kernel/sched/cpuacct.c

   1 #include <linux/cgroup.h>
   2 #include <linux/slab.h>
   3 #include <linux/percpu.h>
   4 #include <linux/spinlock.h>
   5 #include <linux/cpumask.h>
   6 #include <linux/seq_file.h>
   7 #include <linux/rcupdate.h>
   8 #include <linux/kernel_stat.h>
   9 #include <linux/err.h>
  10
  11 #include "sched.h"
  12
  13 /*
  14  * CPU accounting code for task groups.
  15  *
  16  * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
  17  * (balbir@in.ibm.com).
  18  */
  19
  20 /* Time spent by the tasks of the cpu accounting group executing in ... */
  21 enum cpuacct_stat_index {
  22         CPUACCT_STAT_USER,      /* ... user mode */
  23         CPUACCT_STAT_SYSTEM,    /* ... kernel mode */
  24
  25         CPUACCT_STAT_NSTATS,
  26 };
  27
  28 /* track cpu usage of a group of tasks and its child groups */
  29 struct cpuacct {
  30         struct cgroup_subsys_state css;
  31         /* cpuusage holds pointer to a u64-type object on every cpu */
  32         u64 __percpu *cpuusage;
  33         struct kernel_cpustat __percpu *cpustat;
  34 };
  35
  36 static inline struct cpuacct *css_ca(struct cgroup_subsys_state *css)
  37 {
  38         return css ? container_of(css, struct cpuacct, css) : NULL;
  39 }
  40
  41 /* return cpu accounting group corresponding to this container */
  42 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
  43 {
  44         return css_ca(cgroup_css(cgrp, cpuacct_subsys_id));
  45 }
  46
  47 /* return cpu accounting group to which this task belongs */
  48 static inline struct cpuacct *task_ca(struct task_struct *tsk)
  49 {
  50         return css_ca(task_css(tsk, cpuacct_subsys_id));
  51 }
  52
  53 static inline struct cpuacct *__parent_ca(struct cpuacct *ca)
  54 {
  55         return cgroup_ca(ca->css.cgroup->parent);
  56 }
  57
  58 static inline struct cpuacct *parent_ca(struct cpuacct *ca)
  59 {
  60         if (!ca->css.cgroup->parent)
  61                 return NULL;
  62         return cgroup_ca(ca->css.cgroup->parent);
  63 }
  64
  65 static DEFINE_PER_CPU(u64, root_cpuacct_cpuusage);
  66 static struct cpuacct root_cpuacct = {
  67         .cpustat        = &kernel_cpustat,
  68         .cpuusage       = &root_cpuacct_cpuusage,
  69 };
  70
  71 /* create a new cpu accounting group */
  72 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
  73 {
  74         struct cpuacct *ca;
  75
  76         if (!cgrp->parent)
  77                 return &root_cpuacct.css;
  78
  79         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
  80         if (!ca)
  81                 goto out;
  82
  83         ca->cpuusage = alloc_percpu(u64);
  84         if (!ca->cpuusage)
  85                 goto out_free_ca;
  86
  87         ca->cpustat = alloc_percpu(struct kernel_cpustat);
  88         if (!ca->cpustat)
  89                 goto out_free_cpuusage;
  90
  91         return &ca->css;
  92
  93 out_free_cpuusage:
  94         free_percpu(ca->cpuusage);
  95 out_free_ca:
  96         kfree(ca);
  97 out:
  98         return ERR_PTR(-ENOMEM);
  99 }
 100
 101 /* destroy an existing cpu accounting group */
 102 static void cpuacct_css_free(struct cgroup *cgrp)
 103 {
 104         struct cpuacct *ca = cgroup_ca(cgrp);
 105
 106         free_percpu(ca->cpustat);
 107         free_percpu(ca->cpuusage);
 108         kfree(ca);
 109 }
 110
 111 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
 112 {
 113         u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
 114         u64 data;
 115
 116 #ifndef CONFIG_64BIT
 117         /*
 118          * Take rq->lock to make 64-bit read safe on 32-bit platforms.
 119          */
 120         raw_spin_lock_irq(&cpu_rq(cpu)->lock);
 121         data = *cpuusage;
 122         raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
 123 #else
 124         data = *cpuusage;
 125 #endif
 126
 127         return data;
 128 }
 129
 130 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
 131 {
 132         u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
 133
 134 #ifndef CONFIG_64BIT
 135         /*
 136          * Take rq->lock to make 64-bit write safe on 32-bit platforms.
 137          */
 138         raw_spin_lock_irq(&cpu_rq(cpu)->lock);
 139         *cpuusage = val;
 140         raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
 141 #else
 142         *cpuusage = val;
 143 #endif
 144 }
 145
 146 /* return total cpu usage (in nanoseconds) of a group */
 147 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
 148 {
 149         struct cpuacct *ca = cgroup_ca(cgrp);
 150         u64 totalcpuusage = 0;
 151         int i;
 152
 153         for_each_present_cpu(i)
 154                 totalcpuusage += cpuacct_cpuusage_read(ca, i);
 155
 156         return totalcpuusage;
 157 }
 158
 159 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
 160                                                                 u64 reset)
 161 {
 162         struct cpuacct *ca = cgroup_ca(cgrp);
 163         int err = 0;
 164         int i;
 165
 166         if (reset) {
 167                 err = -EINVAL;
 168                 goto out;
 169         }
 170
 171         for_each_present_cpu(i)
 172                 cpuacct_cpuusage_write(ca, i, 0);
 173
 174 out:
 175         return err;
 176 }
 177
 178 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
 179                                    struct seq_file *m)
 180 {
 181         struct cpuacct *ca = cgroup_ca(cgroup);
 182         u64 percpu;
 183         int i;
 184
 185         for_each_present_cpu(i) {
 186                 percpu = cpuacct_cpuusage_read(ca, i);
 187                 seq_printf(m, "%llu ", (unsigned long long) percpu);
 188         }
 189         seq_printf(m, "\n");
 190         return 0;
 191 }
 192
 193 static const char * const cpuacct_stat_desc[] = {
 194         [CPUACCT_STAT_USER] = "user",
 195         [CPUACCT_STAT_SYSTEM] = "system",
 196 };
 197
 198 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
 199                               struct cgroup_map_cb *cb)
 200 {
 201         struct cpuacct *ca = cgroup_ca(cgrp);
 202         int cpu;
 203         s64 val = 0;
 204
 205         for_each_online_cpu(cpu) {
 206                 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
 207                 val += kcpustat->cpustat[CPUTIME_USER];
 208                 val += kcpustat->cpustat[CPUTIME_NICE];
 209         }
 210         val = cputime64_to_clock_t(val);
 211         cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
 212
 213         val = 0;
 214         for_each_online_cpu(cpu) {
 215                 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
 216                 val += kcpustat->cpustat[CPUTIME_SYSTEM];
 217                 val += kcpustat->cpustat[CPUTIME_IRQ];
 218                 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
 219         }
 220
 221         val = cputime64_to_clock_t(val);
 222         cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
 223
 224         return 0;
 225 }
 226
 227 static struct cftype files[] = {
 228         {
 229                 .name = "usage",
 230                 .read_u64 = cpuusage_read,
 231                 .write_u64 = cpuusage_write,
 232         },
 233         {
 234                 .name = "usage_percpu",
 235                 .read_seq_string = cpuacct_percpu_seq_read,
 236         },
 237         {
 238                 .name = "stat",
 239                 .read_map = cpuacct_stats_show,
 240         },
 241         { }     /* terminate */
 242 };
 243
 244 /*
 245  * charge this task's execution time to its accounting group.
 246  *
 247  * called with rq->lock held.
 248  */
 249 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 250 {
 251         struct cpuacct *ca;
 252         int cpu;
 253
 254         cpu = task_cpu(tsk);
 255
 256         rcu_read_lock();
 257
 258         ca = task_ca(tsk);
 259
 260         while (true) {
 261                 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
 262                 *cpuusage += cputime;
 263
 264                 ca = parent_ca(ca);
 265                 if (!ca)
 266                         break;
 267         }
 268
 269         rcu_read_unlock();
 270 }
 271
 272 /*
 273  * Add user/system time to cpuacct.
 274  *
 275  * Note: it's the caller that updates the account of the root cgroup.
 276  */
 277 void cpuacct_account_field(struct task_struct *p, int index, u64 val)
 278 {
 279         struct kernel_cpustat *kcpustat;
 280         struct cpuacct *ca;
 281
 282         rcu_read_lock();
 283         ca = task_ca(p);
 284         while (ca != &root_cpuacct) {
 285                 kcpustat = this_cpu_ptr(ca->cpustat);
 286                 kcpustat->cpustat[index] += val;
 287                 ca = __parent_ca(ca);
 288         }
 289         rcu_read_unlock();
 290 }
 291
 292 struct cgroup_subsys cpuacct_subsys = {
 293         .name           = "cpuacct",
 294         .css_alloc      = cpuacct_css_alloc,
 295         .css_free       = cpuacct_css_free,
 296         .subsys_id      = cpuacct_subsys_id,
 297         .base_cftypes   = files,
 298         .early_init     = 1,
 299 };