xref: /openbmc/linux/io_uring/msg_ring.c (revision b868a02e)
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 	src_file = (struct file *) (file_ptr & FFS_MASK);
99 	get_file(src_file);
100 
101 	ret = __io_fixed_fd_install(target_ctx, src_file, msg->dst_fd);
102 	if (ret < 0) {
103 		fput(src_file);
104 		goto out_unlock;
105 	}
106 
107 	if (msg->flags & IORING_MSG_RING_CQE_SKIP)
108 		goto out_unlock;
109 
110 	/*
111 	 * If this fails, the target still received the file descriptor but
112 	 * wasn't notified of the fact. This means that if this request
113 	 * completes with -EOVERFLOW, then the sender must ensure that a
114 	 * later IORING_OP_MSG_RING delivers the message.
115 	 */
116 	if (!io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0, true))
117 		ret = -EOVERFLOW;
118 out_unlock:
119 	io_double_unlock_ctx(ctx, target_ctx, issue_flags);
120 	return ret;
121 }
122 
123 int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
124 {
125 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
126 
127 	if (unlikely(sqe->buf_index || sqe->personality))
128 		return -EINVAL;
129 
130 	msg->user_data = READ_ONCE(sqe->off);
131 	msg->len = READ_ONCE(sqe->len);
132 	msg->cmd = READ_ONCE(sqe->addr);
133 	msg->src_fd = READ_ONCE(sqe->addr3);
134 	msg->dst_fd = READ_ONCE(sqe->file_index);
135 	msg->flags = READ_ONCE(sqe->msg_ring_flags);
136 	if (msg->flags & ~IORING_MSG_RING_CQE_SKIP)
137 		return -EINVAL;
138 
139 	return 0;
140 }
141 
142 int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
143 {
144 	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
145 	int ret;
146 
147 	ret = -EBADFD;
148 	if (!io_is_uring_fops(req->file))
149 		goto done;
150 
151 	switch (msg->cmd) {
152 	case IORING_MSG_DATA:
153 		ret = io_msg_ring_data(req);
154 		break;
155 	case IORING_MSG_SEND_FD:
156 		ret = io_msg_send_fd(req, issue_flags);
157 		break;
158 	default:
159 		ret = -EINVAL;
160 		break;
161 	}
162 
163 done:
164 	if (ret < 0)
165 		req_set_fail(req);
166 	io_req_set_res(req, ret, 0);
167 	/* put file to avoid an attempt to IOPOLL the req */
168 	if (!(req->flags & REQ_F_FIXED_FILE))
169 		io_put_file(req->file);
170 	req->file = NULL;
171 	return IOU_OK;
172 }
173