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 static DEFINE_SPINLOCK(cgroup_rstat_lock);
11 static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
12 
13 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
14 
cgroup_rstat_cpu(struct cgroup * cgrp,int cpu)15 static 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  */
cgroup_rstat_updated(struct cgroup * cgrp,int cpu)29 __bpf_kfunc void 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  */
cgroup_rstat_cpu_pop_updated(struct cgroup * pos,struct cgroup * root,int cpu)91 static 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 
bpf_rstat_flush(struct cgroup * cgrp,struct cgroup * parent,int cpu)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() */
cgroup_rstat_flush_locked(struct cgroup * cgrp)174 static void cgroup_rstat_flush_locked(struct cgroup *cgrp)
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 		/* play nice and yield if necessary */
211 		if (need_resched() || spin_needbreak(&cgroup_rstat_lock)) {
212 			spin_unlock_irq(&cgroup_rstat_lock);
213 			if (!cond_resched())
214 				cpu_relax();
215 			spin_lock_irq(&cgroup_rstat_lock);
216 		}
217 	}
218 }
219 
220 /**
221  * cgroup_rstat_flush - flush stats in @cgrp's subtree
222  * @cgrp: target cgroup
223  *
224  * Collect all per-cpu stats in @cgrp's subtree into the global counters
225  * and propagate them upwards.  After this function returns, all cgroups in
226  * the subtree have up-to-date ->stat.
227  *
228  * This also gets all cgroups in the subtree including @cgrp off the
229  * ->updated_children lists.
230  *
231  * This function may block.
232  */
cgroup_rstat_flush(struct cgroup * cgrp)233 __bpf_kfunc void cgroup_rstat_flush(struct cgroup *cgrp)
234 {
235 	might_sleep();
236 
237 	spin_lock_irq(&cgroup_rstat_lock);
238 	cgroup_rstat_flush_locked(cgrp);
239 	spin_unlock_irq(&cgroup_rstat_lock);
240 }
241 
242 /**
243  * cgroup_rstat_flush_hold - flush stats in @cgrp's subtree and hold
244  * @cgrp: target cgroup
245  *
246  * Flush stats in @cgrp's subtree and prevent further flushes.  Must be
247  * paired with cgroup_rstat_flush_release().
248  *
249  * This function may block.
250  */
cgroup_rstat_flush_hold(struct cgroup * cgrp)251 void cgroup_rstat_flush_hold(struct cgroup *cgrp)
252 	__acquires(&cgroup_rstat_lock)
253 {
254 	might_sleep();
255 	spin_lock_irq(&cgroup_rstat_lock);
256 	cgroup_rstat_flush_locked(cgrp);
257 }
258 
259 /**
260  * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
261  */
cgroup_rstat_flush_release(void)262 void cgroup_rstat_flush_release(void)
263 	__releases(&cgroup_rstat_lock)
264 {
265 	spin_unlock_irq(&cgroup_rstat_lock);
266 }
267 
cgroup_rstat_init(struct cgroup * cgrp)268 int cgroup_rstat_init(struct cgroup *cgrp)
269 {
270 	int cpu;
271 
272 	/* the root cgrp has rstat_cpu preallocated */
273 	if (!cgrp->rstat_cpu) {
274 		cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
275 		if (!cgrp->rstat_cpu)
276 			return -ENOMEM;
277 	}
278 
279 	/* ->updated_children list is self terminated */
280 	for_each_possible_cpu(cpu) {
281 		struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
282 
283 		rstatc->updated_children = cgrp;
284 		u64_stats_init(&rstatc->bsync);
285 	}
286 
287 	return 0;
288 }
289 
cgroup_rstat_exit(struct cgroup * cgrp)290 void cgroup_rstat_exit(struct cgroup *cgrp)
291 {
292 	int cpu;
293 
294 	cgroup_rstat_flush(cgrp);
295 
296 	/* sanity check */
297 	for_each_possible_cpu(cpu) {
298 		struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
299 
300 		if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
301 		    WARN_ON_ONCE(rstatc->updated_next))
302 			return;
303 	}
304 
305 	free_percpu(cgrp->rstat_cpu);
306 	cgrp->rstat_cpu = NULL;
307 }
308 
cgroup_rstat_boot(void)309 void __init cgroup_rstat_boot(void)
310 {
311 	int cpu;
312 
313 	for_each_possible_cpu(cpu)
314 		raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
315 }
316 
317 /*
318  * Functions for cgroup basic resource statistics implemented on top of
319  * rstat.
320  */
cgroup_base_stat_add(struct cgroup_base_stat * dst_bstat,struct cgroup_base_stat * src_bstat)321 static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
322 				 struct cgroup_base_stat *src_bstat)
323 {
324 	dst_bstat->cputime.utime += src_bstat->cputime.utime;
325 	dst_bstat->cputime.stime += src_bstat->cputime.stime;
326 	dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
327 #ifdef CONFIG_SCHED_CORE
328 	dst_bstat->forceidle_sum += src_bstat->forceidle_sum;
329 #endif
330 }
331 
cgroup_base_stat_sub(struct cgroup_base_stat * dst_bstat,struct cgroup_base_stat * src_bstat)332 static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
333 				 struct cgroup_base_stat *src_bstat)
334 {
335 	dst_bstat->cputime.utime -= src_bstat->cputime.utime;
336 	dst_bstat->cputime.stime -= src_bstat->cputime.stime;
337 	dst_bstat->cputime.sum_exec_runtime -= src_bstat->cputime.sum_exec_runtime;
338 #ifdef CONFIG_SCHED_CORE
339 	dst_bstat->forceidle_sum -= src_bstat->forceidle_sum;
340 #endif
341 }
342 
cgroup_base_stat_flush(struct cgroup * cgrp,int cpu)343 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
344 {
345 	struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
346 	struct cgroup *parent = cgroup_parent(cgrp);
347 	struct cgroup_rstat_cpu *prstatc;
348 	struct cgroup_base_stat delta;
349 	unsigned seq;
350 
351 	/* Root-level stats are sourced from system-wide CPU stats */
352 	if (!parent)
353 		return;
354 
355 	/* fetch the current per-cpu values */
356 	do {
357 		seq = __u64_stats_fetch_begin(&rstatc->bsync);
358 		delta = rstatc->bstat;
359 	} while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
360 
361 	/* propagate per-cpu delta to cgroup and per-cpu global statistics */
362 	cgroup_base_stat_sub(&delta, &rstatc->last_bstat);
363 	cgroup_base_stat_add(&cgrp->bstat, &delta);
364 	cgroup_base_stat_add(&rstatc->last_bstat, &delta);
365 	cgroup_base_stat_add(&rstatc->subtree_bstat, &delta);
366 
367 	/* propagate cgroup and per-cpu global delta to parent (unless that's root) */
368 	if (cgroup_parent(parent)) {
369 		delta = cgrp->bstat;
370 		cgroup_base_stat_sub(&delta, &cgrp->last_bstat);
371 		cgroup_base_stat_add(&parent->bstat, &delta);
372 		cgroup_base_stat_add(&cgrp->last_bstat, &delta);
373 
374 		delta = rstatc->subtree_bstat;
375 		prstatc = cgroup_rstat_cpu(parent, cpu);
376 		cgroup_base_stat_sub(&delta, &rstatc->last_subtree_bstat);
377 		cgroup_base_stat_add(&prstatc->subtree_bstat, &delta);
378 		cgroup_base_stat_add(&rstatc->last_subtree_bstat, &delta);
379 	}
380 }
381 
382 static struct cgroup_rstat_cpu *
cgroup_base_stat_cputime_account_begin(struct cgroup * cgrp,unsigned long * flags)383 cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp, unsigned long *flags)
384 {
385 	struct cgroup_rstat_cpu *rstatc;
386 
387 	rstatc = get_cpu_ptr(cgrp->rstat_cpu);
388 	*flags = u64_stats_update_begin_irqsave(&rstatc->bsync);
389 	return rstatc;
390 }
391 
cgroup_base_stat_cputime_account_end(struct cgroup * cgrp,struct cgroup_rstat_cpu * rstatc,unsigned long flags)392 static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
393 						 struct cgroup_rstat_cpu *rstatc,
394 						 unsigned long flags)
395 {
396 	u64_stats_update_end_irqrestore(&rstatc->bsync, flags);
397 	cgroup_rstat_updated(cgrp, smp_processor_id());
398 	put_cpu_ptr(rstatc);
399 }
400 
__cgroup_account_cputime(struct cgroup * cgrp,u64 delta_exec)401 void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
402 {
403 	struct cgroup_rstat_cpu *rstatc;
404 	unsigned long flags;
405 
406 	rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
407 	rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
408 	cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
409 }
410 
__cgroup_account_cputime_field(struct cgroup * cgrp,enum cpu_usage_stat index,u64 delta_exec)411 void __cgroup_account_cputime_field(struct cgroup *cgrp,
412 				    enum cpu_usage_stat index, u64 delta_exec)
413 {
414 	struct cgroup_rstat_cpu *rstatc;
415 	unsigned long flags;
416 
417 	rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
418 
419 	switch (index) {
420 	case CPUTIME_USER:
421 	case CPUTIME_NICE:
422 		rstatc->bstat.cputime.utime += delta_exec;
423 		break;
424 	case CPUTIME_SYSTEM:
425 	case CPUTIME_IRQ:
426 	case CPUTIME_SOFTIRQ:
427 		rstatc->bstat.cputime.stime += delta_exec;
428 		break;
429 #ifdef CONFIG_SCHED_CORE
430 	case CPUTIME_FORCEIDLE:
431 		rstatc->bstat.forceidle_sum += delta_exec;
432 		break;
433 #endif
434 	default:
435 		break;
436 	}
437 
438 	cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
439 }
440 
441 /*
442  * compute the cputime for the root cgroup by getting the per cpu data
443  * at a global level, then categorizing the fields in a manner consistent
444  * with how it is done by __cgroup_account_cputime_field for each bit of
445  * cpu time attributed to a cgroup.
446  */
root_cgroup_cputime(struct cgroup_base_stat * bstat)447 static void root_cgroup_cputime(struct cgroup_base_stat *bstat)
448 {
449 	struct task_cputime *cputime = &bstat->cputime;
450 	int i;
451 
452 	memset(bstat, 0, sizeof(*bstat));
453 	for_each_possible_cpu(i) {
454 		struct kernel_cpustat kcpustat;
455 		u64 *cpustat = kcpustat.cpustat;
456 		u64 user = 0;
457 		u64 sys = 0;
458 
459 		kcpustat_cpu_fetch(&kcpustat, i);
460 
461 		user += cpustat[CPUTIME_USER];
462 		user += cpustat[CPUTIME_NICE];
463 		cputime->utime += user;
464 
465 		sys += cpustat[CPUTIME_SYSTEM];
466 		sys += cpustat[CPUTIME_IRQ];
467 		sys += cpustat[CPUTIME_SOFTIRQ];
468 		cputime->stime += sys;
469 
470 		cputime->sum_exec_runtime += user;
471 		cputime->sum_exec_runtime += sys;
472 		cputime->sum_exec_runtime += cpustat[CPUTIME_STEAL];
473 
474 #ifdef CONFIG_SCHED_CORE
475 		bstat->forceidle_sum += cpustat[CPUTIME_FORCEIDLE];
476 #endif
477 	}
478 }
479 
cgroup_base_stat_cputime_show(struct seq_file * seq)480 void cgroup_base_stat_cputime_show(struct seq_file *seq)
481 {
482 	struct cgroup *cgrp = seq_css(seq)->cgroup;
483 	u64 usage, utime, stime;
484 	struct cgroup_base_stat bstat;
485 #ifdef CONFIG_SCHED_CORE
486 	u64 forceidle_time;
487 #endif
488 
489 	if (cgroup_parent(cgrp)) {
490 		cgroup_rstat_flush_hold(cgrp);
491 		usage = cgrp->bstat.cputime.sum_exec_runtime;
492 		cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
493 			       &utime, &stime);
494 #ifdef CONFIG_SCHED_CORE
495 		forceidle_time = cgrp->bstat.forceidle_sum;
496 #endif
497 		cgroup_rstat_flush_release();
498 	} else {
499 		root_cgroup_cputime(&bstat);
500 		usage = bstat.cputime.sum_exec_runtime;
501 		utime = bstat.cputime.utime;
502 		stime = bstat.cputime.stime;
503 #ifdef CONFIG_SCHED_CORE
504 		forceidle_time = bstat.forceidle_sum;
505 #endif
506 	}
507 
508 	do_div(usage, NSEC_PER_USEC);
509 	do_div(utime, NSEC_PER_USEC);
510 	do_div(stime, NSEC_PER_USEC);
511 #ifdef CONFIG_SCHED_CORE
512 	do_div(forceidle_time, NSEC_PER_USEC);
513 #endif
514 
515 	seq_printf(seq, "usage_usec %llu\n"
516 		   "user_usec %llu\n"
517 		   "system_usec %llu\n",
518 		   usage, utime, stime);
519 
520 #ifdef CONFIG_SCHED_CORE
521 	seq_printf(seq, "core_sched.force_idle_usec %llu\n", forceidle_time);
522 #endif
523 }
524 
525 /* Add bpf kfuncs for cgroup_rstat_updated() and cgroup_rstat_flush() */
526 BTF_SET8_START(bpf_rstat_kfunc_ids)
527 BTF_ID_FLAGS(func, cgroup_rstat_updated)
528 BTF_ID_FLAGS(func, cgroup_rstat_flush, KF_SLEEPABLE)
529 BTF_SET8_END(bpf_rstat_kfunc_ids)
530 
531 static const struct btf_kfunc_id_set bpf_rstat_kfunc_set = {
532 	.owner          = THIS_MODULE,
533 	.set            = &bpf_rstat_kfunc_ids,
534 };
535 
bpf_rstat_kfunc_init(void)536 static int __init bpf_rstat_kfunc_init(void)
537 {
538 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
539 					 &bpf_rstat_kfunc_set);
540 }
541 late_initcall(bpf_rstat_kfunc_init);
542