Loading...
1#include "sched.h"
2
3#include <linux/proc_fs.h>
4#include <linux/seq_file.h>
5#include <linux/kallsyms.h>
6#include <linux/utsname.h>
7#include <linux/security.h>
8#include <linux/export.h>
9
10unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
11static struct autogroup autogroup_default;
12static atomic_t autogroup_seq_nr;
13
14void __init autogroup_init(struct task_struct *init_task)
15{
16 autogroup_default.tg = &root_task_group;
17 kref_init(&autogroup_default.kref);
18 init_rwsem(&autogroup_default.lock);
19 init_task->signal->autogroup = &autogroup_default;
20}
21
22void autogroup_free(struct task_group *tg)
23{
24 kfree(tg->autogroup);
25}
26
27static inline void autogroup_destroy(struct kref *kref)
28{
29 struct autogroup *ag = container_of(kref, struct autogroup, kref);
30
31#ifdef CONFIG_RT_GROUP_SCHED
32 /* We've redirected RT tasks to the root task group... */
33 ag->tg->rt_se = NULL;
34 ag->tg->rt_rq = NULL;
35#endif
36 sched_offline_group(ag->tg);
37 sched_destroy_group(ag->tg);
38}
39
40static inline void autogroup_kref_put(struct autogroup *ag)
41{
42 kref_put(&ag->kref, autogroup_destroy);
43}
44
45static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
46{
47 kref_get(&ag->kref);
48 return ag;
49}
50
51static inline struct autogroup *autogroup_task_get(struct task_struct *p)
52{
53 struct autogroup *ag;
54 unsigned long flags;
55
56 if (!lock_task_sighand(p, &flags))
57 return autogroup_kref_get(&autogroup_default);
58
59 ag = autogroup_kref_get(p->signal->autogroup);
60 unlock_task_sighand(p, &flags);
61
62 return ag;
63}
64
65static inline struct autogroup *autogroup_create(void)
66{
67 struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
68 struct task_group *tg;
69
70 if (!ag)
71 goto out_fail;
72
73 tg = sched_create_group(&root_task_group);
74
75 if (IS_ERR(tg))
76 goto out_free;
77
78 kref_init(&ag->kref);
79 init_rwsem(&ag->lock);
80 ag->id = atomic_inc_return(&autogroup_seq_nr);
81 ag->tg = tg;
82#ifdef CONFIG_RT_GROUP_SCHED
83 /*
84 * Autogroup RT tasks are redirected to the root task group
85 * so we don't have to move tasks around upon policy change,
86 * or flail around trying to allocate bandwidth on the fly.
87 * A bandwidth exception in __sched_setscheduler() allows
88 * the policy change to proceed.
89 */
90 free_rt_sched_group(tg);
91 tg->rt_se = root_task_group.rt_se;
92 tg->rt_rq = root_task_group.rt_rq;
93#endif
94 tg->autogroup = ag;
95
96 sched_online_group(tg, &root_task_group);
97 return ag;
98
99out_free:
100 kfree(ag);
101out_fail:
102 if (printk_ratelimit()) {
103 printk(KERN_WARNING "autogroup_create: %s failure.\n",
104 ag ? "sched_create_group()" : "kmalloc()");
105 }
106
107 return autogroup_kref_get(&autogroup_default);
108}
109
110bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
111{
112 if (tg != &root_task_group)
113 return false;
114 /*
115 * If we race with autogroup_move_group() the caller can use the old
116 * value of signal->autogroup but in this case sched_move_task() will
117 * be called again before autogroup_kref_put().
118 *
119 * However, there is no way sched_autogroup_exit_task() could tell us
120 * to avoid autogroup->tg, so we abuse PF_EXITING flag for this case.
121 */
122 if (p->flags & PF_EXITING)
123 return false;
124
125 return true;
126}
127
128void sched_autogroup_exit_task(struct task_struct *p)
129{
130 /*
131 * We are going to call exit_notify() and autogroup_move_group() can't
132 * see this thread after that: we can no longer use signal->autogroup.
133 * See the PF_EXITING check in task_wants_autogroup().
134 */
135 sched_move_task(p);
136}
137
138static void
139autogroup_move_group(struct task_struct *p, struct autogroup *ag)
140{
141 struct autogroup *prev;
142 struct task_struct *t;
143 unsigned long flags;
144
145 BUG_ON(!lock_task_sighand(p, &flags));
146
147 prev = p->signal->autogroup;
148 if (prev == ag) {
149 unlock_task_sighand(p, &flags);
150 return;
151 }
152
153 p->signal->autogroup = autogroup_kref_get(ag);
154 /*
155 * We can't avoid sched_move_task() after we changed signal->autogroup,
156 * this process can already run with task_group() == prev->tg or we can
157 * race with cgroup code which can read autogroup = prev under rq->lock.
158 * In the latter case for_each_thread() can not miss a migrating thread,
159 * cpu_cgroup_attach() must not be possible after cgroup_exit() and it
160 * can't be removed from thread list, we hold ->siglock.
161 *
162 * If an exiting thread was already removed from thread list we rely on
163 * sched_autogroup_exit_task().
164 */
165 for_each_thread(p, t)
166 sched_move_task(t);
167
168 unlock_task_sighand(p, &flags);
169 autogroup_kref_put(prev);
170}
171
172/* Allocates GFP_KERNEL, cannot be called under any spinlock */
173void sched_autogroup_create_attach(struct task_struct *p)
174{
175 struct autogroup *ag = autogroup_create();
176
177 autogroup_move_group(p, ag);
178 /* drop extra reference added by autogroup_create() */
179 autogroup_kref_put(ag);
180}
181EXPORT_SYMBOL(sched_autogroup_create_attach);
182
183/* Cannot be called under siglock. Currently has no users */
184void sched_autogroup_detach(struct task_struct *p)
185{
186 autogroup_move_group(p, &autogroup_default);
187}
188EXPORT_SYMBOL(sched_autogroup_detach);
189
190void sched_autogroup_fork(struct signal_struct *sig)
191{
192 sig->autogroup = autogroup_task_get(current);
193}
194
195void sched_autogroup_exit(struct signal_struct *sig)
196{
197 autogroup_kref_put(sig->autogroup);
198}
199
200static int __init setup_autogroup(char *str)
201{
202 sysctl_sched_autogroup_enabled = 0;
203
204 return 1;
205}
206
207__setup("noautogroup", setup_autogroup);
208
209#ifdef CONFIG_PROC_FS
210
211int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
212{
213 static unsigned long next = INITIAL_JIFFIES;
214 struct autogroup *ag;
215 unsigned long shares;
216 int err;
217
218 if (nice < MIN_NICE || nice > MAX_NICE)
219 return -EINVAL;
220
221 err = security_task_setnice(current, nice);
222 if (err)
223 return err;
224
225 if (nice < 0 && !can_nice(current, nice))
226 return -EPERM;
227
228 /* this is a heavy operation taking global locks.. */
229 if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
230 return -EAGAIN;
231
232 next = HZ / 10 + jiffies;
233 ag = autogroup_task_get(p);
234 shares = scale_load(sched_prio_to_weight[nice + 20]);
235
236 down_write(&ag->lock);
237 err = sched_group_set_shares(ag->tg, shares);
238 if (!err)
239 ag->nice = nice;
240 up_write(&ag->lock);
241
242 autogroup_kref_put(ag);
243
244 return err;
245}
246
247void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
248{
249 struct autogroup *ag = autogroup_task_get(p);
250
251 if (!task_group_is_autogroup(ag->tg))
252 goto out;
253
254 down_read(&ag->lock);
255 seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
256 up_read(&ag->lock);
257
258out:
259 autogroup_kref_put(ag);
260}
261#endif /* CONFIG_PROC_FS */
262
263#ifdef CONFIG_SCHED_DEBUG
264int autogroup_path(struct task_group *tg, char *buf, int buflen)
265{
266 if (!task_group_is_autogroup(tg))
267 return 0;
268
269 return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
270}
271#endif /* CONFIG_SCHED_DEBUG */
1#ifdef CONFIG_SCHED_AUTOGROUP
2
3#include "sched.h"
4
5#include <linux/proc_fs.h>
6#include <linux/seq_file.h>
7#include <linux/kallsyms.h>
8#include <linux/utsname.h>
9#include <linux/security.h>
10#include <linux/export.h>
11
12unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
13static struct autogroup autogroup_default;
14static atomic_t autogroup_seq_nr;
15
16void __init autogroup_init(struct task_struct *init_task)
17{
18 autogroup_default.tg = &root_task_group;
19 kref_init(&autogroup_default.kref);
20 init_rwsem(&autogroup_default.lock);
21 init_task->signal->autogroup = &autogroup_default;
22}
23
24void autogroup_free(struct task_group *tg)
25{
26 kfree(tg->autogroup);
27}
28
29static inline void autogroup_destroy(struct kref *kref)
30{
31 struct autogroup *ag = container_of(kref, struct autogroup, kref);
32
33#ifdef CONFIG_RT_GROUP_SCHED
34 /* We've redirected RT tasks to the root task group... */
35 ag->tg->rt_se = NULL;
36 ag->tg->rt_rq = NULL;
37#endif
38 sched_offline_group(ag->tg);
39 sched_destroy_group(ag->tg);
40}
41
42static inline void autogroup_kref_put(struct autogroup *ag)
43{
44 kref_put(&ag->kref, autogroup_destroy);
45}
46
47static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
48{
49 kref_get(&ag->kref);
50 return ag;
51}
52
53static inline struct autogroup *autogroup_task_get(struct task_struct *p)
54{
55 struct autogroup *ag;
56 unsigned long flags;
57
58 if (!lock_task_sighand(p, &flags))
59 return autogroup_kref_get(&autogroup_default);
60
61 ag = autogroup_kref_get(p->signal->autogroup);
62 unlock_task_sighand(p, &flags);
63
64 return ag;
65}
66
67static inline struct autogroup *autogroup_create(void)
68{
69 struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
70 struct task_group *tg;
71
72 if (!ag)
73 goto out_fail;
74
75 tg = sched_create_group(&root_task_group);
76
77 if (IS_ERR(tg))
78 goto out_free;
79
80 kref_init(&ag->kref);
81 init_rwsem(&ag->lock);
82 ag->id = atomic_inc_return(&autogroup_seq_nr);
83 ag->tg = tg;
84#ifdef CONFIG_RT_GROUP_SCHED
85 /*
86 * Autogroup RT tasks are redirected to the root task group
87 * so we don't have to move tasks around upon policy change,
88 * or flail around trying to allocate bandwidth on the fly.
89 * A bandwidth exception in __sched_setscheduler() allows
90 * the policy change to proceed. Thereafter, task_group()
91 * returns &root_task_group, so zero bandwidth is required.
92 */
93 free_rt_sched_group(tg);
94 tg->rt_se = root_task_group.rt_se;
95 tg->rt_rq = root_task_group.rt_rq;
96#endif
97 tg->autogroup = ag;
98
99 sched_online_group(tg, &root_task_group);
100 return ag;
101
102out_free:
103 kfree(ag);
104out_fail:
105 if (printk_ratelimit()) {
106 printk(KERN_WARNING "autogroup_create: %s failure.\n",
107 ag ? "sched_create_group()" : "kmalloc()");
108 }
109
110 return autogroup_kref_get(&autogroup_default);
111}
112
113bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
114{
115 if (tg != &root_task_group)
116 return false;
117
118 if (p->sched_class != &fair_sched_class)
119 return false;
120
121 /*
122 * We can only assume the task group can't go away on us if
123 * autogroup_move_group() can see us on ->thread_group list.
124 */
125 if (p->flags & PF_EXITING)
126 return false;
127
128 return true;
129}
130
131static void
132autogroup_move_group(struct task_struct *p, struct autogroup *ag)
133{
134 struct autogroup *prev;
135 struct task_struct *t;
136 unsigned long flags;
137
138 BUG_ON(!lock_task_sighand(p, &flags));
139
140 prev = p->signal->autogroup;
141 if (prev == ag) {
142 unlock_task_sighand(p, &flags);
143 return;
144 }
145
146 p->signal->autogroup = autogroup_kref_get(ag);
147
148 if (!ACCESS_ONCE(sysctl_sched_autogroup_enabled))
149 goto out;
150
151 t = p;
152 do {
153 sched_move_task(t);
154 } while_each_thread(p, t);
155
156out:
157 unlock_task_sighand(p, &flags);
158 autogroup_kref_put(prev);
159}
160
161/* Allocates GFP_KERNEL, cannot be called under any spinlock */
162void sched_autogroup_create_attach(struct task_struct *p)
163{
164 struct autogroup *ag = autogroup_create();
165
166 autogroup_move_group(p, ag);
167 /* drop extra reference added by autogroup_create() */
168 autogroup_kref_put(ag);
169}
170EXPORT_SYMBOL(sched_autogroup_create_attach);
171
172/* Cannot be called under siglock. Currently has no users */
173void sched_autogroup_detach(struct task_struct *p)
174{
175 autogroup_move_group(p, &autogroup_default);
176}
177EXPORT_SYMBOL(sched_autogroup_detach);
178
179void sched_autogroup_fork(struct signal_struct *sig)
180{
181 sig->autogroup = autogroup_task_get(current);
182}
183
184void sched_autogroup_exit(struct signal_struct *sig)
185{
186 autogroup_kref_put(sig->autogroup);
187}
188
189static int __init setup_autogroup(char *str)
190{
191 sysctl_sched_autogroup_enabled = 0;
192
193 return 1;
194}
195
196__setup("noautogroup", setup_autogroup);
197
198#ifdef CONFIG_PROC_FS
199
200int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
201{
202 static unsigned long next = INITIAL_JIFFIES;
203 struct autogroup *ag;
204 int err;
205
206 if (nice < MIN_NICE || nice > MAX_NICE)
207 return -EINVAL;
208
209 err = security_task_setnice(current, nice);
210 if (err)
211 return err;
212
213 if (nice < 0 && !can_nice(current, nice))
214 return -EPERM;
215
216 /* this is a heavy operation taking global locks.. */
217 if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
218 return -EAGAIN;
219
220 next = HZ / 10 + jiffies;
221 ag = autogroup_task_get(p);
222
223 down_write(&ag->lock);
224 err = sched_group_set_shares(ag->tg, prio_to_weight[nice + 20]);
225 if (!err)
226 ag->nice = nice;
227 up_write(&ag->lock);
228
229 autogroup_kref_put(ag);
230
231 return err;
232}
233
234void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
235{
236 struct autogroup *ag = autogroup_task_get(p);
237
238 if (!task_group_is_autogroup(ag->tg))
239 goto out;
240
241 down_read(&ag->lock);
242 seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
243 up_read(&ag->lock);
244
245out:
246 autogroup_kref_put(ag);
247}
248#endif /* CONFIG_PROC_FS */
249
250#ifdef CONFIG_SCHED_DEBUG
251int autogroup_path(struct task_group *tg, char *buf, int buflen)
252{
253 if (!task_group_is_autogroup(tg))
254 return 0;
255
256 return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
257}
258#endif /* CONFIG_SCHED_DEBUG */
259
260#endif /* CONFIG_SCHED_AUTOGROUP */