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;
55 struct compat_msghdr __user *umsg_compat;
56 struct user_msghdr __user *umsg;
63 /* initialised and used only by !msg send variants */
66 /* used only for send zerocopy */
67 struct io_kiocb *notif;
70 static inline bool io_check_multishot(struct io_kiocb *req,
71 unsigned int issue_flags)
74 * When ->locked_cq is set we only allow to post CQEs from the original
75 * task context. Usual request completions will be handled in other
76 * generic paths but multipoll may decide to post extra cqes.
78 return !(issue_flags & IO_URING_F_IOWQ) ||
79 !(issue_flags & IO_URING_F_MULTISHOT) ||
80 !req->ctx->task_complete;
83 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
85 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
87 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
88 sqe->buf_index || sqe->splice_fd_in))
91 shutdown->how = READ_ONCE(sqe->len);
95 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
97 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
101 if (issue_flags & IO_URING_F_NONBLOCK)
104 sock = sock_from_file(req->file);
108 ret = __sys_shutdown_sock(sock, shutdown->how);
109 io_req_set_res(req, ret, 0);
113 static bool io_net_retry(struct socket *sock, int flags)
115 if (!(flags & MSG_WAITALL))
117 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
120 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
122 struct io_async_msghdr *hdr = req->async_data;
124 if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
127 /* Let normal cleanup path reap it if we fail adding to the cache */
128 if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
129 req->async_data = NULL;
130 req->flags &= ~REQ_F_ASYNC_DATA;
134 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
135 unsigned int issue_flags)
137 struct io_ring_ctx *ctx = req->ctx;
138 struct io_cache_entry *entry;
139 struct io_async_msghdr *hdr;
141 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
142 entry = io_alloc_cache_get(&ctx->netmsg_cache);
144 hdr = container_of(entry, struct io_async_msghdr, cache);
145 hdr->free_iov = NULL;
146 req->flags |= REQ_F_ASYNC_DATA;
147 req->async_data = hdr;
152 if (!io_alloc_async_data(req)) {
153 hdr = req->async_data;
154 hdr->free_iov = NULL;
160 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
162 /* ->prep_async is always called from the submission context */
163 return io_msg_alloc_async(req, 0);
166 static int io_setup_async_msg(struct io_kiocb *req,
167 struct io_async_msghdr *kmsg,
168 unsigned int issue_flags)
170 struct io_async_msghdr *async_msg;
172 if (req_has_async_data(req))
174 async_msg = io_msg_alloc_async(req, issue_flags);
176 kfree(kmsg->free_iov);
179 req->flags |= REQ_F_NEED_CLEANUP;
180 memcpy(async_msg, kmsg, sizeof(*kmsg));
181 if (async_msg->msg.msg_name)
182 async_msg->msg.msg_name = &async_msg->addr;
183 /* if were using fast_iov, set it to the new one */
184 if (!kmsg->free_iov) {
185 size_t fast_idx = kmsg->msg.msg_iter.iov - kmsg->fast_iov;
186 async_msg->msg.msg_iter.iov = &async_msg->fast_iov[fast_idx];
192 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
193 struct io_async_msghdr *iomsg)
195 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
197 iomsg->msg.msg_name = &iomsg->addr;
198 iomsg->free_iov = iomsg->fast_iov;
199 return sendmsg_copy_msghdr(&iomsg->msg, sr->umsg, sr->msg_flags,
203 int io_send_prep_async(struct io_kiocb *req)
205 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
206 struct io_async_msghdr *io;
209 if (!zc->addr || req_has_async_data(req))
211 io = io_msg_alloc_async_prep(req);
214 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
218 static int io_setup_async_addr(struct io_kiocb *req,
219 struct sockaddr_storage *addr_storage,
220 unsigned int issue_flags)
222 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
223 struct io_async_msghdr *io;
225 if (!sr->addr || req_has_async_data(req))
227 io = io_msg_alloc_async(req, issue_flags);
230 memcpy(&io->addr, addr_storage, sizeof(io->addr));
234 int io_sendmsg_prep_async(struct io_kiocb *req)
238 if (!io_msg_alloc_async_prep(req))
240 ret = io_sendmsg_copy_hdr(req, req->async_data);
242 req->flags |= REQ_F_NEED_CLEANUP;
246 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
248 struct io_async_msghdr *io = req->async_data;
253 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
255 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
257 if (req->opcode == IORING_OP_SEND) {
258 if (READ_ONCE(sqe->__pad3[0]))
260 sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
261 sr->addr_len = READ_ONCE(sqe->addr_len);
262 } else if (sqe->addr2 || sqe->file_index) {
266 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
267 sr->len = READ_ONCE(sqe->len);
268 sr->flags = READ_ONCE(sqe->ioprio);
269 if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
271 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
272 if (sr->msg_flags & MSG_DONTWAIT)
273 req->flags |= REQ_F_NOWAIT;
276 if (req->ctx->compat)
277 sr->msg_flags |= MSG_CMSG_COMPAT;
283 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
285 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
286 struct io_async_msghdr iomsg, *kmsg;
292 sock = sock_from_file(req->file);
296 if (req_has_async_data(req)) {
297 kmsg = req->async_data;
299 ret = io_sendmsg_copy_hdr(req, &iomsg);
305 if (!(req->flags & REQ_F_POLLED) &&
306 (sr->flags & IORING_RECVSEND_POLL_FIRST))
307 return io_setup_async_msg(req, kmsg, issue_flags);
309 flags = sr->msg_flags;
310 if (issue_flags & IO_URING_F_NONBLOCK)
311 flags |= MSG_DONTWAIT;
312 if (flags & MSG_WAITALL)
313 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
315 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
318 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
319 return io_setup_async_msg(req, kmsg, issue_flags);
320 if (ret > 0 && io_net_retry(sock, flags)) {
322 req->flags |= REQ_F_PARTIAL_IO;
323 return io_setup_async_msg(req, kmsg, issue_flags);
325 if (ret == -ERESTARTSYS)
329 /* fast path, check for non-NULL to avoid function call */
331 kfree(kmsg->free_iov);
332 req->flags &= ~REQ_F_NEED_CLEANUP;
333 io_netmsg_recycle(req, issue_flags);
336 else if (sr->done_io)
338 io_req_set_res(req, ret, 0);
342 int io_send(struct io_kiocb *req, unsigned int issue_flags)
344 struct sockaddr_storage __address;
345 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
354 msg.msg_control = NULL;
355 msg.msg_controllen = 0;
360 if (req_has_async_data(req)) {
361 struct io_async_msghdr *io = req->async_data;
363 msg.msg_name = &io->addr;
365 ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
366 if (unlikely(ret < 0))
368 msg.msg_name = (struct sockaddr *)&__address;
370 msg.msg_namelen = sr->addr_len;
373 if (!(req->flags & REQ_F_POLLED) &&
374 (sr->flags & IORING_RECVSEND_POLL_FIRST))
375 return io_setup_async_addr(req, &__address, issue_flags);
377 sock = sock_from_file(req->file);
381 ret = import_single_range(ITER_SOURCE, sr->buf, sr->len, &iov, &msg.msg_iter);
385 flags = sr->msg_flags;
386 if (issue_flags & IO_URING_F_NONBLOCK)
387 flags |= MSG_DONTWAIT;
388 if (flags & MSG_WAITALL)
389 min_ret = iov_iter_count(&msg.msg_iter);
391 msg.msg_flags = flags;
392 ret = sock_sendmsg(sock, &msg);
394 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
395 return io_setup_async_addr(req, &__address, issue_flags);
397 if (ret > 0 && io_net_retry(sock, flags)) {
401 req->flags |= REQ_F_PARTIAL_IO;
402 return io_setup_async_addr(req, &__address, issue_flags);
404 if (ret == -ERESTARTSYS)
410 else if (sr->done_io)
412 io_req_set_res(req, ret, 0);
416 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
420 if (iomsg->namelen < 0)
422 if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
423 iomsg->namelen, &hdr))
425 if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
431 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
432 struct io_async_msghdr *iomsg)
434 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
435 struct user_msghdr msg;
438 if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
441 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
445 if (req->flags & REQ_F_BUFFER_SELECT) {
446 if (msg.msg_iovlen == 0) {
447 sr->len = iomsg->fast_iov[0].iov_len = 0;
448 iomsg->fast_iov[0].iov_base = NULL;
449 iomsg->free_iov = NULL;
450 } else if (msg.msg_iovlen > 1) {
453 if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
455 sr->len = iomsg->fast_iov[0].iov_len;
456 iomsg->free_iov = NULL;
459 if (req->flags & REQ_F_APOLL_MULTISHOT) {
460 iomsg->namelen = msg.msg_namelen;
461 iomsg->controllen = msg.msg_controllen;
462 if (io_recvmsg_multishot_overflow(iomsg))
466 iomsg->free_iov = iomsg->fast_iov;
467 ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
468 &iomsg->free_iov, &iomsg->msg.msg_iter,
478 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
479 struct io_async_msghdr *iomsg)
481 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
482 struct compat_msghdr msg;
483 struct compat_iovec __user *uiov;
486 if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
489 ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
493 uiov = compat_ptr(msg.msg_iov);
494 if (req->flags & REQ_F_BUFFER_SELECT) {
497 if (msg.msg_iovlen == 0) {
499 } else if (msg.msg_iovlen > 1) {
502 if (!access_ok(uiov, sizeof(*uiov)))
504 if (__get_user(clen, &uiov->iov_len))
511 if (req->flags & REQ_F_APOLL_MULTISHOT) {
512 iomsg->namelen = msg.msg_namelen;
513 iomsg->controllen = msg.msg_controllen;
514 if (io_recvmsg_multishot_overflow(iomsg))
518 iomsg->free_iov = iomsg->fast_iov;
519 ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
520 UIO_FASTIOV, &iomsg->free_iov,
521 &iomsg->msg.msg_iter, true);
530 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
531 struct io_async_msghdr *iomsg)
533 iomsg->msg.msg_name = &iomsg->addr;
536 if (req->ctx->compat)
537 return __io_compat_recvmsg_copy_hdr(req, iomsg);
540 return __io_recvmsg_copy_hdr(req, iomsg);
543 int io_recvmsg_prep_async(struct io_kiocb *req)
547 if (!io_msg_alloc_async_prep(req))
549 ret = io_recvmsg_copy_hdr(req, req->async_data);
551 req->flags |= REQ_F_NEED_CLEANUP;
555 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
557 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
559 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
561 if (unlikely(sqe->file_index || sqe->addr2))
564 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
565 sr->len = READ_ONCE(sqe->len);
566 sr->flags = READ_ONCE(sqe->ioprio);
567 if (sr->flags & ~(RECVMSG_FLAGS))
569 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
570 if (sr->msg_flags & MSG_DONTWAIT)
571 req->flags |= REQ_F_NOWAIT;
572 if (sr->msg_flags & MSG_ERRQUEUE)
573 req->flags |= REQ_F_CLEAR_POLLIN;
574 if (sr->flags & IORING_RECV_MULTISHOT) {
575 if (!(req->flags & REQ_F_BUFFER_SELECT))
577 if (sr->msg_flags & MSG_WAITALL)
579 if (req->opcode == IORING_OP_RECV && sr->len)
581 req->flags |= REQ_F_APOLL_MULTISHOT;
585 if (req->ctx->compat)
586 sr->msg_flags |= MSG_CMSG_COMPAT;
592 static inline void io_recv_prep_retry(struct io_kiocb *req)
594 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
597 sr->len = 0; /* get from the provided buffer */
601 * Finishes io_recv and io_recvmsg.
603 * Returns true if it is actually finished, or false if it should run
604 * again (for multishot).
606 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
607 unsigned int cflags, bool mshot_finished,
608 unsigned issue_flags)
610 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
611 io_req_set_res(req, *ret, cflags);
616 if (!mshot_finished) {
617 if (io_aux_cqe(req->ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
618 req->cqe.user_data, *ret, cflags | IORING_CQE_F_MORE, true)) {
619 io_recv_prep_retry(req);
622 /* Otherwise stop multishot but use the current result. */
625 io_req_set_res(req, *ret, cflags);
627 if (issue_flags & IO_URING_F_MULTISHOT)
628 *ret = IOU_STOP_MULTISHOT;
634 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
635 struct io_sr_msg *sr, void __user **buf,
638 unsigned long ubuf = (unsigned long) *buf;
641 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
646 if (kmsg->controllen) {
647 unsigned long control = ubuf + hdr - kmsg->controllen;
649 kmsg->msg.msg_control_user = (void __user *) control;
650 kmsg->msg.msg_controllen = kmsg->controllen;
653 sr->buf = *buf; /* stash for later copy */
654 *buf = (void __user *) (ubuf + hdr);
655 kmsg->payloadlen = *len = *len - hdr;
659 struct io_recvmsg_multishot_hdr {
660 struct io_uring_recvmsg_out msg;
661 struct sockaddr_storage addr;
664 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
665 struct io_async_msghdr *kmsg,
666 unsigned int flags, bool *finished)
670 struct io_recvmsg_multishot_hdr hdr;
673 kmsg->msg.msg_name = &hdr.addr;
674 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
675 kmsg->msg.msg_namelen = 0;
677 if (sock->file->f_flags & O_NONBLOCK)
678 flags |= MSG_DONTWAIT;
680 err = sock_recvmsg(sock, &kmsg->msg, flags);
681 *finished = err <= 0;
685 hdr.msg = (struct io_uring_recvmsg_out) {
686 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
687 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
690 hdr.msg.payloadlen = err;
691 if (err > kmsg->payloadlen)
692 err = kmsg->payloadlen;
694 copy_len = sizeof(struct io_uring_recvmsg_out);
695 if (kmsg->msg.msg_namelen > kmsg->namelen)
696 copy_len += kmsg->namelen;
698 copy_len += kmsg->msg.msg_namelen;
701 * "fromlen shall refer to the value before truncation.."
704 hdr.msg.namelen = kmsg->msg.msg_namelen;
706 /* ensure that there is no gap between hdr and sockaddr_storage */
707 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
708 sizeof(struct io_uring_recvmsg_out));
709 if (copy_to_user(io->buf, &hdr, copy_len)) {
714 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
715 kmsg->controllen + err;
718 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
720 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
721 struct io_async_msghdr iomsg, *kmsg;
725 int ret, min_ret = 0;
726 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
727 bool mshot_finished = true;
729 sock = sock_from_file(req->file);
733 if (req_has_async_data(req)) {
734 kmsg = req->async_data;
736 ret = io_recvmsg_copy_hdr(req, &iomsg);
742 if (!(req->flags & REQ_F_POLLED) &&
743 (sr->flags & IORING_RECVSEND_POLL_FIRST))
744 return io_setup_async_msg(req, kmsg, issue_flags);
746 if (!io_check_multishot(req, issue_flags))
747 return io_setup_async_msg(req, kmsg, issue_flags);
750 if (io_do_buffer_select(req)) {
752 size_t len = sr->len;
754 buf = io_buffer_select(req, &len, issue_flags);
758 if (req->flags & REQ_F_APOLL_MULTISHOT) {
759 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
761 io_kbuf_recycle(req, issue_flags);
766 kmsg->fast_iov[0].iov_base = buf;
767 kmsg->fast_iov[0].iov_len = len;
768 iov_iter_init(&kmsg->msg.msg_iter, ITER_DEST, kmsg->fast_iov, 1,
772 flags = sr->msg_flags;
774 flags |= MSG_DONTWAIT;
775 if (flags & MSG_WAITALL)
776 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
778 kmsg->msg.msg_get_inq = 1;
779 if (req->flags & REQ_F_APOLL_MULTISHOT)
780 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
783 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
787 if (ret == -EAGAIN && force_nonblock) {
788 ret = io_setup_async_msg(req, kmsg, issue_flags);
789 if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
790 io_kbuf_recycle(req, issue_flags);
791 return IOU_ISSUE_SKIP_COMPLETE;
795 if (ret > 0 && io_net_retry(sock, flags)) {
797 req->flags |= REQ_F_PARTIAL_IO;
798 return io_setup_async_msg(req, kmsg, issue_flags);
800 if (ret == -ERESTARTSYS)
803 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
809 else if (sr->done_io)
812 io_kbuf_recycle(req, issue_flags);
814 cflags = io_put_kbuf(req, issue_flags);
815 if (kmsg->msg.msg_inq)
816 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
818 if (!io_recv_finish(req, &ret, cflags, mshot_finished, issue_flags))
819 goto retry_multishot;
821 if (mshot_finished) {
822 io_netmsg_recycle(req, issue_flags);
823 /* fast path, check for non-NULL to avoid function call */
825 kfree(kmsg->free_iov);
826 req->flags &= ~REQ_F_NEED_CLEANUP;
832 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
834 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
840 int ret, min_ret = 0;
841 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
842 size_t len = sr->len;
844 if (!(req->flags & REQ_F_POLLED) &&
845 (sr->flags & IORING_RECVSEND_POLL_FIRST))
848 if (!io_check_multishot(req, issue_flags))
851 sock = sock_from_file(req->file);
856 if (io_do_buffer_select(req)) {
859 buf = io_buffer_select(req, &len, issue_flags);
865 ret = import_single_range(ITER_DEST, sr->buf, len, &iov, &msg.msg_iter);
871 msg.msg_control = NULL;
874 msg.msg_controllen = 0;
878 flags = sr->msg_flags;
880 flags |= MSG_DONTWAIT;
881 if (flags & MSG_WAITALL)
882 min_ret = iov_iter_count(&msg.msg_iter);
884 ret = sock_recvmsg(sock, &msg, flags);
886 if (ret == -EAGAIN && force_nonblock) {
887 if (issue_flags & IO_URING_F_MULTISHOT) {
888 io_kbuf_recycle(req, issue_flags);
889 return IOU_ISSUE_SKIP_COMPLETE;
894 if (ret > 0 && io_net_retry(sock, flags)) {
898 req->flags |= REQ_F_PARTIAL_IO;
901 if (ret == -ERESTARTSYS)
904 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
911 else if (sr->done_io)
914 io_kbuf_recycle(req, issue_flags);
916 cflags = io_put_kbuf(req, issue_flags);
918 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
920 if (!io_recv_finish(req, &ret, cflags, ret <= 0, issue_flags))
921 goto retry_multishot;
926 void io_send_zc_cleanup(struct io_kiocb *req)
928 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
929 struct io_async_msghdr *io;
931 if (req_has_async_data(req)) {
932 io = req->async_data;
933 /* might be ->fast_iov if *msg_copy_hdr failed */
934 if (io->free_iov != io->fast_iov)
938 io_notif_flush(zc->notif);
943 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
944 #define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
946 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
948 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
949 struct io_ring_ctx *ctx = req->ctx;
950 struct io_kiocb *notif;
952 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
954 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
955 if (req->flags & REQ_F_CQE_SKIP)
958 notif = zc->notif = io_alloc_notif(ctx);
961 notif->cqe.user_data = req->cqe.user_data;
963 notif->cqe.flags = IORING_CQE_F_NOTIF;
964 req->flags |= REQ_F_NEED_CLEANUP;
966 zc->flags = READ_ONCE(sqe->ioprio);
967 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
968 if (zc->flags & ~IO_ZC_FLAGS_VALID)
970 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
971 io_notif_set_extended(notif);
972 io_notif_to_data(notif)->zc_report = true;
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);
986 if (req->opcode == IORING_OP_SEND_ZC) {
987 if (READ_ONCE(sqe->__pad3[0]))
989 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
990 zc->addr_len = READ_ONCE(sqe->addr_len);
992 if (unlikely(sqe->addr2 || sqe->file_index))
994 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
998 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
999 zc->len = READ_ONCE(sqe->len);
1000 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1001 if (zc->msg_flags & MSG_DONTWAIT)
1002 req->flags |= REQ_F_NOWAIT;
1006 #ifdef CONFIG_COMPAT
1007 if (req->ctx->compat)
1008 zc->msg_flags |= MSG_CMSG_COMPAT;
1013 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1014 struct iov_iter *from, size_t length)
1016 skb_zcopy_downgrade_managed(skb);
1017 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1020 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1021 struct iov_iter *from, size_t length)
1023 struct skb_shared_info *shinfo = skb_shinfo(skb);
1024 int frag = shinfo->nr_frags;
1026 struct bvec_iter bi;
1028 unsigned long truesize = 0;
1031 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1032 else if (unlikely(!skb_zcopy_managed(skb)))
1033 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1035 bi.bi_size = min(from->count, length);
1036 bi.bi_bvec_done = from->iov_offset;
1039 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1040 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1043 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1044 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1045 v.bv_offset, v.bv_len);
1046 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1051 shinfo->nr_frags = frag;
1052 from->bvec += bi.bi_idx;
1053 from->nr_segs -= bi.bi_idx;
1054 from->count -= copied;
1055 from->iov_offset = bi.bi_bvec_done;
1057 skb->data_len += copied;
1059 skb->truesize += truesize;
1061 if (sk && sk->sk_type == SOCK_STREAM) {
1062 sk_wmem_queued_add(sk, truesize);
1063 if (!skb_zcopy_pure(skb))
1064 sk_mem_charge(sk, truesize);
1066 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1071 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1073 struct sockaddr_storage __address;
1074 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1077 struct socket *sock;
1079 int ret, min_ret = 0;
1081 sock = sock_from_file(req->file);
1082 if (unlikely(!sock))
1084 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1087 msg.msg_name = NULL;
1088 msg.msg_control = NULL;
1089 msg.msg_controllen = 0;
1090 msg.msg_namelen = 0;
1093 if (req_has_async_data(req)) {
1094 struct io_async_msghdr *io = req->async_data;
1096 msg.msg_name = &io->addr;
1098 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1099 if (unlikely(ret < 0))
1101 msg.msg_name = (struct sockaddr *)&__address;
1103 msg.msg_namelen = zc->addr_len;
1106 if (!(req->flags & REQ_F_POLLED) &&
1107 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1108 return io_setup_async_addr(req, &__address, issue_flags);
1110 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1111 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1112 (u64)(uintptr_t)zc->buf, zc->len);
1115 msg.sg_from_iter = io_sg_from_iter;
1117 io_notif_set_extended(zc->notif);
1118 ret = import_single_range(ITER_SOURCE, zc->buf, zc->len, &iov,
1122 ret = io_notif_account_mem(zc->notif, zc->len);
1125 msg.sg_from_iter = io_sg_from_iter_iovec;
1128 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1129 if (issue_flags & IO_URING_F_NONBLOCK)
1130 msg_flags |= MSG_DONTWAIT;
1131 if (msg_flags & MSG_WAITALL)
1132 min_ret = iov_iter_count(&msg.msg_iter);
1134 msg.msg_flags = msg_flags;
1135 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1136 ret = sock_sendmsg(sock, &msg);
1138 if (unlikely(ret < min_ret)) {
1139 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1140 return io_setup_async_addr(req, &__address, issue_flags);
1142 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1146 req->flags |= REQ_F_PARTIAL_IO;
1147 return io_setup_async_addr(req, &__address, issue_flags);
1149 if (ret == -ERESTARTSYS)
1156 else if (zc->done_io)
1160 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1161 * flushing notif to io_send_zc_cleanup()
1163 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1164 io_notif_flush(zc->notif);
1165 req->flags &= ~REQ_F_NEED_CLEANUP;
1167 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1171 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1173 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1174 struct io_async_msghdr iomsg, *kmsg;
1175 struct socket *sock;
1177 int ret, min_ret = 0;
1179 io_notif_set_extended(sr->notif);
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;
1302 if (!io_check_multishot(req, issue_flags))
1306 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1307 if (unlikely(fd < 0))
1310 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1315 ret = PTR_ERR(file);
1316 if (ret == -EAGAIN && force_nonblock) {
1318 * if it's multishot and polled, we don't need to
1319 * return EAGAIN to arm the poll infra since it
1320 * has already been done
1322 if (issue_flags & IO_URING_F_MULTISHOT)
1323 ret = IOU_ISSUE_SKIP_COMPLETE;
1326 if (ret == -ERESTARTSYS)
1329 } else if (!fixed) {
1330 fd_install(fd, file);
1333 ret = io_fixed_fd_install(req, issue_flags, file,
1337 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1338 io_req_set_res(req, ret, 0);
1344 if (io_aux_cqe(ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
1345 req->cqe.user_data, ret, IORING_CQE_F_MORE, true))
1351 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1353 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1355 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1358 sock->domain = READ_ONCE(sqe->fd);
1359 sock->type = READ_ONCE(sqe->off);
1360 sock->protocol = READ_ONCE(sqe->len);
1361 sock->file_slot = READ_ONCE(sqe->file_index);
1362 sock->nofile = rlimit(RLIMIT_NOFILE);
1364 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1365 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1367 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1372 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1374 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1375 bool fixed = !!sock->file_slot;
1380 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1381 if (unlikely(fd < 0))
1384 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1388 ret = PTR_ERR(file);
1389 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1391 if (ret == -ERESTARTSYS)
1394 } else if (!fixed) {
1395 fd_install(fd, file);
1398 ret = io_fixed_fd_install(req, issue_flags, file,
1401 io_req_set_res(req, ret, 0);
1405 int io_connect_prep_async(struct io_kiocb *req)
1407 struct io_async_connect *io = req->async_data;
1408 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1410 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1413 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1415 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1417 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1420 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1421 conn->addr_len = READ_ONCE(sqe->addr2);
1422 conn->in_progress = false;
1426 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1428 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1429 struct io_async_connect __io, *io;
1430 unsigned file_flags;
1432 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1434 if (connect->in_progress) {
1435 struct socket *socket;
1438 socket = sock_from_file(req->file);
1440 ret = sock_error(socket->sk);
1444 if (req_has_async_data(req)) {
1445 io = req->async_data;
1447 ret = move_addr_to_kernel(connect->addr,
1455 file_flags = force_nonblock ? O_NONBLOCK : 0;
1457 ret = __sys_connect_file(req->file, &io->address,
1458 connect->addr_len, file_flags);
1459 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
1460 if (ret == -EINPROGRESS) {
1461 connect->in_progress = true;
1463 if (req_has_async_data(req))
1465 if (io_alloc_async_data(req)) {
1469 memcpy(req->async_data, &__io, sizeof(__io));
1473 if (ret == -ERESTARTSYS)
1478 io_req_set_res(req, ret, 0);
1482 void io_netmsg_cache_free(struct io_cache_entry *entry)
1484 kfree(container_of(entry, struct io_async_msghdr, cache));