Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2#include "cgroup-internal.h"
3
4#include <linux/sched/cputime.h>
5
6#include <linux/bpf.h>
7#include <linux/btf.h>
8#include <linux/btf_ids.h>
9
10static DEFINE_SPINLOCK(cgroup_rstat_lock);
11static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
12
13static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
14
15static struct cgroup_rstat_cpu *cgroup_rstat_cpu(struct cgroup *cgrp, int cpu)
16{
17 return per_cpu_ptr(cgrp->rstat_cpu, cpu);
18}
19
20/**
21 * cgroup_rstat_updated - keep track of updated rstat_cpu
22 * @cgrp: target cgroup
23 * @cpu: cpu on which rstat_cpu was updated
24 *
25 * @cgrp's rstat_cpu on @cpu was updated. Put it on the parent's matching
26 * rstat_cpu->updated_children list. See the comment on top of
27 * cgroup_rstat_cpu definition for details.
28 */
29void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
30{
31 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
32 unsigned long flags;
33
34 /*
35 * Speculative already-on-list test. This may race leading to
36 * temporary inaccuracies, which is fine.
37 *
38 * Because @parent's updated_children is terminated with @parent
39 * instead of NULL, we can tell whether @cgrp is on the list by
40 * testing the next pointer for NULL.
41 */
42 if (data_race(cgroup_rstat_cpu(cgrp, cpu)->updated_next))
43 return;
44
45 raw_spin_lock_irqsave(cpu_lock, flags);
46
47 /* put @cgrp and all ancestors on the corresponding updated lists */
48 while (true) {
49 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
50 struct cgroup *parent = cgroup_parent(cgrp);
51 struct cgroup_rstat_cpu *prstatc;
52
53 /*
54 * Both additions and removals are bottom-up. If a cgroup
55 * is already in the tree, all ancestors are.
56 */
57 if (rstatc->updated_next)
58 break;
59
60 /* Root has no parent to link it to, but mark it busy */
61 if (!parent) {
62 rstatc->updated_next = cgrp;
63 break;
64 }
65
66 prstatc = cgroup_rstat_cpu(parent, cpu);
67 rstatc->updated_next = prstatc->updated_children;
68 prstatc->updated_children = cgrp;
69
70 cgrp = parent;
71 }
72
73 raw_spin_unlock_irqrestore(cpu_lock, flags);
74}
75
76/**
77 * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
78 * @pos: current position
79 * @root: root of the tree to traversal
80 * @cpu: target cpu
81 *
82 * Walks the updated rstat_cpu tree on @cpu from @root. %NULL @pos starts
83 * the traversal and %NULL return indicates the end. During traversal,
84 * each returned cgroup is unlinked from the tree. Must be called with the
85 * matching cgroup_rstat_cpu_lock held.
86 *
87 * The only ordering guarantee is that, for a parent and a child pair
88 * covered by a given traversal, if a child is visited, its parent is
89 * guaranteed to be visited afterwards.
90 */
91static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
92 struct cgroup *root, int cpu)
93{
94 struct cgroup_rstat_cpu *rstatc;
95 struct cgroup *parent;
96
97 if (pos == root)
98 return NULL;
99
100 /*
101 * We're gonna walk down to the first leaf and visit/remove it. We
102 * can pick whatever unvisited node as the starting point.
103 */
104 if (!pos) {
105 pos = root;
106 /* return NULL if this subtree is not on-list */
107 if (!cgroup_rstat_cpu(pos, cpu)->updated_next)
108 return NULL;
109 } else {
110 pos = cgroup_parent(pos);
111 }
112
113 /* walk down to the first leaf */
114 while (true) {
115 rstatc = cgroup_rstat_cpu(pos, cpu);
116 if (rstatc->updated_children == pos)
117 break;
118 pos = rstatc->updated_children;
119 }
120
121 /*
122 * Unlink @pos from the tree. As the updated_children list is
123 * singly linked, we have to walk it to find the removal point.
124 * However, due to the way we traverse, @pos will be the first
125 * child in most cases. The only exception is @root.
126 */
127 parent = cgroup_parent(pos);
128 if (parent) {
129 struct cgroup_rstat_cpu *prstatc;
130 struct cgroup **nextp;
131
132 prstatc = cgroup_rstat_cpu(parent, cpu);
133 nextp = &prstatc->updated_children;
134 while (*nextp != pos) {
135 struct cgroup_rstat_cpu *nrstatc;
136
137 nrstatc = cgroup_rstat_cpu(*nextp, cpu);
138 WARN_ON_ONCE(*nextp == parent);
139 nextp = &nrstatc->updated_next;
140 }
141 *nextp = rstatc->updated_next;
142 }
143
144 rstatc->updated_next = NULL;
145 return pos;
146}
147
148/*
149 * A hook for bpf stat collectors to attach to and flush their stats.
150 * Together with providing bpf kfuncs for cgroup_rstat_updated() and
151 * cgroup_rstat_flush(), this enables a complete workflow where bpf progs that
152 * collect cgroup stats can integrate with rstat for efficient flushing.
153 *
154 * A static noinline declaration here could cause the compiler to optimize away
155 * the function. A global noinline declaration will keep the definition, but may
156 * optimize away the callsite. Therefore, __weak is needed to ensure that the
157 * call is still emitted, by telling the compiler that we don't know what the
158 * function might eventually be.
159 *
160 * __diag_* below are needed to dismiss the missing prototype warning.
161 */
162__diag_push();
163__diag_ignore_all("-Wmissing-prototypes",
164 "kfuncs which will be used in BPF programs");
165
166__weak noinline void bpf_rstat_flush(struct cgroup *cgrp,
167 struct cgroup *parent, int cpu)
168{
169}
170
171__diag_pop();
172
173/* see cgroup_rstat_flush() */
174static void cgroup_rstat_flush_locked(struct cgroup *cgrp, bool may_sleep)
175 __releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock)
176{
177 int cpu;
178
179 lockdep_assert_held(&cgroup_rstat_lock);
180
181 for_each_possible_cpu(cpu) {
182 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock,
183 cpu);
184 struct cgroup *pos = NULL;
185 unsigned long flags;
186
187 /*
188 * The _irqsave() is needed because cgroup_rstat_lock is
189 * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
190 * this lock with the _irq() suffix only disables interrupts on
191 * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
192 * interrupts on both configurations. The _irqsave() ensures
193 * that interrupts are always disabled and later restored.
194 */
195 raw_spin_lock_irqsave(cpu_lock, flags);
196 while ((pos = cgroup_rstat_cpu_pop_updated(pos, cgrp, cpu))) {
197 struct cgroup_subsys_state *css;
198
199 cgroup_base_stat_flush(pos, cpu);
200 bpf_rstat_flush(pos, cgroup_parent(pos), cpu);
201
202 rcu_read_lock();
203 list_for_each_entry_rcu(css, &pos->rstat_css_list,
204 rstat_css_node)
205 css->ss->css_rstat_flush(css, cpu);
206 rcu_read_unlock();
207 }
208 raw_spin_unlock_irqrestore(cpu_lock, flags);
209
210 /* if @may_sleep, play nice and yield if necessary */
211 if (may_sleep && (need_resched() ||
212 spin_needbreak(&cgroup_rstat_lock))) {
213 spin_unlock_irq(&cgroup_rstat_lock);
214 if (!cond_resched())
215 cpu_relax();
216 spin_lock_irq(&cgroup_rstat_lock);
217 }
218 }
219}
220
221/**
222 * cgroup_rstat_flush - flush stats in @cgrp's subtree
223 * @cgrp: target cgroup
224 *
225 * Collect all per-cpu stats in @cgrp's subtree into the global counters
226 * and propagate them upwards. After this function returns, all cgroups in
227 * the subtree have up-to-date ->stat.
228 *
229 * This also gets all cgroups in the subtree including @cgrp off the
230 * ->updated_children lists.
231 *
232 * This function may block.
233 */
234void cgroup_rstat_flush(struct cgroup *cgrp)
235{
236 might_sleep();
237
238 spin_lock_irq(&cgroup_rstat_lock);
239 cgroup_rstat_flush_locked(cgrp, true);
240 spin_unlock_irq(&cgroup_rstat_lock);
241}
242
243/**
244 * cgroup_rstat_flush_irqsafe - irqsafe version of cgroup_rstat_flush()
245 * @cgrp: target cgroup
246 *
247 * This function can be called from any context.
248 */
249void cgroup_rstat_flush_irqsafe(struct cgroup *cgrp)
250{
251 unsigned long flags;
252
253 spin_lock_irqsave(&cgroup_rstat_lock, flags);
254 cgroup_rstat_flush_locked(cgrp, false);
255 spin_unlock_irqrestore(&cgroup_rstat_lock, flags);
256}
257
258/**
259 * cgroup_rstat_flush_hold - flush stats in @cgrp's subtree and hold
260 * @cgrp: target cgroup
261 *
262 * Flush stats in @cgrp's subtree and prevent further flushes. Must be
263 * paired with cgroup_rstat_flush_release().
264 *
265 * This function may block.
266 */
267void cgroup_rstat_flush_hold(struct cgroup *cgrp)
268 __acquires(&cgroup_rstat_lock)
269{
270 might_sleep();
271 spin_lock_irq(&cgroup_rstat_lock);
272 cgroup_rstat_flush_locked(cgrp, true);
273}
274
275/**
276 * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
277 */
278void cgroup_rstat_flush_release(void)
279 __releases(&cgroup_rstat_lock)
280{
281 spin_unlock_irq(&cgroup_rstat_lock);
282}
283
284int cgroup_rstat_init(struct cgroup *cgrp)
285{
286 int cpu;
287
288 /* the root cgrp has rstat_cpu preallocated */
289 if (!cgrp->rstat_cpu) {
290 cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
291 if (!cgrp->rstat_cpu)
292 return -ENOMEM;
293 }
294
295 /* ->updated_children list is self terminated */
296 for_each_possible_cpu(cpu) {
297 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
298
299 rstatc->updated_children = cgrp;
300 u64_stats_init(&rstatc->bsync);
301 }
302
303 return 0;
304}
305
306void cgroup_rstat_exit(struct cgroup *cgrp)
307{
308 int cpu;
309
310 cgroup_rstat_flush(cgrp);
311
312 /* sanity check */
313 for_each_possible_cpu(cpu) {
314 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
315
316 if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
317 WARN_ON_ONCE(rstatc->updated_next))
318 return;
319 }
320
321 free_percpu(cgrp->rstat_cpu);
322 cgrp->rstat_cpu = NULL;
323}
324
325void __init cgroup_rstat_boot(void)
326{
327 int cpu;
328
329 for_each_possible_cpu(cpu)
330 raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
331}
332
333/*
334 * Functions for cgroup basic resource statistics implemented on top of
335 * rstat.
336 */
337static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
338 struct cgroup_base_stat *src_bstat)
339{
340 dst_bstat->cputime.utime += src_bstat->cputime.utime;
341 dst_bstat->cputime.stime += src_bstat->cputime.stime;
342 dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
343#ifdef CONFIG_SCHED_CORE
344 dst_bstat->forceidle_sum += src_bstat->forceidle_sum;
345#endif
346}
347
348static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
349 struct cgroup_base_stat *src_bstat)
350{
351 dst_bstat->cputime.utime -= src_bstat->cputime.utime;
352 dst_bstat->cputime.stime -= src_bstat->cputime.stime;
353 dst_bstat->cputime.sum_exec_runtime -= src_bstat->cputime.sum_exec_runtime;
354#ifdef CONFIG_SCHED_CORE
355 dst_bstat->forceidle_sum -= src_bstat->forceidle_sum;
356#endif
357}
358
359static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
360{
361 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
362 struct cgroup *parent = cgroup_parent(cgrp);
363 struct cgroup_base_stat delta;
364 unsigned seq;
365
366 /* Root-level stats are sourced from system-wide CPU stats */
367 if (!parent)
368 return;
369
370 /* fetch the current per-cpu values */
371 do {
372 seq = __u64_stats_fetch_begin(&rstatc->bsync);
373 delta = rstatc->bstat;
374 } while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
375
376 /* propagate percpu delta to global */
377 cgroup_base_stat_sub(&delta, &rstatc->last_bstat);
378 cgroup_base_stat_add(&cgrp->bstat, &delta);
379 cgroup_base_stat_add(&rstatc->last_bstat, &delta);
380
381 /* propagate global delta to parent (unless that's root) */
382 if (cgroup_parent(parent)) {
383 delta = cgrp->bstat;
384 cgroup_base_stat_sub(&delta, &cgrp->last_bstat);
385 cgroup_base_stat_add(&parent->bstat, &delta);
386 cgroup_base_stat_add(&cgrp->last_bstat, &delta);
387 }
388}
389
390static struct cgroup_rstat_cpu *
391cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp, unsigned long *flags)
392{
393 struct cgroup_rstat_cpu *rstatc;
394
395 rstatc = get_cpu_ptr(cgrp->rstat_cpu);
396 *flags = u64_stats_update_begin_irqsave(&rstatc->bsync);
397 return rstatc;
398}
399
400static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
401 struct cgroup_rstat_cpu *rstatc,
402 unsigned long flags)
403{
404 u64_stats_update_end_irqrestore(&rstatc->bsync, flags);
405 cgroup_rstat_updated(cgrp, smp_processor_id());
406 put_cpu_ptr(rstatc);
407}
408
409void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
410{
411 struct cgroup_rstat_cpu *rstatc;
412 unsigned long flags;
413
414 rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
415 rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
416 cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
417}
418
419void __cgroup_account_cputime_field(struct cgroup *cgrp,
420 enum cpu_usage_stat index, u64 delta_exec)
421{
422 struct cgroup_rstat_cpu *rstatc;
423 unsigned long flags;
424
425 rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
426
427 switch (index) {
428 case CPUTIME_USER:
429 case CPUTIME_NICE:
430 rstatc->bstat.cputime.utime += delta_exec;
431 break;
432 case CPUTIME_SYSTEM:
433 case CPUTIME_IRQ:
434 case CPUTIME_SOFTIRQ:
435 rstatc->bstat.cputime.stime += delta_exec;
436 break;
437#ifdef CONFIG_SCHED_CORE
438 case CPUTIME_FORCEIDLE:
439 rstatc->bstat.forceidle_sum += delta_exec;
440 break;
441#endif
442 default:
443 break;
444 }
445
446 cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
447}
448
449/*
450 * compute the cputime for the root cgroup by getting the per cpu data
451 * at a global level, then categorizing the fields in a manner consistent
452 * with how it is done by __cgroup_account_cputime_field for each bit of
453 * cpu time attributed to a cgroup.
454 */
455static void root_cgroup_cputime(struct cgroup_base_stat *bstat)
456{
457 struct task_cputime *cputime = &bstat->cputime;
458 int i;
459
460 cputime->stime = 0;
461 cputime->utime = 0;
462 cputime->sum_exec_runtime = 0;
463 for_each_possible_cpu(i) {
464 struct kernel_cpustat kcpustat;
465 u64 *cpustat = kcpustat.cpustat;
466 u64 user = 0;
467 u64 sys = 0;
468
469 kcpustat_cpu_fetch(&kcpustat, i);
470
471 user += cpustat[CPUTIME_USER];
472 user += cpustat[CPUTIME_NICE];
473 cputime->utime += user;
474
475 sys += cpustat[CPUTIME_SYSTEM];
476 sys += cpustat[CPUTIME_IRQ];
477 sys += cpustat[CPUTIME_SOFTIRQ];
478 cputime->stime += sys;
479
480 cputime->sum_exec_runtime += user;
481 cputime->sum_exec_runtime += sys;
482 cputime->sum_exec_runtime += cpustat[CPUTIME_STEAL];
483
484#ifdef CONFIG_SCHED_CORE
485 bstat->forceidle_sum += cpustat[CPUTIME_FORCEIDLE];
486#endif
487 }
488}
489
490void cgroup_base_stat_cputime_show(struct seq_file *seq)
491{
492 struct cgroup *cgrp = seq_css(seq)->cgroup;
493 u64 usage, utime, stime;
494 struct cgroup_base_stat bstat;
495#ifdef CONFIG_SCHED_CORE
496 u64 forceidle_time;
497#endif
498
499 if (cgroup_parent(cgrp)) {
500 cgroup_rstat_flush_hold(cgrp);
501 usage = cgrp->bstat.cputime.sum_exec_runtime;
502 cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
503 &utime, &stime);
504#ifdef CONFIG_SCHED_CORE
505 forceidle_time = cgrp->bstat.forceidle_sum;
506#endif
507 cgroup_rstat_flush_release();
508 } else {
509 root_cgroup_cputime(&bstat);
510 usage = bstat.cputime.sum_exec_runtime;
511 utime = bstat.cputime.utime;
512 stime = bstat.cputime.stime;
513#ifdef CONFIG_SCHED_CORE
514 forceidle_time = bstat.forceidle_sum;
515#endif
516 }
517
518 do_div(usage, NSEC_PER_USEC);
519 do_div(utime, NSEC_PER_USEC);
520 do_div(stime, NSEC_PER_USEC);
521#ifdef CONFIG_SCHED_CORE
522 do_div(forceidle_time, NSEC_PER_USEC);
523#endif
524
525 seq_printf(seq, "usage_usec %llu\n"
526 "user_usec %llu\n"
527 "system_usec %llu\n",
528 usage, utime, stime);
529
530#ifdef CONFIG_SCHED_CORE
531 seq_printf(seq, "core_sched.force_idle_usec %llu\n", forceidle_time);
532#endif
533}
534
535/* Add bpf kfuncs for cgroup_rstat_updated() and cgroup_rstat_flush() */
536BTF_SET8_START(bpf_rstat_kfunc_ids)
537BTF_ID_FLAGS(func, cgroup_rstat_updated)
538BTF_ID_FLAGS(func, cgroup_rstat_flush, KF_SLEEPABLE)
539BTF_SET8_END(bpf_rstat_kfunc_ids)
540
541static const struct btf_kfunc_id_set bpf_rstat_kfunc_set = {
542 .owner = THIS_MODULE,
543 .set = &bpf_rstat_kfunc_ids,
544};
545
546static int __init bpf_rstat_kfunc_init(void)
547{
548 return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
549 &bpf_rstat_kfunc_set);
550}
551late_initcall(bpf_rstat_kfunc_init);
1// SPDX-License-Identifier: GPL-2.0-only
2#include "cgroup-internal.h"
3
4#include <linux/sched/cputime.h>
5
6#include <linux/bpf.h>
7#include <linux/btf.h>
8#include <linux/btf_ids.h>
9
10#include <trace/events/cgroup.h>
11
12static DEFINE_SPINLOCK(cgroup_rstat_lock);
13static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
14
15static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
16
17static struct cgroup_rstat_cpu *cgroup_rstat_cpu(struct cgroup *cgrp, int cpu)
18{
19 return per_cpu_ptr(cgrp->rstat_cpu, cpu);
20}
21
22/*
23 * Helper functions for rstat per CPU lock (cgroup_rstat_cpu_lock).
24 *
25 * This makes it easier to diagnose locking issues and contention in
26 * production environments. The parameter @fast_path determine the
27 * tracepoints being added, allowing us to diagnose "flush" related
28 * operations without handling high-frequency fast-path "update" events.
29 */
30static __always_inline
31unsigned long _cgroup_rstat_cpu_lock(raw_spinlock_t *cpu_lock, int cpu,
32 struct cgroup *cgrp, const bool fast_path)
33{
34 unsigned long flags;
35 bool contended;
36
37 /*
38 * The _irqsave() is needed because cgroup_rstat_lock is
39 * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
40 * this lock with the _irq() suffix only disables interrupts on
41 * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
42 * interrupts on both configurations. The _irqsave() ensures
43 * that interrupts are always disabled and later restored.
44 */
45 contended = !raw_spin_trylock_irqsave(cpu_lock, flags);
46 if (contended) {
47 if (fast_path)
48 trace_cgroup_rstat_cpu_lock_contended_fastpath(cgrp, cpu, contended);
49 else
50 trace_cgroup_rstat_cpu_lock_contended(cgrp, cpu, contended);
51
52 raw_spin_lock_irqsave(cpu_lock, flags);
53 }
54
55 if (fast_path)
56 trace_cgroup_rstat_cpu_locked_fastpath(cgrp, cpu, contended);
57 else
58 trace_cgroup_rstat_cpu_locked(cgrp, cpu, contended);
59
60 return flags;
61}
62
63static __always_inline
64void _cgroup_rstat_cpu_unlock(raw_spinlock_t *cpu_lock, int cpu,
65 struct cgroup *cgrp, unsigned long flags,
66 const bool fast_path)
67{
68 if (fast_path)
69 trace_cgroup_rstat_cpu_unlock_fastpath(cgrp, cpu, false);
70 else
71 trace_cgroup_rstat_cpu_unlock(cgrp, cpu, false);
72
73 raw_spin_unlock_irqrestore(cpu_lock, flags);
74}
75
76/**
77 * cgroup_rstat_updated - keep track of updated rstat_cpu
78 * @cgrp: target cgroup
79 * @cpu: cpu on which rstat_cpu was updated
80 *
81 * @cgrp's rstat_cpu on @cpu was updated. Put it on the parent's matching
82 * rstat_cpu->updated_children list. See the comment on top of
83 * cgroup_rstat_cpu definition for details.
84 */
85__bpf_kfunc void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
86{
87 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
88 unsigned long flags;
89
90 /*
91 * Speculative already-on-list test. This may race leading to
92 * temporary inaccuracies, which is fine.
93 *
94 * Because @parent's updated_children is terminated with @parent
95 * instead of NULL, we can tell whether @cgrp is on the list by
96 * testing the next pointer for NULL.
97 */
98 if (data_race(cgroup_rstat_cpu(cgrp, cpu)->updated_next))
99 return;
100
101 flags = _cgroup_rstat_cpu_lock(cpu_lock, cpu, cgrp, true);
102
103 /* put @cgrp and all ancestors on the corresponding updated lists */
104 while (true) {
105 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
106 struct cgroup *parent = cgroup_parent(cgrp);
107 struct cgroup_rstat_cpu *prstatc;
108
109 /*
110 * Both additions and removals are bottom-up. If a cgroup
111 * is already in the tree, all ancestors are.
112 */
113 if (rstatc->updated_next)
114 break;
115
116 /* Root has no parent to link it to, but mark it busy */
117 if (!parent) {
118 rstatc->updated_next = cgrp;
119 break;
120 }
121
122 prstatc = cgroup_rstat_cpu(parent, cpu);
123 rstatc->updated_next = prstatc->updated_children;
124 prstatc->updated_children = cgrp;
125
126 cgrp = parent;
127 }
128
129 _cgroup_rstat_cpu_unlock(cpu_lock, cpu, cgrp, flags, true);
130}
131
132/**
133 * cgroup_rstat_push_children - push children cgroups into the given list
134 * @head: current head of the list (= subtree root)
135 * @child: first child of the root
136 * @cpu: target cpu
137 * Return: A new singly linked list of cgroups to be flush
138 *
139 * Iteratively traverse down the cgroup_rstat_cpu updated tree level by
140 * level and push all the parents first before their next level children
141 * into a singly linked list built from the tail backward like "pushing"
142 * cgroups into a stack. The root is pushed by the caller.
143 */
144static struct cgroup *cgroup_rstat_push_children(struct cgroup *head,
145 struct cgroup *child, int cpu)
146{
147 struct cgroup *chead = child; /* Head of child cgroup level */
148 struct cgroup *ghead = NULL; /* Head of grandchild cgroup level */
149 struct cgroup *parent, *grandchild;
150 struct cgroup_rstat_cpu *crstatc;
151
152 child->rstat_flush_next = NULL;
153
154next_level:
155 while (chead) {
156 child = chead;
157 chead = child->rstat_flush_next;
158 parent = cgroup_parent(child);
159
160 /* updated_next is parent cgroup terminated */
161 while (child != parent) {
162 child->rstat_flush_next = head;
163 head = child;
164 crstatc = cgroup_rstat_cpu(child, cpu);
165 grandchild = crstatc->updated_children;
166 if (grandchild != child) {
167 /* Push the grand child to the next level */
168 crstatc->updated_children = child;
169 grandchild->rstat_flush_next = ghead;
170 ghead = grandchild;
171 }
172 child = crstatc->updated_next;
173 crstatc->updated_next = NULL;
174 }
175 }
176
177 if (ghead) {
178 chead = ghead;
179 ghead = NULL;
180 goto next_level;
181 }
182 return head;
183}
184
185/**
186 * cgroup_rstat_updated_list - return a list of updated cgroups to be flushed
187 * @root: root of the cgroup subtree to traverse
188 * @cpu: target cpu
189 * Return: A singly linked list of cgroups to be flushed
190 *
191 * Walks the updated rstat_cpu tree on @cpu from @root. During traversal,
192 * each returned cgroup is unlinked from the updated tree.
193 *
194 * The only ordering guarantee is that, for a parent and a child pair
195 * covered by a given traversal, the child is before its parent in
196 * the list.
197 *
198 * Note that updated_children is self terminated and points to a list of
199 * child cgroups if not empty. Whereas updated_next is like a sibling link
200 * within the children list and terminated by the parent cgroup. An exception
201 * here is the cgroup root whose updated_next can be self terminated.
202 */
203static struct cgroup *cgroup_rstat_updated_list(struct cgroup *root, int cpu)
204{
205 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
206 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(root, cpu);
207 struct cgroup *head = NULL, *parent, *child;
208 unsigned long flags;
209
210 flags = _cgroup_rstat_cpu_lock(cpu_lock, cpu, root, false);
211
212 /* Return NULL if this subtree is not on-list */
213 if (!rstatc->updated_next)
214 goto unlock_ret;
215
216 /*
217 * Unlink @root from its parent. As the updated_children list is
218 * singly linked, we have to walk it to find the removal point.
219 */
220 parent = cgroup_parent(root);
221 if (parent) {
222 struct cgroup_rstat_cpu *prstatc;
223 struct cgroup **nextp;
224
225 prstatc = cgroup_rstat_cpu(parent, cpu);
226 nextp = &prstatc->updated_children;
227 while (*nextp != root) {
228 struct cgroup_rstat_cpu *nrstatc;
229
230 nrstatc = cgroup_rstat_cpu(*nextp, cpu);
231 WARN_ON_ONCE(*nextp == parent);
232 nextp = &nrstatc->updated_next;
233 }
234 *nextp = rstatc->updated_next;
235 }
236
237 rstatc->updated_next = NULL;
238
239 /* Push @root to the list first before pushing the children */
240 head = root;
241 root->rstat_flush_next = NULL;
242 child = rstatc->updated_children;
243 rstatc->updated_children = root;
244 if (child != root)
245 head = cgroup_rstat_push_children(head, child, cpu);
246unlock_ret:
247 _cgroup_rstat_cpu_unlock(cpu_lock, cpu, root, flags, false);
248 return head;
249}
250
251/*
252 * A hook for bpf stat collectors to attach to and flush their stats.
253 * Together with providing bpf kfuncs for cgroup_rstat_updated() and
254 * cgroup_rstat_flush(), this enables a complete workflow where bpf progs that
255 * collect cgroup stats can integrate with rstat for efficient flushing.
256 *
257 * A static noinline declaration here could cause the compiler to optimize away
258 * the function. A global noinline declaration will keep the definition, but may
259 * optimize away the callsite. Therefore, __weak is needed to ensure that the
260 * call is still emitted, by telling the compiler that we don't know what the
261 * function might eventually be.
262 */
263
264__bpf_hook_start();
265
266__weak noinline void bpf_rstat_flush(struct cgroup *cgrp,
267 struct cgroup *parent, int cpu)
268{
269}
270
271__bpf_hook_end();
272
273/*
274 * Helper functions for locking cgroup_rstat_lock.
275 *
276 * This makes it easier to diagnose locking issues and contention in
277 * production environments. The parameter @cpu_in_loop indicate lock
278 * was released and re-taken when collection data from the CPUs. The
279 * value -1 is used when obtaining the main lock else this is the CPU
280 * number processed last.
281 */
282static inline void __cgroup_rstat_lock(struct cgroup *cgrp, int cpu_in_loop)
283 __acquires(&cgroup_rstat_lock)
284{
285 bool contended;
286
287 contended = !spin_trylock_irq(&cgroup_rstat_lock);
288 if (contended) {
289 trace_cgroup_rstat_lock_contended(cgrp, cpu_in_loop, contended);
290 spin_lock_irq(&cgroup_rstat_lock);
291 }
292 trace_cgroup_rstat_locked(cgrp, cpu_in_loop, contended);
293}
294
295static inline void __cgroup_rstat_unlock(struct cgroup *cgrp, int cpu_in_loop)
296 __releases(&cgroup_rstat_lock)
297{
298 trace_cgroup_rstat_unlock(cgrp, cpu_in_loop, false);
299 spin_unlock_irq(&cgroup_rstat_lock);
300}
301
302/* see cgroup_rstat_flush() */
303static void cgroup_rstat_flush_locked(struct cgroup *cgrp)
304 __releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock)
305{
306 int cpu;
307
308 lockdep_assert_held(&cgroup_rstat_lock);
309
310 for_each_possible_cpu(cpu) {
311 struct cgroup *pos = cgroup_rstat_updated_list(cgrp, cpu);
312
313 for (; pos; pos = pos->rstat_flush_next) {
314 struct cgroup_subsys_state *css;
315
316 cgroup_base_stat_flush(pos, cpu);
317 bpf_rstat_flush(pos, cgroup_parent(pos), cpu);
318
319 rcu_read_lock();
320 list_for_each_entry_rcu(css, &pos->rstat_css_list,
321 rstat_css_node)
322 css->ss->css_rstat_flush(css, cpu);
323 rcu_read_unlock();
324 }
325
326 /* play nice and yield if necessary */
327 if (need_resched() || spin_needbreak(&cgroup_rstat_lock)) {
328 __cgroup_rstat_unlock(cgrp, cpu);
329 if (!cond_resched())
330 cpu_relax();
331 __cgroup_rstat_lock(cgrp, cpu);
332 }
333 }
334}
335
336/**
337 * cgroup_rstat_flush - flush stats in @cgrp's subtree
338 * @cgrp: target cgroup
339 *
340 * Collect all per-cpu stats in @cgrp's subtree into the global counters
341 * and propagate them upwards. After this function returns, all cgroups in
342 * the subtree have up-to-date ->stat.
343 *
344 * This also gets all cgroups in the subtree including @cgrp off the
345 * ->updated_children lists.
346 *
347 * This function may block.
348 */
349__bpf_kfunc void cgroup_rstat_flush(struct cgroup *cgrp)
350{
351 might_sleep();
352
353 __cgroup_rstat_lock(cgrp, -1);
354 cgroup_rstat_flush_locked(cgrp);
355 __cgroup_rstat_unlock(cgrp, -1);
356}
357
358/**
359 * cgroup_rstat_flush_hold - flush stats in @cgrp's subtree and hold
360 * @cgrp: target cgroup
361 *
362 * Flush stats in @cgrp's subtree and prevent further flushes. Must be
363 * paired with cgroup_rstat_flush_release().
364 *
365 * This function may block.
366 */
367void cgroup_rstat_flush_hold(struct cgroup *cgrp)
368 __acquires(&cgroup_rstat_lock)
369{
370 might_sleep();
371 __cgroup_rstat_lock(cgrp, -1);
372 cgroup_rstat_flush_locked(cgrp);
373}
374
375/**
376 * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
377 * @cgrp: cgroup used by tracepoint
378 */
379void cgroup_rstat_flush_release(struct cgroup *cgrp)
380 __releases(&cgroup_rstat_lock)
381{
382 __cgroup_rstat_unlock(cgrp, -1);
383}
384
385int cgroup_rstat_init(struct cgroup *cgrp)
386{
387 int cpu;
388
389 /* the root cgrp has rstat_cpu preallocated */
390 if (!cgrp->rstat_cpu) {
391 cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
392 if (!cgrp->rstat_cpu)
393 return -ENOMEM;
394 }
395
396 /* ->updated_children list is self terminated */
397 for_each_possible_cpu(cpu) {
398 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
399
400 rstatc->updated_children = cgrp;
401 u64_stats_init(&rstatc->bsync);
402 }
403
404 return 0;
405}
406
407void cgroup_rstat_exit(struct cgroup *cgrp)
408{
409 int cpu;
410
411 cgroup_rstat_flush(cgrp);
412
413 /* sanity check */
414 for_each_possible_cpu(cpu) {
415 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
416
417 if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
418 WARN_ON_ONCE(rstatc->updated_next))
419 return;
420 }
421
422 free_percpu(cgrp->rstat_cpu);
423 cgrp->rstat_cpu = NULL;
424}
425
426void __init cgroup_rstat_boot(void)
427{
428 int cpu;
429
430 for_each_possible_cpu(cpu)
431 raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
432}
433
434/*
435 * Functions for cgroup basic resource statistics implemented on top of
436 * rstat.
437 */
438static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
439 struct cgroup_base_stat *src_bstat)
440{
441 dst_bstat->cputime.utime += src_bstat->cputime.utime;
442 dst_bstat->cputime.stime += src_bstat->cputime.stime;
443 dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
444#ifdef CONFIG_SCHED_CORE
445 dst_bstat->forceidle_sum += src_bstat->forceidle_sum;
446#endif
447 dst_bstat->ntime += src_bstat->ntime;
448}
449
450static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
451 struct cgroup_base_stat *src_bstat)
452{
453 dst_bstat->cputime.utime -= src_bstat->cputime.utime;
454 dst_bstat->cputime.stime -= src_bstat->cputime.stime;
455 dst_bstat->cputime.sum_exec_runtime -= src_bstat->cputime.sum_exec_runtime;
456#ifdef CONFIG_SCHED_CORE
457 dst_bstat->forceidle_sum -= src_bstat->forceidle_sum;
458#endif
459 dst_bstat->ntime -= src_bstat->ntime;
460}
461
462static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
463{
464 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
465 struct cgroup *parent = cgroup_parent(cgrp);
466 struct cgroup_rstat_cpu *prstatc;
467 struct cgroup_base_stat delta;
468 unsigned seq;
469
470 /* Root-level stats are sourced from system-wide CPU stats */
471 if (!parent)
472 return;
473
474 /* fetch the current per-cpu values */
475 do {
476 seq = __u64_stats_fetch_begin(&rstatc->bsync);
477 delta = rstatc->bstat;
478 } while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
479
480 /* propagate per-cpu delta to cgroup and per-cpu global statistics */
481 cgroup_base_stat_sub(&delta, &rstatc->last_bstat);
482 cgroup_base_stat_add(&cgrp->bstat, &delta);
483 cgroup_base_stat_add(&rstatc->last_bstat, &delta);
484 cgroup_base_stat_add(&rstatc->subtree_bstat, &delta);
485
486 /* propagate cgroup and per-cpu global delta to parent (unless that's root) */
487 if (cgroup_parent(parent)) {
488 delta = cgrp->bstat;
489 cgroup_base_stat_sub(&delta, &cgrp->last_bstat);
490 cgroup_base_stat_add(&parent->bstat, &delta);
491 cgroup_base_stat_add(&cgrp->last_bstat, &delta);
492
493 delta = rstatc->subtree_bstat;
494 prstatc = cgroup_rstat_cpu(parent, cpu);
495 cgroup_base_stat_sub(&delta, &rstatc->last_subtree_bstat);
496 cgroup_base_stat_add(&prstatc->subtree_bstat, &delta);
497 cgroup_base_stat_add(&rstatc->last_subtree_bstat, &delta);
498 }
499}
500
501static struct cgroup_rstat_cpu *
502cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp, unsigned long *flags)
503{
504 struct cgroup_rstat_cpu *rstatc;
505
506 rstatc = get_cpu_ptr(cgrp->rstat_cpu);
507 *flags = u64_stats_update_begin_irqsave(&rstatc->bsync);
508 return rstatc;
509}
510
511static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
512 struct cgroup_rstat_cpu *rstatc,
513 unsigned long flags)
514{
515 u64_stats_update_end_irqrestore(&rstatc->bsync, flags);
516 cgroup_rstat_updated(cgrp, smp_processor_id());
517 put_cpu_ptr(rstatc);
518}
519
520void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
521{
522 struct cgroup_rstat_cpu *rstatc;
523 unsigned long flags;
524
525 rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
526 rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
527 cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
528}
529
530void __cgroup_account_cputime_field(struct cgroup *cgrp,
531 enum cpu_usage_stat index, u64 delta_exec)
532{
533 struct cgroup_rstat_cpu *rstatc;
534 unsigned long flags;
535
536 rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
537
538 switch (index) {
539 case CPUTIME_NICE:
540 rstatc->bstat.ntime += delta_exec;
541 fallthrough;
542 case CPUTIME_USER:
543 rstatc->bstat.cputime.utime += delta_exec;
544 break;
545 case CPUTIME_SYSTEM:
546 case CPUTIME_IRQ:
547 case CPUTIME_SOFTIRQ:
548 rstatc->bstat.cputime.stime += delta_exec;
549 break;
550#ifdef CONFIG_SCHED_CORE
551 case CPUTIME_FORCEIDLE:
552 rstatc->bstat.forceidle_sum += delta_exec;
553 break;
554#endif
555 default:
556 break;
557 }
558
559 cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
560}
561
562/*
563 * compute the cputime for the root cgroup by getting the per cpu data
564 * at a global level, then categorizing the fields in a manner consistent
565 * with how it is done by __cgroup_account_cputime_field for each bit of
566 * cpu time attributed to a cgroup.
567 */
568static void root_cgroup_cputime(struct cgroup_base_stat *bstat)
569{
570 struct task_cputime *cputime = &bstat->cputime;
571 int i;
572
573 memset(bstat, 0, sizeof(*bstat));
574 for_each_possible_cpu(i) {
575 struct kernel_cpustat kcpustat;
576 u64 *cpustat = kcpustat.cpustat;
577 u64 user = 0;
578 u64 sys = 0;
579
580 kcpustat_cpu_fetch(&kcpustat, i);
581
582 user += cpustat[CPUTIME_USER];
583 user += cpustat[CPUTIME_NICE];
584 cputime->utime += user;
585
586 sys += cpustat[CPUTIME_SYSTEM];
587 sys += cpustat[CPUTIME_IRQ];
588 sys += cpustat[CPUTIME_SOFTIRQ];
589 cputime->stime += sys;
590
591 cputime->sum_exec_runtime += user;
592 cputime->sum_exec_runtime += sys;
593
594#ifdef CONFIG_SCHED_CORE
595 bstat->forceidle_sum += cpustat[CPUTIME_FORCEIDLE];
596#endif
597 bstat->ntime += cpustat[CPUTIME_NICE];
598 }
599}
600
601
602static void cgroup_force_idle_show(struct seq_file *seq, struct cgroup_base_stat *bstat)
603{
604#ifdef CONFIG_SCHED_CORE
605 u64 forceidle_time = bstat->forceidle_sum;
606
607 do_div(forceidle_time, NSEC_PER_USEC);
608 seq_printf(seq, "core_sched.force_idle_usec %llu\n", forceidle_time);
609#endif
610}
611
612void cgroup_base_stat_cputime_show(struct seq_file *seq)
613{
614 struct cgroup *cgrp = seq_css(seq)->cgroup;
615 u64 usage, utime, stime, ntime;
616
617 if (cgroup_parent(cgrp)) {
618 cgroup_rstat_flush_hold(cgrp);
619 usage = cgrp->bstat.cputime.sum_exec_runtime;
620 cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
621 &utime, &stime);
622 ntime = cgrp->bstat.ntime;
623 cgroup_rstat_flush_release(cgrp);
624 } else {
625 /* cgrp->bstat of root is not actually used, reuse it */
626 root_cgroup_cputime(&cgrp->bstat);
627 usage = cgrp->bstat.cputime.sum_exec_runtime;
628 utime = cgrp->bstat.cputime.utime;
629 stime = cgrp->bstat.cputime.stime;
630 ntime = cgrp->bstat.ntime;
631 }
632
633 do_div(usage, NSEC_PER_USEC);
634 do_div(utime, NSEC_PER_USEC);
635 do_div(stime, NSEC_PER_USEC);
636 do_div(ntime, NSEC_PER_USEC);
637
638 seq_printf(seq, "usage_usec %llu\n"
639 "user_usec %llu\n"
640 "system_usec %llu\n"
641 "nice_usec %llu\n",
642 usage, utime, stime, ntime);
643
644 cgroup_force_idle_show(seq, &cgrp->bstat);
645}
646
647/* Add bpf kfuncs for cgroup_rstat_updated() and cgroup_rstat_flush() */
648BTF_KFUNCS_START(bpf_rstat_kfunc_ids)
649BTF_ID_FLAGS(func, cgroup_rstat_updated)
650BTF_ID_FLAGS(func, cgroup_rstat_flush, KF_SLEEPABLE)
651BTF_KFUNCS_END(bpf_rstat_kfunc_ids)
652
653static const struct btf_kfunc_id_set bpf_rstat_kfunc_set = {
654 .owner = THIS_MODULE,
655 .set = &bpf_rstat_kfunc_ids,
656};
657
658static int __init bpf_rstat_kfunc_init(void)
659{
660 return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
661 &bpf_rstat_kfunc_set);
662}
663late_initcall(bpf_rstat_kfunc_init);