xref: /openbmc/linux/io_uring/msg_ring.c (revision 8eab9be5)
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>
6 #include <linux/nospec.h>
7 #include <linux/io_uring.h>
8 
9 #include <uapi/linux/io_uring.h>
10 
11 #include "io_uring.h"
12 #include "rsrc.h"
13 #include "filetable.h"
14 #include "msg_ring.h"
15 
16 struct io_msg {
17 	struct file			*file;
18 	u64 user_data;
19 	u32 len;
20 	u32 cmd;
21 	u32 src_fd;
22 	u32 dst_fd;
23 	u32 flags;
24 };
25 
26 static int io_msg_ring_data(struct io_kiocb *req)
27 {
28 	struct io_ring_ctx *target_ctx = req->file->private_data;
29 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
30 
31 	if (msg->src_fd || msg->dst_fd || msg->flags)
32 		return -EINVAL;
33 
34 	if (io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0, true))
35 		return 0;
36 
37 	return -EOVERFLOW;
38 }
39 
40 static void io_double_unlock_ctx(struct io_ring_ctx *ctx,
41 				 struct io_ring_ctx *octx,
42 				 unsigned int issue_flags)
43 {
44 	if (issue_flags & IO_URING_F_UNLOCKED)
45 		mutex_unlock(&ctx->uring_lock);
46 	mutex_unlock(&octx->uring_lock);
47 }
48 
49 static int io_double_lock_ctx(struct io_ring_ctx *ctx,
50 			      struct io_ring_ctx *octx,
51 			      unsigned int issue_flags)
52 {
53 	/*
54 	 * To ensure proper ordering between the two ctxs, we can only
55 	 * attempt a trylock on the target. If that fails and we already have
56 	 * the source ctx lock, punt to io-wq.
57 	 */
58 	if (!(issue_flags & IO_URING_F_UNLOCKED)) {
59 		if (!mutex_trylock(&octx->uring_lock))
60 			return -EAGAIN;
61 		return 0;
62 	}
63 
64 	/* Always grab smallest value ctx first. We know ctx != octx. */
65 	if (ctx < octx) {
66 		mutex_lock(&ctx->uring_lock);
67 		mutex_lock(&octx->uring_lock);
68 	} else {
69 		mutex_lock(&octx->uring_lock);
70 		mutex_lock(&ctx->uring_lock);
71 	}
72 
73 	return 0;
74 }
75 
76 static int io_msg_send_fd(struct io_kiocb *req, unsigned int issue_flags)
77 {
78 	struct io_ring_ctx *target_ctx = req->file->private_data;
79 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
80 	struct io_ring_ctx *ctx = req->ctx;
81 	unsigned long file_ptr;
82 	struct file *src_file;
83 	int ret;
84 
85 	if (target_ctx == ctx)
86 		return -EINVAL;
87 
88 	ret = io_double_lock_ctx(ctx, target_ctx, issue_flags);
89 	if (unlikely(ret))
90 		return ret;
91 
92 	ret = -EBADF;
93 	if (unlikely(msg->src_fd >= ctx->nr_user_files))
94 		goto out_unlock;
95 
96 	msg->src_fd = array_index_nospec(msg->src_fd, ctx->nr_user_files);
97 	file_ptr = io_fixed_file_slot(&ctx->file_table, msg->src_fd)->file_ptr;
98 	if (!file_ptr)
99 		goto out_unlock;
100 
101 	src_file = (struct file *) (file_ptr & FFS_MASK);
102 	get_file(src_file);
103 
104 	ret = __io_fixed_fd_install(target_ctx, src_file, msg->dst_fd);
105 	if (ret < 0) {
106 		fput(src_file);
107 		goto out_unlock;
108 	}
109 
110 	if (msg->flags & IORING_MSG_RING_CQE_SKIP)
111 		goto out_unlock;
112 
113 	/*
114 	 * If this fails, the target still received the file descriptor but
115 	 * wasn't notified of the fact. This means that if this request
116 	 * completes with -EOVERFLOW, then the sender must ensure that a
117 	 * later IORING_OP_MSG_RING delivers the message.
118 	 */
119 	if (!io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0, true))
120 		ret = -EOVERFLOW;
121 out_unlock:
122 	io_double_unlock_ctx(ctx, target_ctx, issue_flags);
123 	return ret;
124 }
125 
126 int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
127 {
128 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
129 
130 	if (unlikely(sqe->buf_index || sqe->personality))
131 		return -EINVAL;
132 
133 	msg->user_data = READ_ONCE(sqe->off);
134 	msg->len = READ_ONCE(sqe->len);
135 	msg->cmd = READ_ONCE(sqe->addr);
136 	msg->src_fd = READ_ONCE(sqe->addr3);
137 	msg->dst_fd = READ_ONCE(sqe->file_index);
138 	msg->flags = READ_ONCE(sqe->msg_ring_flags);
139 	if (msg->flags & ~IORING_MSG_RING_CQE_SKIP)
140 		return -EINVAL;
141 
142 	return 0;
143 }
144 
145 int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
146 {
147 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
148 	int ret;
149 
150 	ret = -EBADFD;
151 	if (!io_is_uring_fops(req->file))
152 		goto done;
153 
154 	switch (msg->cmd) {
155 	case IORING_MSG_DATA:
156 		ret = io_msg_ring_data(req);
157 		break;
158 	case IORING_MSG_SEND_FD:
159 		ret = io_msg_send_fd(req, issue_flags);
160 		break;
161 	default:
162 		ret = -EINVAL;
163 		break;
164 	}
165 
166 done:
167 	if (ret < 0)
168 		req_set_fail(req);
169 	io_req_set_res(req, ret, 0);
170 	/* put file to avoid an attempt to IOPOLL the req */
171 	if (!(req->flags & REQ_F_FIXED_FILE))
172 		io_put_file(req->file);
173 	req->file = NULL;
174 	return IOU_OK;
175 }
176