Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
vdso/timens: Move functions to new file
As a preparation of the untangling of time namespaces and the vDSO, move
the glue functions between those subsystems into a new file.
While at it, switch the mutex lock and mmap_read_lock() in the vDSO
namespace code to guard().
Signed-off-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Link: https://patch.msgid.link/20260326-vdso-timens-decoupling-v2-1-c82693a7775f@linutronix.de
authored by
Thomas Weißschuh
and committed by
Thomas Gleixner
5dc9cf83
bed0053a
+166
-154
+2
MAINTAINERS
+2
MAINTAINERS
···
10768
10768
T: git git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git timers/vdso
10769
10769
F: include/asm-generic/vdso/vsyscall.h
10770
10770
F: include/vdso/
10771
+
F: kernel/time/namespace_vdso.c
10771
10772
F: kernel/time/vsyscall.c
10772
10773
F: lib/vdso/
10773
10774
F: tools/testing/selftests/vDSO/
···
21001
21000
F: kernel/time/itimer.c
21002
21001
F: kernel/time/posix-*
21003
21002
F: kernel/time/namespace.c
21003
+
F: kernel/time/namespace_vdso.c
21004
21004
21005
21005
POWER MANAGEMENT CORE
21006
21006
M: "Rafael J. Wysocki" <rafael@kernel.org>
-8
include/linux/time_namespace.h
-8
include/linux/time_namespace.h
···
38
38
return container_of(ns, struct time_namespace, ns);
39
39
}
40
40
void __init time_ns_init(void);
41
-
extern int vdso_join_timens(struct task_struct *task,
42
-
struct time_namespace *ns);
43
41
extern void timens_commit(struct task_struct *tsk, struct time_namespace *ns);
44
42
45
43
static inline struct time_namespace *get_time_ns(struct time_namespace *ns)
···
113
115
#else
114
116
static inline void __init time_ns_init(void)
115
117
{
116
-
}
117
-
118
-
static inline int vdso_join_timens(struct task_struct *task,
119
-
struct time_namespace *ns)
120
-
{
121
-
return 0;
122
118
}
123
119
124
120
static inline void timens_commit(struct task_struct *tsk,
+1
-1
kernel/time/Makefile
+1
-1
kernel/time/Makefile
···
29
29
obj-$(CONFIG_GENERIC_GETTIMEOFDAY) += vsyscall.o
30
30
obj-$(CONFIG_DEBUG_FS) += timekeeping_debug.o
31
31
obj-$(CONFIG_TEST_UDELAY) += test_udelay.o
32
-
obj-$(CONFIG_TIME_NS) += namespace.o
32
+
obj-$(CONFIG_TIME_NS) += namespace.o namespace_vdso.o
33
33
obj-$(CONFIG_TEST_CLOCKSOURCE_WATCHDOG) += clocksource-wdtest.o
34
34
obj-$(CONFIG_TIME_KUNIT_TEST) += time_test.o
+4
-120
kernel/time/namespace.c
+4
-120
kernel/time/namespace.c
···
19
19
#include <linux/err.h>
20
20
#include <linux/mm.h>
21
21
22
-
#include <vdso/datapage.h>
22
+
#include "namespace_internal.h"
23
23
24
24
ktime_t do_timens_ktime_to_host(clockid_t clockid, ktime_t tim,
25
25
struct timens_offsets *ns_offsets)
···
138
138
return clone_time_ns(user_ns, old_ns);
139
139
}
140
140
141
-
static struct timens_offset offset_from_ts(struct timespec64 off)
142
-
{
143
-
struct timens_offset ret;
144
-
145
-
ret.sec = off.tv_sec;
146
-
ret.nsec = off.tv_nsec;
147
-
148
-
return ret;
149
-
}
150
-
151
-
/*
152
-
* A time namespace VVAR page has the same layout as the VVAR page which
153
-
* contains the system wide VDSO data.
154
-
*
155
-
* For a normal task the VVAR pages are installed in the normal ordering:
156
-
* VVAR
157
-
* PVCLOCK
158
-
* HVCLOCK
159
-
* TIMENS <- Not really required
160
-
*
161
-
* Now for a timens task the pages are installed in the following order:
162
-
* TIMENS
163
-
* PVCLOCK
164
-
* HVCLOCK
165
-
* VVAR
166
-
*
167
-
* The check for vdso_clock->clock_mode is in the unlikely path of
168
-
* the seq begin magic. So for the non-timens case most of the time
169
-
* 'seq' is even, so the branch is not taken.
170
-
*
171
-
* If 'seq' is odd, i.e. a concurrent update is in progress, the extra check
172
-
* for vdso_clock->clock_mode is a non-issue. The task is spin waiting for the
173
-
* update to finish and for 'seq' to become even anyway.
174
-
*
175
-
* Timens page has vdso_clock->clock_mode set to VDSO_CLOCKMODE_TIMENS which
176
-
* enforces the time namespace handling path.
177
-
*/
178
-
static void timens_setup_vdso_clock_data(struct vdso_clock *vc,
179
-
struct time_namespace *ns)
180
-
{
181
-
struct timens_offset *offset = vc->offset;
182
-
struct timens_offset monotonic = offset_from_ts(ns->offsets.monotonic);
183
-
struct timens_offset boottime = offset_from_ts(ns->offsets.boottime);
184
-
185
-
vc->seq = 1;
186
-
vc->clock_mode = VDSO_CLOCKMODE_TIMENS;
187
-
offset[CLOCK_MONOTONIC] = monotonic;
188
-
offset[CLOCK_MONOTONIC_RAW] = monotonic;
189
-
offset[CLOCK_MONOTONIC_COARSE] = monotonic;
190
-
offset[CLOCK_BOOTTIME] = boottime;
191
-
offset[CLOCK_BOOTTIME_ALARM] = boottime;
192
-
}
193
-
194
-
struct page *find_timens_vvar_page(struct vm_area_struct *vma)
195
-
{
196
-
if (likely(vma->vm_mm == current->mm))
197
-
return current->nsproxy->time_ns->vvar_page;
198
-
199
-
/*
200
-
* VM_PFNMAP | VM_IO protect .fault() handler from being called
201
-
* through interfaces like /proc/$pid/mem or
202
-
* process_vm_{readv,writev}() as long as there's no .access()
203
-
* in special_mapping_vmops().
204
-
* For more details check_vma_flags() and __access_remote_vm()
205
-
*/
206
-
207
-
WARN(1, "vvar_page accessed remotely");
208
-
209
-
return NULL;
210
-
}
211
-
212
-
/*
213
-
* Protects possibly multiple offsets writers racing each other
214
-
* and tasks entering the namespace.
215
-
*/
216
-
static DEFINE_MUTEX(offset_lock);
217
-
218
-
static void timens_set_vvar_page(struct task_struct *task,
219
-
struct time_namespace *ns)
220
-
{
221
-
struct vdso_time_data *vdata;
222
-
struct vdso_clock *vc;
223
-
unsigned int i;
224
-
225
-
if (ns == &init_time_ns)
226
-
return;
227
-
228
-
/* Fast-path, taken by every task in namespace except the first. */
229
-
if (likely(ns->frozen_offsets))
230
-
return;
231
-
232
-
mutex_lock(&offset_lock);
233
-
/* Nothing to-do: vvar_page has been already initialized. */
234
-
if (ns->frozen_offsets)
235
-
goto out;
236
-
237
-
ns->frozen_offsets = true;
238
-
vdata = page_address(ns->vvar_page);
239
-
vc = vdata->clock_data;
240
-
241
-
for (i = 0; i < CS_BASES; i++)
242
-
timens_setup_vdso_clock_data(&vc[i], ns);
243
-
244
-
if (IS_ENABLED(CONFIG_POSIX_AUX_CLOCKS)) {
245
-
for (i = 0; i < ARRAY_SIZE(vdata->aux_clock_data); i++)
246
-
timens_setup_vdso_clock_data(&vdata->aux_clock_data[i], ns);
247
-
}
248
-
249
-
out:
250
-
mutex_unlock(&offset_lock);
251
-
}
141
+
DEFINE_MUTEX(timens_offset_lock);
252
142
253
143
void free_time_ns(struct time_namespace *ns)
254
144
{
···
186
296
static void timens_put(struct ns_common *ns)
187
297
{
188
298
put_time_ns(to_time_ns(ns));
189
-
}
190
-
191
-
void timens_commit(struct task_struct *tsk, struct time_namespace *ns)
192
-
{
193
-
timens_set_vvar_page(tsk, ns);
194
-
vdso_join_timens(tsk, ns);
195
299
}
196
300
197
301
static int timens_install(struct nsset *nsset, struct ns_common *new)
···
312
428
goto out;
313
429
}
314
430
315
-
mutex_lock(&offset_lock);
431
+
mutex_lock(&timens_offset_lock);
316
432
if (time_ns->frozen_offsets) {
317
433
err = -EACCES;
318
434
goto out_unlock;
···
337
453
}
338
454
339
455
out_unlock:
340
-
mutex_unlock(&offset_lock);
456
+
mutex_unlock(&timens_offset_lock);
341
457
out:
342
458
put_time_ns(time_ns);
343
459
+13
kernel/time/namespace_internal.h
+13
kernel/time/namespace_internal.h
···
1
+
/* SPDX-License-Identifier: GPL-2.0 */
2
+
#ifndef _TIME_NAMESPACE_INTERNAL_H
3
+
#define _TIME_NAMESPACE_INTERNAL_H
4
+
5
+
#include <linux/mutex.h>
6
+
7
+
/*
8
+
* Protects possibly multiple offsets writers racing each other
9
+
* and tasks entering the namespace.
10
+
*/
11
+
extern struct mutex timens_offset_lock;
12
+
13
+
#endif /* _TIME_NAMESPACE_INTERNAL_H */
+146
kernel/time/namespace_vdso.c
+146
kernel/time/namespace_vdso.c
···
1
+
// SPDX-License-Identifier: GPL-2.0
2
+
/*
3
+
* Author: Andrei Vagin <avagin@openvz.org>
4
+
* Author: Dmitry Safonov <dima@arista.com>
5
+
*/
6
+
7
+
#include <linux/cleanup.h>
8
+
#include <linux/mm.h>
9
+
#include <linux/time_namespace.h>
10
+
#include <linux/time.h>
11
+
#include <linux/vdso_datastore.h>
12
+
13
+
#include <vdso/clocksource.h>
14
+
#include <vdso/datapage.h>
15
+
16
+
#include "namespace_internal.h"
17
+
18
+
static struct timens_offset offset_from_ts(struct timespec64 off)
19
+
{
20
+
struct timens_offset ret;
21
+
22
+
ret.sec = off.tv_sec;
23
+
ret.nsec = off.tv_nsec;
24
+
25
+
return ret;
26
+
}
27
+
28
+
/*
29
+
* A time namespace VVAR page has the same layout as the VVAR page which
30
+
* contains the system wide VDSO data.
31
+
*
32
+
* For a normal task the VVAR pages are installed in the normal ordering:
33
+
* VVAR
34
+
* PVCLOCK
35
+
* HVCLOCK
36
+
* TIMENS <- Not really required
37
+
*
38
+
* Now for a timens task the pages are installed in the following order:
39
+
* TIMENS
40
+
* PVCLOCK
41
+
* HVCLOCK
42
+
* VVAR
43
+
*
44
+
* The check for vdso_clock->clock_mode is in the unlikely path of
45
+
* the seq begin magic. So for the non-timens case most of the time
46
+
* 'seq' is even, so the branch is not taken.
47
+
*
48
+
* If 'seq' is odd, i.e. a concurrent update is in progress, the extra check
49
+
* for vdso_clock->clock_mode is a non-issue. The task is spin waiting for the
50
+
* update to finish and for 'seq' to become even anyway.
51
+
*
52
+
* Timens page has vdso_clock->clock_mode set to VDSO_CLOCKMODE_TIMENS which
53
+
* enforces the time namespace handling path.
54
+
*/
55
+
static void timens_setup_vdso_clock_data(struct vdso_clock *vc,
56
+
struct time_namespace *ns)
57
+
{
58
+
struct timens_offset *offset = vc->offset;
59
+
struct timens_offset monotonic = offset_from_ts(ns->offsets.monotonic);
60
+
struct timens_offset boottime = offset_from_ts(ns->offsets.boottime);
61
+
62
+
vc->seq = 1;
63
+
vc->clock_mode = VDSO_CLOCKMODE_TIMENS;
64
+
offset[CLOCK_MONOTONIC] = monotonic;
65
+
offset[CLOCK_MONOTONIC_RAW] = monotonic;
66
+
offset[CLOCK_MONOTONIC_COARSE] = monotonic;
67
+
offset[CLOCK_BOOTTIME] = boottime;
68
+
offset[CLOCK_BOOTTIME_ALARM] = boottime;
69
+
}
70
+
71
+
struct page *find_timens_vvar_page(struct vm_area_struct *vma)
72
+
{
73
+
if (likely(vma->vm_mm == current->mm))
74
+
return current->nsproxy->time_ns->vvar_page;
75
+
76
+
/*
77
+
* VM_PFNMAP | VM_IO protect .fault() handler from being called
78
+
* through interfaces like /proc/$pid/mem or
79
+
* process_vm_{readv,writev}() as long as there's no .access()
80
+
* in special_mapping_vmops().
81
+
* For more details check_vma_flags() and __access_remote_vm()
82
+
*/
83
+
84
+
WARN(1, "vvar_page accessed remotely");
85
+
86
+
return NULL;
87
+
}
88
+
89
+
static void timens_set_vvar_page(struct task_struct *task,
90
+
struct time_namespace *ns)
91
+
{
92
+
struct vdso_time_data *vdata;
93
+
struct vdso_clock *vc;
94
+
unsigned int i;
95
+
96
+
if (ns == &init_time_ns)
97
+
return;
98
+
99
+
/* Fast-path, taken by every task in namespace except the first. */
100
+
if (likely(ns->frozen_offsets))
101
+
return;
102
+
103
+
guard(mutex)(&timens_offset_lock);
104
+
/* Nothing to-do: vvar_page has been already initialized. */
105
+
if (ns->frozen_offsets)
106
+
return;
107
+
108
+
ns->frozen_offsets = true;
109
+
vdata = page_address(ns->vvar_page);
110
+
vc = vdata->clock_data;
111
+
112
+
for (i = 0; i < CS_BASES; i++)
113
+
timens_setup_vdso_clock_data(&vc[i], ns);
114
+
115
+
if (IS_ENABLED(CONFIG_POSIX_AUX_CLOCKS)) {
116
+
for (i = 0; i < ARRAY_SIZE(vdata->aux_clock_data); i++)
117
+
timens_setup_vdso_clock_data(&vdata->aux_clock_data[i], ns);
118
+
}
119
+
}
120
+
121
+
/*
122
+
* The vvar page layout depends on whether a task belongs to the root or
123
+
* non-root time namespace. Whenever a task changes its namespace, the VVAR
124
+
* page tables are cleared and then they will be re-faulted with a
125
+
* corresponding layout.
126
+
* See also the comment near timens_setup_vdso_clock_data() for details.
127
+
*/
128
+
static int vdso_join_timens(struct task_struct *task, struct time_namespace *ns)
129
+
{
130
+
struct mm_struct *mm = task->mm;
131
+
struct vm_area_struct *vma;
132
+
VMA_ITERATOR(vmi, mm, 0);
133
+
134
+
guard(mmap_read_lock)(mm);
135
+
for_each_vma(vmi, vma) {
136
+
if (vma_is_special_mapping(vma, &vdso_vvar_mapping))
137
+
zap_vma_pages(vma);
138
+
}
139
+
return 0;
140
+
}
141
+
142
+
void timens_commit(struct task_struct *tsk, struct time_namespace *ns)
143
+
{
144
+
timens_set_vvar_page(tsk, ns);
145
+
vdso_join_timens(tsk, ns);
146
+
}
-25
lib/vdso/datastore.c
-25
lib/vdso/datastore.c
···
132
132
VM_MIXEDMAP | VM_SEALED_SYSMAP,
133
133
&vdso_vvar_mapping);
134
134
}
135
-
136
-
#ifdef CONFIG_TIME_NS
137
-
/*
138
-
* The vvar page layout depends on whether a task belongs to the root or
139
-
* non-root time namespace. Whenever a task changes its namespace, the VVAR
140
-
* page tables are cleared and then they will be re-faulted with a
141
-
* corresponding layout.
142
-
* See also the comment near timens_setup_vdso_clock_data() for details.
143
-
*/
144
-
int vdso_join_timens(struct task_struct *task, struct time_namespace *ns)
145
-
{
146
-
struct mm_struct *mm = task->mm;
147
-
struct vm_area_struct *vma;
148
-
VMA_ITERATOR(vmi, mm, 0);
149
-
150
-
mmap_read_lock(mm);
151
-
for_each_vma(vmi, vma) {
152
-
if (vma_is_special_mapping(vma, &vdso_vvar_mapping))
153
-
zap_vma_pages(vma);
154
-
}
155
-
mmap_read_unlock(mm);
156
-
157
-
return 0;
158
-
}
159
-
#endif