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
 16struct io_msg {
 17	struct file			*file;
 18	struct file			*src_file;
 19	struct callback_head		tw;
 20	u64 user_data;
 21	u32 len;
 22	u32 cmd;
 23	u32 src_fd;
 24	u32 dst_fd;
 25	u32 flags;
 26};
 27
 28static void io_double_unlock_ctx(struct io_ring_ctx *octx)
 29{
 30	mutex_unlock(&octx->uring_lock);
 31}
 32
 33static int io_double_lock_ctx(struct io_ring_ctx *octx,
 34			      unsigned int issue_flags)
 35{
 36	/*
 37	 * To ensure proper ordering between the two ctxs, we can only
 38	 * attempt a trylock on the target. If that fails and we already have
 39	 * the source ctx lock, punt to io-wq.
 40	 */
 41	if (!(issue_flags & IO_URING_F_UNLOCKED)) {
 42		if (!mutex_trylock(&octx->uring_lock))
 43			return -EAGAIN;
 44		return 0;
 45	}
 46	mutex_lock(&octx->uring_lock);
 47	return 0;
 48}
 49
 50void io_msg_ring_cleanup(struct io_kiocb *req)
 51{
 52	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
 53
 54	if (WARN_ON_ONCE(!msg->src_file))
 55		return;
 56
 57	fput(msg->src_file);
 58	msg->src_file = NULL;
 59}
 60
 61static inline bool io_msg_need_remote(struct io_ring_ctx *target_ctx)
 62{
 63	if (!target_ctx->task_complete)
 64		return false;
 65	return current != target_ctx->submitter_task;
 66}
 67
 68static int io_msg_exec_remote(struct io_kiocb *req, task_work_func_t func)
 69{
 70	struct io_ring_ctx *ctx = req->file->private_data;
 71	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
 72	struct task_struct *task = READ_ONCE(ctx->submitter_task);
 73
 74	if (unlikely(!task))
 75		return -EOWNERDEAD;
 76
 77	init_task_work(&msg->tw, func);
 78	if (task_work_add(ctx->submitter_task, &msg->tw, TWA_SIGNAL))
 79		return -EOWNERDEAD;
 80
 81	return IOU_ISSUE_SKIP_COMPLETE;
 82}
 83
 84static void io_msg_tw_complete(struct callback_head *head)
 85{
 86	struct io_msg *msg = container_of(head, struct io_msg, tw);
 87	struct io_kiocb *req = cmd_to_io_kiocb(msg);
 88	struct io_ring_ctx *target_ctx = req->file->private_data;
 89	int ret = 0;
 90
 91	if (current->flags & PF_EXITING) {
 92		ret = -EOWNERDEAD;
 93	} else {
 94		/*
 95		 * If the target ring is using IOPOLL mode, then we need to be
 96		 * holding the uring_lock for posting completions. Other ring
 97		 * types rely on the regular completion locking, which is
 98		 * handled while posting.
 99		 */
100		if (target_ctx->flags & IORING_SETUP_IOPOLL)
101			mutex_lock(&target_ctx->uring_lock);
102		if (!io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0))
103			ret = -EOVERFLOW;
104		if (target_ctx->flags & IORING_SETUP_IOPOLL)
105			mutex_unlock(&target_ctx->uring_lock);
106	}
107
108	if (ret < 0)
109		req_set_fail(req);
110	io_req_queue_tw_complete(req, ret);
111}
112
113static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags)
114{
115	struct io_ring_ctx *target_ctx = req->file->private_data;
116	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
117	int ret;
118
119	if (msg->src_fd || msg->dst_fd || msg->flags)
120		return -EINVAL;
121	if (target_ctx->flags & IORING_SETUP_R_DISABLED)
122		return -EBADFD;
123
124	if (io_msg_need_remote(target_ctx))
125		return io_msg_exec_remote(req, io_msg_tw_complete);
126
127	ret = -EOVERFLOW;
128	if (target_ctx->flags & IORING_SETUP_IOPOLL) {
129		if (unlikely(io_double_lock_ctx(target_ctx, issue_flags)))
130			return -EAGAIN;
131		if (io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0))
132			ret = 0;
133		io_double_unlock_ctx(target_ctx);
134	} else {
135		if (io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0))
136			ret = 0;
137	}
138	return ret;
139}
140
141static struct file *io_msg_grab_file(struct io_kiocb *req, unsigned int issue_flags)
142{
143	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
144	struct io_ring_ctx *ctx = req->ctx;
145	struct file *file = NULL;
146	unsigned long file_ptr;
147	int idx = msg->src_fd;
148
149	io_ring_submit_lock(ctx, issue_flags);
150	if (likely(idx < ctx->nr_user_files)) {
151		idx = array_index_nospec(idx, ctx->nr_user_files);
152		file_ptr = io_fixed_file_slot(&ctx->file_table, idx)->file_ptr;
153		file = (struct file *) (file_ptr & FFS_MASK);
154		if (file)
155			get_file(file);
156	}
157	io_ring_submit_unlock(ctx, issue_flags);
158	return file;
159}
160
161static int io_msg_install_complete(struct io_kiocb *req, unsigned int issue_flags)
162{
163	struct io_ring_ctx *target_ctx = req->file->private_data;
164	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
165	struct file *src_file = msg->src_file;
166	int ret;
167
168	if (unlikely(io_double_lock_ctx(target_ctx, issue_flags)))
169		return -EAGAIN;
170
171	ret = __io_fixed_fd_install(target_ctx, src_file, msg->dst_fd);
172	if (ret < 0)
173		goto out_unlock;
174
175	msg->src_file = NULL;
176	req->flags &= ~REQ_F_NEED_CLEANUP;
177
178	if (msg->flags & IORING_MSG_RING_CQE_SKIP)
179		goto out_unlock;
180	/*
181	 * If this fails, the target still received the file descriptor but
182	 * wasn't notified of the fact. This means that if this request
183	 * completes with -EOVERFLOW, then the sender must ensure that a
184	 * later IORING_OP_MSG_RING delivers the message.
185	 */
186	if (!io_post_aux_cqe(target_ctx, msg->user_data, msg->len, 0))
187		ret = -EOVERFLOW;
188out_unlock:
189	io_double_unlock_ctx(target_ctx);
190	return ret;
191}
192
193static void io_msg_tw_fd_complete(struct callback_head *head)
194{
195	struct io_msg *msg = container_of(head, struct io_msg, tw);
196	struct io_kiocb *req = cmd_to_io_kiocb(msg);
197	int ret = -EOWNERDEAD;
198
199	if (!(current->flags & PF_EXITING))
200		ret = io_msg_install_complete(req, IO_URING_F_UNLOCKED);
201	if (ret < 0)
202		req_set_fail(req);
203	io_req_queue_tw_complete(req, ret);
204}
205
206static int io_msg_send_fd(struct io_kiocb *req, unsigned int issue_flags)
207{
208	struct io_ring_ctx *target_ctx = req->file->private_data;
209	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
210	struct io_ring_ctx *ctx = req->ctx;
211	struct file *src_file = msg->src_file;
212
213	if (target_ctx == ctx)
214		return -EINVAL;
215	if (target_ctx->flags & IORING_SETUP_R_DISABLED)
216		return -EBADFD;
217	if (!src_file) {
218		src_file = io_msg_grab_file(req, issue_flags);
219		if (!src_file)
220			return -EBADF;
221		msg->src_file = src_file;
222		req->flags |= REQ_F_NEED_CLEANUP;
223	}
224
225	if (io_msg_need_remote(target_ctx))
226		return io_msg_exec_remote(req, io_msg_tw_fd_complete);
227	return io_msg_install_complete(req, issue_flags);
228}
229
230int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
231{
232	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
233
234	if (unlikely(sqe->buf_index || sqe->personality))
235		return -EINVAL;
236
237	msg->src_file = NULL;
238	msg->user_data = READ_ONCE(sqe->off);
239	msg->len = READ_ONCE(sqe->len);
240	msg->cmd = READ_ONCE(sqe->addr);
241	msg->src_fd = READ_ONCE(sqe->addr3);
242	msg->dst_fd = READ_ONCE(sqe->file_index);
243	msg->flags = READ_ONCE(sqe->msg_ring_flags);
244	if (msg->flags & ~IORING_MSG_RING_CQE_SKIP)
245		return -EINVAL;
246
247	return 0;
248}
249
250int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
251{
252	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
253	int ret;
254
255	ret = -EBADFD;
256	if (!io_is_uring_fops(req->file))
257		goto done;
258
259	switch (msg->cmd) {
260	case IORING_MSG_DATA:
261		ret = io_msg_ring_data(req, issue_flags);
262		break;
263	case IORING_MSG_SEND_FD:
264		ret = io_msg_send_fd(req, issue_flags);
265		break;
266	default:
267		ret = -EINVAL;
268		break;
269	}
270
271done:
272	if (ret < 0) {
273		if (ret == -EAGAIN || ret == IOU_ISSUE_SKIP_COMPLETE)
274			return ret;
275		req_set_fail(req);
276	}
277	io_req_set_res(req, ret, 0);
278	return IOU_OK;
279}