1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * CPU accounting code for task groups.
  4 *
  5 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
  6 * (balbir@in.ibm.com).
  7 */
  8#include <asm/irq_regs.h>
  9#include "sched.h"
 10
 11/* Time spent by the tasks of the CPU accounting group executing in ... */
 12enum cpuacct_stat_index {
 13	CPUACCT_STAT_USER,	/* ... user mode */
 14	CPUACCT_STAT_SYSTEM,	/* ... kernel mode */
 15
 16	CPUACCT_STAT_NSTATS,
 17};
 18
 19static const char * const cpuacct_stat_desc[] = {
 20	[CPUACCT_STAT_USER] = "user",
 21	[CPUACCT_STAT_SYSTEM] = "system",
 22};
 23
 24struct cpuacct_usage {
 25	u64	usages[CPUACCT_STAT_NSTATS];
 26};
 27
 28/* track CPU usage of a group of tasks and its child groups */
 29struct cpuacct {
 30	struct cgroup_subsys_state	css;
 31	/* cpuusage holds pointer to a u64-type object on every CPU */
 32	struct cpuacct_usage __percpu	*cpuusage;
 33	struct kernel_cpustat __percpu	*cpustat;
 34};
 35
 36static inline struct cpuacct *css_ca(struct cgroup_subsys_state *css)
 37{
 38	return css ? container_of(css, struct cpuacct, css) : NULL;
 39}
 40
 41/* Return CPU accounting group to which this task belongs */
 42static inline struct cpuacct *task_ca(struct task_struct *tsk)
 43{
 44	return css_ca(task_css(tsk, cpuacct_cgrp_id));
 45}
 46
 47static inline struct cpuacct *parent_ca(struct cpuacct *ca)
 48{
 49	return css_ca(ca->css.parent);
 50}
 51
 52static DEFINE_PER_CPU(struct cpuacct_usage, root_cpuacct_cpuusage);
 53static struct cpuacct root_cpuacct = {
 54	.cpustat	= &kernel_cpustat,
 55	.cpuusage	= &root_cpuacct_cpuusage,
 56};
 57
 58/* Create a new CPU accounting group */
 59static struct cgroup_subsys_state *
 60cpuacct_css_alloc(struct cgroup_subsys_state *parent_css)
 61{
 62	struct cpuacct *ca;
 63
 64	if (!parent_css)
 65		return &root_cpuacct.css;
 66
 67	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
 68	if (!ca)
 69		goto out;
 70
 71	ca->cpuusage = alloc_percpu(struct cpuacct_usage);
 72	if (!ca->cpuusage)
 73		goto out_free_ca;
 74
 75	ca->cpustat = alloc_percpu(struct kernel_cpustat);
 76	if (!ca->cpustat)
 77		goto out_free_cpuusage;
 78
 79	return &ca->css;
 80
 81out_free_cpuusage:
 82	free_percpu(ca->cpuusage);
 83out_free_ca:
 84	kfree(ca);
 85out:
 86	return ERR_PTR(-ENOMEM);
 87}
 88
 89/* Destroy an existing CPU accounting group */
 90static void cpuacct_css_free(struct cgroup_subsys_state *css)
 91{
 92	struct cpuacct *ca = css_ca(css);
 93
 94	free_percpu(ca->cpustat);
 95	free_percpu(ca->cpuusage);
 96	kfree(ca);
 97}
 98
 99static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu,
100				 enum cpuacct_stat_index index)
101{
102	struct cpuacct_usage *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
103	u64 data;
104
105	/*
106	 * We allow index == CPUACCT_STAT_NSTATS here to read
107	 * the sum of usages.
108	 */
109	BUG_ON(index > CPUACCT_STAT_NSTATS);
110
111#ifndef CONFIG_64BIT
112	/*
113	 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
114	 */
115	raw_spin_rq_lock_irq(cpu_rq(cpu));
116#endif
117
118	if (index == CPUACCT_STAT_NSTATS) {
119		int i = 0;
120
121		data = 0;
122		for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
123			data += cpuusage->usages[i];
124	} else {
125		data = cpuusage->usages[index];
126	}
127
128#ifndef CONFIG_64BIT
129	raw_spin_rq_unlock_irq(cpu_rq(cpu));
130#endif
131
132	return data;
133}
134
135static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
136{
137	struct cpuacct_usage *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
138	int i;
139
140#ifndef CONFIG_64BIT
141	/*
142	 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
143	 */
144	raw_spin_rq_lock_irq(cpu_rq(cpu));
145#endif
146
147	for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
148		cpuusage->usages[i] = val;
149
150#ifndef CONFIG_64BIT
151	raw_spin_rq_unlock_irq(cpu_rq(cpu));
152#endif
153}
154
155/* Return total CPU usage (in nanoseconds) of a group */
156static u64 __cpuusage_read(struct cgroup_subsys_state *css,
157			   enum cpuacct_stat_index index)
158{
159	struct cpuacct *ca = css_ca(css);
160	u64 totalcpuusage = 0;
161	int i;
162
163	for_each_possible_cpu(i)
164		totalcpuusage += cpuacct_cpuusage_read(ca, i, index);
165
166	return totalcpuusage;
167}
168
169static u64 cpuusage_user_read(struct cgroup_subsys_state *css,
170			      struct cftype *cft)
171{
172	return __cpuusage_read(css, CPUACCT_STAT_USER);
173}
174
175static u64 cpuusage_sys_read(struct cgroup_subsys_state *css,
176			     struct cftype *cft)
177{
178	return __cpuusage_read(css, CPUACCT_STAT_SYSTEM);
179}
180
181static u64 cpuusage_read(struct cgroup_subsys_state *css, struct cftype *cft)
182{
183	return __cpuusage_read(css, CPUACCT_STAT_NSTATS);
184}
185
186static int cpuusage_write(struct cgroup_subsys_state *css, struct cftype *cft,
187			  u64 val)
188{
189	struct cpuacct *ca = css_ca(css);
190	int cpu;
191
192	/*
193	 * Only allow '0' here to do a reset.
194	 */
195	if (val)
196		return -EINVAL;
197
198	for_each_possible_cpu(cpu)
199		cpuacct_cpuusage_write(ca, cpu, 0);
200
201	return 0;
202}
203
204static int __cpuacct_percpu_seq_show(struct seq_file *m,
205				     enum cpuacct_stat_index index)
206{
207	struct cpuacct *ca = css_ca(seq_css(m));
208	u64 percpu;
209	int i;
210
211	for_each_possible_cpu(i) {
212		percpu = cpuacct_cpuusage_read(ca, i, index);
213		seq_printf(m, "%llu ", (unsigned long long) percpu);
214	}
215	seq_printf(m, "\n");
216	return 0;
217}
218
219static int cpuacct_percpu_user_seq_show(struct seq_file *m, void *V)
220{
221	return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_USER);
222}
223
224static int cpuacct_percpu_sys_seq_show(struct seq_file *m, void *V)
225{
226	return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_SYSTEM);
227}
228
229static int cpuacct_percpu_seq_show(struct seq_file *m, void *V)
230{
231	return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_NSTATS);
232}
233
234static int cpuacct_all_seq_show(struct seq_file *m, void *V)
235{
236	struct cpuacct *ca = css_ca(seq_css(m));
237	int index;
238	int cpu;
239
240	seq_puts(m, "cpu");
241	for (index = 0; index < CPUACCT_STAT_NSTATS; index++)
242		seq_printf(m, " %s", cpuacct_stat_desc[index]);
243	seq_puts(m, "\n");
244
245	for_each_possible_cpu(cpu) {
246		struct cpuacct_usage *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
247
248		seq_printf(m, "%d", cpu);
249
250		for (index = 0; index < CPUACCT_STAT_NSTATS; index++) {
251#ifndef CONFIG_64BIT
252			/*
253			 * Take rq->lock to make 64-bit read safe on 32-bit
254			 * platforms.
255			 */
256			raw_spin_rq_lock_irq(cpu_rq(cpu));
257#endif
258
259			seq_printf(m, " %llu", cpuusage->usages[index]);
260
261#ifndef CONFIG_64BIT
262			raw_spin_rq_unlock_irq(cpu_rq(cpu));
263#endif
264		}
265		seq_puts(m, "\n");
266	}
267	return 0;
268}
269
270static int cpuacct_stats_show(struct seq_file *sf, void *v)
271{
272	struct cpuacct *ca = css_ca(seq_css(sf));
273	s64 val[CPUACCT_STAT_NSTATS];
274	int cpu;
275	int stat;
276
277	memset(val, 0, sizeof(val));
278	for_each_possible_cpu(cpu) {
279		u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
280
281		val[CPUACCT_STAT_USER]   += cpustat[CPUTIME_USER];
282		val[CPUACCT_STAT_USER]   += cpustat[CPUTIME_NICE];
283		val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SYSTEM];
284		val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_IRQ];
285		val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SOFTIRQ];
286	}
287
288	for (stat = 0; stat < CPUACCT_STAT_NSTATS; stat++) {
289		seq_printf(sf, "%s %lld\n",
290			   cpuacct_stat_desc[stat],
291			   (long long)nsec_to_clock_t(val[stat]));
292	}
293
294	return 0;
295}
296
297static struct cftype files[] = {
298	{
299		.name = "usage",
300		.read_u64 = cpuusage_read,
301		.write_u64 = cpuusage_write,
302	},
303	{
304		.name = "usage_user",
305		.read_u64 = cpuusage_user_read,
306	},
307	{
308		.name = "usage_sys",
309		.read_u64 = cpuusage_sys_read,
310	},
311	{
312		.name = "usage_percpu",
313		.seq_show = cpuacct_percpu_seq_show,
314	},
315	{
316		.name = "usage_percpu_user",
317		.seq_show = cpuacct_percpu_user_seq_show,
318	},
319	{
320		.name = "usage_percpu_sys",
321		.seq_show = cpuacct_percpu_sys_seq_show,
322	},
323	{
324		.name = "usage_all",
325		.seq_show = cpuacct_all_seq_show,
326	},
327	{
328		.name = "stat",
329		.seq_show = cpuacct_stats_show,
330	},
331	{ }	/* terminate */
332};
333
334/*
335 * charge this task's execution time to its accounting group.
336 *
337 * called with rq->lock held.
338 */
339void cpuacct_charge(struct task_struct *tsk, u64 cputime)
340{
341	struct cpuacct *ca;
342	int index = CPUACCT_STAT_SYSTEM;
343	struct pt_regs *regs = get_irq_regs() ? : task_pt_regs(tsk);
344
345	if (regs && user_mode(regs))
346		index = CPUACCT_STAT_USER;
347
348	rcu_read_lock();
349
350	for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
351		__this_cpu_add(ca->cpuusage->usages[index], cputime);
352
353	rcu_read_unlock();
354}
355
356/*
357 * Add user/system time to cpuacct.
358 *
359 * Note: it's the caller that updates the account of the root cgroup.
360 */
361void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
362{
363	struct cpuacct *ca;
364
365	rcu_read_lock();
366	for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
367		__this_cpu_add(ca->cpustat->cpustat[index], val);
368	rcu_read_unlock();
369}
370
371struct cgroup_subsys cpuacct_cgrp_subsys = {
372	.css_alloc	= cpuacct_css_alloc,
373	.css_free	= cpuacct_css_free,
374	.legacy_cftypes	= files,
375	.early_init	= true,
376};