xref: /linux/io_uring/futex.c (revision 1e525507)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/fs.h>
5 #include <linux/file.h>
6 #include <linux/io_uring.h>
7 
8 #include <uapi/linux/io_uring.h>
9 
10 #include "../kernel/futex/futex.h"
11 #include "io_uring.h"
12 #include "rsrc.h"
13 #include "futex.h"
14 
15 struct io_futex {
16 	struct file	*file;
17 	union {
18 		u32 __user			*uaddr;
19 		struct futex_waitv __user	*uwaitv;
20 	};
21 	unsigned long	futex_val;
22 	unsigned long	futex_mask;
23 	unsigned long	futexv_owned;
24 	u32		futex_flags;
25 	unsigned int	futex_nr;
26 	bool		futexv_unqueued;
27 };
28 
29 struct io_futex_data {
30 	union {
31 		struct futex_q		q;
32 		struct io_cache_entry	cache;
33 	};
34 	struct io_kiocb	*req;
35 };
36 
37 void io_futex_cache_init(struct io_ring_ctx *ctx)
38 {
39 	io_alloc_cache_init(&ctx->futex_cache, IO_NODE_ALLOC_CACHE_MAX,
40 				sizeof(struct io_futex_data));
41 }
42 
43 static void io_futex_cache_entry_free(struct io_cache_entry *entry)
44 {
45 	kfree(container_of(entry, struct io_futex_data, cache));
46 }
47 
48 void io_futex_cache_free(struct io_ring_ctx *ctx)
49 {
50 	io_alloc_cache_free(&ctx->futex_cache, io_futex_cache_entry_free);
51 }
52 
53 static void __io_futex_complete(struct io_kiocb *req, struct io_tw_state *ts)
54 {
55 	req->async_data = NULL;
56 	hlist_del_init(&req->hash_node);
57 	io_req_task_complete(req, ts);
58 }
59 
60 static void io_futex_complete(struct io_kiocb *req, struct io_tw_state *ts)
61 {
62 	struct io_futex_data *ifd = req->async_data;
63 	struct io_ring_ctx *ctx = req->ctx;
64 
65 	io_tw_lock(ctx, ts);
66 	if (!io_alloc_cache_put(&ctx->futex_cache, &ifd->cache))
67 		kfree(ifd);
68 	__io_futex_complete(req, ts);
69 }
70 
71 static void io_futexv_complete(struct io_kiocb *req, struct io_tw_state *ts)
72 {
73 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
74 	struct futex_vector *futexv = req->async_data;
75 
76 	io_tw_lock(req->ctx, ts);
77 
78 	if (!iof->futexv_unqueued) {
79 		int res;
80 
81 		res = futex_unqueue_multiple(futexv, iof->futex_nr);
82 		if (res != -1)
83 			io_req_set_res(req, res, 0);
84 	}
85 
86 	kfree(req->async_data);
87 	req->flags &= ~REQ_F_ASYNC_DATA;
88 	__io_futex_complete(req, ts);
89 }
90 
91 static bool io_futexv_claim(struct io_futex *iof)
92 {
93 	if (test_bit(0, &iof->futexv_owned) ||
94 	    test_and_set_bit_lock(0, &iof->futexv_owned))
95 		return false;
96 	return true;
97 }
98 
99 static bool __io_futex_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
100 {
101 	/* futex wake already done or in progress */
102 	if (req->opcode == IORING_OP_FUTEX_WAIT) {
103 		struct io_futex_data *ifd = req->async_data;
104 
105 		if (!futex_unqueue(&ifd->q))
106 			return false;
107 		req->io_task_work.func = io_futex_complete;
108 	} else {
109 		struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
110 
111 		if (!io_futexv_claim(iof))
112 			return false;
113 		req->io_task_work.func = io_futexv_complete;
114 	}
115 
116 	hlist_del_init(&req->hash_node);
117 	io_req_set_res(req, -ECANCELED, 0);
118 	io_req_task_work_add(req);
119 	return true;
120 }
121 
122 int io_futex_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
123 		    unsigned int issue_flags)
124 {
125 	struct hlist_node *tmp;
126 	struct io_kiocb *req;
127 	int nr = 0;
128 
129 	if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_FD_FIXED))
130 		return -ENOENT;
131 
132 	io_ring_submit_lock(ctx, issue_flags);
133 	hlist_for_each_entry_safe(req, tmp, &ctx->futex_list, hash_node) {
134 		if (req->cqe.user_data != cd->data &&
135 		    !(cd->flags & IORING_ASYNC_CANCEL_ANY))
136 			continue;
137 		if (__io_futex_cancel(ctx, req))
138 			nr++;
139 		if (!(cd->flags & IORING_ASYNC_CANCEL_ALL))
140 			break;
141 	}
142 	io_ring_submit_unlock(ctx, issue_flags);
143 
144 	if (nr)
145 		return nr;
146 
147 	return -ENOENT;
148 }
149 
150 bool io_futex_remove_all(struct io_ring_ctx *ctx, struct task_struct *task,
151 			 bool cancel_all)
152 {
153 	struct hlist_node *tmp;
154 	struct io_kiocb *req;
155 	bool found = false;
156 
157 	lockdep_assert_held(&ctx->uring_lock);
158 
159 	hlist_for_each_entry_safe(req, tmp, &ctx->futex_list, hash_node) {
160 		if (!io_match_task_safe(req, task, cancel_all))
161 			continue;
162 		hlist_del_init(&req->hash_node);
163 		__io_futex_cancel(ctx, req);
164 		found = true;
165 	}
166 
167 	return found;
168 }
169 
170 int io_futex_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
171 {
172 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
173 	u32 flags;
174 
175 	if (unlikely(sqe->len || sqe->futex_flags || sqe->buf_index ||
176 		     sqe->file_index))
177 		return -EINVAL;
178 
179 	iof->uaddr = u64_to_user_ptr(READ_ONCE(sqe->addr));
180 	iof->futex_val = READ_ONCE(sqe->addr2);
181 	iof->futex_mask = READ_ONCE(sqe->addr3);
182 	flags = READ_ONCE(sqe->fd);
183 
184 	if (flags & ~FUTEX2_VALID_MASK)
185 		return -EINVAL;
186 
187 	iof->futex_flags = futex2_to_flags(flags);
188 	if (!futex_flags_valid(iof->futex_flags))
189 		return -EINVAL;
190 
191 	if (!futex_validate_input(iof->futex_flags, iof->futex_val) ||
192 	    !futex_validate_input(iof->futex_flags, iof->futex_mask))
193 		return -EINVAL;
194 
195 	return 0;
196 }
197 
198 static void io_futex_wakev_fn(struct wake_q_head *wake_q, struct futex_q *q)
199 {
200 	struct io_kiocb *req = q->wake_data;
201 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
202 
203 	if (!io_futexv_claim(iof))
204 		return;
205 	if (unlikely(!__futex_wake_mark(q)))
206 		return;
207 
208 	io_req_set_res(req, 0, 0);
209 	req->io_task_work.func = io_futexv_complete;
210 	io_req_task_work_add(req);
211 }
212 
213 int io_futexv_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
214 {
215 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
216 	struct futex_vector *futexv;
217 	int ret;
218 
219 	/* No flags or mask supported for waitv */
220 	if (unlikely(sqe->fd || sqe->buf_index || sqe->file_index ||
221 		     sqe->addr2 || sqe->futex_flags || sqe->addr3))
222 		return -EINVAL;
223 
224 	iof->uaddr = u64_to_user_ptr(READ_ONCE(sqe->addr));
225 	iof->futex_nr = READ_ONCE(sqe->len);
226 	if (!iof->futex_nr || iof->futex_nr > FUTEX_WAITV_MAX)
227 		return -EINVAL;
228 
229 	futexv = kcalloc(iof->futex_nr, sizeof(*futexv), GFP_KERNEL);
230 	if (!futexv)
231 		return -ENOMEM;
232 
233 	ret = futex_parse_waitv(futexv, iof->uwaitv, iof->futex_nr,
234 				io_futex_wakev_fn, req);
235 	if (ret) {
236 		kfree(futexv);
237 		return ret;
238 	}
239 
240 	iof->futexv_owned = 0;
241 	iof->futexv_unqueued = 0;
242 	req->flags |= REQ_F_ASYNC_DATA;
243 	req->async_data = futexv;
244 	return 0;
245 }
246 
247 static void io_futex_wake_fn(struct wake_q_head *wake_q, struct futex_q *q)
248 {
249 	struct io_futex_data *ifd = container_of(q, struct io_futex_data, q);
250 	struct io_kiocb *req = ifd->req;
251 
252 	if (unlikely(!__futex_wake_mark(q)))
253 		return;
254 
255 	io_req_set_res(req, 0, 0);
256 	req->io_task_work.func = io_futex_complete;
257 	io_req_task_work_add(req);
258 }
259 
260 static struct io_futex_data *io_alloc_ifd(struct io_ring_ctx *ctx)
261 {
262 	struct io_cache_entry *entry;
263 
264 	entry = io_alloc_cache_get(&ctx->futex_cache);
265 	if (entry)
266 		return container_of(entry, struct io_futex_data, cache);
267 
268 	return kmalloc(sizeof(struct io_futex_data), GFP_NOWAIT);
269 }
270 
271 int io_futexv_wait(struct io_kiocb *req, unsigned int issue_flags)
272 {
273 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
274 	struct futex_vector *futexv = req->async_data;
275 	struct io_ring_ctx *ctx = req->ctx;
276 	int ret, woken = -1;
277 
278 	io_ring_submit_lock(ctx, issue_flags);
279 
280 	ret = futex_wait_multiple_setup(futexv, iof->futex_nr, &woken);
281 
282 	/*
283 	 * Error case, ret is < 0. Mark the request as failed.
284 	 */
285 	if (unlikely(ret < 0)) {
286 		io_ring_submit_unlock(ctx, issue_flags);
287 		req_set_fail(req);
288 		io_req_set_res(req, ret, 0);
289 		kfree(futexv);
290 		req->async_data = NULL;
291 		req->flags &= ~REQ_F_ASYNC_DATA;
292 		return IOU_OK;
293 	}
294 
295 	/*
296 	 * 0 return means that we successfully setup the waiters, and that
297 	 * nobody triggered a wakeup while we were doing so. If the wakeup
298 	 * happened post setup, the task_work will be run post this issue and
299 	 * under the submission lock. 1 means We got woken while setting up,
300 	 * let that side do the completion. Note that
301 	 * futex_wait_multiple_setup() will have unqueued all the futexes in
302 	 * this case. Mark us as having done that already, since this is
303 	 * different from normal wakeup.
304 	 */
305 	if (!ret) {
306 		/*
307 		 * If futex_wait_multiple_setup() returns 0 for a
308 		 * successful setup, then the task state will not be
309 		 * runnable. This is fine for the sync syscall, as
310 		 * it'll be blocking unless we already got one of the
311 		 * futexes woken, but it obviously won't work for an
312 		 * async invocation. Mark us runnable again.
313 		 */
314 		__set_current_state(TASK_RUNNING);
315 		hlist_add_head(&req->hash_node, &ctx->futex_list);
316 	} else {
317 		iof->futexv_unqueued = 1;
318 		if (woken != -1)
319 			io_req_set_res(req, woken, 0);
320 	}
321 
322 	io_ring_submit_unlock(ctx, issue_flags);
323 	return IOU_ISSUE_SKIP_COMPLETE;
324 }
325 
326 int io_futex_wait(struct io_kiocb *req, unsigned int issue_flags)
327 {
328 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
329 	struct io_ring_ctx *ctx = req->ctx;
330 	struct io_futex_data *ifd = NULL;
331 	struct futex_hash_bucket *hb;
332 	int ret;
333 
334 	if (!iof->futex_mask) {
335 		ret = -EINVAL;
336 		goto done;
337 	}
338 
339 	io_ring_submit_lock(ctx, issue_flags);
340 	ifd = io_alloc_ifd(ctx);
341 	if (!ifd) {
342 		ret = -ENOMEM;
343 		goto done_unlock;
344 	}
345 
346 	req->async_data = ifd;
347 	ifd->q = futex_q_init;
348 	ifd->q.bitset = iof->futex_mask;
349 	ifd->q.wake = io_futex_wake_fn;
350 	ifd->req = req;
351 
352 	ret = futex_wait_setup(iof->uaddr, iof->futex_val, iof->futex_flags,
353 			       &ifd->q, &hb);
354 	if (!ret) {
355 		hlist_add_head(&req->hash_node, &ctx->futex_list);
356 		io_ring_submit_unlock(ctx, issue_flags);
357 
358 		futex_queue(&ifd->q, hb);
359 		return IOU_ISSUE_SKIP_COMPLETE;
360 	}
361 
362 done_unlock:
363 	io_ring_submit_unlock(ctx, issue_flags);
364 done:
365 	if (ret < 0)
366 		req_set_fail(req);
367 	io_req_set_res(req, ret, 0);
368 	kfree(ifd);
369 	return IOU_OK;
370 }
371 
372 int io_futex_wake(struct io_kiocb *req, unsigned int issue_flags)
373 {
374 	struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex);
375 	int ret;
376 
377 	/*
378 	 * Strict flags - ensure that waking 0 futexes yields a 0 result.
379 	 * See commit 43adf8449510 ("futex: FLAGS_STRICT") for details.
380 	 */
381 	ret = futex_wake(iof->uaddr, FLAGS_STRICT | iof->futex_flags,
382 			 iof->futex_val, iof->futex_mask);
383 	if (ret < 0)
384 		req_set_fail(req);
385 	io_req_set_res(req, ret, 0);
386 	return IOU_OK;
387 }
388