kernel/cgroup/rstat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #include "cgroup-internal.h"
   3
   4 #include <linux/sched/cputime.h>
   5
   6 static DEFINE_SPINLOCK(cgroup_rstat_lock);
   7 static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
   8
   9 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
  10
  11 static struct cgroup_rstat_cpu *cgroup_rstat_cpu(struct cgroup *cgrp, int cpu)
  12 {
  13         return per_cpu_ptr(cgrp->rstat_cpu, cpu);
  14 }
  15
  16 /**
  17  * cgroup_rstat_updated - keep track of updated rstat_cpu
  18  * @cgrp: target cgroup
  19  * @cpu: cpu on which rstat_cpu was updated
  20  *
  21  * @cgrp's rstat_cpu on @cpu was updated.  Put it on the parent's matching
  22  * rstat_cpu->updated_children list.  See the comment on top of
  23  * cgroup_rstat_cpu definition for details.
  24  */
  25 void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
  26 {
  27         raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
  28         unsigned long flags;
  29
  30         /*
  31          * Speculative already-on-list test. This may race leading to
  32          * temporary inaccuracies, which is fine.
  33          *
  34          * Because @parent's updated_children is terminated with @parent
  35          * instead of NULL, we can tell whether @cgrp is on the list by
  36          * testing the next pointer for NULL.
  37          */
  38         if (cgroup_rstat_cpu(cgrp, cpu)->updated_next)
  39                 return;
  40
  41         raw_spin_lock_irqsave(cpu_lock, flags);
  42
  43         /* put @cgrp and all ancestors on the corresponding updated lists */
  44         while (true) {
  45                 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
  46                 struct cgroup *parent = cgroup_parent(cgrp);
  47                 struct cgroup_rstat_cpu *prstatc;
  48
  49                 /*
  50                  * Both additions and removals are bottom-up.  If a cgroup
  51                  * is already in the tree, all ancestors are.
  52                  */
  53                 if (rstatc->updated_next)
  54                         break;
  55
  56                 /* Root has no parent to link it to, but mark it busy */
  57                 if (!parent) {
  58                         rstatc->updated_next = cgrp;
  59                         break;
  60                 }
  61
  62                 prstatc = cgroup_rstat_cpu(parent, cpu);
  63                 rstatc->updated_next = prstatc->updated_children;
  64                 prstatc->updated_children = cgrp;
  65
  66                 cgrp = parent;
  67         }
  68
  69         raw_spin_unlock_irqrestore(cpu_lock, flags);
  70 }
  71
  72 /**
  73  * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
  74  * @pos: current position
  75  * @root: root of the tree to traversal
  76  * @cpu: target cpu
  77  *
  78  * Walks the updated rstat_cpu tree on @cpu from @root.  %NULL @pos starts
  79  * the traversal and %NULL return indicates the end.  During traversal,
  80  * each returned cgroup is unlinked from the tree.  Must be called with the
  81  * matching cgroup_rstat_cpu_lock held.
  82  *
  83  * The only ordering guarantee is that, for a parent and a child pair
  84  * covered by a given traversal, if a child is visited, its parent is
  85  * guaranteed to be visited afterwards.
  86  */
  87 static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
  88                                                    struct cgroup *root, int cpu)
  89 {
  90         struct cgroup_rstat_cpu *rstatc;
  91
  92         if (pos == root)
  93                 return NULL;
  94
  95         /*
  96          * We're gonna walk down to the first leaf and visit/remove it.  We
  97          * can pick whatever unvisited node as the starting point.
  98          */
  99         if (!pos)
 100                 pos = root;
 101         else
 102                 pos = cgroup_parent(pos);
 103
 104         /* walk down to the first leaf */
 105         while (true) {
 106                 rstatc = cgroup_rstat_cpu(pos, cpu);
 107                 if (rstatc->updated_children == pos)
 108                         break;
 109                 pos = rstatc->updated_children;
 110         }
 111
 112         /*
 113          * Unlink @pos from the tree.  As the updated_children list is
 114          * singly linked, we have to walk it to find the removal point.
 115          * However, due to the way we traverse, @pos will be the first
 116          * child in most cases. The only exception is @root.
 117          */
 118         if (rstatc->updated_next) {
 119                 struct cgroup *parent = cgroup_parent(pos);
 120
 121                 if (parent) {
 122                         struct cgroup_rstat_cpu *prstatc;
 123                         struct cgroup **nextp;
 124
 125                         prstatc = cgroup_rstat_cpu(parent, cpu);
 126                         nextp = &prstatc->updated_children;
 127                         while (true) {
 128                                 struct cgroup_rstat_cpu *nrstatc;
 129
 130                                 nrstatc = cgroup_rstat_cpu(*nextp, cpu);
 131                                 if (*nextp == pos)
 132                                         break;
 133                                 WARN_ON_ONCE(*nextp == parent);
 134                                 nextp = &nrstatc->updated_next;
 135                         }
 136                         *nextp = rstatc->updated_next;
 137                 }
 138
 139                 rstatc->updated_next = NULL;
 140                 return pos;
 141         }
 142
 143         /* only happens for @root */
 144         return NULL;
 145 }
 146
 147 /* see cgroup_rstat_flush() */
 148 static void cgroup_rstat_flush_locked(struct cgroup *cgrp, bool may_sleep)
 149         __releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock)
 150 {
 151         int cpu;
 152
 153         lockdep_assert_held(&cgroup_rstat_lock);
 154
 155         for_each_possible_cpu(cpu) {
 156                 raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock,
 157                                                        cpu);
 158                 struct cgroup *pos = NULL;
 159                 unsigned long flags;
 160
 161                 raw_spin_lock_irqsave(cpu_lock, flags);
 162                 while ((pos = cgroup_rstat_cpu_pop_updated(pos, cgrp, cpu))) {
 163                         struct cgroup_subsys_state *css;
 164
 165                         cgroup_base_stat_flush(pos, cpu);
 166
 167                         rcu_read_lock();
 168                         list_for_each_entry_rcu(css, &pos->rstat_css_list,
 169                                                 rstat_css_node)
 170                                 css->ss->css_rstat_flush(css, cpu);
 171                         rcu_read_unlock();
 172                 }
 173                 raw_spin_unlock_irqrestore(cpu_lock, flags);
 174
 175                 /* if @may_sleep, play nice and yield if necessary */
 176                 if (may_sleep && (need_resched() ||
 177                                   spin_needbreak(&cgroup_rstat_lock))) {
 178                         spin_unlock_irq(&cgroup_rstat_lock);
 179                         if (!cond_resched())
 180                                 cpu_relax();
 181                         spin_lock_irq(&cgroup_rstat_lock);
 182                 }
 183         }
 184 }
 185
 186 /**
 187  * cgroup_rstat_flush - flush stats in @cgrp's subtree
 188  * @cgrp: target cgroup
 189  *
 190  * Collect all per-cpu stats in @cgrp's subtree into the global counters
 191  * and propagate them upwards.  After this function returns, all cgroups in
 192  * the subtree have up-to-date ->stat.
 193  *
 194  * This also gets all cgroups in the subtree including @cgrp off the
 195  * ->updated_children lists.
 196  *
 197  * This function may block.
 198  */
 199 void cgroup_rstat_flush(struct cgroup *cgrp)
 200 {
 201         might_sleep();
 202
 203         spin_lock_irq(&cgroup_rstat_lock);
 204         cgroup_rstat_flush_locked(cgrp, true);
 205         spin_unlock_irq(&cgroup_rstat_lock);
 206 }
 207
 208 /**
 209  * cgroup_rstat_flush_irqsafe - irqsafe version of cgroup_rstat_flush()
 210  * @cgrp: target cgroup
 211  *
 212  * This function can be called from any context.
 213  */
 214 void cgroup_rstat_flush_irqsafe(struct cgroup *cgrp)
 215 {
 216         unsigned long flags;
 217
 218         spin_lock_irqsave(&cgroup_rstat_lock, flags);
 219         cgroup_rstat_flush_locked(cgrp, false);
 220         spin_unlock_irqrestore(&cgroup_rstat_lock, flags);
 221 }
 222
 223 /**
 224  * cgroup_rstat_flush_hold - flush stats in @cgrp's subtree and hold
 225  * @cgrp: target cgroup
 226  *
 227  * Flush stats in @cgrp's subtree and prevent further flushes.  Must be
 228  * paired with cgroup_rstat_flush_release().
 229  *
 230  * This function may block.
 231  */
 232 void cgroup_rstat_flush_hold(struct cgroup *cgrp)
 233         __acquires(&cgroup_rstat_lock)
 234 {
 235         might_sleep();
 236         spin_lock_irq(&cgroup_rstat_lock);
 237         cgroup_rstat_flush_locked(cgrp, true);
 238 }
 239
 240 /**
 241  * cgroup_rstat_flush_release - release cgroup_rstat_flush_hold()
 242  */
 243 void cgroup_rstat_flush_release(void)
 244         __releases(&cgroup_rstat_lock)
 245 {
 246         spin_unlock_irq(&cgroup_rstat_lock);
 247 }
 248
 249 int cgroup_rstat_init(struct cgroup *cgrp)
 250 {
 251         int cpu;
 252
 253         /* the root cgrp has rstat_cpu preallocated */
 254         if (!cgrp->rstat_cpu) {
 255                 cgrp->rstat_cpu = alloc_percpu(struct cgroup_rstat_cpu);
 256                 if (!cgrp->rstat_cpu)
 257                         return -ENOMEM;
 258         }
 259
 260         /* ->updated_children list is self terminated */
 261         for_each_possible_cpu(cpu) {
 262                 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
 263
 264                 rstatc->updated_children = cgrp;
 265                 u64_stats_init(&rstatc->bsync);
 266         }
 267
 268         return 0;
 269 }
 270
 271 void cgroup_rstat_exit(struct cgroup *cgrp)
 272 {
 273         int cpu;
 274
 275         cgroup_rstat_flush(cgrp);
 276
 277         /* sanity check */
 278         for_each_possible_cpu(cpu) {
 279                 struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
 280
 281                 if (WARN_ON_ONCE(rstatc->updated_children != cgrp) ||
 282                     WARN_ON_ONCE(rstatc->updated_next))
 283                         return;
 284         }
 285
 286         free_percpu(cgrp->rstat_cpu);
 287         cgrp->rstat_cpu = NULL;
 288 }
 289
 290 void __init cgroup_rstat_boot(void)
 291 {
 292         int cpu;
 293
 294         for_each_possible_cpu(cpu)
 295                 raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
 296 }
 297
 298 /*
 299  * Functions for cgroup basic resource statistics implemented on top of
 300  * rstat.
 301  */
 302 static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
 303                                  struct cgroup_base_stat *src_bstat)
 304 {
 305         dst_bstat->cputime.utime += src_bstat->cputime.utime;
 306         dst_bstat->cputime.stime += src_bstat->cputime.stime;
 307         dst_bstat->cputime.sum_exec_runtime += src_bstat->cputime.sum_exec_runtime;
 308 }
 309
 310 static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
 311                                  struct cgroup_base_stat *src_bstat)
 312 {
 313         dst_bstat->cputime.utime -= src_bstat->cputime.utime;
 314         dst_bstat->cputime.stime -= src_bstat->cputime.stime;
 315         dst_bstat->cputime.sum_exec_runtime -= src_bstat->cputime.sum_exec_runtime;
 316 }
 317
 318 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
 319 {
 320         struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(cgrp, cpu);
 321         struct cgroup *parent = cgroup_parent(cgrp);
 322         struct cgroup_base_stat cur, delta;
 323         unsigned seq;
 324
 325         /* Root-level stats are sourced from system-wide CPU stats */
 326         if (!parent)
 327                 return;
 328
 329         /* fetch the current per-cpu values */
 330         do {
 331                 seq = __u64_stats_fetch_begin(&rstatc->bsync);
 332                 cur.cputime = rstatc->bstat.cputime;
 333         } while (__u64_stats_fetch_retry(&rstatc->bsync, seq));
 334
 335         /* propagate percpu delta to global */
 336         delta = cur;
 337         cgroup_base_stat_sub(&delta, &rstatc->last_bstat);
 338         cgroup_base_stat_add(&cgrp->bstat, &delta);
 339         cgroup_base_stat_add(&rstatc->last_bstat, &delta);
 340
 341         /* propagate global delta to parent (unless that's root) */
 342         if (cgroup_parent(parent)) {
 343                 delta = cgrp->bstat;
 344                 cgroup_base_stat_sub(&delta, &cgrp->last_bstat);
 345                 cgroup_base_stat_add(&parent->bstat, &delta);
 346                 cgroup_base_stat_add(&cgrp->last_bstat, &delta);
 347         }
 348 }
 349
 350 static struct cgroup_rstat_cpu *
 351 cgroup_base_stat_cputime_account_begin(struct cgroup *cgrp, unsigned long *flags)
 352 {
 353         struct cgroup_rstat_cpu *rstatc;
 354
 355         rstatc = get_cpu_ptr(cgrp->rstat_cpu);
 356         *flags = u64_stats_update_begin_irqsave(&rstatc->bsync);
 357         return rstatc;
 358 }
 359
 360 static void cgroup_base_stat_cputime_account_end(struct cgroup *cgrp,
 361                                                  struct cgroup_rstat_cpu *rstatc,
 362                                                  unsigned long flags)
 363 {
 364         u64_stats_update_end_irqrestore(&rstatc->bsync, flags);
 365         cgroup_rstat_updated(cgrp, smp_processor_id());
 366         put_cpu_ptr(rstatc);
 367 }
 368
 369 void __cgroup_account_cputime(struct cgroup *cgrp, u64 delta_exec)
 370 {
 371         struct cgroup_rstat_cpu *rstatc;
 372         unsigned long flags;
 373
 374         rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
 375         rstatc->bstat.cputime.sum_exec_runtime += delta_exec;
 376         cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
 377 }
 378
 379 void __cgroup_account_cputime_field(struct cgroup *cgrp,
 380                                     enum cpu_usage_stat index, u64 delta_exec)
 381 {
 382         struct cgroup_rstat_cpu *rstatc;
 383         unsigned long flags;
 384
 385         rstatc = cgroup_base_stat_cputime_account_begin(cgrp, &flags);
 386
 387         switch (index) {
 388         case CPUTIME_USER:
 389         case CPUTIME_NICE:
 390                 rstatc->bstat.cputime.utime += delta_exec;
 391                 break;
 392         case CPUTIME_SYSTEM:
 393         case CPUTIME_IRQ:
 394         case CPUTIME_SOFTIRQ:
 395                 rstatc->bstat.cputime.stime += delta_exec;
 396                 break;
 397         default:
 398                 break;
 399         }
 400
 401         cgroup_base_stat_cputime_account_end(cgrp, rstatc, flags);
 402 }
 403
 404 /*
 405  * compute the cputime for the root cgroup by getting the per cpu data
 406  * at a global level, then categorizing the fields in a manner consistent
 407  * with how it is done by __cgroup_account_cputime_field for each bit of
 408  * cpu time attributed to a cgroup.
 409  */
 410 static void root_cgroup_cputime(struct task_cputime *cputime)
 411 {
 412         int i;
 413
 414         cputime->stime = 0;
 415         cputime->utime = 0;
 416         cputime->sum_exec_runtime = 0;
 417         for_each_possible_cpu(i) {
 418                 struct kernel_cpustat kcpustat;
 419                 u64 *cpustat = kcpustat.cpustat;
 420                 u64 user = 0;
 421                 u64 sys = 0;
 422
 423                 kcpustat_cpu_fetch(&kcpustat, i);
 424
 425                 user += cpustat[CPUTIME_USER];
 426                 user += cpustat[CPUTIME_NICE];
 427                 cputime->utime += user;
 428
 429                 sys += cpustat[CPUTIME_SYSTEM];
 430                 sys += cpustat[CPUTIME_IRQ];
 431                 sys += cpustat[CPUTIME_SOFTIRQ];
 432                 cputime->stime += sys;
 433
 434                 cputime->sum_exec_runtime += user;
 435                 cputime->sum_exec_runtime += sys;
 436                 cputime->sum_exec_runtime += cpustat[CPUTIME_STEAL];
 437         }
 438 }
 439
 440 void cgroup_base_stat_cputime_show(struct seq_file *seq)
 441 {
 442         struct cgroup *cgrp = seq_css(seq)->cgroup;
 443         u64 usage, utime, stime;
 444         struct task_cputime cputime;
 445
 446         if (cgroup_parent(cgrp)) {
 447                 cgroup_rstat_flush_hold(cgrp);
 448                 usage = cgrp->bstat.cputime.sum_exec_runtime;
 449                 cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
 450                                &utime, &stime);
 451                 cgroup_rstat_flush_release();
 452         } else {
 453                 root_cgroup_cputime(&cputime);
 454                 usage = cputime.sum_exec_runtime;
 455                 utime = cputime.utime;
 456                 stime = cputime.stime;
 457         }
 458
 459         do_div(usage, NSEC_PER_USEC);
 460         do_div(utime, NSEC_PER_USEC);
 461         do_div(stime, NSEC_PER_USEC);
 462
 463         seq_printf(seq, "usage_usec %llu\n"
 464                    "user_usec %llu\n"
 465                    "system_usec %llu\n",
 466                    usage, utime, stime);
 467 }