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 */
67 /* used only for send zerocopy */
68 struct io_kiocb *notif;
71 static inline bool io_check_multishot(struct io_kiocb *req,
72 unsigned int issue_flags)
75 * When ->locked_cq is set we only allow to post CQEs from the original
76 * task context. Usual request completions will be handled in other
77 * generic paths but multipoll may decide to post extra cqes.
79 return !(issue_flags & IO_URING_F_IOWQ) ||
80 !(issue_flags & IO_URING_F_MULTISHOT) ||
81 !req->ctx->task_complete;
84 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
86 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
88 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
89 sqe->buf_index || sqe->splice_fd_in))
92 shutdown->how = READ_ONCE(sqe->len);
96 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
98 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
102 if (issue_flags & IO_URING_F_NONBLOCK)
105 sock = sock_from_file(req->file);
109 ret = __sys_shutdown_sock(sock, shutdown->how);
110 io_req_set_res(req, ret, 0);
114 static bool io_net_retry(struct socket *sock, int flags)
116 if (!(flags & MSG_WAITALL))
118 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
121 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
123 struct io_async_msghdr *hdr = req->async_data;
125 if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
128 /* Let normal cleanup path reap it if we fail adding to the cache */
129 if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
130 req->async_data = NULL;
131 req->flags &= ~REQ_F_ASYNC_DATA;
135 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
136 unsigned int issue_flags)
138 struct io_ring_ctx *ctx = req->ctx;
139 struct io_cache_entry *entry;
140 struct io_async_msghdr *hdr;
142 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
143 entry = io_alloc_cache_get(&ctx->netmsg_cache);
145 hdr = container_of(entry, struct io_async_msghdr, cache);
146 hdr->free_iov = NULL;
147 req->flags |= REQ_F_ASYNC_DATA;
148 req->async_data = hdr;
153 if (!io_alloc_async_data(req)) {
154 hdr = req->async_data;
155 hdr->free_iov = NULL;
161 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
163 /* ->prep_async is always called from the submission context */
164 return io_msg_alloc_async(req, 0);
167 static int io_setup_async_msg(struct io_kiocb *req,
168 struct io_async_msghdr *kmsg,
169 unsigned int issue_flags)
171 struct io_async_msghdr *async_msg;
173 if (req_has_async_data(req))
175 async_msg = io_msg_alloc_async(req, issue_flags);
177 kfree(kmsg->free_iov);
180 req->flags |= REQ_F_NEED_CLEANUP;
181 memcpy(async_msg, kmsg, sizeof(*kmsg));
182 if (async_msg->msg.msg_name)
183 async_msg->msg.msg_name = &async_msg->addr;
184 /* if were using fast_iov, set it to the new one */
185 if (!kmsg->free_iov) {
186 size_t fast_idx = kmsg->msg.msg_iter.iov - kmsg->fast_iov;
187 async_msg->msg.msg_iter.iov = &async_msg->fast_iov[fast_idx];
193 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
194 struct io_async_msghdr *iomsg)
196 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
198 iomsg->msg.msg_name = &iomsg->addr;
199 iomsg->free_iov = iomsg->fast_iov;
200 return sendmsg_copy_msghdr(&iomsg->msg, sr->umsg, sr->msg_flags,
204 int io_send_prep_async(struct io_kiocb *req)
206 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
207 struct io_async_msghdr *io;
210 if (!zc->addr || req_has_async_data(req))
212 io = io_msg_alloc_async_prep(req);
215 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
219 static int io_setup_async_addr(struct io_kiocb *req,
220 struct sockaddr_storage *addr_storage,
221 unsigned int issue_flags)
223 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
224 struct io_async_msghdr *io;
226 if (!sr->addr || req_has_async_data(req))
228 io = io_msg_alloc_async(req, issue_flags);
231 memcpy(&io->addr, addr_storage, sizeof(io->addr));
235 int io_sendmsg_prep_async(struct io_kiocb *req)
239 if (!io_msg_alloc_async_prep(req))
241 ret = io_sendmsg_copy_hdr(req, req->async_data);
243 req->flags |= REQ_F_NEED_CLEANUP;
247 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
249 struct io_async_msghdr *io = req->async_data;
254 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
256 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
258 if (req->opcode == IORING_OP_SEND) {
259 if (READ_ONCE(sqe->__pad3[0]))
261 sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
262 sr->addr_len = READ_ONCE(sqe->addr_len);
263 } else if (sqe->addr2 || sqe->file_index) {
267 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
268 sr->len = READ_ONCE(sqe->len);
269 sr->flags = READ_ONCE(sqe->ioprio);
270 if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
272 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
273 if (sr->msg_flags & MSG_DONTWAIT)
274 req->flags |= REQ_F_NOWAIT;
277 if (req->ctx->compat)
278 sr->msg_flags |= MSG_CMSG_COMPAT;
284 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
286 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
287 struct io_async_msghdr iomsg, *kmsg;
293 sock = sock_from_file(req->file);
297 if (req_has_async_data(req)) {
298 kmsg = req->async_data;
300 ret = io_sendmsg_copy_hdr(req, &iomsg);
306 if (!(req->flags & REQ_F_POLLED) &&
307 (sr->flags & IORING_RECVSEND_POLL_FIRST))
308 return io_setup_async_msg(req, kmsg, issue_flags);
310 flags = sr->msg_flags;
311 if (issue_flags & IO_URING_F_NONBLOCK)
312 flags |= MSG_DONTWAIT;
313 if (flags & MSG_WAITALL)
314 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
316 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
319 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
320 return io_setup_async_msg(req, kmsg, issue_flags);
321 if (ret > 0 && io_net_retry(sock, flags)) {
323 req->flags |= REQ_F_PARTIAL_IO;
324 return io_setup_async_msg(req, kmsg, issue_flags);
326 if (ret == -ERESTARTSYS)
330 /* fast path, check for non-NULL to avoid function call */
332 kfree(kmsg->free_iov);
333 req->flags &= ~REQ_F_NEED_CLEANUP;
334 io_netmsg_recycle(req, issue_flags);
337 else if (sr->done_io)
339 io_req_set_res(req, ret, 0);
343 int io_send(struct io_kiocb *req, unsigned int issue_flags)
345 struct sockaddr_storage __address;
346 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
355 msg.msg_control = NULL;
356 msg.msg_controllen = 0;
361 if (req_has_async_data(req)) {
362 struct io_async_msghdr *io = req->async_data;
364 msg.msg_name = &io->addr;
366 ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
367 if (unlikely(ret < 0))
369 msg.msg_name = (struct sockaddr *)&__address;
371 msg.msg_namelen = sr->addr_len;
374 if (!(req->flags & REQ_F_POLLED) &&
375 (sr->flags & IORING_RECVSEND_POLL_FIRST))
376 return io_setup_async_addr(req, &__address, issue_flags);
378 sock = sock_from_file(req->file);
382 ret = import_single_range(ITER_SOURCE, sr->buf, sr->len, &iov, &msg.msg_iter);
386 flags = sr->msg_flags;
387 if (issue_flags & IO_URING_F_NONBLOCK)
388 flags |= MSG_DONTWAIT;
389 if (flags & MSG_WAITALL)
390 min_ret = iov_iter_count(&msg.msg_iter);
392 msg.msg_flags = flags;
393 ret = sock_sendmsg(sock, &msg);
395 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
396 return io_setup_async_addr(req, &__address, issue_flags);
398 if (ret > 0 && io_net_retry(sock, flags)) {
402 req->flags |= REQ_F_PARTIAL_IO;
403 return io_setup_async_addr(req, &__address, issue_flags);
405 if (ret == -ERESTARTSYS)
411 else if (sr->done_io)
413 io_req_set_res(req, ret, 0);
417 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
421 if (iomsg->namelen < 0)
423 if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
424 iomsg->namelen, &hdr))
426 if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
432 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
433 struct io_async_msghdr *iomsg)
435 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
436 struct user_msghdr msg;
439 if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
442 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
446 if (req->flags & REQ_F_BUFFER_SELECT) {
447 if (msg.msg_iovlen == 0) {
448 sr->len = iomsg->fast_iov[0].iov_len = 0;
449 iomsg->fast_iov[0].iov_base = NULL;
450 iomsg->free_iov = NULL;
451 } else if (msg.msg_iovlen > 1) {
454 if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
456 sr->len = iomsg->fast_iov[0].iov_len;
457 iomsg->free_iov = NULL;
460 if (req->flags & REQ_F_APOLL_MULTISHOT) {
461 iomsg->namelen = msg.msg_namelen;
462 iomsg->controllen = msg.msg_controllen;
463 if (io_recvmsg_multishot_overflow(iomsg))
467 iomsg->free_iov = iomsg->fast_iov;
468 ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
469 &iomsg->free_iov, &iomsg->msg.msg_iter,
479 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
480 struct io_async_msghdr *iomsg)
482 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
483 struct compat_msghdr msg;
484 struct compat_iovec __user *uiov;
487 if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
490 ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
494 uiov = compat_ptr(msg.msg_iov);
495 if (req->flags & REQ_F_BUFFER_SELECT) {
498 iomsg->free_iov = NULL;
499 if (msg.msg_iovlen == 0) {
501 } else if (msg.msg_iovlen > 1) {
504 if (!access_ok(uiov, sizeof(*uiov)))
506 if (__get_user(clen, &uiov->iov_len))
513 if (req->flags & REQ_F_APOLL_MULTISHOT) {
514 iomsg->namelen = msg.msg_namelen;
515 iomsg->controllen = msg.msg_controllen;
516 if (io_recvmsg_multishot_overflow(iomsg))
520 iomsg->free_iov = iomsg->fast_iov;
521 ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
522 UIO_FASTIOV, &iomsg->free_iov,
523 &iomsg->msg.msg_iter, true);
532 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
533 struct io_async_msghdr *iomsg)
535 iomsg->msg.msg_name = &iomsg->addr;
538 if (req->ctx->compat)
539 return __io_compat_recvmsg_copy_hdr(req, iomsg);
542 return __io_recvmsg_copy_hdr(req, iomsg);
545 int io_recvmsg_prep_async(struct io_kiocb *req)
549 if (!io_msg_alloc_async_prep(req))
551 ret = io_recvmsg_copy_hdr(req, req->async_data);
553 req->flags |= REQ_F_NEED_CLEANUP;
557 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
559 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
561 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
563 if (unlikely(sqe->file_index || sqe->addr2))
566 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
567 sr->len = READ_ONCE(sqe->len);
568 sr->flags = READ_ONCE(sqe->ioprio);
569 if (sr->flags & ~(RECVMSG_FLAGS))
571 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
572 if (sr->msg_flags & MSG_DONTWAIT)
573 req->flags |= REQ_F_NOWAIT;
574 if (sr->msg_flags & MSG_ERRQUEUE)
575 req->flags |= REQ_F_CLEAR_POLLIN;
576 if (sr->flags & IORING_RECV_MULTISHOT) {
577 if (!(req->flags & REQ_F_BUFFER_SELECT))
579 if (sr->msg_flags & MSG_WAITALL)
581 if (req->opcode == IORING_OP_RECV && sr->len)
583 req->flags |= REQ_F_APOLL_MULTISHOT;
585 * Store the buffer group for this multishot receive separately,
586 * as if we end up doing an io-wq based issue that selects a
587 * buffer, it has to be committed immediately and that will
588 * clear ->buf_list. This means we lose the link to the buffer
589 * list, and the eventual buffer put on completion then cannot
592 sr->buf_group = req->buf_index;
596 if (req->ctx->compat)
597 sr->msg_flags |= MSG_CMSG_COMPAT;
603 static inline void io_recv_prep_retry(struct io_kiocb *req)
605 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
608 sr->len = 0; /* get from the provided buffer */
609 req->buf_index = sr->buf_group;
613 * Finishes io_recv and io_recvmsg.
615 * Returns true if it is actually finished, or false if it should run
616 * again (for multishot).
618 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
619 unsigned int cflags, bool mshot_finished,
620 unsigned issue_flags)
622 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
623 io_req_set_res(req, *ret, cflags);
628 if (!mshot_finished) {
629 if (io_aux_cqe(req->ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
630 req->cqe.user_data, *ret, cflags | IORING_CQE_F_MORE, true)) {
631 io_recv_prep_retry(req);
634 /* Otherwise stop multishot but use the current result. */
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);
758 if (!io_check_multishot(req, issue_flags))
759 return io_setup_async_msg(req, kmsg, issue_flags);
762 if (io_do_buffer_select(req)) {
764 size_t len = sr->len;
766 buf = io_buffer_select(req, &len, issue_flags);
770 if (req->flags & REQ_F_APOLL_MULTISHOT) {
771 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
773 io_kbuf_recycle(req, issue_flags);
778 kmsg->fast_iov[0].iov_base = buf;
779 kmsg->fast_iov[0].iov_len = len;
780 iov_iter_init(&kmsg->msg.msg_iter, ITER_DEST, kmsg->fast_iov, 1,
784 flags = sr->msg_flags;
786 flags |= MSG_DONTWAIT;
787 if (flags & MSG_WAITALL)
788 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
790 kmsg->msg.msg_get_inq = 1;
791 if (req->flags & REQ_F_APOLL_MULTISHOT)
792 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
795 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 if (!io_check_multishot(req, issue_flags))
863 sock = sock_from_file(req->file);
868 if (io_do_buffer_select(req)) {
871 buf = io_buffer_select(req, &len, issue_flags);
877 ret = import_single_range(ITER_DEST, sr->buf, len, &iov, &msg.msg_iter);
883 msg.msg_control = NULL;
886 msg.msg_controllen = 0;
890 flags = sr->msg_flags;
892 flags |= MSG_DONTWAIT;
893 if (flags & MSG_WAITALL)
894 min_ret = iov_iter_count(&msg.msg_iter);
896 ret = sock_recvmsg(sock, &msg, flags);
898 if (ret == -EAGAIN && force_nonblock) {
899 if (issue_flags & IO_URING_F_MULTISHOT) {
900 io_kbuf_recycle(req, issue_flags);
901 return IOU_ISSUE_SKIP_COMPLETE;
906 if (ret > 0 && io_net_retry(sock, flags)) {
910 req->flags |= REQ_F_PARTIAL_IO;
913 if (ret == -ERESTARTSYS)
916 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
923 else if (sr->done_io)
926 io_kbuf_recycle(req, issue_flags);
928 cflags = io_put_kbuf(req, issue_flags);
930 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
932 if (!io_recv_finish(req, &ret, cflags, ret <= 0, issue_flags))
933 goto retry_multishot;
938 void io_send_zc_cleanup(struct io_kiocb *req)
940 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
941 struct io_async_msghdr *io;
943 if (req_has_async_data(req)) {
944 io = req->async_data;
945 /* might be ->fast_iov if *msg_copy_hdr failed */
946 if (io->free_iov != io->fast_iov)
950 io_notif_flush(zc->notif);
955 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
956 #define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
958 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
960 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
961 struct io_ring_ctx *ctx = req->ctx;
962 struct io_kiocb *notif;
964 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
966 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
967 if (req->flags & REQ_F_CQE_SKIP)
970 notif = zc->notif = io_alloc_notif(ctx);
973 notif->cqe.user_data = req->cqe.user_data;
975 notif->cqe.flags = IORING_CQE_F_NOTIF;
976 req->flags |= REQ_F_NEED_CLEANUP;
978 zc->flags = READ_ONCE(sqe->ioprio);
979 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
980 if (zc->flags & ~IO_ZC_FLAGS_VALID)
982 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
983 io_notif_set_extended(notif);
984 io_notif_to_data(notif)->zc_report = true;
988 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
989 unsigned idx = READ_ONCE(sqe->buf_index);
991 if (unlikely(idx >= ctx->nr_user_bufs))
993 idx = array_index_nospec(idx, ctx->nr_user_bufs);
994 req->imu = READ_ONCE(ctx->user_bufs[idx]);
995 io_req_set_rsrc_node(notif, ctx, 0);
998 if (req->opcode == IORING_OP_SEND_ZC) {
999 if (READ_ONCE(sqe->__pad3[0]))
1001 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1002 zc->addr_len = READ_ONCE(sqe->addr_len);
1004 if (unlikely(sqe->addr2 || sqe->file_index))
1006 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1010 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1011 zc->len = READ_ONCE(sqe->len);
1012 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1013 if (zc->msg_flags & MSG_DONTWAIT)
1014 req->flags |= REQ_F_NOWAIT;
1018 #ifdef CONFIG_COMPAT
1019 if (req->ctx->compat)
1020 zc->msg_flags |= MSG_CMSG_COMPAT;
1025 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1026 struct iov_iter *from, size_t length)
1028 skb_zcopy_downgrade_managed(skb);
1029 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1032 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1033 struct iov_iter *from, size_t length)
1035 struct skb_shared_info *shinfo = skb_shinfo(skb);
1036 int frag = shinfo->nr_frags;
1038 struct bvec_iter bi;
1040 unsigned long truesize = 0;
1043 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1044 else if (unlikely(!skb_zcopy_managed(skb)))
1045 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1047 bi.bi_size = min(from->count, length);
1048 bi.bi_bvec_done = from->iov_offset;
1051 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1052 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1055 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1056 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1057 v.bv_offset, v.bv_len);
1058 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1063 shinfo->nr_frags = frag;
1064 from->bvec += bi.bi_idx;
1065 from->nr_segs -= bi.bi_idx;
1066 from->count -= copied;
1067 from->iov_offset = bi.bi_bvec_done;
1069 skb->data_len += copied;
1071 skb->truesize += truesize;
1073 if (sk && sk->sk_type == SOCK_STREAM) {
1074 sk_wmem_queued_add(sk, truesize);
1075 if (!skb_zcopy_pure(skb))
1076 sk_mem_charge(sk, truesize);
1078 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1083 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1085 struct sockaddr_storage __address;
1086 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1089 struct socket *sock;
1091 int ret, min_ret = 0;
1093 sock = sock_from_file(req->file);
1094 if (unlikely(!sock))
1096 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1099 msg.msg_name = NULL;
1100 msg.msg_control = NULL;
1101 msg.msg_controllen = 0;
1102 msg.msg_namelen = 0;
1105 if (req_has_async_data(req)) {
1106 struct io_async_msghdr *io = req->async_data;
1108 msg.msg_name = &io->addr;
1110 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1111 if (unlikely(ret < 0))
1113 msg.msg_name = (struct sockaddr *)&__address;
1115 msg.msg_namelen = zc->addr_len;
1118 if (!(req->flags & REQ_F_POLLED) &&
1119 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1120 return io_setup_async_addr(req, &__address, issue_flags);
1122 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1123 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1124 (u64)(uintptr_t)zc->buf, zc->len);
1127 msg.sg_from_iter = io_sg_from_iter;
1129 io_notif_set_extended(zc->notif);
1130 ret = import_single_range(ITER_SOURCE, zc->buf, zc->len, &iov,
1134 ret = io_notif_account_mem(zc->notif, zc->len);
1137 msg.sg_from_iter = io_sg_from_iter_iovec;
1140 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1141 if (issue_flags & IO_URING_F_NONBLOCK)
1142 msg_flags |= MSG_DONTWAIT;
1143 if (msg_flags & MSG_WAITALL)
1144 min_ret = iov_iter_count(&msg.msg_iter);
1146 msg.msg_flags = msg_flags;
1147 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1148 ret = sock_sendmsg(sock, &msg);
1150 if (unlikely(ret < min_ret)) {
1151 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1152 return io_setup_async_addr(req, &__address, issue_flags);
1154 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1158 req->flags |= REQ_F_PARTIAL_IO;
1159 return io_setup_async_addr(req, &__address, issue_flags);
1161 if (ret == -ERESTARTSYS)
1168 else if (zc->done_io)
1172 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1173 * flushing notif to io_send_zc_cleanup()
1175 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1176 io_notif_flush(zc->notif);
1177 req->flags &= ~REQ_F_NEED_CLEANUP;
1179 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1183 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1185 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1186 struct io_async_msghdr iomsg, *kmsg;
1187 struct socket *sock;
1189 int ret, min_ret = 0;
1191 io_notif_set_extended(sr->notif);
1193 sock = sock_from_file(req->file);
1194 if (unlikely(!sock))
1196 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1199 if (req_has_async_data(req)) {
1200 kmsg = req->async_data;
1202 ret = io_sendmsg_copy_hdr(req, &iomsg);
1208 if (!(req->flags & REQ_F_POLLED) &&
1209 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1210 return io_setup_async_msg(req, kmsg, issue_flags);
1212 flags = sr->msg_flags | MSG_ZEROCOPY;
1213 if (issue_flags & IO_URING_F_NONBLOCK)
1214 flags |= MSG_DONTWAIT;
1215 if (flags & MSG_WAITALL)
1216 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1218 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1219 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1220 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1222 if (unlikely(ret < min_ret)) {
1223 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1224 return io_setup_async_msg(req, kmsg, issue_flags);
1226 if (ret > 0 && io_net_retry(sock, flags)) {
1228 req->flags |= REQ_F_PARTIAL_IO;
1229 return io_setup_async_msg(req, kmsg, issue_flags);
1231 if (ret == -ERESTARTSYS)
1235 /* fast path, check for non-NULL to avoid function call */
1236 if (kmsg->free_iov) {
1237 kfree(kmsg->free_iov);
1238 kmsg->free_iov = NULL;
1241 io_netmsg_recycle(req, issue_flags);
1244 else if (sr->done_io)
1248 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1249 * flushing notif to io_send_zc_cleanup()
1251 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1252 io_notif_flush(sr->notif);
1253 req->flags &= ~REQ_F_NEED_CLEANUP;
1255 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1259 void io_sendrecv_fail(struct io_kiocb *req)
1261 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1263 if (req->flags & REQ_F_PARTIAL_IO)
1264 req->cqe.res = sr->done_io;
1266 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1267 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1268 req->cqe.flags |= IORING_CQE_F_MORE;
1271 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1273 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1276 if (sqe->len || sqe->buf_index)
1279 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1280 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1281 accept->flags = READ_ONCE(sqe->accept_flags);
1282 accept->nofile = rlimit(RLIMIT_NOFILE);
1283 flags = READ_ONCE(sqe->ioprio);
1284 if (flags & ~IORING_ACCEPT_MULTISHOT)
1287 accept->file_slot = READ_ONCE(sqe->file_index);
1288 if (accept->file_slot) {
1289 if (accept->flags & SOCK_CLOEXEC)
1291 if (flags & IORING_ACCEPT_MULTISHOT &&
1292 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1295 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1297 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1298 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1299 if (flags & IORING_ACCEPT_MULTISHOT)
1300 req->flags |= REQ_F_APOLL_MULTISHOT;
1304 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1306 struct io_ring_ctx *ctx = req->ctx;
1307 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1308 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1309 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1310 bool fixed = !!accept->file_slot;
1314 if (!io_check_multishot(req, issue_flags))
1318 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1319 if (unlikely(fd < 0))
1322 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1327 ret = PTR_ERR(file);
1328 if (ret == -EAGAIN && force_nonblock) {
1330 * if it's multishot and polled, we don't need to
1331 * return EAGAIN to arm the poll infra since it
1332 * has already been done
1334 if (issue_flags & IO_URING_F_MULTISHOT)
1335 ret = IOU_ISSUE_SKIP_COMPLETE;
1338 if (ret == -ERESTARTSYS)
1341 } else if (!fixed) {
1342 fd_install(fd, file);
1345 ret = io_fixed_fd_install(req, issue_flags, file,
1349 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1350 io_req_set_res(req, ret, 0);
1356 if (io_aux_cqe(ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
1357 req->cqe.user_data, ret, IORING_CQE_F_MORE, true))
1363 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1365 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1367 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1370 sock->domain = READ_ONCE(sqe->fd);
1371 sock->type = READ_ONCE(sqe->off);
1372 sock->protocol = READ_ONCE(sqe->len);
1373 sock->file_slot = READ_ONCE(sqe->file_index);
1374 sock->nofile = rlimit(RLIMIT_NOFILE);
1376 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1377 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1379 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1384 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1386 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1387 bool fixed = !!sock->file_slot;
1392 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1393 if (unlikely(fd < 0))
1396 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1400 ret = PTR_ERR(file);
1401 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1403 if (ret == -ERESTARTSYS)
1406 } else if (!fixed) {
1407 fd_install(fd, file);
1410 ret = io_fixed_fd_install(req, issue_flags, file,
1413 io_req_set_res(req, ret, 0);
1417 int io_connect_prep_async(struct io_kiocb *req)
1419 struct io_async_connect *io = req->async_data;
1420 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1422 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1425 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1427 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1429 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1432 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1433 conn->addr_len = READ_ONCE(sqe->addr2);
1434 conn->in_progress = false;
1438 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1440 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1441 struct io_async_connect __io, *io;
1442 unsigned file_flags;
1444 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1446 if (connect->in_progress) {
1447 struct socket *socket;
1450 socket = sock_from_file(req->file);
1452 ret = sock_error(socket->sk);
1456 if (req_has_async_data(req)) {
1457 io = req->async_data;
1459 ret = move_addr_to_kernel(connect->addr,
1467 file_flags = force_nonblock ? O_NONBLOCK : 0;
1469 ret = __sys_connect_file(req->file, &io->address,
1470 connect->addr_len, file_flags);
1471 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
1472 if (ret == -EINPROGRESS) {
1473 connect->in_progress = true;
1475 if (req_has_async_data(req))
1477 if (io_alloc_async_data(req)) {
1481 memcpy(req->async_data, &__io, sizeof(__io));
1485 if (ret == -ERESTARTSYS)
1490 io_req_set_res(req, ret, 0);
1494 void io_netmsg_cache_free(struct io_cache_entry *entry)
1496 kfree(container_of(entry, struct io_async_msghdr, cache));