sched/accounting: Re-use scheduler statistics for the root cgroup
authorGlauber Costa <glommer@parallels.com>
Mon, 28 Nov 2011 16:45:19 +0000 (14:45 -0200)
committerIngo Molnar <mingo@elte.hu>
Tue, 6 Dec 2011 19:51:21 +0000 (20:51 +0100)
Right now, after we collect tick statistics for user and system and store them
in a well known location, we keep the same statistics again for cpuacct.
Since cpuacct is hierarchical, the numbers for the root cgroup should be
absolutely equal to the system-wide numbers.

So it would be better to just use it: this patch changes cpuacct accounting
in a way that the cpustat statistics are kept in a struct kernel_cpustat percpu
array. In the root cgroup case, we just point it to the main array. The rest of
the hierarchy walk can be totally disabled later with a static branch - but I am
not doing it here.

Signed-off-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Tuner <pjt@google.com>
Link: http://lkml.kernel.org/r/1322498719-2255-4-git-send-email-glommer@parallels.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched/core.c
kernel/sched/sched.h

index a727c4ea9a3ebb1e3b4fa812bd7ae2ecdea06689..3e078f26cb6772ab9cab833c18b42082053ec568 100644 (file)
@@ -2556,6 +2556,42 @@ unsigned long long task_sched_runtime(struct task_struct *p)
        return ns;
 }
 
+#ifdef CONFIG_CGROUP_CPUACCT
+struct cgroup_subsys cpuacct_subsys;
+struct cpuacct root_cpuacct;
+#endif
+
+static inline void task_group_account_field(struct task_struct *p,
+                                            u64 tmp, int index)
+{
+#ifdef CONFIG_CGROUP_CPUACCT
+       struct kernel_cpustat *kcpustat;
+       struct cpuacct *ca;
+#endif
+       /*
+        * Since all updates are sure to touch the root cgroup, we
+        * get ourselves ahead and touch it first. If the root cgroup
+        * is the only cgroup, then nothing else should be necessary.
+        *
+        */
+       __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
+
+#ifdef CONFIG_CGROUP_CPUACCT
+       if (unlikely(!cpuacct_subsys.active))
+               return;
+
+       rcu_read_lock();
+       ca = task_ca(p);
+       while (ca && (ca != &root_cpuacct)) {
+               kcpustat = this_cpu_ptr(ca->cpustat);
+               kcpustat->cpustat[index] += tmp;
+               ca = parent_ca(ca);
+       }
+       rcu_read_unlock();
+#endif
+}
+
+
 /*
  * Account user cpu time to a process.
  * @p: the process that the cpu time gets accounted to
@@ -2580,7 +2616,7 @@ void account_user_time(struct task_struct *p, cputime_t cputime,
        index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
        cpustat[index] += tmp;
 
-       cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
+       task_group_account_field(p, index, cputime);
        /* Account for user time used */
        acct_update_integrals(p);
 }
@@ -2636,7 +2672,7 @@ void __account_system_time(struct task_struct *p, cputime_t cputime,
 
        /* Add system time to cpustat. */
        cpustat[index] += tmp;
-       cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
+       task_group_account_field(p, index, cputime);
 
        /* Account for system time used */
        acct_update_integrals(p);
@@ -6781,8 +6817,15 @@ void __init sched_init(void)
        INIT_LIST_HEAD(&root_task_group.children);
        INIT_LIST_HEAD(&root_task_group.siblings);
        autogroup_init(&init_task);
+
 #endif /* CONFIG_CGROUP_SCHED */
 
+#ifdef CONFIG_CGROUP_CPUACCT
+       root_cpuacct.cpustat = &kernel_cpustat;
+       root_cpuacct.cpuusage = alloc_percpu(u64);
+       /* Too early, not expected to fail */
+       BUG_ON(!root_cpuacct.cpuusage);
+#endif
        for_each_possible_cpu(i) {
                struct rq *rq;
 
@@ -7843,44 +7886,16 @@ struct cgroup_subsys cpu_cgroup_subsys = {
  * (balbir@in.ibm.com).
  */
 
-/* track cpu usage of a group of tasks and its child groups */
-struct cpuacct {
-       struct cgroup_subsys_state css;
-       /* cpuusage holds pointer to a u64-type object on every cpu */
-       u64 __percpu *cpuusage;
-       struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
-};
-
-struct cgroup_subsys cpuacct_subsys;
-
-/* return cpu accounting group corresponding to this container */
-static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
-{
-       return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
-                           struct cpuacct, css);
-}
-
-/* return cpu accounting group to which this task belongs */
-static inline struct cpuacct *task_ca(struct task_struct *tsk)
-{
-       return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
-                           struct cpuacct, css);
-}
-
-static inline struct cpuacct *parent_ca(struct cpuacct *ca)
-{
-       if (!ca || !ca->css.cgroup->parent)
-               return NULL;
-       return cgroup_ca(ca->css.cgroup->parent);
-}
-
 /* create a new cpu accounting group */
 static struct cgroup_subsys_state *cpuacct_create(
        struct cgroup_subsys *ss, struct cgroup *cgrp)
 {
-       struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
-       int i;
+       struct cpuacct *ca;
 
+       if (!cgrp->parent)
+               return &root_cpuacct.css;
+
+       ca = kzalloc(sizeof(*ca), GFP_KERNEL);
        if (!ca)
                goto out;
 
@@ -7888,15 +7903,13 @@ static struct cgroup_subsys_state *cpuacct_create(
        if (!ca->cpuusage)
                goto out_free_ca;
 
-       for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
-               if (percpu_counter_init(&ca->cpustat[i], 0))
-                       goto out_free_counters;
+       ca->cpustat = alloc_percpu(struct kernel_cpustat);
+       if (!ca->cpustat)
+               goto out_free_cpuusage;
 
        return &ca->css;
 
-out_free_counters:
-       while (--i >= 0)
-               percpu_counter_destroy(&ca->cpustat[i]);
+out_free_cpuusage:
        free_percpu(ca->cpuusage);
 out_free_ca:
        kfree(ca);
@@ -7909,10 +7922,8 @@ static void
 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
 {
        struct cpuacct *ca = cgroup_ca(cgrp);
-       int i;
 
-       for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
-               percpu_counter_destroy(&ca->cpustat[i]);
+       free_percpu(ca->cpustat);
        free_percpu(ca->cpuusage);
        kfree(ca);
 }
@@ -8005,16 +8016,31 @@ static const char *cpuacct_stat_desc[] = {
 };
 
 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
-               struct cgroup_map_cb *cb)
+                             struct cgroup_map_cb *cb)
 {
        struct cpuacct *ca = cgroup_ca(cgrp);
-       int i;
+       int cpu;
+       s64 val = 0;
 
-       for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
-               s64 val = percpu_counter_read(&ca->cpustat[i]);
-               val = cputime64_to_clock_t(val);
-               cb->fill(cb, cpuacct_stat_desc[i], val);
+       for_each_online_cpu(cpu) {
+               struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
+               val += kcpustat->cpustat[CPUTIME_USER];
+               val += kcpustat->cpustat[CPUTIME_NICE];
        }
+       val = cputime64_to_clock_t(val);
+       cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
+
+       val = 0;
+       for_each_online_cpu(cpu) {
+               struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
+               val += kcpustat->cpustat[CPUTIME_SYSTEM];
+               val += kcpustat->cpustat[CPUTIME_IRQ];
+               val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
+       }
+
+       val = cputime64_to_clock_t(val);
+       cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
+
        return 0;
 }
 
@@ -8066,45 +8092,6 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
        rcu_read_unlock();
 }
 
-/*
- * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
- * in cputime_t units. As a result, cpuacct_update_stats calls
- * percpu_counter_add with values large enough to always overflow the
- * per cpu batch limit causing bad SMP scalability.
- *
- * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
- * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
- * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
- */
-#ifdef CONFIG_SMP
-#define CPUACCT_BATCH  \
-       min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
-#else
-#define CPUACCT_BATCH  0
-#endif
-
-/*
- * Charge the system/user time to the task's accounting group.
- */
-void cpuacct_update_stats(struct task_struct *tsk,
-               enum cpuacct_stat_index idx, cputime_t val)
-{
-       struct cpuacct *ca;
-       int batch = CPUACCT_BATCH;
-
-       if (unlikely(!cpuacct_subsys.active))
-               return;
-
-       rcu_read_lock();
-       ca = task_ca(tsk);
-
-       do {
-               __percpu_counter_add(&ca->cpustat[idx], val, batch);
-               ca = parent_ca(ca);
-       } while (ca);
-       rcu_read_unlock();
-}
-
 struct cgroup_subsys cpuacct_subsys = {
        .name = "cpuacct",
        .create = cpuacct_create,
index d88545c667e3ff9753e16d5a71c68fa6328ed0a8..c248016362194df4017c62bc8117d95db5e20a35 100644 (file)
@@ -830,13 +830,39 @@ extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime
 extern void update_cpu_load(struct rq *this_rq);
 
 #ifdef CONFIG_CGROUP_CPUACCT
+#include <linux/cgroup.h>
+/* track cpu usage of a group of tasks and its child groups */
+struct cpuacct {
+       struct cgroup_subsys_state css;
+       /* cpuusage holds pointer to a u64-type object on every cpu */
+       u64 __percpu *cpuusage;
+       struct kernel_cpustat __percpu *cpustat;
+};
+
+/* return cpu accounting group corresponding to this container */
+static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
+{
+       return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
+                           struct cpuacct, css);
+}
+
+/* return cpu accounting group to which this task belongs */
+static inline struct cpuacct *task_ca(struct task_struct *tsk)
+{
+       return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
+                           struct cpuacct, css);
+}
+
+static inline struct cpuacct *parent_ca(struct cpuacct *ca)
+{
+       if (!ca || !ca->css.cgroup->parent)
+               return NULL;
+       return cgroup_ca(ca->css.cgroup->parent);
+}
+
 extern void cpuacct_charge(struct task_struct *tsk, u64 cputime);
-extern void cpuacct_update_stats(struct task_struct *tsk,
-               enum cpuacct_stat_index idx, cputime_t val);
 #else
 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
-static inline void cpuacct_update_stats(struct task_struct *tsk,
-               enum cpuacct_stat_index idx, cputime_t val) {}
 #endif
 
 static inline void inc_nr_running(struct rq *rq)