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_tw_req tw_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
io_waitid_free(struct io_kiocb * req)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
io_waitid_compat_copy_si(struct io_waitid * iw,int signo)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
io_waitid_copy_si(struct io_kiocb * req,int signo)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
io_waitid_finish(struct io_kiocb * req,int ret)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
io_waitid_remove_wq(struct io_kiocb * req)112 static void io_waitid_remove_wq(struct io_kiocb *req)
113 {
114 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
115 struct wait_queue_head *head;
116
117 head = smp_load_acquire(&iw->head);
118 if (head) {
119 struct io_waitid_async *iwa = req->async_data;
120
121 smp_store_release(&iw->head, NULL);
122 spin_lock_irq(&head->lock);
123 list_del_init(&iwa->wo.child_wait.entry);
124 spin_unlock_irq(&head->lock);
125 }
126 }
127
io_waitid_complete(struct io_kiocb * req,int ret,bool copy_si)128 static void io_waitid_complete(struct io_kiocb *req, int ret, bool copy_si)
129 {
130 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
131
132 /* anyone completing better be holding a reference */
133 WARN_ON_ONCE(!(atomic_read(&iw->refs) & IO_WAITID_REF_MASK));
134
135 lockdep_assert_held(&req->ctx->uring_lock);
136
137 hlist_del_init(&req->hash_node);
138 io_waitid_remove_wq(req);
139
140 if (copy_si)
141 ret = io_waitid_finish(req, ret);
142 else
143 io_waitid_free(req);
144 if (ret < 0)
145 req_set_fail(req);
146 io_req_set_res(req, ret, 0);
147 }
148
__io_waitid_cancel(struct io_kiocb * req,bool copy_si)149 static bool __io_waitid_cancel(struct io_kiocb *req, bool copy_si)
150 {
151 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
152
153 lockdep_assert_held(&req->ctx->uring_lock);
154
155 /*
156 * Mark us canceled regardless of ownership. This will prevent a
157 * potential retry from a spurious wakeup.
158 */
159 atomic_or(IO_WAITID_CANCEL_FLAG, &iw->refs);
160
161 /* claim ownership */
162 if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
163 return false;
164
165 io_waitid_complete(req, -ECANCELED, copy_si);
166 io_req_queue_tw_complete(req, -ECANCELED);
167 return true;
168 }
169
io_waitid_cancel_cb(struct io_kiocb * req)170 static bool io_waitid_cancel_cb(struct io_kiocb *req)
171 {
172 return __io_waitid_cancel(req, true);
173 }
174
io_waitid_cancel_nocopy_cb(struct io_kiocb * req)175 static bool io_waitid_cancel_nocopy_cb(struct io_kiocb *req)
176 {
177 return __io_waitid_cancel(req, false);
178 }
179
io_waitid_cancel(struct io_ring_ctx * ctx,struct io_cancel_data * cd,unsigned int issue_flags)180 int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
181 unsigned int issue_flags)
182 {
183 return io_cancel_remove(ctx, cd, issue_flags, &ctx->waitid_list, io_waitid_cancel_cb);
184 }
185
io_waitid_remove_all(struct io_ring_ctx * ctx,struct io_uring_task * tctx,bool cancel_all)186 bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
187 bool cancel_all)
188 {
189 return io_cancel_remove_all(ctx, tctx, &ctx->waitid_list, cancel_all,
190 tctx ? io_waitid_cancel_cb : io_waitid_cancel_nocopy_cb);
191 }
192
io_waitid_drop_issue_ref(struct io_kiocb * req)193 static inline bool io_waitid_drop_issue_ref(struct io_kiocb *req)
194 {
195 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
196
197 if (!atomic_sub_return(1, &iw->refs))
198 return false;
199
200 io_waitid_remove_wq(req);
201
202 /*
203 * Wakeup triggered, racing with us. It was prevented from
204 * completing because of that, queue up the tw to do that.
205 */
206 req->io_task_work.func = io_waitid_cb;
207 io_req_task_work_add(req);
208 return true;
209 }
210
io_waitid_cb(struct io_tw_req tw_req,io_tw_token_t tw)211 static void io_waitid_cb(struct io_tw_req tw_req, io_tw_token_t tw)
212 {
213 struct io_kiocb *req = tw_req.req;
214 struct io_waitid_async *iwa = req->async_data;
215 struct io_ring_ctx *ctx = req->ctx;
216 int ret;
217
218 io_tw_lock(ctx, tw);
219 if (unlikely(tw.cancel)) {
220 io_waitid_complete(req, -ECANCELED, false);
221 io_req_task_complete(tw_req, tw);
222 return;
223 }
224
225 ret = __do_wait(&iwa->wo);
226
227 /*
228 * If we get -ERESTARTSYS here, we need to re-arm and check again
229 * to ensure we get another callback. If the retry works, then we can
230 * just remove ourselves from the waitqueue again and finish the
231 * request.
232 */
233 if (unlikely(ret == -ERESTARTSYS)) {
234 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
235
236 /* Don't retry if cancel found it meanwhile */
237 ret = -ECANCELED;
238 if (!(atomic_read(&iw->refs) & IO_WAITID_CANCEL_FLAG)) {
239 iw->head = ¤t->signal->wait_chldexit;
240 add_wait_queue(iw->head, &iwa->wo.child_wait);
241 ret = __do_wait(&iwa->wo);
242 if (ret == -ERESTARTSYS) {
243 /* retry armed, drop our ref */
244 io_waitid_drop_issue_ref(req);
245 return;
246 }
247 /* fall through to complete, will kill waitqueue */
248 }
249 }
250
251 io_waitid_complete(req, ret, true);
252 io_req_task_complete(tw_req, tw);
253 }
254
io_waitid_wait(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)255 static int io_waitid_wait(struct wait_queue_entry *wait, unsigned mode,
256 int sync, void *key)
257 {
258 struct wait_opts *wo = container_of(wait, struct wait_opts, child_wait);
259 struct io_waitid_async *iwa = container_of(wo, struct io_waitid_async, wo);
260 struct io_kiocb *req = iwa->req;
261 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
262 struct task_struct *p = key;
263
264 if (!pid_child_should_wake(wo, p))
265 return 0;
266
267 list_del_init(&wait->entry);
268 smp_store_release(&iw->head, NULL);
269
270 /* cancel is in progress */
271 if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
272 return 1;
273
274 req->io_task_work.func = io_waitid_cb;
275 __io_req_task_work_add(req, IOU_F_TWQ_IN_WAKE);
276 return 1;
277 }
278
io_waitid_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)279 int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
280 {
281 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
282 struct io_waitid_async *iwa;
283
284 if (sqe->addr || sqe->buf_index || sqe->addr3 || sqe->waitid_flags)
285 return -EINVAL;
286
287 iwa = io_uring_alloc_async_data(NULL, req);
288 if (unlikely(!iwa))
289 return -ENOMEM;
290 iwa->req = req;
291
292 iw->which = READ_ONCE(sqe->len);
293 iw->upid = READ_ONCE(sqe->fd);
294 iw->options = READ_ONCE(sqe->file_index);
295 iw->head = NULL;
296 iw->infop = u64_to_user_ptr(READ_ONCE(sqe->addr2));
297 memset(&iw->info, 0, sizeof(iw->info));
298 return 0;
299 }
300
io_waitid(struct io_kiocb * req,unsigned int issue_flags)301 int io_waitid(struct io_kiocb *req, unsigned int issue_flags)
302 {
303 struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
304 struct io_waitid_async *iwa = req->async_data;
305 struct io_ring_ctx *ctx = req->ctx;
306 int ret;
307
308 ret = kernel_waitid_prepare(&iwa->wo, iw->which, iw->upid, &iw->info,
309 iw->options, NULL);
310 if (ret)
311 goto done;
312
313 /*
314 * Mark the request as busy upfront, in case we're racing with the
315 * wakeup. If we are, then we'll notice when we drop this initial
316 * reference again after arming.
317 */
318 atomic_set(&iw->refs, 1);
319
320 /*
321 * Cancel must hold the ctx lock, so there's no risk of cancelation
322 * finding us until a) we remain on the list, and b) the lock is
323 * dropped. We only need to worry about racing with the wakeup
324 * callback.
325 */
326 io_ring_submit_lock(ctx, issue_flags);
327
328 /*
329 * iw->head is valid under the ring lock, and as long as the request
330 * is on the waitid_list where cancelations may find it.
331 */
332 iw->head = ¤t->signal->wait_chldexit;
333 hlist_add_head(&req->hash_node, &ctx->waitid_list);
334
335 init_waitqueue_func_entry(&iwa->wo.child_wait, io_waitid_wait);
336 iwa->wo.child_wait.private = req->tctx->task;
337 add_wait_queue(iw->head, &iwa->wo.child_wait);
338
339 ret = __do_wait(&iwa->wo);
340 if (ret == -ERESTARTSYS) {
341 /*
342 * Nobody else grabbed a reference, it'll complete when we get
343 * a waitqueue callback, or if someone cancels it.
344 */
345 if (!io_waitid_drop_issue_ref(req)) {
346 io_ring_submit_unlock(ctx, issue_flags);
347 return IOU_ISSUE_SKIP_COMPLETE;
348 }
349
350 /*
351 * Wakeup triggered, racing with us. It was prevented from
352 * completing because of that, queue up the tw to do that.
353 */
354 io_ring_submit_unlock(ctx, issue_flags);
355 return IOU_ISSUE_SKIP_COMPLETE;
356 }
357
358 hlist_del_init(&req->hash_node);
359 io_waitid_remove_wq(req);
360 ret = io_waitid_finish(req, ret);
361
362 io_ring_submit_unlock(ctx, issue_flags);
363 done:
364 if (ret < 0)
365 req_set_fail(req);
366 io_req_set_res(req, ret, 0);
367 return IOU_COMPLETE;
368 }
369