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 static inline bool io_check_multishot(struct io_kiocb *req,
74 unsigned int issue_flags)
77 * When ->locked_cq is set we only allow to post CQEs from the original
78 * task context. Usual request completions will be handled in other
79 * generic paths but multipoll may decide to post extra cqes.
81 return !(issue_flags & IO_URING_F_IOWQ) ||
82 !(issue_flags & IO_URING_F_MULTISHOT) ||
83 !req->ctx->task_complete;
86 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
88 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
90 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
91 sqe->buf_index || sqe->splice_fd_in))
94 shutdown->how = READ_ONCE(sqe->len);
95 req->flags |= REQ_F_FORCE_ASYNC;
99 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
101 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
105 WARN_ON_ONCE(issue_flags & IO_URING_F_NONBLOCK);
107 sock = sock_from_file(req->file);
111 ret = __sys_shutdown_sock(sock, shutdown->how);
112 io_req_set_res(req, ret, 0);
116 static bool io_net_retry(struct socket *sock, int flags)
118 if (!(flags & MSG_WAITALL))
120 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
123 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
125 struct io_async_msghdr *hdr = req->async_data;
127 if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
130 /* Let normal cleanup path reap it if we fail adding to the cache */
131 if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
132 req->async_data = NULL;
133 req->flags &= ~REQ_F_ASYNC_DATA;
137 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
138 unsigned int issue_flags)
140 struct io_ring_ctx *ctx = req->ctx;
141 struct io_cache_entry *entry;
142 struct io_async_msghdr *hdr;
144 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
145 entry = io_alloc_cache_get(&ctx->netmsg_cache);
147 hdr = container_of(entry, struct io_async_msghdr, cache);
148 hdr->free_iov = NULL;
149 req->flags |= REQ_F_ASYNC_DATA;
150 req->async_data = hdr;
155 if (!io_alloc_async_data(req)) {
156 hdr = req->async_data;
157 hdr->free_iov = NULL;
163 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
165 /* ->prep_async is always called from the submission context */
166 return io_msg_alloc_async(req, 0);
169 static int io_setup_async_msg(struct io_kiocb *req,
170 struct io_async_msghdr *kmsg,
171 unsigned int issue_flags)
173 struct io_async_msghdr *async_msg;
175 if (req_has_async_data(req))
177 async_msg = io_msg_alloc_async(req, issue_flags);
179 kfree(kmsg->free_iov);
182 req->flags |= REQ_F_NEED_CLEANUP;
183 memcpy(async_msg, kmsg, sizeof(*kmsg));
184 if (async_msg->msg.msg_name)
185 async_msg->msg.msg_name = &async_msg->addr;
186 /* if were using fast_iov, set it to the new one */
187 if (iter_is_iovec(&kmsg->msg.msg_iter) && !kmsg->free_iov) {
188 size_t fast_idx = iter_iov(&kmsg->msg.msg_iter) - kmsg->fast_iov;
189 async_msg->msg.msg_iter.__iov = &async_msg->fast_iov[fast_idx];
195 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
196 struct io_async_msghdr *iomsg)
198 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
201 iomsg->msg.msg_name = &iomsg->addr;
202 iomsg->free_iov = iomsg->fast_iov;
203 ret = sendmsg_copy_msghdr(&iomsg->msg, sr->umsg, sr->msg_flags,
205 /* save msg_control as sys_sendmsg() overwrites it */
206 sr->msg_control = iomsg->msg.msg_control_user;
210 int io_send_prep_async(struct io_kiocb *req)
212 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
213 struct io_async_msghdr *io;
216 if (!zc->addr || req_has_async_data(req))
218 io = io_msg_alloc_async_prep(req);
221 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
225 static int io_setup_async_addr(struct io_kiocb *req,
226 struct sockaddr_storage *addr_storage,
227 unsigned int issue_flags)
229 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
230 struct io_async_msghdr *io;
232 if (!sr->addr || req_has_async_data(req))
234 io = io_msg_alloc_async(req, issue_flags);
237 memcpy(&io->addr, addr_storage, sizeof(io->addr));
241 int io_sendmsg_prep_async(struct io_kiocb *req)
245 if (!io_msg_alloc_async_prep(req))
247 ret = io_sendmsg_copy_hdr(req, req->async_data);
249 req->flags |= REQ_F_NEED_CLEANUP;
253 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
255 struct io_async_msghdr *io = req->async_data;
260 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
262 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
264 if (req->opcode == IORING_OP_SEND) {
265 if (READ_ONCE(sqe->__pad3[0]))
267 sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
268 sr->addr_len = READ_ONCE(sqe->addr_len);
269 } else if (sqe->addr2 || sqe->file_index) {
273 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
274 sr->len = READ_ONCE(sqe->len);
275 sr->flags = READ_ONCE(sqe->ioprio);
276 if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
278 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
279 if (sr->msg_flags & MSG_DONTWAIT)
280 req->flags |= REQ_F_NOWAIT;
283 if (req->ctx->compat)
284 sr->msg_flags |= MSG_CMSG_COMPAT;
290 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
292 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
293 struct io_async_msghdr iomsg, *kmsg;
299 sock = sock_from_file(req->file);
303 if (req_has_async_data(req)) {
304 kmsg = req->async_data;
305 kmsg->msg.msg_control_user = sr->msg_control;
307 ret = io_sendmsg_copy_hdr(req, &iomsg);
313 if (!(req->flags & REQ_F_POLLED) &&
314 (sr->flags & IORING_RECVSEND_POLL_FIRST))
315 return io_setup_async_msg(req, kmsg, issue_flags);
317 flags = sr->msg_flags;
318 if (issue_flags & IO_URING_F_NONBLOCK)
319 flags |= MSG_DONTWAIT;
320 if (flags & MSG_WAITALL)
321 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
323 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
326 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
327 return io_setup_async_msg(req, kmsg, issue_flags);
328 if (ret > 0 && io_net_retry(sock, flags)) {
329 kmsg->msg.msg_controllen = 0;
330 kmsg->msg.msg_control = NULL;
332 req->flags |= REQ_F_PARTIAL_IO;
333 return io_setup_async_msg(req, kmsg, issue_flags);
335 if (ret == -ERESTARTSYS)
339 /* fast path, check for non-NULL to avoid function call */
341 kfree(kmsg->free_iov);
342 req->flags &= ~REQ_F_NEED_CLEANUP;
343 io_netmsg_recycle(req, issue_flags);
346 else if (sr->done_io)
348 io_req_set_res(req, ret, 0);
352 int io_send(struct io_kiocb *req, unsigned int issue_flags)
354 struct sockaddr_storage __address;
355 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
363 msg.msg_control = NULL;
364 msg.msg_controllen = 0;
369 if (req_has_async_data(req)) {
370 struct io_async_msghdr *io = req->async_data;
372 msg.msg_name = &io->addr;
374 ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
375 if (unlikely(ret < 0))
377 msg.msg_name = (struct sockaddr *)&__address;
379 msg.msg_namelen = sr->addr_len;
382 if (!(req->flags & REQ_F_POLLED) &&
383 (sr->flags & IORING_RECVSEND_POLL_FIRST))
384 return io_setup_async_addr(req, &__address, issue_flags);
386 sock = sock_from_file(req->file);
390 ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len, &msg.msg_iter);
394 flags = sr->msg_flags;
395 if (issue_flags & IO_URING_F_NONBLOCK)
396 flags |= MSG_DONTWAIT;
397 if (flags & MSG_WAITALL)
398 min_ret = iov_iter_count(&msg.msg_iter);
400 flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
401 msg.msg_flags = flags;
402 ret = sock_sendmsg(sock, &msg);
404 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
405 return io_setup_async_addr(req, &__address, issue_flags);
407 if (ret > 0 && io_net_retry(sock, flags)) {
411 req->flags |= REQ_F_PARTIAL_IO;
412 return io_setup_async_addr(req, &__address, issue_flags);
414 if (ret == -ERESTARTSYS)
420 else if (sr->done_io)
422 io_req_set_res(req, ret, 0);
426 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
430 if (iomsg->namelen < 0)
432 if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
433 iomsg->namelen, &hdr))
435 if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
441 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
442 struct io_async_msghdr *iomsg)
444 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
445 struct user_msghdr msg;
448 if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
451 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
455 if (req->flags & REQ_F_BUFFER_SELECT) {
456 if (msg.msg_iovlen == 0) {
457 sr->len = iomsg->fast_iov[0].iov_len = 0;
458 iomsg->fast_iov[0].iov_base = NULL;
459 iomsg->free_iov = NULL;
460 } else if (msg.msg_iovlen > 1) {
463 if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
465 sr->len = iomsg->fast_iov[0].iov_len;
466 iomsg->free_iov = NULL;
469 if (req->flags & REQ_F_APOLL_MULTISHOT) {
470 iomsg->namelen = msg.msg_namelen;
471 iomsg->controllen = msg.msg_controllen;
472 if (io_recvmsg_multishot_overflow(iomsg))
476 iomsg->free_iov = iomsg->fast_iov;
477 ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
478 &iomsg->free_iov, &iomsg->msg.msg_iter,
488 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
489 struct io_async_msghdr *iomsg)
491 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
492 struct compat_msghdr msg;
493 struct compat_iovec __user *uiov;
496 if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
499 ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
503 uiov = compat_ptr(msg.msg_iov);
504 if (req->flags & REQ_F_BUFFER_SELECT) {
507 iomsg->free_iov = NULL;
508 if (msg.msg_iovlen == 0) {
510 } else if (msg.msg_iovlen > 1) {
513 if (!access_ok(uiov, sizeof(*uiov)))
515 if (__get_user(clen, &uiov->iov_len))
522 if (req->flags & REQ_F_APOLL_MULTISHOT) {
523 iomsg->namelen = msg.msg_namelen;
524 iomsg->controllen = msg.msg_controllen;
525 if (io_recvmsg_multishot_overflow(iomsg))
529 iomsg->free_iov = iomsg->fast_iov;
530 ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
531 UIO_FASTIOV, &iomsg->free_iov,
532 &iomsg->msg.msg_iter, true);
541 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
542 struct io_async_msghdr *iomsg)
544 iomsg->msg.msg_name = &iomsg->addr;
547 if (req->ctx->compat)
548 return __io_compat_recvmsg_copy_hdr(req, iomsg);
551 return __io_recvmsg_copy_hdr(req, iomsg);
554 int io_recvmsg_prep_async(struct io_kiocb *req)
558 if (!io_msg_alloc_async_prep(req))
560 ret = io_recvmsg_copy_hdr(req, req->async_data);
562 req->flags |= REQ_F_NEED_CLEANUP;
566 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
568 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
570 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
572 if (unlikely(sqe->file_index || sqe->addr2))
575 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
576 sr->len = READ_ONCE(sqe->len);
577 sr->flags = READ_ONCE(sqe->ioprio);
578 if (sr->flags & ~(RECVMSG_FLAGS))
580 sr->msg_flags = READ_ONCE(sqe->msg_flags);
581 if (sr->msg_flags & MSG_DONTWAIT)
582 req->flags |= REQ_F_NOWAIT;
583 if (sr->msg_flags & MSG_ERRQUEUE)
584 req->flags |= REQ_F_CLEAR_POLLIN;
585 if (sr->flags & IORING_RECV_MULTISHOT) {
586 if (!(req->flags & REQ_F_BUFFER_SELECT))
588 if (sr->msg_flags & MSG_WAITALL)
590 if (req->opcode == IORING_OP_RECV && sr->len)
592 req->flags |= REQ_F_APOLL_MULTISHOT;
594 * Store the buffer group for this multishot receive separately,
595 * as if we end up doing an io-wq based issue that selects a
596 * buffer, it has to be committed immediately and that will
597 * clear ->buf_list. This means we lose the link to the buffer
598 * list, and the eventual buffer put on completion then cannot
601 sr->buf_group = req->buf_index;
605 if (req->ctx->compat)
606 sr->msg_flags |= MSG_CMSG_COMPAT;
612 static inline void io_recv_prep_retry(struct io_kiocb *req)
614 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
617 sr->len = 0; /* get from the provided buffer */
618 req->buf_index = sr->buf_group;
622 * Finishes io_recv and io_recvmsg.
624 * Returns true if it is actually finished, or false if it should run
625 * again (for multishot).
627 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
628 unsigned int cflags, bool mshot_finished,
629 unsigned issue_flags)
631 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
632 io_req_set_res(req, *ret, cflags);
637 if (!mshot_finished) {
638 if (io_aux_cqe(req->ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
639 req->cqe.user_data, *ret, cflags | IORING_CQE_F_MORE, true)) {
640 io_recv_prep_retry(req);
643 /* Otherwise stop multishot but use the current result. */
646 io_req_set_res(req, *ret, cflags);
648 if (issue_flags & IO_URING_F_MULTISHOT)
649 *ret = IOU_STOP_MULTISHOT;
655 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
656 struct io_sr_msg *sr, void __user **buf,
659 unsigned long ubuf = (unsigned long) *buf;
662 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
667 if (kmsg->controllen) {
668 unsigned long control = ubuf + hdr - kmsg->controllen;
670 kmsg->msg.msg_control_user = (void __user *) control;
671 kmsg->msg.msg_controllen = kmsg->controllen;
674 sr->buf = *buf; /* stash for later copy */
675 *buf = (void __user *) (ubuf + hdr);
676 kmsg->payloadlen = *len = *len - hdr;
680 struct io_recvmsg_multishot_hdr {
681 struct io_uring_recvmsg_out msg;
682 struct sockaddr_storage addr;
685 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
686 struct io_async_msghdr *kmsg,
687 unsigned int flags, bool *finished)
691 struct io_recvmsg_multishot_hdr hdr;
694 kmsg->msg.msg_name = &hdr.addr;
695 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
696 kmsg->msg.msg_namelen = 0;
698 if (sock->file->f_flags & O_NONBLOCK)
699 flags |= MSG_DONTWAIT;
701 err = sock_recvmsg(sock, &kmsg->msg, flags);
702 *finished = err <= 0;
706 hdr.msg = (struct io_uring_recvmsg_out) {
707 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
708 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
711 hdr.msg.payloadlen = err;
712 if (err > kmsg->payloadlen)
713 err = kmsg->payloadlen;
715 copy_len = sizeof(struct io_uring_recvmsg_out);
716 if (kmsg->msg.msg_namelen > kmsg->namelen)
717 copy_len += kmsg->namelen;
719 copy_len += kmsg->msg.msg_namelen;
722 * "fromlen shall refer to the value before truncation.."
725 hdr.msg.namelen = kmsg->msg.msg_namelen;
727 /* ensure that there is no gap between hdr and sockaddr_storage */
728 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
729 sizeof(struct io_uring_recvmsg_out));
730 if (copy_to_user(io->buf, &hdr, copy_len)) {
735 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
736 kmsg->controllen + err;
739 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
741 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
742 struct io_async_msghdr iomsg, *kmsg;
746 int ret, min_ret = 0;
747 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
748 bool mshot_finished = true;
750 sock = sock_from_file(req->file);
754 if (req_has_async_data(req)) {
755 kmsg = req->async_data;
757 ret = io_recvmsg_copy_hdr(req, &iomsg);
763 if (!(req->flags & REQ_F_POLLED) &&
764 (sr->flags & IORING_RECVSEND_POLL_FIRST))
765 return io_setup_async_msg(req, kmsg, issue_flags);
767 if (!io_check_multishot(req, issue_flags))
768 return io_setup_async_msg(req, kmsg, issue_flags);
771 if (io_do_buffer_select(req)) {
773 size_t len = sr->len;
775 buf = io_buffer_select(req, &len, issue_flags);
779 if (req->flags & REQ_F_APOLL_MULTISHOT) {
780 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
782 io_kbuf_recycle(req, issue_flags);
787 iov_iter_ubuf(&kmsg->msg.msg_iter, ITER_DEST, buf, len);
790 flags = sr->msg_flags;
792 flags |= MSG_DONTWAIT;
794 kmsg->msg.msg_get_inq = 1;
795 if (req->flags & REQ_F_APOLL_MULTISHOT) {
796 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
799 /* disable partial retry for recvmsg with cmsg attached */
800 if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
801 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
803 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
808 if (ret == -EAGAIN && force_nonblock) {
809 ret = io_setup_async_msg(req, kmsg, issue_flags);
810 if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
811 io_kbuf_recycle(req, issue_flags);
812 return IOU_ISSUE_SKIP_COMPLETE;
816 if (ret > 0 && io_net_retry(sock, flags)) {
818 req->flags |= REQ_F_PARTIAL_IO;
819 return io_setup_async_msg(req, kmsg, issue_flags);
821 if (ret == -ERESTARTSYS)
824 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
830 else if (sr->done_io)
833 io_kbuf_recycle(req, issue_flags);
835 cflags = io_put_kbuf(req, issue_flags);
836 if (kmsg->msg.msg_inq)
837 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
839 if (!io_recv_finish(req, &ret, cflags, mshot_finished, issue_flags))
840 goto retry_multishot;
842 if (mshot_finished) {
843 /* fast path, check for non-NULL to avoid function call */
845 kfree(kmsg->free_iov);
846 io_netmsg_recycle(req, issue_flags);
847 req->flags &= ~REQ_F_NEED_CLEANUP;
853 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
855 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
860 int ret, min_ret = 0;
861 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
862 size_t len = sr->len;
864 if (!(req->flags & REQ_F_POLLED) &&
865 (sr->flags & IORING_RECVSEND_POLL_FIRST))
868 if (!io_check_multishot(req, issue_flags))
871 sock = sock_from_file(req->file);
876 if (io_do_buffer_select(req)) {
879 buf = io_buffer_select(req, &len, issue_flags);
885 ret = import_ubuf(ITER_DEST, sr->buf, len, &msg.msg_iter);
891 msg.msg_control = NULL;
894 msg.msg_controllen = 0;
898 flags = sr->msg_flags;
900 flags |= MSG_DONTWAIT;
901 if (flags & MSG_WAITALL)
902 min_ret = iov_iter_count(&msg.msg_iter);
904 ret = sock_recvmsg(sock, &msg, flags);
906 if (ret == -EAGAIN && force_nonblock) {
907 if (issue_flags & IO_URING_F_MULTISHOT) {
908 io_kbuf_recycle(req, issue_flags);
909 return IOU_ISSUE_SKIP_COMPLETE;
914 if (ret > 0 && io_net_retry(sock, flags)) {
918 req->flags |= REQ_F_PARTIAL_IO;
921 if (ret == -ERESTARTSYS)
924 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
931 else if (sr->done_io)
934 io_kbuf_recycle(req, issue_flags);
936 cflags = io_put_kbuf(req, issue_flags);
938 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
940 if (!io_recv_finish(req, &ret, cflags, ret <= 0, issue_flags))
941 goto retry_multishot;
946 void io_send_zc_cleanup(struct io_kiocb *req)
948 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
949 struct io_async_msghdr *io;
951 if (req_has_async_data(req)) {
952 io = req->async_data;
953 /* might be ->fast_iov if *msg_copy_hdr failed */
954 if (io->free_iov != io->fast_iov)
958 io_notif_flush(zc->notif);
963 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
964 #define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
966 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
968 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
969 struct io_ring_ctx *ctx = req->ctx;
970 struct io_kiocb *notif;
972 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
974 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
975 if (req->flags & REQ_F_CQE_SKIP)
978 notif = zc->notif = io_alloc_notif(ctx);
981 notif->cqe.user_data = req->cqe.user_data;
983 notif->cqe.flags = IORING_CQE_F_NOTIF;
984 req->flags |= REQ_F_NEED_CLEANUP;
986 zc->flags = READ_ONCE(sqe->ioprio);
987 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
988 if (zc->flags & ~IO_ZC_FLAGS_VALID)
990 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
991 io_notif_set_extended(notif);
992 io_notif_to_data(notif)->zc_report = true;
996 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
997 unsigned idx = READ_ONCE(sqe->buf_index);
999 if (unlikely(idx >= ctx->nr_user_bufs))
1001 idx = array_index_nospec(idx, ctx->nr_user_bufs);
1002 req->imu = READ_ONCE(ctx->user_bufs[idx]);
1003 io_req_set_rsrc_node(notif, ctx, 0);
1006 if (req->opcode == IORING_OP_SEND_ZC) {
1007 if (READ_ONCE(sqe->__pad3[0]))
1009 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1010 zc->addr_len = READ_ONCE(sqe->addr_len);
1012 if (unlikely(sqe->addr2 || sqe->file_index))
1014 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1018 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1019 zc->len = READ_ONCE(sqe->len);
1020 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1021 if (zc->msg_flags & MSG_DONTWAIT)
1022 req->flags |= REQ_F_NOWAIT;
1026 #ifdef CONFIG_COMPAT
1027 if (req->ctx->compat)
1028 zc->msg_flags |= MSG_CMSG_COMPAT;
1033 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1034 struct iov_iter *from, size_t length)
1036 skb_zcopy_downgrade_managed(skb);
1037 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1040 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1041 struct iov_iter *from, size_t length)
1043 struct skb_shared_info *shinfo = skb_shinfo(skb);
1044 int frag = shinfo->nr_frags;
1046 struct bvec_iter bi;
1048 unsigned long truesize = 0;
1051 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1052 else if (unlikely(!skb_zcopy_managed(skb)))
1053 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1055 bi.bi_size = min(from->count, length);
1056 bi.bi_bvec_done = from->iov_offset;
1059 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1060 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1063 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1064 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1065 v.bv_offset, v.bv_len);
1066 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1071 shinfo->nr_frags = frag;
1072 from->bvec += bi.bi_idx;
1073 from->nr_segs -= bi.bi_idx;
1074 from->count -= copied;
1075 from->iov_offset = bi.bi_bvec_done;
1077 skb->data_len += copied;
1079 skb->truesize += truesize;
1081 if (sk && sk->sk_type == SOCK_STREAM) {
1082 sk_wmem_queued_add(sk, truesize);
1083 if (!skb_zcopy_pure(skb))
1084 sk_mem_charge(sk, truesize);
1086 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1091 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1093 struct sockaddr_storage __address;
1094 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1096 struct socket *sock;
1098 int ret, min_ret = 0;
1100 sock = sock_from_file(req->file);
1101 if (unlikely(!sock))
1103 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1106 msg.msg_name = NULL;
1107 msg.msg_control = NULL;
1108 msg.msg_controllen = 0;
1109 msg.msg_namelen = 0;
1112 if (req_has_async_data(req)) {
1113 struct io_async_msghdr *io = req->async_data;
1115 msg.msg_name = &io->addr;
1117 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1118 if (unlikely(ret < 0))
1120 msg.msg_name = (struct sockaddr *)&__address;
1122 msg.msg_namelen = zc->addr_len;
1125 if (!(req->flags & REQ_F_POLLED) &&
1126 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1127 return io_setup_async_addr(req, &__address, issue_flags);
1129 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1130 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1131 (u64)(uintptr_t)zc->buf, zc->len);
1134 msg.sg_from_iter = io_sg_from_iter;
1136 io_notif_set_extended(zc->notif);
1137 ret = import_ubuf(ITER_SOURCE, zc->buf, zc->len, &msg.msg_iter);
1140 ret = io_notif_account_mem(zc->notif, zc->len);
1143 msg.sg_from_iter = io_sg_from_iter_iovec;
1146 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1147 if (issue_flags & IO_URING_F_NONBLOCK)
1148 msg_flags |= MSG_DONTWAIT;
1149 if (msg_flags & MSG_WAITALL)
1150 min_ret = iov_iter_count(&msg.msg_iter);
1151 msg_flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
1153 msg.msg_flags = msg_flags;
1154 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1155 ret = sock_sendmsg(sock, &msg);
1157 if (unlikely(ret < min_ret)) {
1158 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1159 return io_setup_async_addr(req, &__address, issue_flags);
1161 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1165 req->flags |= REQ_F_PARTIAL_IO;
1166 return io_setup_async_addr(req, &__address, issue_flags);
1168 if (ret == -ERESTARTSYS)
1175 else if (zc->done_io)
1179 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1180 * flushing notif to io_send_zc_cleanup()
1182 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1183 io_notif_flush(zc->notif);
1184 req->flags &= ~REQ_F_NEED_CLEANUP;
1186 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1190 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1192 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1193 struct io_async_msghdr iomsg, *kmsg;
1194 struct socket *sock;
1196 int ret, min_ret = 0;
1198 io_notif_set_extended(sr->notif);
1200 sock = sock_from_file(req->file);
1201 if (unlikely(!sock))
1203 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1206 if (req_has_async_data(req)) {
1207 kmsg = req->async_data;
1209 ret = io_sendmsg_copy_hdr(req, &iomsg);
1215 if (!(req->flags & REQ_F_POLLED) &&
1216 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1217 return io_setup_async_msg(req, kmsg, issue_flags);
1219 flags = sr->msg_flags | MSG_ZEROCOPY;
1220 if (issue_flags & IO_URING_F_NONBLOCK)
1221 flags |= MSG_DONTWAIT;
1222 if (flags & MSG_WAITALL)
1223 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1225 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1226 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1227 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1229 if (unlikely(ret < min_ret)) {
1230 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1231 return io_setup_async_msg(req, kmsg, issue_flags);
1233 if (ret > 0 && io_net_retry(sock, flags)) {
1235 req->flags |= REQ_F_PARTIAL_IO;
1236 return io_setup_async_msg(req, kmsg, issue_flags);
1238 if (ret == -ERESTARTSYS)
1242 /* fast path, check for non-NULL to avoid function call */
1243 if (kmsg->free_iov) {
1244 kfree(kmsg->free_iov);
1245 kmsg->free_iov = NULL;
1248 io_netmsg_recycle(req, issue_flags);
1251 else if (sr->done_io)
1255 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1256 * flushing notif to io_send_zc_cleanup()
1258 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1259 io_notif_flush(sr->notif);
1260 req->flags &= ~REQ_F_NEED_CLEANUP;
1262 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1266 void io_sendrecv_fail(struct io_kiocb *req)
1268 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1270 if (req->flags & REQ_F_PARTIAL_IO)
1271 req->cqe.res = sr->done_io;
1273 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1274 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1275 req->cqe.flags |= IORING_CQE_F_MORE;
1278 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1280 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1283 if (sqe->len || sqe->buf_index)
1286 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1287 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1288 accept->flags = READ_ONCE(sqe->accept_flags);
1289 accept->nofile = rlimit(RLIMIT_NOFILE);
1290 flags = READ_ONCE(sqe->ioprio);
1291 if (flags & ~IORING_ACCEPT_MULTISHOT)
1294 accept->file_slot = READ_ONCE(sqe->file_index);
1295 if (accept->file_slot) {
1296 if (accept->flags & SOCK_CLOEXEC)
1298 if (flags & IORING_ACCEPT_MULTISHOT &&
1299 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1302 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1304 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1305 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1306 if (flags & IORING_ACCEPT_MULTISHOT)
1307 req->flags |= REQ_F_APOLL_MULTISHOT;
1311 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1313 struct io_ring_ctx *ctx = req->ctx;
1314 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1315 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1316 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1317 bool fixed = !!accept->file_slot;
1321 if (!io_check_multishot(req, issue_flags))
1325 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1326 if (unlikely(fd < 0))
1329 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1334 ret = PTR_ERR(file);
1335 if (ret == -EAGAIN && force_nonblock) {
1337 * if it's multishot and polled, we don't need to
1338 * return EAGAIN to arm the poll infra since it
1339 * has already been done
1341 if (issue_flags & IO_URING_F_MULTISHOT)
1342 ret = IOU_ISSUE_SKIP_COMPLETE;
1345 if (ret == -ERESTARTSYS)
1348 } else if (!fixed) {
1349 fd_install(fd, file);
1352 ret = io_fixed_fd_install(req, issue_flags, file,
1356 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1357 io_req_set_res(req, ret, 0);
1363 if (io_aux_cqe(ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
1364 req->cqe.user_data, ret, IORING_CQE_F_MORE, true))
1370 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1372 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1374 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1377 sock->domain = READ_ONCE(sqe->fd);
1378 sock->type = READ_ONCE(sqe->off);
1379 sock->protocol = READ_ONCE(sqe->len);
1380 sock->file_slot = READ_ONCE(sqe->file_index);
1381 sock->nofile = rlimit(RLIMIT_NOFILE);
1383 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1384 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1386 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1391 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1393 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1394 bool fixed = !!sock->file_slot;
1399 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1400 if (unlikely(fd < 0))
1403 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1407 ret = PTR_ERR(file);
1408 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1410 if (ret == -ERESTARTSYS)
1413 } else if (!fixed) {
1414 fd_install(fd, file);
1417 ret = io_fixed_fd_install(req, issue_flags, file,
1420 io_req_set_res(req, ret, 0);
1424 int io_connect_prep_async(struct io_kiocb *req)
1426 struct io_async_connect *io = req->async_data;
1427 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1429 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1432 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1434 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1436 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1439 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1440 conn->addr_len = READ_ONCE(sqe->addr2);
1441 conn->in_progress = conn->seen_econnaborted = false;
1445 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1447 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1448 struct io_async_connect __io, *io;
1449 unsigned file_flags;
1451 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1453 if (connect->in_progress) {
1454 struct socket *socket;
1457 socket = sock_from_file(req->file);
1459 ret = sock_error(socket->sk);
1463 if (req_has_async_data(req)) {
1464 io = req->async_data;
1466 ret = move_addr_to_kernel(connect->addr,
1474 file_flags = force_nonblock ? O_NONBLOCK : 0;
1476 ret = __sys_connect_file(req->file, &io->address,
1477 connect->addr_len, file_flags);
1478 if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1479 && force_nonblock) {
1480 if (ret == -EINPROGRESS) {
1481 connect->in_progress = true;
1484 if (ret == -ECONNABORTED) {
1485 if (connect->seen_econnaborted)
1487 connect->seen_econnaborted = true;
1489 if (req_has_async_data(req))
1491 if (io_alloc_async_data(req)) {
1495 memcpy(req->async_data, &__io, sizeof(__io));
1498 if (ret == -ERESTARTSYS)
1503 io_req_set_res(req, ret, 0);
1507 void io_netmsg_cache_free(struct io_cache_entry *entry)
1509 kfree(container_of(entry, struct io_async_msghdr, cache));