1 // SPDX-License-Identifier: GPL-2.0-only
2 #include "cgroup-internal.h"
4 #include <linux/sched/cputime.h>
6 static DEFINE_SPINLOCK(cgroup_rstat_lock);
7 static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
9 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
11 static struct cgroup_rstat_cpu *cgroup_rstat_cpu(struct cgroup *cgrp, int cpu)
13 return per_cpu_ptr(cgrp->rstat_cpu, cpu);
17 * cgroup_rstat_updated - keep track of updated rstat_cpu
18 * @cgrp: target cgroup
19 * @cpu: cpu on which rstat_cpu was updated
21 * @cgrp's rstat_cpu on @cpu was updated. Put it on the parent's matching
22 * rstat_cpu->updated_children list. See the comment on top of
23 * cgroup_rstat_cpu definition for details.
25 void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
27 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
28 struct cgroup *parent;
31 /* nothing to do for root */
32 if (!cgroup_parent(cgrp))
36 * Paired with the one in cgroup_rstat_cpu_pop_updated(). Either we
37 * see NULL updated_next or they see our updated stat.
42 * Because @parent's updated_children is terminated with @parent
43 * instead of NULL, we can tell whether @cgrp is on the list by
44 * testing the next pointer for NULL.
46 if (cgroup_rstat_cpu(cgrp, cpu)->updated_next)
49 raw_spin_lock_irqsave(cpu_lock, flags);
51 /* put @cgrp and all ancestors on the corresponding updated lists */
52 for (parent = cgroup_parent(cgrp); parent;
53 cgrp = parent, parent = cgroup_parent(cgrp)) {
54 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
55 struct cgroup_rstat_cpu *prstatc = cgroup_rstat_cpu(parent, cpu);
58 * Both additions and removals are bottom-up. If a cgroup
59 * is already in the tree, all ancestors are.
61 if (rstatc->updated_next)
64 rstatc->updated_next = prstatc->updated_children;
65 prstatc->updated_children = cgrp;
68 raw_spin_unlock_irqrestore(cpu_lock, flags);
70 EXPORT_SYMBOL_GPL(cgroup_rstat_updated);
73 * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
74 * @pos: current position
75 * @root: root of the tree to traversal
78 * Walks the udpated rstat_cpu tree on @cpu from @root. %NULL @pos starts
79 * the traversal and %NULL return indicates the end. During traversal,
80 * each returned cgroup is unlinked from the tree. Must be called with the
81 * matching cgroup_rstat_cpu_lock held.
83 * The only ordering guarantee is that, for a parent and a child pair
84 * covered by a given traversal, if a child is visited, its parent is
85 * guaranteed to be visited afterwards.
87 static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
88 struct cgroup *root, int cpu)
90 struct cgroup_rstat_cpu *rstatc;
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.
102 pos = cgroup_parent(pos);
104 /* walk down to the first leaf */
106 rstatc = cgroup_rstat_cpu(pos, cpu);
107 if (rstatc->updated_children == pos)
109 pos = rstatc->updated_children;
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.
118 if (rstatc->updated_next) {
119 struct cgroup *parent = cgroup_parent(pos);
120 struct cgroup_rstat_cpu *prstatc = cgroup_rstat_cpu(parent, cpu);
121 struct cgroup_rstat_cpu *nrstatc;
122 struct cgroup **nextp;
124 nextp = &prstatc->updated_children;
126 nrstatc = cgroup_rstat_cpu(*nextp, cpu);
130 WARN_ON_ONCE(*nextp == parent);
131 nextp = &nrstatc->updated_next;
134 *nextp = rstatc->updated_next;
135 rstatc->updated_next = NULL;
138 * Paired with the one in cgroup_rstat_cpu_updated().
139 * Either they see NULL updated_next or we see their
147 /* only happens for @root */
151 /* see cgroup_rstat_flush() */
152 static void cgroup_rstat_flush_locked(struct cgroup *cgrp, bool may_sleep)
153 __releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock)
157 lockdep_assert_held(&cgroup_rstat_lock);
159 for_each_possible_cpu(cpu) {
160 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock,
162 struct cgroup *pos = NULL;
164 raw_spin_lock(cpu_lock);
165 while ((pos = cgroup_rstat_cpu_pop_updated(pos, cgrp, cpu))) {
166 struct cgroup_subsys_state *css;
168 cgroup_base_stat_flush(pos, cpu);
171 list_for_each_entry_rcu(css, &pos->rstat_css_list,
173 css->ss->css_rstat_flush(css, cpu);
176 raw_spin_unlock(cpu_lock);
178 /* if @may_sleep, play nice and yield if necessary */
179 if (may_sleep && (need_resched() ||
180 spin_needbreak(&cgroup_rstat_lock))) {
181 spin_unlock_irq(&cgroup_rstat_lock);
184 spin_lock_irq(&cgroup_rstat_lock);
190 * cgroup_rstat_flush - flush stats in @cgrp's subtree
191 * @cgrp: target cgroup
193 * Collect all per-cpu stats in @cgrp's subtree into the global counters
194 * and propagate them upwards. After this function returns, all cgroups in
195 * the subtree have up-to-date ->stat.
197 * This also gets all cgroups in the subtree including @cgrp off the
198 * ->updated_children lists.
200 * This function may block.
202 void cgroup_rstat_flush(struct cgroup *cgrp)
206 spin_lock_irq(&cgroup_rstat_lock);
207 cgroup_rstat_flush_locked(cgrp, true);
208 spin_unlock_irq(&cgroup_rstat_lock);
212 * cgroup_rstat_flush_irqsafe - irqsafe version of cgroup_rstat_flush()
213 * @cgrp: target cgroup
215 * This function can be called from any context.
217 void cgroup_rstat_flush_irqsafe(struct cgroup *cgrp)
221 spin_lock_irqsave(&cgroup_rstat_lock, flags);
222 cgroup_rstat_flush_locked(cgrp, false);
223 spin_unlock_irqrestore(&cgroup_rstat_lock, flags);
227 * cgroup_rstat_flush_begin - flush stats in @cgrp's subtree and hold
228 * @cgrp: target cgroup
230 * Flush stats in @cgrp's subtree and prevent further flushes. Must be
231 * paired with cgroup_rstat_flush_release().
233 * This function may block.
235 void cgroup_rstat_flush_hold(struct cgroup *cgrp)
236 __acquires(&cgroup_rstat_lock)
239 spin_lock_irq(&cgroup_rstat_lock);
240 cgroup_rstat_flush_locked(cgrp, true);
244 * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
246 void cgroup_rstat_flush_release(void)
247 __releases(&cgroup_rstat_lock)
249 spin_unlock_irq(&cgroup_rstat_lock);
252 int cgroup_rstat_init(struct cgroup *cgrp)
256 /* the root cgrp has rstat_cpu preallocated */
257 if (!cgrp->rstat_cpu) {
258 cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
259 if (!cgrp->rstat_cpu)
263 /* ->updated_children list is self terminated */
264 for_each_possible_cpu(cpu) {
265 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
267 rstatc->updated_children = cgrp;
268 u64_stats_init(&rstatc->bsync);
274 void cgroup_rstat_exit(struct cgroup *cgrp)
278 cgroup_rstat_flush(cgrp);
281 for_each_possible_cpu(cpu) {
282 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
284 if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
285 WARN_ON_ONCE(rstatc->updated_next))
289 free_percpu(cgrp->rstat_cpu);
290 cgrp->rstat_cpu = NULL;
293 void __init cgroup_rstat_boot(void)
297 for_each_possible_cpu(cpu)
298 raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
300 BUG_ON(cgroup_rstat_init(&cgrp_dfl_root.cgrp));
304 * Functions for cgroup basic resource statistics implemented on top of
307 static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
308 struct cgroup_base_stat *src_bstat)
310 dst_bstat->cputime.utime += src_bstat->cputime.utime;
311 dst_bstat->cputime.stime += src_bstat->cputime.stime;
312 dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
315 static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
316 struct cgroup_base_stat *src_bstat)
318 dst_bstat->cputime.utime -= src_bstat->cputime.utime;
319 dst_bstat->cputime.stime -= src_bstat->cputime.stime;
320 dst_bstat->cputime.sum_exec_runtime -= src_bstat->cputime.sum_exec_runtime;
323 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
325 struct cgroup *parent = cgroup_parent(cgrp);
326 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
327 struct cgroup_base_stat cur, delta;
330 /* fetch the current per-cpu values */
332 seq = __u64_stats_fetch_begin(&rstatc->bsync);
333 cur.cputime = rstatc->bstat.cputime;
334 } while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
336 /* propagate percpu delta to global */
338 cgroup_base_stat_sub(&delta, &rstatc->last_bstat);
339 cgroup_base_stat_add(&cgrp->bstat, &delta);
340 cgroup_base_stat_add(&rstatc->last_bstat, &delta);
342 /* propagate global delta to parent */
345 cgroup_base_stat_sub(&delta, &cgrp->last_bstat);
346 cgroup_base_stat_add(&parent->bstat, &delta);
347 cgroup_base_stat_add(&cgrp->last_bstat, &delta);
351 static struct cgroup_rstat_cpu *
352 cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp)
354 struct cgroup_rstat_cpu *rstatc;
356 rstatc = get_cpu_ptr(cgrp->rstat_cpu);
357 u64_stats_update_begin(&rstatc->bsync);
361 static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
362 struct cgroup_rstat_cpu *rstatc)
364 u64_stats_update_end(&rstatc->bsync);
365 cgroup_rstat_updated(cgrp, smp_processor_id());
369 void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
371 struct cgroup_rstat_cpu *rstatc;
373 rstatc = cgroup_base_stat_cputime_account_begin(cgrp);
374 rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
375 cgroup_base_stat_cputime_account_end(cgrp, rstatc);
378 void __cgroup_account_cputime_field(struct cgroup *cgrp,
379 enum cpu_usage_stat index, u64 delta_exec)
381 struct cgroup_rstat_cpu *rstatc;
383 rstatc = cgroup_base_stat_cputime_account_begin(cgrp);
388 rstatc->bstat.cputime.utime += delta_exec;
392 case CPUTIME_SOFTIRQ:
393 rstatc->bstat.cputime.stime += delta_exec;
399 cgroup_base_stat_cputime_account_end(cgrp, rstatc);
402 void cgroup_base_stat_cputime_show(struct seq_file *seq)
404 struct cgroup *cgrp = seq_css(seq)->cgroup;
405 u64 usage, utime, stime;
407 if (!cgroup_parent(cgrp))
410 cgroup_rstat_flush_hold(cgrp);
411 usage = cgrp->bstat.cputime.sum_exec_runtime;
412 cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime, &utime, &stime);
413 cgroup_rstat_flush_release();
415 do_div(usage, NSEC_PER_USEC);
416 do_div(utime, NSEC_PER_USEC);
417 do_div(stime, NSEC_PER_USEC);
419 seq_printf(seq, "usage_usec %llu\n"
421 "system_usec %llu\n",
422 usage, utime, stime);