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>
7 #include <linux/compat.h>
8 #include <net/compat.h>
9 #include <linux/io_uring.h>
11 #include <uapi/linux/io_uring.h>
15 #include "alloc_cache.h"
20 #if defined(CONFIG_NET)
28 struct sockaddr __user *addr;
47 struct sockaddr __user *addr;
50 bool seen_econnaborted;
56 struct compat_msghdr __user *umsg_compat;
57 struct user_msghdr __user *umsg;
64 /* initialised and used only by !msg send variants */
68 void __user *msg_control;
69 /* used only for send zerocopy */
70 struct io_kiocb *notif;
73 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
75 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
77 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
78 sqe->buf_index || sqe->splice_fd_in))
81 shutdown->how = READ_ONCE(sqe->len);
85 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
87 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
91 if (issue_flags & IO_URING_F_NONBLOCK)
94 sock = sock_from_file(req->file);
98 ret = __sys_shutdown_sock(sock, shutdown->how);
99 io_req_set_res(req, ret, 0);
103 static bool io_net_retry(struct socket *sock, int flags)
105 if (!(flags & MSG_WAITALL))
107 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
110 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
112 struct io_async_msghdr *hdr = req->async_data;
114 if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
117 /* Let normal cleanup path reap it if we fail adding to the cache */
118 if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
119 req->async_data = NULL;
120 req->flags &= ~REQ_F_ASYNC_DATA;
124 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
125 unsigned int issue_flags)
127 struct io_ring_ctx *ctx = req->ctx;
128 struct io_cache_entry *entry;
129 struct io_async_msghdr *hdr;
131 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
132 entry = io_alloc_cache_get(&ctx->netmsg_cache);
134 hdr = container_of(entry, struct io_async_msghdr, cache);
135 hdr->free_iov = NULL;
136 req->flags |= REQ_F_ASYNC_DATA;
137 req->async_data = hdr;
142 if (!io_alloc_async_data(req)) {
143 hdr = req->async_data;
144 hdr->free_iov = NULL;
150 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
152 /* ->prep_async is always called from the submission context */
153 return io_msg_alloc_async(req, 0);
156 static int io_setup_async_msg(struct io_kiocb *req,
157 struct io_async_msghdr *kmsg,
158 unsigned int issue_flags)
160 struct io_async_msghdr *async_msg;
162 if (req_has_async_data(req))
164 async_msg = io_msg_alloc_async(req, issue_flags);
166 kfree(kmsg->free_iov);
169 req->flags |= REQ_F_NEED_CLEANUP;
170 memcpy(async_msg, kmsg, sizeof(*kmsg));
171 if (async_msg->msg.msg_name)
172 async_msg->msg.msg_name = &async_msg->addr;
173 /* if were using fast_iov, set it to the new one */
174 if (!kmsg->free_iov) {
175 size_t fast_idx = kmsg->msg.msg_iter.iov - kmsg->fast_iov;
176 async_msg->msg.msg_iter.iov = &async_msg->fast_iov[fast_idx];
182 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
183 struct io_async_msghdr *iomsg)
185 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
188 iomsg->msg.msg_name = &iomsg->addr;
189 iomsg->free_iov = iomsg->fast_iov;
190 ret = sendmsg_copy_msghdr(&iomsg->msg, sr->umsg, sr->msg_flags,
192 /* save msg_control as sys_sendmsg() overwrites it */
193 sr->msg_control = iomsg->msg.msg_control_user;
197 int io_send_prep_async(struct io_kiocb *req)
199 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
200 struct io_async_msghdr *io;
203 if (!zc->addr || req_has_async_data(req))
205 io = io_msg_alloc_async_prep(req);
208 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
212 static int io_setup_async_addr(struct io_kiocb *req,
213 struct sockaddr_storage *addr_storage,
214 unsigned int issue_flags)
216 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
217 struct io_async_msghdr *io;
219 if (!sr->addr || req_has_async_data(req))
221 io = io_msg_alloc_async(req, issue_flags);
224 memcpy(&io->addr, addr_storage, sizeof(io->addr));
228 int io_sendmsg_prep_async(struct io_kiocb *req)
232 if (!io_msg_alloc_async_prep(req))
234 ret = io_sendmsg_copy_hdr(req, req->async_data);
236 req->flags |= REQ_F_NEED_CLEANUP;
240 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
242 struct io_async_msghdr *io = req->async_data;
247 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
249 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
251 if (req->opcode == IORING_OP_SEND) {
252 if (READ_ONCE(sqe->__pad3[0]))
254 sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
255 sr->addr_len = READ_ONCE(sqe->addr_len);
256 } else if (sqe->addr2 || sqe->file_index) {
260 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
261 sr->len = READ_ONCE(sqe->len);
262 sr->flags = READ_ONCE(sqe->ioprio);
263 if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
265 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
266 if (sr->msg_flags & MSG_DONTWAIT)
267 req->flags |= REQ_F_NOWAIT;
270 if (req->ctx->compat)
271 sr->msg_flags |= MSG_CMSG_COMPAT;
277 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
279 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
280 struct io_async_msghdr iomsg, *kmsg;
286 sock = sock_from_file(req->file);
290 if (req_has_async_data(req)) {
291 kmsg = req->async_data;
292 kmsg->msg.msg_control_user = sr->msg_control;
294 ret = io_sendmsg_copy_hdr(req, &iomsg);
300 if (!(req->flags & REQ_F_POLLED) &&
301 (sr->flags & IORING_RECVSEND_POLL_FIRST))
302 return io_setup_async_msg(req, kmsg, issue_flags);
304 flags = sr->msg_flags;
305 if (issue_flags & IO_URING_F_NONBLOCK)
306 flags |= MSG_DONTWAIT;
307 if (flags & MSG_WAITALL)
308 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
310 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
313 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
314 return io_setup_async_msg(req, kmsg, issue_flags);
315 if (ret > 0 && io_net_retry(sock, flags)) {
316 kmsg->msg.msg_controllen = 0;
317 kmsg->msg.msg_control = NULL;
319 req->flags |= REQ_F_PARTIAL_IO;
320 return io_setup_async_msg(req, kmsg, issue_flags);
322 if (ret == -ERESTARTSYS)
326 /* fast path, check for non-NULL to avoid function call */
328 kfree(kmsg->free_iov);
329 req->flags &= ~REQ_F_NEED_CLEANUP;
330 io_netmsg_recycle(req, issue_flags);
333 else if (sr->done_io)
335 io_req_set_res(req, ret, 0);
339 int io_send(struct io_kiocb *req, unsigned int issue_flags)
341 struct sockaddr_storage __address;
342 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
351 msg.msg_control = NULL;
352 msg.msg_controllen = 0;
357 if (req_has_async_data(req)) {
358 struct io_async_msghdr *io = req->async_data;
360 msg.msg_name = &io->addr;
362 ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
363 if (unlikely(ret < 0))
365 msg.msg_name = (struct sockaddr *)&__address;
367 msg.msg_namelen = sr->addr_len;
370 if (!(req->flags & REQ_F_POLLED) &&
371 (sr->flags & IORING_RECVSEND_POLL_FIRST))
372 return io_setup_async_addr(req, &__address, issue_flags);
374 sock = sock_from_file(req->file);
378 ret = import_single_range(ITER_SOURCE, sr->buf, sr->len, &iov, &msg.msg_iter);
382 flags = sr->msg_flags;
383 if (issue_flags & IO_URING_F_NONBLOCK)
384 flags |= MSG_DONTWAIT;
385 if (flags & MSG_WAITALL)
386 min_ret = iov_iter_count(&msg.msg_iter);
388 msg.msg_flags = flags;
389 ret = sock_sendmsg(sock, &msg);
391 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
392 return io_setup_async_addr(req, &__address, issue_flags);
394 if (ret > 0 && io_net_retry(sock, flags)) {
398 req->flags |= REQ_F_PARTIAL_IO;
399 return io_setup_async_addr(req, &__address, issue_flags);
401 if (ret == -ERESTARTSYS)
407 else if (sr->done_io)
409 io_req_set_res(req, ret, 0);
413 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
417 if (iomsg->namelen < 0)
419 if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
420 iomsg->namelen, &hdr))
422 if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
428 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
429 struct io_async_msghdr *iomsg)
431 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
432 struct user_msghdr msg;
435 if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
438 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
442 if (req->flags & REQ_F_BUFFER_SELECT) {
443 if (msg.msg_iovlen == 0) {
444 sr->len = iomsg->fast_iov[0].iov_len = 0;
445 iomsg->fast_iov[0].iov_base = NULL;
446 iomsg->free_iov = NULL;
447 } else if (msg.msg_iovlen > 1) {
450 if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
452 sr->len = iomsg->fast_iov[0].iov_len;
453 iomsg->free_iov = NULL;
456 if (req->flags & REQ_F_APOLL_MULTISHOT) {
457 iomsg->namelen = msg.msg_namelen;
458 iomsg->controllen = msg.msg_controllen;
459 if (io_recvmsg_multishot_overflow(iomsg))
463 iomsg->free_iov = iomsg->fast_iov;
464 ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
465 &iomsg->free_iov, &iomsg->msg.msg_iter,
475 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
476 struct io_async_msghdr *iomsg)
478 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
479 struct compat_msghdr msg;
480 struct compat_iovec __user *uiov;
483 if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
486 ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
490 uiov = compat_ptr(msg.msg_iov);
491 if (req->flags & REQ_F_BUFFER_SELECT) {
494 iomsg->free_iov = NULL;
495 if (msg.msg_iovlen == 0) {
497 } else if (msg.msg_iovlen > 1) {
500 if (!access_ok(uiov, sizeof(*uiov)))
502 if (__get_user(clen, &uiov->iov_len))
509 if (req->flags & REQ_F_APOLL_MULTISHOT) {
510 iomsg->namelen = msg.msg_namelen;
511 iomsg->controllen = msg.msg_controllen;
512 if (io_recvmsg_multishot_overflow(iomsg))
516 iomsg->free_iov = iomsg->fast_iov;
517 ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
518 UIO_FASTIOV, &iomsg->free_iov,
519 &iomsg->msg.msg_iter, true);
528 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
529 struct io_async_msghdr *iomsg)
531 iomsg->msg.msg_name = &iomsg->addr;
534 if (req->ctx->compat)
535 return __io_compat_recvmsg_copy_hdr(req, iomsg);
538 return __io_recvmsg_copy_hdr(req, iomsg);
541 int io_recvmsg_prep_async(struct io_kiocb *req)
545 if (!io_msg_alloc_async_prep(req))
547 ret = io_recvmsg_copy_hdr(req, req->async_data);
549 req->flags |= REQ_F_NEED_CLEANUP;
553 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
555 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
557 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
559 if (unlikely(sqe->file_index || sqe->addr2))
562 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
563 sr->len = READ_ONCE(sqe->len);
564 sr->flags = READ_ONCE(sqe->ioprio);
565 if (sr->flags & ~(RECVMSG_FLAGS))
567 sr->msg_flags = READ_ONCE(sqe->msg_flags);
568 if (sr->msg_flags & MSG_DONTWAIT)
569 req->flags |= REQ_F_NOWAIT;
570 if (sr->msg_flags & MSG_ERRQUEUE)
571 req->flags |= REQ_F_CLEAR_POLLIN;
572 if (sr->flags & IORING_RECV_MULTISHOT) {
573 if (!(req->flags & REQ_F_BUFFER_SELECT))
575 if (sr->msg_flags & MSG_WAITALL)
577 if (req->opcode == IORING_OP_RECV && sr->len)
579 req->flags |= REQ_F_APOLL_MULTISHOT;
581 * Store the buffer group for this multishot receive separately,
582 * as if we end up doing an io-wq based issue that selects a
583 * buffer, it has to be committed immediately and that will
584 * clear ->buf_list. This means we lose the link to the buffer
585 * list, and the eventual buffer put on completion then cannot
588 sr->buf_group = req->buf_index;
592 if (req->ctx->compat)
593 sr->msg_flags |= MSG_CMSG_COMPAT;
599 static inline void io_recv_prep_retry(struct io_kiocb *req)
601 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
604 sr->len = 0; /* get from the provided buffer */
605 req->buf_index = sr->buf_group;
609 * Finishes io_recv and io_recvmsg.
611 * Returns true if it is actually finished, or false if it should run
612 * again (for multishot).
614 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
615 unsigned int cflags, bool mshot_finished,
616 unsigned issue_flags)
618 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
619 io_req_set_res(req, *ret, cflags);
624 if (!mshot_finished) {
625 if (io_post_aux_cqe(req->ctx, req->cqe.user_data, *ret,
626 cflags | IORING_CQE_F_MORE, false)) {
627 io_recv_prep_retry(req);
631 * Otherwise stop multishot but use the current result.
632 * Probably will end up going into overflow, but this means
633 * we cannot trust the ordering anymore
637 io_req_set_res(req, *ret, cflags);
639 if (issue_flags & IO_URING_F_MULTISHOT)
640 *ret = IOU_STOP_MULTISHOT;
646 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
647 struct io_sr_msg *sr, void __user **buf,
650 unsigned long ubuf = (unsigned long) *buf;
653 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
658 if (kmsg->controllen) {
659 unsigned long control = ubuf + hdr - kmsg->controllen;
661 kmsg->msg.msg_control_user = (void __user *) control;
662 kmsg->msg.msg_controllen = kmsg->controllen;
665 sr->buf = *buf; /* stash for later copy */
666 *buf = (void __user *) (ubuf + hdr);
667 kmsg->payloadlen = *len = *len - hdr;
671 struct io_recvmsg_multishot_hdr {
672 struct io_uring_recvmsg_out msg;
673 struct sockaddr_storage addr;
676 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
677 struct io_async_msghdr *kmsg,
678 unsigned int flags, bool *finished)
682 struct io_recvmsg_multishot_hdr hdr;
685 kmsg->msg.msg_name = &hdr.addr;
686 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
687 kmsg->msg.msg_namelen = 0;
689 if (sock->file->f_flags & O_NONBLOCK)
690 flags |= MSG_DONTWAIT;
692 err = sock_recvmsg(sock, &kmsg->msg, flags);
693 *finished = err <= 0;
697 hdr.msg = (struct io_uring_recvmsg_out) {
698 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
699 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
702 hdr.msg.payloadlen = err;
703 if (err > kmsg->payloadlen)
704 err = kmsg->payloadlen;
706 copy_len = sizeof(struct io_uring_recvmsg_out);
707 if (kmsg->msg.msg_namelen > kmsg->namelen)
708 copy_len += kmsg->namelen;
710 copy_len += kmsg->msg.msg_namelen;
713 * "fromlen shall refer to the value before truncation.."
716 hdr.msg.namelen = kmsg->msg.msg_namelen;
718 /* ensure that there is no gap between hdr and sockaddr_storage */
719 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
720 sizeof(struct io_uring_recvmsg_out));
721 if (copy_to_user(io->buf, &hdr, copy_len)) {
726 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
727 kmsg->controllen + err;
730 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
732 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
733 struct io_async_msghdr iomsg, *kmsg;
737 int ret, min_ret = 0;
738 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
739 bool mshot_finished = true;
741 sock = sock_from_file(req->file);
745 if (req_has_async_data(req)) {
746 kmsg = req->async_data;
748 ret = io_recvmsg_copy_hdr(req, &iomsg);
754 if (!(req->flags & REQ_F_POLLED) &&
755 (sr->flags & IORING_RECVSEND_POLL_FIRST))
756 return io_setup_async_msg(req, kmsg, issue_flags);
759 if (io_do_buffer_select(req)) {
761 size_t len = sr->len;
763 buf = io_buffer_select(req, &len, issue_flags);
767 if (req->flags & REQ_F_APOLL_MULTISHOT) {
768 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
770 io_kbuf_recycle(req, issue_flags);
775 kmsg->fast_iov[0].iov_base = buf;
776 kmsg->fast_iov[0].iov_len = len;
777 iov_iter_init(&kmsg->msg.msg_iter, ITER_DEST, kmsg->fast_iov, 1,
781 flags = sr->msg_flags;
783 flags |= MSG_DONTWAIT;
785 kmsg->msg.msg_get_inq = 1;
786 if (req->flags & REQ_F_APOLL_MULTISHOT) {
787 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
790 /* disable partial retry for recvmsg with cmsg attached */
791 if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
792 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
794 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
799 if (ret == -EAGAIN && force_nonblock) {
800 ret = io_setup_async_msg(req, kmsg, issue_flags);
801 if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
802 io_kbuf_recycle(req, issue_flags);
803 return IOU_ISSUE_SKIP_COMPLETE;
807 if (ret > 0 && io_net_retry(sock, flags)) {
809 req->flags |= REQ_F_PARTIAL_IO;
810 return io_setup_async_msg(req, kmsg, issue_flags);
812 if (ret == -ERESTARTSYS)
815 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
821 else if (sr->done_io)
824 io_kbuf_recycle(req, issue_flags);
826 cflags = io_put_kbuf(req, issue_flags);
827 if (kmsg->msg.msg_inq)
828 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
830 if (!io_recv_finish(req, &ret, cflags, mshot_finished, issue_flags))
831 goto retry_multishot;
833 if (mshot_finished) {
834 /* fast path, check for non-NULL to avoid function call */
836 kfree(kmsg->free_iov);
837 io_netmsg_recycle(req, issue_flags);
838 req->flags &= ~REQ_F_NEED_CLEANUP;
844 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
846 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
852 int ret, min_ret = 0;
853 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
854 size_t len = sr->len;
856 if (!(req->flags & REQ_F_POLLED) &&
857 (sr->flags & IORING_RECVSEND_POLL_FIRST))
860 sock = sock_from_file(req->file);
865 if (io_do_buffer_select(req)) {
868 buf = io_buffer_select(req, &len, issue_flags);
874 ret = import_single_range(ITER_DEST, sr->buf, len, &iov, &msg.msg_iter);
880 msg.msg_control = NULL;
883 msg.msg_controllen = 0;
887 flags = sr->msg_flags;
889 flags |= MSG_DONTWAIT;
890 if (flags & MSG_WAITALL)
891 min_ret = iov_iter_count(&msg.msg_iter);
893 ret = sock_recvmsg(sock, &msg, flags);
895 if (ret == -EAGAIN && force_nonblock) {
896 if (issue_flags & IO_URING_F_MULTISHOT) {
897 io_kbuf_recycle(req, issue_flags);
898 return IOU_ISSUE_SKIP_COMPLETE;
903 if (ret > 0 && io_net_retry(sock, flags)) {
907 req->flags |= REQ_F_PARTIAL_IO;
910 if (ret == -ERESTARTSYS)
913 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
920 else if (sr->done_io)
923 io_kbuf_recycle(req, issue_flags);
925 cflags = io_put_kbuf(req, issue_flags);
927 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
929 if (!io_recv_finish(req, &ret, cflags, ret <= 0, issue_flags))
930 goto retry_multishot;
935 void io_send_zc_cleanup(struct io_kiocb *req)
937 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
938 struct io_async_msghdr *io;
940 if (req_has_async_data(req)) {
941 io = req->async_data;
942 /* might be ->fast_iov if *msg_copy_hdr failed */
943 if (io->free_iov != io->fast_iov)
947 io_notif_flush(zc->notif);
952 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
954 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
955 struct io_ring_ctx *ctx = req->ctx;
956 struct io_kiocb *notif;
958 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
960 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
961 if (req->flags & REQ_F_CQE_SKIP)
964 zc->flags = READ_ONCE(sqe->ioprio);
965 if (zc->flags & ~(IORING_RECVSEND_POLL_FIRST |
966 IORING_RECVSEND_FIXED_BUF |
967 IORING_SEND_ZC_REPORT_USAGE))
969 notif = zc->notif = io_alloc_notif(ctx);
972 notif->cqe.user_data = req->cqe.user_data;
974 notif->cqe.flags = IORING_CQE_F_NOTIF;
975 req->flags |= REQ_F_NEED_CLEANUP;
976 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
977 unsigned idx = READ_ONCE(sqe->buf_index);
979 if (unlikely(idx >= ctx->nr_user_bufs))
981 idx = array_index_nospec(idx, ctx->nr_user_bufs);
982 req->imu = READ_ONCE(ctx->user_bufs[idx]);
983 io_req_set_rsrc_node(notif, ctx, 0);
985 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
986 io_notif_to_data(notif)->zc_report = true;
989 if (req->opcode == IORING_OP_SEND_ZC) {
990 if (READ_ONCE(sqe->__pad3[0]))
992 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
993 zc->addr_len = READ_ONCE(sqe->addr_len);
995 if (unlikely(sqe->addr2 || sqe->file_index))
997 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1001 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1002 zc->len = READ_ONCE(sqe->len);
1003 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1004 if (zc->msg_flags & MSG_DONTWAIT)
1005 req->flags |= REQ_F_NOWAIT;
1009 #ifdef CONFIG_COMPAT
1010 if (req->ctx->compat)
1011 zc->msg_flags |= MSG_CMSG_COMPAT;
1016 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1017 struct iov_iter *from, size_t length)
1019 skb_zcopy_downgrade_managed(skb);
1020 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1023 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1024 struct iov_iter *from, size_t length)
1026 struct skb_shared_info *shinfo = skb_shinfo(skb);
1027 int frag = shinfo->nr_frags;
1029 struct bvec_iter bi;
1031 unsigned long truesize = 0;
1034 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1035 else if (unlikely(!skb_zcopy_managed(skb)))
1036 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1038 bi.bi_size = min(from->count, length);
1039 bi.bi_bvec_done = from->iov_offset;
1042 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1043 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1046 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1047 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1048 v.bv_offset, v.bv_len);
1049 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1054 shinfo->nr_frags = frag;
1055 from->bvec += bi.bi_idx;
1056 from->nr_segs -= bi.bi_idx;
1057 from->count -= copied;
1058 from->iov_offset = bi.bi_bvec_done;
1060 skb->data_len += copied;
1062 skb->truesize += truesize;
1064 if (sk && sk->sk_type == SOCK_STREAM) {
1065 sk_wmem_queued_add(sk, truesize);
1066 if (!skb_zcopy_pure(skb))
1067 sk_mem_charge(sk, truesize);
1069 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1074 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1076 struct sockaddr_storage __address;
1077 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1080 struct socket *sock;
1082 int ret, min_ret = 0;
1084 sock = sock_from_file(req->file);
1085 if (unlikely(!sock))
1087 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1090 msg.msg_name = NULL;
1091 msg.msg_control = NULL;
1092 msg.msg_controllen = 0;
1093 msg.msg_namelen = 0;
1096 if (req_has_async_data(req)) {
1097 struct io_async_msghdr *io = req->async_data;
1099 msg.msg_name = &io->addr;
1101 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1102 if (unlikely(ret < 0))
1104 msg.msg_name = (struct sockaddr *)&__address;
1106 msg.msg_namelen = zc->addr_len;
1109 if (!(req->flags & REQ_F_POLLED) &&
1110 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1111 return io_setup_async_addr(req, &__address, issue_flags);
1113 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1114 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1115 (u64)(uintptr_t)zc->buf, zc->len);
1118 msg.sg_from_iter = io_sg_from_iter;
1120 ret = import_single_range(ITER_SOURCE, zc->buf, zc->len, &iov,
1124 ret = io_notif_account_mem(zc->notif, zc->len);
1127 msg.sg_from_iter = io_sg_from_iter_iovec;
1130 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1131 if (issue_flags & IO_URING_F_NONBLOCK)
1132 msg_flags |= MSG_DONTWAIT;
1133 if (msg_flags & MSG_WAITALL)
1134 min_ret = iov_iter_count(&msg.msg_iter);
1136 msg.msg_flags = msg_flags;
1137 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1138 ret = sock_sendmsg(sock, &msg);
1140 if (unlikely(ret < min_ret)) {
1141 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1142 return io_setup_async_addr(req, &__address, issue_flags);
1144 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1148 req->flags |= REQ_F_PARTIAL_IO;
1149 return io_setup_async_addr(req, &__address, issue_flags);
1151 if (ret == -ERESTARTSYS)
1158 else if (zc->done_io)
1162 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1163 * flushing notif to io_send_zc_cleanup()
1165 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1166 io_notif_flush(zc->notif);
1167 req->flags &= ~REQ_F_NEED_CLEANUP;
1169 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1173 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1175 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1176 struct io_async_msghdr iomsg, *kmsg;
1177 struct socket *sock;
1179 int ret, min_ret = 0;
1181 sock = sock_from_file(req->file);
1182 if (unlikely(!sock))
1184 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1187 if (req_has_async_data(req)) {
1188 kmsg = req->async_data;
1190 ret = io_sendmsg_copy_hdr(req, &iomsg);
1196 if (!(req->flags & REQ_F_POLLED) &&
1197 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1198 return io_setup_async_msg(req, kmsg, issue_flags);
1200 flags = sr->msg_flags | MSG_ZEROCOPY;
1201 if (issue_flags & IO_URING_F_NONBLOCK)
1202 flags |= MSG_DONTWAIT;
1203 if (flags & MSG_WAITALL)
1204 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1206 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1207 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1208 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1210 if (unlikely(ret < min_ret)) {
1211 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1212 return io_setup_async_msg(req, kmsg, issue_flags);
1214 if (ret > 0 && io_net_retry(sock, flags)) {
1216 req->flags |= REQ_F_PARTIAL_IO;
1217 return io_setup_async_msg(req, kmsg, issue_flags);
1219 if (ret == -ERESTARTSYS)
1223 /* fast path, check for non-NULL to avoid function call */
1224 if (kmsg->free_iov) {
1225 kfree(kmsg->free_iov);
1226 kmsg->free_iov = NULL;
1229 io_netmsg_recycle(req, issue_flags);
1232 else if (sr->done_io)
1236 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1237 * flushing notif to io_send_zc_cleanup()
1239 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1240 io_notif_flush(sr->notif);
1241 req->flags &= ~REQ_F_NEED_CLEANUP;
1243 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1247 void io_sendrecv_fail(struct io_kiocb *req)
1249 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1251 if (req->flags & REQ_F_PARTIAL_IO)
1252 req->cqe.res = sr->done_io;
1254 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1255 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1256 req->cqe.flags |= IORING_CQE_F_MORE;
1259 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1261 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1264 if (sqe->len || sqe->buf_index)
1267 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1268 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1269 accept->flags = READ_ONCE(sqe->accept_flags);
1270 accept->nofile = rlimit(RLIMIT_NOFILE);
1271 flags = READ_ONCE(sqe->ioprio);
1272 if (flags & ~IORING_ACCEPT_MULTISHOT)
1275 accept->file_slot = READ_ONCE(sqe->file_index);
1276 if (accept->file_slot) {
1277 if (accept->flags & SOCK_CLOEXEC)
1279 if (flags & IORING_ACCEPT_MULTISHOT &&
1280 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1283 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1285 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1286 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1287 if (flags & IORING_ACCEPT_MULTISHOT)
1288 req->flags |= REQ_F_APOLL_MULTISHOT;
1292 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1294 struct io_ring_ctx *ctx = req->ctx;
1295 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1296 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1297 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1298 bool fixed = !!accept->file_slot;
1304 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1305 if (unlikely(fd < 0))
1308 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1313 ret = PTR_ERR(file);
1314 if (ret == -EAGAIN && force_nonblock) {
1316 * if it's multishot and polled, we don't need to
1317 * return EAGAIN to arm the poll infra since it
1318 * has already been done
1320 if (issue_flags & IO_URING_F_MULTISHOT)
1321 ret = IOU_ISSUE_SKIP_COMPLETE;
1324 if (ret == -ERESTARTSYS)
1327 } else if (!fixed) {
1328 fd_install(fd, file);
1331 ret = io_fixed_fd_install(req, issue_flags, file,
1335 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1336 io_req_set_res(req, ret, 0);
1341 io_post_aux_cqe(ctx, req->cqe.user_data, ret, IORING_CQE_F_MORE, false))
1344 io_req_set_res(req, ret, 0);
1345 return (issue_flags & IO_URING_F_MULTISHOT) ? IOU_STOP_MULTISHOT : IOU_OK;
1348 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1350 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1352 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1355 sock->domain = READ_ONCE(sqe->fd);
1356 sock->type = READ_ONCE(sqe->off);
1357 sock->protocol = READ_ONCE(sqe->len);
1358 sock->file_slot = READ_ONCE(sqe->file_index);
1359 sock->nofile = rlimit(RLIMIT_NOFILE);
1361 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1362 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1364 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1369 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1371 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1372 bool fixed = !!sock->file_slot;
1377 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1378 if (unlikely(fd < 0))
1381 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1385 ret = PTR_ERR(file);
1386 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1388 if (ret == -ERESTARTSYS)
1391 } else if (!fixed) {
1392 fd_install(fd, file);
1395 ret = io_fixed_fd_install(req, issue_flags, file,
1398 io_req_set_res(req, ret, 0);
1402 int io_connect_prep_async(struct io_kiocb *req)
1404 struct io_async_connect *io = req->async_data;
1405 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1407 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1410 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1412 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1414 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1417 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1418 conn->addr_len = READ_ONCE(sqe->addr2);
1419 conn->in_progress = conn->seen_econnaborted = false;
1423 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1425 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1426 struct io_async_connect __io, *io;
1427 unsigned file_flags;
1429 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1431 if (connect->in_progress) {
1432 struct socket *socket;
1435 socket = sock_from_file(req->file);
1437 ret = sock_error(socket->sk);
1441 if (req_has_async_data(req)) {
1442 io = req->async_data;
1444 ret = move_addr_to_kernel(connect->addr,
1452 file_flags = force_nonblock ? O_NONBLOCK : 0;
1454 ret = __sys_connect_file(req->file, &io->address,
1455 connect->addr_len, file_flags);
1456 if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1457 && force_nonblock) {
1458 if (ret == -EINPROGRESS) {
1459 connect->in_progress = true;
1462 if (ret == -ECONNABORTED) {
1463 if (connect->seen_econnaborted)
1465 connect->seen_econnaborted = true;
1467 if (req_has_async_data(req))
1469 if (io_alloc_async_data(req)) {
1473 memcpy(req->async_data, &__io, sizeof(__io));
1476 if (ret == -ERESTARTSYS)
1481 io_req_set_res(req, ret, 0);
1485 void io_netmsg_cache_free(struct io_cache_entry *entry)
1487 kfree(container_of(entry, struct io_async_msghdr, cache));