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 msg.msg_flags = flags;
401 ret = sock_sendmsg(sock, &msg);
403 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
404 return io_setup_async_addr(req, &__address, issue_flags);
406 if (ret > 0 && io_net_retry(sock, flags)) {
410 req->flags |= REQ_F_PARTIAL_IO;
411 return io_setup_async_addr(req, &__address, issue_flags);
413 if (ret == -ERESTARTSYS)
419 else if (sr->done_io)
421 io_req_set_res(req, ret, 0);
425 static bool io_recvmsg_multishot_overflow(struct io_async_msghdr *iomsg)
429 if (iomsg->namelen < 0)
431 if (check_add_overflow((int)sizeof(struct io_uring_recvmsg_out),
432 iomsg->namelen, &hdr))
434 if (check_add_overflow(hdr, (int)iomsg->controllen, &hdr))
440 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
441 struct io_async_msghdr *iomsg)
443 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
444 struct user_msghdr msg;
447 if (copy_from_user(&msg, sr->umsg, sizeof(*sr->umsg)))
450 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
454 if (req->flags & REQ_F_BUFFER_SELECT) {
455 if (msg.msg_iovlen == 0) {
456 sr->len = iomsg->fast_iov[0].iov_len = 0;
457 iomsg->fast_iov[0].iov_base = NULL;
458 iomsg->free_iov = NULL;
459 } else if (msg.msg_iovlen > 1) {
462 if (copy_from_user(iomsg->fast_iov, msg.msg_iov, sizeof(*msg.msg_iov)))
464 sr->len = iomsg->fast_iov[0].iov_len;
465 iomsg->free_iov = NULL;
468 if (req->flags & REQ_F_APOLL_MULTISHOT) {
469 iomsg->namelen = msg.msg_namelen;
470 iomsg->controllen = msg.msg_controllen;
471 if (io_recvmsg_multishot_overflow(iomsg))
475 iomsg->free_iov = iomsg->fast_iov;
476 ret = __import_iovec(ITER_DEST, msg.msg_iov, msg.msg_iovlen, UIO_FASTIOV,
477 &iomsg->free_iov, &iomsg->msg.msg_iter,
487 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
488 struct io_async_msghdr *iomsg)
490 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
491 struct compat_msghdr msg;
492 struct compat_iovec __user *uiov;
495 if (copy_from_user(&msg, sr->umsg_compat, sizeof(msg)))
498 ret = __get_compat_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
502 uiov = compat_ptr(msg.msg_iov);
503 if (req->flags & REQ_F_BUFFER_SELECT) {
506 iomsg->free_iov = NULL;
507 if (msg.msg_iovlen == 0) {
509 } else if (msg.msg_iovlen > 1) {
512 if (!access_ok(uiov, sizeof(*uiov)))
514 if (__get_user(clen, &uiov->iov_len))
521 if (req->flags & REQ_F_APOLL_MULTISHOT) {
522 iomsg->namelen = msg.msg_namelen;
523 iomsg->controllen = msg.msg_controllen;
524 if (io_recvmsg_multishot_overflow(iomsg))
528 iomsg->free_iov = iomsg->fast_iov;
529 ret = __import_iovec(ITER_DEST, (struct iovec __user *)uiov, msg.msg_iovlen,
530 UIO_FASTIOV, &iomsg->free_iov,
531 &iomsg->msg.msg_iter, true);
540 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
541 struct io_async_msghdr *iomsg)
543 iomsg->msg.msg_name = &iomsg->addr;
546 if (req->ctx->compat)
547 return __io_compat_recvmsg_copy_hdr(req, iomsg);
550 return __io_recvmsg_copy_hdr(req, iomsg);
553 int io_recvmsg_prep_async(struct io_kiocb *req)
557 if (!io_msg_alloc_async_prep(req))
559 ret = io_recvmsg_copy_hdr(req, req->async_data);
561 req->flags |= REQ_F_NEED_CLEANUP;
565 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
567 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
569 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
571 if (unlikely(sqe->file_index || sqe->addr2))
574 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
575 sr->len = READ_ONCE(sqe->len);
576 sr->flags = READ_ONCE(sqe->ioprio);
577 if (sr->flags & ~(RECVMSG_FLAGS))
579 sr->msg_flags = READ_ONCE(sqe->msg_flags);
580 if (sr->msg_flags & MSG_DONTWAIT)
581 req->flags |= REQ_F_NOWAIT;
582 if (sr->msg_flags & MSG_ERRQUEUE)
583 req->flags |= REQ_F_CLEAR_POLLIN;
584 if (sr->flags & IORING_RECV_MULTISHOT) {
585 if (!(req->flags & REQ_F_BUFFER_SELECT))
587 if (sr->msg_flags & MSG_WAITALL)
589 if (req->opcode == IORING_OP_RECV && sr->len)
591 req->flags |= REQ_F_APOLL_MULTISHOT;
593 * Store the buffer group for this multishot receive separately,
594 * as if we end up doing an io-wq based issue that selects a
595 * buffer, it has to be committed immediately and that will
596 * clear ->buf_list. This means we lose the link to the buffer
597 * list, and the eventual buffer put on completion then cannot
600 sr->buf_group = req->buf_index;
604 if (req->ctx->compat)
605 sr->msg_flags |= MSG_CMSG_COMPAT;
611 static inline void io_recv_prep_retry(struct io_kiocb *req)
613 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
616 sr->len = 0; /* get from the provided buffer */
617 req->buf_index = sr->buf_group;
621 * Finishes io_recv and io_recvmsg.
623 * Returns true if it is actually finished, or false if it should run
624 * again (for multishot).
626 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
627 unsigned int cflags, bool mshot_finished,
628 unsigned issue_flags)
630 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
631 io_req_set_res(req, *ret, cflags);
636 if (!mshot_finished) {
637 if (io_aux_cqe(req->ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
638 req->cqe.user_data, *ret, cflags | IORING_CQE_F_MORE, true)) {
639 io_recv_prep_retry(req);
642 /* Otherwise stop multishot but use the current result. */
645 io_req_set_res(req, *ret, cflags);
647 if (issue_flags & IO_URING_F_MULTISHOT)
648 *ret = IOU_STOP_MULTISHOT;
654 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
655 struct io_sr_msg *sr, void __user **buf,
658 unsigned long ubuf = (unsigned long) *buf;
661 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
666 if (kmsg->controllen) {
667 unsigned long control = ubuf + hdr - kmsg->controllen;
669 kmsg->msg.msg_control_user = (void __user *) control;
670 kmsg->msg.msg_controllen = kmsg->controllen;
673 sr->buf = *buf; /* stash for later copy */
674 *buf = (void __user *) (ubuf + hdr);
675 kmsg->payloadlen = *len = *len - hdr;
679 struct io_recvmsg_multishot_hdr {
680 struct io_uring_recvmsg_out msg;
681 struct sockaddr_storage addr;
684 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
685 struct io_async_msghdr *kmsg,
686 unsigned int flags, bool *finished)
690 struct io_recvmsg_multishot_hdr hdr;
693 kmsg->msg.msg_name = &hdr.addr;
694 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
695 kmsg->msg.msg_namelen = 0;
697 if (sock->file->f_flags & O_NONBLOCK)
698 flags |= MSG_DONTWAIT;
700 err = sock_recvmsg(sock, &kmsg->msg, flags);
701 *finished = err <= 0;
705 hdr.msg = (struct io_uring_recvmsg_out) {
706 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
707 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
710 hdr.msg.payloadlen = err;
711 if (err > kmsg->payloadlen)
712 err = kmsg->payloadlen;
714 copy_len = sizeof(struct io_uring_recvmsg_out);
715 if (kmsg->msg.msg_namelen > kmsg->namelen)
716 copy_len += kmsg->namelen;
718 copy_len += kmsg->msg.msg_namelen;
721 * "fromlen shall refer to the value before truncation.."
724 hdr.msg.namelen = kmsg->msg.msg_namelen;
726 /* ensure that there is no gap between hdr and sockaddr_storage */
727 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
728 sizeof(struct io_uring_recvmsg_out));
729 if (copy_to_user(io->buf, &hdr, copy_len)) {
734 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
735 kmsg->controllen + err;
738 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
740 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
741 struct io_async_msghdr iomsg, *kmsg;
745 int ret, min_ret = 0;
746 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
747 bool mshot_finished = true;
749 sock = sock_from_file(req->file);
753 if (req_has_async_data(req)) {
754 kmsg = req->async_data;
756 ret = io_recvmsg_copy_hdr(req, &iomsg);
762 if (!(req->flags & REQ_F_POLLED) &&
763 (sr->flags & IORING_RECVSEND_POLL_FIRST))
764 return io_setup_async_msg(req, kmsg, issue_flags);
766 if (!io_check_multishot(req, issue_flags))
767 return io_setup_async_msg(req, kmsg, issue_flags);
770 if (io_do_buffer_select(req)) {
772 size_t len = sr->len;
774 buf = io_buffer_select(req, &len, issue_flags);
778 if (req->flags & REQ_F_APOLL_MULTISHOT) {
779 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
781 io_kbuf_recycle(req, issue_flags);
786 iov_iter_ubuf(&kmsg->msg.msg_iter, ITER_DEST, buf, len);
789 flags = sr->msg_flags;
791 flags |= MSG_DONTWAIT;
793 kmsg->msg.msg_get_inq = 1;
794 if (req->flags & REQ_F_APOLL_MULTISHOT) {
795 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
798 /* disable partial retry for recvmsg with cmsg attached */
799 if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
800 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
802 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
807 if (ret == -EAGAIN && force_nonblock) {
808 ret = io_setup_async_msg(req, kmsg, issue_flags);
809 if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
810 io_kbuf_recycle(req, issue_flags);
811 return IOU_ISSUE_SKIP_COMPLETE;
815 if (ret > 0 && io_net_retry(sock, flags)) {
817 req->flags |= REQ_F_PARTIAL_IO;
818 return io_setup_async_msg(req, kmsg, issue_flags);
820 if (ret == -ERESTARTSYS)
823 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
829 else if (sr->done_io)
832 io_kbuf_recycle(req, issue_flags);
834 cflags = io_put_kbuf(req, issue_flags);
835 if (kmsg->msg.msg_inq)
836 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
838 if (!io_recv_finish(req, &ret, cflags, mshot_finished, issue_flags))
839 goto retry_multishot;
841 if (mshot_finished) {
842 /* fast path, check for non-NULL to avoid function call */
844 kfree(kmsg->free_iov);
845 io_netmsg_recycle(req, issue_flags);
846 req->flags &= ~REQ_F_NEED_CLEANUP;
852 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
854 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
859 int ret, min_ret = 0;
860 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
861 size_t len = sr->len;
863 if (!(req->flags & REQ_F_POLLED) &&
864 (sr->flags & IORING_RECVSEND_POLL_FIRST))
867 if (!io_check_multishot(req, issue_flags))
870 sock = sock_from_file(req->file);
875 if (io_do_buffer_select(req)) {
878 buf = io_buffer_select(req, &len, issue_flags);
884 ret = import_ubuf(ITER_DEST, sr->buf, len, &msg.msg_iter);
890 msg.msg_control = NULL;
893 msg.msg_controllen = 0;
897 flags = sr->msg_flags;
899 flags |= MSG_DONTWAIT;
900 if (flags & MSG_WAITALL)
901 min_ret = iov_iter_count(&msg.msg_iter);
903 ret = sock_recvmsg(sock, &msg, flags);
905 if (ret == -EAGAIN && force_nonblock) {
906 if (issue_flags & IO_URING_F_MULTISHOT) {
907 io_kbuf_recycle(req, issue_flags);
908 return IOU_ISSUE_SKIP_COMPLETE;
913 if (ret > 0 && io_net_retry(sock, flags)) {
917 req->flags |= REQ_F_PARTIAL_IO;
920 if (ret == -ERESTARTSYS)
923 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
930 else if (sr->done_io)
933 io_kbuf_recycle(req, issue_flags);
935 cflags = io_put_kbuf(req, issue_flags);
937 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
939 if (!io_recv_finish(req, &ret, cflags, ret <= 0, issue_flags))
940 goto retry_multishot;
945 void io_send_zc_cleanup(struct io_kiocb *req)
947 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
948 struct io_async_msghdr *io;
950 if (req_has_async_data(req)) {
951 io = req->async_data;
952 /* might be ->fast_iov if *msg_copy_hdr failed */
953 if (io->free_iov != io->fast_iov)
957 io_notif_flush(zc->notif);
962 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
963 #define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
965 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
967 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
968 struct io_ring_ctx *ctx = req->ctx;
969 struct io_kiocb *notif;
971 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
973 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
974 if (req->flags & REQ_F_CQE_SKIP)
977 notif = zc->notif = io_alloc_notif(ctx);
980 notif->cqe.user_data = req->cqe.user_data;
982 notif->cqe.flags = IORING_CQE_F_NOTIF;
983 req->flags |= REQ_F_NEED_CLEANUP;
985 zc->flags = READ_ONCE(sqe->ioprio);
986 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
987 if (zc->flags & ~IO_ZC_FLAGS_VALID)
989 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
990 io_notif_set_extended(notif);
991 io_notif_to_data(notif)->zc_report = true;
995 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
996 unsigned idx = READ_ONCE(sqe->buf_index);
998 if (unlikely(idx >= ctx->nr_user_bufs))
1000 idx = array_index_nospec(idx, ctx->nr_user_bufs);
1001 req->imu = READ_ONCE(ctx->user_bufs[idx]);
1002 io_req_set_rsrc_node(notif, ctx, 0);
1005 if (req->opcode == IORING_OP_SEND_ZC) {
1006 if (READ_ONCE(sqe->__pad3[0]))
1008 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1009 zc->addr_len = READ_ONCE(sqe->addr_len);
1011 if (unlikely(sqe->addr2 || sqe->file_index))
1013 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1017 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1018 zc->len = READ_ONCE(sqe->len);
1019 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1020 if (zc->msg_flags & MSG_DONTWAIT)
1021 req->flags |= REQ_F_NOWAIT;
1025 #ifdef CONFIG_COMPAT
1026 if (req->ctx->compat)
1027 zc->msg_flags |= MSG_CMSG_COMPAT;
1032 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1033 struct iov_iter *from, size_t length)
1035 skb_zcopy_downgrade_managed(skb);
1036 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1039 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1040 struct iov_iter *from, size_t length)
1042 struct skb_shared_info *shinfo = skb_shinfo(skb);
1043 int frag = shinfo->nr_frags;
1045 struct bvec_iter bi;
1047 unsigned long truesize = 0;
1050 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1051 else if (unlikely(!skb_zcopy_managed(skb)))
1052 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1054 bi.bi_size = min(from->count, length);
1055 bi.bi_bvec_done = from->iov_offset;
1058 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1059 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1062 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1063 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1064 v.bv_offset, v.bv_len);
1065 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1070 shinfo->nr_frags = frag;
1071 from->bvec += bi.bi_idx;
1072 from->nr_segs -= bi.bi_idx;
1073 from->count -= copied;
1074 from->iov_offset = bi.bi_bvec_done;
1076 skb->data_len += copied;
1078 skb->truesize += truesize;
1080 if (sk && sk->sk_type == SOCK_STREAM) {
1081 sk_wmem_queued_add(sk, truesize);
1082 if (!skb_zcopy_pure(skb))
1083 sk_mem_charge(sk, truesize);
1085 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1090 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1092 struct sockaddr_storage __address;
1093 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1095 struct socket *sock;
1097 int ret, min_ret = 0;
1099 sock = sock_from_file(req->file);
1100 if (unlikely(!sock))
1102 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1105 msg.msg_name = NULL;
1106 msg.msg_control = NULL;
1107 msg.msg_controllen = 0;
1108 msg.msg_namelen = 0;
1111 if (req_has_async_data(req)) {
1112 struct io_async_msghdr *io = req->async_data;
1114 msg.msg_name = &io->addr;
1116 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1117 if (unlikely(ret < 0))
1119 msg.msg_name = (struct sockaddr *)&__address;
1121 msg.msg_namelen = zc->addr_len;
1124 if (!(req->flags & REQ_F_POLLED) &&
1125 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1126 return io_setup_async_addr(req, &__address, issue_flags);
1128 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1129 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1130 (u64)(uintptr_t)zc->buf, zc->len);
1133 msg.sg_from_iter = io_sg_from_iter;
1135 io_notif_set_extended(zc->notif);
1136 ret = import_ubuf(ITER_SOURCE, zc->buf, zc->len, &msg.msg_iter);
1139 ret = io_notif_account_mem(zc->notif, zc->len);
1142 msg.sg_from_iter = io_sg_from_iter_iovec;
1145 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1146 if (issue_flags & IO_URING_F_NONBLOCK)
1147 msg_flags |= MSG_DONTWAIT;
1148 if (msg_flags & MSG_WAITALL)
1149 min_ret = iov_iter_count(&msg.msg_iter);
1151 msg.msg_flags = msg_flags;
1152 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1153 ret = sock_sendmsg(sock, &msg);
1155 if (unlikely(ret < min_ret)) {
1156 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1157 return io_setup_async_addr(req, &__address, issue_flags);
1159 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1163 req->flags |= REQ_F_PARTIAL_IO;
1164 return io_setup_async_addr(req, &__address, issue_flags);
1166 if (ret == -ERESTARTSYS)
1173 else if (zc->done_io)
1177 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1178 * flushing notif to io_send_zc_cleanup()
1180 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1181 io_notif_flush(zc->notif);
1182 req->flags &= ~REQ_F_NEED_CLEANUP;
1184 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1188 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1190 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1191 struct io_async_msghdr iomsg, *kmsg;
1192 struct socket *sock;
1194 int ret, min_ret = 0;
1196 io_notif_set_extended(sr->notif);
1198 sock = sock_from_file(req->file);
1199 if (unlikely(!sock))
1201 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1204 if (req_has_async_data(req)) {
1205 kmsg = req->async_data;
1207 ret = io_sendmsg_copy_hdr(req, &iomsg);
1213 if (!(req->flags & REQ_F_POLLED) &&
1214 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1215 return io_setup_async_msg(req, kmsg, issue_flags);
1217 flags = sr->msg_flags | MSG_ZEROCOPY;
1218 if (issue_flags & IO_URING_F_NONBLOCK)
1219 flags |= MSG_DONTWAIT;
1220 if (flags & MSG_WAITALL)
1221 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1223 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1224 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1225 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1227 if (unlikely(ret < min_ret)) {
1228 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1229 return io_setup_async_msg(req, kmsg, issue_flags);
1231 if (ret > 0 && io_net_retry(sock, flags)) {
1233 req->flags |= REQ_F_PARTIAL_IO;
1234 return io_setup_async_msg(req, kmsg, issue_flags);
1236 if (ret == -ERESTARTSYS)
1240 /* fast path, check for non-NULL to avoid function call */
1241 if (kmsg->free_iov) {
1242 kfree(kmsg->free_iov);
1243 kmsg->free_iov = NULL;
1246 io_netmsg_recycle(req, issue_flags);
1249 else if (sr->done_io)
1253 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1254 * flushing notif to io_send_zc_cleanup()
1256 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1257 io_notif_flush(sr->notif);
1258 req->flags &= ~REQ_F_NEED_CLEANUP;
1260 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1264 void io_sendrecv_fail(struct io_kiocb *req)
1266 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1268 if (req->flags & REQ_F_PARTIAL_IO)
1269 req->cqe.res = sr->done_io;
1271 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1272 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1273 req->cqe.flags |= IORING_CQE_F_MORE;
1276 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1278 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1281 if (sqe->len || sqe->buf_index)
1284 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1285 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1286 accept->flags = READ_ONCE(sqe->accept_flags);
1287 accept->nofile = rlimit(RLIMIT_NOFILE);
1288 flags = READ_ONCE(sqe->ioprio);
1289 if (flags & ~IORING_ACCEPT_MULTISHOT)
1292 accept->file_slot = READ_ONCE(sqe->file_index);
1293 if (accept->file_slot) {
1294 if (accept->flags & SOCK_CLOEXEC)
1296 if (flags & IORING_ACCEPT_MULTISHOT &&
1297 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1300 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1302 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1303 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1304 if (flags & IORING_ACCEPT_MULTISHOT)
1305 req->flags |= REQ_F_APOLL_MULTISHOT;
1309 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1311 struct io_ring_ctx *ctx = req->ctx;
1312 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1313 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1314 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1315 bool fixed = !!accept->file_slot;
1319 if (!io_check_multishot(req, issue_flags))
1323 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1324 if (unlikely(fd < 0))
1327 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1332 ret = PTR_ERR(file);
1333 if (ret == -EAGAIN && force_nonblock) {
1335 * if it's multishot and polled, we don't need to
1336 * return EAGAIN to arm the poll infra since it
1337 * has already been done
1339 if (issue_flags & IO_URING_F_MULTISHOT)
1340 ret = IOU_ISSUE_SKIP_COMPLETE;
1343 if (ret == -ERESTARTSYS)
1346 } else if (!fixed) {
1347 fd_install(fd, file);
1350 ret = io_fixed_fd_install(req, issue_flags, file,
1354 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1355 io_req_set_res(req, ret, 0);
1361 if (io_aux_cqe(ctx, issue_flags & IO_URING_F_COMPLETE_DEFER,
1362 req->cqe.user_data, ret, IORING_CQE_F_MORE, true))
1368 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1370 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1372 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1375 sock->domain = READ_ONCE(sqe->fd);
1376 sock->type = READ_ONCE(sqe->off);
1377 sock->protocol = READ_ONCE(sqe->len);
1378 sock->file_slot = READ_ONCE(sqe->file_index);
1379 sock->nofile = rlimit(RLIMIT_NOFILE);
1381 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1382 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1384 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1389 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1391 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1392 bool fixed = !!sock->file_slot;
1397 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1398 if (unlikely(fd < 0))
1401 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1405 ret = PTR_ERR(file);
1406 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1408 if (ret == -ERESTARTSYS)
1411 } else if (!fixed) {
1412 fd_install(fd, file);
1415 ret = io_fixed_fd_install(req, issue_flags, file,
1418 io_req_set_res(req, ret, 0);
1422 int io_connect_prep_async(struct io_kiocb *req)
1424 struct io_async_connect *io = req->async_data;
1425 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1427 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1430 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1432 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1434 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1437 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1438 conn->addr_len = READ_ONCE(sqe->addr2);
1439 conn->in_progress = conn->seen_econnaborted = false;
1443 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1445 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1446 struct io_async_connect __io, *io;
1447 unsigned file_flags;
1449 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1451 if (connect->in_progress) {
1452 struct socket *socket;
1455 socket = sock_from_file(req->file);
1457 ret = sock_error(socket->sk);
1461 if (req_has_async_data(req)) {
1462 io = req->async_data;
1464 ret = move_addr_to_kernel(connect->addr,
1472 file_flags = force_nonblock ? O_NONBLOCK : 0;
1474 ret = __sys_connect_file(req->file, &io->address,
1475 connect->addr_len, file_flags);
1476 if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1477 && force_nonblock) {
1478 if (ret == -EINPROGRESS) {
1479 connect->in_progress = true;
1482 if (ret == -ECONNABORTED) {
1483 if (connect->seen_econnaborted)
1485 connect->seen_econnaborted = true;
1487 if (req_has_async_data(req))
1489 if (io_alloc_async_data(req)) {
1493 memcpy(req->async_data, &__io, sizeof(__io));
1496 if (ret == -ERESTARTSYS)
1501 io_req_set_res(req, ret, 0);
1505 void io_netmsg_cache_free(struct io_cache_entry *entry)
1507 kfree(container_of(entry, struct io_async_msghdr, cache));