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         if (WARN_ON_ONCE(index > CPUACCT_STAT_NSTATS))
 107                 return 0;
 108
 109 #ifndef CONFIG_64BIT
 110         /*
 111          * Take rq->lock to make 64-bit read safe on 32-bit platforms.
 112          */
 113         raw_spin_rq_lock_irq(cpu_rq(cpu));
 114 #endif
 115
 116         switch (index) {
 117         case CPUACCT_STAT_USER:
 118                 data = cpustat[CPUTIME_USER] + cpustat[CPUTIME_NICE];
 119                 break;
 120         case CPUACCT_STAT_SYSTEM:
 121                 data = cpustat[CPUTIME_SYSTEM] + cpustat[CPUTIME_IRQ] +
 122                         cpustat[CPUTIME_SOFTIRQ];
 123                 break;
 124         case CPUACCT_STAT_NSTATS:
 125                 data = *cpuusage;
 126                 break;
 127         }
 128
 129 #ifndef CONFIG_64BIT
 130         raw_spin_rq_unlock_irq(cpu_rq(cpu));
 131 #endif
 132
 133         return data;
 134 }
 135
 136 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu)
 137 {
 138         u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
 139         u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
 140
 141         /* Don't allow to reset global kernel_cpustat */
 142         if (ca == &root_cpuacct)
 143                 return;
 144
 145 #ifndef CONFIG_64BIT
 146         /*
 147          * Take rq->lock to make 64-bit write safe on 32-bit platforms.
 148          */
 149         raw_spin_rq_lock_irq(cpu_rq(cpu));
 150 #endif
 151         *cpuusage = 0;
 152         cpustat[CPUTIME_USER] = cpustat[CPUTIME_NICE] = 0;
 153         cpustat[CPUTIME_SYSTEM] = cpustat[CPUTIME_IRQ] = 0;
 154         cpustat[CPUTIME_SOFTIRQ] = 0;
 155
 156 #ifndef CONFIG_64BIT
 157         raw_spin_rq_unlock_irq(cpu_rq(cpu));
 158 #endif
 159 }
 160
 161 /* Return total CPU usage (in nanoseconds) of a group */
 162 static u64 __cpuusage_read(struct cgroup_subsys_state *css,
 163                            enum cpuacct_stat_index index)
 164 {
 165         struct cpuacct *ca = css_ca(css);
 166         u64 totalcpuusage = 0;
 167         int i;
 168
 169         for_each_possible_cpu(i)
 170                 totalcpuusage += cpuacct_cpuusage_read(ca, i, index);
 171
 172         return totalcpuusage;
 173 }
 174
 175 static u64 cpuusage_user_read(struct cgroup_subsys_state *css,
 176                               struct cftype *cft)
 177 {
 178         return __cpuusage_read(css, CPUACCT_STAT_USER);
 179 }
 180
 181 static u64 cpuusage_sys_read(struct cgroup_subsys_state *css,
 182                              struct cftype *cft)
 183 {
 184         return __cpuusage_read(css, CPUACCT_STAT_SYSTEM);
 185 }
 186
 187 static u64 cpuusage_read(struct cgroup_subsys_state *css, struct cftype *cft)
 188 {
 189         return __cpuusage_read(css, CPUACCT_STAT_NSTATS);
 190 }
 191
 192 static int cpuusage_write(struct cgroup_subsys_state *css, struct cftype *cft,
 193                           u64 val)
 194 {
 195         struct cpuacct *ca = css_ca(css);
 196         int cpu;
 197
 198         /*
 199          * Only allow '0' here to do a reset.
 200          */
 201         if (val)
 202                 return -EINVAL;
 203
 204         for_each_possible_cpu(cpu)
 205                 cpuacct_cpuusage_write(ca, cpu);
 206
 207         return 0;
 208 }
 209
 210 static int __cpuacct_percpu_seq_show(struct seq_file *m,
 211                                      enum cpuacct_stat_index index)
 212 {
 213         struct cpuacct *ca = css_ca(seq_css(m));
 214         u64 percpu;
 215         int i;
 216
 217         for_each_possible_cpu(i) {
 218                 percpu = cpuacct_cpuusage_read(ca, i, index);
 219                 seq_printf(m, "%llu ", (unsigned long long) percpu);
 220         }
 221         seq_printf(m, "\n");
 222         return 0;
 223 }
 224
 225 static int cpuacct_percpu_user_seq_show(struct seq_file *m, void *V)
 226 {
 227         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_USER);
 228 }
 229
 230 static int cpuacct_percpu_sys_seq_show(struct seq_file *m, void *V)
 231 {
 232         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_SYSTEM);
 233 }
 234
 235 static int cpuacct_percpu_seq_show(struct seq_file *m, void *V)
 236 {
 237         return __cpuacct_percpu_seq_show(m, CPUACCT_STAT_NSTATS);
 238 }
 239
 240 static int cpuacct_all_seq_show(struct seq_file *m, void *V)
 241 {
 242         struct cpuacct *ca = css_ca(seq_css(m));
 243         int index;
 244         int cpu;
 245
 246         seq_puts(m, "cpu");
 247         for (index = 0; index < CPUACCT_STAT_NSTATS; index++)
 248                 seq_printf(m, " %s", cpuacct_stat_desc[index]);
 249         seq_puts(m, "\n");
 250
 251         for_each_possible_cpu(cpu) {
 252                 seq_printf(m, "%d", cpu);
 253                 for (index = 0; index < CPUACCT_STAT_NSTATS; index++)
 254                         seq_printf(m, " %llu",
 255                                    cpuacct_cpuusage_read(ca, cpu, index));
 256                 seq_puts(m, "\n");
 257         }
 258         return 0;
 259 }
 260
 261 static int cpuacct_stats_show(struct seq_file *sf, void *v)
 262 {
 263         struct cpuacct *ca = css_ca(seq_css(sf));
 264         struct task_cputime cputime;
 265         u64 val[CPUACCT_STAT_NSTATS];
 266         int cpu;
 267         int stat;
 268
 269         memset(&cputime, 0, sizeof(cputime));
 270         for_each_possible_cpu(cpu) {
 271                 u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
 272
 273                 cputime.utime += cpustat[CPUTIME_USER];
 274                 cputime.utime += cpustat[CPUTIME_NICE];
 275                 cputime.stime += cpustat[CPUTIME_SYSTEM];
 276                 cputime.stime += cpustat[CPUTIME_IRQ];
 277                 cputime.stime += cpustat[CPUTIME_SOFTIRQ];
 278
 279                 cputime.sum_exec_runtime += *per_cpu_ptr(ca->cpuusage, cpu);
 280         }
 281
 282         cputime_adjust(&cputime, &seq_css(sf)->cgroup->prev_cputime,
 283                 &val[CPUACCT_STAT_USER], &val[CPUACCT_STAT_SYSTEM]);
 284
 285         for (stat = 0; stat < CPUACCT_STAT_NSTATS; stat++) {
 286                 seq_printf(sf, "%s %llu\n", cpuacct_stat_desc[stat],
 287                         nsec_to_clock_t(val[stat]));
 288         }
 289
 290         return 0;
 291 }
 292
 293 static struct cftype files[] = {
 294         {
 295                 .name = "usage",
 296                 .read_u64 = cpuusage_read,
 297                 .write_u64 = cpuusage_write,
 298         },
 299         {
 300                 .name = "usage_user",
 301                 .read_u64 = cpuusage_user_read,
 302         },
 303         {
 304                 .name = "usage_sys",
 305                 .read_u64 = cpuusage_sys_read,
 306         },
 307         {
 308                 .name = "usage_percpu",
 309                 .seq_show = cpuacct_percpu_seq_show,
 310         },
 311         {
 312                 .name = "usage_percpu_user",
 313                 .seq_show = cpuacct_percpu_user_seq_show,
 314         },
 315         {
 316                 .name = "usage_percpu_sys",
 317                 .seq_show = cpuacct_percpu_sys_seq_show,
 318         },
 319         {
 320                 .name = "usage_all",
 321                 .seq_show = cpuacct_all_seq_show,
 322         },
 323         {
 324                 .name = "stat",
 325                 .seq_show = cpuacct_stats_show,
 326         },
 327         { }     /* terminate */
 328 };
 329
 330 /*
 331  * charge this task's execution time to its accounting group.
 332  *
 333  * called with rq->lock held.
 334  */
 335 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 336 {
 337         unsigned int cpu = task_cpu(tsk);
 338         struct cpuacct *ca;
 339
 340         rcu_read_lock();
 341
 342         for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
 343                 *per_cpu_ptr(ca->cpuusage, cpu) += cputime;
 344
 345         rcu_read_unlock();
 346 }
 347
 348 /*
 349  * Add user/system time to cpuacct.
 350  *
 351  * Note: it's the caller that updates the account of the root cgroup.
 352  */
 353 void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
 354 {
 355         struct cpuacct *ca;
 356
 357         rcu_read_lock();
 358         for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
 359                 __this_cpu_add(ca->cpustat->cpustat[index], val);
 360         rcu_read_unlock();
 361 }
 362
 363 struct cgroup_subsys cpuacct_cgrp_subsys = {
 364         .css_alloc      = cpuacct_css_alloc,
 365         .css_free       = cpuacct_css_free,
 366         .legacy_cftypes = files,
 367         .early_init     = true,
 368 };