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;
63 unsigned nr_multishot_loops;
65 /* initialised and used only by !msg send variants */
69 void __user *msg_control;
70 /* used only for send zerocopy */
71 struct io_kiocb *notif;
75 * Number of times we'll try and do receives if there's more data. If we
76 * exceed this limit, then add us to the back of the queue and retry from
77 * there. This helps fairness between flooding clients.
79 #define MULTISHOT_MAX_RETRY 32
81 int io_shutdown_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
83 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
85 if (unlikely(sqe->off || sqe->addr || sqe->rw_flags ||
86 sqe->buf_index || sqe->splice_fd_in))
89 shutdown->how = READ_ONCE(sqe->len);
90 req->flags |= REQ_F_FORCE_ASYNC;
94 int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
96 struct io_shutdown *shutdown = io_kiocb_to_cmd(req, struct io_shutdown);
100 WARN_ON_ONCE(issue_flags & IO_URING_F_NONBLOCK);
102 sock = sock_from_file(req->file);
106 ret = __sys_shutdown_sock(sock, shutdown->how);
107 io_req_set_res(req, ret, 0);
111 static bool io_net_retry(struct socket *sock, int flags)
113 if (!(flags & MSG_WAITALL))
115 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
118 static void io_netmsg_recycle(struct io_kiocb *req, unsigned int issue_flags)
120 struct io_async_msghdr *hdr = req->async_data;
122 if (!req_has_async_data(req) || issue_flags & IO_URING_F_UNLOCKED)
125 /* Let normal cleanup path reap it if we fail adding to the cache */
126 if (io_alloc_cache_put(&req->ctx->netmsg_cache, &hdr->cache)) {
127 req->async_data = NULL;
128 req->flags &= ~REQ_F_ASYNC_DATA;
132 static struct io_async_msghdr *io_msg_alloc_async(struct io_kiocb *req,
133 unsigned int issue_flags)
135 struct io_ring_ctx *ctx = req->ctx;
136 struct io_cache_entry *entry;
137 struct io_async_msghdr *hdr;
139 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
140 entry = io_alloc_cache_get(&ctx->netmsg_cache);
142 hdr = container_of(entry, struct io_async_msghdr, cache);
143 hdr->free_iov = NULL;
144 req->flags |= REQ_F_ASYNC_DATA;
145 req->async_data = hdr;
150 if (!io_alloc_async_data(req)) {
151 hdr = req->async_data;
152 hdr->free_iov = NULL;
158 static inline struct io_async_msghdr *io_msg_alloc_async_prep(struct io_kiocb *req)
160 /* ->prep_async is always called from the submission context */
161 return io_msg_alloc_async(req, 0);
164 static int io_setup_async_msg(struct io_kiocb *req,
165 struct io_async_msghdr *kmsg,
166 unsigned int issue_flags)
168 struct io_async_msghdr *async_msg;
170 if (req_has_async_data(req))
172 async_msg = io_msg_alloc_async(req, issue_flags);
174 kfree(kmsg->free_iov);
177 req->flags |= REQ_F_NEED_CLEANUP;
178 memcpy(async_msg, kmsg, sizeof(*kmsg));
179 if (async_msg->msg.msg_name)
180 async_msg->msg.msg_name = &async_msg->addr;
182 if ((req->flags & REQ_F_BUFFER_SELECT) && !async_msg->msg.msg_iter.nr_segs)
185 /* if were using fast_iov, set it to the new one */
186 if (iter_is_iovec(&kmsg->msg.msg_iter) && !kmsg->free_iov) {
187 size_t fast_idx = iter_iov(&kmsg->msg.msg_iter) - kmsg->fast_iov;
188 async_msg->msg.msg_iter.__iov = &async_msg->fast_iov[fast_idx];
195 static int io_compat_msg_copy_hdr(struct io_kiocb *req,
196 struct io_async_msghdr *iomsg,
197 struct compat_msghdr *msg, int ddir)
199 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
200 struct compat_iovec __user *uiov;
203 if (copy_from_user(msg, sr->umsg_compat, sizeof(*msg)))
206 uiov = compat_ptr(msg->msg_iov);
207 if (req->flags & REQ_F_BUFFER_SELECT) {
210 iomsg->free_iov = NULL;
211 if (msg->msg_iovlen == 0) {
213 } else if (msg->msg_iovlen > 1) {
216 if (!access_ok(uiov, sizeof(*uiov)))
218 if (__get_user(clen, &uiov->iov_len))
228 iomsg->free_iov = iomsg->fast_iov;
229 ret = __import_iovec(ddir, (struct iovec __user *)uiov, msg->msg_iovlen,
230 UIO_FASTIOV, &iomsg->free_iov,
231 &iomsg->msg.msg_iter, true);
232 if (unlikely(ret < 0))
239 static int io_msg_copy_hdr(struct io_kiocb *req, struct io_async_msghdr *iomsg,
240 struct user_msghdr *msg, int ddir)
242 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
245 if (!user_access_begin(sr->umsg, sizeof(*sr->umsg)))
249 unsafe_get_user(msg->msg_name, &sr->umsg->msg_name, ua_end);
250 unsafe_get_user(msg->msg_namelen, &sr->umsg->msg_namelen, ua_end);
251 unsafe_get_user(msg->msg_iov, &sr->umsg->msg_iov, ua_end);
252 unsafe_get_user(msg->msg_iovlen, &sr->umsg->msg_iovlen, ua_end);
253 unsafe_get_user(msg->msg_control, &sr->umsg->msg_control, ua_end);
254 unsafe_get_user(msg->msg_controllen, &sr->umsg->msg_controllen, ua_end);
257 if (req->flags & REQ_F_BUFFER_SELECT) {
258 if (msg->msg_iovlen == 0) {
259 sr->len = iomsg->fast_iov[0].iov_len = 0;
260 iomsg->fast_iov[0].iov_base = NULL;
261 iomsg->free_iov = NULL;
262 } else if (msg->msg_iovlen > 1) {
266 /* we only need the length for provided buffers */
267 if (!access_ok(&msg->msg_iov[0].iov_len, sizeof(__kernel_size_t)))
269 unsafe_get_user(iomsg->fast_iov[0].iov_len,
270 &msg->msg_iov[0].iov_len, ua_end);
271 sr->len = iomsg->fast_iov[0].iov_len;
272 iomsg->free_iov = NULL;
281 iomsg->free_iov = iomsg->fast_iov;
282 ret = __import_iovec(ddir, msg->msg_iov, msg->msg_iovlen, UIO_FASTIOV,
283 &iomsg->free_iov, &iomsg->msg.msg_iter, false);
284 if (unlikely(ret < 0))
290 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
291 struct io_async_msghdr *iomsg)
293 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
294 struct user_msghdr msg;
297 iomsg->msg.msg_name = &iomsg->addr;
298 iomsg->msg.msg_iter.nr_segs = 0;
301 if (unlikely(req->ctx->compat)) {
302 struct compat_msghdr cmsg;
304 ret = io_compat_msg_copy_hdr(req, iomsg, &cmsg, ITER_SOURCE);
308 return __get_compat_msghdr(&iomsg->msg, &cmsg, NULL);
312 ret = io_msg_copy_hdr(req, iomsg, &msg, ITER_SOURCE);
316 ret = __copy_msghdr(&iomsg->msg, &msg, NULL);
318 /* save msg_control as sys_sendmsg() overwrites it */
319 sr->msg_control = iomsg->msg.msg_control_user;
323 int io_send_prep_async(struct io_kiocb *req)
325 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
326 struct io_async_msghdr *io;
329 if (req_has_async_data(req))
334 io = io_msg_alloc_async_prep(req);
337 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &io->addr);
341 static int io_setup_async_addr(struct io_kiocb *req,
342 struct sockaddr_storage *addr_storage,
343 unsigned int issue_flags)
345 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
346 struct io_async_msghdr *io;
348 if (!sr->addr || req_has_async_data(req))
350 io = io_msg_alloc_async(req, issue_flags);
353 memcpy(&io->addr, addr_storage, sizeof(io->addr));
357 int io_sendmsg_prep_async(struct io_kiocb *req)
359 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
363 if (!io_msg_alloc_async_prep(req))
365 ret = io_sendmsg_copy_hdr(req, req->async_data);
367 req->flags |= REQ_F_NEED_CLEANUP;
371 void io_sendmsg_recvmsg_cleanup(struct io_kiocb *req)
373 struct io_async_msghdr *io = req->async_data;
378 int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
380 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
384 if (req->opcode == IORING_OP_SEND) {
385 if (READ_ONCE(sqe->__pad3[0]))
387 sr->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
388 sr->addr_len = READ_ONCE(sqe->addr_len);
389 } else if (sqe->addr2 || sqe->file_index) {
393 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
394 sr->len = READ_ONCE(sqe->len);
395 sr->flags = READ_ONCE(sqe->ioprio);
396 if (sr->flags & ~IORING_RECVSEND_POLL_FIRST)
398 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
399 if (sr->msg_flags & MSG_DONTWAIT)
400 req->flags |= REQ_F_NOWAIT;
403 if (req->ctx->compat)
404 sr->msg_flags |= MSG_CMSG_COMPAT;
409 static void io_req_msg_cleanup(struct io_kiocb *req,
410 struct io_async_msghdr *kmsg,
411 unsigned int issue_flags)
413 req->flags &= ~REQ_F_NEED_CLEANUP;
414 /* fast path, check for non-NULL to avoid function call */
416 kfree(kmsg->free_iov);
417 io_netmsg_recycle(req, issue_flags);
420 int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
422 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
423 struct io_async_msghdr iomsg, *kmsg;
429 sock = sock_from_file(req->file);
433 if (req_has_async_data(req)) {
434 kmsg = req->async_data;
435 kmsg->msg.msg_control_user = sr->msg_control;
437 ret = io_sendmsg_copy_hdr(req, &iomsg);
443 if (!(req->flags & REQ_F_POLLED) &&
444 (sr->flags & IORING_RECVSEND_POLL_FIRST))
445 return io_setup_async_msg(req, kmsg, issue_flags);
447 flags = sr->msg_flags;
448 if (issue_flags & IO_URING_F_NONBLOCK)
449 flags |= MSG_DONTWAIT;
450 if (flags & MSG_WAITALL)
451 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
453 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
456 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
457 return io_setup_async_msg(req, kmsg, issue_flags);
458 if (ret > 0 && io_net_retry(sock, flags)) {
459 kmsg->msg.msg_controllen = 0;
460 kmsg->msg.msg_control = NULL;
462 req->flags |= REQ_F_BL_NO_RECYCLE;
463 return io_setup_async_msg(req, kmsg, issue_flags);
465 if (ret == -ERESTARTSYS)
469 io_req_msg_cleanup(req, kmsg, issue_flags);
472 else if (sr->done_io)
474 io_req_set_res(req, ret, 0);
478 int io_send(struct io_kiocb *req, unsigned int issue_flags)
480 struct sockaddr_storage __address;
481 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
489 msg.msg_control = NULL;
490 msg.msg_controllen = 0;
495 if (req_has_async_data(req)) {
496 struct io_async_msghdr *io = req->async_data;
498 msg.msg_name = &io->addr;
500 ret = move_addr_to_kernel(sr->addr, sr->addr_len, &__address);
501 if (unlikely(ret < 0))
503 msg.msg_name = (struct sockaddr *)&__address;
505 msg.msg_namelen = sr->addr_len;
508 if (!(req->flags & REQ_F_POLLED) &&
509 (sr->flags & IORING_RECVSEND_POLL_FIRST))
510 return io_setup_async_addr(req, &__address, issue_flags);
512 sock = sock_from_file(req->file);
516 ret = import_ubuf(ITER_SOURCE, sr->buf, sr->len, &msg.msg_iter);
520 flags = sr->msg_flags;
521 if (issue_flags & IO_URING_F_NONBLOCK)
522 flags |= MSG_DONTWAIT;
523 if (flags & MSG_WAITALL)
524 min_ret = iov_iter_count(&msg.msg_iter);
526 flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
527 msg.msg_flags = flags;
528 ret = sock_sendmsg(sock, &msg);
530 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
531 return io_setup_async_addr(req, &__address, issue_flags);
533 if (ret > 0 && io_net_retry(sock, flags)) {
537 req->flags |= REQ_F_BL_NO_RECYCLE;
538 return io_setup_async_addr(req, &__address, issue_flags);
540 if (ret == -ERESTARTSYS)
546 else if (sr->done_io)
548 io_req_set_res(req, ret, 0);
552 static int io_recvmsg_mshot_prep(struct io_kiocb *req,
553 struct io_async_msghdr *iomsg,
554 int namelen, size_t controllen)
556 if ((req->flags & (REQ_F_APOLL_MULTISHOT|REQ_F_BUFFER_SELECT)) ==
557 (REQ_F_APOLL_MULTISHOT|REQ_F_BUFFER_SELECT)) {
560 if (unlikely(namelen < 0))
562 if (check_add_overflow(sizeof(struct io_uring_recvmsg_out),
565 if (check_add_overflow(hdr, controllen, &hdr))
568 iomsg->namelen = namelen;
569 iomsg->controllen = controllen;
576 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
577 struct io_async_msghdr *iomsg)
579 struct user_msghdr msg;
582 iomsg->msg.msg_name = &iomsg->addr;
583 iomsg->msg.msg_iter.nr_segs = 0;
586 if (unlikely(req->ctx->compat)) {
587 struct compat_msghdr cmsg;
589 ret = io_compat_msg_copy_hdr(req, iomsg, &cmsg, ITER_DEST);
593 ret = __get_compat_msghdr(&iomsg->msg, &cmsg, &iomsg->uaddr);
597 return io_recvmsg_mshot_prep(req, iomsg, cmsg.msg_namelen,
598 cmsg.msg_controllen);
602 ret = io_msg_copy_hdr(req, iomsg, &msg, ITER_DEST);
606 ret = __copy_msghdr(&iomsg->msg, &msg, &iomsg->uaddr);
610 return io_recvmsg_mshot_prep(req, iomsg, msg.msg_namelen,
614 int io_recvmsg_prep_async(struct io_kiocb *req)
616 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
617 struct io_async_msghdr *iomsg;
621 if (!io_msg_alloc_async_prep(req))
623 iomsg = req->async_data;
624 ret = io_recvmsg_copy_hdr(req, iomsg);
626 req->flags |= REQ_F_NEED_CLEANUP;
630 #define RECVMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECV_MULTISHOT)
632 int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
634 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
638 if (unlikely(sqe->file_index || sqe->addr2))
641 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
642 sr->len = READ_ONCE(sqe->len);
643 sr->flags = READ_ONCE(sqe->ioprio);
644 if (sr->flags & ~(RECVMSG_FLAGS))
646 sr->msg_flags = READ_ONCE(sqe->msg_flags);
647 if (sr->msg_flags & MSG_DONTWAIT)
648 req->flags |= REQ_F_NOWAIT;
649 if (sr->msg_flags & MSG_ERRQUEUE)
650 req->flags |= REQ_F_CLEAR_POLLIN;
651 if (sr->flags & IORING_RECV_MULTISHOT) {
652 if (!(req->flags & REQ_F_BUFFER_SELECT))
654 if (sr->msg_flags & MSG_WAITALL)
656 if (req->opcode == IORING_OP_RECV && sr->len)
658 req->flags |= REQ_F_APOLL_MULTISHOT;
660 * Store the buffer group for this multishot receive separately,
661 * as if we end up doing an io-wq based issue that selects a
662 * buffer, it has to be committed immediately and that will
663 * clear ->buf_list. This means we lose the link to the buffer
664 * list, and the eventual buffer put on completion then cannot
667 sr->buf_group = req->buf_index;
671 if (req->ctx->compat)
672 sr->msg_flags |= MSG_CMSG_COMPAT;
674 sr->nr_multishot_loops = 0;
678 static inline void io_recv_prep_retry(struct io_kiocb *req)
680 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
682 req->flags &= ~REQ_F_BL_EMPTY;
684 sr->len = 0; /* get from the provided buffer */
685 req->buf_index = sr->buf_group;
689 * Finishes io_recv and io_recvmsg.
691 * Returns true if it is actually finished, or false if it should run
692 * again (for multishot).
694 static inline bool io_recv_finish(struct io_kiocb *req, int *ret,
695 struct msghdr *msg, bool mshot_finished,
696 unsigned issue_flags)
700 cflags = io_put_kbuf(req, issue_flags);
701 if (msg->msg_inq > 0)
702 cflags |= IORING_CQE_F_SOCK_NONEMPTY;
705 * Fill CQE for this receive and see if we should keep trying to
706 * receive from this socket.
708 if ((req->flags & REQ_F_APOLL_MULTISHOT) && !mshot_finished &&
709 io_fill_cqe_req_aux(req, issue_flags & IO_URING_F_COMPLETE_DEFER,
710 *ret, cflags | IORING_CQE_F_MORE)) {
711 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
712 int mshot_retry_ret = IOU_ISSUE_SKIP_COMPLETE;
714 io_recv_prep_retry(req);
715 /* Known not-empty or unknown state, retry */
716 if (cflags & IORING_CQE_F_SOCK_NONEMPTY || msg->msg_inq < 0) {
717 if (sr->nr_multishot_loops++ < MULTISHOT_MAX_RETRY)
719 /* mshot retries exceeded, force a requeue */
720 sr->nr_multishot_loops = 0;
721 mshot_retry_ret = IOU_REQUEUE;
723 if (issue_flags & IO_URING_F_MULTISHOT)
724 *ret = mshot_retry_ret;
730 /* Finish the request / stop multishot. */
731 io_req_set_res(req, *ret, cflags);
733 if (issue_flags & IO_URING_F_MULTISHOT)
734 *ret = IOU_STOP_MULTISHOT;
740 static int io_recvmsg_prep_multishot(struct io_async_msghdr *kmsg,
741 struct io_sr_msg *sr, void __user **buf,
744 unsigned long ubuf = (unsigned long) *buf;
747 hdr = sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
752 if (kmsg->controllen) {
753 unsigned long control = ubuf + hdr - kmsg->controllen;
755 kmsg->msg.msg_control_user = (void __user *) control;
756 kmsg->msg.msg_controllen = kmsg->controllen;
759 sr->buf = *buf; /* stash for later copy */
760 *buf = (void __user *) (ubuf + hdr);
761 kmsg->payloadlen = *len = *len - hdr;
765 struct io_recvmsg_multishot_hdr {
766 struct io_uring_recvmsg_out msg;
767 struct sockaddr_storage addr;
770 static int io_recvmsg_multishot(struct socket *sock, struct io_sr_msg *io,
771 struct io_async_msghdr *kmsg,
772 unsigned int flags, bool *finished)
776 struct io_recvmsg_multishot_hdr hdr;
779 kmsg->msg.msg_name = &hdr.addr;
780 kmsg->msg.msg_flags = flags & (MSG_CMSG_CLOEXEC|MSG_CMSG_COMPAT);
781 kmsg->msg.msg_namelen = 0;
783 if (sock->file->f_flags & O_NONBLOCK)
784 flags |= MSG_DONTWAIT;
786 err = sock_recvmsg(sock, &kmsg->msg, flags);
787 *finished = err <= 0;
791 hdr.msg = (struct io_uring_recvmsg_out) {
792 .controllen = kmsg->controllen - kmsg->msg.msg_controllen,
793 .flags = kmsg->msg.msg_flags & ~MSG_CMSG_COMPAT
796 hdr.msg.payloadlen = err;
797 if (err > kmsg->payloadlen)
798 err = kmsg->payloadlen;
800 copy_len = sizeof(struct io_uring_recvmsg_out);
801 if (kmsg->msg.msg_namelen > kmsg->namelen)
802 copy_len += kmsg->namelen;
804 copy_len += kmsg->msg.msg_namelen;
807 * "fromlen shall refer to the value before truncation.."
810 hdr.msg.namelen = kmsg->msg.msg_namelen;
812 /* ensure that there is no gap between hdr and sockaddr_storage */
813 BUILD_BUG_ON(offsetof(struct io_recvmsg_multishot_hdr, addr) !=
814 sizeof(struct io_uring_recvmsg_out));
815 if (copy_to_user(io->buf, &hdr, copy_len)) {
820 return sizeof(struct io_uring_recvmsg_out) + kmsg->namelen +
821 kmsg->controllen + err;
824 int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
826 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
827 struct io_async_msghdr iomsg, *kmsg;
830 int ret, min_ret = 0;
831 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
832 bool mshot_finished = true;
834 sock = sock_from_file(req->file);
838 if (req_has_async_data(req)) {
839 kmsg = req->async_data;
841 ret = io_recvmsg_copy_hdr(req, &iomsg);
847 if (!(req->flags & REQ_F_POLLED) &&
848 (sr->flags & IORING_RECVSEND_POLL_FIRST))
849 return io_setup_async_msg(req, kmsg, issue_flags);
851 flags = sr->msg_flags;
853 flags |= MSG_DONTWAIT;
856 if (io_do_buffer_select(req)) {
858 size_t len = sr->len;
860 buf = io_buffer_select(req, &len, issue_flags);
864 if (req->flags & REQ_F_APOLL_MULTISHOT) {
865 ret = io_recvmsg_prep_multishot(kmsg, sr, &buf, &len);
867 io_kbuf_recycle(req, issue_flags);
872 iov_iter_ubuf(&kmsg->msg.msg_iter, ITER_DEST, buf, len);
875 kmsg->msg.msg_get_inq = 1;
876 kmsg->msg.msg_inq = -1;
877 if (req->flags & REQ_F_APOLL_MULTISHOT) {
878 ret = io_recvmsg_multishot(sock, sr, kmsg, flags,
881 /* disable partial retry for recvmsg with cmsg attached */
882 if (flags & MSG_WAITALL && !kmsg->msg.msg_controllen)
883 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
885 ret = __sys_recvmsg_sock(sock, &kmsg->msg, sr->umsg,
890 if (ret == -EAGAIN && force_nonblock) {
891 ret = io_setup_async_msg(req, kmsg, issue_flags);
892 if (ret == -EAGAIN && (issue_flags & IO_URING_F_MULTISHOT)) {
893 io_kbuf_recycle(req, issue_flags);
894 return IOU_ISSUE_SKIP_COMPLETE;
898 if (ret > 0 && io_net_retry(sock, flags)) {
900 req->flags |= REQ_F_BL_NO_RECYCLE;
901 return io_setup_async_msg(req, kmsg, issue_flags);
903 if (ret == -ERESTARTSYS)
906 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
912 else if (sr->done_io)
915 io_kbuf_recycle(req, issue_flags);
917 if (!io_recv_finish(req, &ret, &kmsg->msg, mshot_finished, issue_flags))
918 goto retry_multishot;
921 io_req_msg_cleanup(req, kmsg, issue_flags);
922 else if (ret == -EAGAIN)
923 return io_setup_async_msg(req, kmsg, issue_flags);
928 int io_recv(struct io_kiocb *req, unsigned int issue_flags)
930 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
934 int ret, min_ret = 0;
935 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
936 size_t len = sr->len;
938 if (!(req->flags & REQ_F_POLLED) &&
939 (sr->flags & IORING_RECVSEND_POLL_FIRST))
942 sock = sock_from_file(req->file);
948 msg.msg_control = NULL;
950 msg.msg_controllen = 0;
954 flags = sr->msg_flags;
956 flags |= MSG_DONTWAIT;
959 if (io_do_buffer_select(req)) {
962 buf = io_buffer_select(req, &len, issue_flags);
969 ret = import_ubuf(ITER_DEST, sr->buf, len, &msg.msg_iter);
976 if (flags & MSG_WAITALL)
977 min_ret = iov_iter_count(&msg.msg_iter);
979 ret = sock_recvmsg(sock, &msg, flags);
981 if (ret == -EAGAIN && force_nonblock) {
982 if (issue_flags & IO_URING_F_MULTISHOT) {
983 io_kbuf_recycle(req, issue_flags);
984 return IOU_ISSUE_SKIP_COMPLETE;
989 if (ret > 0 && io_net_retry(sock, flags)) {
993 req->flags |= REQ_F_BL_NO_RECYCLE;
996 if (ret == -ERESTARTSYS)
999 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
1006 else if (sr->done_io)
1009 io_kbuf_recycle(req, issue_flags);
1011 if (!io_recv_finish(req, &ret, &msg, ret <= 0, issue_flags))
1012 goto retry_multishot;
1017 void io_send_zc_cleanup(struct io_kiocb *req)
1019 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1020 struct io_async_msghdr *io;
1022 if (req_has_async_data(req)) {
1023 io = req->async_data;
1024 /* might be ->fast_iov if *msg_copy_hdr failed */
1025 if (io->free_iov != io->fast_iov)
1026 kfree(io->free_iov);
1029 io_notif_flush(zc->notif);
1034 #define IO_ZC_FLAGS_COMMON (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_FIXED_BUF)
1035 #define IO_ZC_FLAGS_VALID (IO_ZC_FLAGS_COMMON | IORING_SEND_ZC_REPORT_USAGE)
1037 int io_send_zc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1039 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1040 struct io_ring_ctx *ctx = req->ctx;
1041 struct io_kiocb *notif;
1045 if (unlikely(READ_ONCE(sqe->__pad2[0]) || READ_ONCE(sqe->addr3)))
1047 /* we don't support IOSQE_CQE_SKIP_SUCCESS just yet */
1048 if (req->flags & REQ_F_CQE_SKIP)
1051 notif = zc->notif = io_alloc_notif(ctx);
1054 notif->cqe.user_data = req->cqe.user_data;
1056 notif->cqe.flags = IORING_CQE_F_NOTIF;
1057 req->flags |= REQ_F_NEED_CLEANUP;
1059 zc->flags = READ_ONCE(sqe->ioprio);
1060 if (unlikely(zc->flags & ~IO_ZC_FLAGS_COMMON)) {
1061 if (zc->flags & ~IO_ZC_FLAGS_VALID)
1063 if (zc->flags & IORING_SEND_ZC_REPORT_USAGE) {
1064 io_notif_set_extended(notif);
1065 io_notif_to_data(notif)->zc_report = true;
1069 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1070 unsigned idx = READ_ONCE(sqe->buf_index);
1072 if (unlikely(idx >= ctx->nr_user_bufs))
1074 idx = array_index_nospec(idx, ctx->nr_user_bufs);
1075 req->imu = READ_ONCE(ctx->user_bufs[idx]);
1076 io_req_set_rsrc_node(notif, ctx, 0);
1079 if (req->opcode == IORING_OP_SEND_ZC) {
1080 if (READ_ONCE(sqe->__pad3[0]))
1082 zc->addr = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1083 zc->addr_len = READ_ONCE(sqe->addr_len);
1085 if (unlikely(sqe->addr2 || sqe->file_index))
1087 if (unlikely(zc->flags & IORING_RECVSEND_FIXED_BUF))
1091 zc->buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1092 zc->len = READ_ONCE(sqe->len);
1093 zc->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
1094 if (zc->msg_flags & MSG_DONTWAIT)
1095 req->flags |= REQ_F_NOWAIT;
1097 #ifdef CONFIG_COMPAT
1098 if (req->ctx->compat)
1099 zc->msg_flags |= MSG_CMSG_COMPAT;
1104 static int io_sg_from_iter_iovec(struct sock *sk, struct sk_buff *skb,
1105 struct iov_iter *from, size_t length)
1107 skb_zcopy_downgrade_managed(skb);
1108 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1111 static int io_sg_from_iter(struct sock *sk, struct sk_buff *skb,
1112 struct iov_iter *from, size_t length)
1114 struct skb_shared_info *shinfo = skb_shinfo(skb);
1115 int frag = shinfo->nr_frags;
1117 struct bvec_iter bi;
1119 unsigned long truesize = 0;
1122 shinfo->flags |= SKBFL_MANAGED_FRAG_REFS;
1123 else if (unlikely(!skb_zcopy_managed(skb)))
1124 return __zerocopy_sg_from_iter(NULL, sk, skb, from, length);
1126 bi.bi_size = min(from->count, length);
1127 bi.bi_bvec_done = from->iov_offset;
1130 while (bi.bi_size && frag < MAX_SKB_FRAGS) {
1131 struct bio_vec v = mp_bvec_iter_bvec(from->bvec, bi);
1134 truesize += PAGE_ALIGN(v.bv_len + v.bv_offset);
1135 __skb_fill_page_desc_noacc(shinfo, frag++, v.bv_page,
1136 v.bv_offset, v.bv_len);
1137 bvec_iter_advance_single(from->bvec, &bi, v.bv_len);
1142 shinfo->nr_frags = frag;
1143 from->bvec += bi.bi_idx;
1144 from->nr_segs -= bi.bi_idx;
1145 from->count -= copied;
1146 from->iov_offset = bi.bi_bvec_done;
1148 skb->data_len += copied;
1150 skb->truesize += truesize;
1152 if (sk && sk->sk_type == SOCK_STREAM) {
1153 sk_wmem_queued_add(sk, truesize);
1154 if (!skb_zcopy_pure(skb))
1155 sk_mem_charge(sk, truesize);
1157 refcount_add(truesize, &skb->sk->sk_wmem_alloc);
1162 int io_send_zc(struct io_kiocb *req, unsigned int issue_flags)
1164 struct sockaddr_storage __address;
1165 struct io_sr_msg *zc = io_kiocb_to_cmd(req, struct io_sr_msg);
1167 struct socket *sock;
1169 int ret, min_ret = 0;
1171 sock = sock_from_file(req->file);
1172 if (unlikely(!sock))
1174 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1177 msg.msg_name = NULL;
1178 msg.msg_control = NULL;
1179 msg.msg_controllen = 0;
1180 msg.msg_namelen = 0;
1183 if (req_has_async_data(req)) {
1184 struct io_async_msghdr *io = req->async_data;
1186 msg.msg_name = &io->addr;
1188 ret = move_addr_to_kernel(zc->addr, zc->addr_len, &__address);
1189 if (unlikely(ret < 0))
1191 msg.msg_name = (struct sockaddr *)&__address;
1193 msg.msg_namelen = zc->addr_len;
1196 if (!(req->flags & REQ_F_POLLED) &&
1197 (zc->flags & IORING_RECVSEND_POLL_FIRST))
1198 return io_setup_async_addr(req, &__address, issue_flags);
1200 if (zc->flags & IORING_RECVSEND_FIXED_BUF) {
1201 ret = io_import_fixed(ITER_SOURCE, &msg.msg_iter, req->imu,
1202 (u64)(uintptr_t)zc->buf, zc->len);
1205 msg.sg_from_iter = io_sg_from_iter;
1207 io_notif_set_extended(zc->notif);
1208 ret = import_ubuf(ITER_SOURCE, zc->buf, zc->len, &msg.msg_iter);
1211 ret = io_notif_account_mem(zc->notif, zc->len);
1214 msg.sg_from_iter = io_sg_from_iter_iovec;
1217 msg_flags = zc->msg_flags | MSG_ZEROCOPY;
1218 if (issue_flags & IO_URING_F_NONBLOCK)
1219 msg_flags |= MSG_DONTWAIT;
1220 if (msg_flags & MSG_WAITALL)
1221 min_ret = iov_iter_count(&msg.msg_iter);
1222 msg_flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
1224 msg.msg_flags = msg_flags;
1225 msg.msg_ubuf = &io_notif_to_data(zc->notif)->uarg;
1226 ret = sock_sendmsg(sock, &msg);
1228 if (unlikely(ret < min_ret)) {
1229 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1230 return io_setup_async_addr(req, &__address, issue_flags);
1232 if (ret > 0 && io_net_retry(sock, msg.msg_flags)) {
1236 req->flags |= REQ_F_BL_NO_RECYCLE;
1237 return io_setup_async_addr(req, &__address, issue_flags);
1239 if (ret == -ERESTARTSYS)
1246 else if (zc->done_io)
1250 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1251 * flushing notif to io_send_zc_cleanup()
1253 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1254 io_notif_flush(zc->notif);
1255 req->flags &= ~REQ_F_NEED_CLEANUP;
1257 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1261 int io_sendmsg_zc(struct io_kiocb *req, unsigned int issue_flags)
1263 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1264 struct io_async_msghdr iomsg, *kmsg;
1265 struct socket *sock;
1267 int ret, min_ret = 0;
1269 io_notif_set_extended(sr->notif);
1271 sock = sock_from_file(req->file);
1272 if (unlikely(!sock))
1274 if (!test_bit(SOCK_SUPPORT_ZC, &sock->flags))
1277 if (req_has_async_data(req)) {
1278 kmsg = req->async_data;
1279 kmsg->msg.msg_control_user = sr->msg_control;
1281 ret = io_sendmsg_copy_hdr(req, &iomsg);
1287 if (!(req->flags & REQ_F_POLLED) &&
1288 (sr->flags & IORING_RECVSEND_POLL_FIRST))
1289 return io_setup_async_msg(req, kmsg, issue_flags);
1291 flags = sr->msg_flags | MSG_ZEROCOPY;
1292 if (issue_flags & IO_URING_F_NONBLOCK)
1293 flags |= MSG_DONTWAIT;
1294 if (flags & MSG_WAITALL)
1295 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
1297 kmsg->msg.msg_ubuf = &io_notif_to_data(sr->notif)->uarg;
1298 kmsg->msg.sg_from_iter = io_sg_from_iter_iovec;
1299 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
1301 if (unlikely(ret < min_ret)) {
1302 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1303 return io_setup_async_msg(req, kmsg, issue_flags);
1305 if (ret > 0 && io_net_retry(sock, flags)) {
1307 req->flags |= REQ_F_BL_NO_RECYCLE;
1308 return io_setup_async_msg(req, kmsg, issue_flags);
1310 if (ret == -ERESTARTSYS)
1314 /* fast path, check for non-NULL to avoid function call */
1315 if (kmsg->free_iov) {
1316 kfree(kmsg->free_iov);
1317 kmsg->free_iov = NULL;
1320 io_netmsg_recycle(req, issue_flags);
1323 else if (sr->done_io)
1327 * If we're in io-wq we can't rely on tw ordering guarantees, defer
1328 * flushing notif to io_send_zc_cleanup()
1330 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1331 io_notif_flush(sr->notif);
1332 req->flags &= ~REQ_F_NEED_CLEANUP;
1334 io_req_set_res(req, ret, IORING_CQE_F_MORE);
1338 void io_sendrecv_fail(struct io_kiocb *req)
1340 struct io_sr_msg *sr = io_kiocb_to_cmd(req, struct io_sr_msg);
1343 req->cqe.res = sr->done_io;
1345 if ((req->flags & REQ_F_NEED_CLEANUP) &&
1346 (req->opcode == IORING_OP_SEND_ZC || req->opcode == IORING_OP_SENDMSG_ZC))
1347 req->cqe.flags |= IORING_CQE_F_MORE;
1350 int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1352 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1355 if (sqe->len || sqe->buf_index)
1358 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1359 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
1360 accept->flags = READ_ONCE(sqe->accept_flags);
1361 accept->nofile = rlimit(RLIMIT_NOFILE);
1362 flags = READ_ONCE(sqe->ioprio);
1363 if (flags & ~IORING_ACCEPT_MULTISHOT)
1366 accept->file_slot = READ_ONCE(sqe->file_index);
1367 if (accept->file_slot) {
1368 if (accept->flags & SOCK_CLOEXEC)
1370 if (flags & IORING_ACCEPT_MULTISHOT &&
1371 accept->file_slot != IORING_FILE_INDEX_ALLOC)
1374 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1376 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
1377 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
1378 if (flags & IORING_ACCEPT_MULTISHOT)
1379 req->flags |= REQ_F_APOLL_MULTISHOT;
1383 int io_accept(struct io_kiocb *req, unsigned int issue_flags)
1385 struct io_accept *accept = io_kiocb_to_cmd(req, struct io_accept);
1386 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1387 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
1388 bool fixed = !!accept->file_slot;
1394 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
1395 if (unlikely(fd < 0))
1398 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
1403 ret = PTR_ERR(file);
1404 if (ret == -EAGAIN && force_nonblock) {
1406 * if it's multishot and polled, we don't need to
1407 * return EAGAIN to arm the poll infra since it
1408 * has already been done
1410 if (issue_flags & IO_URING_F_MULTISHOT)
1411 return IOU_ISSUE_SKIP_COMPLETE;
1414 if (ret == -ERESTARTSYS)
1417 } else if (!fixed) {
1418 fd_install(fd, file);
1421 ret = io_fixed_fd_install(req, issue_flags, file,
1425 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1426 io_req_set_res(req, ret, 0);
1432 if (io_fill_cqe_req_aux(req, issue_flags & IO_URING_F_COMPLETE_DEFER,
1433 ret, IORING_CQE_F_MORE))
1436 io_req_set_res(req, ret, 0);
1437 return IOU_STOP_MULTISHOT;
1440 int io_socket_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1442 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1444 if (sqe->addr || sqe->rw_flags || sqe->buf_index)
1447 sock->domain = READ_ONCE(sqe->fd);
1448 sock->type = READ_ONCE(sqe->off);
1449 sock->protocol = READ_ONCE(sqe->len);
1450 sock->file_slot = READ_ONCE(sqe->file_index);
1451 sock->nofile = rlimit(RLIMIT_NOFILE);
1453 sock->flags = sock->type & ~SOCK_TYPE_MASK;
1454 if (sock->file_slot && (sock->flags & SOCK_CLOEXEC))
1456 if (sock->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
1461 int io_socket(struct io_kiocb *req, unsigned int issue_flags)
1463 struct io_socket *sock = io_kiocb_to_cmd(req, struct io_socket);
1464 bool fixed = !!sock->file_slot;
1469 fd = __get_unused_fd_flags(sock->flags, sock->nofile);
1470 if (unlikely(fd < 0))
1473 file = __sys_socket_file(sock->domain, sock->type, sock->protocol);
1477 ret = PTR_ERR(file);
1478 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
1480 if (ret == -ERESTARTSYS)
1483 } else if (!fixed) {
1484 fd_install(fd, file);
1487 ret = io_fixed_fd_install(req, issue_flags, file,
1490 io_req_set_res(req, ret, 0);
1494 int io_connect_prep_async(struct io_kiocb *req)
1496 struct io_async_connect *io = req->async_data;
1497 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1499 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
1502 int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1504 struct io_connect *conn = io_kiocb_to_cmd(req, struct io_connect);
1506 if (sqe->len || sqe->buf_index || sqe->rw_flags || sqe->splice_fd_in)
1509 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
1510 conn->addr_len = READ_ONCE(sqe->addr2);
1511 conn->in_progress = conn->seen_econnaborted = false;
1515 int io_connect(struct io_kiocb *req, unsigned int issue_flags)
1517 struct io_connect *connect = io_kiocb_to_cmd(req, struct io_connect);
1518 struct io_async_connect __io, *io;
1519 unsigned file_flags;
1521 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1523 if (req_has_async_data(req)) {
1524 io = req->async_data;
1526 ret = move_addr_to_kernel(connect->addr,
1534 file_flags = force_nonblock ? O_NONBLOCK : 0;
1536 ret = __sys_connect_file(req->file, &io->address,
1537 connect->addr_len, file_flags);
1538 if ((ret == -EAGAIN || ret == -EINPROGRESS || ret == -ECONNABORTED)
1539 && force_nonblock) {
1540 if (ret == -EINPROGRESS) {
1541 connect->in_progress = true;
1542 } else if (ret == -ECONNABORTED) {
1543 if (connect->seen_econnaborted)
1545 connect->seen_econnaborted = true;
1547 if (req_has_async_data(req))
1549 if (io_alloc_async_data(req)) {
1553 memcpy(req->async_data, &__io, sizeof(__io));
1556 if (connect->in_progress) {
1558 * At least bluetooth will return -EBADFD on a re-connect
1559 * attempt, and it's (supposedly) also valid to get -EISCONN
1560 * which means the previous result is good. For both of these,
1561 * grab the sock_error() and use that for the completion.
1563 if (ret == -EBADFD || ret == -EISCONN)
1564 ret = sock_error(sock_from_file(req->file)->sk);
1566 if (ret == -ERESTARTSYS)
1571 io_req_set_res(req, ret, 0);
1575 void io_netmsg_cache_free(struct io_cache_entry *entry)
1577 kfree(container_of(entry, struct io_async_msghdr, cache));