msg_ring.c

  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	/*
129	 * Keep IORING_SETUP_R_DISABLED check before submitter_task load
130	 * in io_msg_data_remote() -> io_msg_remote_post()
131	 */
132	if (smp_load_acquire(&target_ctx->flags) & IORING_SETUP_R_DISABLED)
133		return -EBADFD;
134
135	if (io_msg_need_remote(target_ctx))
136		return io_msg_data_remote(target_ctx, msg);
137
138	if (msg->flags & IORING_MSG_RING_FLAGS_PASS)
139		flags = msg->cqe_flags;
140
141	ret = -EOVERFLOW;
142	if (target_ctx->flags & IORING_SETUP_IOPOLL) {
143		if (unlikely(io_lock_external_ctx(target_ctx, issue_flags)))
144			return -EAGAIN;
145	}
146	if (io_post_aux_cqe(target_ctx, msg->user_data, msg->len, flags))
147		ret = 0;
148	if (target_ctx->flags & IORING_SETUP_IOPOLL)
149		io_double_unlock_ctx(target_ctx);
150	return ret;
151}
152
153static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags)
154{
155	struct io_ring_ctx *target_ctx = req->file->private_data;
156	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
157
158	return __io_msg_ring_data(target_ctx, msg, issue_flags);
159}
160
161static int io_msg_grab_file(struct io_kiocb *req, unsigned int issue_flags)
162{
163	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
164	struct io_ring_ctx *ctx = req->ctx;
165	struct io_rsrc_node *node;
166	int ret = -EBADF;
167
168	io_ring_submit_lock(ctx, issue_flags);
169	node = io_rsrc_node_lookup(&ctx->file_table.data, msg->src_fd);
170	if (node) {
171		msg->src_file = io_slot_file(node);
172		if (msg->src_file)
173			get_file(msg->src_file);
174		req->flags |= REQ_F_NEED_CLEANUP;
175		ret = 0;
176	}
177	io_ring_submit_unlock(ctx, issue_flags);
178	return ret;
179}
180
181static int io_msg_install_complete(struct io_kiocb *req, unsigned int issue_flags)
182{
183	struct io_ring_ctx *target_ctx = req->file->private_data;
184	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
185	struct file *src_file = msg->src_file;
186	int ret;
187
188	if (unlikely(io_lock_external_ctx(target_ctx, issue_flags)))
189		return -EAGAIN;
190
191	ret = __io_fixed_fd_install(target_ctx, src_file, msg->dst_fd);
192	if (ret < 0)
193		goto out_unlock;
194
195	msg->src_file = NULL;
196	req->flags &= ~REQ_F_NEED_CLEANUP;
197
198	if (msg->flags & IORING_MSG_RING_CQE_SKIP)
199		goto out_unlock;
200	/*
201	 * If this fails, the target still received the file descriptor but
202	 * wasn't notified of the fact. This means that if this request
203	 * completes with -EOVERFLOW, then the sender must ensure that a
204	 * later IORING_OP_MSG_RING delivers the message.
205	 */
206	if (!io_post_aux_cqe(target_ctx, msg->user_data, ret, 0))
207		ret = -EOVERFLOW;
208out_unlock:
209	io_double_unlock_ctx(target_ctx);
210	return ret;
211}
212
213static void io_msg_tw_fd_complete(struct callback_head *head)
214{
215	struct io_msg *msg = container_of(head, struct io_msg, tw);
216	struct io_kiocb *req = cmd_to_io_kiocb(msg);
217	int ret = -EOWNERDEAD;
218
219	if (!(current->flags & PF_EXITING))
220		ret = io_msg_install_complete(req, IO_URING_F_UNLOCKED);
221	if (ret < 0)
222		req_set_fail(req);
223	io_req_queue_tw_complete(req, ret);
224}
225
226static int io_msg_fd_remote(struct io_kiocb *req)
227{
228	struct io_ring_ctx *ctx = req->file->private_data;
229	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
230	struct task_struct *task = READ_ONCE(ctx->submitter_task);
231
232	if (unlikely(!task))
233		return -EOWNERDEAD;
234
235	init_task_work(&msg->tw, io_msg_tw_fd_complete);
236	if (task_work_add(task, &msg->tw, TWA_SIGNAL))
237		return -EOWNERDEAD;
238
239	return IOU_ISSUE_SKIP_COMPLETE;
240}
241
242static int io_msg_send_fd(struct io_kiocb *req, unsigned int issue_flags)
243{
244	struct io_ring_ctx *target_ctx = req->file->private_data;
245	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
246	struct io_ring_ctx *ctx = req->ctx;
247
248	if (msg->len)
249		return -EINVAL;
250	if (target_ctx == ctx)
251		return -EINVAL;
252	/*
253	 * Keep IORING_SETUP_R_DISABLED check before submitter_task load
254	 * in io_msg_fd_remote()
255	 */
256	if (smp_load_acquire(&target_ctx->flags) & IORING_SETUP_R_DISABLED)
257		return -EBADFD;
258	if (!msg->src_file) {
259		int ret = io_msg_grab_file(req, issue_flags);
260		if (unlikely(ret))
261			return ret;
262	}
263
264	if (io_msg_need_remote(target_ctx))
265		return io_msg_fd_remote(req);
266	return io_msg_install_complete(req, issue_flags);
267}
268
269static int __io_msg_ring_prep(struct io_msg *msg, const struct io_uring_sqe *sqe)
270{
271	if (unlikely(sqe->buf_index || sqe->personality))
272		return -EINVAL;
273
274	msg->src_file = NULL;
275	msg->user_data = READ_ONCE(sqe->off);
276	msg->len = READ_ONCE(sqe->len);
277	msg->cmd = READ_ONCE(sqe->addr);
278	msg->src_fd = READ_ONCE(sqe->addr3);
279	msg->dst_fd = READ_ONCE(sqe->file_index);
280	msg->flags = READ_ONCE(sqe->msg_ring_flags);
281	if (msg->flags & ~IORING_MSG_RING_MASK)
282		return -EINVAL;
283
284	return 0;
285}
286
287int io_msg_ring_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
288{
289	return __io_msg_ring_prep(io_kiocb_to_cmd(req, struct io_msg), sqe);
290}
291
292int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
293{
294	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
295	int ret;
296
297	ret = -EBADFD;
298	if (!io_is_uring_fops(req->file))
299		goto done;
300
301	switch (msg->cmd) {
302	case IORING_MSG_DATA:
303		ret = io_msg_ring_data(req, issue_flags);
304		break;
305	case IORING_MSG_SEND_FD:
306		ret = io_msg_send_fd(req, issue_flags);
307		break;
308	default:
309		ret = -EINVAL;
310		break;
311	}
312
313done:
314	if (ret < 0) {
315		if (ret == -EAGAIN || ret == IOU_ISSUE_SKIP_COMPLETE)
316			return ret;
317		req_set_fail(req);
318	}
319	io_req_set_res(req, ret, 0);
320	return IOU_COMPLETE;
321}
322
323int io_uring_sync_msg_ring(struct io_uring_sqe *sqe)
324{
325	struct io_msg io_msg = { };
326	int ret;
327
328	ret = __io_msg_ring_prep(&io_msg, sqe);
329	if (unlikely(ret))
330		return ret;
331
332	/*
333	 * Only data sending supported, not IORING_MSG_SEND_FD as that one
334	 * doesn't make sense without a source ring to send files from.
335	 */
336	if (io_msg.cmd != IORING_MSG_DATA)
337		return -EINVAL;
338
339	CLASS(fd, f)(sqe->fd);
340	if (fd_empty(f))
341		return -EBADF;
342	if (!io_is_uring_fops(fd_file(f)))
343		return -EBADFD;
344	return  __io_msg_ring_data(fd_file(f)->private_data,
345				   &io_msg, IO_URING_F_UNLOCKED);
346}