1#include "cgroup-internal.h"
  2
  3#include <linux/sched/cputime.h>
  4
  5static DEFINE_MUTEX(cgroup_stat_mutex);
  6static DEFINE_PER_CPU(raw_spinlock_t, cgroup_cpu_stat_lock);
  7
  8static struct cgroup_cpu_stat *cgroup_cpu_stat(struct cgroup *cgrp, int cpu)
  9{
 10	return per_cpu_ptr(cgrp->cpu_stat, cpu);
 11}
 12
 13/**
 14 * cgroup_cpu_stat_updated - keep track of updated cpu_stat
 15 * @cgrp: target cgroup
 16 * @cpu: cpu on which cpu_stat was updated
 17 *
 18 * @cgrp's cpu_stat on @cpu was updated.  Put it on the parent's matching
 19 * cpu_stat->updated_children list.  See the comment on top of
 20 * cgroup_cpu_stat definition for details.
 21 */
 22static void cgroup_cpu_stat_updated(struct cgroup *cgrp, int cpu)
 23{
 24	raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_cpu_stat_lock, cpu);
 25	struct cgroup *parent;
 26	unsigned long flags;
 27
 28	/*
 29	 * Speculative already-on-list test.  This may race leading to
 30	 * temporary inaccuracies, which is fine.
 31	 *
 32	 * Because @parent's updated_children is terminated with @parent
 33	 * instead of NULL, we can tell whether @cgrp is on the list by
 34	 * testing the next pointer for NULL.
 35	 */
 36	if (cgroup_cpu_stat(cgrp, cpu)->updated_next)
 37		return;
 38
 39	raw_spin_lock_irqsave(cpu_lock, flags);
 40
 41	/* put @cgrp and all ancestors on the corresponding updated lists */
 42	for (parent = cgroup_parent(cgrp); parent;
 43	     cgrp = parent, parent = cgroup_parent(cgrp)) {
 44		struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
 45		struct cgroup_cpu_stat *pcstat = cgroup_cpu_stat(parent, cpu);
 46
 47		/*
 48		 * Both additions and removals are bottom-up.  If a cgroup
 49		 * is already in the tree, all ancestors are.
 50		 */
 51		if (cstat->updated_next)
 52			break;
 53
 54		cstat->updated_next = pcstat->updated_children;
 55		pcstat->updated_children = cgrp;
 56	}
 57
 58	raw_spin_unlock_irqrestore(cpu_lock, flags);
 59}
 60
 61/**
 62 * cgroup_cpu_stat_pop_updated - iterate and dismantle cpu_stat updated tree
 63 * @pos: current position
 64 * @root: root of the tree to traversal
 65 * @cpu: target cpu
 66 *
 67 * Walks the udpated cpu_stat tree on @cpu from @root.  %NULL @pos starts
 68 * the traversal and %NULL return indicates the end.  During traversal,
 69 * each returned cgroup is unlinked from the tree.  Must be called with the
 70 * matching cgroup_cpu_stat_lock held.
 71 *
 72 * The only ordering guarantee is that, for a parent and a child pair
 73 * covered by a given traversal, if a child is visited, its parent is
 74 * guaranteed to be visited afterwards.
 75 */
 76static struct cgroup *cgroup_cpu_stat_pop_updated(struct cgroup *pos,
 77						  struct cgroup *root, int cpu)
 78{
 79	struct cgroup_cpu_stat *cstat;
 80	struct cgroup *parent;
 81
 82	if (pos == root)
 83		return NULL;
 84
 85	/*
 86	 * We're gonna walk down to the first leaf and visit/remove it.  We
 87	 * can pick whatever unvisited node as the starting point.
 88	 */
 89	if (!pos)
 90		pos = root;
 91	else
 92		pos = cgroup_parent(pos);
 93
 94	/* walk down to the first leaf */
 95	while (true) {
 96		cstat = cgroup_cpu_stat(pos, cpu);
 97		if (cstat->updated_children == pos)
 98			break;
 99		pos = cstat->updated_children;
100	}
101
102	/*
103	 * Unlink @pos from the tree.  As the updated_children list is
104	 * singly linked, we have to walk it to find the removal point.
105	 * However, due to the way we traverse, @pos will be the first
106	 * child in most cases. The only exception is @root.
107	 */
108	parent = cgroup_parent(pos);
109	if (parent && cstat->updated_next) {
110		struct cgroup_cpu_stat *pcstat = cgroup_cpu_stat(parent, cpu);
111		struct cgroup_cpu_stat *ncstat;
112		struct cgroup **nextp;
113
114		nextp = &pcstat->updated_children;
115		while (true) {
116			ncstat = cgroup_cpu_stat(*nextp, cpu);
117			if (*nextp == pos)
118				break;
119
120			WARN_ON_ONCE(*nextp == parent);
121			nextp = &ncstat->updated_next;
122		}
123
124		*nextp = cstat->updated_next;
125		cstat->updated_next = NULL;
126	}
127
128	return pos;
129}
130
131static void cgroup_stat_accumulate(struct cgroup_stat *dst_stat,
132				   struct cgroup_stat *src_stat)
133{
134	dst_stat->cputime.utime += src_stat->cputime.utime;
135	dst_stat->cputime.stime += src_stat->cputime.stime;
136	dst_stat->cputime.sum_exec_runtime += src_stat->cputime.sum_exec_runtime;
137}
138
139static void cgroup_cpu_stat_flush_one(struct cgroup *cgrp, int cpu)
140{
141	struct cgroup *parent = cgroup_parent(cgrp);
142	struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
143	struct task_cputime *last_cputime = &cstat->last_cputime;
144	struct task_cputime cputime;
145	struct cgroup_stat delta;
146	unsigned seq;
147
148	lockdep_assert_held(&cgroup_stat_mutex);
149
150	/* fetch the current per-cpu values */
151	do {
152		seq = __u64_stats_fetch_begin(&cstat->sync);
153		cputime = cstat->cputime;
154	} while (__u64_stats_fetch_retry(&cstat->sync, seq));
155
156	/* accumulate the deltas to propgate */
157	delta.cputime.utime = cputime.utime - last_cputime->utime;
158	delta.cputime.stime = cputime.stime - last_cputime->stime;
159	delta.cputime.sum_exec_runtime = cputime.sum_exec_runtime -
160					 last_cputime->sum_exec_runtime;
161	*last_cputime = cputime;
162
163	/* transfer the pending stat into delta */
164	cgroup_stat_accumulate(&delta, &cgrp->pending_stat);
165	memset(&cgrp->pending_stat, 0, sizeof(cgrp->pending_stat));
166
167	/* propagate delta into the global stat and the parent's pending */
168	cgroup_stat_accumulate(&cgrp->stat, &delta);
169	if (parent)
170		cgroup_stat_accumulate(&parent->pending_stat, &delta);
171}
172
173/* see cgroup_stat_flush() */
174static void cgroup_stat_flush_locked(struct cgroup *cgrp)
175{
176	int cpu;
177
178	lockdep_assert_held(&cgroup_stat_mutex);
179
180	for_each_possible_cpu(cpu) {
181		raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_cpu_stat_lock, cpu);
182		struct cgroup *pos = NULL;
183
184		raw_spin_lock_irq(cpu_lock);
185		while ((pos = cgroup_cpu_stat_pop_updated(pos, cgrp, cpu)))
186			cgroup_cpu_stat_flush_one(pos, cpu);
187		raw_spin_unlock_irq(cpu_lock);
188	}
189}
190
191/**
192 * cgroup_stat_flush - flush stats in @cgrp's subtree
193 * @cgrp: target cgroup
194 *
195 * Collect all per-cpu stats in @cgrp's subtree into the global counters
196 * and propagate them upwards.  After this function returns, all cgroups in
197 * the subtree have up-to-date ->stat.
198 *
199 * This also gets all cgroups in the subtree including @cgrp off the
200 * ->updated_children lists.
201 */
202void cgroup_stat_flush(struct cgroup *cgrp)
203{
204	mutex_lock(&cgroup_stat_mutex);
205	cgroup_stat_flush_locked(cgrp);
206	mutex_unlock(&cgroup_stat_mutex);
207}
208
209static struct cgroup_cpu_stat *cgroup_cpu_stat_account_begin(struct cgroup *cgrp)
210{
211	struct cgroup_cpu_stat *cstat;
212
213	cstat = get_cpu_ptr(cgrp->cpu_stat);
214	u64_stats_update_begin(&cstat->sync);
215	return cstat;
216}
217
218static void cgroup_cpu_stat_account_end(struct cgroup *cgrp,
219					struct cgroup_cpu_stat *cstat)
220{
221	u64_stats_update_end(&cstat->sync);
222	cgroup_cpu_stat_updated(cgrp, smp_processor_id());
223	put_cpu_ptr(cstat);
224}
225
226void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
227{
228	struct cgroup_cpu_stat *cstat;
229
230	cstat = cgroup_cpu_stat_account_begin(cgrp);
231	cstat->cputime.sum_exec_runtime += delta_exec;
232	cgroup_cpu_stat_account_end(cgrp, cstat);
233}
234
235void __cgroup_account_cputime_field(struct cgroup *cgrp,
236				    enum cpu_usage_stat index, u64 delta_exec)
237{
238	struct cgroup_cpu_stat *cstat;
239
240	cstat = cgroup_cpu_stat_account_begin(cgrp);
241
242	switch (index) {
243	case CPUTIME_USER:
244	case CPUTIME_NICE:
245		cstat->cputime.utime += delta_exec;
246		break;
247	case CPUTIME_SYSTEM:
248	case CPUTIME_IRQ:
249	case CPUTIME_SOFTIRQ:
250		cstat->cputime.stime += delta_exec;
251		break;
252	default:
253		break;
254	}
255
256	cgroup_cpu_stat_account_end(cgrp, cstat);
257}
258
259void cgroup_stat_show_cputime(struct seq_file *seq)
260{
261	struct cgroup *cgrp = seq_css(seq)->cgroup;
262	u64 usage, utime, stime;
263
264	if (!cgroup_parent(cgrp))
265		return;
266
267	mutex_lock(&cgroup_stat_mutex);
268
269	cgroup_stat_flush_locked(cgrp);
270
271	usage = cgrp->stat.cputime.sum_exec_runtime;
272	cputime_adjust(&cgrp->stat.cputime, &cgrp->stat.prev_cputime,
273		       &utime, &stime);
274
275	mutex_unlock(&cgroup_stat_mutex);
276
277	do_div(usage, NSEC_PER_USEC);
278	do_div(utime, NSEC_PER_USEC);
279	do_div(stime, NSEC_PER_USEC);
280
281	seq_printf(seq, "usage_usec %llu\n"
282		   "user_usec %llu\n"
283		   "system_usec %llu\n",
284		   usage, utime, stime);
285}
286
287int cgroup_stat_init(struct cgroup *cgrp)
288{
289	int cpu;
290
291	/* the root cgrp has cpu_stat preallocated */
292	if (!cgrp->cpu_stat) {
293		cgrp->cpu_stat = alloc_percpu(struct cgroup_cpu_stat);
294		if (!cgrp->cpu_stat)
295			return -ENOMEM;
296	}
297
298	/* ->updated_children list is self terminated */
299	for_each_possible_cpu(cpu) {
300		struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
301
302		cstat->updated_children = cgrp;
303		u64_stats_init(&cstat->sync);
304	}
305
306	prev_cputime_init(&cgrp->stat.prev_cputime);
307
308	return 0;
309}
310
311void cgroup_stat_exit(struct cgroup *cgrp)
312{
313	int cpu;
314
315	cgroup_stat_flush(cgrp);
316
317	/* sanity check */
318	for_each_possible_cpu(cpu) {
319		struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
320
321		if (WARN_ON_ONCE(cstat->updated_children != cgrp) ||
322		    WARN_ON_ONCE(cstat->updated_next))
323			return;
324	}
325
326	free_percpu(cgrp->cpu_stat);
327	cgrp->cpu_stat = NULL;
328}
329
330void __init cgroup_stat_boot(void)
331{
332	int cpu;
333
334	for_each_possible_cpu(cpu)
335		raw_spin_lock_init(per_cpu_ptr(&cgroup_cpu_stat_lock, cpu));
336
337	BUG_ON(cgroup_stat_init(&cgrp_dfl_root.cgrp));
338}