Loading...
1//SPDX-License-Identifier: GPL-2.0
2#include <linux/cgroup.h>
3#include <linux/sched.h>
4#include <linux/sched/task.h>
5#include <linux/sched/signal.h>
6
7#include "cgroup-internal.h"
8
9#include <trace/events/cgroup.h>
10
11/*
12 * Propagate the cgroup frozen state upwards by the cgroup tree.
13 */
14static void cgroup_propagate_frozen(struct cgroup *cgrp, bool frozen)
15{
16 int desc = 1;
17
18 /*
19 * If the new state is frozen, some freezing ancestor cgroups may change
20 * their state too, depending on if all their descendants are frozen.
21 *
22 * Otherwise, all ancestor cgroups are forced into the non-frozen state.
23 */
24 while ((cgrp = cgroup_parent(cgrp))) {
25 if (frozen) {
26 cgrp->freezer.nr_frozen_descendants += desc;
27 if (!test_bit(CGRP_FROZEN, &cgrp->flags) &&
28 test_bit(CGRP_FREEZE, &cgrp->flags) &&
29 cgrp->freezer.nr_frozen_descendants ==
30 cgrp->nr_descendants) {
31 set_bit(CGRP_FROZEN, &cgrp->flags);
32 cgroup_file_notify(&cgrp->events_file);
33 TRACE_CGROUP_PATH(notify_frozen, cgrp, 1);
34 desc++;
35 }
36 } else {
37 cgrp->freezer.nr_frozen_descendants -= desc;
38 if (test_bit(CGRP_FROZEN, &cgrp->flags)) {
39 clear_bit(CGRP_FROZEN, &cgrp->flags);
40 cgroup_file_notify(&cgrp->events_file);
41 TRACE_CGROUP_PATH(notify_frozen, cgrp, 0);
42 desc++;
43 }
44 }
45 }
46}
47
48/*
49 * Revisit the cgroup frozen state.
50 * Checks if the cgroup is really frozen and perform all state transitions.
51 */
52void cgroup_update_frozen(struct cgroup *cgrp)
53{
54 bool frozen;
55
56 lockdep_assert_held(&css_set_lock);
57
58 /*
59 * If the cgroup has to be frozen (CGRP_FREEZE bit set),
60 * and all tasks are frozen and/or stopped, let's consider
61 * the cgroup frozen. Otherwise it's not frozen.
62 */
63 frozen = test_bit(CGRP_FREEZE, &cgrp->flags) &&
64 cgrp->freezer.nr_frozen_tasks == __cgroup_task_count(cgrp);
65
66 if (frozen) {
67 /* Already there? */
68 if (test_bit(CGRP_FROZEN, &cgrp->flags))
69 return;
70
71 set_bit(CGRP_FROZEN, &cgrp->flags);
72 } else {
73 /* Already there? */
74 if (!test_bit(CGRP_FROZEN, &cgrp->flags))
75 return;
76
77 clear_bit(CGRP_FROZEN, &cgrp->flags);
78 }
79 cgroup_file_notify(&cgrp->events_file);
80 TRACE_CGROUP_PATH(notify_frozen, cgrp, frozen);
81
82 /* Update the state of ancestor cgroups. */
83 cgroup_propagate_frozen(cgrp, frozen);
84}
85
86/*
87 * Increment cgroup's nr_frozen_tasks.
88 */
89static void cgroup_inc_frozen_cnt(struct cgroup *cgrp)
90{
91 cgrp->freezer.nr_frozen_tasks++;
92}
93
94/*
95 * Decrement cgroup's nr_frozen_tasks.
96 */
97static void cgroup_dec_frozen_cnt(struct cgroup *cgrp)
98{
99 cgrp->freezer.nr_frozen_tasks--;
100 WARN_ON_ONCE(cgrp->freezer.nr_frozen_tasks < 0);
101}
102
103/*
104 * Enter frozen/stopped state, if not yet there. Update cgroup's counters,
105 * and revisit the state of the cgroup, if necessary.
106 */
107void cgroup_enter_frozen(void)
108{
109 struct cgroup *cgrp;
110
111 if (current->frozen)
112 return;
113
114 spin_lock_irq(&css_set_lock);
115 current->frozen = true;
116 cgrp = task_dfl_cgroup(current);
117 cgroup_inc_frozen_cnt(cgrp);
118 cgroup_update_frozen(cgrp);
119 spin_unlock_irq(&css_set_lock);
120}
121
122/*
123 * Conditionally leave frozen/stopped state. Update cgroup's counters,
124 * and revisit the state of the cgroup, if necessary.
125 *
126 * If always_leave is not set, and the cgroup is freezing,
127 * we're racing with the cgroup freezing. In this case, we don't
128 * drop the frozen counter to avoid a transient switch to
129 * the unfrozen state.
130 */
131void cgroup_leave_frozen(bool always_leave)
132{
133 struct cgroup *cgrp;
134
135 spin_lock_irq(&css_set_lock);
136 cgrp = task_dfl_cgroup(current);
137 if (always_leave || !test_bit(CGRP_FREEZE, &cgrp->flags)) {
138 cgroup_dec_frozen_cnt(cgrp);
139 cgroup_update_frozen(cgrp);
140 WARN_ON_ONCE(!current->frozen);
141 current->frozen = false;
142 } else if (!(current->jobctl & JOBCTL_TRAP_FREEZE)) {
143 spin_lock(¤t->sighand->siglock);
144 current->jobctl |= JOBCTL_TRAP_FREEZE;
145 set_thread_flag(TIF_SIGPENDING);
146 spin_unlock(¤t->sighand->siglock);
147 }
148 spin_unlock_irq(&css_set_lock);
149}
150
151/*
152 * Freeze or unfreeze the task by setting or clearing the JOBCTL_TRAP_FREEZE
153 * jobctl bit.
154 */
155static void cgroup_freeze_task(struct task_struct *task, bool freeze)
156{
157 unsigned long flags;
158
159 /* If the task is about to die, don't bother with freezing it. */
160 if (!lock_task_sighand(task, &flags))
161 return;
162
163 if (freeze) {
164 task->jobctl |= JOBCTL_TRAP_FREEZE;
165 signal_wake_up(task, false);
166 } else {
167 task->jobctl &= ~JOBCTL_TRAP_FREEZE;
168 wake_up_process(task);
169 }
170
171 unlock_task_sighand(task, &flags);
172}
173
174/*
175 * Freeze or unfreeze all tasks in the given cgroup.
176 */
177static void cgroup_do_freeze(struct cgroup *cgrp, bool freeze)
178{
179 struct css_task_iter it;
180 struct task_struct *task;
181
182 lockdep_assert_held(&cgroup_mutex);
183
184 spin_lock_irq(&css_set_lock);
185 if (freeze)
186 set_bit(CGRP_FREEZE, &cgrp->flags);
187 else
188 clear_bit(CGRP_FREEZE, &cgrp->flags);
189 spin_unlock_irq(&css_set_lock);
190
191 if (freeze)
192 TRACE_CGROUP_PATH(freeze, cgrp);
193 else
194 TRACE_CGROUP_PATH(unfreeze, cgrp);
195
196 css_task_iter_start(&cgrp->self, 0, &it);
197 while ((task = css_task_iter_next(&it))) {
198 /*
199 * Ignore kernel threads here. Freezing cgroups containing
200 * kthreads isn't supported.
201 */
202 if (task->flags & PF_KTHREAD)
203 continue;
204 cgroup_freeze_task(task, freeze);
205 }
206 css_task_iter_end(&it);
207
208 /*
209 * Cgroup state should be revisited here to cover empty leaf cgroups
210 * and cgroups which descendants are already in the desired state.
211 */
212 spin_lock_irq(&css_set_lock);
213 if (cgrp->nr_descendants == cgrp->freezer.nr_frozen_descendants)
214 cgroup_update_frozen(cgrp);
215 spin_unlock_irq(&css_set_lock);
216}
217
218/*
219 * Adjust the task state (freeze or unfreeze) and revisit the state of
220 * source and destination cgroups.
221 */
222void cgroup_freezer_migrate_task(struct task_struct *task,
223 struct cgroup *src, struct cgroup *dst)
224{
225 lockdep_assert_held(&css_set_lock);
226
227 /*
228 * Kernel threads are not supposed to be frozen at all.
229 */
230 if (task->flags & PF_KTHREAD)
231 return;
232
233 /*
234 * Adjust counters of freezing and frozen tasks.
235 * Note, that if the task is frozen, but the destination cgroup is not
236 * frozen, we bump both counters to keep them balanced.
237 */
238 if (task->frozen) {
239 cgroup_inc_frozen_cnt(dst);
240 cgroup_dec_frozen_cnt(src);
241 }
242 cgroup_update_frozen(dst);
243 cgroup_update_frozen(src);
244
245 /*
246 * Force the task to the desired state.
247 */
248 cgroup_freeze_task(task, test_bit(CGRP_FREEZE, &dst->flags));
249}
250
251void cgroup_freeze(struct cgroup *cgrp, bool freeze)
252{
253 struct cgroup_subsys_state *css;
254 struct cgroup *dsct;
255 bool applied = false;
256
257 lockdep_assert_held(&cgroup_mutex);
258
259 /*
260 * Nothing changed? Just exit.
261 */
262 if (cgrp->freezer.freeze == freeze)
263 return;
264
265 cgrp->freezer.freeze = freeze;
266
267 /*
268 * Propagate changes downwards the cgroup tree.
269 */
270 css_for_each_descendant_pre(css, &cgrp->self) {
271 dsct = css->cgroup;
272
273 if (cgroup_is_dead(dsct))
274 continue;
275
276 if (freeze) {
277 dsct->freezer.e_freeze++;
278 /*
279 * Already frozen because of ancestor's settings?
280 */
281 if (dsct->freezer.e_freeze > 1)
282 continue;
283 } else {
284 dsct->freezer.e_freeze--;
285 /*
286 * Still frozen because of ancestor's settings?
287 */
288 if (dsct->freezer.e_freeze > 0)
289 continue;
290
291 WARN_ON_ONCE(dsct->freezer.e_freeze < 0);
292 }
293
294 /*
295 * Do change actual state: freeze or unfreeze.
296 */
297 cgroup_do_freeze(dsct, freeze);
298 applied = true;
299 }
300
301 /*
302 * Even if the actual state hasn't changed, let's notify a user.
303 * The state can be enforced by an ancestor cgroup: the cgroup
304 * can already be in the desired state or it can be locked in the
305 * opposite state, so that the transition will never happen.
306 * In both cases it's better to notify a user, that there is
307 * nothing to wait for.
308 */
309 if (!applied) {
310 TRACE_CGROUP_PATH(notify_frozen, cgrp,
311 test_bit(CGRP_FROZEN, &cgrp->flags));
312 cgroup_file_notify(&cgrp->events_file);
313 }
314}
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/cgroup.h>
3#include <linux/sched.h>
4#include <linux/sched/task.h>
5#include <linux/sched/signal.h>
6
7#include "cgroup-internal.h"
8
9#include <trace/events/cgroup.h>
10
11/*
12 * Update CGRP_FROZEN of cgroup.flag
13 * Return true if flags is updated; false if flags has no change
14 */
15static bool cgroup_update_frozen_flag(struct cgroup *cgrp, bool frozen)
16{
17 lockdep_assert_held(&css_set_lock);
18
19 /* Already there? */
20 if (test_bit(CGRP_FROZEN, &cgrp->flags) == frozen)
21 return false;
22
23 if (frozen)
24 set_bit(CGRP_FROZEN, &cgrp->flags);
25 else
26 clear_bit(CGRP_FROZEN, &cgrp->flags);
27
28 cgroup_file_notify(&cgrp->events_file);
29 TRACE_CGROUP_PATH(notify_frozen, cgrp, frozen);
30 return true;
31}
32
33/*
34 * Propagate the cgroup frozen state upwards by the cgroup tree.
35 */
36static void cgroup_propagate_frozen(struct cgroup *cgrp, bool frozen)
37{
38 int desc = 1;
39
40 /*
41 * If the new state is frozen, some freezing ancestor cgroups may change
42 * their state too, depending on if all their descendants are frozen.
43 *
44 * Otherwise, all ancestor cgroups are forced into the non-frozen state.
45 */
46 while ((cgrp = cgroup_parent(cgrp))) {
47 if (frozen) {
48 cgrp->freezer.nr_frozen_descendants += desc;
49 if (!test_bit(CGRP_FREEZE, &cgrp->flags) ||
50 (cgrp->freezer.nr_frozen_descendants !=
51 cgrp->nr_descendants))
52 continue;
53 } else {
54 cgrp->freezer.nr_frozen_descendants -= desc;
55 }
56
57 if (cgroup_update_frozen_flag(cgrp, frozen))
58 desc++;
59 }
60}
61
62/*
63 * Revisit the cgroup frozen state.
64 * Checks if the cgroup is really frozen and perform all state transitions.
65 */
66void cgroup_update_frozen(struct cgroup *cgrp)
67{
68 bool frozen;
69
70 /*
71 * If the cgroup has to be frozen (CGRP_FREEZE bit set),
72 * and all tasks are frozen and/or stopped, let's consider
73 * the cgroup frozen. Otherwise it's not frozen.
74 */
75 frozen = test_bit(CGRP_FREEZE, &cgrp->flags) &&
76 cgrp->freezer.nr_frozen_tasks == __cgroup_task_count(cgrp);
77
78 /* If flags is updated, update the state of ancestor cgroups. */
79 if (cgroup_update_frozen_flag(cgrp, frozen))
80 cgroup_propagate_frozen(cgrp, frozen);
81}
82
83/*
84 * Increment cgroup's nr_frozen_tasks.
85 */
86static void cgroup_inc_frozen_cnt(struct cgroup *cgrp)
87{
88 cgrp->freezer.nr_frozen_tasks++;
89}
90
91/*
92 * Decrement cgroup's nr_frozen_tasks.
93 */
94static void cgroup_dec_frozen_cnt(struct cgroup *cgrp)
95{
96 cgrp->freezer.nr_frozen_tasks--;
97 WARN_ON_ONCE(cgrp->freezer.nr_frozen_tasks < 0);
98}
99
100/*
101 * Enter frozen/stopped state, if not yet there. Update cgroup's counters,
102 * and revisit the state of the cgroup, if necessary.
103 */
104void cgroup_enter_frozen(void)
105{
106 struct cgroup *cgrp;
107
108 if (current->frozen)
109 return;
110
111 spin_lock_irq(&css_set_lock);
112 current->frozen = true;
113 cgrp = task_dfl_cgroup(current);
114 cgroup_inc_frozen_cnt(cgrp);
115 cgroup_update_frozen(cgrp);
116 spin_unlock_irq(&css_set_lock);
117}
118
119/*
120 * Conditionally leave frozen/stopped state. Update cgroup's counters,
121 * and revisit the state of the cgroup, if necessary.
122 *
123 * If always_leave is not set, and the cgroup is freezing,
124 * we're racing with the cgroup freezing. In this case, we don't
125 * drop the frozen counter to avoid a transient switch to
126 * the unfrozen state.
127 */
128void cgroup_leave_frozen(bool always_leave)
129{
130 struct cgroup *cgrp;
131
132 spin_lock_irq(&css_set_lock);
133 cgrp = task_dfl_cgroup(current);
134 if (always_leave || !test_bit(CGRP_FREEZE, &cgrp->flags)) {
135 cgroup_dec_frozen_cnt(cgrp);
136 cgroup_update_frozen(cgrp);
137 WARN_ON_ONCE(!current->frozen);
138 current->frozen = false;
139 } else if (!(current->jobctl & JOBCTL_TRAP_FREEZE)) {
140 spin_lock(¤t->sighand->siglock);
141 current->jobctl |= JOBCTL_TRAP_FREEZE;
142 set_thread_flag(TIF_SIGPENDING);
143 spin_unlock(¤t->sighand->siglock);
144 }
145 spin_unlock_irq(&css_set_lock);
146}
147
148/*
149 * Freeze or unfreeze the task by setting or clearing the JOBCTL_TRAP_FREEZE
150 * jobctl bit.
151 */
152static void cgroup_freeze_task(struct task_struct *task, bool freeze)
153{
154 unsigned long flags;
155
156 /* If the task is about to die, don't bother with freezing it. */
157 if (!lock_task_sighand(task, &flags))
158 return;
159
160 if (freeze) {
161 task->jobctl |= JOBCTL_TRAP_FREEZE;
162 signal_wake_up(task, false);
163 } else {
164 task->jobctl &= ~JOBCTL_TRAP_FREEZE;
165 wake_up_process(task);
166 }
167
168 unlock_task_sighand(task, &flags);
169}
170
171/*
172 * Freeze or unfreeze all tasks in the given cgroup.
173 */
174static void cgroup_do_freeze(struct cgroup *cgrp, bool freeze)
175{
176 struct css_task_iter it;
177 struct task_struct *task;
178
179 lockdep_assert_held(&cgroup_mutex);
180
181 spin_lock_irq(&css_set_lock);
182 if (freeze)
183 set_bit(CGRP_FREEZE, &cgrp->flags);
184 else
185 clear_bit(CGRP_FREEZE, &cgrp->flags);
186 spin_unlock_irq(&css_set_lock);
187
188 if (freeze)
189 TRACE_CGROUP_PATH(freeze, cgrp);
190 else
191 TRACE_CGROUP_PATH(unfreeze, cgrp);
192
193 css_task_iter_start(&cgrp->self, 0, &it);
194 while ((task = css_task_iter_next(&it))) {
195 /*
196 * Ignore kernel threads here. Freezing cgroups containing
197 * kthreads isn't supported.
198 */
199 if (task->flags & PF_KTHREAD)
200 continue;
201 cgroup_freeze_task(task, freeze);
202 }
203 css_task_iter_end(&it);
204
205 /*
206 * Cgroup state should be revisited here to cover empty leaf cgroups
207 * and cgroups which descendants are already in the desired state.
208 */
209 spin_lock_irq(&css_set_lock);
210 if (cgrp->nr_descendants == cgrp->freezer.nr_frozen_descendants)
211 cgroup_update_frozen(cgrp);
212 spin_unlock_irq(&css_set_lock);
213}
214
215/*
216 * Adjust the task state (freeze or unfreeze) and revisit the state of
217 * source and destination cgroups.
218 */
219void cgroup_freezer_migrate_task(struct task_struct *task,
220 struct cgroup *src, struct cgroup *dst)
221{
222 lockdep_assert_held(&css_set_lock);
223
224 /*
225 * Kernel threads are not supposed to be frozen at all.
226 */
227 if (task->flags & PF_KTHREAD)
228 return;
229
230 /*
231 * It's not necessary to do changes if both of the src and dst cgroups
232 * are not freezing and task is not frozen.
233 */
234 if (!test_bit(CGRP_FREEZE, &src->flags) &&
235 !test_bit(CGRP_FREEZE, &dst->flags) &&
236 !task->frozen)
237 return;
238
239 /*
240 * Adjust counters of freezing and frozen tasks.
241 * Note, that if the task is frozen, but the destination cgroup is not
242 * frozen, we bump both counters to keep them balanced.
243 */
244 if (task->frozen) {
245 cgroup_inc_frozen_cnt(dst);
246 cgroup_dec_frozen_cnt(src);
247 }
248 cgroup_update_frozen(dst);
249 cgroup_update_frozen(src);
250
251 /*
252 * Force the task to the desired state.
253 */
254 cgroup_freeze_task(task, test_bit(CGRP_FREEZE, &dst->flags));
255}
256
257void cgroup_freeze(struct cgroup *cgrp, bool freeze)
258{
259 struct cgroup_subsys_state *css;
260 struct cgroup *parent;
261 struct cgroup *dsct;
262 bool applied = false;
263 bool old_e;
264
265 lockdep_assert_held(&cgroup_mutex);
266
267 /*
268 * Nothing changed? Just exit.
269 */
270 if (cgrp->freezer.freeze == freeze)
271 return;
272
273 cgrp->freezer.freeze = freeze;
274
275 /*
276 * Propagate changes downwards the cgroup tree.
277 */
278 css_for_each_descendant_pre(css, &cgrp->self) {
279 dsct = css->cgroup;
280
281 if (cgroup_is_dead(dsct))
282 continue;
283
284 /*
285 * e_freeze is affected by parent's e_freeze and dst's freeze.
286 * If old e_freeze eq new e_freeze, no change, its children
287 * will not be affected. So do nothing and skip the subtree
288 */
289 old_e = dsct->freezer.e_freeze;
290 parent = cgroup_parent(dsct);
291 dsct->freezer.e_freeze = (dsct->freezer.freeze ||
292 parent->freezer.e_freeze);
293 if (dsct->freezer.e_freeze == old_e) {
294 css = css_rightmost_descendant(css);
295 continue;
296 }
297
298 /*
299 * Do change actual state: freeze or unfreeze.
300 */
301 cgroup_do_freeze(dsct, freeze);
302 applied = true;
303 }
304
305 /*
306 * Even if the actual state hasn't changed, let's notify a user.
307 * The state can be enforced by an ancestor cgroup: the cgroup
308 * can already be in the desired state or it can be locked in the
309 * opposite state, so that the transition will never happen.
310 * In both cases it's better to notify a user, that there is
311 * nothing to wait for.
312 */
313 if (!applied) {
314 TRACE_CGROUP_PATH(notify_frozen, cgrp,
315 test_bit(CGRP_FROZEN, &cgrp->flags));
316 cgroup_file_notify(&cgrp->events_file);
317 }
318}