xref: /openbmc/linux/kernel/sched/core_sched.c (revision f05643a0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 /*
4  * A simple wrapper around refcount. An allocated sched_core_cookie's
5  * address is used to compute the cookie of the task.
6  */
7 struct sched_core_cookie {
8 	refcount_t refcnt;
9 };
10 
11 static unsigned long sched_core_alloc_cookie(void)
12 {
13 	struct sched_core_cookie *ck = kmalloc(sizeof(*ck), GFP_KERNEL);
14 	if (!ck)
15 		return 0;
16 
17 	refcount_set(&ck->refcnt, 1);
18 	sched_core_get();
19 
20 	return (unsigned long)ck;
21 }
22 
23 static void sched_core_put_cookie(unsigned long cookie)
24 {
25 	struct sched_core_cookie *ptr = (void *)cookie;
26 
27 	if (ptr && refcount_dec_and_test(&ptr->refcnt)) {
28 		kfree(ptr);
29 		sched_core_put();
30 	}
31 }
32 
33 static unsigned long sched_core_get_cookie(unsigned long cookie)
34 {
35 	struct sched_core_cookie *ptr = (void *)cookie;
36 
37 	if (ptr)
38 		refcount_inc(&ptr->refcnt);
39 
40 	return cookie;
41 }
42 
43 /*
44  * sched_core_update_cookie - replace the cookie on a task
45  * @p: the task to update
46  * @cookie: the new cookie
47  *
48  * Effectively exchange the task cookie; caller is responsible for lifetimes on
49  * both ends.
50  *
51  * Returns: the old cookie
52  */
53 static unsigned long sched_core_update_cookie(struct task_struct *p,
54 					      unsigned long cookie)
55 {
56 	unsigned long old_cookie;
57 	struct rq_flags rf;
58 	struct rq *rq;
59 	bool enqueued;
60 
61 	rq = task_rq_lock(p, &rf);
62 
63 	/*
64 	 * Since creating a cookie implies sched_core_get(), and we cannot set
65 	 * a cookie until after we've created it, similarly, we cannot destroy
66 	 * a cookie until after we've removed it, we must have core scheduling
67 	 * enabled here.
68 	 */
69 	SCHED_WARN_ON((p->core_cookie || cookie) && !sched_core_enabled(rq));
70 
71 	enqueued = sched_core_enqueued(p);
72 	if (enqueued)
73 		sched_core_dequeue(rq, p, DEQUEUE_SAVE);
74 
75 	old_cookie = p->core_cookie;
76 	p->core_cookie = cookie;
77 
78 	if (enqueued)
79 		sched_core_enqueue(rq, p);
80 
81 	/*
82 	 * If task is currently running, it may not be compatible anymore after
83 	 * the cookie change, so enter the scheduler on its CPU to schedule it
84 	 * away.
85 	 *
86 	 * Note that it is possible that as a result of this cookie change, the
87 	 * core has now entered/left forced idle state. Defer accounting to the
88 	 * next scheduling edge, rather than always forcing a reschedule here.
89 	 */
90 	if (task_running(rq, p))
91 		resched_curr(rq);
92 
93 	task_rq_unlock(rq, p, &rf);
94 
95 	return old_cookie;
96 }
97 
98 static unsigned long sched_core_clone_cookie(struct task_struct *p)
99 {
100 	unsigned long cookie, flags;
101 
102 	raw_spin_lock_irqsave(&p->pi_lock, flags);
103 	cookie = sched_core_get_cookie(p->core_cookie);
104 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
105 
106 	return cookie;
107 }
108 
109 void sched_core_fork(struct task_struct *p)
110 {
111 	RB_CLEAR_NODE(&p->core_node);
112 	p->core_cookie = sched_core_clone_cookie(current);
113 }
114 
115 void sched_core_free(struct task_struct *p)
116 {
117 	sched_core_put_cookie(p->core_cookie);
118 }
119 
120 static void __sched_core_set(struct task_struct *p, unsigned long cookie)
121 {
122 	cookie = sched_core_get_cookie(cookie);
123 	cookie = sched_core_update_cookie(p, cookie);
124 	sched_core_put_cookie(cookie);
125 }
126 
127 /* Called from prctl interface: PR_SCHED_CORE */
128 int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,
129 			 unsigned long uaddr)
130 {
131 	unsigned long cookie = 0, id = 0;
132 	struct task_struct *task, *p;
133 	struct pid *grp;
134 	int err = 0;
135 
136 	if (!static_branch_likely(&sched_smt_present))
137 		return -ENODEV;
138 
139 	BUILD_BUG_ON(PR_SCHED_CORE_SCOPE_THREAD != PIDTYPE_PID);
140 	BUILD_BUG_ON(PR_SCHED_CORE_SCOPE_THREAD_GROUP != PIDTYPE_TGID);
141 	BUILD_BUG_ON(PR_SCHED_CORE_SCOPE_PROCESS_GROUP != PIDTYPE_PGID);
142 
143 	if (type > PIDTYPE_PGID || cmd >= PR_SCHED_CORE_MAX || pid < 0 ||
144 	    (cmd != PR_SCHED_CORE_GET && uaddr))
145 		return -EINVAL;
146 
147 	rcu_read_lock();
148 	if (pid == 0) {
149 		task = current;
150 	} else {
151 		task = find_task_by_vpid(pid);
152 		if (!task) {
153 			rcu_read_unlock();
154 			return -ESRCH;
155 		}
156 	}
157 	get_task_struct(task);
158 	rcu_read_unlock();
159 
160 	/*
161 	 * Check if this process has the right to modify the specified
162 	 * process. Use the regular "ptrace_may_access()" checks.
163 	 */
164 	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
165 		err = -EPERM;
166 		goto out;
167 	}
168 
169 	switch (cmd) {
170 	case PR_SCHED_CORE_GET:
171 		if (type != PIDTYPE_PID || uaddr & 7) {
172 			err = -EINVAL;
173 			goto out;
174 		}
175 		cookie = sched_core_clone_cookie(task);
176 		if (cookie) {
177 			/* XXX improve ? */
178 			ptr_to_hashval((void *)cookie, &id);
179 		}
180 		err = put_user(id, (u64 __user *)uaddr);
181 		goto out;
182 
183 	case PR_SCHED_CORE_CREATE:
184 		cookie = sched_core_alloc_cookie();
185 		if (!cookie) {
186 			err = -ENOMEM;
187 			goto out;
188 		}
189 		break;
190 
191 	case PR_SCHED_CORE_SHARE_TO:
192 		cookie = sched_core_clone_cookie(current);
193 		break;
194 
195 	case PR_SCHED_CORE_SHARE_FROM:
196 		if (type != PIDTYPE_PID) {
197 			err = -EINVAL;
198 			goto out;
199 		}
200 		cookie = sched_core_clone_cookie(task);
201 		__sched_core_set(current, cookie);
202 		goto out;
203 
204 	default:
205 		err = -EINVAL;
206 		goto out;
207 	};
208 
209 	if (type == PIDTYPE_PID) {
210 		__sched_core_set(task, cookie);
211 		goto out;
212 	}
213 
214 	read_lock(&tasklist_lock);
215 	grp = task_pid_type(task, type);
216 
217 	do_each_pid_thread(grp, type, p) {
218 		if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS)) {
219 			err = -EPERM;
220 			goto out_tasklist;
221 		}
222 	} while_each_pid_thread(grp, type, p);
223 
224 	do_each_pid_thread(grp, type, p) {
225 		__sched_core_set(p, cookie);
226 	} while_each_pid_thread(grp, type, p);
227 out_tasklist:
228 	read_unlock(&tasklist_lock);
229 
230 out:
231 	sched_core_put_cookie(cookie);
232 	put_task_struct(task);
233 	return err;
234 }
235 
236 #ifdef CONFIG_SCHEDSTATS
237 
238 /* REQUIRES: rq->core's clock recently updated. */
239 void __sched_core_account_forceidle(struct rq *rq)
240 {
241 	const struct cpumask *smt_mask = cpu_smt_mask(cpu_of(rq));
242 	u64 delta, now = rq_clock(rq->core);
243 	struct rq *rq_i;
244 	struct task_struct *p;
245 	int i;
246 
247 	lockdep_assert_rq_held(rq);
248 
249 	WARN_ON_ONCE(!rq->core->core_forceidle_count);
250 
251 	if (rq->core->core_forceidle_start == 0)
252 		return;
253 
254 	delta = now - rq->core->core_forceidle_start;
255 	if (unlikely((s64)delta <= 0))
256 		return;
257 
258 	rq->core->core_forceidle_start = now;
259 
260 	if (WARN_ON_ONCE(!rq->core->core_forceidle_occupation)) {
261 		/* can't be forced idle without a running task */
262 	} else if (rq->core->core_forceidle_count > 1 ||
263 		   rq->core->core_forceidle_occupation > 1) {
264 		/*
265 		 * For larger SMT configurations, we need to scale the charged
266 		 * forced idle amount since there can be more than one forced
267 		 * idle sibling and more than one running cookied task.
268 		 */
269 		delta *= rq->core->core_forceidle_count;
270 		delta = div_u64(delta, rq->core->core_forceidle_occupation);
271 	}
272 
273 	for_each_cpu(i, smt_mask) {
274 		rq_i = cpu_rq(i);
275 		p = rq_i->core_pick ?: rq_i->curr;
276 
277 		if (p == rq_i->idle)
278 			continue;
279 
280 		__schedstat_add(p->stats.core_forceidle_sum, delta);
281 	}
282 }
283 
284 void __sched_core_tick(struct rq *rq)
285 {
286 	if (!rq->core->core_forceidle_count)
287 		return;
288 
289 	if (rq != rq->core)
290 		update_rq_clock(rq->core);
291 
292 	__sched_core_account_forceidle(rq);
293 }
294 
295 #endif /* CONFIG_SCHEDSTATS */
296