tools/perf/bench/numa.c

   1 /*
   2  * numa.c
   3  *
   4  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5  */
   6
   7 #include "../perf.h"
   8 #include "../builtin.h"
   9 #include "../util/util.h"
  10 #include <subcmd/parse-options.h>
  11 #include "../util/cloexec.h"
  12
  13 #include "bench.h"
  14
  15 #include <errno.h>
  16 #include <sched.h>
  17 #include <stdio.h>
  18 #include <assert.h>
  19 #include <malloc.h>
  20 #include <signal.h>
  21 #include <stdlib.h>
  22 #include <string.h>
  23 #include <unistd.h>
  24 #include <pthread.h>
  25 #include <sys/mman.h>
  26 #include <sys/time.h>
  27 #include <sys/resource.h>
  28 #include <sys/wait.h>
  29 #include <sys/prctl.h>
  30 #include <sys/types.h>
  31
  32 #include <numa.h>
  33 #include <numaif.h>
  34
  35 /*
  36  * Regular printout to the terminal, supressed if -q is specified:
  37  */
  38 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  39
  40 /*
  41  * Debug printf:
  42  */
  43 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  44
  45 struct thread_data {
  46         int                     curr_cpu;
  47         cpu_set_t               bind_cpumask;
  48         int                     bind_node;
  49         u8                      *process_data;
  50         int                     process_nr;
  51         int                     thread_nr;
  52         int                     task_nr;
  53         unsigned int            loops_done;
  54         u64                     val;
  55         u64                     runtime_ns;
  56         u64                     system_time_ns;
  57         u64                     user_time_ns;
  58         double                  speed_gbs;
  59         pthread_mutex_t         *process_lock;
  60 };
  61
  62 /* Parameters set by options: */
  63
  64 struct params {
  65         /* Startup synchronization: */
  66         bool                    serialize_startup;
  67
  68         /* Task hierarchy: */
  69         int                     nr_proc;
  70         int                     nr_threads;
  71
  72         /* Working set sizes: */
  73         const char              *mb_global_str;
  74         const char              *mb_proc_str;
  75         const char              *mb_proc_locked_str;
  76         const char              *mb_thread_str;
  77
  78         double                  mb_global;
  79         double                  mb_proc;
  80         double                  mb_proc_locked;
  81         double                  mb_thread;
  82
  83         /* Access patterns to the working set: */
  84         bool                    data_reads;
  85         bool                    data_writes;
  86         bool                    data_backwards;
  87         bool                    data_zero_memset;
  88         bool                    data_rand_walk;
  89         u32                     nr_loops;
  90         u32                     nr_secs;
  91         u32                     sleep_usecs;
  92
  93         /* Working set initialization: */
  94         bool                    init_zero;
  95         bool                    init_random;
  96         bool                    init_cpu0;
  97
  98         /* Misc options: */
  99         int                     show_details;
 100         int                     run_all;
 101         int                     thp;
 102
 103         long                    bytes_global;
 104         long                    bytes_process;
 105         long                    bytes_process_locked;
 106         long                    bytes_thread;
 107
 108         int                     nr_tasks;
 109         bool                    show_quiet;
 110
 111         bool                    show_convergence;
 112         bool                    measure_convergence;
 113
 114         int                     perturb_secs;
 115         int                     nr_cpus;
 116         int                     nr_nodes;
 117
 118         /* Affinity options -C and -N: */
 119         char                    *cpu_list_str;
 120         char                    *node_list_str;
 121 };
 122
 123
 124 /* Global, read-writable area, accessible to all processes and threads: */
 125
 126 struct global_info {
 127         u8                      *data;
 128
 129         pthread_mutex_t         startup_mutex;
 130         int                     nr_tasks_started;
 131
 132         pthread_mutex_t         startup_done_mutex;
 133
 134         pthread_mutex_t         start_work_mutex;
 135         int                     nr_tasks_working;
 136
 137         pthread_mutex_t         stop_work_mutex;
 138         u64                     bytes_done;
 139
 140         struct thread_data      *threads;
 141
 142         /* Convergence latency measurement: */
 143         bool                    all_converged;
 144         bool                    stop_work;
 145
 146         int                     print_once;
 147
 148         struct params           p;
 149 };
 150
 151 static struct global_info       *g = NULL;
 152
 153 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 154 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 155
 156 struct params p0;
 157
 158 static const struct option options[] = {
 159         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 160         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 161
 162         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 163         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 164         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 165         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 166
 167         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 168         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 169         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 170
 171         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via writes (can be mixed with -W)"),
 172         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 173         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 174         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 175         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 176
 177
 178         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 179         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 180         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 181         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 182
 183         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 184         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 185         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 186         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
 187         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 188         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 189         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 190
 191         /* Special option string parsing callbacks: */
 192         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 193                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 194                         parse_cpus_opt),
 195         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 196                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 197                         parse_nodes_opt),
 198         OPT_END()
 199 };
 200
 201 static const char * const bench_numa_usage[] = {
 202         "perf bench numa <options>",
 203         NULL
 204 };
 205
 206 static const char * const numa_usage[] = {
 207         "perf bench numa mem [<options>]",
 208         NULL
 209 };
 210
 211 static cpu_set_t bind_to_cpu(int target_cpu)
 212 {
 213         cpu_set_t orig_mask, mask;
 214         int ret;
 215
 216         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 217         BUG_ON(ret);
 218
 219         CPU_ZERO(&mask);
 220
 221         if (target_cpu == -1) {
 222                 int cpu;
 223
 224                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 225                         CPU_SET(cpu, &mask);
 226         } else {
 227                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 228                 CPU_SET(target_cpu, &mask);
 229         }
 230
 231         ret = sched_setaffinity(0, sizeof(mask), &mask);
 232         BUG_ON(ret);
 233
 234         return orig_mask;
 235 }
 236
 237 static cpu_set_t bind_to_node(int target_node)
 238 {
 239         int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 240         cpu_set_t orig_mask, mask;
 241         int cpu;
 242         int ret;
 243
 244         BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 245         BUG_ON(!cpus_per_node);
 246
 247         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 248         BUG_ON(ret);
 249
 250         CPU_ZERO(&mask);
 251
 252         if (target_node == -1) {
 253                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 254                         CPU_SET(cpu, &mask);
 255         } else {
 256                 int cpu_start = (target_node + 0) * cpus_per_node;
 257                 int cpu_stop  = (target_node + 1) * cpus_per_node;
 258
 259                 BUG_ON(cpu_stop > g->p.nr_cpus);
 260
 261                 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 262                         CPU_SET(cpu, &mask);
 263         }
 264
 265         ret = sched_setaffinity(0, sizeof(mask), &mask);
 266         BUG_ON(ret);
 267
 268         return orig_mask;
 269 }
 270
 271 static void bind_to_cpumask(cpu_set_t mask)
 272 {
 273         int ret;
 274
 275         ret = sched_setaffinity(0, sizeof(mask), &mask);
 276         BUG_ON(ret);
 277 }
 278
 279 static void mempol_restore(void)
 280 {
 281         int ret;
 282
 283         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 284
 285         BUG_ON(ret);
 286 }
 287
 288 static void bind_to_memnode(int node)
 289 {
 290         unsigned long nodemask;
 291         int ret;
 292
 293         if (node == -1)
 294                 return;
 295
 296         BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask));
 297         nodemask = 1L << node;
 298
 299         ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 300         dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 301
 302         BUG_ON(ret);
 303 }
 304
 305 #define HPSIZE (2*1024*1024)
 306
 307 #define set_taskname(fmt...)                            \
 308 do {                                                    \
 309         char name[20];                                  \
 310                                                         \
 311         snprintf(name, 20, fmt);                        \
 312         prctl(PR_SET_NAME, name);                       \
 313 } while (0)
 314
 315 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 316                       int init_zero, int init_cpu0, int thp, int init_random)
 317 {
 318         cpu_set_t orig_mask;
 319         ssize_t bytes;
 320         u8 *buf;
 321         int ret;
 322
 323         if (!bytes0)
 324                 return NULL;
 325
 326         /* Allocate and initialize all memory on CPU#0: */
 327         if (init_cpu0) {
 328                 orig_mask = bind_to_node(0);
 329                 bind_to_memnode(0);
 330         }
 331
 332         bytes = bytes0 + HPSIZE;
 333
 334         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 335         BUG_ON(buf == (void *)-1);
 336
 337         if (map_flags == MAP_PRIVATE) {
 338                 if (thp > 0) {
 339                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 340                         if (ret && !g->print_once) {
 341                                 g->print_once = 1;
 342                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 343                         }
 344                 }
 345                 if (thp < 0) {
 346                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 347                         if (ret && !g->print_once) {
 348                                 g->print_once = 1;
 349                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 350                         }
 351                 }
 352         }
 353
 354         if (init_zero) {
 355                 bzero(buf, bytes);
 356         } else {
 357                 /* Initialize random contents, different in each word: */
 358                 if (init_random) {
 359                         u64 *wbuf = (void *)buf;
 360                         long off = rand();
 361                         long i;
 362
 363                         for (i = 0; i < bytes/8; i++)
 364                                 wbuf[i] = i + off;
 365                 }
 366         }
 367
 368         /* Align to 2MB boundary: */
 369         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 370
 371         /* Restore affinity: */
 372         if (init_cpu0) {
 373                 bind_to_cpumask(orig_mask);
 374                 mempol_restore();
 375         }
 376
 377         return buf;
 378 }
 379
 380 static void free_data(void *data, ssize_t bytes)
 381 {
 382         int ret;
 383
 384         if (!data)
 385                 return;
 386
 387         ret = munmap(data, bytes);
 388         BUG_ON(ret);
 389 }
 390
 391 /*
 392  * Create a shared memory buffer that can be shared between processes, zeroed:
 393  */
 394 static void * zalloc_shared_data(ssize_t bytes)
 395 {
 396         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 397 }
 398
 399 /*
 400  * Create a shared memory buffer that can be shared between processes:
 401  */
 402 static void * setup_shared_data(ssize_t bytes)
 403 {
 404         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 405 }
 406
 407 /*
 408  * Allocate process-local memory - this will either be shared between
 409  * threads of this process, or only be accessed by this thread:
 410  */
 411 static void * setup_private_data(ssize_t bytes)
 412 {
 413         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 414 }
 415
 416 /*
 417  * Return a process-shared (global) mutex:
 418  */
 419 static void init_global_mutex(pthread_mutex_t *mutex)
 420 {
 421         pthread_mutexattr_t attr;
 422
 423         pthread_mutexattr_init(&attr);
 424         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 425         pthread_mutex_init(mutex, &attr);
 426 }
 427
 428 static int parse_cpu_list(const char *arg)
 429 {
 430         p0.cpu_list_str = strdup(arg);
 431
 432         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 433
 434         return 0;
 435 }
 436
 437 static int parse_setup_cpu_list(void)
 438 {
 439         struct thread_data *td;
 440         char *str0, *str;
 441         int t;
 442
 443         if (!g->p.cpu_list_str)
 444                 return 0;
 445
 446         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 447
 448         str0 = str = strdup(g->p.cpu_list_str);
 449         t = 0;
 450
 451         BUG_ON(!str);
 452
 453         tprintf("# binding tasks to CPUs:\n");
 454         tprintf("#  ");
 455
 456         while (true) {
 457                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 458                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 459                 int bind_len;
 460                 int step;
 461                 int mul;
 462
 463                 tok = strsep(&str, ",");
 464                 if (!tok)
 465                         break;
 466
 467                 tok_end = strstr(tok, "-");
 468
 469                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 470                 if (!tok_end) {
 471                         /* Single CPU specified: */
 472                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 473                 } else {
 474                         /* CPU range specified (for example: "5-11"): */
 475                         bind_cpu_0 = atol(tok);
 476                         bind_cpu_1 = atol(tok_end + 1);
 477                 }
 478
 479                 step = 1;
 480                 tok_step = strstr(tok, "#");
 481                 if (tok_step) {
 482                         step = atol(tok_step + 1);
 483                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 484                 }
 485
 486                 /*
 487                  * Mask length.
 488                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 489                  * where the _4 means the next 4 CPUs are allowed.
 490                  */
 491                 bind_len = 1;
 492                 tok_len = strstr(tok, "_");
 493                 if (tok_len) {
 494                         bind_len = atol(tok_len + 1);
 495                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 496                 }
 497
 498                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 499                 mul = 1;
 500                 tok_mul = strstr(tok, "x");
 501                 if (tok_mul) {
 502                         mul = atol(tok_mul + 1);
 503                         BUG_ON(mul <= 0);
 504                 }
 505
 506                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 507
 508                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 509                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 510                         return -1;
 511                 }
 512
 513                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 514                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 515
 516                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 517                         int i;
 518
 519                         for (i = 0; i < mul; i++) {
 520                                 int cpu;
 521
 522                                 if (t >= g->p.nr_tasks) {
 523                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 524                                         goto out;
 525                                 }
 526                                 td = g->threads + t;
 527
 528                                 if (t)
 529                                         tprintf(",");
 530                                 if (bind_len > 1) {
 531                                         tprintf("%2d/%d", bind_cpu, bind_len);
 532                                 } else {
 533                                         tprintf("%2d", bind_cpu);
 534                                 }
 535
 536                                 CPU_ZERO(&td->bind_cpumask);
 537                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 538                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 539                                         CPU_SET(cpu, &td->bind_cpumask);
 540                                 }
 541                                 t++;
 542                         }
 543                 }
 544         }
 545 out:
 546
 547         tprintf("\n");
 548
 549         if (t < g->p.nr_tasks)
 550                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 551
 552         free(str0);
 553         return 0;
 554 }
 555
 556 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 557                           const char *arg, int unset __maybe_unused)
 558 {
 559         if (!arg)
 560                 return -1;
 561
 562         return parse_cpu_list(arg);
 563 }
 564
 565 static int parse_node_list(const char *arg)
 566 {
 567         p0.node_list_str = strdup(arg);
 568
 569         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 570
 571         return 0;
 572 }
 573
 574 static int parse_setup_node_list(void)
 575 {
 576         struct thread_data *td;
 577         char *str0, *str;
 578         int t;
 579
 580         if (!g->p.node_list_str)
 581                 return 0;
 582
 583         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 584
 585         str0 = str = strdup(g->p.node_list_str);
 586         t = 0;
 587
 588         BUG_ON(!str);
 589
 590         tprintf("# binding tasks to NODEs:\n");
 591         tprintf("# ");
 592
 593         while (true) {
 594                 int bind_node, bind_node_0, bind_node_1;
 595                 char *tok, *tok_end, *tok_step, *tok_mul;
 596                 int step;
 597                 int mul;
 598
 599                 tok = strsep(&str, ",");
 600                 if (!tok)
 601                         break;
 602
 603                 tok_end = strstr(tok, "-");
 604
 605                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 606                 if (!tok_end) {
 607                         /* Single NODE specified: */
 608                         bind_node_0 = bind_node_1 = atol(tok);
 609                 } else {
 610                         /* NODE range specified (for example: "5-11"): */
 611                         bind_node_0 = atol(tok);
 612                         bind_node_1 = atol(tok_end + 1);
 613                 }
 614
 615                 step = 1;
 616                 tok_step = strstr(tok, "#");
 617                 if (tok_step) {
 618                         step = atol(tok_step + 1);
 619                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 620                 }
 621
 622                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 623                 mul = 1;
 624                 tok_mul = strstr(tok, "x");
 625                 if (tok_mul) {
 626                         mul = atol(tok_mul + 1);
 627                         BUG_ON(mul <= 0);
 628                 }
 629
 630                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 631
 632                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 633                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 634                         return -1;
 635                 }
 636
 637                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 638                 BUG_ON(bind_node_0 > bind_node_1);
 639
 640                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 641                         int i;
 642
 643                         for (i = 0; i < mul; i++) {
 644                                 if (t >= g->p.nr_tasks) {
 645                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 646                                         goto out;
 647                                 }
 648                                 td = g->threads + t;
 649
 650                                 if (!t)
 651                                         tprintf(" %2d", bind_node);
 652                                 else
 653                                         tprintf(",%2d", bind_node);
 654
 655                                 td->bind_node = bind_node;
 656                                 t++;
 657                         }
 658                 }
 659         }
 660 out:
 661
 662         tprintf("\n");
 663
 664         if (t < g->p.nr_tasks)
 665                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 666
 667         free(str0);
 668         return 0;
 669 }
 670
 671 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 672                           const char *arg, int unset __maybe_unused)
 673 {
 674         if (!arg)
 675                 return -1;
 676
 677         return parse_node_list(arg);
 678
 679         return 0;
 680 }
 681
 682 #define BIT(x) (1ul << x)
 683
 684 static inline uint32_t lfsr_32(uint32_t lfsr)
 685 {
 686         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 687         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 688 }
 689
 690 /*
 691  * Make sure there's real data dependency to RAM (when read
 692  * accesses are enabled), so the compiler, the CPU and the
 693  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 694  * accesses:
 695  */
 696 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 697 {
 698         if (g->p.data_reads)
 699                 val += *data;
 700         if (g->p.data_writes)
 701                 *data = val + 1;
 702         return val;
 703 }
 704
 705 /*
 706  * The worker process does two types of work, a forwards going
 707  * loop and a backwards going loop.
 708  *
 709  * We do this so that on multiprocessor systems we do not create
 710  * a 'train' of processing, with highly synchronized processes,
 711  * skewing the whole benchmark.
 712  */
 713 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 714 {
 715         long words = bytes/sizeof(u64);
 716         u64 *data = (void *)__data;
 717         long chunk_0, chunk_1;
 718         u64 *d0, *d, *d1;
 719         long off;
 720         long i;
 721
 722         BUG_ON(!data && words);
 723         BUG_ON(data && !words);
 724
 725         if (!data)
 726                 return val;
 727
 728         /* Very simple memset() work variant: */
 729         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 730                 bzero(data, bytes);
 731                 return val;
 732         }
 733
 734         /* Spread out by PID/TID nr and by loop nr: */
 735         chunk_0 = words/nr_max;
 736         chunk_1 = words/g->p.nr_loops;
 737         off = nr*chunk_0 + loop*chunk_1;
 738
 739         while (off >= words)
 740                 off -= words;
 741
 742         if (g->p.data_rand_walk) {
 743                 u32 lfsr = nr + loop + val;
 744                 int j;
 745
 746                 for (i = 0; i < words/1024; i++) {
 747                         long start, end;
 748
 749                         lfsr = lfsr_32(lfsr);
 750
 751                         start = lfsr % words;
 752                         end = min(start + 1024, words-1);
 753
 754                         if (g->p.data_zero_memset) {
 755                                 bzero(data + start, (end-start) * sizeof(u64));
 756                         } else {
 757                                 for (j = start; j < end; j++)
 758                                         val = access_data(data + j, val);
 759                         }
 760                 }
 761         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 762
 763                 d0 = data + off;
 764                 d  = data + off + 1;
 765                 d1 = data + words;
 766
 767                 /* Process data forwards: */
 768                 for (;;) {
 769                         if (unlikely(d >= d1))
 770                                 d = data;
 771                         if (unlikely(d == d0))
 772                                 break;
 773
 774                         val = access_data(d, val);
 775
 776                         d++;
 777                 }
 778         } else {
 779                 /* Process data backwards: */
 780
 781                 d0 = data + off;
 782                 d  = data + off - 1;
 783                 d1 = data + words;
 784
 785                 /* Process data forwards: */
 786                 for (;;) {
 787                         if (unlikely(d < data))
 788                                 d = data + words-1;
 789                         if (unlikely(d == d0))
 790                                 break;
 791
 792                         val = access_data(d, val);
 793
 794                         d--;
 795                 }
 796         }
 797
 798         return val;
 799 }
 800
 801 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 802 {
 803         unsigned int cpu;
 804
 805         cpu = sched_getcpu();
 806
 807         g->threads[task_nr].curr_cpu = cpu;
 808         prctl(0, bytes_worked);
 809 }
 810
 811 #define MAX_NR_NODES    64
 812
 813 /*
 814  * Count the number of nodes a process's threads
 815  * are spread out on.
 816  *
 817  * A count of 1 means that the process is compressed
 818  * to a single node. A count of g->p.nr_nodes means it's
 819  * spread out on the whole system.
 820  */
 821 static int count_process_nodes(int process_nr)
 822 {
 823         char node_present[MAX_NR_NODES] = { 0, };
 824         int nodes;
 825         int n, t;
 826
 827         for (t = 0; t < g->p.nr_threads; t++) {
 828                 struct thread_data *td;
 829                 int task_nr;
 830                 int node;
 831
 832                 task_nr = process_nr*g->p.nr_threads + t;
 833                 td = g->threads + task_nr;
 834
 835                 node = numa_node_of_cpu(td->curr_cpu);
 836                 if (node < 0) /* curr_cpu was likely still -1 */
 837                         return 0;
 838
 839                 node_present[node] = 1;
 840         }
 841
 842         nodes = 0;
 843
 844         for (n = 0; n < MAX_NR_NODES; n++)
 845                 nodes += node_present[n];
 846
 847         return nodes;
 848 }
 849
 850 /*
 851  * Count the number of distinct process-threads a node contains.
 852  *
 853  * A count of 1 means that the node contains only a single
 854  * process. If all nodes on the system contain at most one
 855  * process then we are well-converged.
 856  */
 857 static int count_node_processes(int node)
 858 {
 859         int processes = 0;
 860         int t, p;
 861
 862         for (p = 0; p < g->p.nr_proc; p++) {
 863                 for (t = 0; t < g->p.nr_threads; t++) {
 864                         struct thread_data *td;
 865                         int task_nr;
 866                         int n;
 867
 868                         task_nr = p*g->p.nr_threads + t;
 869                         td = g->threads + task_nr;
 870
 871                         n = numa_node_of_cpu(td->curr_cpu);
 872                         if (n == node) {
 873                                 processes++;
 874                                 break;
 875                         }
 876                 }
 877         }
 878
 879         return processes;
 880 }
 881
 882 static void calc_convergence_compression(int *strong)
 883 {
 884         unsigned int nodes_min, nodes_max;
 885         int p;
 886
 887         nodes_min = -1;
 888         nodes_max =  0;
 889
 890         for (p = 0; p < g->p.nr_proc; p++) {
 891                 unsigned int nodes = count_process_nodes(p);
 892
 893                 if (!nodes) {
 894                         *strong = 0;
 895                         return;
 896                 }
 897
 898                 nodes_min = min(nodes, nodes_min);
 899                 nodes_max = max(nodes, nodes_max);
 900         }
 901
 902         /* Strong convergence: all threads compress on a single node: */
 903         if (nodes_min == 1 && nodes_max == 1) {
 904                 *strong = 1;
 905         } else {
 906                 *strong = 0;
 907                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 908         }
 909 }
 910
 911 static void calc_convergence(double runtime_ns_max, double *convergence)
 912 {
 913         unsigned int loops_done_min, loops_done_max;
 914         int process_groups;
 915         int nodes[MAX_NR_NODES];
 916         int distance;
 917         int nr_min;
 918         int nr_max;
 919         int strong;
 920         int sum;
 921         int nr;
 922         int node;
 923         int cpu;
 924         int t;
 925
 926         if (!g->p.show_convergence && !g->p.measure_convergence)
 927                 return;
 928
 929         for (node = 0; node < g->p.nr_nodes; node++)
 930                 nodes[node] = 0;
 931
 932         loops_done_min = -1;
 933         loops_done_max = 0;
 934
 935         for (t = 0; t < g->p.nr_tasks; t++) {
 936                 struct thread_data *td = g->threads + t;
 937                 unsigned int loops_done;
 938
 939                 cpu = td->curr_cpu;
 940
 941                 /* Not all threads have written it yet: */
 942                 if (cpu < 0)
 943                         continue;
 944
 945                 node = numa_node_of_cpu(cpu);
 946
 947                 nodes[node]++;
 948
 949                 loops_done = td->loops_done;
 950                 loops_done_min = min(loops_done, loops_done_min);
 951                 loops_done_max = max(loops_done, loops_done_max);
 952         }
 953
 954         nr_max = 0;
 955         nr_min = g->p.nr_tasks;
 956         sum = 0;
 957
 958         for (node = 0; node < g->p.nr_nodes; node++) {
 959                 nr = nodes[node];
 960                 nr_min = min(nr, nr_min);
 961                 nr_max = max(nr, nr_max);
 962                 sum += nr;
 963         }
 964         BUG_ON(nr_min > nr_max);
 965
 966         BUG_ON(sum > g->p.nr_tasks);
 967
 968         if (0 && (sum < g->p.nr_tasks))
 969                 return;
 970
 971         /*
 972          * Count the number of distinct process groups present
 973          * on nodes - when we are converged this will decrease
 974          * to g->p.nr_proc:
 975          */
 976         process_groups = 0;
 977
 978         for (node = 0; node < g->p.nr_nodes; node++) {
 979                 int processes = count_node_processes(node);
 980
 981                 nr = nodes[node];
 982                 tprintf(" %2d/%-2d", nr, processes);
 983
 984                 process_groups += processes;
 985         }
 986
 987         distance = nr_max - nr_min;
 988
 989         tprintf(" [%2d/%-2d]", distance, process_groups);
 990
 991         tprintf(" l:%3d-%-3d (%3d)",
 992                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 993
 994         if (loops_done_min && loops_done_max) {
 995                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
 996
 997                 tprintf(" [%4.1f%%]", skew * 100.0);
 998         }
 999
1000         calc_convergence_compression(&strong);
1001
1002         if (strong && process_groups == g->p.nr_proc) {
1003                 if (!*convergence) {
1004                         *convergence = runtime_ns_max;
1005                         tprintf(" (%6.1fs converged)\n", *convergence/1e9);
1006                         if (g->p.measure_convergence) {
1007                                 g->all_converged = true;
1008                                 g->stop_work = true;
1009                         }
1010                 }
1011         } else {
1012                 if (*convergence) {
1013                         tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
1014                         *convergence = 0;
1015                 }
1016                 tprintf("\n");
1017         }
1018 }
1019
1020 static void show_summary(double runtime_ns_max, int l, double *convergence)
1021 {
1022         tprintf("\r #  %5.1f%%  [%.1f mins]",
1023                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1024
1025         calc_convergence(runtime_ns_max, convergence);
1026
1027         if (g->p.show_details >= 0)
1028                 fflush(stdout);
1029 }
1030
1031 static void *worker_thread(void *__tdata)
1032 {
1033         struct thread_data *td = __tdata;
1034         struct timeval start0, start, stop, diff;
1035         int process_nr = td->process_nr;
1036         int thread_nr = td->thread_nr;
1037         unsigned long last_perturbance;
1038         int task_nr = td->task_nr;
1039         int details = g->p.show_details;
1040         int first_task, last_task;
1041         double convergence = 0;
1042         u64 val = td->val;
1043         double runtime_ns_max;
1044         u8 *global_data;
1045         u8 *process_data;
1046         u8 *thread_data;
1047         u64 bytes_done;
1048         long work_done;
1049         u32 l;
1050         struct rusage rusage;
1051
1052         bind_to_cpumask(td->bind_cpumask);
1053         bind_to_memnode(td->bind_node);
1054
1055         set_taskname("thread %d/%d", process_nr, thread_nr);
1056
1057         global_data = g->data;
1058         process_data = td->process_data;
1059         thread_data = setup_private_data(g->p.bytes_thread);
1060
1061         bytes_done = 0;
1062
1063         last_task = 0;
1064         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1065                 last_task = 1;
1066
1067         first_task = 0;
1068         if (process_nr == 0 && thread_nr == 0)
1069                 first_task = 1;
1070
1071         if (details >= 2) {
1072                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1073                         process_nr, thread_nr, global_data, process_data, thread_data);
1074         }
1075
1076         if (g->p.serialize_startup) {
1077                 pthread_mutex_lock(&g->startup_mutex);
1078                 g->nr_tasks_started++;
1079                 pthread_mutex_unlock(&g->startup_mutex);
1080
1081                 /* Here we will wait for the main process to start us all at once: */
1082                 pthread_mutex_lock(&g->start_work_mutex);
1083                 g->nr_tasks_working++;
1084
1085                 /* Last one wake the main process: */
1086                 if (g->nr_tasks_working == g->p.nr_tasks)
1087                         pthread_mutex_unlock(&g->startup_done_mutex);
1088
1089                 pthread_mutex_unlock(&g->start_work_mutex);
1090         }
1091
1092         gettimeofday(&start0, NULL);
1093
1094         start = stop = start0;
1095         last_perturbance = start.tv_sec;
1096
1097         for (l = 0; l < g->p.nr_loops; l++) {
1098                 start = stop;
1099
1100                 if (g->stop_work)
1101                         break;
1102
1103                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1104                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1105                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1106
1107                 if (g->p.sleep_usecs) {
1108                         pthread_mutex_lock(td->process_lock);
1109                         usleep(g->p.sleep_usecs);
1110                         pthread_mutex_unlock(td->process_lock);
1111                 }
1112                 /*
1113                  * Amount of work to be done under a process-global lock:
1114                  */
1115                 if (g->p.bytes_process_locked) {
1116                         pthread_mutex_lock(td->process_lock);
1117                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1118                         pthread_mutex_unlock(td->process_lock);
1119                 }
1120
1121                 work_done = g->p.bytes_global + g->p.bytes_process +
1122                             g->p.bytes_process_locked + g->p.bytes_thread;
1123
1124                 update_curr_cpu(task_nr, work_done);
1125                 bytes_done += work_done;
1126
1127                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1128                         continue;
1129
1130                 td->loops_done = l;
1131
1132                 gettimeofday(&stop, NULL);
1133
1134                 /* Check whether our max runtime timed out: */
1135                 if (g->p.nr_secs) {
1136                         timersub(&stop, &start0, &diff);
1137                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1138                                 g->stop_work = true;
1139                                 break;
1140                         }
1141                 }
1142
1143                 /* Update the summary at most once per second: */
1144                 if (start.tv_sec == stop.tv_sec)
1145                         continue;
1146
1147                 /*
1148                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1149                  * by migrating to CPU#0:
1150                  */
1151                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1152                         cpu_set_t orig_mask;
1153                         int target_cpu;
1154                         int this_cpu;
1155
1156                         last_perturbance = stop.tv_sec;
1157
1158                         /*
1159                          * Depending on where we are running, move into
1160                          * the other half of the system, to create some
1161                          * real disturbance:
1162                          */
1163                         this_cpu = g->threads[task_nr].curr_cpu;
1164                         if (this_cpu < g->p.nr_cpus/2)
1165                                 target_cpu = g->p.nr_cpus-1;
1166                         else
1167                                 target_cpu = 0;
1168
1169                         orig_mask = bind_to_cpu(target_cpu);
1170
1171                         /* Here we are running on the target CPU already */
1172                         if (details >= 1)
1173                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1174
1175                         bind_to_cpumask(orig_mask);
1176                 }
1177
1178                 if (details >= 3) {
1179                         timersub(&stop, &start, &diff);
1180                         runtime_ns_max = diff.tv_sec * 1000000000;
1181                         runtime_ns_max += diff.tv_usec * 1000;
1182
1183                         if (details >= 0) {
1184                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1185                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1186                         }
1187                         fflush(stdout);
1188                 }
1189                 if (!last_task)
1190                         continue;
1191
1192                 timersub(&stop, &start0, &diff);
1193                 runtime_ns_max = diff.tv_sec * 1000000000ULL;
1194                 runtime_ns_max += diff.tv_usec * 1000ULL;
1195
1196                 show_summary(runtime_ns_max, l, &convergence);
1197         }
1198
1199         gettimeofday(&stop, NULL);
1200         timersub(&stop, &start0, &diff);
1201         td->runtime_ns = diff.tv_sec * 1000000000ULL;
1202         td->runtime_ns += diff.tv_usec * 1000ULL;
1203         td->speed_gbs = bytes_done / (td->runtime_ns / 1e9) / 1e9;
1204
1205         getrusage(RUSAGE_THREAD, &rusage);
1206         td->system_time_ns = rusage.ru_stime.tv_sec * 1000000000ULL;
1207         td->system_time_ns += rusage.ru_stime.tv_usec * 1000ULL;
1208         td->user_time_ns = rusage.ru_utime.tv_sec * 1000000000ULL;
1209         td->user_time_ns += rusage.ru_utime.tv_usec * 1000ULL;
1210
1211         free_data(thread_data, g->p.bytes_thread);
1212
1213         pthread_mutex_lock(&g->stop_work_mutex);
1214         g->bytes_done += bytes_done;
1215         pthread_mutex_unlock(&g->stop_work_mutex);
1216
1217         return NULL;
1218 }
1219
1220 /*
1221  * A worker process starts a couple of threads:
1222  */
1223 static void worker_process(int process_nr)
1224 {
1225         pthread_mutex_t process_lock;
1226         struct thread_data *td;
1227         pthread_t *pthreads;
1228         u8 *process_data;
1229         int task_nr;
1230         int ret;
1231         int t;
1232
1233         pthread_mutex_init(&process_lock, NULL);
1234         set_taskname("process %d", process_nr);
1235
1236         /*
1237          * Pick up the memory policy and the CPU binding of our first thread,
1238          * so that we initialize memory accordingly:
1239          */
1240         task_nr = process_nr*g->p.nr_threads;
1241         td = g->threads + task_nr;
1242
1243         bind_to_memnode(td->bind_node);
1244         bind_to_cpumask(td->bind_cpumask);
1245
1246         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1247         process_data = setup_private_data(g->p.bytes_process);
1248
1249         if (g->p.show_details >= 3) {
1250                 printf(" # process %2d global mem: %p, process mem: %p\n",
1251                         process_nr, g->data, process_data);
1252         }
1253
1254         for (t = 0; t < g->p.nr_threads; t++) {
1255                 task_nr = process_nr*g->p.nr_threads + t;
1256                 td = g->threads + task_nr;
1257
1258                 td->process_data = process_data;
1259                 td->process_nr   = process_nr;
1260                 td->thread_nr    = t;
1261                 td->task_nr      = task_nr;
1262                 td->val          = rand();
1263                 td->curr_cpu     = -1;
1264                 td->process_lock = &process_lock;
1265
1266                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1267                 BUG_ON(ret);
1268         }
1269
1270         for (t = 0; t < g->p.nr_threads; t++) {
1271                 ret = pthread_join(pthreads[t], NULL);
1272                 BUG_ON(ret);
1273         }
1274
1275         free_data(process_data, g->p.bytes_process);
1276         free(pthreads);
1277 }
1278
1279 static void print_summary(void)
1280 {
1281         if (g->p.show_details < 0)
1282                 return;
1283
1284         printf("\n ###\n");
1285         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1286                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1287         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1288                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1289         printf(" #      %5dx %5ldMB process shared mem operations\n",
1290                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1291         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1292                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1293
1294         printf(" ###\n");
1295
1296         printf("\n ###\n"); fflush(stdout);
1297 }
1298
1299 static void init_thread_data(void)
1300 {
1301         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1302         int t;
1303
1304         g->threads = zalloc_shared_data(size);
1305
1306         for (t = 0; t < g->p.nr_tasks; t++) {
1307                 struct thread_data *td = g->threads + t;
1308                 int cpu;
1309
1310                 /* Allow all nodes by default: */
1311                 td->bind_node = -1;
1312
1313                 /* Allow all CPUs by default: */
1314                 CPU_ZERO(&td->bind_cpumask);
1315                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1316                         CPU_SET(cpu, &td->bind_cpumask);
1317         }
1318 }
1319
1320 static void deinit_thread_data(void)
1321 {
1322         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1323
1324         free_data(g->threads, size);
1325 }
1326
1327 static int init(void)
1328 {
1329         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1330
1331         /* Copy over options: */
1332         g->p = p0;
1333
1334         g->p.nr_cpus = numa_num_configured_cpus();
1335
1336         g->p.nr_nodes = numa_max_node() + 1;
1337
1338         /* char array in count_process_nodes(): */
1339         BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1340
1341         if (g->p.show_quiet && !g->p.show_details)
1342                 g->p.show_details = -1;
1343
1344         /* Some memory should be specified: */
1345         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1346                 return -1;
1347
1348         if (g->p.mb_global_str) {
1349                 g->p.mb_global = atof(g->p.mb_global_str);
1350                 BUG_ON(g->p.mb_global < 0);
1351         }
1352
1353         if (g->p.mb_proc_str) {
1354                 g->p.mb_proc = atof(g->p.mb_proc_str);
1355                 BUG_ON(g->p.mb_proc < 0);
1356         }
1357
1358         if (g->p.mb_proc_locked_str) {
1359                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1360                 BUG_ON(g->p.mb_proc_locked < 0);
1361                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1362         }
1363
1364         if (g->p.mb_thread_str) {
1365                 g->p.mb_thread = atof(g->p.mb_thread_str);
1366                 BUG_ON(g->p.mb_thread < 0);
1367         }
1368
1369         BUG_ON(g->p.nr_threads <= 0);
1370         BUG_ON(g->p.nr_proc <= 0);
1371
1372         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1373
1374         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1375         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1376         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1377         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1378
1379         g->data = setup_shared_data(g->p.bytes_global);
1380
1381         /* Startup serialization: */
1382         init_global_mutex(&g->start_work_mutex);
1383         init_global_mutex(&g->startup_mutex);
1384         init_global_mutex(&g->startup_done_mutex);
1385         init_global_mutex(&g->stop_work_mutex);
1386
1387         init_thread_data();
1388
1389         tprintf("#\n");
1390         if (parse_setup_cpu_list() || parse_setup_node_list())
1391                 return -1;
1392         tprintf("#\n");
1393
1394         print_summary();
1395
1396         return 0;
1397 }
1398
1399 static void deinit(void)
1400 {
1401         free_data(g->data, g->p.bytes_global);
1402         g->data = NULL;
1403
1404         deinit_thread_data();
1405
1406         free_data(g, sizeof(*g));
1407         g = NULL;
1408 }
1409
1410 /*
1411  * Print a short or long result, depending on the verbosity setting:
1412  */
1413 static void print_res(const char *name, double val,
1414                       const char *txt_unit, const char *txt_short, const char *txt_long)
1415 {
1416         if (!name)
1417                 name = "main,";
1418
1419         if (!g->p.show_quiet)
1420                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1421         else
1422                 printf(" %14.3f %s\n", val, txt_long);
1423 }
1424
1425 static int __bench_numa(const char *name)
1426 {
1427         struct timeval start, stop, diff;
1428         u64 runtime_ns_min, runtime_ns_sum;
1429         pid_t *pids, pid, wpid;
1430         double delta_runtime;
1431         double runtime_avg;
1432         double runtime_sec_max;
1433         double runtime_sec_min;
1434         int wait_stat;
1435         double bytes;
1436         int i, t, p;
1437
1438         if (init())
1439                 return -1;
1440
1441         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1442         pid = -1;
1443
1444         /* All threads try to acquire it, this way we can wait for them to start up: */
1445         pthread_mutex_lock(&g->start_work_mutex);
1446
1447         if (g->p.serialize_startup) {
1448                 tprintf(" #\n");
1449                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1450         }
1451
1452         gettimeofday(&start, NULL);
1453
1454         for (i = 0; i < g->p.nr_proc; i++) {
1455                 pid = fork();
1456                 dprintf(" # process %2d: PID %d\n", i, pid);
1457
1458                 BUG_ON(pid < 0);
1459                 if (!pid) {
1460                         /* Child process: */
1461                         worker_process(i);
1462
1463                         exit(0);
1464                 }
1465                 pids[i] = pid;
1466
1467         }
1468         /* Wait for all the threads to start up: */
1469         while (g->nr_tasks_started != g->p.nr_tasks)
1470                 usleep(1000);
1471
1472         BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1473
1474         if (g->p.serialize_startup) {
1475                 double startup_sec;
1476
1477                 pthread_mutex_lock(&g->startup_done_mutex);
1478
1479                 /* This will start all threads: */
1480                 pthread_mutex_unlock(&g->start_work_mutex);
1481
1482                 /* This mutex is locked - the last started thread will wake us: */
1483                 pthread_mutex_lock(&g->startup_done_mutex);
1484
1485                 gettimeofday(&stop, NULL);
1486
1487                 timersub(&stop, &start, &diff);
1488
1489                 startup_sec = diff.tv_sec * 1000000000.0;
1490                 startup_sec += diff.tv_usec * 1000.0;
1491                 startup_sec /= 1e9;
1492
1493                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1494                 tprintf(" #\n");
1495
1496                 start = stop;
1497                 pthread_mutex_unlock(&g->startup_done_mutex);
1498         } else {
1499                 gettimeofday(&start, NULL);
1500         }
1501
1502         /* Parent process: */
1503
1504
1505         for (i = 0; i < g->p.nr_proc; i++) {
1506                 wpid = waitpid(pids[i], &wait_stat, 0);
1507                 BUG_ON(wpid < 0);
1508                 BUG_ON(!WIFEXITED(wait_stat));
1509
1510         }
1511
1512         runtime_ns_sum = 0;
1513         runtime_ns_min = -1LL;
1514
1515         for (t = 0; t < g->p.nr_tasks; t++) {
1516                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1517
1518                 runtime_ns_sum += thread_runtime_ns;
1519                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1520         }
1521
1522         gettimeofday(&stop, NULL);
1523         timersub(&stop, &start, &diff);
1524
1525         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1526
1527         tprintf("\n ###\n");
1528         tprintf("\n");
1529
1530         runtime_sec_max = diff.tv_sec * 1000000000.0;
1531         runtime_sec_max += diff.tv_usec * 1000.0;
1532         runtime_sec_max /= 1e9;
1533
1534         runtime_sec_min = runtime_ns_min/1e9;
1535
1536         bytes = g->bytes_done;
1537         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1538
1539         if (g->p.measure_convergence) {
1540                 print_res(name, runtime_sec_max,
1541                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1542         }
1543
1544         print_res(name, runtime_sec_max,
1545                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1546
1547         print_res(name, runtime_sec_min,
1548                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1549
1550         print_res(name, runtime_avg,
1551                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1552
1553         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1554         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1555                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1556
1557         print_res(name, bytes / g->p.nr_tasks / 1e9,
1558                 "GB,", "data/thread",           "GB data processed, per thread");
1559
1560         print_res(name, bytes / 1e9,
1561                 "GB,", "data-total",            "GB data processed, total");
1562
1563         print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1564                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1565
1566         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1567                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1568
1569         print_res(name, bytes / runtime_sec_max / 1e9,
1570                 "GB/sec,", "total-speed",       "GB/sec total speed");
1571
1572         if (g->p.show_details >= 2) {
1573                 char tname[32];
1574                 struct thread_data *td;
1575                 for (p = 0; p < g->p.nr_proc; p++) {
1576                         for (t = 0; t < g->p.nr_threads; t++) {
1577                                 memset(tname, 0, 32);
1578                                 td = g->threads + p*g->p.nr_threads + t;
1579                                 snprintf(tname, 32, "process%d:thread%d", p, t);
1580                                 print_res(tname, td->speed_gbs,
1581                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1582                                 print_res(tname, td->system_time_ns / 1e9,
1583                                         "secs", "thread-system-time", "system CPU time/thread");
1584                                 print_res(tname, td->user_time_ns / 1e9,
1585                                         "secs", "thread-user-time", "user CPU time/thread");
1586                         }
1587                 }
1588         }
1589
1590         free(pids);
1591
1592         deinit();
1593
1594         return 0;
1595 }
1596
1597 #define MAX_ARGS 50
1598
1599 static int command_size(const char **argv)
1600 {
1601         int size = 0;
1602
1603         while (*argv) {
1604                 size++;
1605                 argv++;
1606         }
1607
1608         BUG_ON(size >= MAX_ARGS);
1609
1610         return size;
1611 }
1612
1613 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1614 {
1615         int i;
1616
1617         printf("\n # Running %s \"perf bench numa", name);
1618
1619         for (i = 0; i < argc; i++)
1620                 printf(" %s", argv[i]);
1621
1622         printf("\"\n");
1623
1624         memset(p, 0, sizeof(*p));
1625
1626         /* Initialize nonzero defaults: */
1627
1628         p->serialize_startup            = 1;
1629         p->data_reads                   = true;
1630         p->data_writes                  = true;
1631         p->data_backwards               = true;
1632         p->data_rand_walk               = true;
1633         p->nr_loops                     = -1;
1634         p->init_random                  = true;
1635         p->mb_global_str                = "1";
1636         p->nr_proc                      = 1;
1637         p->nr_threads                   = 1;
1638         p->nr_secs                      = 5;
1639         p->run_all                      = argc == 1;
1640 }
1641
1642 static int run_bench_numa(const char *name, const char **argv)
1643 {
1644         int argc = command_size(argv);
1645
1646         init_params(&p0, name, argc, argv);
1647         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1648         if (argc)
1649                 goto err;
1650
1651         if (__bench_numa(name))
1652                 goto err;
1653
1654         return 0;
1655
1656 err:
1657         return -1;
1658 }
1659
1660 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1661 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1662
1663 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1664 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1665
1666 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1667 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1668
1669 /*
1670  * The built-in test-suite executed by "perf bench numa -a".
1671  *
1672  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1673  */
1674 static const char *tests[][MAX_ARGS] = {
1675    /* Basic single-stream NUMA bandwidth measurements: */
1676    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1677                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1678    { "RAM-bw-local-NOTHP,",
1679                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1680                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1681    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1682                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1683
1684    /* 2-stream NUMA bandwidth measurements: */
1685    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1686                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1687    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1688                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1689
1690    /* Cross-stream NUMA bandwidth measurement: */
1691    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1692                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1693
1694    /* Convergence latency measurements: */
1695    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1696    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1697    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1698    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1699    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1700    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1701    { " 4x4-convergence-NOTHP,",
1702                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1703    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1704    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1705    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1706    { " 8x4-convergence-NOTHP,",
1707                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1708    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1709    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1710    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1711    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1712    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1713
1714    /* Various NUMA process/thread layout bandwidth measurements: */
1715    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1716    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1717    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1718    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1719    { " 8x1-bw-process-NOTHP,",
1720                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1721    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1722
1723    { " 4x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1724    { " 8x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1725    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1726    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1727
1728    { " 2x3-bw-thread,",   "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1729    { " 4x4-bw-thread,",   "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1730    { " 4x6-bw-thread,",   "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1731    { " 4x8-bw-thread,",   "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1732    { " 4x8-bw-thread-NOTHP,",
1733                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1734    { " 3x3-bw-thread,",   "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1735    { " 5x5-bw-thread,",   "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1736
1737    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1738    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1739
1740    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1741    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1742    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1743    { "numa01-bw-thread-NOTHP,",
1744                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1745 };
1746
1747 static int bench_all(void)
1748 {
1749         int nr = ARRAY_SIZE(tests);
1750         int ret;
1751         int i;
1752
1753         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1754         BUG_ON(ret < 0);
1755
1756         for (i = 0; i < nr; i++) {
1757                 run_bench_numa(tests[i][0], tests[i] + 1);
1758         }
1759
1760         printf("\n");
1761
1762         return 0;
1763 }
1764
1765 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1766 {
1767         init_params(&p0, "main,", argc, argv);
1768         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1769         if (argc)
1770                 goto err;
1771
1772         if (p0.run_all)
1773                 return bench_all();
1774
1775         if (__bench_numa(NULL))
1776                 goto err;
1777
1778         return 0;
1779
1780 err:
1781         usage_with_options(numa_usage, options);
1782         return -1;
1783 }