Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
io_uring: add IORING_OP_WAITID support
This adds support for an async version of waitid(2), in a fully async
version. If an event isn't immediately available, wait for a callback
to trigger a retry.
The format of the sqe is as follows:
sqe->len The 'which', the idtype being queried/waited for.
sqe->fd The 'pid' (or id) being waited for.
sqe->file_index The 'options' being set.
sqe->addr2 A pointer to siginfo_t, if any, being filled in.
buf_index, add3, and waitid_flags are reserved/unused for now.
waitid_flags will be used for options for this request type. One
interesting use case may be to add multi-shot support, so that the
request stays armed and posts a notification every time a monitored
process state change occurs.
Note that this does not support rusage, on Arnd's recommendation.
See the waitid(2) man page for details on the arguments.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
+410
-1
+2
include/linux/io_uring_types.h
+2
include/linux/io_uring_types.h
+2
include/uapi/linux/io_uring.h
+2
include/uapi/linux/io_uring.h
···
65
65
__u32 xattr_flags;
66
66
__u32 msg_ring_flags;
67
67
__u32 uring_cmd_flags;
68
+
__u32 waitid_flags;
68
69
};
69
70
__u64 user_data; /* data to be passed back at completion time */
70
71
/* pack this to avoid bogus arm OABI complaints */
···
242
241
IORING_OP_SEND_ZC,
243
242
IORING_OP_SENDMSG_ZC,
244
243
IORING_OP_READ_MULTISHOT,
244
+
IORING_OP_WAITID,
245
245
246
246
/* this goes last, obviously */
247
247
IORING_OP_LAST,
+2
-1
io_uring/Makefile
+2
-1
io_uring/Makefile
+5
io_uring/cancel.c
+5
io_uring/cancel.c
···
15
15
#include "tctx.h"
16
16
#include "poll.h"
17
17
#include "timeout.h"
18
+
#include "waitid.h"
18
19
#include "cancel.h"
19
20
20
21
struct io_cancel {
···
117
116
return 0;
118
117
119
118
ret = io_poll_cancel(ctx, cd, issue_flags);
119
+
if (ret != -ENOENT)
120
+
return ret;
121
+
122
+
ret = io_waitid_cancel(ctx, cd, issue_flags);
120
123
if (ret != -ENOENT)
121
124
return ret;
122
125
+3
io_uring/io_uring.c
+3
io_uring/io_uring.c
···
92
92
#include "cancel.h"
93
93
#include "net.h"
94
94
#include "notif.h"
95
+
#include "waitid.h"
95
96
96
97
#include "timeout.h"
97
98
#include "poll.h"
···
349
348
INIT_LIST_HEAD(&ctx->tctx_list);
350
349
ctx->submit_state.free_list.next = NULL;
351
350
INIT_WQ_LIST(&ctx->locked_free_list);
351
+
INIT_HLIST_HEAD(&ctx->waitid_list);
352
352
INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
353
353
INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
354
354
return ctx;
···
3305
3303
ret |= io_cancel_defer_files(ctx, task, cancel_all);
3306
3304
mutex_lock(&ctx->uring_lock);
3307
3305
ret |= io_poll_remove_all(ctx, task, cancel_all);
3306
+
ret |= io_waitid_remove_all(ctx, task, cancel_all);
3308
3307
mutex_unlock(&ctx->uring_lock);
3309
3308
ret |= io_kill_timeouts(ctx, task, cancel_all);
3310
3309
if (task)
+9
io_uring/opdef.c
+9
io_uring/opdef.c
···
33
33
#include "poll.h"
34
34
#include "cancel.h"
35
35
#include "rw.h"
36
+
#include "waitid.h"
36
37
37
38
static int io_no_issue(struct io_kiocb *req, unsigned int issue_flags)
38
39
{
···
440
439
.prep = io_read_mshot_prep,
441
440
.issue = io_read_mshot,
442
441
},
442
+
[IORING_OP_WAITID] = {
443
+
.prep = io_waitid_prep,
444
+
.issue = io_waitid,
445
+
},
443
446
};
444
447
445
448
const struct io_cold_def io_cold_defs[] = {
···
665
660
},
666
661
[IORING_OP_READ_MULTISHOT] = {
667
662
.name = "READ_MULTISHOT",
663
+
},
664
+
[IORING_OP_WAITID] = {
665
+
.name = "WAITID",
666
+
.async_size = sizeof(struct io_waitid_async),
668
667
},
669
668
};
670
669
+372
io_uring/waitid.c
+372
io_uring/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
+
19
+
static void io_waitid_cb(struct io_kiocb *req, struct io_tw_state *ts);
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
+
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
+
kfree(req->async_data);
41
+
req->async_data = NULL;
42
+
req->flags &= ~REQ_F_ASYNC_DATA;
43
+
}
44
+
45
+
#ifdef CONFIG_COMPAT
46
+
static bool io_waitid_compat_copy_si(struct io_waitid *iw, int signo)
47
+
{
48
+
struct compat_siginfo __user *infop;
49
+
bool ret;
50
+
51
+
infop = (struct compat_siginfo __user *) iw->infop;
52
+
53
+
if (!user_write_access_begin(infop, sizeof(*infop)))
54
+
return false;
55
+
56
+
unsafe_put_user(signo, &infop->si_signo, Efault);
57
+
unsafe_put_user(0, &infop->si_errno, Efault);
58
+
unsafe_put_user(iw->info.cause, &infop->si_code, Efault);
59
+
unsafe_put_user(iw->info.pid, &infop->si_pid, Efault);
60
+
unsafe_put_user(iw->info.uid, &infop->si_uid, Efault);
61
+
unsafe_put_user(iw->info.status, &infop->si_status, Efault);
62
+
ret = true;
63
+
done:
64
+
user_write_access_end();
65
+
return ret;
66
+
Efault:
67
+
ret = false;
68
+
goto done;
69
+
}
70
+
#endif
71
+
72
+
static bool io_waitid_copy_si(struct io_kiocb *req, int signo)
73
+
{
74
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
75
+
bool ret;
76
+
77
+
if (!iw->infop)
78
+
return true;
79
+
80
+
#ifdef CONFIG_COMPAT
81
+
if (req->ctx->compat)
82
+
return io_waitid_compat_copy_si(iw, signo);
83
+
#endif
84
+
85
+
if (!user_write_access_begin(iw->infop, sizeof(*iw->infop)))
86
+
return false;
87
+
88
+
unsafe_put_user(signo, &iw->infop->si_signo, Efault);
89
+
unsafe_put_user(0, &iw->infop->si_errno, Efault);
90
+
unsafe_put_user(iw->info.cause, &iw->infop->si_code, Efault);
91
+
unsafe_put_user(iw->info.pid, &iw->infop->si_pid, Efault);
92
+
unsafe_put_user(iw->info.uid, &iw->infop->si_uid, Efault);
93
+
unsafe_put_user(iw->info.status, &iw->infop->si_status, Efault);
94
+
ret = true;
95
+
done:
96
+
user_write_access_end();
97
+
return ret;
98
+
Efault:
99
+
ret = false;
100
+
goto done;
101
+
}
102
+
103
+
static int io_waitid_finish(struct io_kiocb *req, int ret)
104
+
{
105
+
int signo = 0;
106
+
107
+
if (ret > 0) {
108
+
signo = SIGCHLD;
109
+
ret = 0;
110
+
}
111
+
112
+
if (!io_waitid_copy_si(req, signo))
113
+
ret = -EFAULT;
114
+
io_waitid_free(req);
115
+
return ret;
116
+
}
117
+
118
+
static void io_waitid_complete(struct io_kiocb *req, int ret)
119
+
{
120
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
121
+
struct io_tw_state ts = { .locked = true };
122
+
123
+
/* anyone completing better be holding a reference */
124
+
WARN_ON_ONCE(!(atomic_read(&iw->refs) & IO_WAITID_REF_MASK));
125
+
126
+
lockdep_assert_held(&req->ctx->uring_lock);
127
+
128
+
/*
129
+
* Did cancel find it meanwhile?
130
+
*/
131
+
if (hlist_unhashed(&req->hash_node))
132
+
return;
133
+
134
+
hlist_del_init(&req->hash_node);
135
+
136
+
ret = io_waitid_finish(req, ret);
137
+
if (ret < 0)
138
+
req_set_fail(req);
139
+
io_req_set_res(req, ret, 0);
140
+
io_req_task_complete(req, &ts);
141
+
}
142
+
143
+
static bool __io_waitid_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
144
+
{
145
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
146
+
struct io_waitid_async *iwa = req->async_data;
147
+
148
+
/*
149
+
* Mark us canceled regardless of ownership. This will prevent a
150
+
* potential retry from a spurious wakeup.
151
+
*/
152
+
atomic_or(IO_WAITID_CANCEL_FLAG, &iw->refs);
153
+
154
+
/* claim ownership */
155
+
if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
156
+
return false;
157
+
158
+
spin_lock_irq(&iw->head->lock);
159
+
list_del_init(&iwa->wo.child_wait.entry);
160
+
spin_unlock_irq(&iw->head->lock);
161
+
io_waitid_complete(req, -ECANCELED);
162
+
return true;
163
+
}
164
+
165
+
int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
166
+
unsigned int issue_flags)
167
+
{
168
+
struct hlist_node *tmp;
169
+
struct io_kiocb *req;
170
+
int nr = 0;
171
+
172
+
if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_FD_FIXED))
173
+
return -ENOENT;
174
+
175
+
io_ring_submit_lock(ctx, issue_flags);
176
+
hlist_for_each_entry_safe(req, tmp, &ctx->waitid_list, hash_node) {
177
+
if (req->cqe.user_data != cd->data &&
178
+
!(cd->flags & IORING_ASYNC_CANCEL_ANY))
179
+
continue;
180
+
if (__io_waitid_cancel(ctx, req))
181
+
nr++;
182
+
if (!(cd->flags & IORING_ASYNC_CANCEL_ALL))
183
+
break;
184
+
}
185
+
io_ring_submit_unlock(ctx, issue_flags);
186
+
187
+
if (nr)
188
+
return nr;
189
+
190
+
return -ENOENT;
191
+
}
192
+
193
+
bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct task_struct *task,
194
+
bool cancel_all)
195
+
{
196
+
struct hlist_node *tmp;
197
+
struct io_kiocb *req;
198
+
bool found = false;
199
+
200
+
lockdep_assert_held(&ctx->uring_lock);
201
+
202
+
hlist_for_each_entry_safe(req, tmp, &ctx->waitid_list, hash_node) {
203
+
if (!io_match_task_safe(req, task, cancel_all))
204
+
continue;
205
+
__io_waitid_cancel(ctx, req);
206
+
found = true;
207
+
}
208
+
209
+
return found;
210
+
}
211
+
212
+
static inline bool io_waitid_drop_issue_ref(struct io_kiocb *req)
213
+
{
214
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
215
+
struct io_waitid_async *iwa = req->async_data;
216
+
217
+
if (!atomic_sub_return(1, &iw->refs))
218
+
return false;
219
+
220
+
/*
221
+
* Wakeup triggered, racing with us. It was prevented from
222
+
* completing because of that, queue up the tw to do that.
223
+
*/
224
+
req->io_task_work.func = io_waitid_cb;
225
+
io_req_task_work_add(req);
226
+
remove_wait_queue(iw->head, &iwa->wo.child_wait);
227
+
return true;
228
+
}
229
+
230
+
static void io_waitid_cb(struct io_kiocb *req, struct io_tw_state *ts)
231
+
{
232
+
struct io_waitid_async *iwa = req->async_data;
233
+
struct io_ring_ctx *ctx = req->ctx;
234
+
int ret;
235
+
236
+
io_tw_lock(ctx, ts);
237
+
238
+
ret = __do_wait(&iwa->wo);
239
+
240
+
/*
241
+
* If we get -ERESTARTSYS here, we need to re-arm and check again
242
+
* to ensure we get another callback. If the retry works, then we can
243
+
* just remove ourselves from the waitqueue again and finish the
244
+
* request.
245
+
*/
246
+
if (unlikely(ret == -ERESTARTSYS)) {
247
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
248
+
249
+
/* Don't retry if cancel found it meanwhile */
250
+
ret = -ECANCELED;
251
+
if (!(atomic_read(&iw->refs) & IO_WAITID_CANCEL_FLAG)) {
252
+
iw->head = ¤t->signal->wait_chldexit;
253
+
add_wait_queue(iw->head, &iwa->wo.child_wait);
254
+
ret = __do_wait(&iwa->wo);
255
+
if (ret == -ERESTARTSYS) {
256
+
/* retry armed, drop our ref */
257
+
io_waitid_drop_issue_ref(req);
258
+
return;
259
+
}
260
+
261
+
remove_wait_queue(iw->head, &iwa->wo.child_wait);
262
+
}
263
+
}
264
+
265
+
io_waitid_complete(req, ret);
266
+
}
267
+
268
+
static int io_waitid_wait(struct wait_queue_entry *wait, unsigned mode,
269
+
int sync, void *key)
270
+
{
271
+
struct wait_opts *wo = container_of(wait, struct wait_opts, child_wait);
272
+
struct io_waitid_async *iwa = container_of(wo, struct io_waitid_async, wo);
273
+
struct io_kiocb *req = iwa->req;
274
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
275
+
struct task_struct *p = key;
276
+
277
+
if (!pid_child_should_wake(wo, p))
278
+
return 0;
279
+
280
+
/* cancel is in progress */
281
+
if (atomic_fetch_inc(&iw->refs) & IO_WAITID_REF_MASK)
282
+
return 1;
283
+
284
+
req->io_task_work.func = io_waitid_cb;
285
+
io_req_task_work_add(req);
286
+
list_del_init(&wait->entry);
287
+
return 1;
288
+
}
289
+
290
+
int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
291
+
{
292
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
293
+
294
+
if (sqe->addr || sqe->buf_index || sqe->addr3 || sqe->waitid_flags)
295
+
return -EINVAL;
296
+
297
+
iw->which = READ_ONCE(sqe->len);
298
+
iw->upid = READ_ONCE(sqe->fd);
299
+
iw->options = READ_ONCE(sqe->file_index);
300
+
iw->infop = u64_to_user_ptr(READ_ONCE(sqe->addr2));
301
+
return 0;
302
+
}
303
+
304
+
int io_waitid(struct io_kiocb *req, unsigned int issue_flags)
305
+
{
306
+
struct io_waitid *iw = io_kiocb_to_cmd(req, struct io_waitid);
307
+
struct io_ring_ctx *ctx = req->ctx;
308
+
struct io_waitid_async *iwa;
309
+
int ret;
310
+
311
+
if (io_alloc_async_data(req))
312
+
return -ENOMEM;
313
+
314
+
iwa = req->async_data;
315
+
iwa->req = req;
316
+
317
+
ret = kernel_waitid_prepare(&iwa->wo, iw->which, iw->upid, &iw->info,
318
+
iw->options, NULL);
319
+
if (ret)
320
+
goto done;
321
+
322
+
/*
323
+
* Mark the request as busy upfront, in case we're racing with the
324
+
* wakeup. If we are, then we'll notice when we drop this initial
325
+
* reference again after arming.
326
+
*/
327
+
atomic_set(&iw->refs, 1);
328
+
329
+
/*
330
+
* Cancel must hold the ctx lock, so there's no risk of cancelation
331
+
* finding us until a) we remain on the list, and b) the lock is
332
+
* dropped. We only need to worry about racing with the wakeup
333
+
* callback.
334
+
*/
335
+
io_ring_submit_lock(ctx, issue_flags);
336
+
hlist_add_head(&req->hash_node, &ctx->waitid_list);
337
+
338
+
init_waitqueue_func_entry(&iwa->wo.child_wait, io_waitid_wait);
339
+
iwa->wo.child_wait.private = req->task;
340
+
iw->head = ¤t->signal->wait_chldexit;
341
+
add_wait_queue(iw->head, &iwa->wo.child_wait);
342
+
343
+
ret = __do_wait(&iwa->wo);
344
+
if (ret == -ERESTARTSYS) {
345
+
/*
346
+
* Nobody else grabbed a reference, it'll complete when we get
347
+
* a waitqueue callback, or if someone cancels it.
348
+
*/
349
+
if (!io_waitid_drop_issue_ref(req)) {
350
+
io_ring_submit_unlock(ctx, issue_flags);
351
+
return IOU_ISSUE_SKIP_COMPLETE;
352
+
}
353
+
354
+
/*
355
+
* Wakeup triggered, racing with us. It was prevented from
356
+
* completing because of that, queue up the tw to do that.
357
+
*/
358
+
io_ring_submit_unlock(ctx, issue_flags);
359
+
return IOU_ISSUE_SKIP_COMPLETE;
360
+
}
361
+
362
+
hlist_del_init(&req->hash_node);
363
+
remove_wait_queue(iw->head, &iwa->wo.child_wait);
364
+
ret = io_waitid_finish(req, ret);
365
+
366
+
io_ring_submit_unlock(ctx, issue_flags);
367
+
done:
368
+
if (ret < 0)
369
+
req_set_fail(req);
370
+
io_req_set_res(req, ret, 0);
371
+
return IOU_OK;
372
+
}
+15
io_uring/waitid.h
+15
io_uring/waitid.h
···
1
+
// SPDX-License-Identifier: GPL-2.0
2
+
3
+
#include "../kernel/exit.h"
4
+
5
+
struct io_waitid_async {
6
+
struct io_kiocb *req;
7
+
struct wait_opts wo;
8
+
};
9
+
10
+
int io_waitid_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
11
+
int io_waitid(struct io_kiocb *req, unsigned int issue_flags);
12
+
int io_waitid_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
13
+
unsigned int issue_flags);
14
+
bool io_waitid_remove_all(struct io_ring_ctx *ctx, struct task_struct *task,
15
+
bool cancel_all);