cgroup: Implement cgroup2 basic CPU usage accounting

In cgroup1, while cpuacct isn't actually controlling any resources, it
is a separate controller due to combination of two factors -
1. enabling cpu controller has significant side effects, and 2. we
have to pick one of the hierarchies to account CPU usages on.  cpuacct
controller is effectively used to designate a hierarchy to track CPU
usages on.

cgroup2's unified hierarchy removes the second reason and we can
account basic CPU usages by default.  While we can use cpuacct for
this purpose, both its interface and implementation leave a lot to be
desired - it collects and exposes two sources of truth which don't
agree with each other and some of the exposed statistics don't make
much sense.  Also, it propagates all the way up the hierarchy on each
accounting event which is unnecessary.

This patch adds basic resource accounting mechanism to cgroup2's
unified hierarchy and accounts CPU usages using it.

* All accountings are done per-cpu and don't propagate immediately.
  It just bumps the per-cgroup per-cpu counters and links to the
  parent's updated list if not already on it.

* On a read, the per-cpu counters are collected into the global ones
  and then propagated upwards.  Only the per-cpu counters which have
  changed since the last read are propagated.

* CPU usage stats are collected and shown in "cgroup.stat" with "cpu."
  prefix.  Total usage is collected from scheduling events.  User/sys
  breakdown is sourced from tick sampling and adjusted to the usage
  using cputime_adjust().

This keeps the accounting side hot path O(1) and per-cpu and the read
side O(nr_updated_since_last_read).

v2: Minor changes and documentation updates as suggested by Waiman and
    Roman.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Roman Gushchin <guro@fb.com>

Documentation/cgroup-v2.txt

@@ -886,6 +886,15 @@ All cgroup core files are prefixed with "cgroup."
 		A dying cgroup can consume system resources not exceeding
 		limits, which were active at the moment of cgroup deletion.
 
+	  cpu.usage_usec
+		CPU time consumed in the subtree.
+
+	  cpu.user_usec
+		User CPU time consumed in the subtree.
+
+	  cpu.system_usec
+		System CPU time consumed in the subtree.
+
 
 Controllers
 ===========

include/linux/cgroup-defs.h

@@ -16,6 +16,7 @@
 #include <linux/refcount.h>
 #include <linux/percpu-refcount.h>
 #include <linux/percpu-rwsem.h>
+#include <linux/u64_stats_sync.h>
 #include <linux/workqueue.h>
 #include <linux/bpf-cgroup.h>
 
@@ -254,6 +255,57 @@ struct css_set {
 	struct rcu_head rcu_head;
 };
 
+/*
+ * cgroup basic resource usage statistics.  Accounting is done per-cpu in
+ * cgroup_cpu_stat which is then lazily propagated up the hierarchy on
+ * reads.
+ *
+ * When a stat gets updated, the cgroup_cpu_stat and its ancestors are
+ * linked into the updated tree.  On the following read, propagation only
+ * considers and consumes the updated tree.  This makes reading O(the
+ * number of descendants which have been active since last read) instead of
+ * O(the total number of descendants).
+ *
+ * This is important because there can be a lot of (draining) cgroups which
+ * aren't active and stat may be read frequently.  The combination can
+ * become very expensive.  By propagating selectively, increasing reading
+ * frequency decreases the cost of each read.
+ */
+struct cgroup_cpu_stat {
+	/*
+	 * ->sync protects all the current counters.  These are the only
+	 * fields which get updated in the hot path.
+	 */
+	struct u64_stats_sync sync;
+	struct task_cputime cputime;
+
+	/*
+	 * Snapshots at the last reading.  These are used to calculate the
+	 * deltas to propagate to the global counters.
+	 */
+	struct task_cputime last_cputime;
+
+	/*
+	 * Child cgroups with stat updates on this cpu since the last read
+	 * are linked on the parent's ->updated_children through
+	 * ->updated_next.
+	 *
+	 * In addition to being more compact, singly-linked list pointing
+	 * to the cgroup makes it unnecessary for each per-cpu struct to
+	 * point back to the associated cgroup.
+	 *
+	 * Protected by per-cpu cgroup_cpu_stat_lock.
+	 */
+	struct cgroup *updated_children;	/* terminated by self cgroup */
+	struct cgroup *updated_next;		/* NULL iff not on the list */
+};
+
+struct cgroup_stat {
+	/* per-cpu statistics are collected into the folowing global counters */
+	struct task_cputime cputime;
+	struct prev_cputime prev_cputime;
+};
+
 struct cgroup {
 	/* self css with NULL ->ss, points back to this cgroup */
 	struct cgroup_subsys_state self;
@@ -353,6 +405,11 @@ struct cgroup {
 	 */
 	struct cgroup *dom_cgrp;
 
+	/* cgroup basic resource statistics */
+	struct cgroup_cpu_stat __percpu *cpu_stat;
+	struct cgroup_stat pending_stat;	/* pending from children */
+	struct cgroup_stat stat;
+
 	/*
 	 * list of pidlists, up to two for each namespace (one for procs, one
 	 * for tasks); created on demand.

include/linux/cgroup.h

@@ -703,17 +703,39 @@ static inline void cpuacct_account_field(struct task_struct *tsk, int index,
 					 u64 val) {}
 #endif
 
+void cgroup_stat_show_cputime(struct seq_file *seq, const char *prefix);
+
+void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec);
+void __cgroup_account_cputime_field(struct cgroup *cgrp,
+				    enum cpu_usage_stat index, u64 delta_exec);
+
 static inline void cgroup_account_cputime(struct task_struct *task,
 					  u64 delta_exec)
 {
+	struct cgroup *cgrp;
+
 	cpuacct_charge(task, delta_exec);
+
+	rcu_read_lock();
+	cgrp = task_dfl_cgroup(task);
+	if (cgroup_parent(cgrp))
+		__cgroup_account_cputime(cgrp, delta_exec);
+	rcu_read_unlock();
 }
 
 static inline void cgroup_account_cputime_field(struct task_struct *task,
 						enum cpu_usage_stat index,
 						u64 delta_exec)
 {
+	struct cgroup *cgrp;
+
 	cpuacct_account_field(task, index, delta_exec);
+
+	rcu_read_lock();
+	cgrp = task_dfl_cgroup(task);
+	if (cgroup_parent(cgrp))
+		__cgroup_account_cputime_field(cgrp, index, delta_exec);
+	rcu_read_unlock();
 }
 
 #else	/* CONFIG_CGROUPS */

kernel/cgroup/Makefile

-obj-y := cgroup.o namespace.o cgroup-v1.o
+obj-y := cgroup.o stat.o namespace.o cgroup-v1.o
 
 obj-$(CONFIG_CGROUP_FREEZER) += freezer.o
 obj-$(CONFIG_CGROUP_PIDS) += pids.o

kernel/cgroup/cgroup-internal.h

@@ -199,6 +199,14 @@ int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
 
 int cgroup_task_count(const struct cgroup *cgrp);
 
+/*
+ * stat.c
+ */
+void cgroup_stat_flush(struct cgroup *cgrp);
+int cgroup_stat_init(struct cgroup *cgrp);
+void cgroup_stat_exit(struct cgroup *cgrp);
+void cgroup_stat_boot(void);
+
 /*
  * namespace.c
  */

kernel/cgroup/cgroup.c

@@ -142,12 +142,14 @@ static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
 };
 #undef SUBSYS
 
+static DEFINE_PER_CPU(struct cgroup_cpu_stat, cgrp_dfl_root_cpu_stat);
+
 /*
  * The default hierarchy, reserved for the subsystems that are otherwise
  * unattached - it never has more than a single cgroup, and all tasks are
  * part of that cgroup.
  */
-struct cgroup_root cgrp_dfl_root;
+struct cgroup_root cgrp_dfl_root = { .cgrp.cpu_stat = &cgrp_dfl_root_cpu_stat };
 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
 
 /*
@@ -3301,6 +3303,8 @@ static int cgroup_stat_show(struct seq_file *seq, void *v)
 	seq_printf(seq, "nr_dying_descendants %d\n",
 		   cgroup->nr_dying_descendants);
 
+	cgroup_stat_show_cputime(seq, "cpu.");
+
 	return 0;
 }
 
@@ -4471,6 +4475,8 @@ static void css_free_work_fn(struct work_struct *work)
 			 */
 			cgroup_put(cgroup_parent(cgrp));
 			kernfs_put(cgrp->kn);
+			if (cgroup_on_dfl(cgrp))
+				cgroup_stat_exit(cgrp);
 			kfree(cgrp);
 		} else {
 			/*
@@ -4515,6 +4521,9 @@ static void css_release_work_fn(struct work_struct *work)
 		/* cgroup release path */
 		trace_cgroup_release(cgrp);
 
+		if (cgroup_on_dfl(cgrp))
+			cgroup_stat_flush(cgrp);
+
 		for (tcgrp = cgroup_parent(cgrp); tcgrp;
 		     tcgrp = cgroup_parent(tcgrp))
 			tcgrp->nr_dying_descendants--;
@@ -4698,6 +4707,12 @@ static struct cgroup *cgroup_create(struct cgroup *parent)
 	if (ret)
 		goto out_free_cgrp;
 
+	if (cgroup_on_dfl(parent)) {
+		ret = cgroup_stat_init(cgrp);
+		if (ret)
+			goto out_cancel_ref;
+	}
+
 	/*
 	 * Temporarily set the pointer to NULL, so idr_find() won't return
 	 * a half-baked cgroup.
@@ -4705,7 +4720,7 @@ static struct cgroup *cgroup_create(struct cgroup *parent)
 	cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
 	if (cgrp->id < 0) {
 		ret = -ENOMEM;
-		goto out_cancel_ref;
+		goto out_stat_exit;
 	}
 
 	init_cgroup_housekeeping(cgrp);
@@ -4754,6 +4769,9 @@ static struct cgroup *cgroup_create(struct cgroup *parent)
 
 	return cgrp;
 
+out_stat_exit:
+	if (cgroup_on_dfl(parent))
+		cgroup_stat_exit(cgrp);
 out_cancel_ref:
 	percpu_ref_exit(&cgrp->self.refcnt);
 out_free_cgrp:
@@ -5148,6 +5166,8 @@ int __init cgroup_init(void)
 	BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
 	BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
 
+	cgroup_stat_boot();
+
 	/*
 	 * The latency of the synchronize_sched() is too high for cgroups,
 	 * avoid it at the cost of forcing all readers into the slow path.

kernel/cgroup/stat.c

+#include "cgroup-internal.h"
+
+#include <linux/sched/cputime.h>
+
+static DEFINE_MUTEX(cgroup_stat_mutex);
+static DEFINE_PER_CPU(raw_spinlock_t, cgroup_cpu_stat_lock);
+
+static struct cgroup_cpu_stat *cgroup_cpu_stat(struct cgroup *cgrp, int cpu)
+{
+	return per_cpu_ptr(cgrp->cpu_stat, cpu);
+}
+
+/**
+ * cgroup_cpu_stat_updated - keep track of updated cpu_stat
+ * @cgrp: target cgroup
+ * @cpu: cpu on which cpu_stat was updated
+ *
+ * @cgrp's cpu_stat on @cpu was updated.  Put it on the parent's matching
+ * cpu_stat->updated_children list.  See the comment on top of
+ * cgroup_cpu_stat definition for details.
+ */
+static void cgroup_cpu_stat_updated(struct cgroup *cgrp, int cpu)
+{
+	raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_cpu_stat_lock, cpu);
+	struct cgroup *parent;
+	unsigned long flags;
+
+	/*
+	 * Speculative already-on-list test.  This may race leading to
+	 * temporary inaccuracies, which is fine.
+	 *
+	 * Because @parent's updated_children is terminated with @parent
+	 * instead of NULL, we can tell whether @cgrp is on the list by
+	 * testing the next pointer for NULL.
+	 */
+	if (cgroup_cpu_stat(cgrp, cpu)->updated_next)
+		return;
+
+	raw_spin_lock_irqsave(cpu_lock, flags);
+
+	/* put @cgrp and all ancestors on the corresponding updated lists */
+	for (parent = cgroup_parent(cgrp); parent;
+	     cgrp = parent, parent = cgroup_parent(cgrp)) {
+		struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
+		struct cgroup_cpu_stat *pcstat = cgroup_cpu_stat(parent, cpu);
+
+		/*
+		 * Both additions and removals are bottom-up.  If a cgroup
+		 * is already in the tree, all ancestors are.
+		 */
+		if (cstat->updated_next)
+			break;
+
+		cstat->updated_next = pcstat->updated_children;
+		pcstat->updated_children = cgrp;
+	}
+
+	raw_spin_unlock_irqrestore(cpu_lock, flags);
+}
+
+/**
+ * cgroup_cpu_stat_pop_updated - iterate and dismantle cpu_stat updated tree
+ * @pos: current position
+ * @root: root of the tree to traversal
+ * @cpu: target cpu
+ *
+ * Walks the udpated cpu_stat tree on @cpu from @root.  %NULL @pos starts
+ * the traversal and %NULL return indicates the end.  During traversal,
+ * each returned cgroup is unlinked from the tree.  Must be called with the
+ * matching cgroup_cpu_stat_lock held.
+ *
+ * The only ordering guarantee is that, for a parent and a child pair
+ * covered by a given traversal, if a child is visited, its parent is
+ * guaranteed to be visited afterwards.
+ */
+static struct cgroup *cgroup_cpu_stat_pop_updated(struct cgroup *pos,
+						  struct cgroup *root, int cpu)
+{
+	struct cgroup_cpu_stat *cstat;
+	struct cgroup *parent;
+
+	if (pos == root)
+		return NULL;
+
+	/*
+	 * We're gonna walk down to the first leaf and visit/remove it.  We
+	 * can pick whatever unvisited node as the starting point.
+	 */
+	if (!pos)
+		pos = root;
+	else
+		pos = cgroup_parent(pos);
+
+	/* walk down to the first leaf */
+	while (true) {
+		cstat = cgroup_cpu_stat(pos, cpu);
+		if (cstat->updated_children == pos)
+			break;
+		pos = cstat->updated_children;
+	}
+
+	/*
+	 * Unlink @pos from the tree.  As the updated_children list is
+	 * singly linked, we have to walk it to find the removal point.
+	 * However, due to the way we traverse, @pos will be the first
+	 * child in most cases. The only exception is @root.
+	 */
+	parent = cgroup_parent(pos);
+	if (parent && cstat->updated_next) {
+		struct cgroup_cpu_stat *pcstat = cgroup_cpu_stat(parent, cpu);
+		struct cgroup_cpu_stat *ncstat;
+		struct cgroup **nextp;
+
+		nextp = &pcstat->updated_children;
+		while (true) {
+			ncstat = cgroup_cpu_stat(*nextp, cpu);
+			if (*nextp == pos)
+				break;
+
+			WARN_ON_ONCE(*nextp == parent);
+			nextp = &ncstat->updated_next;
+		}
+
+		*nextp = cstat->updated_next;
+		cstat->updated_next = NULL;
+	}
+
+	return pos;
+}
+
+static void cgroup_stat_accumulate(struct cgroup_stat *dst_stat,
+				   struct cgroup_stat *src_stat)
+{
+	dst_stat->cputime.utime += src_stat->cputime.utime;
+	dst_stat->cputime.stime += src_stat->cputime.stime;
+	dst_stat->cputime.sum_exec_runtime += src_stat->cputime.sum_exec_runtime;
+}
+
+static void cgroup_cpu_stat_flush_one(struct cgroup *cgrp, int cpu)
+{
+	struct cgroup *parent = cgroup_parent(cgrp);
+	struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
+	struct task_cputime *last_cputime = &cstat->last_cputime;
+	struct task_cputime cputime;
+	struct cgroup_stat delta;
+	unsigned seq;
+
+	lockdep_assert_held(&cgroup_stat_mutex);
+
+	/* fetch the current per-cpu values */
+	do {
+		seq = __u64_stats_fetch_begin(&cstat->sync);
+		cputime = cstat->cputime;
+	} while (__u64_stats_fetch_retry(&cstat->sync, seq));
+
+	/* accumulate the deltas to propgate */
+	delta.cputime.utime = cputime.utime - last_cputime->utime;
+	delta.cputime.stime = cputime.stime - last_cputime->stime;
+	delta.cputime.sum_exec_runtime = cputime.sum_exec_runtime -
+					 last_cputime->sum_exec_runtime;
+	*last_cputime = cputime;
+
+	/* transfer the pending stat into delta */
+	cgroup_stat_accumulate(&delta, &cgrp->pending_stat);
+	memset(&cgrp->pending_stat, 0, sizeof(cgrp->pending_stat));
+
+	/* propagate delta into the global stat and the parent's pending */
+	cgroup_stat_accumulate(&cgrp->stat, &delta);
+	if (parent)
+		cgroup_stat_accumulate(&parent->pending_stat, &delta);
+}
+
+/* see cgroup_stat_flush() */
+static void cgroup_stat_flush_locked(struct cgroup *cgrp)
+{
+	int cpu;
+
+	lockdep_assert_held(&cgroup_stat_mutex);
+
+	for_each_possible_cpu(cpu) {
+		raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_cpu_stat_lock, cpu);
+		struct cgroup *pos = NULL;
+
+		raw_spin_lock_irq(cpu_lock);
+		while ((pos = cgroup_cpu_stat_pop_updated(pos, cgrp, cpu)))
+			cgroup_cpu_stat_flush_one(pos, cpu);
+		raw_spin_unlock_irq(cpu_lock);
+	}
+}
+
+/**
+ * cgroup_stat_flush - flush stats in @cgrp's subtree
+ * @cgrp: target cgroup
+ *
+ * Collect all per-cpu stats in @cgrp's subtree into the global counters
+ * and propagate them upwards.  After this function returns, all cgroups in
+ * the subtree have up-to-date ->stat.
+ *
+ * This also gets all cgroups in the subtree including @cgrp off the
+ * ->updated_children lists.
+ */
+void cgroup_stat_flush(struct cgroup *cgrp)
+{
+	mutex_lock(&cgroup_stat_mutex);
+	cgroup_stat_flush_locked(cgrp);
+	mutex_unlock(&cgroup_stat_mutex);
+}
+
+static struct cgroup_cpu_stat *cgroup_cpu_stat_account_begin(struct cgroup *cgrp)
+{
+	struct cgroup_cpu_stat *cstat;
+
+	cstat = get_cpu_ptr(cgrp->cpu_stat);
+	u64_stats_update_begin(&cstat->sync);
+	return cstat;
+}
+
+static void cgroup_cpu_stat_account_end(struct cgroup *cgrp,
+					struct cgroup_cpu_stat *cstat)
+{
+	u64_stats_update_end(&cstat->sync);
+	cgroup_cpu_stat_updated(cgrp, smp_processor_id());
+	put_cpu_ptr(cstat);
+}
+
+void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
+{
+	struct cgroup_cpu_stat *cstat;
+
+	cstat = cgroup_cpu_stat_account_begin(cgrp);
+	cstat->cputime.sum_exec_runtime += delta_exec;
+	cgroup_cpu_stat_account_end(cgrp, cstat);
+}
+
+void __cgroup_account_cputime_field(struct cgroup *cgrp,
+				    enum cpu_usage_stat index, u64 delta_exec)
+{
+	struct cgroup_cpu_stat *cstat;
+
+	cstat = cgroup_cpu_stat_account_begin(cgrp);
+
+	switch (index) {
+	case CPUTIME_USER:
+	case CPUTIME_NICE:
+		cstat->cputime.utime += delta_exec;
+		break;
+	case CPUTIME_SYSTEM:
+	case CPUTIME_IRQ:
+	case CPUTIME_SOFTIRQ:
+		cstat->cputime.stime += delta_exec;
+		break;
+	default:
+		break;
+	}
+
+	cgroup_cpu_stat_account_end(cgrp, cstat);
+}
+
+void cgroup_stat_show_cputime(struct seq_file *seq, const char *prefix)
+{
+	struct cgroup *cgrp = seq_css(seq)->cgroup;
+	u64 usage, utime, stime;
+
+	if (!cgroup_parent(cgrp))
+		return;
+
+	mutex_lock(&cgroup_stat_mutex);
+
+	cgroup_stat_flush_locked(cgrp);
+
+	usage = cgrp->stat.cputime.sum_exec_runtime;
+	cputime_adjust(&cgrp->stat.cputime, &cgrp->stat.prev_cputime,
+		       &utime, &stime);
+
+	mutex_unlock(&cgroup_stat_mutex);
+
+	do_div(usage, NSEC_PER_USEC);
+	do_div(utime, NSEC_PER_USEC);
+	do_div(stime, NSEC_PER_USEC);
+
+	seq_printf(seq, "%susage_usec %llu\n"
+		   "%suser_usec %llu\n"
+		   "%ssystem_usec %llu\n",
+		   prefix, usage, prefix, utime, prefix, stime);
+}
+
+int cgroup_stat_init(struct cgroup *cgrp)
+{
+	int cpu;
+
+	/* the root cgrp has cpu_stat preallocated */
+	if (!cgrp->cpu_stat) {
+		cgrp->cpu_stat = alloc_percpu(struct cgroup_cpu_stat);
+		if (!cgrp->cpu_stat)
+			return -ENOMEM;
+	}
+
+	/* ->updated_children list is self terminated */
+	for_each_possible_cpu(cpu)
+		cgroup_cpu_stat(cgrp, cpu)->updated_children = cgrp;
+
+	prev_cputime_init(&cgrp->stat.prev_cputime);
+
+	return 0;
+}
+
+void cgroup_stat_exit(struct cgroup *cgrp)
+{
+	int cpu;
+
+	cgroup_stat_flush(cgrp);
+
+	/* sanity check */
+	for_each_possible_cpu(cpu) {
+		struct cgroup_cpu_stat *cstat = cgroup_cpu_stat(cgrp, cpu);
+
+		if (WARN_ON_ONCE(cstat->updated_children != cgrp) ||
+		    WARN_ON_ONCE(cstat->updated_next))
+			return;
+	}
+
+	free_percpu(cgrp->cpu_stat);
+	cgrp->cpu_stat = NULL;
+}
+
+void __init cgroup_stat_boot(void)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		raw_spin_lock_init(per_cpu_ptr(&cgroup_cpu_stat_lock, cpu));
+
+	BUG_ON(cgroup_stat_init(&cgrp_dfl_root.cgrp));
+}