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
19static 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
24struct 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
35static void io_waitid_free(struct io_kiocb *req)
36{
37 struct io_waitid_async *iwa = req->async_data;
38
39 put_pid(pid: iwa->wo.wo_pid);
40 kfree(objp: req->async_data);
41 req->async_data = NULL;
42 req->flags &= ~REQ_F_ASYNC_DATA;
43}
44
45static bool io_waitid_compat_copy_si(struct io_waitid *iw, int signo)
46{
47 struct compat_siginfo __user *infop;
48 bool ret;
49
50 infop = (struct compat_siginfo __user *) iw->infop;
51
52 if (!user_write_access_begin(infop, sizeof(*infop)))
53 return false;
54
55 unsafe_put_user(signo, &infop->si_signo, Efault);
56 unsafe_put_user(0, &infop->si_errno, Efault);
57 unsafe_put_user(iw->info.cause, &infop->si_code, Efault);
58 unsafe_put_user(iw->info.pid, &infop->si_pid, Efault);
59 unsafe_put_user(iw->info.uid, &infop->si_uid, Efault);
60 unsafe_put_user(iw->info.status, &infop->si_status, Efault);
61 ret = true;
62done:
63 user_write_access_end();
64 return ret;
65Efault:
66 ret = false;
67 goto done;
68}
69
70static bool io_waitid_copy_si(struct io_kiocb *req, int signo)
71{
72 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
73 bool ret;
74
75 if (!iw->infop)
76 return true;
77
78 if (io_is_compat(ctx: req->ctx))
79 return io_waitid_compat_copy_si(iw, signo);
80
81 if (!user_write_access_begin(iw->infop, sizeof(*iw->infop)))
82 return false;
83
84 unsafe_put_user(signo, &iw->infop->si_signo, Efault);
85 unsafe_put_user(0, &iw->infop->si_errno, Efault);
86 unsafe_put_user(iw->info.cause, &iw->infop->si_code, Efault);
87 unsafe_put_user(iw->info.pid, &iw->infop->si_pid, Efault);
88 unsafe_put_user(iw->info.uid, &iw->infop->si_uid, Efault);
89 unsafe_put_user(iw->info.status, &iw->infop->si_status, Efault);
90 ret = true;
91done:
92 user_write_access_end();
93 return ret;
94Efault:
95 ret = false;
96 goto done;
97}
98
99static int io_waitid_finish(struct io_kiocb *req, int ret)
100{
101 int signo = 0;
102
103 if (ret > 0) {
104 signo = SIGCHLD;
105 ret = 0;
106 }
107
108 if (!io_waitid_copy_si(req, signo))
109 ret = -EFAULT;
110 io_waitid_free(req);
111 return ret;
112}
113
114static void io_waitid_complete(struct io_kiocb *req, int ret)
115{
116 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
117
118 /* anyone completing better be holding a reference */
119 WARN_ON_ONCE(!(atomic_read(&iw->refs) & IO_WAITID_REF_MASK));
120
121 lockdep_assert_held(&req->ctx->uring_lock);
122
123 hlist_del_init(n: &req->hash_node);
124
125 ret = io_waitid_finish(req, ret);
126 if (ret < 0)
127 req_set_fail(req);
128 io_req_set_res(req, res: ret, cflags: 0);
129}
130
131static bool __io_waitid_cancel(struct io_kiocb *req)
132{
133 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
134 struct io_waitid_async *iwa = req->async_data;
135
136 /*
137 * Mark us canceled regardless of ownership. This will prevent a
138 * potential retry from a spurious wakeup.
139 */
140 atomic_or(IO_WAITID_CANCEL_FLAG, v: &iw->refs);
141
142 /* claim ownership */
143 if (atomic_fetch_inc(v: &iw->refs) & IO_WAITID_REF_MASK)
144 return false;
145
146 spin_lock_irq(lock: &iw->head->lock);
147 list_del_init(entry: &iwa->wo.child_wait.entry);
148 spin_unlock_irq(lock: &iw->head->lock);
149 io_waitid_complete(req, ret: -ECANCELED);
150 io_req_queue_tw_complete(req, res: -ECANCELED);
151 return true;
152}
153
154int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
155 unsigned int issue_flags)
156{
157 return io_cancel_remove(ctx, cd, issue_flags, list: &ctx->waitid_list, cancel: __io_waitid_cancel);
158}
159
160bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
161 bool cancel_all)
162{
163 return io_cancel_remove_all(ctx, tctx, list: &ctx->waitid_list, cancel_all, cancel: __io_waitid_cancel);
164}
165
166static inline bool io_waitid_drop_issue_ref(struct io_kiocb *req)
167{
168 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
169 struct io_waitid_async *iwa = req->async_data;
170
171 if (!atomic_sub_return(i: 1, v: &iw->refs))
172 return false;
173
174 /*
175 * Wakeup triggered, racing with us. It was prevented from
176 * completing because of that, queue up the tw to do that.
177 */
178 req->io_task_work.func = io_waitid_cb;
179 io_req_task_work_add(req);
180 remove_wait_queue(wq_head: iw->head, wq_entry: &iwa->wo.child_wait);
181 return true;
182}
183
184static void io_waitid_cb(struct io_kiocb *req, io_tw_token_t tw)
185{
186 struct io_waitid_async *iwa = req->async_data;
187 struct io_ring_ctx *ctx = req->ctx;
188 int ret;
189
190 io_tw_lock(ctx, tw);
191
192 ret = __do_wait(wo: &iwa->wo);
193
194 /*
195 * If we get -ERESTARTSYS here, we need to re-arm and check again
196 * to ensure we get another callback. If the retry works, then we can
197 * just remove ourselves from the waitqueue again and finish the
198 * request.
199 */
200 if (unlikely(ret == -ERESTARTSYS)) {
201 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
202
203 /* Don't retry if cancel found it meanwhile */
204 ret = -ECANCELED;
205 if (!(atomic_read(v: &iw->refs) & IO_WAITID_CANCEL_FLAG)) {
206 iw->head = &current->signal->wait_chldexit;
207 add_wait_queue(wq_head: iw->head, wq_entry: &iwa->wo.child_wait);
208 ret = __do_wait(wo: &iwa->wo);
209 if (ret == -ERESTARTSYS) {
210 /* retry armed, drop our ref */
211 io_waitid_drop_issue_ref(req);
212 return;
213 }
214
215 remove_wait_queue(wq_head: iw->head, wq_entry: &iwa->wo.child_wait);
216 }
217 }
218
219 io_waitid_complete(req, ret);
220 io_req_task_complete(req, tw);
221}
222
223static int io_waitid_wait(struct wait_queue_entry *wait, unsigned mode,
224 int sync, void *key)
225{
226 struct wait_opts *wo = container_of(wait, struct wait_opts, child_wait);
227 struct io_waitid_async *iwa = container_of(wo, struct io_waitid_async, wo);
228 struct io_kiocb *req = iwa->req;
229 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
230 struct task_struct *p = key;
231
232 if (!pid_child_should_wake(wo, p))
233 return 0;
234
235 /* cancel is in progress */
236 if (atomic_fetch_inc(v: &iw->refs) & IO_WAITID_REF_MASK)
237 return 1;
238
239 req->io_task_work.func = io_waitid_cb;
240 io_req_task_work_add(req);
241 list_del_init(entry: &wait->entry);
242 return 1;
243}
244
245int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
246{
247 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
248 struct io_waitid_async *iwa;
249
250 if (sqe->addr || sqe->buf_index || sqe->addr3 || sqe->waitid_flags)
251 return -EINVAL;
252
253 iwa = io_uring_alloc_async_data(NULL, req);
254 if (!unlikely(iwa))
255 return -ENOMEM;
256 iwa->req = req;
257
258 iw->which = READ_ONCE(sqe->len);
259 iw->upid = READ_ONCE(sqe->fd);
260 iw->options = READ_ONCE(sqe->file_index);
261 iw->infop = u64_to_user_ptr(READ_ONCE(sqe->addr2));
262 return 0;
263}
264
265int io_waitid(struct io_kiocb *req, unsigned int issue_flags)
266{
267 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
268 struct io_waitid_async *iwa = req->async_data;
269 struct io_ring_ctx *ctx = req->ctx;
270 int ret;
271
272 ret = kernel_waitid_prepare(wo: &iwa->wo, which: iw->which, upid: iw->upid, infop: &iw->info,
273 options: iw->options, NULL);
274 if (ret)
275 goto done;
276
277 /*
278 * Mark the request as busy upfront, in case we're racing with the
279 * wakeup. If we are, then we'll notice when we drop this initial
280 * reference again after arming.
281 */
282 atomic_set(v: &iw->refs, i: 1);
283
284 /*
285 * Cancel must hold the ctx lock, so there's no risk of cancelation
286 * finding us until a) we remain on the list, and b) the lock is
287 * dropped. We only need to worry about racing with the wakeup
288 * callback.
289 */
290 io_ring_submit_lock(ctx, issue_flags);
291 hlist_add_head(n: &req->hash_node, h: &ctx->waitid_list);
292
293 init_waitqueue_func_entry(wq_entry: &iwa->wo.child_wait, func: io_waitid_wait);
294 iwa->wo.child_wait.private = req->tctx->task;
295 iw->head = &current->signal->wait_chldexit;
296 add_wait_queue(wq_head: iw->head, wq_entry: &iwa->wo.child_wait);
297
298 ret = __do_wait(wo: &iwa->wo);
299 if (ret == -ERESTARTSYS) {
300 /*
301 * Nobody else grabbed a reference, it'll complete when we get
302 * a waitqueue callback, or if someone cancels it.
303 */
304 if (!io_waitid_drop_issue_ref(req)) {
305 io_ring_submit_unlock(ctx, issue_flags);
306 return IOU_ISSUE_SKIP_COMPLETE;
307 }
308
309 /*
310 * Wakeup triggered, racing with us. It was prevented from
311 * completing because of that, queue up the tw to do that.
312 */
313 io_ring_submit_unlock(ctx, issue_flags);
314 return IOU_ISSUE_SKIP_COMPLETE;
315 }
316
317 hlist_del_init(n: &req->hash_node);
318 remove_wait_queue(wq_head: iw->head, wq_entry: &iwa->wo.child_wait);
319 ret = io_waitid_finish(req, ret);
320
321 io_ring_submit_unlock(ctx, issue_flags);
322done:
323 if (ret < 0)
324 req_set_fail(req);
325 io_req_set_res(req, res: ret, cflags: 0);
326 return IOU_COMPLETE;
327}
328

source code of linux/io_uring/waitid.c