Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
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/* All valid masks for MSG_RING */
17#define IORING_MSG_RING_MASK (IORING_MSG_RING_CQE_SKIP | \
18 IORING_MSG_RING_FLAGS_PASS)
19
20struct io_msg {
21 struct file *file;
22 struct file *src_file;
23 struct callback_head tw;
24 u64 user_data;
25 u32 len;
26 u32 cmd;
27 u32 src_fd;
28 union {
29 u32 dst_fd;
30 u32 cqe_flags;
31 };
32 u32 flags;
33};
34
35static void io_double_unlock_ctx(struct io_ring_ctx *octx)
36{
37 mutex_unlock(&octx->uring_lock);
38}
39
40static int io_lock_external_ctx(struct io_ring_ctx *octx,
41 unsigned int issue_flags)
42{
43 /*
44 * To ensure proper ordering between the two ctxs, we can only
45 * attempt a trylock on the target. If that fails and we already have
46 * the source ctx lock, punt to io-wq.
47 */
48 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
49 if (!mutex_trylock(&octx->uring_lock))
50 return -EAGAIN;
51 return 0;
52 }
53 mutex_lock(&octx->uring_lock);
54 return 0;
55}
56
57void io_msg_ring_cleanup(struct io_kiocb *req)
58{
59 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
60
61 if (WARN_ON_ONCE(!msg->src_file))
62 return;
63
64 fput(msg->src_file);
65 msg->src_file = NULL;
66}
67
68static inline bool io_msg_need_remote(struct io_ring_ctx *target_ctx)
69{
70 return target_ctx->task_complete;
71}
72
73static void io_msg_tw_complete(struct io_tw_req tw_req, io_tw_token_t tw)
74{
75 struct io_kiocb *req = tw_req.req;
76 struct io_ring_ctx *ctx = req->ctx;
77
78 io_add_aux_cqe(ctx, req->cqe.user_data, req->cqe.res, req->cqe.flags);
79 kfree_rcu(req, rcu_head);
80 percpu_ref_put(&ctx->refs);
81}
82
83static int io_msg_remote_post(struct io_ring_ctx *ctx, struct io_kiocb *req,
84 int res, u32 cflags, u64 user_data)
85{
86 if (!READ_ONCE(ctx->submitter_task)) {
87 kfree_rcu(req, rcu_head);
88 return -EOWNERDEAD;
89 }
90 req->opcode = IORING_OP_NOP;
91 req->cqe.user_data = user_data;
92 io_req_set_res(req, res, cflags);
93 percpu_ref_get(&ctx->refs);
94 req->ctx = ctx;
95 req->tctx = NULL;
96 req->io_task_work.func = io_msg_tw_complete;
97 io_req_task_work_add_remote(req, IOU_F_TWQ_LAZY_WAKE);
98 return 0;
99}
100
101static int io_msg_data_remote(struct io_ring_ctx *target_ctx,
102 struct io_msg *msg)
103{
104 struct io_kiocb *target;
105 u32 flags = 0;
106
107 target = kmem_cache_alloc(req_cachep, GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO) ;
108 if (unlikely(!target))
109 return -ENOMEM;
110
111 if (msg->flags & IORING_MSG_RING_FLAGS_PASS)
112 flags = msg->cqe_flags;
113
114 return io_msg_remote_post(target_ctx, target, msg->len, flags,
115 msg->user_data);
116}
117
118static int __io_msg_ring_data(struct io_ring_ctx *target_ctx,
119 struct io_msg *msg, unsigned int issue_flags)
120{
121 u32 flags = 0;
122 int ret;
123
124 if (msg->src_fd || msg->flags & ~IORING_MSG_RING_FLAGS_PASS)
125 return -EINVAL;
126 if (!(msg->flags & IORING_MSG_RING_FLAGS_PASS) && msg->dst_fd)
127 return -EINVAL;
128 if (target_ctx->flags & IORING_SETUP_R_DISABLED)
129 return -EBADFD;
130
131 if (io_msg_need_remote(target_ctx))
132 return io_msg_data_remote(target_ctx, msg);
133
134 if (msg->flags & IORING_MSG_RING_FLAGS_PASS)
135 flags = msg->cqe_flags;
136
137 ret = -EOVERFLOW;
138 if (target_ctx->flags & IORING_SETUP_IOPOLL) {
139 if (unlikely(io_lock_external_ctx(target_ctx, issue_flags)))
140 return -EAGAIN;
141 }
142 if (io_post_aux_cqe(target_ctx, msg->user_data, msg->len, flags))
143 ret = 0;
144 if (target_ctx->flags & IORING_SETUP_IOPOLL)
145 io_double_unlock_ctx(target_ctx);
146 return ret;
147}
148
149static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags)
150{
151 struct io_ring_ctx *target_ctx = req->file->private_data;
152 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
153
154 return __io_msg_ring_data(target_ctx, msg, issue_flags);
155}
156
157static int io_msg_grab_file(struct io_kiocb *req, unsigned int issue_flags)
158{
159 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
160 struct io_ring_ctx *ctx = req->ctx;
161 struct io_rsrc_node *node;
162 int ret = -EBADF;
163
164 io_ring_submit_lock(ctx, issue_flags);
165 node = io_rsrc_node_lookup(&ctx->file_table.data, msg->src_fd);
166 if (node) {
167 msg->src_file = io_slot_file(node);
168 if (msg->src_file)
169 get_file(msg->src_file);
170 req->flags |= REQ_F_NEED_CLEANUP;
171 ret = 0;
172 }
173 io_ring_submit_unlock(ctx, issue_flags);
174 return ret;
175}
176
177static int io_msg_install_complete(struct io_kiocb *req, unsigned int issue_flags)
178{
179 struct io_ring_ctx *target_ctx = req->file->private_data;
180 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
181 struct file *src_file = msg->src_file;
182 int ret;
183
184 if (unlikely(io_lock_external_ctx(target_ctx, issue_flags)))
185 return -EAGAIN;
186
187 ret = __io_fixed_fd_install(target_ctx, src_file, msg->dst_fd);
188 if (ret < 0)
189 goto out_unlock;
190
191 msg->src_file = NULL;
192 req->flags &= ~REQ_F_NEED_CLEANUP;
193
194 if (msg->flags & IORING_MSG_RING_CQE_SKIP)
195 goto out_unlock;
196 /*
197 * If this fails, the target still received the file descriptor but
198 * wasn't notified of the fact. This means that if this request
199 * completes with -EOVERFLOW, then the sender must ensure that a
200 * later IORING_OP_MSG_RING delivers the message.
201 */
202 if (!io_post_aux_cqe(target_ctx, msg->user_data, ret, 0))
203 ret = -EOVERFLOW;
204out_unlock:
205 io_double_unlock_ctx(target_ctx);
206 return ret;
207}
208
209static void io_msg_tw_fd_complete(struct callback_head *head)
210{
211 struct io_msg *msg = container_of(head, struct io_msg, tw);
212 struct io_kiocb *req = cmd_to_io_kiocb(msg);
213 int ret = -EOWNERDEAD;
214
215 if (!(current->flags & PF_EXITING))
216 ret = io_msg_install_complete(req, IO_URING_F_UNLOCKED);
217 if (ret < 0)
218 req_set_fail(req);
219 io_req_queue_tw_complete(req, ret);
220}
221
222static int io_msg_fd_remote(struct io_kiocb *req)
223{
224 struct io_ring_ctx *ctx = req->file->private_data;
225 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
226 struct task_struct *task = READ_ONCE(ctx->submitter_task);
227
228 if (unlikely(!task))
229 return -EOWNERDEAD;
230
231 init_task_work(&msg->tw, io_msg_tw_fd_complete);
232 if (task_work_add(task, &msg->tw, TWA_SIGNAL))
233 return -EOWNERDEAD;
234
235 return IOU_ISSUE_SKIP_COMPLETE;
236}
237
238static int io_msg_send_fd(struct io_kiocb *req, unsigned int issue_flags)
239{
240 struct io_ring_ctx *target_ctx = req->file->private_data;
241 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
242 struct io_ring_ctx *ctx = req->ctx;
243
244 if (msg->len)
245 return -EINVAL;
246 if (target_ctx == ctx)
247 return -EINVAL;
248 if (target_ctx->flags & IORING_SETUP_R_DISABLED)
249 return -EBADFD;
250 if (!msg->src_file) {
251 int ret = io_msg_grab_file(req, issue_flags);
252 if (unlikely(ret))
253 return ret;
254 }
255
256 if (io_msg_need_remote(target_ctx))
257 return io_msg_fd_remote(req);
258 return io_msg_install_complete(req, issue_flags);
259}
260
261static int __io_msg_ring_prep(struct io_msg *msg, const struct io_uring_sqe *sqe)
262{
263 if (unlikely(sqe->buf_index || sqe->personality))
264 return -EINVAL;
265
266 msg->src_file = NULL;
267 msg->user_data = READ_ONCE(sqe->off);
268 msg->len = READ_ONCE(sqe->len);
269 msg->cmd = READ_ONCE(sqe->addr);
270 msg->src_fd = READ_ONCE(sqe->addr3);
271 msg->dst_fd = READ_ONCE(sqe->file_index);
272 msg->flags = READ_ONCE(sqe->msg_ring_flags);
273 if (msg->flags & ~IORING_MSG_RING_MASK)
274 return -EINVAL;
275
276 return 0;
277}
278
279int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
280{
281 return __io_msg_ring_prep(io_kiocb_to_cmd(req, struct io_msg), sqe);
282}
283
284int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
285{
286 struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
287 int ret;
288
289 ret = -EBADFD;
290 if (!io_is_uring_fops(req->file))
291 goto done;
292
293 switch (msg->cmd) {
294 case IORING_MSG_DATA:
295 ret = io_msg_ring_data(req, issue_flags);
296 break;
297 case IORING_MSG_SEND_FD:
298 ret = io_msg_send_fd(req, issue_flags);
299 break;
300 default:
301 ret = -EINVAL;
302 break;
303 }
304
305done:
306 if (ret < 0) {
307 if (ret == -EAGAIN || ret == IOU_ISSUE_SKIP_COMPLETE)
308 return ret;
309 req_set_fail(req);
310 }
311 io_req_set_res(req, ret, 0);
312 return IOU_COMPLETE;
313}
314
315int io_uring_sync_msg_ring(struct io_uring_sqe *sqe)
316{
317 struct io_msg io_msg = { };
318 int ret;
319
320 ret = __io_msg_ring_prep(&io_msg, sqe);
321 if (unlikely(ret))
322 return ret;
323
324 /*
325 * Only data sending supported, not IORING_MSG_SEND_FD as that one
326 * doesn't make sense without a source ring to send files from.
327 */
328 if (io_msg.cmd != IORING_MSG_DATA)
329 return -EINVAL;
330
331 CLASS(fd, f)(sqe->fd);
332 if (fd_empty(f))
333 return -EBADF;
334 if (!io_is_uring_fops(fd_file(f)))
335 return -EBADFD;
336 return __io_msg_ring_data(fd_file(f)->private_data,
337 &io_msg, IO_URING_F_UNLOCKED);
338}