1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright 2018 Linaro Limited
4  *
5  * Author: Daniel Lezcano <daniel.lezcano@linaro.org>
6  *
7  * The idle injection framework provides a way to force CPUs to enter idle
8  * states for a specified fraction of time over a specified period.
9  *
10  * It relies on the smpboot kthreads feature providing common code for CPU
11  * hotplug and thread [un]parking.
12  *
13  * All of the kthreads used for idle injection are created at init time.
14  *
15  * Next, the users of the the idle injection framework provide a cpumask via
16  * its register function. The kthreads will be synchronized with respect to
17  * this cpumask.
18  *
19  * The idle + run duration is specified via separate helpers and that allows
20  * idle injection to be started.
21  *
22  * The idle injection kthreads will call play_idle_precise() with the idle
23  * duration and max allowed latency specified as per the above.
24  *
25  * After all of them have been woken up, a timer is set to start the next idle
26  * injection cycle.
27  *
28  * The timer interrupt handler will wake up the idle injection kthreads for
29  * all of the CPUs in the cpumask provided by the user.
30  *
31  * Idle injection is stopped synchronously and no leftover idle injection
32  * kthread activity after its completion is guaranteed.
33  *
34  * It is up to the user of this framework to provide a lock for higher-level
35  * synchronization to prevent race conditions like starting idle injection
36  * while unregistering from the framework.
37  */
38 #define pr_fmt(fmt) "ii_dev: " fmt
39 
40 #include <linux/cpu.h>
41 #include <linux/hrtimer.h>
42 #include <linux/kthread.h>
43 #include <linux/sched.h>
44 #include <linux/slab.h>
45 #include <linux/smpboot.h>
46 #include <linux/idle_inject.h>
47 
48 #include <uapi/linux/sched/types.h>
49 
50 /**
51  * struct idle_inject_thread - task on/off switch structure
52  * @tsk: task injecting the idle cycles
53  * @should_run: whether or not to run the task (for the smpboot kthread API)
54  */
55 struct idle_inject_thread {
56 	struct task_struct *tsk;
57 	int should_run;
58 };
59 
60 /**
61  * struct idle_inject_device - idle injection data
62  * @timer: idle injection period timer
63  * @idle_duration_us: duration of CPU idle time to inject
64  * @run_duration_us: duration of CPU run time to allow
65  * @latency_us: max allowed latency
66  * @cpumask: mask of CPUs affected by idle injection
67  */
68 struct idle_inject_device {
69 	struct hrtimer timer;
70 	unsigned int idle_duration_us;
71 	unsigned int run_duration_us;
72 	unsigned int latency_us;
73 	unsigned long cpumask[];
74 };
75 
76 static DEFINE_PER_CPU(struct idle_inject_thread, idle_inject_thread);
77 static DEFINE_PER_CPU(struct idle_inject_device *, idle_inject_device);
78 
79 /**
80  * idle_inject_wakeup - Wake up idle injection threads
81  * @ii_dev: target idle injection device
82  *
83  * Every idle injection task associated with the given idle injection device
84  * and running on an online CPU will be woken up.
85  */
idle_inject_wakeup(struct idle_inject_device * ii_dev)86 static void idle_inject_wakeup(struct idle_inject_device *ii_dev)
87 {
88 	struct idle_inject_thread *iit;
89 	unsigned int cpu;
90 
91 	for_each_cpu_and(cpu, to_cpumask(ii_dev->cpumask), cpu_online_mask) {
92 		iit = per_cpu_ptr(&idle_inject_thread, cpu);
93 		iit->should_run = 1;
94 		wake_up_process(iit->tsk);
95 	}
96 }
97 
98 /**
99  * idle_inject_timer_fn - idle injection timer function
100  * @timer: idle injection hrtimer
101  *
102  * This function is called when the idle injection timer expires.  It wakes up
103  * idle injection tasks associated with the timer and they, in turn, invoke
104  * play_idle_precise() to inject a specified amount of CPU idle time.
105  *
106  * Return: HRTIMER_RESTART.
107  */
idle_inject_timer_fn(struct hrtimer * timer)108 static enum hrtimer_restart idle_inject_timer_fn(struct hrtimer *timer)
109 {
110 	unsigned int duration_us;
111 	struct idle_inject_device *ii_dev =
112 		container_of(timer, struct idle_inject_device, timer);
113 
114 	duration_us = READ_ONCE(ii_dev->run_duration_us);
115 	duration_us += READ_ONCE(ii_dev->idle_duration_us);
116 
117 	idle_inject_wakeup(ii_dev);
118 
119 	hrtimer_forward_now(timer, ns_to_ktime(duration_us * NSEC_PER_USEC));
120 
121 	return HRTIMER_RESTART;
122 }
123 
124 /**
125  * idle_inject_fn - idle injection work function
126  * @cpu: the CPU owning the task
127  *
128  * This function calls play_idle_precise() to inject a specified amount of CPU
129  * idle time.
130  */
idle_inject_fn(unsigned int cpu)131 static void idle_inject_fn(unsigned int cpu)
132 {
133 	struct idle_inject_device *ii_dev;
134 	struct idle_inject_thread *iit;
135 
136 	ii_dev = per_cpu(idle_inject_device, cpu);
137 	iit = per_cpu_ptr(&idle_inject_thread, cpu);
138 
139 	/*
140 	 * Let the smpboot main loop know that the task should not run again.
141 	 */
142 	iit->should_run = 0;
143 
144 	play_idle_precise(READ_ONCE(ii_dev->idle_duration_us) * NSEC_PER_USEC,
145 			  READ_ONCE(ii_dev->latency_us) * NSEC_PER_USEC);
146 }
147 
148 /**
149  * idle_inject_set_duration - idle and run duration update helper
150  * @run_duration_us: CPU run time to allow in microseconds
151  * @idle_duration_us: CPU idle time to inject in microseconds
152  */
idle_inject_set_duration(struct idle_inject_device * ii_dev,unsigned int run_duration_us,unsigned int idle_duration_us)153 void idle_inject_set_duration(struct idle_inject_device *ii_dev,
154 			      unsigned int run_duration_us,
155 			      unsigned int idle_duration_us)
156 {
157 	if (run_duration_us && idle_duration_us) {
158 		WRITE_ONCE(ii_dev->run_duration_us, run_duration_us);
159 		WRITE_ONCE(ii_dev->idle_duration_us, idle_duration_us);
160 	}
161 }
162 
163 /**
164  * idle_inject_get_duration - idle and run duration retrieval helper
165  * @run_duration_us: memory location to store the current CPU run time
166  * @idle_duration_us: memory location to store the current CPU idle time
167  */
idle_inject_get_duration(struct idle_inject_device * ii_dev,unsigned int * run_duration_us,unsigned int * idle_duration_us)168 void idle_inject_get_duration(struct idle_inject_device *ii_dev,
169 			      unsigned int *run_duration_us,
170 			      unsigned int *idle_duration_us)
171 {
172 	*run_duration_us = READ_ONCE(ii_dev->run_duration_us);
173 	*idle_duration_us = READ_ONCE(ii_dev->idle_duration_us);
174 }
175 
176 /**
177  * idle_inject_set_latency - set the maximum latency allowed
178  * @latency_us: set the latency requirement for the idle state
179  */
idle_inject_set_latency(struct idle_inject_device * ii_dev,unsigned int latency_us)180 void idle_inject_set_latency(struct idle_inject_device *ii_dev,
181 			     unsigned int latency_us)
182 {
183 	WRITE_ONCE(ii_dev->latency_us, latency_us);
184 }
185 
186 /**
187  * idle_inject_start - start idle injections
188  * @ii_dev: idle injection control device structure
189  *
190  * The function starts idle injection by first waking up all of the idle
191  * injection kthreads associated with @ii_dev to let them inject CPU idle time
192  * sets up a timer to start the next idle injection period.
193  *
194  * Return: -EINVAL if the CPU idle or CPU run time is not set or 0 on success.
195  */
idle_inject_start(struct idle_inject_device * ii_dev)196 int idle_inject_start(struct idle_inject_device *ii_dev)
197 {
198 	unsigned int idle_duration_us = READ_ONCE(ii_dev->idle_duration_us);
199 	unsigned int run_duration_us = READ_ONCE(ii_dev->run_duration_us);
200 
201 	if (!idle_duration_us || !run_duration_us)
202 		return -EINVAL;
203 
204 	pr_debug("Starting injecting idle cycles on CPUs '%*pbl'\n",
205 		 cpumask_pr_args(to_cpumask(ii_dev->cpumask)));
206 
207 	idle_inject_wakeup(ii_dev);
208 
209 	hrtimer_start(&ii_dev->timer,
210 		      ns_to_ktime((idle_duration_us + run_duration_us) *
211 				  NSEC_PER_USEC),
212 		      HRTIMER_MODE_REL);
213 
214 	return 0;
215 }
216 
217 /**
218  * idle_inject_stop - stops idle injections
219  * @ii_dev: idle injection control device structure
220  *
221  * The function stops idle injection and waits for the threads to finish work.
222  * If CPU idle time is being injected when this function runs, then it will
223  * wait until the end of the cycle.
224  *
225  * When it returns, there is no more idle injection kthread activity.  The
226  * kthreads are scheduled out and the periodic timer is off.
227  */
idle_inject_stop(struct idle_inject_device * ii_dev)228 void idle_inject_stop(struct idle_inject_device *ii_dev)
229 {
230 	struct idle_inject_thread *iit;
231 	unsigned int cpu;
232 
233 	pr_debug("Stopping idle injection on CPUs '%*pbl'\n",
234 		 cpumask_pr_args(to_cpumask(ii_dev->cpumask)));
235 
236 	hrtimer_cancel(&ii_dev->timer);
237 
238 	/*
239 	 * Stopping idle injection requires all of the idle injection kthreads
240 	 * associated with the given cpumask to be parked and stay that way, so
241 	 * prevent CPUs from going online at this point.  Any CPUs going online
242 	 * after the loop below will be covered by clearing the should_run flag
243 	 * that will cause the smpboot main loop to schedule them out.
244 	 */
245 	cpu_hotplug_disable();
246 
247 	/*
248 	 * Iterate over all (online + offline) CPUs here in case one of them
249 	 * goes offline with the should_run flag set so as to prevent its idle
250 	 * injection kthread from running when the CPU goes online again after
251 	 * the ii_dev has been freed.
252 	 */
253 	for_each_cpu(cpu, to_cpumask(ii_dev->cpumask)) {
254 		iit = per_cpu_ptr(&idle_inject_thread, cpu);
255 		iit->should_run = 0;
256 
257 		wait_task_inactive(iit->tsk, 0);
258 	}
259 
260 	cpu_hotplug_enable();
261 }
262 
263 /**
264  * idle_inject_setup - prepare the current task for idle injection
265  * @cpu: not used
266  *
267  * Called once, this function is in charge of setting the current task's
268  * scheduler parameters to make it an RT task.
269  */
idle_inject_setup(unsigned int cpu)270 static void idle_inject_setup(unsigned int cpu)
271 {
272 	sched_set_fifo(current);
273 }
274 
275 /**
276  * idle_inject_should_run - function helper for the smpboot API
277  * @cpu: CPU the kthread is running on
278  *
279  * Return: whether or not the thread can run.
280  */
idle_inject_should_run(unsigned int cpu)281 static int idle_inject_should_run(unsigned int cpu)
282 {
283 	struct idle_inject_thread *iit =
284 		per_cpu_ptr(&idle_inject_thread, cpu);
285 
286 	return iit->should_run;
287 }
288 
289 /**
290  * idle_inject_register - initialize idle injection on a set of CPUs
291  * @cpumask: CPUs to be affected by idle injection
292  *
293  * This function creates an idle injection control device structure for the
294  * given set of CPUs and initializes the timer associated with it.  It does not
295  * start any injection cycles.
296  *
297  * Return: NULL if memory allocation fails, idle injection control device
298  * pointer on success.
299  */
idle_inject_register(struct cpumask * cpumask)300 struct idle_inject_device *idle_inject_register(struct cpumask *cpumask)
301 {
302 	struct idle_inject_device *ii_dev;
303 	int cpu, cpu_rb;
304 
305 	ii_dev = kzalloc(sizeof(*ii_dev) + cpumask_size(), GFP_KERNEL);
306 	if (!ii_dev)
307 		return NULL;
308 
309 	cpumask_copy(to_cpumask(ii_dev->cpumask), cpumask);
310 	hrtimer_init(&ii_dev->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
311 	ii_dev->timer.function = idle_inject_timer_fn;
312 	ii_dev->latency_us = UINT_MAX;
313 
314 	for_each_cpu(cpu, to_cpumask(ii_dev->cpumask)) {
315 
316 		if (per_cpu(idle_inject_device, cpu)) {
317 			pr_err("cpu%d is already registered\n", cpu);
318 			goto out_rollback;
319 		}
320 
321 		per_cpu(idle_inject_device, cpu) = ii_dev;
322 	}
323 
324 	return ii_dev;
325 
326 out_rollback:
327 	for_each_cpu(cpu_rb, to_cpumask(ii_dev->cpumask)) {
328 		if (cpu == cpu_rb)
329 			break;
330 		per_cpu(idle_inject_device, cpu_rb) = NULL;
331 	}
332 
333 	kfree(ii_dev);
334 
335 	return NULL;
336 }
337 
338 /**
339  * idle_inject_unregister - unregister idle injection control device
340  * @ii_dev: idle injection control device to unregister
341  *
342  * The function stops idle injection for the given control device,
343  * unregisters its kthreads and frees memory allocated when that device was
344  * created.
345  */
idle_inject_unregister(struct idle_inject_device * ii_dev)346 void idle_inject_unregister(struct idle_inject_device *ii_dev)
347 {
348 	unsigned int cpu;
349 
350 	idle_inject_stop(ii_dev);
351 
352 	for_each_cpu(cpu, to_cpumask(ii_dev->cpumask))
353 		per_cpu(idle_inject_device, cpu) = NULL;
354 
355 	kfree(ii_dev);
356 }
357 
358 static struct smp_hotplug_thread idle_inject_threads = {
359 	.store = &idle_inject_thread.tsk,
360 	.setup = idle_inject_setup,
361 	.thread_fn = idle_inject_fn,
362 	.thread_comm = "idle_inject/%u",
363 	.thread_should_run = idle_inject_should_run,
364 };
365 
idle_inject_init(void)366 static int __init idle_inject_init(void)
367 {
368 	return smpboot_register_percpu_thread(&idle_inject_threads);
369 }
370 early_initcall(idle_inject_init);
371