req->opcode == IORING_OP_SENDMSG)
434 return -EINVAL;
435 if (!(req->flags & REQ_F_BUFFER_SELECT))
436 return -EINVAL;
437 sr->msg_flags |= MSG_WAITALL;
438 sr->buf_group = req->buf_index;
439 req->buf_list = NULL;
440 }
441
442#ifdef CONFIG_COMPAT
443 if (req->ctx->compat)
444 sr->msg_flags |= MSG_CMSG_COMPAT;
445#endif
446 if (unlikely(!io_msg_alloc_async(req)))
447 return -ENOMEM;
448 if (req->opcode != IORING_OP_SENDMSG)
449 return io_send_setup(req, sqe);
450 return io_sendmsg_setup(req, sqe);
451}
452
453static void io_req_msg_cleanup(struct io_kiocb *req,
454 unsigned int issue_flags)
455{
456 req->flags &= ~REQ_F_NEED_CLEANUP;
457 io_netmsg_recycle(req, issue_flags);
458}
459
460/*
461 * For bundle completions, we need to figure out how many segments we consumed.
462 * A bundle could be using a single ITER_UBUF if that's all we mapped, or it
463 * could be using an ITER_IOVEC. If the latter, then if we consumed all of
464 * the segments, then it's a trivial questiont o answer. If we have residual
465 * data in the iter, then loop the segments to figure out how much we
466 * transferred.
467 */
468static int io_bundle_nbufs(struct io_async_msghdr *kmsg, int ret)
469{
470 struct iovec *iov;
471 int nbufs;
472
473 /* no data is always zero segments, and a ubuf is always 1 segment */
474 if (ret <= 0)
475 return 0;
476 if (iter_is_ubuf(&kmsg->msg.msg_iter))
477 return 1;
478
479 iov = kmsg->free_iov;
480 if (!iov)
481 iov = &kmsg->fast_iov;
482
483 /* if all data was transferred, it's basic pointer math */
484 if (!iov_iter_count(&kmsg->msg.msg_iter))
485 return iter_iov(&kmsg->msg.msg_iter) - iov;
486
487 /* short transfer, count segments */
488 nbufs = 0;
489 do {
490 int this_len = min_t(int, iov[nbufs].iov_len, ret);
491
492 nbufs++;
493 ret -= this_len;
494 } while (ret);
495
496 return nbufs;
497}
498
499static inline bool io_send_finish(struct io_kiocb *req, int *ret,
500 struct io_async_msghdr *kmsg,
501 unsigned issue_flags)
502{
503 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
504 bool bundle_finished = *ret <= 0;
505 unsigned int cflags;
506
507 if (!(sr->flags & IORING_RECVSEND_BUNDLE)) {
508 cflags = io_put_kbuf(req, *ret, issue_flags);
509 goto finish;
510 }
511
512 cflags = io_put_kbufs(req, *ret, io_bundle_nbufs(kmsg, *ret), issue_flags);
513
514 if (bundle_finished || req->flags & REQ_F_BL_EMPTY)
515 goto finish;
516
517 /*
518 * Fill CQE for this receive and see if we should keep trying to
519 * receive from this socket.
520 */
521 if (io_req_post_cqe(req, *ret, cflags | IORING_CQE_F_MORE)) {
522 io_mshot_prep_retry(req, kmsg);
523 return false;
524 }
525
526 /* Otherwise stop bundle and use the current result. */
527finish:
528 io_req_set_res(req, *ret, cflags);
529 *ret = IOU_OK;
530 return true;
531}
532
533int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
534{
535 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
536 struct io_async_msghdr *kmsg = req->async_data;
537 struct socket *sock;
538 unsigned flags;
539 int min_ret = 0;
540 int ret;
541
542 sock = sock_from_file(req->file);
543 if (unlikely(!sock))
544 return -ENOTSOCK;
545
546 if (!(req->flags & REQ_F_POLLED) &&
547 (sr->flags & IORING_RECVSEND_POLL_FIRST))
548 return -EAGAIN;
549
550 flags = sr->msg_flags;
551 if (issue_flags & IO_URING_F_NONBLOCK)
552 flags |= MSG_DONTWAIT;
553 if (flags & MSG_WAITALL)
554 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
555
556 kmsg->msg.msg_control_user = sr->msg_control;
557
558 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
559
560 if (ret < min_ret) {
561 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
562 return -EAGAIN;
563 if (ret > 0 && io_net_retry(sock, flags)) {
564 kmsg->msg.msg_controllen = 0;
565 kmsg->msg.msg_control = NULL;
566 sr->done_io += ret;
567 req->flags |= REQ_F_BL_NO_RECYCLE;
568 return -EAGAIN;
569 }
570 if (ret == -ERESTARTSYS)
571 ret = -EINTR;
572 req_set_fail(req);
573 }
574 io_req_msg_cleanup(req, issue_flags);
575 if (ret >= 0)
576 ret += sr->done_io;
577 else if (sr->done_io)
578 ret = sr->done_io;
579 io_req_set_res(req, ret, 0);
580 return IOU_OK;
581}
582
583int io_send(struct io_kiocb *req, unsigned int issue_flags)
584{
585 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
586 struct io_async_msghdr *kmsg = req->async_data;
587 struct socket *sock;
588 unsigned flags;
589 int min_ret = 0;
590 int ret;
591
592 sock = sock_from_file(req->file);
593 if (unlikely(!sock))
594 return -ENOTSOCK;
595
596 if (!(req->flags & REQ_F_POLLED) &&
597 (sr->flags & IORING_RECVSEND_POLL_FIRST))
598 return -EAGAIN;
599
600 flags = sr->msg_flags;
601 if (issue_flags & IO_URING_F_NONBLOCK)
602 flags |= MSG_DONTWAIT;
603
604retry_bundle:
605 if (io_do_buffer_select(req)) {
606 struct buf_sel_arg arg = {
607 .iovs = &kmsg->fast_iov,
608 .max_len = min_not_zero(sr->len, INT_MAX),
609 .nr_iovs = 1,
610 };
611
612 if (kmsg->free_iov) {
613 arg.nr_iovs = kmsg->free_iov_nr;
614 arg.iovs = kmsg->free_iov;
615 arg.mode = KBUF_MODE_FREE;
616 }
617
618 if (!(sr->flags & IORING_RECVSEND_BUNDLE))
619 arg.nr_iovs = 1;
620 else
621 arg.mode |= KBUF_MODE_EXPAND;
622
623 ret = io_buffers_select(req, &arg, issue_flags);
624 if (unlikely(ret < 0))
625 return ret;
626
627 if (arg.iovs != &kmsg->fast_iov && arg.iovs != kmsg->free_iov) {
628 kmsg->free_iov_nr = ret;
629 kmsg->free_iov = arg.iovs;
630 req->flags |= REQ_F_NEED_CLEANUP;
631 }
632 sr->len = arg.out_len;
633
634 if (ret == 1) {
635 sr->buf = arg.iovs[0].iov_base;
636 ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len,
637 &kmsg->msg.msg_iter);
638 if (unlikely(ret))
639 return ret;
640 } else {
641 iov_iter_init(&kmsg->msg.msg_iter, ITER_SOURCE,
642 arg.iovs, ret, arg.out_len);
643 }
644 }
645
646 /*
647 * If MSG_WAITALL is set, or this is a bundle send, then we need
648 * the full amount. If just bundle is set, if we do a short send
649 * then we complete the bundle sequence rather than continue on.
650 */
651 if (flags & MSG_WAITALL || sr->flags & IORING_RECVSEND_BUNDLE)
652 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
653
654 flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
655 kmsg->msg.msg_flags = flags;
656 ret = sock_sendmsg(sock, &kmsg->msg);
657 if (ret < min_ret) {
658 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
659 return -EAGAIN;
660
661 if (ret > 0 && io_net_retry(sock, flags)) {
662 sr->len -= ret;
663 sr->buf += ret;
664 sr->done_io += ret;
665 req->flags |= REQ_F_BL_NO_RECYCLE;
666 return -EAGAIN;
667 }
668 if (ret == -ERESTARTSYS)
669 ret = -EINTR;
670 req_set_fail(req);
671 }
672 if (ret >= 0)
673 ret += sr->done_io;
674 else if (sr->done_io)
675 ret = sr->done_io;
676
677 if (!io_send_finish(req, &ret, kmsg, issue_flags))
678 goto retry_bundle;
679
680 io_req_msg_cleanup(req, issue_flags);
681 return ret;
682}
683
684static int io_recvmsg_mshot_prep(struct io_kiocb *req,
685 struct io_async_msghdr *iomsg,
686 int namelen, size_t controllen)
687{
688 if ((req->flags & (REQ_F_APOLL_MULTISHOT|REQ_F_BUFFER_SELECT)) ==
689 (REQ_F_APOLL_MULTISHOT|REQ_F_BUFFER_SELECT)) {
690 int hdr;
691
692 if (unlikely(namelen < 0))
693 return -EOVERFLOW;
694 if (check_add_overflow(sizeof(struct io_uring_recvmsg_out),
695 namelen, &hdr))
696 return -EOVERFLOW;
697 if (check_add_overflow(hdr, controllen, &hdr))
698 return -EOVERFLOW;
699
700 iomsg->namelen = namelen;
701 iomsg->controllen = controllen;
702 return 0;
703 }
704
705 return 0;
706}
707
708static int io_recvmsg_copy_hdr(struct io_kiocb *req,
709 struct io_async_msghdr *iomsg)
710{
711 struct user_msghdr msg;
712 int ret;
713
714 iomsg->msg.msg_name = &iomsg->addr;
715 iomsg->msg.msg_iter.nr_segs = 0;
716
717#ifdef CONFIG_COMPAT
718 if (unlikely(req->ctx->compat)) {
719 struct compat_msghdr cmsg;
720
721 ret = io_compat_msg_copy_hdr(req, iomsg, &cmsg, ITER_DEST);
722 if (unlikely(ret))
723 return ret;
724
725 ret = __get_compat_msghdr(&iomsg->msg, &cmsg, &iomsg->uaddr);
726 if (unlikely(ret))
727 return ret;
728
729 return io_recvmsg_mshot_prep(req, iomsg, cmsg.msg_namelen,
730 cmsg.msg_controllen);
731 }
732#endif
733
734 ret = io_msg_copy_hdr(req, iomsg, &msg, ITER_DEST);
735 if (unlikely(ret))
736 return ret;
737
738 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
739 if (unlikely(ret))
740 return ret;
741
742 return io_recvmsg_mshot_prep(req, iomsg, msg.msg_namelen,
743 msg.msg_controllen);
744}
745
746static int io_recvmsg_prep_setup(struct io_kiocb *req)
747{
748 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
749 struct io_async_msghdr *kmsg;
750 int ret;
751
752 kmsg = io_msg_alloc_async(req);
753 if (unlikely(!kmsg))
754 return -ENOMEM;
755
756 if (req->opcode == IORING_OP_RECV) {
757 kmsg->msg.msg_name = NULL;
758 kmsg->msg.msg_namelen = 0;
759 kmsg->msg.msg_inq = 0;
760 kmsg->msg.msg_control = NULL;
761 kmsg->msg.msg_get_inq = 1;
762 kmsg->msg.msg_controllen = 0;
763 kmsg->msg.msg_iocb = NULL;
764 kmsg->msg.msg_ubuf = NULL;
765
766 if (!io_do_buffer_select(req)) {
767 ret = import_ubuf(ITER_DEST, sr->buf, sr->len,
768 &kmsg->msg.msg_iter);
769 if (unlikely(ret))
770 return ret;
771 }
772 return 0;
773 }
774
775 ret = io_recvmsg_copy_hdr(req, kmsg);
776 if (!ret)
777 req->flags |= REQ_F_NEED_CLEANUP;
778 return ret;
779}
780
781#define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT | \
782 IORING_RECVSEND_BUNDLE)
783
784int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
785{
786 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
787
788 sr->done_io = 0;
789
790 if (unlikely(sqe->file_index || sqe->addr2))
791 return -EINVAL;
792
793 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
794 sr->len = READ_ONCE(sqe->len);
795 sr->flags = READ_ONCE(sqe->ioprio);
796 if (sr->flags & ~RECVMSG_FLAGS)
797 return -EINVAL;
798 sr->msg_flags = READ_ONCE(sqe->msg_flags);
799 if (sr->msg_flags & MSG_DONTWAIT)
800 req->flags |= REQ_F_NOWAIT;
801 if (sr->msg_flags & MSG_ERRQUEUE)
802 req->flags |= REQ_F_CLEAR_POLLIN;
803 if (req->flags & REQ_F_BUFFER_SELECT) {
804 /*
805 * Store the buffer group for this multishot receive separately,
806 * as if we end up doing an io-wq based issue that selects a
807 * buffer, it has to be committed immediately and that will
808 * clear ->buf_list. This means we lose the link to the buffer
809 * list, and the eventual buffer put on completion then cannot
810 * restore it.
811 */
812 sr->buf_group = req->buf_index;
813 req->buf_list = NULL;
814 }
815 if (sr->flags & IORING_RECV_MULTISHOT) {
816 if (!(req->flags & REQ_F_BUFFER_SELECT))
817 return -EINVAL;
818 if (sr->msg_flags & MSG_WAITALL)
819 return -EINVAL;
820 if (req->opcode == IORING_OP_RECV && sr->len)
821 return -EINVAL;
822 req->flags |= REQ_F_APOLL_MULTISHOT;
823 }
824 if (sr->flags & IORING_RECVSEND_BUNDLE) {
825 if (req->opcode == IORING_OP_RECVMSG)
826 return -EINVAL;
827 }
828
829#ifdef CONFIG_COMPAT
830 if (req->ctx->compat)
831 sr->msg_flags |= MSG_CMSG_COMPAT;
832#endif
833 sr->nr_multishot_loops = 0;
834 return io_recvmsg_prep_setup(req);
835}
836
837/*
838 * Finishes io_recv and io_recvmsg.
839 *
840 * Returns true if it is actually finished, or false if it should run
841 * again (for multishot).
842 */
843static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
844 struct io_async_msghdr *kmsg,
845 bool mshot_finished, unsigned issue_flags)
846{
847 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
848 unsigned int cflags = 0;
849
850 if (kmsg->msg.msg_inq > 0)
851 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
852
853 if (sr->flags & IORING_RECVSEND_BUNDLE) {
854 cflags |= io_put_kbufs(req, *ret, io_bundle_nbufs(kmsg, *ret),
855 issue_flags);
856 /* bundle with no more immediate buffers, we're done */
857 if (req->flags & REQ_F_BL_EMPTY)
858 goto finish;
859 } else {
860 cflags |= io_put_kbuf(req, *ret, issue_flags);
861 }
862
863 /*
864 * Fill CQE for this receive and see if we should keep trying to
865 * receive from this socket.
866 */
867 if ((req->flags & REQ_F_APOLL_MULTISHOT) && !mshot_finished &&
868 io_req_post_cqe(req, *ret, cflags | IORING_CQE_F_MORE)) {
869 int mshot_retry_ret = IOU_ISSUE_SKIP_COMPLETE;
870
871 io_mshot_prep_retry(req, kmsg);
872 /* Known not-empty or unknown state, retry */
873 if (cflags & IORING_CQE_F_SOCK_NONEMPTY || kmsg->msg.msg_inq < 0) {
874 if (sr->nr_multishot_loops++ < MULTISHOT_MAX_RETRY)
875 return false;
876 /* mshot retries exceeded, force a requeue */
877 sr->nr_multishot_loops = 0;
878 mshot_retry_ret = IOU_REQUEUE;
879 }
880 if (issue_flags & IO_URING_F_MULTISHOT)
881 *ret = mshot_retry_ret;
882 else
883 *ret = -EAGAIN;
884 return true;
885 }
886
887 /* Finish the request / stop multishot. */
888finish:
889 io_req_set_res(req, *ret, cflags);
890
891 if (issue_flags & IO_URING_F_MULTISHOT)
892 *ret = IOU_STOP_MULTISHOT;
893 else
894 *ret = IOU_OK;
895 io_req_msg_cleanup(req, issue_flags);
896 return true;
897}
898
899static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
900 struct io_sr_msg *sr, void __user **buf,
901 size_t *len)
902{
903 unsigned long ubuf = (unsigned long) *buf;
904 unsigned long hdr;
905
906 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
907 kmsg->controllen;
908 if (*len < hdr)
909 return -EFAULT;
910
911 if (kmsg->controllen) {
912 unsigned long control = ubuf + hdr - kmsg->controllen;
913
914 kmsg->msg.msg_control_user = (void __user *) control;
915 kmsg->msg.msg_controllen = kmsg->controllen;
916 }
917
918 sr->buf = *buf; /* stash for later copy */
919 *buf = (void __user *) (ubuf + hdr);
920 kmsg->payloadlen = *len = *len - hdr;
921 return 0;
922}
923
924struct io_recvmsg_multishot_hdr {
925 struct io_uring_recvmsg_out msg;
926 struct sockaddr_storage addr;
927};
928
929static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
930 struct io_async_msghdr *kmsg,
931 unsigned int flags, bool *finished)
932{
933 int err;
934 int copy_len;
935 struct io_recvmsg_multishot_hdr hdr;
936
937 if (kmsg->namelen)
938 kmsg->msg.msg_name = &hdr.addr;
939 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
940 kmsg->msg.msg_namelen = 0;
941
942 if (sock->file->f_flags & O_NONBLOCK)
943 flags |= MSG_DONTWAIT;
944
945 err = sock_recvmsg(sock, &kmsg->msg, flags);
946 *finished = err <= 0;
947 if (err < 0)
948 return err;
949
950 hdr.msg = (struct io_uring_recvmsg_out) {
951 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
952 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
953 };
954
955 hdr.msg.payloadlen = err;
956 if (err > kmsg->payloadlen)
957 err = kmsg->payloadlen;
958
959 copy_len = sizeof(struct io_uring_recvmsg_out);
960 if (kmsg->msg.msg_namelen > kmsg->namelen)
961 copy_len += kmsg->namelen;
962 else
963 copy_len += kmsg->msg.msg_namelen;
964
965 /*
966 * "fromlen shall refer to the value before truncation.."
967 * 1003.1g
968 */
969 hdr.msg.namelen = kmsg->msg.msg_namelen;
970
971 /* ensure that there is no gap between hdr and sockaddr_storage */
972 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
973 sizeof(struct io_uring_recvmsg_out));
974 if (copy_to_user(io->buf, &hdr, copy_len)) {
975 *finished = true;
976 return -EFAULT;
977 }
978
979 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
980 kmsg->controllen + err;
981}
982
983int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
984{
985 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
986 struct io_async_msghdr *kmsg = req->async_data;
987 struct socket *sock;
988 unsigned flags;
989 int ret, min_ret = 0;
990 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
991 bool mshot_finished = true;
992
993 sock = sock_from_file(req->file);
994 if (unlikely(!sock))
995 return -ENOTSOCK;
996
997 if (!(req->flags & REQ_F_POLLED) &&
998 (sr->flags & IORING_RECVSEND_POLL_FIRST))
999 return -EAGAIN;
1000
1001 flags = sr->msg_flags;
1002 if (force_nonblock)
1003 flags |= MSG_DONTWAIT;
1004
1005retry_multishot:
1006 if (io_do_buffer_select(req)) {
1007 void __user *buf;
1008 size_t len = sr->len;
1009
1010 buf = io_buffer_select(req, &len, issue_flags);
1011 if (!buf)
1012 return -ENOBUFS;
1013
1014 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1015 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
1016 if (ret) {
1017 io_kbuf_recycle(req, issue_flags);
1018 return ret;
1019 }
1020 }
1021
1022 iov_iter_ubuf(&kmsg->msg.msg_iter, ITER_DEST, buf, len);
1023 }
1024
1025 kmsg->msg.msg_get_inq = 1;
1026 kmsg->msg.msg_inq = -1;
1027 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1028 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
1029 &mshot_finished);
1030 } else {
1031 /* disable partial retry for recvmsg with cmsg attached */
1032 if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
1033 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1034
1035 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
1036 kmsg->uaddr, flags);
1037 }
1038
1039 if (ret < min_ret) {
1040 if (ret == -EAGAIN && force_nonblock) {
1041 if (issue_flags & IO_URING_F_MULTISHOT) {
1042 io_kbuf_recycle(req, issue_flags);
1043 return IOU_ISSUE_SKIP_COMPLETE;
1044 }
1045 return -EAGAIN;
1046 }
1047 if (ret > 0 && io_net_retry(sock, flags)) {
1048 sr->done_io += ret;
1049 req->flags |= REQ_F_BL_NO_RECYCLE;
1050 return -EAGAIN;
1051 }
1052 if (ret == -ERESTARTSYS)
1053 ret = -EINTR;
1054 req_set_fail(req);
1055 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
1056 req_set_fail(req);
1057 }
1058
1059 if (ret > 0)
1060 ret += sr->done_io;
1061 else if (sr->done_io)
1062 ret = sr->done_io;
1063 else
1064 io_kbuf_recycle(req, issue_flags);
1065
1066 if (!io_recv_finish(req, &ret, kmsg, mshot_finished, issue_flags))
1067 goto retry_multishot;
1068
1069 return ret;
1070}
1071
1072static int io_recv_buf_select(struct io_kiocb *req, struct io_async_msghdr *kmsg,
1073 size_t *len, unsigned int issue_flags)
1074{
1075 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1076 int ret;
1077
1078 /*
1079 * If the ring isn't locked, then don't use the peek interface
1080 * to grab multiple buffers as we will lock/unlock between
1081 * this selection and posting the buffers.
1082 */
1083 if (!(issue_flags & IO_URING_F_UNLOCKED) &&
1084 sr->flags & IORING_RECVSEND_BUNDLE) {
1085 struct buf_sel_arg arg = {
1086 .iovs = &kmsg->fast_iov,
1087 .nr_iovs = 1,
1088 .mode = KBUF_MODE_EXPAND,
1089 };
1090
1091 if (kmsg->free_iov) {
1092 arg.nr_iovs = kmsg->free_iov_nr;
1093 arg.iovs = kmsg->free_iov;
1094 arg.mode |= KBUF_MODE_FREE;
1095 }
1096
1097 if (kmsg->msg.msg_inq > 0)
1098 arg.max_len = min_not_zero(sr->len, kmsg->msg.msg_inq);
1099
1100 ret = io_buffers_peek(req, &arg);
1101 if (unlikely(ret < 0))
1102 return ret;
1103
1104 /* special case 1 vec, can be a fast path */
1105 if (ret == 1) {
1106 sr->buf = arg.iovs[0].iov_base;
1107 sr->len = arg.iovs[0].iov_len;
1108 goto map_ubuf;
1109 }
1110 iov_iter_init(&kmsg->msg.msg_iter, ITER_DEST, arg.iovs, ret,
1111 arg.out_len);
1112 if (arg.iovs != &kmsg->fast_iov && arg.iovs != kmsg->free_iov) {
1113 kmsg->free_iov_nr = ret;
1114 kmsg->free_iov = arg.iovs;
1115 req->flags |= REQ_F_NEED_CLEANUP;
1116 }
1117 } else {
1118 void __user *buf;
1119
1120 *len = sr->len;
1121 buf = io_buffer_select(req, len, issue_flags);
1122 if (!buf)
1123 return -ENOBUFS;
1124 sr->buf = buf;
1125 sr->len = *len;
1126map_ubuf:
1127 ret = import_ubuf(ITER_DEST, sr->buf, sr->len,
1128 &kmsg->msg.msg_iter);
1129 if (unlikely(ret))
1130 return ret;
1131 }
1132
1133 return 0;
1134}
1135
1136int io_recv(struct io_kiocb *req, unsigned int issue_flags)
1137{
1138 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1139 struct io_async_msghdr *kmsg = req->async_data;
1140 struct socket *sock;
1141 unsigned flags;
1142 int ret, min_ret = 0;
1143 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1144 size_t len = sr->len;
1145 bool mshot_finished;
1146
1147 if (!(req->flags & REQ_F_POLLED) &&
1148 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1149 return -EAGAIN;
1150
1151 sock = sock_from_file(req->file);
1152 if (unlikely(!sock))
1153 return -ENOTSOCK;
1154
1155 flags = sr->msg_flags;
1156 if (force_nonblock)
1157 flags |= MSG_DONTWAIT;
1158
1159retry_multishot:
1160 if (io_do_buffer_select(req)) {
1161 ret = io_recv_buf_select(req, kmsg, &len, issue_flags);
1162 if (unlikely(ret)) {
1163 kmsg->msg.msg_inq = -1;
1164 goto out_free;
1165 }
1166 sr->buf = NULL;
1167 }
1168
1169 kmsg->msg.msg_flags = 0;
1170 kmsg->msg.msg_inq = -1;
1171
1172 if (flags & MSG_WAITALL)
1173 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1174
1175 ret = sock_recvmsg(sock, &kmsg->msg, flags);
1176 if (ret < min_ret) {
1177 if (ret == -EAGAIN && force_nonblock) {
1178 if (issue_flags & IO_URING_F_MULTISHOT) {
1179 io_kbuf_recycle(req, issue_flags);
1180 return IOU_ISSUE_SKIP_COMPLETE;
1181 }
1182
1183 return -EAGAIN;
1184 }
1185 if (ret > 0 && io_net_retry(sock, flags)) {
1186 sr->len -= ret;
1187 sr->buf += ret;
1188 sr->done_io += ret;
1189 req->flags |= REQ_F_BL_NO_RECYCLE;
1190 return -EAGAIN;
1191 }
1192 if (ret == -ERESTARTSYS)
1193 ret = -EINTR;
1194 req_set_fail(req);
1195 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
1196out_free:
1197 req_set_fail(req);
1198 }
1199
1200 mshot_finished = ret <= 0;
1201 if (ret > 0)
1202 ret += sr->done_io;
1203 else if (sr->done_io)
1204 ret = sr->done_io;
1205 else
1206 io_kbuf_recycle(req, issue_flags);
1207
1208 if (!io_recv_finish(req, &ret, kmsg, mshot_finished, issue_flags))
1209 goto retry_multishot;
1210
1211 return ret;
1212}
1213
1214void io_send_zc_cleanup(struct io_kiocb *req)
1215{
1216 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1217 struct io_async_msghdr *io = req->async_data;
1218
1219 if (req_has_async_data(req))
1220 io_netmsg_iovec_free(io);
1221 if (zc->notif) {
1222 io_notif_flush(zc->notif);
1223 zc->notif = NULL;
1224 }
1225}
1226
1227#define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
1228#define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
1229
1230int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1231{
1232 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1233 struct io_ring_ctx *ctx = req->ctx;
1234 struct io_kiocb *notif;
1235
1236 zc->done_io = 0;
1237 req->flags |= REQ_F_POLL_NO_LAZY;
1238
1239 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
1240 return -EINVAL;
1241 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
1242 if (req->flags & REQ_F_CQE_SKIP)
1243 return -EINVAL;
1244
1245 notif = zc->notif = io_alloc_notif(ctx);
1246 if (!notif)
1247 return -ENOMEM;
1248 notif->cqe.user_data = req->cqe.user_data;
1249 notif->cqe.res = 0;
1250 notif->cqe.flags = IORING_CQE_F_NOTIF;
1251 req->flags |= REQ_F_NEED_CLEANUP;
1252
1253 zc->flags = READ_ONCE(sqe->ioprio);
1254 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
1255 if (zc->flags & ~IO_ZC_FLAGS_VALID)
1256 return -EINVAL;
1257 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
1258 struct io_notif_data *nd = io_notif_to_data(notif);
1259
1260 nd->zc_report = true;
1261 nd->zc_used = false;
1262 nd->zc_copied = false;
1263 }
1264 }
1265
1266 if (req->opcode != IORING_OP_SEND_ZC) {
1267 if (unlikely(sqe->addr2 || sqe->file_index))
1268 return -EINVAL;
1269 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1270 return -EINVAL;
1271 }
1272
1273 zc->len = READ_ONCE(sqe->len);
1274 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL | MSG_ZEROCOPY;
1275 zc->buf_index = READ_ONCE(sqe->buf_index);
1276 if (zc->msg_flags & MSG_DONTWAIT)
1277 req->flags |= REQ_F_NOWAIT;
1278
1279#ifdef CONFIG_COMPAT
1280 if (req->ctx->compat)
1281 zc->msg_flags |= MSG_CMSG_COMPAT;
1282#endif
1283 if (unlikely(!io_msg_alloc_async(req)))
1284 return -ENOMEM;
1285 if (req->opcode != IORING_OP_SENDMSG_ZC)
1286 return io_send_setup(req, sqe);
1287 return io_sendmsg_setup(req, sqe);
1288}
1289
1290static int io_sg_from_iter_iovec(struct sk_buff *skb,
1291 struct iov_iter *from, size_t length)
1292{
1293 skb_zcopy_downgrade_managed(skb);
1294 return zerocopy_fill_skb_from_iter(skb, from, length);
1295}
1296
1297static int io_sg_from_iter(struct sk_buff *skb,
1298 struct iov_iter *from, size_t length)
1299{
1300 struct skb_shared_info *shinfo = skb_shinfo(skb);
1301 int frag = shinfo->nr_frags;
1302 int ret = 0;
1303 struct bvec_iter bi;
1304 ssize_t copied = 0;
1305 unsigned long truesize = 0;
1306
1307 if (!frag)
1308 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1309 else if (unlikely(!skb_zcopy_managed(skb)))
1310 return zerocopy_fill_skb_from_iter(skb, from, length);
1311
1312 bi.bi_size = min(from->count, length);
1313 bi.bi_bvec_done = from->iov_offset;
1314 bi.bi_idx = 0;
1315
1316 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1317 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1318
1319 copied += v.bv_len;
1320 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1321 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1322 v.bv_offset, v.bv_len);
1323 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1324 }
1325 if (bi.bi_size)
1326 ret = -EMSGSIZE;
1327
1328 shinfo->nr_frags = frag;
1329 from->bvec += bi.bi_idx;
1330 from->nr_segs -= bi.bi_idx;
1331 from->count -= copied;
1332 from->iov_offset = bi.bi_bvec_done;
1333
1334 skb->data_len += copied;
1335 skb->len += copied;
1336 skb->truesize += truesize;
1337 return ret;
1338}
1339
1340static int io_send_zc_import(struct io_kiocb *req, unsigned int issue_flags)
1341{
1342 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1343 struct io_async_msghdr *kmsg = req->async_data;
1344 int ret;
1345
1346 if (sr->flags & IORING_RECVSEND_FIXED_BUF) {
1347 struct io_ring_ctx *ctx = req->ctx;
1348 struct io_rsrc_node *node;
1349
1350 ret = -EFAULT;
1351 io_ring_submit_lock(ctx, issue_flags);
1352 node = io_rsrc_node_lookup(&ctx->buf_table, sr->buf_index);
1353 if (node) {
1354 io_req_assign_buf_node(sr->notif, node);
1355 ret = 0;
1356 }
1357 io_ring_submit_unlock(ctx, issue_flags);
1358
1359 if (unlikely(ret))
1360 return ret;
1361
1362 ret = io_import_fixed(ITER_SOURCE, &kmsg->msg.msg_iter,
1363 node->buf, (u64)(uintptr_t)sr->buf,
1364 sr->len);
1365 if (unlikely(ret))
1366 return ret;
1367 kmsg->msg.sg_from_iter = io_sg_from_iter;
1368 } else {
1369 ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len, &kmsg->msg.msg_iter);
1370 if (unlikely(ret))
1371 return ret;
1372 ret = io_notif_account_mem(sr->notif, sr->len);
1373 if (unlikely(ret))
1374 return ret;
1375 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1376 }
1377
1378 return ret;
1379}
1380
1381int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1382{
1383 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1384 struct io_async_msghdr *kmsg = req->async_data;
1385 struct socket *sock;
1386 unsigned msg_flags;
1387 int ret, min_ret = 0;
1388
1389 sock = sock_from_file(req->file);
1390 if (unlikely(!sock))
1391 return -ENOTSOCK;
1392 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1393 return -EOPNOTSUPP;
1394
1395 if (!(req->flags & REQ_F_POLLED) &&
1396 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1397 return -EAGAIN;
1398
1399 if (!zc->done_io) {
1400 ret = io_send_zc_import(req, issue_flags);
1401 if (unlikely(ret))
1402 return ret;
1403 }
1404
1405 msg_flags = zc->msg_flags;
1406 if (issue_flags & IO_URING_F_NONBLOCK)
1407 msg_flags |= MSG_DONTWAIT;
1408 if (msg_flags & MSG_WAITALL)
1409 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1410 msg_flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
1411
1412 kmsg->msg.msg_flags = msg_flags;
1413 kmsg->msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1414 ret = sock_sendmsg(sock, &kmsg->msg);
1415
1416 if (unlikely(ret < min_ret)) {
1417 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1418 return -EAGAIN;
1419
1420 if (ret > 0 && io_net_retry(sock, kmsg->msg.msg_flags)) {
1421 zc->len -= ret;
1422 zc->buf += ret;
1423 zc->done_io += ret;
1424 req->flags |= REQ_F_BL_NO_RECYCLE;
1425 return -EAGAIN;
1426 }
1427 if (ret == -ERESTARTSYS)
1428 ret = -EINTR;
1429 req_set_fail(req);
1430 }
1431
1432 if (ret >= 0)
1433 ret += zc->done_io;
1434 else if (zc->done_io)
1435 ret = zc->done_io;
1436
1437 /*
1438 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1439 * flushing notif to io_send_zc_cleanup()
1440 */
1441 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1442 io_notif_flush(zc->notif);
1443 io_req_msg_cleanup(req, 0);
1444 }
1445 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1446 return IOU_OK;
1447}
1448
1449int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1450{
1451 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1452 struct io_async_msghdr *kmsg = req->async_data;
1453 struct socket *sock;
1454 unsigned flags;
1455 int ret, min_ret = 0;
1456
1457 sock = sock_from_file(req->file);
1458 if (unlikely(!sock))
1459 return -ENOTSOCK;
1460 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1461 return -EOPNOTSUPP;
1462
1463 if (!(req->flags & REQ_F_POLLED) &&
1464 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1465 return -EAGAIN;
1466
1467 flags = sr->msg_flags;
1468 if (issue_flags & IO_URING_F_NONBLOCK)
1469 flags |= MSG_DONTWAIT;
1470 if (flags & MSG_WAITALL)
1471 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1472
1473 kmsg->msg.msg_control_user = sr->msg_control;
1474 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1475 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1476 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1477
1478 if (unlikely(ret < min_ret)) {
1479 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1480 return -EAGAIN;
1481
1482 if (ret > 0 && io_net_retry(sock, flags)) {
1483 sr->done_io += ret;
1484 req->flags |= REQ_F_BL_NO_RECYCLE;
1485 return -EAGAIN;
1486 }
1487 if (ret == -ERESTARTSYS)
1488 ret = -EINTR;
1489 req_set_fail(req);
1490 }
1491
1492 if (ret >= 0)
1493 ret += sr->done_io;
1494 else if (sr->done_io)
1495 ret = sr->done_io;
1496
1497 /*
1498 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1499 * flushing notif to io_send_zc_cleanup()
1500 */
1501 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1502 io_notif_flush(sr->notif);
1503 io_req_msg_cleanup(req, 0);
1504 }
1505 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1506 return IOU_OK;
1507}
1508
1509void io_sendrecv_fail(struct io_kiocb *req)
1510{
1511 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1512
1513 if (sr->done_io)
1514 req->cqe.res = sr->done_io;
1515
1516 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1517 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1518 req->cqe.flags |= IORING_CQE_F_MORE;
1519}
1520
1521#define ACCEPT_FLAGS (IORING_ACCEPT_MULTISHOT | IORING_ACCEPT_DONTWAIT | \
1522 IORING_ACCEPT_POLL_FIRST)
1523
1524int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1525{
1526 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1527
1528 if (sqe->len || sqe->buf_index)
1529 return -EINVAL;
1530
1531 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1532 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1533 accept->flags = READ_ONCE(sqe->accept_flags);
1534 accept->nofile = rlimit(RLIMIT_NOFILE);
1535 accept->iou_flags = READ_ONCE(sqe->ioprio);
1536 if (accept->iou_flags & ~ACCEPT_FLAGS)
1537 return -EINVAL;
1538
1539 accept->file_slot = READ_ONCE(sqe->file_index);
1540 if (accept->file_slot) {
1541 if (accept->flags & SOCK_CLOEXEC)
1542 return -EINVAL;
1543 if (accept->iou_flags & IORING_ACCEPT_MULTISHOT &&
1544 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1545 return -EINVAL;
1546 }
1547 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1548 return -EINVAL;
1549 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1550 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1551 if (accept->iou_flags & IORING_ACCEPT_MULTISHOT)
1552 req->flags |= REQ_F_APOLL_MULTISHOT;
1553 if (accept->iou_flags & IORING_ACCEPT_DONTWAIT)
1554 req->flags |= REQ_F_NOWAIT;
1555 return 0;
1556}
1557
1558int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1559{
1560 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1561 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1562 bool fixed = !!accept->file_slot;
1563 struct proto_accept_arg arg = {
1564 .flags = force_nonblock ? O_NONBLOCK : 0,
1565 };
1566 struct file *file;
1567 unsigned cflags;
1568 int ret, fd;
1569
1570 if (!(req->flags & REQ_F_POLLED) &&
1571 accept->iou_flags & IORING_ACCEPT_POLL_FIRST)
1572 return -EAGAIN;
1573
1574retry:
1575 if (!fixed) {
1576 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1577 if (unlikely(fd < 0))
1578 return fd;
1579 }
1580 arg.err = 0;
1581 arg.is_empty = -1;
1582 file = do_accept(req->file, &arg, accept->addr, accept->addr_len,
1583 accept->flags);
1584 if (IS_ERR(file)) {
1585 if (!fixed)
1586 put_unused_fd(fd);
1587 ret = PTR_ERR(file);
1588 if (ret == -EAGAIN && force_nonblock &&
1589 !(accept->iou_flags & IORING_ACCEPT_DONTWAIT)) {
1590 /*
1591 * if it's multishot and polled, we don't need to
1592 * return EAGAIN to arm the poll infra since it
1593 * has already been done
1594 */
1595 if (issue_flags & IO_URING_F_MULTISHOT)
1596 return IOU_ISSUE_SKIP_COMPLETE;
1597 return ret;
1598 }
1599 if (ret == -ERESTARTSYS)
1600 ret = -EINTR;
1601 req_set_fail(req);
1602 } else if (!fixed) {
1603 fd_install(fd, file);
1604 ret = fd;
1605 } else {
1606 ret = io_fixed_fd_install(req, issue_flags, file,
1607 accept->file_slot);
1608 }
1609
1610 cflags = 0;
1611 if (!arg.is_empty)
1612 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
1613
1614 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1615 io_req_set_res(req, ret, cflags);
1616 return IOU_OK;
1617 }
1618
1619 if (ret < 0)
1620 return ret;
1621 if (io_req_post_cqe(req, ret, cflags | IORING_CQE_F_MORE)) {
1622 if (cflags & IORING_CQE_F_SOCK_NONEMPTY || arg.is_empty == -1)
1623 goto retry;
1624 if (issue_flags & IO_URING_F_MULTISHOT)
1625 return IOU_ISSUE_SKIP_COMPLETE;
1626 return -EAGAIN;
1627 }
1628
1629 io_req_set_res(req, ret, cflags);
1630 return IOU_STOP_MULTISHOT;
1631}
1632
1633int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1634{
1635 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1636
1637 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1638 return -EINVAL;
1639
1640 sock->domain = READ_ONCE(sqe->fd);
1641 sock->type = READ_ONCE(sqe->off);
1642 sock->protocol = READ_ONCE(sqe->len);
1643 sock->file_slot = READ_ONCE(sqe->file_index);
1644 sock->nofile = rlimit(RLIMIT_NOFILE);
1645
1646 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1647 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1648 return -EINVAL;
1649 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1650 return -EINVAL;
1651 return 0;
1652}
1653
1654int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1655{
1656 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1657 bool fixed = !!sock->file_slot;
1658 struct file *file;
1659 int ret, fd;
1660
1661 if (!fixed) {
1662 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1663 if (unlikely(fd < 0))
1664 return fd;
1665 }
1666 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1667 if (IS_ERR(file)) {
1668 if (!fixed)
1669 put_unused_fd(fd);
1670 ret = PTR_ERR(file);
1671 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1672 return -EAGAIN;
1673 if (ret == -ERESTARTSYS)
1674 ret = -EINTR;
1675 req_set_fail(req);
1676 } else if (!fixed) {
1677 fd_install(fd, file);
1678 ret = fd;
1679 } else {
1680 ret = io_fixed_fd_install(req, issue_flags, file,
1681 sock->file_slot);
1682 }
1683 io_req_set_res(req, ret, 0);
1684 return IOU_OK;
1685}
1686
1687int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1688{
1689 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1690 struct io_async_msghdr *io;
1691
1692 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1693 return -EINVAL;
1694
1695 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1696 conn->addr_len = READ_ONCE(sqe->addr2);
1697 conn->in_progress = conn->seen_econnaborted = false;
1698
1699 io = io_msg_alloc_async(req);
1700 if (unlikely(!io))
1701 return -ENOMEM;
1702
1703 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->addr);
1704}
1705
1706int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1707{
1708 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1709 struct io_async_msghdr *io = req->async_data;
1710 unsigned file_flags;
1711 int ret;
1712 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1713
1714 if (unlikely(req->flags & REQ_F_FAIL)) {
1715 ret = -ECONNRESET;
1716 goto out;
1717 }
1718
1719 file_flags = force_nonblock ? O_NONBLOCK : 0;
1720
1721 ret = __sys_connect_file(req->file, &io->addr, connect->addr_len,
1722 file_flags);
1723 if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1724 && force_nonblock) {
1725 if (ret == -EINPROGRESS) {
1726 connect->in_progress = true;
1727 } else if (ret == -ECONNABORTED) {
1728 if (connect->seen_econnaborted)
1729 goto out;
1730 connect->seen_econnaborted = true;
1731 }
1732 return -EAGAIN;
1733 }
1734 if (connect->in_progress) {
1735 /*
1736 * At least bluetooth will return -EBADFD on a re-connect
1737 * attempt, and it's (supposedly) also valid to get -EISCONN
1738 * which means the previous result is good. For both of these,
1739 * grab the sock_error() and use that for the completion.
1740 */
1741 if (ret == -EBADFD || ret == -EISCONN)
1742 ret = sock_error(sock_from_file(req->file)->sk);
1743 }
1744 if (ret == -ERESTARTSYS)
1745 ret = -EINTR;
1746out:
1747 if (ret < 0)
1748 req_set_fail(req);
1749 io_req_msg_cleanup(req, issue_flags);
1750 io_req_set_res(req, ret, 0);
1751 return IOU_OK;
1752}
1753
1754int io_bind_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1755{
1756 struct io_bind *bind = io_kiocb_to_cmd(req, struct io_bind);
1757 struct sockaddr __user *uaddr;
1758 struct io_async_msghdr *io;
1759
1760 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1761 return -EINVAL;
1762
1763 uaddr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1764 bind->addr_len = READ_ONCE(sqe->addr2);
1765
1766 io = io_msg_alloc_async(req);
1767 if (unlikely(!io))
1768 return -ENOMEM;
1769 return move_addr_to_kernel(uaddr, bind->addr_len, &io->addr);
1770}
1771
1772int io_bind(struct io_kiocb *req, unsigned int issue_flags)
1773{
1774 struct io_bind *bind = io_kiocb_to_cmd(req, struct io_bind);
1775 struct io_async_msghdr *io = req->async_data;
1776 struct socket *sock;
1777 int ret;
1778
1779 sock = sock_from_file(req->file);
1780 if (unlikely(!sock))
1781 return -ENOTSOCK;
1782
1783 ret = __sys_bind_socket(sock, &io->addr, bind->addr_len);
1784 if (ret < 0)
1785 req_set_fail(req);
1786 io_req_set_res(req, ret, 0);
1787 return 0;
1788}
1789
1790int io_listen_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1791{
1792 struct io_listen *listen = io_kiocb_to_cmd(req, struct io_listen);
1793
1794 if (sqe->addr || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in || sqe->addr2)
1795 return -EINVAL;
1796
1797 listen->backlog = READ_ONCE(sqe->len);
1798 return 0;
1799}
1800
1801int io_listen(struct io_kiocb *req, unsigned int issue_flags)
1802{
1803 struct io_listen *listen = io_kiocb_to_cmd(req, struct io_listen);
1804 struct socket *sock;
1805 int ret;
1806
1807 sock = sock_from_file(req->file);
1808 if (unlikely(!sock))
1809 return -ENOTSOCK;
1810
1811 ret = __sys_listen_socket(sock, listen->backlog);
1812 if (ret < 0)
1813 req_set_fail(req);
1814 io_req_set_res(req, ret, 0);
1815 return 0;
1816}
1817
1818void io_netmsg_cache_free(const void *entry)
1819{
1820 struct io_async_msghdr *kmsg = (struct io_async_msghdr *) entry;
1821
1822 if (kmsg->free_iov) {
1823 kasan_mempool_unpoison_object(kmsg->free_iov,
1824 kmsg->free_iov_nr * sizeof(struct iovec));
1825 io_netmsg_iovec_free(kmsg);
1826 }
1827 kfree(kmsg);
1828}
1829#endif