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,bool may_sleep)174 static 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  */
cgroup_rstat_flush(struct cgroup * cgrp)234 __bpf_kfunc void 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  */
cgroup_rstat_flush_irqsafe(struct cgroup * cgrp)249 void 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  */
cgroup_rstat_flush_hold(struct cgroup * cgrp)267 void 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  */
cgroup_rstat_flush_release(void)278 void cgroup_rstat_flush_release(void)
279 	__releases(&cgroup_rstat_lock)
280 {
281 	spin_unlock_irq(&cgroup_rstat_lock);
282 }
283 
cgroup_rstat_init(struct cgroup * cgrp)284 int 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 
cgroup_rstat_exit(struct cgroup * cgrp)306 void 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 
cgroup_rstat_boot(void)325 void __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  */
cgroup_base_stat_add(struct cgroup_base_stat * dst_bstat,struct cgroup_base_stat * src_bstat)337 static 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 
cgroup_base_stat_sub(struct cgroup_base_stat * dst_bstat,struct cgroup_base_stat * src_bstat)348 static 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 
cgroup_base_stat_flush(struct cgroup * cgrp,int cpu)359 static 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 
390 static struct cgroup_rstat_cpu *
cgroup_base_stat_cputime_account_begin(struct cgroup * cgrp,unsigned long * flags)391 cgroup_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 
cgroup_base_stat_cputime_account_end(struct cgroup * cgrp,struct cgroup_rstat_cpu * rstatc,unsigned long flags)400 static 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 
__cgroup_account_cputime(struct cgroup * cgrp,u64 delta_exec)409 void __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 
__cgroup_account_cputime_field(struct cgroup * cgrp,enum cpu_usage_stat index,u64 delta_exec)419 void __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  */
root_cgroup_cputime(struct cgroup_base_stat * bstat)455 static 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 
cgroup_base_stat_cputime_show(struct seq_file * seq)490 void 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() */
536 BTF_SET8_START(bpf_rstat_kfunc_ids)
537 BTF_ID_FLAGS(func, cgroup_rstat_updated)
538 BTF_ID_FLAGS(func, cgroup_rstat_flush, KF_SLEEPABLE)
539 BTF_SET8_END(bpf_rstat_kfunc_ids)
540 
541 static const struct btf_kfunc_id_set bpf_rstat_kfunc_set = {
542 	.owner          = THIS_MODULE,
543 	.set            = &bpf_rstat_kfunc_ids,
544 };
545 
bpf_rstat_kfunc_init(void)546 static 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 }
551 late_initcall(bpf_rstat_kfunc_init);
552