autogroup.c

  1// SPDX-License-Identifier: GPL-2.0
  2
  3/*
  4 * Auto-group scheduling implementation:
  5 */
  6
  7#include "autogroup.h"
  8#include "sched.h"
  9
 10unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
 11static struct autogroup autogroup_default;
 12static atomic_t autogroup_seq_nr;
 13
 14#ifdef CONFIG_SYSCTL
 15static const struct ctl_table sched_autogroup_sysctls[] = {
 16	{
 17		.procname       = "sched_autogroup_enabled",
 18		.data           = &sysctl_sched_autogroup_enabled,
 19		.maxlen         = sizeof(unsigned int),
 20		.mode           = 0644,
 21		.proc_handler   = proc_dointvec_minmax,
 22		.extra1         = SYSCTL_ZERO,
 23		.extra2         = SYSCTL_ONE,
 24	},
 25};
 26
 27static void __init sched_autogroup_sysctl_init(void)
 28{
 29	register_sysctl_init("kernel", sched_autogroup_sysctls);
 30}
 31#else /* !CONFIG_SYSCTL: */
 32#define sched_autogroup_sysctl_init() do { } while (0)
 33#endif /* !CONFIG_SYSCTL */
 34
 35void __init autogroup_init(struct task_struct *init_task)
 36{
 37	autogroup_default.tg = &root_task_group;
 38	kref_init(&autogroup_default.kref);
 39	init_rwsem(&autogroup_default.lock);
 40	init_task->signal->autogroup = &autogroup_default;
 41	sched_autogroup_sysctl_init();
 42}
 43
 44void autogroup_free(struct task_group *tg)
 45{
 46	kfree(tg->autogroup);
 47}
 48
 49static inline void autogroup_destroy(struct kref *kref)
 50{
 51	struct autogroup *ag = container_of(kref, struct autogroup, kref);
 52
 53#ifdef CONFIG_RT_GROUP_SCHED
 54	/* We've redirected RT tasks to the root task group... */
 55	ag->tg->rt_se = NULL;
 56	ag->tg->rt_rq = NULL;
 57#endif
 58	sched_release_group(ag->tg);
 59	sched_destroy_group(ag->tg);
 60}
 61
 62static inline void autogroup_kref_put(struct autogroup *ag)
 63{
 64	kref_put(&ag->kref, autogroup_destroy);
 65}
 66
 67static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
 68{
 69	kref_get(&ag->kref);
 70	return ag;
 71}
 72
 73static inline struct autogroup *autogroup_task_get(struct task_struct *p)
 74{
 75	struct autogroup *ag;
 76	unsigned long flags;
 77
 78	if (!lock_task_sighand(p, &flags))
 79		return autogroup_kref_get(&autogroup_default);
 80
 81	ag = autogroup_kref_get(p->signal->autogroup);
 82	unlock_task_sighand(p, &flags);
 83
 84	return ag;
 85}
 86
 87static inline struct autogroup *autogroup_create(void)
 88{
 89	struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
 90	struct task_group *tg;
 91
 92	if (!ag)
 93		goto out_fail;
 94
 95	tg = sched_create_group(&root_task_group);
 96	if (IS_ERR(tg))
 97		goto out_free;
 98
 99	kref_init(&ag->kref);
100	init_rwsem(&ag->lock);
101	ag->id = atomic_inc_return(&autogroup_seq_nr);
102	ag->tg = tg;
103#ifdef CONFIG_RT_GROUP_SCHED
104	/*
105	 * Autogroup RT tasks are redirected to the root task group
106	 * so we don't have to move tasks around upon policy change,
107	 * or flail around trying to allocate bandwidth on the fly.
108	 * A bandwidth exception in __sched_setscheduler() allows
109	 * the policy change to proceed.
110	 */
111	free_rt_sched_group(tg);
112	tg->rt_se = root_task_group.rt_se;
113	tg->rt_rq = root_task_group.rt_rq;
114#endif /* CONFIG_RT_GROUP_SCHED */
115	tg->autogroup = ag;
116
117	sched_online_group(tg, &root_task_group);
118	return ag;
119
120out_free:
121	kfree(ag);
122out_fail:
123	if (printk_ratelimit()) {
124		printk(KERN_WARNING "autogroup_create: %s failure.\n",
125			ag ? "sched_create_group()" : "kzalloc()");
126	}
127
128	return autogroup_kref_get(&autogroup_default);
129}
130
131bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
132{
133	if (tg != &root_task_group)
134		return false;
135	/*
136	 * If we race with autogroup_move_group() the caller can use the old
137	 * value of signal->autogroup but in this case sched_move_task() will
138	 * be called again before autogroup_kref_put().
139	 *
140	 * However, there is no way sched_autogroup_exit_task() could tell us
141	 * to avoid autogroup->tg, so we abuse PF_EXITING flag for this case.
142	 */
143	if (p->flags & PF_EXITING)
144		return false;
145
146	return true;
147}
148
149void sched_autogroup_exit_task(struct task_struct *p)
150{
151	/*
152	 * We are going to call exit_notify() and autogroup_move_group() can't
153	 * see this thread after that: we can no longer use signal->autogroup.
154	 * See the PF_EXITING check in task_wants_autogroup().
155	 */
156	sched_move_task(p, true);
157}
158
159static void
160autogroup_move_group(struct task_struct *p, struct autogroup *ag)
161{
162	struct autogroup *prev;
163	struct task_struct *t;
164	unsigned long flags;
165
166	if (WARN_ON_ONCE(!lock_task_sighand(p, &flags)))
167		return;
168
169	prev = p->signal->autogroup;
170	if (prev == ag) {
171		unlock_task_sighand(p, &flags);
172		return;
173	}
174
175	p->signal->autogroup = autogroup_kref_get(ag);
176	/*
177	 * We can't avoid sched_move_task() after we changed signal->autogroup,
178	 * this process can already run with task_group() == prev->tg or we can
179	 * race with cgroup code which can read autogroup = prev under rq->lock.
180	 * In the latter case for_each_thread() can not miss a migrating thread,
181	 * cpu_cgroup_attach() must not be possible after cgroup_task_exit()
182	 * and it can't be removed from thread list, we hold ->siglock.
183	 *
184	 * If an exiting thread was already removed from thread list we rely on
185	 * sched_autogroup_exit_task().
186	 */
187	for_each_thread(p, t)
188		sched_move_task(t, true);
189
190	unlock_task_sighand(p, &flags);
191	autogroup_kref_put(prev);
192}
193
194/* Allocates GFP_KERNEL, cannot be called under any spinlock: */
195void sched_autogroup_create_attach(struct task_struct *p)
196{
197	struct autogroup *ag = autogroup_create();
198
199	autogroup_move_group(p, ag);
200
201	/* Drop extra reference added by autogroup_create(): */
202	autogroup_kref_put(ag);
203}
204EXPORT_SYMBOL(sched_autogroup_create_attach);
205
206/* Cannot be called under siglock. Currently has no users: */
207void sched_autogroup_detach(struct task_struct *p)
208{
209	autogroup_move_group(p, &autogroup_default);
210}
211EXPORT_SYMBOL(sched_autogroup_detach);
212
213void sched_autogroup_fork(struct signal_struct *sig)
214{
215	sig->autogroup = autogroup_task_get(current);
216}
217
218void sched_autogroup_exit(struct signal_struct *sig)
219{
220	autogroup_kref_put(sig->autogroup);
221}
222
223static int __init setup_autogroup(char *str)
224{
225	sysctl_sched_autogroup_enabled = 0;
226
227	return 1;
228}
229__setup("noautogroup", setup_autogroup);
230
231#ifdef CONFIG_PROC_FS
232
233int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
234{
235	static unsigned long next = INITIAL_JIFFIES;
236	struct autogroup *ag;
237	unsigned long shares;
238	int err, idx;
239
240	if (nice < MIN_NICE || nice > MAX_NICE)
241		return -EINVAL;
242
243	err = security_task_setnice(current, nice);
244	if (err)
245		return err;
246
247	if (nice < 0 && !can_nice(current, nice))
248		return -EPERM;
249
250	/* This is a heavy operation, taking global locks.. */
251	if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
252		return -EAGAIN;
253
254	next = HZ / 10 + jiffies;
255	ag = autogroup_task_get(p);
256
257	idx = array_index_nospec(nice + 20, 40);
258	shares = scale_load(sched_prio_to_weight[idx]);
259
260	down_write(&ag->lock);
261	err = sched_group_set_shares(ag->tg, shares);
262	if (!err)
263		ag->nice = nice;
264	up_write(&ag->lock);
265
266	autogroup_kref_put(ag);
267
268	return err;
269}
270
271void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
272{
273	struct autogroup *ag = autogroup_task_get(p);
274
275	if (!task_group_is_autogroup(ag->tg))
276		goto out;
277
278	down_read(&ag->lock);
279	seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
280	up_read(&ag->lock);
281
282out:
283	autogroup_kref_put(ag);
284}
285#endif /* CONFIG_PROC_FS */
286
287int autogroup_path(struct task_group *tg, char *buf, int buflen)
288{
289	if (!task_group_is_autogroup(tg))
290		return 0;
291
292	return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
293}