xref: /openbmc/linux/kernel/cgroup/rstat.c (revision d0e22329)
1 #include "cgroup-internal.h"
2 
3 #include <linux/sched/cputime.h>
4 
5 static DEFINE_SPINLOCK(cgroup_rstat_lock);
6 static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
7 
8 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
9 
10 static struct cgroup_rstat_cpu *cgroup_rstat_cpu(struct cgroup *cgrp, int cpu)
11 {
12 	return per_cpu_ptr(cgrp->rstat_cpu, cpu);
13 }
14 
15 /**
16  * cgroup_rstat_updated - keep track of updated rstat_cpu
17  * @cgrp: target cgroup
18  * @cpu: cpu on which rstat_cpu was updated
19  *
20  * @cgrp's rstat_cpu on @cpu was updated.  Put it on the parent's matching
21  * rstat_cpu->updated_children list.  See the comment on top of
22  * cgroup_rstat_cpu definition for details.
23  */
24 void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
25 {
26 	raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
27 	struct cgroup *parent;
28 	unsigned long flags;
29 
30 	/* nothing to do for root */
31 	if (!cgroup_parent(cgrp))
32 		return;
33 
34 	/*
35 	 * Paired with the one in cgroup_rstat_cpu_pop_upated().  Either we
36 	 * see NULL updated_next or they see our updated stat.
37 	 */
38 	smp_mb();
39 
40 	/*
41 	 * Because @parent's updated_children is terminated with @parent
42 	 * instead of NULL, we can tell whether @cgrp is on the list by
43 	 * testing the next pointer for NULL.
44 	 */
45 	if (cgroup_rstat_cpu(cgrp, cpu)->updated_next)
46 		return;
47 
48 	raw_spin_lock_irqsave(cpu_lock, flags);
49 
50 	/* put @cgrp and all ancestors on the corresponding updated lists */
51 	for (parent = cgroup_parent(cgrp); parent;
52 	     cgrp = parent, parent = cgroup_parent(cgrp)) {
53 		struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
54 		struct cgroup_rstat_cpu *prstatc = cgroup_rstat_cpu(parent, cpu);
55 
56 		/*
57 		 * Both additions and removals are bottom-up.  If a cgroup
58 		 * is already in the tree, all ancestors are.
59 		 */
60 		if (rstatc->updated_next)
61 			break;
62 
63 		rstatc->updated_next = prstatc->updated_children;
64 		prstatc->updated_children = cgrp;
65 	}
66 
67 	raw_spin_unlock_irqrestore(cpu_lock, flags);
68 }
69 EXPORT_SYMBOL_GPL(cgroup_rstat_updated);
70 
71 /**
72  * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
73  * @pos: current position
74  * @root: root of the tree to traversal
75  * @cpu: target cpu
76  *
77  * Walks the udpated rstat_cpu tree on @cpu from @root.  %NULL @pos starts
78  * the traversal and %NULL return indicates the end.  During traversal,
79  * each returned cgroup is unlinked from the tree.  Must be called with the
80  * matching cgroup_rstat_cpu_lock held.
81  *
82  * The only ordering guarantee is that, for a parent and a child pair
83  * covered by a given traversal, if a child is visited, its parent is
84  * guaranteed to be visited afterwards.
85  */
86 static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
87 						   struct cgroup *root, int cpu)
88 {
89 	struct cgroup_rstat_cpu *rstatc;
90 	struct cgroup *parent;
91 
92 	if (pos == root)
93 		return NULL;
94 
95 	/*
96 	 * We're gonna walk down to the first leaf and visit/remove it.  We
97 	 * can pick whatever unvisited node as the starting point.
98 	 */
99 	if (!pos)
100 		pos = root;
101 	else
102 		pos = cgroup_parent(pos);
103 
104 	/* walk down to the first leaf */
105 	while (true) {
106 		rstatc = cgroup_rstat_cpu(pos, cpu);
107 		if (rstatc->updated_children == pos)
108 			break;
109 		pos = rstatc->updated_children;
110 	}
111 
112 	/*
113 	 * Unlink @pos from the tree.  As the updated_children list is
114 	 * singly linked, we have to walk it to find the removal point.
115 	 * However, due to the way we traverse, @pos will be the first
116 	 * child in most cases. The only exception is @root.
117 	 */
118 	parent = cgroup_parent(pos);
119 	if (parent && rstatc->updated_next) {
120 		struct cgroup_rstat_cpu *prstatc = cgroup_rstat_cpu(parent, cpu);
121 		struct cgroup_rstat_cpu *nrstatc;
122 		struct cgroup **nextp;
123 
124 		nextp = &prstatc->updated_children;
125 		while (true) {
126 			nrstatc = cgroup_rstat_cpu(*nextp, cpu);
127 			if (*nextp == pos)
128 				break;
129 
130 			WARN_ON_ONCE(*nextp == parent);
131 			nextp = &nrstatc->updated_next;
132 		}
133 
134 		*nextp = rstatc->updated_next;
135 		rstatc->updated_next = NULL;
136 
137 		/*
138 		 * Paired with the one in cgroup_rstat_cpu_updated().
139 		 * Either they see NULL updated_next or we see their
140 		 * updated stat.
141 		 */
142 		smp_mb();
143 	}
144 
145 	return pos;
146 }
147 
148 /* see cgroup_rstat_flush() */
149 static void cgroup_rstat_flush_locked(struct cgroup *cgrp, bool may_sleep)
150 	__releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock)
151 {
152 	int cpu;
153 
154 	lockdep_assert_held(&cgroup_rstat_lock);
155 
156 	for_each_possible_cpu(cpu) {
157 		raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock,
158 						       cpu);
159 		struct cgroup *pos = NULL;
160 
161 		raw_spin_lock(cpu_lock);
162 		while ((pos = cgroup_rstat_cpu_pop_updated(pos, cgrp, cpu))) {
163 			struct cgroup_subsys_state *css;
164 
165 			cgroup_base_stat_flush(pos, cpu);
166 
167 			rcu_read_lock();
168 			list_for_each_entry_rcu(css, &pos->rstat_css_list,
169 						rstat_css_node)
170 				css->ss->css_rstat_flush(css, cpu);
171 			rcu_read_unlock();
172 		}
173 		raw_spin_unlock(cpu_lock);
174 
175 		/* if @may_sleep, play nice and yield if necessary */
176 		if (may_sleep && (need_resched() ||
177 				  spin_needbreak(&cgroup_rstat_lock))) {
178 			spin_unlock_irq(&cgroup_rstat_lock);
179 			if (!cond_resched())
180 				cpu_relax();
181 			spin_lock_irq(&cgroup_rstat_lock);
182 		}
183 	}
184 }
185 
186 /**
187  * cgroup_rstat_flush - flush stats in @cgrp's subtree
188  * @cgrp: target cgroup
189  *
190  * Collect all per-cpu stats in @cgrp's subtree into the global counters
191  * and propagate them upwards.  After this function returns, all cgroups in
192  * the subtree have up-to-date ->stat.
193  *
194  * This also gets all cgroups in the subtree including @cgrp off the
195  * ->updated_children lists.
196  *
197  * This function may block.
198  */
199 void cgroup_rstat_flush(struct cgroup *cgrp)
200 {
201 	might_sleep();
202 
203 	spin_lock_irq(&cgroup_rstat_lock);
204 	cgroup_rstat_flush_locked(cgrp, true);
205 	spin_unlock_irq(&cgroup_rstat_lock);
206 }
207 
208 /**
209  * cgroup_rstat_flush_irqsafe - irqsafe version of cgroup_rstat_flush()
210  * @cgrp: target cgroup
211  *
212  * This function can be called from any context.
213  */
214 void cgroup_rstat_flush_irqsafe(struct cgroup *cgrp)
215 {
216 	unsigned long flags;
217 
218 	spin_lock_irqsave(&cgroup_rstat_lock, flags);
219 	cgroup_rstat_flush_locked(cgrp, false);
220 	spin_unlock_irqrestore(&cgroup_rstat_lock, flags);
221 }
222 
223 /**
224  * cgroup_rstat_flush_begin - flush stats in @cgrp's subtree and hold
225  * @cgrp: target cgroup
226  *
227  * Flush stats in @cgrp's subtree and prevent further flushes.  Must be
228  * paired with cgroup_rstat_flush_release().
229  *
230  * This function may block.
231  */
232 void cgroup_rstat_flush_hold(struct cgroup *cgrp)
233 	__acquires(&cgroup_rstat_lock)
234 {
235 	might_sleep();
236 	spin_lock_irq(&cgroup_rstat_lock);
237 	cgroup_rstat_flush_locked(cgrp, true);
238 }
239 
240 /**
241  * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
242  */
243 void cgroup_rstat_flush_release(void)
244 	__releases(&cgroup_rstat_lock)
245 {
246 	spin_unlock_irq(&cgroup_rstat_lock);
247 }
248 
249 int cgroup_rstat_init(struct cgroup *cgrp)
250 {
251 	int cpu;
252 
253 	/* the root cgrp has rstat_cpu preallocated */
254 	if (!cgrp->rstat_cpu) {
255 		cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
256 		if (!cgrp->rstat_cpu)
257 			return -ENOMEM;
258 	}
259 
260 	/* ->updated_children list is self terminated */
261 	for_each_possible_cpu(cpu) {
262 		struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
263 
264 		rstatc->updated_children = cgrp;
265 		u64_stats_init(&rstatc->bsync);
266 	}
267 
268 	return 0;
269 }
270 
271 void cgroup_rstat_exit(struct cgroup *cgrp)
272 {
273 	int cpu;
274 
275 	cgroup_rstat_flush(cgrp);
276 
277 	/* sanity check */
278 	for_each_possible_cpu(cpu) {
279 		struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
280 
281 		if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
282 		    WARN_ON_ONCE(rstatc->updated_next))
283 			return;
284 	}
285 
286 	free_percpu(cgrp->rstat_cpu);
287 	cgrp->rstat_cpu = NULL;
288 }
289 
290 void __init cgroup_rstat_boot(void)
291 {
292 	int cpu;
293 
294 	for_each_possible_cpu(cpu)
295 		raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
296 
297 	BUG_ON(cgroup_rstat_init(&cgrp_dfl_root.cgrp));
298 }
299 
300 /*
301  * Functions for cgroup basic resource statistics implemented on top of
302  * rstat.
303  */
304 static void cgroup_base_stat_accumulate(struct cgroup_base_stat *dst_bstat,
305 					struct cgroup_base_stat *src_bstat)
306 {
307 	dst_bstat->cputime.utime += src_bstat->cputime.utime;
308 	dst_bstat->cputime.stime += src_bstat->cputime.stime;
309 	dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
310 }
311 
312 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
313 {
314 	struct cgroup *parent = cgroup_parent(cgrp);
315 	struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
316 	struct task_cputime *last_cputime = &rstatc->last_bstat.cputime;
317 	struct task_cputime cputime;
318 	struct cgroup_base_stat delta;
319 	unsigned seq;
320 
321 	/* fetch the current per-cpu values */
322 	do {
323 		seq = __u64_stats_fetch_begin(&rstatc->bsync);
324 		cputime = rstatc->bstat.cputime;
325 	} while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
326 
327 	/* calculate the delta to propgate */
328 	delta.cputime.utime = cputime.utime - last_cputime->utime;
329 	delta.cputime.stime = cputime.stime - last_cputime->stime;
330 	delta.cputime.sum_exec_runtime = cputime.sum_exec_runtime -
331 					 last_cputime->sum_exec_runtime;
332 	*last_cputime = cputime;
333 
334 	/* transfer the pending stat into delta */
335 	cgroup_base_stat_accumulate(&delta, &cgrp->pending_bstat);
336 	memset(&cgrp->pending_bstat, 0, sizeof(cgrp->pending_bstat));
337 
338 	/* propagate delta into the global stat and the parent's pending */
339 	cgroup_base_stat_accumulate(&cgrp->bstat, &delta);
340 	if (parent)
341 		cgroup_base_stat_accumulate(&parent->pending_bstat, &delta);
342 }
343 
344 static struct cgroup_rstat_cpu *
345 cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp)
346 {
347 	struct cgroup_rstat_cpu *rstatc;
348 
349 	rstatc = get_cpu_ptr(cgrp->rstat_cpu);
350 	u64_stats_update_begin(&rstatc->bsync);
351 	return rstatc;
352 }
353 
354 static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
355 						 struct cgroup_rstat_cpu *rstatc)
356 {
357 	u64_stats_update_end(&rstatc->bsync);
358 	cgroup_rstat_updated(cgrp, smp_processor_id());
359 	put_cpu_ptr(rstatc);
360 }
361 
362 void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
363 {
364 	struct cgroup_rstat_cpu *rstatc;
365 
366 	rstatc = cgroup_base_stat_cputime_account_begin(cgrp);
367 	rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
368 	cgroup_base_stat_cputime_account_end(cgrp, rstatc);
369 }
370 
371 void __cgroup_account_cputime_field(struct cgroup *cgrp,
372 				    enum cpu_usage_stat index, u64 delta_exec)
373 {
374 	struct cgroup_rstat_cpu *rstatc;
375 
376 	rstatc = cgroup_base_stat_cputime_account_begin(cgrp);
377 
378 	switch (index) {
379 	case CPUTIME_USER:
380 	case CPUTIME_NICE:
381 		rstatc->bstat.cputime.utime += delta_exec;
382 		break;
383 	case CPUTIME_SYSTEM:
384 	case CPUTIME_IRQ:
385 	case CPUTIME_SOFTIRQ:
386 		rstatc->bstat.cputime.stime += delta_exec;
387 		break;
388 	default:
389 		break;
390 	}
391 
392 	cgroup_base_stat_cputime_account_end(cgrp, rstatc);
393 }
394 
395 void cgroup_base_stat_cputime_show(struct seq_file *seq)
396 {
397 	struct cgroup *cgrp = seq_css(seq)->cgroup;
398 	u64 usage, utime, stime;
399 
400 	if (!cgroup_parent(cgrp))
401 		return;
402 
403 	cgroup_rstat_flush_hold(cgrp);
404 	usage = cgrp->bstat.cputime.sum_exec_runtime;
405 	cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime, &utime, &stime);
406 	cgroup_rstat_flush_release();
407 
408 	do_div(usage, NSEC_PER_USEC);
409 	do_div(utime, NSEC_PER_USEC);
410 	do_div(stime, NSEC_PER_USEC);
411 
412 	seq_printf(seq, "usage_usec %llu\n"
413 		   "user_usec %llu\n"
414 		   "system_usec %llu\n",
415 		   usage, utime, stime);
416 }
417