waitid.c

  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_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
 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(iwa->wo.wo_pid);
 40	io_req_async_data_free(req);
 41}
 42
 43static 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;
 60done:
 61	user_write_access_end();
 62	return ret;
 63Efault:
 64	ret = false;
 65	goto done;
 66}
 67
 68static 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;
 89done:
 90	user_write_access_end();
 91	return ret;
 92Efault:
 93	ret = false;
 94	goto done;
 95}
 96
 97static 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
112static 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
128static void io_waitid_complete(struct io_kiocb *req, int ret)
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	ret = io_waitid_finish(req, ret);
141	if (ret < 0)
142		req_set_fail(req);
143	io_req_set_res(req, ret, 0);
144}
145
146static bool __io_waitid_cancel(struct io_kiocb *req)
147{
148	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
149
150	lockdep_assert_held(&req->ctx->uring_lock);
151
152	/*
153	 * Mark us canceled regardless of ownership. This will prevent a
154	 * potential retry from a spurious wakeup.
155	 */
156	atomic_or(IO_WAITID_CANCEL_FLAG, &iw->refs);
157
158	/* claim ownership */
159	if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
160		return false;
161
162	io_waitid_complete(req, -ECANCELED);
163	io_req_queue_tw_complete(req, -ECANCELED);
164	return true;
165}
166
167int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
168		     unsigned int issue_flags)
169{
170	return io_cancel_remove(ctx, cd, issue_flags, &ctx->waitid_list, __io_waitid_cancel);
171}
172
173bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
174			  bool cancel_all)
175{
176	return io_cancel_remove_all(ctx, tctx, &ctx->waitid_list, cancel_all, __io_waitid_cancel);
177}
178
179static inline bool io_waitid_drop_issue_ref(struct io_kiocb *req)
180{
181	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
182
183	if (!atomic_sub_return(1, &iw->refs))
184		return false;
185
186	io_waitid_remove_wq(req);
187
188	/*
189	 * Wakeup triggered, racing with us. It was prevented from
190	 * completing because of that, queue up the tw to do that.
191	 */
192	req->io_task_work.func = io_waitid_cb;
193	io_req_task_work_add(req);
194	return true;
195}
196
197static void io_waitid_cb(struct io_tw_req tw_req, io_tw_token_t tw)
198{
199	struct io_kiocb *req = tw_req.req;
200	struct io_waitid_async *iwa = req->async_data;
201	struct io_ring_ctx *ctx = req->ctx;
202	int ret;
203
204	io_tw_lock(ctx, tw);
205
206	ret = __do_wait(&iwa->wo);
207
208	/*
209	 * If we get -ERESTARTSYS here, we need to re-arm and check again
210	 * to ensure we get another callback. If the retry works, then we can
211	 * just remove ourselves from the waitqueue again and finish the
212	 * request.
213	 */
214	if (unlikely(ret == -ERESTARTSYS)) {
215		struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
216
217		/* Don't retry if cancel found it meanwhile */
218		ret = -ECANCELED;
219		if (!(atomic_read(&iw->refs) & IO_WAITID_CANCEL_FLAG)) {
220			iw->head = &current->signal->wait_chldexit;
221			add_wait_queue(iw->head, &iwa->wo.child_wait);
222			ret = __do_wait(&iwa->wo);
223			if (ret == -ERESTARTSYS) {
224				/* retry armed, drop our ref */
225				io_waitid_drop_issue_ref(req);
226				return;
227			}
228			/* fall through to complete, will kill waitqueue */
229		}
230	}
231
232	io_waitid_complete(req, ret);
233	io_req_task_complete(tw_req, tw);
234}
235
236static int io_waitid_wait(struct wait_queue_entry *wait, unsigned mode,
237			  int sync, void *key)
238{
239	struct wait_opts *wo = container_of(wait, struct wait_opts, child_wait);
240	struct io_waitid_async *iwa = container_of(wo, struct io_waitid_async, wo);
241	struct io_kiocb *req = iwa->req;
242	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
243	struct task_struct *p = key;
244
245	if (!pid_child_should_wake(wo, p))
246		return 0;
247
248	list_del_init(&wait->entry);
249	smp_store_release(&iw->head, NULL);
250
251	/* cancel is in progress */
252	if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
253		return 1;
254
255	req->io_task_work.func = io_waitid_cb;
256	io_req_task_work_add(req);
257	return 1;
258}
259
260int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
261{
262	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
263	struct io_waitid_async *iwa;
264
265	if (sqe->addr || sqe->buf_index || sqe->addr3 || sqe->waitid_flags)
266		return -EINVAL;
267
268	iwa = io_uring_alloc_async_data(NULL, req);
269	if (unlikely(!iwa))
270		return -ENOMEM;
271	iwa->req = req;
272
273	iw->which = READ_ONCE(sqe->len);
274	iw->upid = READ_ONCE(sqe->fd);
275	iw->options = READ_ONCE(sqe->file_index);
276	iw->head = NULL;
277	iw->infop = u64_to_user_ptr(READ_ONCE(sqe->addr2));
278	return 0;
279}
280
281int io_waitid(struct io_kiocb *req, unsigned int issue_flags)
282{
283	struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
284	struct io_waitid_async *iwa = req->async_data;
285	struct io_ring_ctx *ctx = req->ctx;
286	int ret;
287
288	ret = kernel_waitid_prepare(&iwa->wo, iw->which, iw->upid, &iw->info,
289					iw->options, NULL);
290	if (ret)
291		goto done;
292
293	/*
294	 * Mark the request as busy upfront, in case we're racing with the
295	 * wakeup. If we are, then we'll notice when we drop this initial
296	 * reference again after arming.
297	 */
298	atomic_set(&iw->refs, 1);
299
300	/*
301	 * Cancel must hold the ctx lock, so there's no risk of cancelation
302	 * finding us until a) we remain on the list, and b) the lock is
303	 * dropped. We only need to worry about racing with the wakeup
304	 * callback.
305	 */
306	io_ring_submit_lock(ctx, issue_flags);
307
308	/*
309	 * iw->head is valid under the ring lock, and as long as the request
310	 * is on the waitid_list where cancelations may find it.
311	 */
312	iw->head = &current->signal->wait_chldexit;
313	hlist_add_head(&req->hash_node, &ctx->waitid_list);
314
315	init_waitqueue_func_entry(&iwa->wo.child_wait, io_waitid_wait);
316	iwa->wo.child_wait.private = req->tctx->task;
317	add_wait_queue(iw->head, &iwa->wo.child_wait);
318
319	ret = __do_wait(&iwa->wo);
320	if (ret == -ERESTARTSYS) {
321		/*
322		 * Nobody else grabbed a reference, it'll complete when we get
323		 * a waitqueue callback, or if someone cancels it.
324		 */
325		if (!io_waitid_drop_issue_ref(req)) {
326			io_ring_submit_unlock(ctx, issue_flags);
327			return IOU_ISSUE_SKIP_COMPLETE;
328		}
329
330		/*
331		 * Wakeup triggered, racing with us. It was prevented from
332		 * completing because of that, queue up the tw to do that.
333		 */
334		io_ring_submit_unlock(ctx, issue_flags);
335		return IOU_ISSUE_SKIP_COMPLETE;
336	}
337
338	hlist_del_init(&req->hash_node);
339	io_waitid_remove_wq(req);
340	ret = io_waitid_finish(req, ret);
341
342	io_ring_submit_unlock(ctx, issue_flags);
343done:
344	if (ret < 0)
345		req_set_fail(req);
346	io_req_set_res(req, ret, 0);
347	return IOU_COMPLETE;
348}