kernel/sched/cpuacct.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPU accounting code for task groups.
   4  *
   5  * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
   6  * (balbir@in.ibm.com).
   7  */
   8 #include <asm/irq_regs.h>
   9 #include "sched.h"
  10
  11 /* Time spent by the tasks of the CPU accounting group executing in ... */
  12 enum cpuacct_stat_index {
  13         CPUACCT_STAT_USER,      /* ... user mode */
  14         CPUACCT_STAT_SYSTEM,    /* ... kernel mode */
  15
  16         CPUACCT_STAT_NSTATS,
  17 };
  18
  19 static const char * const cpuacct_stat_desc[] = {
  20         [CPUACCT_STAT_USER] = "user",
  21         [CPUACCT_STAT_SYSTEM] = "system",
  22 };
  23
  24 /* track CPU usage of a group of tasks and its child groups */
  25 struct cpuacct {
  26         struct cgroup_subsys_state      css;
  27         /* cpuusage holds pointer to a u64-type object on every CPU */
  28         u64 __percpu    *cpuusage;
  29         struct kernel_cpustat __percpu  *cpustat;
  30 };
  31
  32 static inline struct cpuacct *css_ca(struct cgroup_subsys_state *css)
  33 {
  34         return css ? container_of(css, struct cpuacct, css) : NULL;
  35 }
  36
  37 /* Return CPU accounting group to which this task belongs */
  38 static inline struct cpuacct *task_ca(struct task_struct *tsk)
  39 {
  40         return css_ca(task_css(tsk, cpuacct_cgrp_id));
  41 }
  42
  43 static inline struct cpuacct *parent_ca(struct cpuacct *ca)
  44 {
  45         return css_ca(ca->css.parent);
  46 }
  47
  48 static DEFINE_PER_CPU(u64, root_cpuacct_cpuusage);
  49 static struct cpuacct root_cpuacct = {
  50         .cpustat        = &kernel_cpustat,
  51         .cpuusage       = &root_cpuacct_cpuusage,
  52 };
  53
  54 /* Create a new CPU accounting group */
  55 static struct cgroup_subsys_state *
  56 cpuacct_css_alloc(struct cgroup_subsys_state *parent_css)
  57 {
  58         struct cpuacct *ca;
  59
  60         if (!parent_css)
  61                 return &root_cpuacct.css;
  62
  63         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
  64         if (!ca)
  65                 goto out;
  66
  67         ca->cpuusage = alloc_percpu(u64);
  68         if (!ca->cpuusage)
  69                 goto out_free_ca;
  70
  71         ca->cpustat = alloc_percpu(struct kernel_cpustat);
  72         if (!ca->cpustat)
  73                 goto out_free_cpuusage;
  74
  75         return &ca->css;
  76
  77 out_free_cpuusage:
  78         free_percpu(ca->cpuusage);
  79 out_free_ca:
  80         kfree(ca);
  81 out:
  82         return ERR_PTR(-ENOMEM);
  83 }
  84
  85 /* Destroy an existing CPU accounting group */
  86 static void cpuacct_css_free(struct cgroup_subsys_state *css)
  87 {
  88         struct cpuacct *ca = css_ca(css);
  89
  90         free_percpu(ca->cpustat);
  91         free_percpu(ca->cpuusage);
  92         kfree(ca);
  93 }
  94
  95 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu,
  96                                  enum cpuacct_stat_index index)
  97 {
  98         u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
  99         u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
 100         u64 data;
 101
 102         /*
 103          * We allow index == CPUACCT_STAT_NSTATS here to read
 104          * the sum of usages.
 105          */
 106         BUG_ON(index > CPUACCT_STAT_NSTATS);
 107
 108 #ifndef CONFIG_64BIT
 109         /*
 110          * Take rq->lock to make 64-bit read safe on 32-bit platforms.
 111          */
 112         raw_spin_rq_lock_irq(cpu_rq(cpu));
 113 #endif
 114
 115         switch (index) {
 116         case CPUACCT_STAT_USER:
 117                 data = cpustat[CPUTIME_USER] + cpustat[CPUTIME_NICE];
 118                 break;
 119         case CPUACCT_STAT_SYSTEM:
 120                 data = cpustat[CPUTIME_SYSTEM] + cpustat[CPUTIME_IRQ] +
 121                         cpustat[CPUTIME_SOFTIRQ];
 122                 break;
 123         case CPUACCT_STAT_NSTATS:
 124                 data = *cpuusage;
 125                 break;
 126         }
 127
 128 #ifndef CONFIG_64BIT
 129         raw_spin_rq_unlock_irq(cpu_rq(cpu));
 130 #endif
 131
 132         return data;
 133 }
 134
 135 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu)
 136 {
 137         u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
 138         u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
 139
 140         /* Don't allow to reset global kernel_cpustat */
 141         if (ca == &root_cpuacct)
 142                 return;
 143
 144 #ifndef CONFIG_64BIT
 145         /*
 146          * Take rq->lock to make 64-bit write safe on 32-bit platforms.
 147          */
 148         raw_spin_rq_lock_irq(cpu_rq(cpu));
 149 #endif
 150         *cpuusage = 0;
 151         cpustat[CPUTIME_USER] = cpustat[CPUTIME_NICE] = 0;
 152         cpustat[CPUTIME_SYSTEM] = cpustat[CPUTIME_IRQ] = 0;
 153         cpustat[CPUTIME_SOFTIRQ] = 0;
 154
 155 #ifndef CONFIG_64BIT
 156         raw_spin_rq_unlock_irq(cpu_rq(cpu));
 157 #endif
 158 }
 159
 160 /* Return total CPU usage (in nanoseconds) of a group */
 161 static u64 __cpuusage_read(struct cgroup_subsys_state *css,
 162                            enum cpuacct_stat_index index)
 163 {
 164         struct cpuacct *ca = css_ca(css);
 165         u64 totalcpuusage = 0;
 166         int i;
 167
 168         for_each_possible_cpu(i)
 169                 totalcpuusage += cpuacct_cpuusage_read(ca, i, index);
 170
 171         return totalcpuusage;
 172 }
 173
 174 static u64 cpuusage_user_read(struct cgroup_subsys_state *css,
 175                               struct cftype *cft)
 176 {
 177         return __cpuusage_read(css, CPUACCT_STAT_USER);
 178 }
 179
 180 static u64 cpuusage_sys_read(struct cgroup_subsys_state *css,
 181                              struct cftype *cft)
 182 {
 183         return __cpuusage_read(css, CPUACCT_STAT_SYSTEM);
 184 }
 185
 186 static u64 cpuusage_read(struct cgroup_subsys_state *css, struct cftype *cft)
 187 {
 188         return __cpuusage_read(css, CPUACCT_STAT_NSTATS);
 189 }
 190
 191 static int cpuusage_write(struct cgroup_subsys_state *css, struct cftype *cft,
 192                           u64 val)
 193 {
 194         struct cpuacct *ca = css_ca(css);
 195         int cpu;
 196
 197         /*
 198          * Only allow '0' here to do a reset.
 199          */
 200         if (val)
 201                 return -EINVAL;
 202
 203         for_each_possible_cpu(cpu)
 204                 cpuacct_cpuusage_write(ca, cpu);
 205
 206         return 0;
 207 }
 208
 209 static int __cpuacct_percpu_seq_show(struct seq_file *m,
 210                                      enum cpuacct_stat_index index)
 211 {
 212         struct cpuacct *ca = css_ca(seq_css(m));
 213         u64 percpu;
 214         int i;
 215
 216         for_each_possible_cpu(i) {
 217                 percpu = cpuacct_cpuusage_read(ca, i, index);
 218                 seq_printf(m, "%llu ", (unsigned long long) percpu);
 219         }
 220         seq_printf(m, "\n");
 221         return 0;
 222 }
 223
 224 static int cpuacct_percpu_user_seq_show(struct seq_file *m, void *V)
 225 {
 226         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_USER);
 227 }
 228
 229 static int cpuacct_percpu_sys_seq_show(struct seq_file *m, void *V)
 230 {
 231         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_SYSTEM);
 232 }
 233
 234 static int cpuacct_percpu_seq_show(struct seq_file *m, void *V)
 235 {
 236         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_NSTATS);
 237 }
 238
 239 static int cpuacct_all_seq_show(struct seq_file *m, void *V)
 240 {
 241         struct cpuacct *ca = css_ca(seq_css(m));
 242         int index;
 243         int cpu;
 244
 245         seq_puts(m, "cpu");
 246         for (index = 0; index < CPUACCT_STAT_NSTATS; index++)
 247                 seq_printf(m, " %s", cpuacct_stat_desc[index]);
 248         seq_puts(m, "\n");
 249
 250         for_each_possible_cpu(cpu) {
 251                 seq_printf(m, "%d", cpu);
 252                 for (index = 0; index < CPUACCT_STAT_NSTATS; index++)
 253                         seq_printf(m, " %llu",
 254                                    cpuacct_cpuusage_read(ca, cpu, index));
 255                 seq_puts(m, "\n");
 256         }
 257         return 0;
 258 }
 259
 260 static int cpuacct_stats_show(struct seq_file *sf, void *v)
 261 {
 262         struct cpuacct *ca = css_ca(seq_css(sf));
 263         s64 val[CPUACCT_STAT_NSTATS];
 264         int cpu;
 265         int stat;
 266
 267         memset(val, 0, sizeof(val));
 268         for_each_possible_cpu(cpu) {
 269                 u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
 270
 271                 val[CPUACCT_STAT_USER]   += cpustat[CPUTIME_USER];
 272                 val[CPUACCT_STAT_USER]   += cpustat[CPUTIME_NICE];
 273                 val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SYSTEM];
 274                 val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_IRQ];
 275                 val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SOFTIRQ];
 276         }
 277
 278         for (stat = 0; stat < CPUACCT_STAT_NSTATS; stat++) {
 279                 seq_printf(sf, "%s %lld\n",
 280                            cpuacct_stat_desc[stat],
 281                            (long long)nsec_to_clock_t(val[stat]));
 282         }
 283
 284         return 0;
 285 }
 286
 287 static struct cftype files[] = {
 288         {
 289                 .name = "usage",
 290                 .read_u64 = cpuusage_read,
 291                 .write_u64 = cpuusage_write,
 292         },
 293         {
 294                 .name = "usage_user",
 295                 .read_u64 = cpuusage_user_read,
 296         },
 297         {
 298                 .name = "usage_sys",
 299                 .read_u64 = cpuusage_sys_read,
 300         },
 301         {
 302                 .name = "usage_percpu",
 303                 .seq_show = cpuacct_percpu_seq_show,
 304         },
 305         {
 306                 .name = "usage_percpu_user",
 307                 .seq_show = cpuacct_percpu_user_seq_show,
 308         },
 309         {
 310                 .name = "usage_percpu_sys",
 311                 .seq_show = cpuacct_percpu_sys_seq_show,
 312         },
 313         {
 314                 .name = "usage_all",
 315                 .seq_show = cpuacct_all_seq_show,
 316         },
 317         {
 318                 .name = "stat",
 319                 .seq_show = cpuacct_stats_show,
 320         },
 321         { }     /* terminate */
 322 };
 323
 324 /*
 325  * charge this task's execution time to its accounting group.
 326  *
 327  * called with rq->lock held.
 328  */
 329 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 330 {
 331         unsigned int cpu = task_cpu(tsk);
 332         struct cpuacct *ca;
 333
 334         rcu_read_lock();
 335
 336         for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
 337                 *per_cpu_ptr(ca->cpuusage, cpu) += cputime;
 338
 339         rcu_read_unlock();
 340 }
 341
 342 /*
 343  * Add user/system time to cpuacct.
 344  *
 345  * Note: it's the caller that updates the account of the root cgroup.
 346  */
 347 void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
 348 {
 349         struct cpuacct *ca;
 350
 351         rcu_read_lock();
 352         for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
 353                 __this_cpu_add(ca->cpustat->cpustat[index], val);
 354         rcu_read_unlock();
 355 }
 356
 357 struct cgroup_subsys cpuacct_cgrp_subsys = {
 358         .css_alloc      = cpuacct_css_alloc,
 359         .css_free       = cpuacct_css_free,
 360         .legacy_cftypes = files,
 361         .early_init     = true,
 362 };