src/proc_cpuview.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #ifndef _GNU_SOURCE
   4 #define _GNU_SOURCE
   5 #endif
   6
   7 #ifndef FUSE_USE_VERSION
   8 #define FUSE_USE_VERSION 26
   9 #endif
  10
  11 #define _FILE_OFFSET_BITS 64
  12
  13 #define __STDC_FORMAT_MACROS
  14 #include <dirent.h>
  15 #include <errno.h>
  16 #include <fcntl.h>
  17 #include <fuse.h>
  18 #include <inttypes.h>
  19 #include <libgen.h>
  20 #include <pthread.h>
  21 #include <sched.h>
  22 #include <stdarg.h>
  23 #include <stdbool.h>
  24 #include <stdint.h>
  25 #include <stdio.h>
  26 #include <stdlib.h>
  27 #include <string.h>
  28 #include <time.h>
  29 #include <unistd.h>
  30 #include <wait.h>
  31 #include <linux/magic.h>
  32 #include <linux/sched.h>
  33 #include <sys/epoll.h>
  34 #include <sys/mman.h>
  35 #include <sys/mount.h>
  36 #include <sys/param.h>
  37 #include <sys/socket.h>
  38 #include <sys/syscall.h>
  39 #include <sys/sysinfo.h>
  40 #include <sys/vfs.h>
  41
  42 #include "bindings.h"
  43 #include "config.h"
  44 #include "cgroup_fuse.h"
  45 #include "cpuset_parse.h"
  46 #include "cgroups/cgroup.h"
  47 #include "cgroups/cgroup_utils.h"
  48 #include "memory_utils.h"
  49 #include "proc_loadavg.h"
  50 #include "utils.h"
  51
  52 /* Data for CPU view */
  53 struct cg_proc_stat {
  54         char *cg;
  55         struct cpuacct_usage *usage;    /* Real usage as read from the host's /proc/stat. */
  56         struct cpuacct_usage *view;     /* Usage stats reported to the container. */
  57         int cpu_count;
  58         pthread_mutex_t lock;           /* For node manipulation. */
  59         struct cg_proc_stat *next;
  60 };
  61
  62 struct cg_proc_stat_head {
  63         struct cg_proc_stat *next;
  64         time_t lastcheck;
  65
  66         /*
  67          * For access to the list. Reading can be parallel, pruning is exclusive.
  68          */
  69         pthread_rwlock_t lock;
  70 };
  71
  72 #define CPUVIEW_HASH_SIZE 100
  73 static struct cg_proc_stat_head *proc_stat_history[CPUVIEW_HASH_SIZE];
  74
  75 static void reset_proc_stat_node(struct cg_proc_stat *node,
  76                                  struct cpuacct_usage *usage, int cpu_count)
  77 {
  78         lxcfs_debug("Resetting stat node for %s\n", node->cg);
  79         memcpy(node->usage, usage, sizeof(struct cpuacct_usage) * cpu_count);
  80
  81         for (int i = 0; i < cpu_count; i++) {
  82                 node->view[i].user = 0;
  83                 node->view[i].system = 0;
  84                 node->view[i].idle = 0;
  85         }
  86
  87         node->cpu_count = cpu_count;
  88 }
  89
  90 static bool expand_proc_stat_node(struct cg_proc_stat *node, int cpu_count)
  91 {
  92         __do_free struct cpuacct_usage *new_usage = NULL, *new_view = NULL;
  93
  94         /* Allocate new memory */
  95         new_usage = zalloc(sizeof(struct cpuacct_usage) * cpu_count);
  96         if (!new_usage)
  97                 return false;
  98
  99         new_view = zalloc(sizeof(struct cpuacct_usage) * cpu_count);
 100         if (!new_view)
 101                 return false;
 102
 103         /* Copy existing data & initialize new elements */
 104         for (int i = 0; i < cpu_count; i++) {
 105                 if (i < node->cpu_count) {
 106                         new_usage[i].user       = node->usage[i].user;
 107                         new_usage[i].system     = node->usage[i].system;
 108                         new_usage[i].idle       = node->usage[i].idle;
 109
 110                         new_view[i].user        = node->view[i].user;
 111                         new_view[i].system      = node->view[i].system;
 112                         new_view[i].idle        = node->view[i].idle;
 113                 }
 114         }
 115
 116         free(node->usage);
 117         node->usage = move_ptr(new_usage);
 118
 119         free(node->view);
 120         node->view = move_ptr(new_view);
 121         node->cpu_count = cpu_count;
 122
 123         return true;
 124 }
 125
 126 static void free_proc_stat_node(struct cg_proc_stat *node)
 127 {
 128         if (node) {
 129                 /*
 130                  * We're abusing the usage pointer to indicate that
 131                  * pthread_mutex_init() was successful. Don't judge me.
 132                  */
 133                 if (node->usage)
 134                         pthread_mutex_destroy(&node->lock);
 135                 free_disarm(node->cg);
 136                 free_disarm(node->usage);
 137                 free_disarm(node->view);
 138                 free_disarm(node);
 139         }
 140 }
 141
 142 define_cleanup_function(struct cg_proc_stat *, free_proc_stat_node);
 143
 144 static struct cg_proc_stat *add_proc_stat_node(struct cg_proc_stat *new_node)
 145 {
 146         call_cleaner(free_proc_stat_node) struct cg_proc_stat *new = new_node;
 147         struct cg_proc_stat *rv = new_node;
 148         int hash = calc_hash(new->cg) % CPUVIEW_HASH_SIZE;
 149         struct cg_proc_stat_head *head = proc_stat_history[hash];
 150         struct cg_proc_stat *cur;
 151
 152         pthread_rwlock_wrlock(&head->lock);
 153
 154         if (!head->next) {
 155                 head->next = move_ptr(new);
 156                 goto out_rwlock_unlock;
 157         }
 158
 159         cur = head->next;
 160
 161         for (;;) {
 162                 /*
 163                  * The node to be added is already present in the list, so
 164                  * free the newly allocated one and return the one we found.
 165                  */
 166                 if (strcmp(cur->cg, new->cg) == 0) {
 167                         rv = cur;
 168                         goto out_rwlock_unlock;
 169                 }
 170
 171                 /* Keep walking. */
 172                 if (cur->next) {
 173                         cur = cur->next;
 174                         continue;
 175                 }
 176
 177                 /* Add new node to end of list. */
 178                 cur->next = move_ptr(new);
 179                 goto out_rwlock_unlock;
 180         }
 181
 182 out_rwlock_unlock:
 183         pthread_rwlock_unlock(&head->lock);
 184         return move_ptr(rv);
 185 }
 186
 187 static struct cg_proc_stat *new_proc_stat_node(struct cpuacct_usage *usage,
 188                                                int cpu_count, const char *cg)
 189 {
 190         call_cleaner(free_proc_stat_node) struct cg_proc_stat *node = NULL;
 191         __do_free struct cpuacct_usage *new_usage = NULL;
 192
 193         node = zalloc(sizeof(struct cg_proc_stat));
 194         if (!node)
 195                 return NULL;
 196
 197         node->cg = strdup(cg);
 198         if (!node->cg)
 199                 return NULL;
 200
 201         new_usage = memdup(usage, sizeof(struct cpuacct_usage) * cpu_count);
 202         if (!new_usage)
 203                 return NULL;
 204
 205         node->view = zalloc(sizeof(struct cpuacct_usage) * cpu_count);
 206         if (!node->view)
 207                 return NULL;
 208
 209         node->cpu_count = cpu_count;
 210
 211         if (pthread_mutex_init(&node->lock, NULL))
 212                 return NULL;
 213         /*
 214          * We're abusing the usage pointer to indicate that
 215          * pthread_mutex_init() was successful. Don't judge me.
 216          */
 217         node->usage = move_ptr(new_usage);
 218
 219         return move_ptr(node);
 220 }
 221
 222 static bool cgfs_param_exist(const char *controller, const char *cgroup,
 223                              const char *file)
 224 {
 225         __do_free char *path = NULL;
 226         int cfd;
 227
 228         cfd = get_cgroup_fd(controller);
 229         if (cfd < 0)
 230                 return false;
 231
 232         path = must_make_path_relative(cgroup, file, NULL);
 233         return (faccessat(cfd, path, F_OK, 0) == 0);
 234 }
 235
 236 static struct cg_proc_stat *prune_proc_stat_list(struct cg_proc_stat *node)
 237 {
 238         struct cg_proc_stat *first = NULL;
 239
 240         for (struct cg_proc_stat *prev = NULL; node; ) {
 241                 if (!cgfs_param_exist("cpu", node->cg, "cpu.shares")) {
 242                         struct cg_proc_stat *tmp = node;
 243
 244                         lxcfs_debug("Removing stat node for %s\n", node->cg);
 245
 246                         if (prev)
 247                                 prev->next = node->next;
 248                         else
 249                                 first = node->next;
 250
 251                         node = node->next;
 252                         free_proc_stat_node(tmp);
 253                 } else {
 254                         if (!first)
 255                                 first = node;
 256                         prev = node;
 257                         node = node->next;
 258                 }
 259         }
 260
 261         return first;
 262 }
 263
 264 #define PROC_STAT_PRUNE_INTERVAL 10
 265 static void prune_proc_stat_history(void)
 266 {
 267         time_t now = time(NULL);
 268
 269         for (int i = 0; i < CPUVIEW_HASH_SIZE; i++) {
 270                 pthread_rwlock_wrlock(&proc_stat_history[i]->lock);
 271
 272                 if ((proc_stat_history[i]->lastcheck + PROC_STAT_PRUNE_INTERVAL) > now) {
 273                         pthread_rwlock_unlock(&proc_stat_history[i]->lock);
 274                         return;
 275                 }
 276
 277                 if (proc_stat_history[i]->next) {
 278                         proc_stat_history[i]->next = prune_proc_stat_list(proc_stat_history[i]->next);
 279                         proc_stat_history[i]->lastcheck = now;
 280                 }
 281
 282                 pthread_rwlock_unlock(&proc_stat_history[i]->lock);
 283         }
 284 }
 285
 286 static struct cg_proc_stat *find_proc_stat_node(struct cg_proc_stat_head *head,
 287                                                 const char *cg)
 288 {
 289         struct cg_proc_stat *node;
 290
 291         pthread_rwlock_rdlock(&head->lock);
 292
 293         if (!head->next) {
 294                 pthread_rwlock_unlock(&head->lock);
 295                 return NULL;
 296         }
 297
 298         node = head->next;
 299
 300         do {
 301                 if (strcmp(cg, node->cg) == 0)
 302                         goto out;
 303         } while ((node = node->next));
 304
 305         node = NULL;
 306
 307 out:
 308         pthread_rwlock_unlock(&head->lock);
 309         prune_proc_stat_history();
 310         return node;
 311 }
 312
 313 static struct cg_proc_stat *find_or_create_proc_stat_node(struct cpuacct_usage *usage, int cpu_count, const char *cg)
 314 {
 315         int hash = calc_hash(cg) % CPUVIEW_HASH_SIZE;
 316         struct cg_proc_stat_head *head = proc_stat_history[hash];
 317         struct cg_proc_stat *node;
 318
 319         node = find_proc_stat_node(head, cg);
 320         if (!node) {
 321                 node = new_proc_stat_node(usage, cpu_count, cg);
 322                 if (!node)
 323                         return NULL;
 324
 325                 node = add_proc_stat_node(node);
 326                 lxcfs_debug("New stat node (%d) for %s\n", cpu_count, cg);
 327         }
 328
 329         pthread_mutex_lock(&node->lock);
 330
 331         /* If additional CPUs on the host have been enabled, CPU usage counter
 332          * arrays have to be expanded */
 333         if (node->cpu_count < cpu_count) {
 334                 lxcfs_debug("Expanding stat node %d->%d for %s\n",
 335                                 node->cpu_count, cpu_count, cg);
 336
 337                 if (!expand_proc_stat_node(node, cpu_count)) {
 338                         pthread_mutex_unlock(&node->lock);
 339                         return log_debug(NULL, "Unable to expand stat node %d->%d for %s", node->cpu_count, cpu_count, cg);
 340                 }
 341         }
 342
 343         return node;
 344 }
 345
 346 static void add_cpu_usage(uint64_t *surplus, struct cpuacct_usage *usage,
 347                           uint64_t *counter, uint64_t threshold)
 348 {
 349         uint64_t free_space, to_add;
 350
 351         free_space = threshold - usage->user - usage->system;
 352
 353         if (free_space > usage->idle)
 354                 free_space = usage->idle;
 355
 356         to_add = free_space > *surplus ? *surplus : free_space;
 357
 358         *counter += to_add;
 359         usage->idle -= to_add;
 360         *surplus -= to_add;
 361 }
 362
 363 static uint64_t diff_cpu_usage(struct cpuacct_usage *older,
 364                                struct cpuacct_usage *newer,
 365                                struct cpuacct_usage *diff, int cpu_count)
 366 {
 367         uint64_t sum = 0;
 368
 369         for (int i = 0; i < cpu_count; i++) {
 370                 if (!newer[i].online)
 371                         continue;
 372
 373                 /*
 374                  * When cpuset is changed on the fly, the CPUs might get
 375                  * reordered. We could either reset all counters, or check
 376                  * that the substractions below will return expected results.
 377                  */
 378                 if (newer[i].user > older[i].user)
 379                         diff[i].user = newer[i].user - older[i].user;
 380                 else
 381                         diff[i].user = 0;
 382
 383                 if (newer[i].system > older[i].system)
 384                         diff[i].system = newer[i].system - older[i].system;
 385                 else
 386                         diff[i].system = 0;
 387
 388                 if (newer[i].idle > older[i].idle)
 389                         diff[i].idle = newer[i].idle - older[i].idle;
 390                 else
 391                         diff[i].idle = 0;
 392
 393                 sum += diff[i].user;
 394                 sum += diff[i].system;
 395                 sum += diff[i].idle;
 396         }
 397
 398         return sum;
 399 }
 400
 401 /*
 402  * Read cgroup CPU quota parameters from `cpu.cfs_quota_us` or
 403  * `cpu.cfs_period_us`, depending on `param`. Parameter value is returned
 404  * throuh `value`.
 405  */
 406 static bool read_cpu_cfs_param(const char *cg, const char *param, int64_t *value)
 407 {
 408         __do_free char *str = NULL;
 409         char file[11 + 6 + 1]; /* cpu.cfs__us + quota/period + \0 */
 410         bool first = true;
 411
 412         if (!pure_unified_layout(cgroup_ops)) {
 413                 snprintf(file, sizeof(file), "cpu.cfs_%s_us", param);
 414         } else {
 415                 strcpy(file, "cpu.max");
 416                 first = !strcmp(param, "quota");
 417         }
 418
 419         if (!cgroup_ops->get(cgroup_ops, "cpu", cg, file, &str))
 420                 return false;
 421
 422         if (sscanf(str, first ? "%" PRId64 : "%*d %" PRId64, value) != 1)
 423                 return false;
 424
 425         return true;
 426 }
 427
 428 /*
 429  * Return the exact number of visible CPUs based on CPU quotas.
 430  * If there is no quota set, zero is returned.
 431  */
 432 static double exact_cpu_count(const char *cg)
 433 {
 434         double rv;
 435         int nprocs;
 436         int64_t cfs_quota, cfs_period;
 437
 438         read_cpu_cfs_param(cg, "quota", &cfs_quota);
 439         read_cpu_cfs_param(cg, "period", &cfs_period);
 440
 441         if (cfs_quota <= 0 || cfs_period <= 0)
 442                 return 0;
 443
 444         rv = (double)cfs_quota / (double)cfs_period;
 445
 446         nprocs = get_nprocs();
 447
 448         if (rv > nprocs)
 449                 rv = nprocs;
 450
 451         return rv;
 452 }
 453
 454 /*
 455  * Return the maximum number of visible CPUs based on CPU quotas.
 456  * If there is no quota set, zero is returned.
 457  */
 458 int max_cpu_count(const char *cg)
 459 {
 460         __do_free char *cpuset = NULL;
 461         int rv, nprocs;
 462         int64_t cfs_quota, cfs_period;
 463         int nr_cpus_in_cpuset = 0;
 464
 465         read_cpu_cfs_param(cg, "quota", &cfs_quota);
 466         read_cpu_cfs_param(cg, "period", &cfs_period);
 467
 468         cpuset = get_cpuset(cg);
 469         if (cpuset)
 470                 nr_cpus_in_cpuset = cpu_number_in_cpuset(cpuset);
 471
 472         if (cfs_quota <= 0 || cfs_period <= 0){
 473                 if (nr_cpus_in_cpuset > 0)
 474                         return nr_cpus_in_cpuset;
 475
 476                 return 0;
 477         }
 478
 479         rv = cfs_quota / cfs_period;
 480
 481         /* In case quota/period does not yield a whole number, add one CPU for
 482          * the remainder.
 483          */
 484         if ((cfs_quota % cfs_period) > 0)
 485                 rv += 1;
 486
 487         nprocs = get_nprocs();
 488         if (rv > nprocs)
 489                 rv = nprocs;
 490
 491         /* use min value in cpu quota and cpuset */
 492         if (nr_cpus_in_cpuset > 0 && nr_cpus_in_cpuset < rv)
 493                 rv = nr_cpus_in_cpuset;
 494
 495         return rv;
 496 }
 497
 498 int cpuview_proc_stat(const char *cg, const char *cpuset,
 499                       struct cpuacct_usage *cg_cpu_usage, int cg_cpu_usage_size,
 500                       FILE *f, char *buf, size_t buf_size)
 501 {
 502         __do_free char *line = NULL;
 503         __do_free struct cpuacct_usage *diff = NULL;
 504         size_t linelen = 0, total_len = 0;
 505         int curcpu = -1; /* cpu numbering starts at 0 */
 506         int physcpu, i;
 507         int cpu_cnt = 0;
 508         uint64_t user = 0, nice = 0, system = 0, idle = 0, iowait = 0, irq = 0,
 509                  softirq = 0, steal = 0, guest = 0, guest_nice = 0;
 510         uint64_t user_sum = 0, system_sum = 0, idle_sum = 0;
 511         uint64_t user_surplus = 0, system_surplus = 0;
 512         int nprocs, max_cpus;
 513         ssize_t l;
 514         uint64_t total_sum, threshold;
 515         struct cg_proc_stat *stat_node;
 516
 517         nprocs = get_nprocs_conf();
 518         if (cg_cpu_usage_size < nprocs)
 519                 nprocs = cg_cpu_usage_size;
 520
 521         /* Read all CPU stats and stop when we've encountered other lines */
 522         while (getline(&line, &linelen, f) != -1) {
 523                 int ret;
 524                 char cpu_char[10]; /* That's a lot of cores */
 525                 uint64_t all_used, cg_used;
 526
 527                 if (strlen(line) == 0)
 528                         continue;
 529
 530                 /* not a ^cpuN line containing a number N */
 531                 if (sscanf(line, "cpu%9[^ ]", cpu_char) != 1)
 532                         break;
 533
 534                 if (sscanf(cpu_char, "%d", &physcpu) != 1)
 535                         continue;
 536
 537                 if (physcpu >= cg_cpu_usage_size)
 538                         continue;
 539
 540                 curcpu++;
 541                 cpu_cnt++;
 542
 543                 if (!cpu_in_cpuset(physcpu, cpuset)) {
 544                         for (i = curcpu; i <= physcpu; i++)
 545                                 cg_cpu_usage[i].online = false;
 546                         continue;
 547                 }
 548
 549                 if (curcpu < physcpu) {
 550                         /* Some CPUs may be disabled */
 551                         for (i = curcpu; i < physcpu; i++)
 552                                 cg_cpu_usage[i].online = false;
 553
 554                         curcpu = physcpu;
 555                 }
 556
 557                 cg_cpu_usage[curcpu].online = true;
 558
 559                 ret = sscanf(line, "%*s %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 " %" PRIu64 "lu",
 560                            &user,
 561                            &nice,
 562                            &system,
 563                            &idle,
 564                            &iowait,
 565                            &irq,
 566                            &softirq,
 567                            &steal,
 568                            &guest,
 569                            &guest_nice);
 570                 if (ret != 10)
 571                         continue;
 572
 573                 all_used = user + nice + system + iowait + irq + softirq + steal + guest + guest_nice;
 574                 cg_used = cg_cpu_usage[curcpu].user + cg_cpu_usage[curcpu].system;
 575
 576                 if (all_used >= cg_used) {
 577                         cg_cpu_usage[curcpu].idle = idle + (all_used - cg_used);
 578
 579                 } else {
 580                         lxcfs_error("cpu%d from %s has unexpected cpu time: %" PRIu64 " in /proc/stat, %" PRIu64 " in cpuacct.usage_all; unable to determine idle time",
 581                                     curcpu, cg, all_used, cg_used);
 582                         cg_cpu_usage[curcpu].idle = idle;
 583                 }
 584         }
 585
 586         /* Cannot use more CPUs than is available in cpuset. */
 587         max_cpus = max_cpu_count(cg);
 588         if (max_cpus > cpu_cnt || !max_cpus)
 589                 max_cpus = cpu_cnt;
 590
 591         stat_node = find_or_create_proc_stat_node(cg_cpu_usage, nprocs, cg);
 592         if (!stat_node)
 593                 return log_error(0, "Failed to find/create stat node for %s", cg);
 594
 595         diff = malloc(sizeof(struct cpuacct_usage) * nprocs);
 596         if (!diff)
 597                 return 0;
 598
 599         /*
 600          * If the new values are LOWER than values stored in memory, it means
 601          * the cgroup has been reset/recreated and we should reset too.
 602          */
 603         for (curcpu = 0; curcpu < nprocs; curcpu++) {
 604                 if (!cg_cpu_usage[curcpu].online)
 605                         continue;
 606
 607                 if (cg_cpu_usage[curcpu].user < stat_node->usage[curcpu].user)
 608                         reset_proc_stat_node(stat_node, cg_cpu_usage, nprocs);
 609
 610                 break;
 611         }
 612
 613         total_sum = diff_cpu_usage(stat_node->usage, cg_cpu_usage, diff, nprocs);
 614
 615         for (curcpu = 0, i = -1; curcpu < nprocs; curcpu++) {
 616                 stat_node->usage[curcpu].online = cg_cpu_usage[curcpu].online;
 617
 618                 if (!stat_node->usage[curcpu].online)
 619                         continue;
 620
 621                 i++;
 622
 623                 stat_node->usage[curcpu].user += diff[curcpu].user;
 624                 stat_node->usage[curcpu].system += diff[curcpu].system;
 625                 stat_node->usage[curcpu].idle += diff[curcpu].idle;
 626
 627                 if (max_cpus > 0 && i >= max_cpus) {
 628                         user_surplus += diff[curcpu].user;
 629                         system_surplus += diff[curcpu].system;
 630                 }
 631         }
 632
 633         /* Calculate usage counters of visible CPUs */
 634         if (max_cpus > 0) {
 635                 uint64_t diff_user = 0;
 636                 uint64_t diff_system = 0;
 637                 uint64_t diff_idle = 0;
 638                 uint64_t max_diff_idle = 0;
 639                 uint64_t max_diff_idle_index = 0;
 640                 double exact_cpus;
 641
 642                 /* threshold = maximum usage per cpu, including idle */
 643                 threshold = total_sum / cpu_cnt * max_cpus;
 644
 645                 for (curcpu = 0, i = -1; curcpu < nprocs; curcpu++) {
 646                         if (!stat_node->usage[curcpu].online)
 647                                 continue;
 648
 649                         i++;
 650
 651                         if (i == max_cpus)
 652                                 break;
 653
 654                         if (diff[curcpu].user + diff[curcpu].system >= threshold)
 655                                 continue;
 656
 657                         /* Add user */
 658                         add_cpu_usage(&user_surplus, &diff[curcpu],
 659                                       &diff[curcpu].user, threshold);
 660
 661                         if (diff[curcpu].user + diff[curcpu].system >= threshold)
 662                                 continue;
 663
 664                         /* If there is still room, add system */
 665                         add_cpu_usage(&system_surplus, &diff[curcpu],
 666                                       &diff[curcpu].system, threshold);
 667                 }
 668
 669                 if (user_surplus > 0)
 670                         lxcfs_debug("leftover user: %lu for %s\n", user_surplus, cg);
 671                 if (system_surplus > 0)
 672                         lxcfs_debug("leftover system: %lu for %s\n", system_surplus, cg);
 673
 674                 for (curcpu = 0, i = -1; curcpu < nprocs; curcpu++) {
 675                         if (!stat_node->usage[curcpu].online)
 676                                 continue;
 677
 678                         i++;
 679
 680                         if (i == max_cpus)
 681                                 break;
 682
 683                         stat_node->view[curcpu].user += diff[curcpu].user;
 684                         stat_node->view[curcpu].system += diff[curcpu].system;
 685                         stat_node->view[curcpu].idle += diff[curcpu].idle;
 686
 687                         user_sum += stat_node->view[curcpu].user;
 688                         system_sum += stat_node->view[curcpu].system;
 689                         idle_sum += stat_node->view[curcpu].idle;
 690
 691                         diff_user += diff[curcpu].user;
 692                         diff_system += diff[curcpu].system;
 693                         diff_idle += diff[curcpu].idle;
 694                         if (diff[curcpu].idle > max_diff_idle) {
 695                                 max_diff_idle = diff[curcpu].idle;
 696                                 max_diff_idle_index = curcpu;
 697                         }
 698
 699                         lxcfs_v("curcpu: %d, diff_user: %lu, diff_system: %lu, diff_idle: %lu\n", curcpu, diff[curcpu].user, diff[curcpu].system, diff[curcpu].idle);
 700                 }
 701                 lxcfs_v("total. diff_user: %lu, diff_system: %lu, diff_idle: %lu\n", diff_user, diff_system, diff_idle);
 702
 703                 /* revise cpu usage view to support partial cpu case. */
 704                 exact_cpus = exact_cpu_count(cg);
 705                 if (exact_cpus < (double)max_cpus){
 706                         uint64_t delta = (uint64_t)((double)(diff_user + diff_system + diff_idle) * (1 - exact_cpus / (double)max_cpus));
 707
 708                         lxcfs_v("revising cpu usage view to match the exact cpu count [%f]\n", exact_cpus);
 709                         lxcfs_v("delta: %lu\n", delta);
 710                         lxcfs_v("idle_sum before: %lu\n", idle_sum);
 711                         idle_sum = idle_sum > delta ? idle_sum - delta : 0;
 712                         lxcfs_v("idle_sum after: %lu\n", idle_sum);
 713
 714                         curcpu = max_diff_idle_index;
 715                         lxcfs_v("curcpu: %d, idle before: %lu\n", curcpu, stat_node->view[curcpu].idle);
 716                         stat_node->view[curcpu].idle = stat_node->view[curcpu].idle > delta ? stat_node->view[curcpu].idle - delta : 0;
 717                         lxcfs_v("curcpu: %d, idle after: %lu\n", curcpu, stat_node->view[curcpu].idle);
 718                 }
 719         } else {
 720                 for (curcpu = 0; curcpu < nprocs; curcpu++) {
 721                         if (!stat_node->usage[curcpu].online)
 722                                 continue;
 723
 724                         stat_node->view[curcpu].user = stat_node->usage[curcpu].user;
 725                         stat_node->view[curcpu].system = stat_node->usage[curcpu].system;
 726                         stat_node->view[curcpu].idle = stat_node->usage[curcpu].idle;
 727
 728                         user_sum += stat_node->view[curcpu].user;
 729                         system_sum += stat_node->view[curcpu].system;
 730                         idle_sum += stat_node->view[curcpu].idle;
 731                 }
 732         }
 733
 734         /* Render the file */
 735         /* cpu-all */
 736         l = snprintf(buf, buf_size,
 737                      "cpu  %" PRIu64 " 0 %" PRIu64 " %" PRIu64 " 0 0 0 0 0 0\n",
 738                      user_sum, system_sum, idle_sum);
 739         lxcfs_v("cpu-all: %s\n", buf);
 740         if (l < 0)
 741                 return log_error(0, "Failed to write cache");
 742         if (l >= buf_size)
 743                 return log_error(0, "Write to cache was truncated");
 744
 745         buf += l;
 746         buf_size -= l;
 747         total_len += l;
 748
 749         /* Render visible CPUs */
 750         for (curcpu = 0, i = -1; curcpu < nprocs; curcpu++) {
 751                 if (!stat_node->usage[curcpu].online)
 752                         continue;
 753
 754                 i++;
 755
 756                 if (max_cpus > 0 && i == max_cpus)
 757                         break;
 758
 759                 l = snprintf(buf, buf_size, "cpu%d %" PRIu64 " 0 %" PRIu64 " %" PRIu64 " 0 0 0 0 0 0\n",
 760                              i,
 761                              stat_node->view[curcpu].user,
 762                              stat_node->view[curcpu].system,
 763                              stat_node->view[curcpu].idle);
 764                 lxcfs_v("cpu: %s\n", buf);
 765                 if (l < 0)
 766                         return log_error(0, "Failed to write cache");
 767                 if (l >= buf_size)
 768                         return log_error(0, "Write to cache was truncated");
 769
 770                 buf += l;
 771                 buf_size -= l;
 772                 total_len += l;
 773         }
 774
 775         /* Pass the rest of /proc/stat, start with the last line read */
 776         l = snprintf(buf, buf_size, "%s", line);
 777         if (l < 0)
 778                 return log_error(0, "Failed to write cache");
 779         if (l >= buf_size)
 780                 return log_error(0, "Write to cache was truncated");
 781
 782         buf += l;
 783         buf_size -= l;
 784         total_len += l;
 785
 786         /* Pass the rest of the host's /proc/stat */
 787         while (getline(&line, &linelen, f) != -1) {
 788                 l = snprintf(buf, buf_size, "%s", line);
 789                 if (l < 0)
 790                         return log_error(0, "Failed to write cache");
 791                 if (l >= buf_size)
 792                         return log_error(0, "Write to cache was truncated");
 793
 794                 buf += l;
 795                 buf_size -= l;
 796                 total_len += l;
 797         }
 798
 799         if (stat_node)
 800                 pthread_mutex_unlock(&stat_node->lock);
 801
 802         return total_len;
 803 }
 804
 805 /*
 806  * check whether this is a '^processor" line in /proc/cpuinfo
 807  */
 808 static inline bool is_processor_line(const char *line)
 809 {
 810         int cpu;
 811         return sscanf(line, "processor       : %d", &cpu) == 1;
 812 }
 813
 814 static inline bool cpuline_in_cpuset(const char *line, const char *cpuset)
 815 {
 816         int cpu;
 817
 818         if (sscanf(line, "processor       : %d", &cpu) == 1)
 819                 return cpu_in_cpuset(cpu, cpuset);
 820
 821         return false;
 822 }
 823
 824 int proc_cpuinfo_read(char *buf, size_t size, off_t offset,
 825                       struct fuse_file_info *fi)
 826 {
 827         __do_free char *cg = NULL, *cpuset = NULL, *line = NULL;
 828         __do_free void *fopen_cache = NULL;
 829         __do_fclose FILE *f = NULL;
 830         struct fuse_context *fc = fuse_get_context();
 831         struct lxcfs_opts *opts = (struct lxcfs_opts *)fc->private_data;
 832         struct file_info *d = INTTYPE_TO_PTR(fi->fh);
 833         size_t linelen = 0, total_len = 0;
 834         bool am_printing = false, firstline = true, is_s390x = false;
 835         int curcpu = -1, cpu, max_cpus = 0;
 836         bool use_view;
 837         char *cache = d->buf;
 838         size_t cache_size = d->buflen;
 839
 840         if (offset) {
 841                 int left;
 842
 843                 if (offset > d->size)
 844                         return -EINVAL;
 845
 846                 if (!d->cached)
 847                         return 0;
 848
 849                 left = d->size - offset;
 850                 total_len = left > size ? size: left;
 851                 memcpy(buf, cache + offset, total_len);
 852
 853                 return total_len;
 854         }
 855
 856         pid_t initpid = lookup_initpid_in_store(fc->pid);
 857         if (initpid <= 1 || is_shared_pidns(initpid))
 858                 initpid = fc->pid;
 859
 860         cg = get_pid_cgroup(initpid, "cpuset");
 861         if (!cg)
 862                 return read_file_fuse("proc/cpuinfo", buf, size, d);
 863         prune_init_slice(cg);
 864
 865         cpuset = get_cpuset(cg);
 866         if (!cpuset)
 867                 return 0;
 868
 869         if (cgroup_ops->can_use_cpuview(cgroup_ops) && opts && opts->use_cfs)
 870                 use_view = true;
 871         else
 872                 use_view = false;
 873         if (use_view)
 874                 max_cpus = max_cpu_count(cg);
 875
 876         f = fopen_cached("/proc/cpuinfo", "re", &fopen_cache);
 877         if (!f)
 878                 return 0;
 879
 880         while (getline(&line, &linelen, f) != -1) {
 881                 ssize_t l;
 882                 if (firstline) {
 883                         firstline = false;
 884                         if (strstr(line, "IBM/S390") != NULL) {
 885                                 is_s390x = true;
 886                                 am_printing = true;
 887                                 continue;
 888                         }
 889                 }
 890
 891                 if (strncmp(line, "# processors:", 12) == 0)
 892                         continue;
 893
 894                 if (is_processor_line(line)) {
 895                         if (use_view && max_cpus > 0 && (curcpu + 1) == max_cpus)
 896                                 break;
 897
 898                         am_printing = cpuline_in_cpuset(line, cpuset);
 899                         if (am_printing) {
 900                                 curcpu++;
 901                                 l = snprintf(cache, cache_size, "processor      : %d\n", curcpu);
 902                                 if (l < 0)
 903                                         return log_error(0, "Failed to write cache");
 904                                 if (l >= cache_size)
 905                                         return log_error(0, "Write to cache was truncated");
 906                                 cache += l;
 907                                 cache_size -= l;
 908                                 total_len += l;
 909                         }
 910                         continue;
 911                 } else if (is_s390x && sscanf(line, "processor %d:", &cpu) == 1) {
 912                         char *p;
 913
 914                         if (use_view && max_cpus > 0 && (curcpu + 1) == max_cpus)
 915                                 break;
 916
 917                         if (!cpu_in_cpuset(cpu, cpuset))
 918                                 continue;
 919
 920                         curcpu ++;
 921                         p = strchr(line, ':');
 922                         if (!p || !*p)
 923                                 return 0;
 924                         p++;
 925
 926                         l = snprintf(cache, cache_size, "processor %d:%s", curcpu, p);
 927                         if (l < 0)
 928                                 return log_error(0, "Failed to write cache");
 929                         if (l >= cache_size)
 930                                 return log_error(0, "Write to cache was truncated");
 931
 932                         cache += l;
 933                         cache_size -= l;
 934                         total_len += l;
 935                         continue;
 936
 937                 }
 938                 if (am_printing) {
 939                         l = snprintf(cache, cache_size, "%s", line);
 940                         if (l < 0)
 941                                 return log_error(0, "Failed to write cache");
 942                         if (l >= cache_size)
 943                                 return log_error(0, "Write to cache was truncated");
 944
 945                         cache += l;
 946                         cache_size -= l;
 947                         total_len += l;
 948                 }
 949         }
 950
 951         if (is_s390x) {
 952                 __do_free char *origcache = d->buf;
 953                 ssize_t l;
 954
 955                 d->buf = malloc(d->buflen);
 956                 if (!d->buf) {
 957                         d->buf = move_ptr(origcache);
 958                         return 0;
 959                 }
 960
 961                 cache = d->buf;
 962                 cache_size = d->buflen;
 963                 total_len = 0;
 964                 l = snprintf(cache, cache_size, "vendor_id       : IBM/S390\n");
 965                 if (l < 0 || l >= cache_size)
 966                         return 0;
 967
 968                 cache_size -= l;
 969                 cache += l;
 970                 total_len += l;
 971                 l = snprintf(cache, cache_size, "# processors    : %d\n", curcpu + 1);
 972                 if (l < 0 || l >= cache_size)
 973                         return 0;
 974
 975                 cache_size -= l;
 976                 cache += l;
 977                 total_len += l;
 978                 l = snprintf(cache, cache_size, "%s", origcache);
 979                 if (l < 0 || l >= cache_size)
 980                         return 0;
 981                 total_len += l;
 982         }
 983
 984         d->cached = 1;
 985         d->size = total_len;
 986         if (total_len > size)
 987                 total_len = size;
 988
 989         /* read from off 0 */
 990         memcpy(buf, d->buf, total_len);
 991
 992         return total_len;
 993 }
 994
 995 /*
 996  * Returns 0 on success.
 997  * It is the caller's responsibility to free `return_usage`, unless this
 998  * function returns an error.
 999  */
1000 int read_cpuacct_usage_all(char *cg, char *cpuset,
1001                            struct cpuacct_usage **return_usage, int *size)
1002 {
1003         __do_free char *usage_str = NULL;
1004         __do_free struct cpuacct_usage *cpu_usage = NULL;
1005         int i = 0, j = 0, read_pos = 0, read_cnt = 0;
1006         int cpucount;
1007         int ret;
1008         int cg_cpu;
1009         uint64_t cg_user, cg_system;
1010         int64_t ticks_per_sec;
1011
1012         ticks_per_sec = sysconf(_SC_CLK_TCK);
1013         if (ticks_per_sec < 0 && errno == EINVAL) {
1014                 lxcfs_debug("%m - Failed to determine number of ticks per second");
1015                 return -1;
1016         }
1017
1018         cpucount = get_nprocs_conf();
1019         cpu_usage = malloc(sizeof(struct cpuacct_usage) * cpucount);
1020         if (!cpu_usage)
1021                 return -ENOMEM;
1022
1023         memset(cpu_usage, 0, sizeof(struct cpuacct_usage) * cpucount);
1024         if (!cgroup_ops->get(cgroup_ops, "cpuacct", cg, "cpuacct.usage_all", &usage_str)) {
1025                 char *sep = " \t\n";
1026                 char *tok;
1027
1028                 /* Read cpuacct.usage_percpu instead. */
1029                 lxcfs_debug("Falling back to cpuacct.usage_percpu");
1030                 if (!cgroup_ops->get(cgroup_ops, "cpuacct", cg, "cpuacct.usage_percpu", &usage_str))
1031                         return -1;
1032
1033                 lxc_iterate_parts(tok, usage_str, sep) {
1034                         uint64_t percpu_user;
1035
1036                         if (i >= cpucount)
1037                                 break;
1038
1039                         tok = trim_whitespace_in_place(tok);
1040                         ret = safe_uint64(tok, &percpu_user, 10);
1041                         if (ret)
1042                                 return -1;
1043
1044                         /* Convert the time from nanoseconds to USER_HZ */
1045                         cpu_usage[i].user = percpu_user / 1000.0 / 1000 / 1000 * ticks_per_sec;
1046                         cpu_usage[i].system = cpu_usage[i].user;
1047                         i++;
1048                         lxcfs_debug("cpu%d with time %s", i, tok);
1049                 }
1050         } else {
1051                 if (sscanf(usage_str, "cpu user system\n%n", &read_cnt) != 0)
1052                         return log_error(-1, "read_cpuacct_usage_all reading first line from %s/cpuacct.usage_all failed", cg);
1053
1054                 read_pos += read_cnt;
1055
1056                 for (i = 0, j = 0; i < cpucount; i++) {
1057                         ret = sscanf(usage_str + read_pos,
1058                                         "%d %" PRIu64 " %" PRIu64 "\n%n", &cg_cpu,
1059                                         &cg_user, &cg_system, &read_cnt);
1060
1061                         if (ret == EOF)
1062                                 break;
1063
1064                         if (ret != 3)
1065                                 return log_error(-EINVAL, "Failed to parse cpuacct.usage_all line %s from cgroup %s",
1066                                                 usage_str + read_pos, cg);
1067
1068                         read_pos += read_cnt;
1069
1070                         /* Convert the time from nanoseconds to USER_HZ */
1071                         cpu_usage[j].user = cg_user / 1000.0 / 1000 / 1000 * ticks_per_sec;
1072                         cpu_usage[j].system = cg_system / 1000.0 / 1000 / 1000 * ticks_per_sec;
1073                         j++;
1074                 }
1075         }
1076
1077         *return_usage = move_ptr(cpu_usage);
1078         *size = cpucount;
1079         return 0;
1080 }
1081
1082 static bool cpuview_init_head(struct cg_proc_stat_head **head)
1083 {
1084         *head = malloc(sizeof(struct cg_proc_stat_head));
1085         if (!(*head))
1086                 return log_error(false, "%s", strerror(errno));
1087
1088         (*head)->lastcheck = time(NULL);
1089         (*head)->next = NULL;
1090
1091         if (pthread_rwlock_init(&(*head)->lock, NULL) != 0) {
1092                 free_disarm(*head);
1093                 return log_error(false, "Failed to initialize list lock");
1094         }
1095
1096         return true;
1097 }
1098
1099 bool init_cpuview(void)
1100 {
1101         int i;
1102
1103         for (i = 0; i < CPUVIEW_HASH_SIZE; i++)
1104                 proc_stat_history[i] = NULL;
1105
1106         for (i = 0; i < CPUVIEW_HASH_SIZE; i++) {
1107                 if (!cpuview_init_head(&proc_stat_history[i]))
1108                         goto err;
1109         }
1110
1111         return true;
1112
1113 err:
1114         for (i = 0; i < CPUVIEW_HASH_SIZE; i++) {
1115                 if (proc_stat_history[i])
1116                         free_disarm(proc_stat_history[i]);
1117         }
1118
1119         return false;
1120 }
1121
1122 static void cpuview_free_head(struct cg_proc_stat_head *head)
1123 {
1124         struct cg_proc_stat *node;
1125
1126         if (head->next) {
1127                 node = head->next;
1128
1129                 for (;;) {
1130                         struct cg_proc_stat *cur = node;
1131                         node = node->next;
1132                         free_proc_stat_node(cur);
1133                         if (!node)
1134                                 break;
1135                 }
1136         }
1137
1138         pthread_rwlock_destroy(&head->lock);
1139         free_disarm(head);
1140 }
1141
1142 void free_cpuview(void)
1143 {
1144         for (int i = 0; i < CPUVIEW_HASH_SIZE; i++)
1145                 if (proc_stat_history[i])
1146                         cpuview_free_head(proc_stat_history[i]);
1147 }