xref: /openbmc/linux/kernel/sched/autogroup.c (revision 2359ccdd)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Auto-group scheduling implementation:
4  */
5 #include "sched.h"
6 
7 unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
8 static struct autogroup autogroup_default;
9 static atomic_t autogroup_seq_nr;
10 
11 void __init autogroup_init(struct task_struct *init_task)
12 {
13 	autogroup_default.tg = &root_task_group;
14 	kref_init(&autogroup_default.kref);
15 	init_rwsem(&autogroup_default.lock);
16 	init_task->signal->autogroup = &autogroup_default;
17 }
18 
19 void autogroup_free(struct task_group *tg)
20 {
21 	kfree(tg->autogroup);
22 }
23 
24 static inline void autogroup_destroy(struct kref *kref)
25 {
26 	struct autogroup *ag = container_of(kref, struct autogroup, kref);
27 
28 #ifdef CONFIG_RT_GROUP_SCHED
29 	/* We've redirected RT tasks to the root task group... */
30 	ag->tg->rt_se = NULL;
31 	ag->tg->rt_rq = NULL;
32 #endif
33 	sched_offline_group(ag->tg);
34 	sched_destroy_group(ag->tg);
35 }
36 
37 static inline void autogroup_kref_put(struct autogroup *ag)
38 {
39 	kref_put(&ag->kref, autogroup_destroy);
40 }
41 
42 static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
43 {
44 	kref_get(&ag->kref);
45 	return ag;
46 }
47 
48 static inline struct autogroup *autogroup_task_get(struct task_struct *p)
49 {
50 	struct autogroup *ag;
51 	unsigned long flags;
52 
53 	if (!lock_task_sighand(p, &flags))
54 		return autogroup_kref_get(&autogroup_default);
55 
56 	ag = autogroup_kref_get(p->signal->autogroup);
57 	unlock_task_sighand(p, &flags);
58 
59 	return ag;
60 }
61 
62 static inline struct autogroup *autogroup_create(void)
63 {
64 	struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
65 	struct task_group *tg;
66 
67 	if (!ag)
68 		goto out_fail;
69 
70 	tg = sched_create_group(&root_task_group);
71 	if (IS_ERR(tg))
72 		goto out_free;
73 
74 	kref_init(&ag->kref);
75 	init_rwsem(&ag->lock);
76 	ag->id = atomic_inc_return(&autogroup_seq_nr);
77 	ag->tg = tg;
78 #ifdef CONFIG_RT_GROUP_SCHED
79 	/*
80 	 * Autogroup RT tasks are redirected to the root task group
81 	 * so we don't have to move tasks around upon policy change,
82 	 * or flail around trying to allocate bandwidth on the fly.
83 	 * A bandwidth exception in __sched_setscheduler() allows
84 	 * the policy change to proceed.
85 	 */
86 	free_rt_sched_group(tg);
87 	tg->rt_se = root_task_group.rt_se;
88 	tg->rt_rq = root_task_group.rt_rq;
89 #endif
90 	tg->autogroup = ag;
91 
92 	sched_online_group(tg, &root_task_group);
93 	return ag;
94 
95 out_free:
96 	kfree(ag);
97 out_fail:
98 	if (printk_ratelimit()) {
99 		printk(KERN_WARNING "autogroup_create: %s failure.\n",
100 			ag ? "sched_create_group()" : "kzalloc()");
101 	}
102 
103 	return autogroup_kref_get(&autogroup_default);
104 }
105 
106 bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
107 {
108 	if (tg != &root_task_group)
109 		return false;
110 	/*
111 	 * If we race with autogroup_move_group() the caller can use the old
112 	 * value of signal->autogroup but in this case sched_move_task() will
113 	 * be called again before autogroup_kref_put().
114 	 *
115 	 * However, there is no way sched_autogroup_exit_task() could tell us
116 	 * to avoid autogroup->tg, so we abuse PF_EXITING flag for this case.
117 	 */
118 	if (p->flags & PF_EXITING)
119 		return false;
120 
121 	return true;
122 }
123 
124 void sched_autogroup_exit_task(struct task_struct *p)
125 {
126 	/*
127 	 * We are going to call exit_notify() and autogroup_move_group() can't
128 	 * see this thread after that: we can no longer use signal->autogroup.
129 	 * See the PF_EXITING check in task_wants_autogroup().
130 	 */
131 	sched_move_task(p);
132 }
133 
134 static void
135 autogroup_move_group(struct task_struct *p, struct autogroup *ag)
136 {
137 	struct autogroup *prev;
138 	struct task_struct *t;
139 	unsigned long flags;
140 
141 	BUG_ON(!lock_task_sighand(p, &flags));
142 
143 	prev = p->signal->autogroup;
144 	if (prev == ag) {
145 		unlock_task_sighand(p, &flags);
146 		return;
147 	}
148 
149 	p->signal->autogroup = autogroup_kref_get(ag);
150 	/*
151 	 * We can't avoid sched_move_task() after we changed signal->autogroup,
152 	 * this process can already run with task_group() == prev->tg or we can
153 	 * race with cgroup code which can read autogroup = prev under rq->lock.
154 	 * In the latter case for_each_thread() can not miss a migrating thread,
155 	 * cpu_cgroup_attach() must not be possible after cgroup_exit() and it
156 	 * can't be removed from thread list, we hold ->siglock.
157 	 *
158 	 * If an exiting thread was already removed from thread list we rely on
159 	 * sched_autogroup_exit_task().
160 	 */
161 	for_each_thread(p, t)
162 		sched_move_task(t);
163 
164 	unlock_task_sighand(p, &flags);
165 	autogroup_kref_put(prev);
166 }
167 
168 /* Allocates GFP_KERNEL, cannot be called under any spinlock: */
169 void sched_autogroup_create_attach(struct task_struct *p)
170 {
171 	struct autogroup *ag = autogroup_create();
172 
173 	autogroup_move_group(p, ag);
174 
175 	/* Drop extra reference added by autogroup_create(): */
176 	autogroup_kref_put(ag);
177 }
178 EXPORT_SYMBOL(sched_autogroup_create_attach);
179 
180 /* Cannot be called under siglock. Currently has no users: */
181 void sched_autogroup_detach(struct task_struct *p)
182 {
183 	autogroup_move_group(p, &autogroup_default);
184 }
185 EXPORT_SYMBOL(sched_autogroup_detach);
186 
187 void sched_autogroup_fork(struct signal_struct *sig)
188 {
189 	sig->autogroup = autogroup_task_get(current);
190 }
191 
192 void sched_autogroup_exit(struct signal_struct *sig)
193 {
194 	autogroup_kref_put(sig->autogroup);
195 }
196 
197 static int __init setup_autogroup(char *str)
198 {
199 	sysctl_sched_autogroup_enabled = 0;
200 
201 	return 1;
202 }
203 __setup("noautogroup", setup_autogroup);
204 
205 #ifdef CONFIG_PROC_FS
206 
207 int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
208 {
209 	static unsigned long next = INITIAL_JIFFIES;
210 	struct autogroup *ag;
211 	unsigned long shares;
212 	int err;
213 
214 	if (nice < MIN_NICE || nice > MAX_NICE)
215 		return -EINVAL;
216 
217 	err = security_task_setnice(current, nice);
218 	if (err)
219 		return err;
220 
221 	if (nice < 0 && !can_nice(current, nice))
222 		return -EPERM;
223 
224 	/* This is a heavy operation, taking global locks.. */
225 	if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
226 		return -EAGAIN;
227 
228 	next = HZ / 10 + jiffies;
229 	ag = autogroup_task_get(p);
230 	shares = scale_load(sched_prio_to_weight[nice + 20]);
231 
232 	down_write(&ag->lock);
233 	err = sched_group_set_shares(ag->tg, shares);
234 	if (!err)
235 		ag->nice = nice;
236 	up_write(&ag->lock);
237 
238 	autogroup_kref_put(ag);
239 
240 	return err;
241 }
242 
243 void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
244 {
245 	struct autogroup *ag = autogroup_task_get(p);
246 
247 	if (!task_group_is_autogroup(ag->tg))
248 		goto out;
249 
250 	down_read(&ag->lock);
251 	seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
252 	up_read(&ag->lock);
253 
254 out:
255 	autogroup_kref_put(ag);
256 }
257 #endif /* CONFIG_PROC_FS */
258 
259 #ifdef CONFIG_SCHED_DEBUG
260 int autogroup_path(struct task_group *tg, char *buf, int buflen)
261 {
262 	if (!task_group_is_autogroup(tg))
263 		return 0;
264 
265 	return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
266 }
267 #endif
268