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