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