xref: /linux/io_uring/uring_cmd.c (revision 7fe03f8ff55d33fe6398637f78a8620dd2a78b38)
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/io_uring/cmd.h>
6 #include <linux/io_uring/net.h>
7 #include <linux/security.h>
8 #include <linux/nospec.h>
9 #include <net/sock.h>
10 
11 #include <uapi/linux/io_uring.h>
12 #include <asm/ioctls.h>
13 
14 #include "io_uring.h"
15 #include "alloc_cache.h"
16 #include "rsrc.h"
17 #include "uring_cmd.h"
18 
19 static void io_req_uring_cleanup(struct io_kiocb *req, unsigned int issue_flags)
20 {
21 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
22 	struct io_uring_cmd_data *cache = req->async_data;
23 
24 	if (cache->op_data) {
25 		kfree(cache->op_data);
26 		cache->op_data = NULL;
27 	}
28 
29 	if (issue_flags & IO_URING_F_UNLOCKED)
30 		return;
31 	if (io_alloc_cache_put(&req->ctx->uring_cache, cache)) {
32 		ioucmd->sqe = NULL;
33 		req->async_data = NULL;
34 		req->flags &= ~REQ_F_ASYNC_DATA;
35 	}
36 }
37 
38 bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
39 				   struct io_uring_task *tctx, bool cancel_all)
40 {
41 	struct hlist_node *tmp;
42 	struct io_kiocb *req;
43 	bool ret = false;
44 
45 	lockdep_assert_held(&ctx->uring_lock);
46 
47 	hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
48 			hash_node) {
49 		struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
50 				struct io_uring_cmd);
51 		struct file *file = req->file;
52 
53 		if (!cancel_all && req->tctx != tctx)
54 			continue;
55 
56 		if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
57 			/* ->sqe isn't available if no async data */
58 			if (!req_has_async_data(req))
59 				cmd->sqe = NULL;
60 			file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL |
61 						   IO_URING_F_COMPLETE_DEFER);
62 			ret = true;
63 		}
64 	}
65 	io_submit_flush_completions(ctx);
66 	return ret;
67 }
68 
69 static void io_uring_cmd_del_cancelable(struct io_uring_cmd *cmd,
70 		unsigned int issue_flags)
71 {
72 	struct io_kiocb *req = cmd_to_io_kiocb(cmd);
73 	struct io_ring_ctx *ctx = req->ctx;
74 
75 	if (!(cmd->flags & IORING_URING_CMD_CANCELABLE))
76 		return;
77 
78 	cmd->flags &= ~IORING_URING_CMD_CANCELABLE;
79 	io_ring_submit_lock(ctx, issue_flags);
80 	hlist_del(&req->hash_node);
81 	io_ring_submit_unlock(ctx, issue_flags);
82 }
83 
84 /*
85  * Mark this command as concelable, then io_uring_try_cancel_uring_cmd()
86  * will try to cancel this issued command by sending ->uring_cmd() with
87  * issue_flags of IO_URING_F_CANCEL.
88  *
89  * The command is guaranteed to not be done when calling ->uring_cmd()
90  * with IO_URING_F_CANCEL, but it is driver's responsibility to deal
91  * with race between io_uring canceling and normal completion.
92  */
93 void io_uring_cmd_mark_cancelable(struct io_uring_cmd *cmd,
94 		unsigned int issue_flags)
95 {
96 	struct io_kiocb *req = cmd_to_io_kiocb(cmd);
97 	struct io_ring_ctx *ctx = req->ctx;
98 
99 	if (!(cmd->flags & IORING_URING_CMD_CANCELABLE)) {
100 		cmd->flags |= IORING_URING_CMD_CANCELABLE;
101 		io_ring_submit_lock(ctx, issue_flags);
102 		hlist_add_head(&req->hash_node, &ctx->cancelable_uring_cmd);
103 		io_ring_submit_unlock(ctx, issue_flags);
104 	}
105 }
106 EXPORT_SYMBOL_GPL(io_uring_cmd_mark_cancelable);
107 
108 static void io_uring_cmd_work(struct io_kiocb *req, struct io_tw_state *ts)
109 {
110 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
111 	unsigned int flags = IO_URING_F_COMPLETE_DEFER;
112 
113 	if (io_should_terminate_tw())
114 		flags |= IO_URING_F_TASK_DEAD;
115 
116 	/* task_work executor checks the deffered list completion */
117 	ioucmd->task_work_cb(ioucmd, flags);
118 }
119 
120 void __io_uring_cmd_do_in_task(struct io_uring_cmd *ioucmd,
121 			void (*task_work_cb)(struct io_uring_cmd *, unsigned),
122 			unsigned flags)
123 {
124 	struct io_kiocb *req = cmd_to_io_kiocb(ioucmd);
125 
126 	ioucmd->task_work_cb = task_work_cb;
127 	req->io_task_work.func = io_uring_cmd_work;
128 	__io_req_task_work_add(req, flags);
129 }
130 EXPORT_SYMBOL_GPL(__io_uring_cmd_do_in_task);
131 
132 static inline void io_req_set_cqe32_extra(struct io_kiocb *req,
133 					  u64 extra1, u64 extra2)
134 {
135 	req->big_cqe.extra1 = extra1;
136 	req->big_cqe.extra2 = extra2;
137 }
138 
139 /*
140  * Called by consumers of io_uring_cmd, if they originally returned
141  * -EIOCBQUEUED upon receiving the command.
142  */
143 void io_uring_cmd_done(struct io_uring_cmd *ioucmd, ssize_t ret, u64 res2,
144 		       unsigned issue_flags)
145 {
146 	struct io_kiocb *req = cmd_to_io_kiocb(ioucmd);
147 
148 	io_uring_cmd_del_cancelable(ioucmd, issue_flags);
149 
150 	if (ret < 0)
151 		req_set_fail(req);
152 
153 	io_req_set_res(req, ret, 0);
154 	if (req->ctx->flags & IORING_SETUP_CQE32)
155 		io_req_set_cqe32_extra(req, res2, 0);
156 	io_req_uring_cleanup(req, issue_flags);
157 	if (req->ctx->flags & IORING_SETUP_IOPOLL) {
158 		/* order with io_iopoll_req_issued() checking ->iopoll_complete */
159 		smp_store_release(&req->iopoll_completed, 1);
160 	} else if (issue_flags & IO_URING_F_COMPLETE_DEFER) {
161 		if (WARN_ON_ONCE(issue_flags & IO_URING_F_UNLOCKED))
162 			return;
163 		io_req_complete_defer(req);
164 	} else {
165 		req->io_task_work.func = io_req_task_complete;
166 		io_req_task_work_add(req);
167 	}
168 }
169 EXPORT_SYMBOL_GPL(io_uring_cmd_done);
170 
171 static int io_uring_cmd_prep_setup(struct io_kiocb *req,
172 				   const struct io_uring_sqe *sqe)
173 {
174 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
175 	struct io_uring_cmd_data *cache;
176 
177 	cache = io_uring_alloc_async_data(&req->ctx->uring_cache, req);
178 	if (!cache)
179 		return -ENOMEM;
180 	cache->op_data = NULL;
181 
182 	if (!(req->flags & REQ_F_FORCE_ASYNC)) {
183 		/* defer memcpy until we need it */
184 		ioucmd->sqe = sqe;
185 		return 0;
186 	}
187 
188 	memcpy(cache->sqes, sqe, uring_sqe_size(req->ctx));
189 	ioucmd->sqe = cache->sqes;
190 	return 0;
191 }
192 
193 int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
194 {
195 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
196 
197 	if (sqe->__pad1)
198 		return -EINVAL;
199 
200 	ioucmd->flags = READ_ONCE(sqe->uring_cmd_flags);
201 	if (ioucmd->flags & ~IORING_URING_CMD_MASK)
202 		return -EINVAL;
203 
204 	if (ioucmd->flags & IORING_URING_CMD_FIXED) {
205 		struct io_ring_ctx *ctx = req->ctx;
206 		struct io_rsrc_node *node;
207 		u16 index = READ_ONCE(sqe->buf_index);
208 
209 		node = io_rsrc_node_lookup(&ctx->buf_table, index);
210 		if (unlikely(!node))
211 			return -EFAULT;
212 		/*
213 		 * Pi node upfront, prior to io_uring_cmd_import_fixed()
214 		 * being called. This prevents destruction of the mapped buffer
215 		 * we'll need at actual import time.
216 		 */
217 		io_req_assign_buf_node(req, node);
218 	}
219 	ioucmd->cmd_op = READ_ONCE(sqe->cmd_op);
220 
221 	return io_uring_cmd_prep_setup(req, sqe);
222 }
223 
224 int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
225 {
226 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
227 	struct io_ring_ctx *ctx = req->ctx;
228 	struct file *file = req->file;
229 	int ret;
230 
231 	if (!file->f_op->uring_cmd)
232 		return -EOPNOTSUPP;
233 
234 	ret = security_uring_cmd(ioucmd);
235 	if (ret)
236 		return ret;
237 
238 	if (ctx->flags & IORING_SETUP_SQE128)
239 		issue_flags |= IO_URING_F_SQE128;
240 	if (ctx->flags & IORING_SETUP_CQE32)
241 		issue_flags |= IO_URING_F_CQE32;
242 	if (ctx->compat)
243 		issue_flags |= IO_URING_F_COMPAT;
244 	if (ctx->flags & IORING_SETUP_IOPOLL) {
245 		if (!file->f_op->uring_cmd_iopoll)
246 			return -EOPNOTSUPP;
247 		issue_flags |= IO_URING_F_IOPOLL;
248 		req->iopoll_completed = 0;
249 	}
250 
251 	ret = file->f_op->uring_cmd(ioucmd, issue_flags);
252 	if (ret == -EAGAIN) {
253 		struct io_uring_cmd_data *cache = req->async_data;
254 
255 		if (ioucmd->sqe != (void *) cache)
256 			memcpy(cache->sqes, ioucmd->sqe, uring_sqe_size(req->ctx));
257 		return -EAGAIN;
258 	} else if (ret == -EIOCBQUEUED) {
259 		return -EIOCBQUEUED;
260 	}
261 
262 	if (ret < 0)
263 		req_set_fail(req);
264 	io_req_uring_cleanup(req, issue_flags);
265 	io_req_set_res(req, ret, 0);
266 	return IOU_OK;
267 }
268 
269 int io_uring_cmd_import_fixed(u64 ubuf, unsigned long len, int rw,
270 			      struct iov_iter *iter, void *ioucmd)
271 {
272 	struct io_kiocb *req = cmd_to_io_kiocb(ioucmd);
273 	struct io_rsrc_node *node = req->buf_node;
274 
275 	/* Must have had rsrc_node assigned at prep time */
276 	if (node)
277 		return io_import_fixed(rw, iter, node->buf, ubuf, len);
278 
279 	return -EFAULT;
280 }
281 EXPORT_SYMBOL_GPL(io_uring_cmd_import_fixed);
282 
283 void io_uring_cmd_issue_blocking(struct io_uring_cmd *ioucmd)
284 {
285 	struct io_kiocb *req = cmd_to_io_kiocb(ioucmd);
286 
287 	io_req_queue_iowq(req);
288 }
289 
290 static inline int io_uring_cmd_getsockopt(struct socket *sock,
291 					  struct io_uring_cmd *cmd,
292 					  unsigned int issue_flags)
293 {
294 	bool compat = !!(issue_flags & IO_URING_F_COMPAT);
295 	int optlen, optname, level, err;
296 	void __user *optval;
297 
298 	level = READ_ONCE(cmd->sqe->level);
299 	if (level != SOL_SOCKET)
300 		return -EOPNOTSUPP;
301 
302 	optval = u64_to_user_ptr(READ_ONCE(cmd->sqe->optval));
303 	optname = READ_ONCE(cmd->sqe->optname);
304 	optlen = READ_ONCE(cmd->sqe->optlen);
305 
306 	err = do_sock_getsockopt(sock, compat, level, optname,
307 				 USER_SOCKPTR(optval),
308 				 KERNEL_SOCKPTR(&optlen));
309 	if (err)
310 		return err;
311 
312 	/* On success, return optlen */
313 	return optlen;
314 }
315 
316 static inline int io_uring_cmd_setsockopt(struct socket *sock,
317 					  struct io_uring_cmd *cmd,
318 					  unsigned int issue_flags)
319 {
320 	bool compat = !!(issue_flags & IO_URING_F_COMPAT);
321 	int optname, optlen, level;
322 	void __user *optval;
323 	sockptr_t optval_s;
324 
325 	optval = u64_to_user_ptr(READ_ONCE(cmd->sqe->optval));
326 	optname = READ_ONCE(cmd->sqe->optname);
327 	optlen = READ_ONCE(cmd->sqe->optlen);
328 	level = READ_ONCE(cmd->sqe->level);
329 	optval_s = USER_SOCKPTR(optval);
330 
331 	return do_sock_setsockopt(sock, compat, level, optname, optval_s,
332 				  optlen);
333 }
334 
335 #if defined(CONFIG_NET)
336 int io_uring_cmd_sock(struct io_uring_cmd *cmd, unsigned int issue_flags)
337 {
338 	struct socket *sock = cmd->file->private_data;
339 	struct sock *sk = sock->sk;
340 	struct proto *prot = READ_ONCE(sk->sk_prot);
341 	int ret, arg = 0;
342 
343 	if (!prot || !prot->ioctl)
344 		return -EOPNOTSUPP;
345 
346 	switch (cmd->cmd_op) {
347 	case SOCKET_URING_OP_SIOCINQ:
348 		ret = prot->ioctl(sk, SIOCINQ, &arg);
349 		if (ret)
350 			return ret;
351 		return arg;
352 	case SOCKET_URING_OP_SIOCOUTQ:
353 		ret = prot->ioctl(sk, SIOCOUTQ, &arg);
354 		if (ret)
355 			return ret;
356 		return arg;
357 	case SOCKET_URING_OP_GETSOCKOPT:
358 		return io_uring_cmd_getsockopt(sock, cmd, issue_flags);
359 	case SOCKET_URING_OP_SETSOCKOPT:
360 		return io_uring_cmd_setsockopt(sock, cmd, issue_flags);
361 	default:
362 		return -EOPNOTSUPP;
363 	}
364 }
365 EXPORT_SYMBOL_GPL(io_uring_cmd_sock);
366 #endif
367