xref: /linux/io_uring/waitid.c (revision 5832d26433f2bd0d28f8b12526e3c2fdb203507f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Support for async notification of waitid
4  */
5 #include <linux/kernel.h>
6 #include <linux/errno.h>
7 #include <linux/fs.h>
8 #include <linux/file.h>
9 #include <linux/compat.h>
10 #include <linux/io_uring.h>
11 
12 #include <uapi/linux/io_uring.h>
13 
14 #include "io_uring.h"
15 #include "cancel.h"
16 #include "waitid.h"
17 #include "../kernel/exit.h"
18 
19 static void io_waitid_cb(struct io_kiocb *req, io_tw_token_t tw);
20 
21 #define IO_WAITID_CANCEL_FLAG	BIT(31)
22 #define IO_WAITID_REF_MASK	GENMASK(30, 0)
23 
24 struct io_waitid {
25 	struct file *file;
26 	int which;
27 	pid_t upid;
28 	int options;
29 	atomic_t refs;
30 	struct wait_queue_head *head;
31 	struct siginfo __user *infop;
32 	struct waitid_info info;
33 };
34 
35 static void io_waitid_free(struct io_kiocb *req)
36 {
37 	struct io_waitid_async *iwa = req->async_data;
38 
39 	put_pid(iwa->wo.wo_pid);
40 	io_req_async_data_free(req);
41 }
42 
43 static bool io_waitid_compat_copy_si(struct io_waitid *iw, int signo)
44 {
45 	struct compat_siginfo __user *infop;
46 	bool ret;
47 
48 	infop = (struct compat_siginfo __user *) iw->infop;
49 
50 	if (!user_write_access_begin(infop, sizeof(*infop)))
51 		return false;
52 
53 	unsafe_put_user(signo, &infop->si_signo, Efault);
54 	unsafe_put_user(0, &infop->si_errno, Efault);
55 	unsafe_put_user(iw->info.cause, &infop->si_code, Efault);
56 	unsafe_put_user(iw->info.pid, &infop->si_pid, Efault);
57 	unsafe_put_user(iw->info.uid, &infop->si_uid, Efault);
58 	unsafe_put_user(iw->info.status, &infop->si_status, Efault);
59 	ret = true;
60 done:
61 	user_write_access_end();
62 	return ret;
63 Efault:
64 	ret = false;
65 	goto done;
66 }
67 
68 static bool io_waitid_copy_si(struct io_kiocb *req, int signo)
69 {
70 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
71 	bool ret;
72 
73 	if (!iw->infop)
74 		return true;
75 
76 	if (io_is_compat(req->ctx))
77 		return io_waitid_compat_copy_si(iw, signo);
78 
79 	if (!user_write_access_begin(iw->infop, sizeof(*iw->infop)))
80 		return false;
81 
82 	unsafe_put_user(signo, &iw->infop->si_signo, Efault);
83 	unsafe_put_user(0, &iw->infop->si_errno, Efault);
84 	unsafe_put_user(iw->info.cause, &iw->infop->si_code, Efault);
85 	unsafe_put_user(iw->info.pid, &iw->infop->si_pid, Efault);
86 	unsafe_put_user(iw->info.uid, &iw->infop->si_uid, Efault);
87 	unsafe_put_user(iw->info.status, &iw->infop->si_status, Efault);
88 	ret = true;
89 done:
90 	user_write_access_end();
91 	return ret;
92 Efault:
93 	ret = false;
94 	goto done;
95 }
96 
97 static int io_waitid_finish(struct io_kiocb *req, int ret)
98 {
99 	int signo = 0;
100 
101 	if (ret > 0) {
102 		signo = SIGCHLD;
103 		ret = 0;
104 	}
105 
106 	if (!io_waitid_copy_si(req, signo))
107 		ret = -EFAULT;
108 	io_waitid_free(req);
109 	return ret;
110 }
111 
112 static void io_waitid_complete(struct io_kiocb *req, int ret)
113 {
114 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
115 
116 	/* anyone completing better be holding a reference */
117 	WARN_ON_ONCE(!(atomic_read(&iw->refs) & IO_WAITID_REF_MASK));
118 
119 	lockdep_assert_held(&req->ctx->uring_lock);
120 
121 	hlist_del_init(&req->hash_node);
122 
123 	ret = io_waitid_finish(req, ret);
124 	if (ret < 0)
125 		req_set_fail(req);
126 	io_req_set_res(req, ret, 0);
127 }
128 
129 static bool __io_waitid_cancel(struct io_kiocb *req)
130 {
131 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
132 	struct io_waitid_async *iwa = req->async_data;
133 
134 	/*
135 	 * Mark us canceled regardless of ownership. This will prevent a
136 	 * potential retry from a spurious wakeup.
137 	 */
138 	atomic_or(IO_WAITID_CANCEL_FLAG, &iw->refs);
139 
140 	/* claim ownership */
141 	if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
142 		return false;
143 
144 	spin_lock_irq(&iw->head->lock);
145 	list_del_init(&iwa->wo.child_wait.entry);
146 	spin_unlock_irq(&iw->head->lock);
147 	io_waitid_complete(req, -ECANCELED);
148 	io_req_queue_tw_complete(req, -ECANCELED);
149 	return true;
150 }
151 
152 int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
153 		     unsigned int issue_flags)
154 {
155 	return io_cancel_remove(ctx, cd, issue_flags, &ctx->waitid_list, __io_waitid_cancel);
156 }
157 
158 bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
159 			  bool cancel_all)
160 {
161 	return io_cancel_remove_all(ctx, tctx, &ctx->waitid_list, cancel_all, __io_waitid_cancel);
162 }
163 
164 static inline bool io_waitid_drop_issue_ref(struct io_kiocb *req)
165 {
166 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
167 	struct io_waitid_async *iwa = req->async_data;
168 
169 	if (!atomic_sub_return(1, &iw->refs))
170 		return false;
171 
172 	/*
173 	 * Wakeup triggered, racing with us. It was prevented from
174 	 * completing because of that, queue up the tw to do that.
175 	 */
176 	req->io_task_work.func = io_waitid_cb;
177 	io_req_task_work_add(req);
178 	remove_wait_queue(iw->head, &iwa->wo.child_wait);
179 	return true;
180 }
181 
182 static void io_waitid_cb(struct io_kiocb *req, io_tw_token_t tw)
183 {
184 	struct io_waitid_async *iwa = req->async_data;
185 	struct io_ring_ctx *ctx = req->ctx;
186 	int ret;
187 
188 	io_tw_lock(ctx, tw);
189 
190 	ret = __do_wait(&iwa->wo);
191 
192 	/*
193 	 * If we get -ERESTARTSYS here, we need to re-arm and check again
194 	 * to ensure we get another callback. If the retry works, then we can
195 	 * just remove ourselves from the waitqueue again and finish the
196 	 * request.
197 	 */
198 	if (unlikely(ret == -ERESTARTSYS)) {
199 		struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
200 
201 		/* Don't retry if cancel found it meanwhile */
202 		ret = -ECANCELED;
203 		if (!(atomic_read(&iw->refs) & IO_WAITID_CANCEL_FLAG)) {
204 			iw->head = &current->signal->wait_chldexit;
205 			add_wait_queue(iw->head, &iwa->wo.child_wait);
206 			ret = __do_wait(&iwa->wo);
207 			if (ret == -ERESTARTSYS) {
208 				/* retry armed, drop our ref */
209 				io_waitid_drop_issue_ref(req);
210 				return;
211 			}
212 
213 			remove_wait_queue(iw->head, &iwa->wo.child_wait);
214 		}
215 	}
216 
217 	io_waitid_complete(req, ret);
218 	io_req_task_complete(req, tw);
219 }
220 
221 static int io_waitid_wait(struct wait_queue_entry *wait, unsigned mode,
222 			  int sync, void *key)
223 {
224 	struct wait_opts *wo = container_of(wait, struct wait_opts, child_wait);
225 	struct io_waitid_async *iwa = container_of(wo, struct io_waitid_async, wo);
226 	struct io_kiocb *req = iwa->req;
227 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
228 	struct task_struct *p = key;
229 
230 	if (!pid_child_should_wake(wo, p))
231 		return 0;
232 
233 	/* cancel is in progress */
234 	if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
235 		return 1;
236 
237 	req->io_task_work.func = io_waitid_cb;
238 	io_req_task_work_add(req);
239 	list_del_init(&wait->entry);
240 	return 1;
241 }
242 
243 int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
244 {
245 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
246 	struct io_waitid_async *iwa;
247 
248 	if (sqe->addr || sqe->buf_index || sqe->addr3 || sqe->waitid_flags)
249 		return -EINVAL;
250 
251 	iwa = io_uring_alloc_async_data(NULL, req);
252 	if (!unlikely(iwa))
253 		return -ENOMEM;
254 	iwa->req = req;
255 
256 	iw->which = READ_ONCE(sqe->len);
257 	iw->upid = READ_ONCE(sqe->fd);
258 	iw->options = READ_ONCE(sqe->file_index);
259 	iw->infop = u64_to_user_ptr(READ_ONCE(sqe->addr2));
260 	return 0;
261 }
262 
263 int io_waitid(struct io_kiocb *req, unsigned int issue_flags)
264 {
265 	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
266 	struct io_waitid_async *iwa = req->async_data;
267 	struct io_ring_ctx *ctx = req->ctx;
268 	int ret;
269 
270 	ret = kernel_waitid_prepare(&iwa->wo, iw->which, iw->upid, &iw->info,
271 					iw->options, NULL);
272 	if (ret)
273 		goto done;
274 
275 	/*
276 	 * Mark the request as busy upfront, in case we're racing with the
277 	 * wakeup. If we are, then we'll notice when we drop this initial
278 	 * reference again after arming.
279 	 */
280 	atomic_set(&iw->refs, 1);
281 
282 	/*
283 	 * Cancel must hold the ctx lock, so there's no risk of cancelation
284 	 * finding us until a) we remain on the list, and b) the lock is
285 	 * dropped. We only need to worry about racing with the wakeup
286 	 * callback.
287 	 */
288 	io_ring_submit_lock(ctx, issue_flags);
289 	hlist_add_head(&req->hash_node, &ctx->waitid_list);
290 
291 	init_waitqueue_func_entry(&iwa->wo.child_wait, io_waitid_wait);
292 	iwa->wo.child_wait.private = req->tctx->task;
293 	iw->head = &current->signal->wait_chldexit;
294 	add_wait_queue(iw->head, &iwa->wo.child_wait);
295 
296 	ret = __do_wait(&iwa->wo);
297 	if (ret == -ERESTARTSYS) {
298 		/*
299 		 * Nobody else grabbed a reference, it'll complete when we get
300 		 * a waitqueue callback, or if someone cancels it.
301 		 */
302 		if (!io_waitid_drop_issue_ref(req)) {
303 			io_ring_submit_unlock(ctx, issue_flags);
304 			return IOU_ISSUE_SKIP_COMPLETE;
305 		}
306 
307 		/*
308 		 * Wakeup triggered, racing with us. It was prevented from
309 		 * completing because of that, queue up the tw to do that.
310 		 */
311 		io_ring_submit_unlock(ctx, issue_flags);
312 		return IOU_ISSUE_SKIP_COMPLETE;
313 	}
314 
315 	hlist_del_init(&req->hash_node);
316 	remove_wait_queue(iw->head, &iwa->wo.child_wait);
317 	ret = io_waitid_finish(req, ret);
318 
319 	io_ring_submit_unlock(ctx, issue_flags);
320 done:
321 	if (ret < 0)
322 		req_set_fail(req);
323 	io_req_set_res(req, ret, 0);
324 	return IOU_COMPLETE;
325 }
326