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
17/* All valid masks for MSG_RING */
18#define IORING_MSG_RING_MASK (IORING_MSG_RING_CQE_SKIP | \
19 IORING_MSG_RING_FLAGS_PASS)
20
21struct io_msg {
22 struct file *file;
23 struct file *src_file;
24 struct callback_head tw;
25 u64 user_data;
26 u32 len;
27 u32 cmd;
28 u32 src_fd;
29 union {
30 u32 dst_fd;
31 u32 cqe_flags;
32 };
33 u32 flags;
34};
35
36static void io_double_unlock_ctx(struct io_ring_ctx *octx)
37{
38 mutex_unlock(lock: &octx->uring_lock);
39}
40
41static int io_double_lock_ctx(struct io_ring_ctx *octx,
42 unsigned int issue_flags)
43{
44 /*
45 * To ensure proper ordering between the two ctxs, we can only
46 * attempt a trylock on the target. If that fails and we already have
47 * the source ctx lock, punt to io-wq.
48 */
49 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
50 if (!mutex_trylock(lock: &octx->uring_lock))
51 return -EAGAIN;
52 return 0;
53 }
54 mutex_lock(&octx->uring_lock);
55 return 0;
56}
57
58void io_msg_ring_cleanup(struct io_kiocb *req)
59{
60 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
61
62 if (WARN_ON_ONCE(!msg->src_file))
63 return;
64
65 fput(msg->src_file);
66 msg->src_file = NULL;
67}
68
69static inline bool io_msg_need_remote(struct io_ring_ctx *target_ctx)
70{
71 if (!target_ctx->task_complete)
72 return false;
73 return current != target_ctx->submitter_task;
74}
75
76static int io_msg_exec_remote(struct io_kiocb *req, task_work_func_t func)
77{
78 struct io_ring_ctx *ctx = req->file->private_data;
79 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
80 struct task_struct *task = READ_ONCE(ctx->submitter_task);
81
82 if (unlikely(!task))
83 return -EOWNERDEAD;
84
85 init_task_work(twork: &msg->tw, func);
86 if (task_work_add(task: ctx->submitter_task, twork: &msg->tw, mode: TWA_SIGNAL))
87 return -EOWNERDEAD;
88
89 return IOU_ISSUE_SKIP_COMPLETE;
90}
91
92static void io_msg_tw_complete(struct callback_head *head)
93{
94 struct io_msg *msg = container_of(head, struct io_msg, tw);
95 struct io_kiocb *req = cmd_to_io_kiocb(msg);
96 struct io_ring_ctx *target_ctx = req->file->private_data;
97 int ret = 0;
98
99 if (current->flags & PF_EXITING) {
100 ret = -EOWNERDEAD;
101 } else {
102 u32 flags = 0;
103
104 if (msg->flags & IORING_MSG_RING_FLAGS_PASS)
105 flags = msg->cqe_flags;
106
107 /*
108 * If the target ring is using IOPOLL mode, then we need to be
109 * holding the uring_lock for posting completions. Other ring
110 * types rely on the regular completion locking, which is
111 * handled while posting.
112 */
113 if (target_ctx->flags & IORING_SETUP_IOPOLL)
114 mutex_lock(&target_ctx->uring_lock);
115 if (!io_post_aux_cqe(ctx: target_ctx, user_data: msg->user_data, res: msg->len, cflags: flags))
116 ret = -EOVERFLOW;
117 if (target_ctx->flags & IORING_SETUP_IOPOLL)
118 mutex_unlock(lock: &target_ctx->uring_lock);
119 }
120
121 if (ret < 0)
122 req_set_fail(req);
123 io_req_queue_tw_complete(req, res: ret);
124}
125
126static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags)
127{
128 struct io_ring_ctx *target_ctx = req->file->private_data;
129 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
130 u32 flags = 0;
131 int ret;
132
133 if (msg->src_fd || msg->flags & ~IORING_MSG_RING_FLAGS_PASS)
134 return -EINVAL;
135 if (!(msg->flags & IORING_MSG_RING_FLAGS_PASS) && msg->dst_fd)
136 return -EINVAL;
137 if (target_ctx->flags & IORING_SETUP_R_DISABLED)
138 return -EBADFD;
139
140 if (io_msg_need_remote(target_ctx))
141 return io_msg_exec_remote(req, func: io_msg_tw_complete);
142
143 if (msg->flags & IORING_MSG_RING_FLAGS_PASS)
144 flags = msg->cqe_flags;
145
146 ret = -EOVERFLOW;
147 if (target_ctx->flags & IORING_SETUP_IOPOLL) {
148 if (unlikely(io_double_lock_ctx(target_ctx, issue_flags)))
149 return -EAGAIN;
150 if (io_post_aux_cqe(ctx: target_ctx, user_data: msg->user_data, res: msg->len, cflags: flags))
151 ret = 0;
152 io_double_unlock_ctx(octx: target_ctx);
153 } else {
154 if (io_post_aux_cqe(ctx: target_ctx, user_data: msg->user_data, res: msg->len, cflags: flags))
155 ret = 0;
156 }
157 return ret;
158}
159
160static struct file *io_msg_grab_file(struct io_kiocb *req, unsigned int issue_flags)
161{
162 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
163 struct io_ring_ctx *ctx = req->ctx;
164 struct file *file = NULL;
165 int idx = msg->src_fd;
166
167 io_ring_submit_lock(ctx, issue_flags);
168 if (likely(idx < ctx->nr_user_files)) {
169 idx = array_index_nospec(idx, ctx->nr_user_files);
170 file = io_file_from_index(table: &ctx->file_table, index: idx);
171 if (file)
172 get_file(f: file);
173 }
174 io_ring_submit_unlock(ctx, issue_flags);
175 return file;
176}
177
178static int io_msg_install_complete(struct io_kiocb *req, unsigned int issue_flags)
179{
180 struct io_ring_ctx *target_ctx = req->file->private_data;
181 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
182 struct file *src_file = msg->src_file;
183 int ret;
184
185 if (unlikely(io_double_lock_ctx(target_ctx, issue_flags)))
186 return -EAGAIN;
187
188 ret = __io_fixed_fd_install(ctx: target_ctx, file: src_file, file_slot: msg->dst_fd);
189 if (ret < 0)
190 goto out_unlock;
191
192 msg->src_file = NULL;
193 req->flags &= ~REQ_F_NEED_CLEANUP;
194
195 if (msg->flags & IORING_MSG_RING_CQE_SKIP)
196 goto out_unlock;
197 /*
198 * If this fails, the target still received the file descriptor but
199 * wasn't notified of the fact. This means that if this request
200 * completes with -EOVERFLOW, then the sender must ensure that a
201 * later IORING_OP_MSG_RING delivers the message.
202 */
203 if (!io_post_aux_cqe(ctx: target_ctx, user_data: msg->user_data, res: ret, cflags: 0))
204 ret = -EOVERFLOW;
205out_unlock:
206 io_double_unlock_ctx(octx: target_ctx);
207 return ret;
208}
209
210static void io_msg_tw_fd_complete(struct callback_head *head)
211{
212 struct io_msg *msg = container_of(head, struct io_msg, tw);
213 struct io_kiocb *req = cmd_to_io_kiocb(msg);
214 int ret = -EOWNERDEAD;
215
216 if (!(current->flags & PF_EXITING))
217 ret = io_msg_install_complete(req, issue_flags: IO_URING_F_UNLOCKED);
218 if (ret < 0)
219 req_set_fail(req);
220 io_req_queue_tw_complete(req, res: ret);
221}
222
223static int io_msg_send_fd(struct io_kiocb *req, unsigned int issue_flags)
224{
225 struct io_ring_ctx *target_ctx = req->file->private_data;
226 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
227 struct io_ring_ctx *ctx = req->ctx;
228 struct file *src_file = msg->src_file;
229
230 if (msg->len)
231 return -EINVAL;
232 if (target_ctx == ctx)
233 return -EINVAL;
234 if (target_ctx->flags & IORING_SETUP_R_DISABLED)
235 return -EBADFD;
236 if (!src_file) {
237 src_file = io_msg_grab_file(req, issue_flags);
238 if (!src_file)
239 return -EBADF;
240 msg->src_file = src_file;
241 req->flags |= REQ_F_NEED_CLEANUP;
242 }
243
244 if (io_msg_need_remote(target_ctx))
245 return io_msg_exec_remote(req, func: io_msg_tw_fd_complete);
246 return io_msg_install_complete(req, issue_flags);
247}
248
249int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
250{
251 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
252
253 if (unlikely(sqe->buf_index || sqe->personality))
254 return -EINVAL;
255
256 msg->src_file = NULL;
257 msg->user_data = READ_ONCE(sqe->off);
258 msg->len = READ_ONCE(sqe->len);
259 msg->cmd = READ_ONCE(sqe->addr);
260 msg->src_fd = READ_ONCE(sqe->addr3);
261 msg->dst_fd = READ_ONCE(sqe->file_index);
262 msg->flags = READ_ONCE(sqe->msg_ring_flags);
263 if (msg->flags & ~IORING_MSG_RING_MASK)
264 return -EINVAL;
265
266 return 0;
267}
268
269int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
270{
271 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
272 int ret;
273
274 ret = -EBADFD;
275 if (!io_is_uring_fops(file: req->file))
276 goto done;
277
278 switch (msg->cmd) {
279 case IORING_MSG_DATA:
280 ret = io_msg_ring_data(req, issue_flags);
281 break;
282 case IORING_MSG_SEND_FD:
283 ret = io_msg_send_fd(req, issue_flags);
284 break;
285 default:
286 ret = -EINVAL;
287 break;
288 }
289
290done:
291 if (ret < 0) {
292 if (ret == -EAGAIN || ret == IOU_ISSUE_SKIP_COMPLETE)
293 return ret;
294 req_set_fail(req);
295 }
296 io_req_set_res(req, res: ret, cflags: 0);
297 return IOU_OK;
298}
299

source code of linux/io_uring/msg_ring.c