cgroup: add cpu.stat file to root cgroup

Currently, the root cgroup does not have a cpu.stat file. Add one which
is consistent with /proc/stat to capture global cpu statistics that
might not fall under cgroup accounting.

We haven't done this in the past because the data are already presented
in /proc/stat and we didn't want to add overhead from collecting root
cgroup stats when cgroups are configured, but no cgroups have been
created.

By keeping the data consistent with /proc/stat, I think we avoid the
first problem, while improving the usability of cgroups stats.
We avoid the second problem by computing the contents of cpu.stat from
existing data collected for /proc/stat anyway.

Signed-off-by: Boris Burkov <boris@bur.io>
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Tejun Heo <tj@kernel.org>

Documentation/admin-guide/cgroup-v2.rst

@@ -714,9 +714,7 @@ Conventions
 - Settings for a single feature should be contained in a single file.
 
 - The root cgroup should be exempt from resource control and thus
-  shouldn't have resource control interface files.  Also,
+  shouldn't have resource control interface files.
-  informational files on the root cgroup which end up showing global
-  information available elsewhere shouldn't exist.
 
 - The default time unit is microseconds.  If a different unit is ever
   used, an explicit unit suffix must be present.
@@ -985,7 +983,7 @@ CPU Interface Files
 All time durations are in microseconds.
 
   cpu.stat
-	A read-only flat-keyed file which exists on non-root cgroups.
+	A read-only flat-keyed file.
 	This file exists whether the controller is enabled or not.
 
 	It always reports the following three stats:

kernel/cgroup/cgroup.c

@@ -4874,7 +4874,6 @@ static struct cftype cgroup_base_files[] = {
 	},
 	{
 		.name = "cpu.stat",
-		.flags = CFTYPE_NOT_ON_ROOT,
 		.seq_show = cpu_stat_show,
 	},
 #ifdef CONFIG_PSI

kernel/cgroup/rstat.c

@@ -389,18 +389,62 @@ void __cgroup_account_cputime_field(struct cgroup *cgrp,
 	cgroup_base_stat_cputime_account_end(cgrp, rstatc);
 }
 
+/*
+ * compute the cputime for the root cgroup by getting the per cpu data
+ * at a global level, then categorizing the fields in a manner consistent
+ * with how it is done by __cgroup_account_cputime_field for each bit of
+ * cpu time attributed to a cgroup.
+ */
+static void root_cgroup_cputime(struct task_cputime *cputime)
+{
+	int i;
+
+	cputime->stime = 0;
+	cputime->utime = 0;
+	cputime->sum_exec_runtime = 0;
+	for_each_possible_cpu(i) {
+		struct kernel_cpustat kcpustat;
+		u64 *cpustat = kcpustat.cpustat;
+		u64 user = 0;
+		u64 sys = 0;
+
+		kcpustat_cpu_fetch(&kcpustat, i);
+
+		user += cpustat[CPUTIME_USER];
+		user += cpustat[CPUTIME_NICE];
+		cputime->utime += user;
+
+		sys += cpustat[CPUTIME_SYSTEM];
+		sys += cpustat[CPUTIME_IRQ];
+		sys += cpustat[CPUTIME_SOFTIRQ];
+		cputime->stime += sys;
+
+		cputime->sum_exec_runtime += user;
+		cputime->sum_exec_runtime += sys;
+		cputime->sum_exec_runtime += cpustat[CPUTIME_STEAL];
+		cputime->sum_exec_runtime += cpustat[CPUTIME_GUEST];
+		cputime->sum_exec_runtime += cpustat[CPUTIME_GUEST_NICE];
+	}
+}
+
 void cgroup_base_stat_cputime_show(struct seq_file *seq)
 {
 	struct cgroup *cgrp = seq_css(seq)->cgroup;
 	u64 usage, utime, stime;
+	struct task_cputime cputime;
 
-	if (!cgroup_parent(cgrp))
+	if (cgroup_parent(cgrp)) {
-		return;
-
 		cgroup_rstat_flush_hold(cgrp);
 		usage = cgrp->bstat.cputime.sum_exec_runtime;
-	cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime, &utime, &stime);
+		cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
+			       &utime, &stime);
 		cgroup_rstat_flush_release();
+	} else {
+		root_cgroup_cputime(&cputime);
+		usage = cputime.sum_exec_runtime;
+		utime = cputime.utime;
+		stime = cputime.stime;
+	}
 
 	do_div(usage, NSEC_PER_USEC);
 	do_div(utime, NSEC_PER_USEC);