tools/perf/bench/numa.c

   1 /*
   2  * numa.c
   3  *
   4  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5  */
   6
   7 /* For the CLR_() macros */
   8 #include <pthread.h>
   9
  10 #include "../perf.h"
  11 #include "../builtin.h"
  12 #include "../util/util.h"
  13 #include <subcmd/parse-options.h>
  14 #include "../util/cloexec.h"
  15
  16 #include "bench.h"
  17
  18 #include <errno.h>
  19 #include <sched.h>
  20 #include <stdio.h>
  21 #include <assert.h>
  22 #include <malloc.h>
  23 #include <signal.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <sys/mman.h>
  28 #include <sys/time.h>
  29 #include <sys/resource.h>
  30 #include <sys/wait.h>
  31 #include <sys/prctl.h>
  32 #include <sys/types.h>
  33 #include <linux/time64.h>
  34
  35 #include <numa.h>
  36 #include <numaif.h>
  37
  38 /*
  39  * Regular printout to the terminal, supressed if -q is specified:
  40  */
  41 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  42
  43 /*
  44  * Debug printf:
  45  */
  46 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  47
  48 struct thread_data {
  49         int                     curr_cpu;
  50         cpu_set_t               bind_cpumask;
  51         int                     bind_node;
  52         u8                      *process_data;
  53         int                     process_nr;
  54         int                     thread_nr;
  55         int                     task_nr;
  56         unsigned int            loops_done;
  57         u64                     val;
  58         u64                     runtime_ns;
  59         u64                     system_time_ns;
  60         u64                     user_time_ns;
  61         double                  speed_gbs;
  62         pthread_mutex_t         *process_lock;
  63 };
  64
  65 /* Parameters set by options: */
  66
  67 struct params {
  68         /* Startup synchronization: */
  69         bool                    serialize_startup;
  70
  71         /* Task hierarchy: */
  72         int                     nr_proc;
  73         int                     nr_threads;
  74
  75         /* Working set sizes: */
  76         const char              *mb_global_str;
  77         const char              *mb_proc_str;
  78         const char              *mb_proc_locked_str;
  79         const char              *mb_thread_str;
  80
  81         double                  mb_global;
  82         double                  mb_proc;
  83         double                  mb_proc_locked;
  84         double                  mb_thread;
  85
  86         /* Access patterns to the working set: */
  87         bool                    data_reads;
  88         bool                    data_writes;
  89         bool                    data_backwards;
  90         bool                    data_zero_memset;
  91         bool                    data_rand_walk;
  92         u32                     nr_loops;
  93         u32                     nr_secs;
  94         u32                     sleep_usecs;
  95
  96         /* Working set initialization: */
  97         bool                    init_zero;
  98         bool                    init_random;
  99         bool                    init_cpu0;
 100
 101         /* Misc options: */
 102         int                     show_details;
 103         int                     run_all;
 104         int                     thp;
 105
 106         long                    bytes_global;
 107         long                    bytes_process;
 108         long                    bytes_process_locked;
 109         long                    bytes_thread;
 110
 111         int                     nr_tasks;
 112         bool                    show_quiet;
 113
 114         bool                    show_convergence;
 115         bool                    measure_convergence;
 116
 117         int                     perturb_secs;
 118         int                     nr_cpus;
 119         int                     nr_nodes;
 120
 121         /* Affinity options -C and -N: */
 122         char                    *cpu_list_str;
 123         char                    *node_list_str;
 124 };
 125
 126
 127 /* Global, read-writable area, accessible to all processes and threads: */
 128
 129 struct global_info {
 130         u8                      *data;
 131
 132         pthread_mutex_t         startup_mutex;
 133         int                     nr_tasks_started;
 134
 135         pthread_mutex_t         startup_done_mutex;
 136
 137         pthread_mutex_t         start_work_mutex;
 138         int                     nr_tasks_working;
 139
 140         pthread_mutex_t         stop_work_mutex;
 141         u64                     bytes_done;
 142
 143         struct thread_data      *threads;
 144
 145         /* Convergence latency measurement: */
 146         bool                    all_converged;
 147         bool                    stop_work;
 148
 149         int                     print_once;
 150
 151         struct params           p;
 152 };
 153
 154 static struct global_info       *g = NULL;
 155
 156 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 157 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 158
 159 struct params p0;
 160
 161 static const struct option options[] = {
 162         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 163         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 164
 165         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 166         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 167         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 168         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 169
 170         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 171         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 172         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 173
 174         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via writes (can be mixed with -W)"),
 175         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 176         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 177         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 178         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 179
 180
 181         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 182         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 183         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 184         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 185
 186         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 187         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 188         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 189         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
 190         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 191         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 192         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 193
 194         /* Special option string parsing callbacks: */
 195         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 196                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 197                         parse_cpus_opt),
 198         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 199                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 200                         parse_nodes_opt),
 201         OPT_END()
 202 };
 203
 204 static const char * const bench_numa_usage[] = {
 205         "perf bench numa <options>",
 206         NULL
 207 };
 208
 209 static const char * const numa_usage[] = {
 210         "perf bench numa mem [<options>]",
 211         NULL
 212 };
 213
 214 static cpu_set_t bind_to_cpu(int target_cpu)
 215 {
 216         cpu_set_t orig_mask, mask;
 217         int ret;
 218
 219         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 220         BUG_ON(ret);
 221
 222         CPU_ZERO(&mask);
 223
 224         if (target_cpu == -1) {
 225                 int cpu;
 226
 227                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 228                         CPU_SET(cpu, &mask);
 229         } else {
 230                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 231                 CPU_SET(target_cpu, &mask);
 232         }
 233
 234         ret = sched_setaffinity(0, sizeof(mask), &mask);
 235         BUG_ON(ret);
 236
 237         return orig_mask;
 238 }
 239
 240 static cpu_set_t bind_to_node(int target_node)
 241 {
 242         int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 243         cpu_set_t orig_mask, mask;
 244         int cpu;
 245         int ret;
 246
 247         BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 248         BUG_ON(!cpus_per_node);
 249
 250         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 251         BUG_ON(ret);
 252
 253         CPU_ZERO(&mask);
 254
 255         if (target_node == -1) {
 256                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 257                         CPU_SET(cpu, &mask);
 258         } else {
 259                 int cpu_start = (target_node + 0) * cpus_per_node;
 260                 int cpu_stop  = (target_node + 1) * cpus_per_node;
 261
 262                 BUG_ON(cpu_stop > g->p.nr_cpus);
 263
 264                 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 265                         CPU_SET(cpu, &mask);
 266         }
 267
 268         ret = sched_setaffinity(0, sizeof(mask), &mask);
 269         BUG_ON(ret);
 270
 271         return orig_mask;
 272 }
 273
 274 static void bind_to_cpumask(cpu_set_t mask)
 275 {
 276         int ret;
 277
 278         ret = sched_setaffinity(0, sizeof(mask), &mask);
 279         BUG_ON(ret);
 280 }
 281
 282 static void mempol_restore(void)
 283 {
 284         int ret;
 285
 286         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 287
 288         BUG_ON(ret);
 289 }
 290
 291 static void bind_to_memnode(int node)
 292 {
 293         unsigned long nodemask;
 294         int ret;
 295
 296         if (node == -1)
 297                 return;
 298
 299         BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 300         nodemask = 1L << node;
 301
 302         ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 303         dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 304
 305         BUG_ON(ret);
 306 }
 307
 308 #define HPSIZE (2*1024*1024)
 309
 310 #define set_taskname(fmt...)                            \
 311 do {                                                    \
 312         char name[20];                                  \
 313                                                         \
 314         snprintf(name, 20, fmt);                        \
 315         prctl(PR_SET_NAME, name);                       \
 316 } while (0)
 317
 318 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 319                       int init_zero, int init_cpu0, int thp, int init_random)
 320 {
 321         cpu_set_t orig_mask;
 322         ssize_t bytes;
 323         u8 *buf;
 324         int ret;
 325
 326         if (!bytes0)
 327                 return NULL;
 328
 329         /* Allocate and initialize all memory on CPU#0: */
 330         if (init_cpu0) {
 331                 orig_mask = bind_to_node(0);
 332                 bind_to_memnode(0);
 333         }
 334
 335         bytes = bytes0 + HPSIZE;
 336
 337         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 338         BUG_ON(buf == (void *)-1);
 339
 340         if (map_flags == MAP_PRIVATE) {
 341                 if (thp > 0) {
 342                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 343                         if (ret && !g->print_once) {
 344                                 g->print_once = 1;
 345                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 346                         }
 347                 }
 348                 if (thp < 0) {
 349                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 350                         if (ret && !g->print_once) {
 351                                 g->print_once = 1;
 352                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 353                         }
 354                 }
 355         }
 356
 357         if (init_zero) {
 358                 bzero(buf, bytes);
 359         } else {
 360                 /* Initialize random contents, different in each word: */
 361                 if (init_random) {
 362                         u64 *wbuf = (void *)buf;
 363                         long off = rand();
 364                         long i;
 365
 366                         for (i = 0; i < bytes/8; i++)
 367                                 wbuf[i] = i + off;
 368                 }
 369         }
 370
 371         /* Align to 2MB boundary: */
 372         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 373
 374         /* Restore affinity: */
 375         if (init_cpu0) {
 376                 bind_to_cpumask(orig_mask);
 377                 mempol_restore();
 378         }
 379
 380         return buf;
 381 }
 382
 383 static void free_data(void *data, ssize_t bytes)
 384 {
 385         int ret;
 386
 387         if (!data)
 388                 return;
 389
 390         ret = munmap(data, bytes);
 391         BUG_ON(ret);
 392 }
 393
 394 /*
 395  * Create a shared memory buffer that can be shared between processes, zeroed:
 396  */
 397 static void * zalloc_shared_data(ssize_t bytes)
 398 {
 399         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 400 }
 401
 402 /*
 403  * Create a shared memory buffer that can be shared between processes:
 404  */
 405 static void * setup_shared_data(ssize_t bytes)
 406 {
 407         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 408 }
 409
 410 /*
 411  * Allocate process-local memory - this will either be shared between
 412  * threads of this process, or only be accessed by this thread:
 413  */
 414 static void * setup_private_data(ssize_t bytes)
 415 {
 416         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 417 }
 418
 419 /*
 420  * Return a process-shared (global) mutex:
 421  */
 422 static void init_global_mutex(pthread_mutex_t *mutex)
 423 {
 424         pthread_mutexattr_t attr;
 425
 426         pthread_mutexattr_init(&attr);
 427         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 428         pthread_mutex_init(mutex, &attr);
 429 }
 430
 431 static int parse_cpu_list(const char *arg)
 432 {
 433         p0.cpu_list_str = strdup(arg);
 434
 435         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 436
 437         return 0;
 438 }
 439
 440 static int parse_setup_cpu_list(void)
 441 {
 442         struct thread_data *td;
 443         char *str0, *str;
 444         int t;
 445
 446         if (!g->p.cpu_list_str)
 447                 return 0;
 448
 449         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 450
 451         str0 = str = strdup(g->p.cpu_list_str);
 452         t = 0;
 453
 454         BUG_ON(!str);
 455
 456         tprintf("# binding tasks to CPUs:\n");
 457         tprintf("#  ");
 458
 459         while (true) {
 460                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 461                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 462                 int bind_len;
 463                 int step;
 464                 int mul;
 465
 466                 tok = strsep(&str, ",");
 467                 if (!tok)
 468                         break;
 469
 470                 tok_end = strstr(tok, "-");
 471
 472                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 473                 if (!tok_end) {
 474                         /* Single CPU specified: */
 475                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 476                 } else {
 477                         /* CPU range specified (for example: "5-11"): */
 478                         bind_cpu_0 = atol(tok);
 479                         bind_cpu_1 = atol(tok_end + 1);
 480                 }
 481
 482                 step = 1;
 483                 tok_step = strstr(tok, "#");
 484                 if (tok_step) {
 485                         step = atol(tok_step + 1);
 486                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 487                 }
 488
 489                 /*
 490                  * Mask length.
 491                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 492                  * where the _4 means the next 4 CPUs are allowed.
 493                  */
 494                 bind_len = 1;
 495                 tok_len = strstr(tok, "_");
 496                 if (tok_len) {
 497                         bind_len = atol(tok_len + 1);
 498                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 499                 }
 500
 501                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 502                 mul = 1;
 503                 tok_mul = strstr(tok, "x");
 504                 if (tok_mul) {
 505                         mul = atol(tok_mul + 1);
 506                         BUG_ON(mul <= 0);
 507                 }
 508
 509                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 510
 511                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 512                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 513                         return -1;
 514                 }
 515
 516                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 517                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 518
 519                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 520                         int i;
 521
 522                         for (i = 0; i < mul; i++) {
 523                                 int cpu;
 524
 525                                 if (t >= g->p.nr_tasks) {
 526                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 527                                         goto out;
 528                                 }
 529                                 td = g->threads + t;
 530
 531                                 if (t)
 532                                         tprintf(",");
 533                                 if (bind_len > 1) {
 534                                         tprintf("%2d/%d", bind_cpu, bind_len);
 535                                 } else {
 536                                         tprintf("%2d", bind_cpu);
 537                                 }
 538
 539                                 CPU_ZERO(&td->bind_cpumask);
 540                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 541                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 542                                         CPU_SET(cpu, &td->bind_cpumask);
 543                                 }
 544                                 t++;
 545                         }
 546                 }
 547         }
 548 out:
 549
 550         tprintf("\n");
 551
 552         if (t < g->p.nr_tasks)
 553                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 554
 555         free(str0);
 556         return 0;
 557 }
 558
 559 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 560                           const char *arg, int unset __maybe_unused)
 561 {
 562         if (!arg)
 563                 return -1;
 564
 565         return parse_cpu_list(arg);
 566 }
 567
 568 static int parse_node_list(const char *arg)
 569 {
 570         p0.node_list_str = strdup(arg);
 571
 572         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 573
 574         return 0;
 575 }
 576
 577 static int parse_setup_node_list(void)
 578 {
 579         struct thread_data *td;
 580         char *str0, *str;
 581         int t;
 582
 583         if (!g->p.node_list_str)
 584                 return 0;
 585
 586         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 587
 588         str0 = str = strdup(g->p.node_list_str);
 589         t = 0;
 590
 591         BUG_ON(!str);
 592
 593         tprintf("# binding tasks to NODEs:\n");
 594         tprintf("# ");
 595
 596         while (true) {
 597                 int bind_node, bind_node_0, bind_node_1;
 598                 char *tok, *tok_end, *tok_step, *tok_mul;
 599                 int step;
 600                 int mul;
 601
 602                 tok = strsep(&str, ",");
 603                 if (!tok)
 604                         break;
 605
 606                 tok_end = strstr(tok, "-");
 607
 608                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 609                 if (!tok_end) {
 610                         /* Single NODE specified: */
 611                         bind_node_0 = bind_node_1 = atol(tok);
 612                 } else {
 613                         /* NODE range specified (for example: "5-11"): */
 614                         bind_node_0 = atol(tok);
 615                         bind_node_1 = atol(tok_end + 1);
 616                 }
 617
 618                 step = 1;
 619                 tok_step = strstr(tok, "#");
 620                 if (tok_step) {
 621                         step = atol(tok_step + 1);
 622                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 623                 }
 624
 625                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 626                 mul = 1;
 627                 tok_mul = strstr(tok, "x");
 628                 if (tok_mul) {
 629                         mul = atol(tok_mul + 1);
 630                         BUG_ON(mul <= 0);
 631                 }
 632
 633                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 634
 635                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 636                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 637                         return -1;
 638                 }
 639
 640                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 641                 BUG_ON(bind_node_0 > bind_node_1);
 642
 643                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 644                         int i;
 645
 646                         for (i = 0; i < mul; i++) {
 647                                 if (t >= g->p.nr_tasks) {
 648                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 649                                         goto out;
 650                                 }
 651                                 td = g->threads + t;
 652
 653                                 if (!t)
 654                                         tprintf(" %2d", bind_node);
 655                                 else
 656                                         tprintf(",%2d", bind_node);
 657
 658                                 td->bind_node = bind_node;
 659                                 t++;
 660                         }
 661                 }
 662         }
 663 out:
 664
 665         tprintf("\n");
 666
 667         if (t < g->p.nr_tasks)
 668                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 669
 670         free(str0);
 671         return 0;
 672 }
 673
 674 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 675                           const char *arg, int unset __maybe_unused)
 676 {
 677         if (!arg)
 678                 return -1;
 679
 680         return parse_node_list(arg);
 681
 682         return 0;
 683 }
 684
 685 #define BIT(x) (1ul << x)
 686
 687 static inline uint32_t lfsr_32(uint32_t lfsr)
 688 {
 689         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 690         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 691 }
 692
 693 /*
 694  * Make sure there's real data dependency to RAM (when read
 695  * accesses are enabled), so the compiler, the CPU and the
 696  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 697  * accesses:
 698  */
 699 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 700 {
 701         if (g->p.data_reads)
 702                 val += *data;
 703         if (g->p.data_writes)
 704                 *data = val + 1;
 705         return val;
 706 }
 707
 708 /*
 709  * The worker process does two types of work, a forwards going
 710  * loop and a backwards going loop.
 711  *
 712  * We do this so that on multiprocessor systems we do not create
 713  * a 'train' of processing, with highly synchronized processes,
 714  * skewing the whole benchmark.
 715  */
 716 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 717 {
 718         long words = bytes/sizeof(u64);
 719         u64 *data = (void *)__data;
 720         long chunk_0, chunk_1;
 721         u64 *d0, *d, *d1;
 722         long off;
 723         long i;
 724
 725         BUG_ON(!data && words);
 726         BUG_ON(data && !words);
 727
 728         if (!data)
 729                 return val;
 730
 731         /* Very simple memset() work variant: */
 732         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 733                 bzero(data, bytes);
 734                 return val;
 735         }
 736
 737         /* Spread out by PID/TID nr and by loop nr: */
 738         chunk_0 = words/nr_max;
 739         chunk_1 = words/g->p.nr_loops;
 740         off = nr*chunk_0 + loop*chunk_1;
 741
 742         while (off >= words)
 743                 off -= words;
 744
 745         if (g->p.data_rand_walk) {
 746                 u32 lfsr = nr + loop + val;
 747                 int j;
 748
 749                 for (i = 0; i < words/1024; i++) {
 750                         long start, end;
 751
 752                         lfsr = lfsr_32(lfsr);
 753
 754                         start = lfsr % words;
 755                         end = min(start + 1024, words-1);
 756
 757                         if (g->p.data_zero_memset) {
 758                                 bzero(data + start, (end-start) * sizeof(u64));
 759                         } else {
 760                                 for (j = start; j < end; j++)
 761                                         val = access_data(data + j, val);
 762                         }
 763                 }
 764         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 765
 766                 d0 = data + off;
 767                 d  = data + off + 1;
 768                 d1 = data + words;
 769
 770                 /* Process data forwards: */
 771                 for (;;) {
 772                         if (unlikely(d >= d1))
 773                                 d = data;
 774                         if (unlikely(d == d0))
 775                                 break;
 776
 777                         val = access_data(d, val);
 778
 779                         d++;
 780                 }
 781         } else {
 782                 /* Process data backwards: */
 783
 784                 d0 = data + off;
 785                 d  = data + off - 1;
 786                 d1 = data + words;
 787
 788                 /* Process data forwards: */
 789                 for (;;) {
 790                         if (unlikely(d < data))
 791                                 d = data + words-1;
 792                         if (unlikely(d == d0))
 793                                 break;
 794
 795                         val = access_data(d, val);
 796
 797                         d--;
 798                 }
 799         }
 800
 801         return val;
 802 }
 803
 804 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 805 {
 806         unsigned int cpu;
 807
 808         cpu = sched_getcpu();
 809
 810         g->threads[task_nr].curr_cpu = cpu;
 811         prctl(0, bytes_worked);
 812 }
 813
 814 #define MAX_NR_NODES    64
 815
 816 /*
 817  * Count the number of nodes a process's threads
 818  * are spread out on.
 819  *
 820  * A count of 1 means that the process is compressed
 821  * to a single node. A count of g->p.nr_nodes means it's
 822  * spread out on the whole system.
 823  */
 824 static int count_process_nodes(int process_nr)
 825 {
 826         char node_present[MAX_NR_NODES] = { 0, };
 827         int nodes;
 828         int n, t;
 829
 830         for (t = 0; t < g->p.nr_threads; t++) {
 831                 struct thread_data *td;
 832                 int task_nr;
 833                 int node;
 834
 835                 task_nr = process_nr*g->p.nr_threads + t;
 836                 td = g->threads + task_nr;
 837
 838                 node = numa_node_of_cpu(td->curr_cpu);
 839                 if (node < 0) /* curr_cpu was likely still -1 */
 840                         return 0;
 841
 842                 node_present[node] = 1;
 843         }
 844
 845         nodes = 0;
 846
 847         for (n = 0; n < MAX_NR_NODES; n++)
 848                 nodes += node_present[n];
 849
 850         return nodes;
 851 }
 852
 853 /*
 854  * Count the number of distinct process-threads a node contains.
 855  *
 856  * A count of 1 means that the node contains only a single
 857  * process. If all nodes on the system contain at most one
 858  * process then we are well-converged.
 859  */
 860 static int count_node_processes(int node)
 861 {
 862         int processes = 0;
 863         int t, p;
 864
 865         for (p = 0; p < g->p.nr_proc; p++) {
 866                 for (t = 0; t < g->p.nr_threads; t++) {
 867                         struct thread_data *td;
 868                         int task_nr;
 869                         int n;
 870
 871                         task_nr = p*g->p.nr_threads + t;
 872                         td = g->threads + task_nr;
 873
 874                         n = numa_node_of_cpu(td->curr_cpu);
 875                         if (n == node) {
 876                                 processes++;
 877                                 break;
 878                         }
 879                 }
 880         }
 881
 882         return processes;
 883 }
 884
 885 static void calc_convergence_compression(int *strong)
 886 {
 887         unsigned int nodes_min, nodes_max;
 888         int p;
 889
 890         nodes_min = -1;
 891         nodes_max =  0;
 892
 893         for (p = 0; p < g->p.nr_proc; p++) {
 894                 unsigned int nodes = count_process_nodes(p);
 895
 896                 if (!nodes) {
 897                         *strong = 0;
 898                         return;
 899                 }
 900
 901                 nodes_min = min(nodes, nodes_min);
 902                 nodes_max = max(nodes, nodes_max);
 903         }
 904
 905         /* Strong convergence: all threads compress on a single node: */
 906         if (nodes_min == 1 && nodes_max == 1) {
 907                 *strong = 1;
 908         } else {
 909                 *strong = 0;
 910                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 911         }
 912 }
 913
 914 static void calc_convergence(double runtime_ns_max, double *convergence)
 915 {
 916         unsigned int loops_done_min, loops_done_max;
 917         int process_groups;
 918         int nodes[MAX_NR_NODES];
 919         int distance;
 920         int nr_min;
 921         int nr_max;
 922         int strong;
 923         int sum;
 924         int nr;
 925         int node;
 926         int cpu;
 927         int t;
 928
 929         if (!g->p.show_convergence && !g->p.measure_convergence)
 930                 return;
 931
 932         for (node = 0; node < g->p.nr_nodes; node++)
 933                 nodes[node] = 0;
 934
 935         loops_done_min = -1;
 936         loops_done_max = 0;
 937
 938         for (t = 0; t < g->p.nr_tasks; t++) {
 939                 struct thread_data *td = g->threads + t;
 940                 unsigned int loops_done;
 941
 942                 cpu = td->curr_cpu;
 943
 944                 /* Not all threads have written it yet: */
 945                 if (cpu < 0)
 946                         continue;
 947
 948                 node = numa_node_of_cpu(cpu);
 949
 950                 nodes[node]++;
 951
 952                 loops_done = td->loops_done;
 953                 loops_done_min = min(loops_done, loops_done_min);
 954                 loops_done_max = max(loops_done, loops_done_max);
 955         }
 956
 957         nr_max = 0;
 958         nr_min = g->p.nr_tasks;
 959         sum = 0;
 960
 961         for (node = 0; node < g->p.nr_nodes; node++) {
 962                 nr = nodes[node];
 963                 nr_min = min(nr, nr_min);
 964                 nr_max = max(nr, nr_max);
 965                 sum += nr;
 966         }
 967         BUG_ON(nr_min > nr_max);
 968
 969         BUG_ON(sum > g->p.nr_tasks);
 970
 971         if (0 && (sum < g->p.nr_tasks))
 972                 return;
 973
 974         /*
 975          * Count the number of distinct process groups present
 976          * on nodes - when we are converged this will decrease
 977          * to g->p.nr_proc:
 978          */
 979         process_groups = 0;
 980
 981         for (node = 0; node < g->p.nr_nodes; node++) {
 982                 int processes = count_node_processes(node);
 983
 984                 nr = nodes[node];
 985                 tprintf(" %2d/%-2d", nr, processes);
 986
 987                 process_groups += processes;
 988         }
 989
 990         distance = nr_max - nr_min;
 991
 992         tprintf(" [%2d/%-2d]", distance, process_groups);
 993
 994         tprintf(" l:%3d-%-3d (%3d)",
 995                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 996
 997         if (loops_done_min && loops_done_max) {
 998                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
 999
1000                 tprintf(" [%4.1f%%]", skew * 100.0);
1001         }
1002
1003         calc_convergence_compression(&strong);
1004
1005         if (strong && process_groups == g->p.nr_proc) {
1006                 if (!*convergence) {
1007                         *convergence = runtime_ns_max;
1008                         tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1009                         if (g->p.measure_convergence) {
1010                                 g->all_converged = true;
1011                                 g->stop_work = true;
1012                         }
1013                 }
1014         } else {
1015                 if (*convergence) {
1016                         tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1017                         *convergence = 0;
1018                 }
1019                 tprintf("\n");
1020         }
1021 }
1022
1023 static void show_summary(double runtime_ns_max, int l, double *convergence)
1024 {
1025         tprintf("\r #  %5.1f%%  [%.1f mins]",
1026                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1027
1028         calc_convergence(runtime_ns_max, convergence);
1029
1030         if (g->p.show_details >= 0)
1031                 fflush(stdout);
1032 }
1033
1034 static void *worker_thread(void *__tdata)
1035 {
1036         struct thread_data *td = __tdata;
1037         struct timeval start0, start, stop, diff;
1038         int process_nr = td->process_nr;
1039         int thread_nr = td->thread_nr;
1040         unsigned long last_perturbance;
1041         int task_nr = td->task_nr;
1042         int details = g->p.show_details;
1043         int first_task, last_task;
1044         double convergence = 0;
1045         u64 val = td->val;
1046         double runtime_ns_max;
1047         u8 *global_data;
1048         u8 *process_data;
1049         u8 *thread_data;
1050         u64 bytes_done;
1051         long work_done;
1052         u32 l;
1053         struct rusage rusage;
1054
1055         bind_to_cpumask(td->bind_cpumask);
1056         bind_to_memnode(td->bind_node);
1057
1058         set_taskname("thread %d/%d", process_nr, thread_nr);
1059
1060         global_data = g->data;
1061         process_data = td->process_data;
1062         thread_data = setup_private_data(g->p.bytes_thread);
1063
1064         bytes_done = 0;
1065
1066         last_task = 0;
1067         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1068                 last_task = 1;
1069
1070         first_task = 0;
1071         if (process_nr == 0 && thread_nr == 0)
1072                 first_task = 1;
1073
1074         if (details >= 2) {
1075                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1076                         process_nr, thread_nr, global_data, process_data, thread_data);
1077         }
1078
1079         if (g->p.serialize_startup) {
1080                 pthread_mutex_lock(&g->startup_mutex);
1081                 g->nr_tasks_started++;
1082                 pthread_mutex_unlock(&g->startup_mutex);
1083
1084                 /* Here we will wait for the main process to start us all at once: */
1085                 pthread_mutex_lock(&g->start_work_mutex);
1086                 g->nr_tasks_working++;
1087
1088                 /* Last one wake the main process: */
1089                 if (g->nr_tasks_working == g->p.nr_tasks)
1090                         pthread_mutex_unlock(&g->startup_done_mutex);
1091
1092                 pthread_mutex_unlock(&g->start_work_mutex);
1093         }
1094
1095         gettimeofday(&start0, NULL);
1096
1097         start = stop = start0;
1098         last_perturbance = start.tv_sec;
1099
1100         for (l = 0; l < g->p.nr_loops; l++) {
1101                 start = stop;
1102
1103                 if (g->stop_work)
1104                         break;
1105
1106                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1107                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1108                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1109
1110                 if (g->p.sleep_usecs) {
1111                         pthread_mutex_lock(td->process_lock);
1112                         usleep(g->p.sleep_usecs);
1113                         pthread_mutex_unlock(td->process_lock);
1114                 }
1115                 /*
1116                  * Amount of work to be done under a process-global lock:
1117                  */
1118                 if (g->p.bytes_process_locked) {
1119                         pthread_mutex_lock(td->process_lock);
1120                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1121                         pthread_mutex_unlock(td->process_lock);
1122                 }
1123
1124                 work_done = g->p.bytes_global + g->p.bytes_process +
1125                             g->p.bytes_process_locked + g->p.bytes_thread;
1126
1127                 update_curr_cpu(task_nr, work_done);
1128                 bytes_done += work_done;
1129
1130                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1131                         continue;
1132
1133                 td->loops_done = l;
1134
1135                 gettimeofday(&stop, NULL);
1136
1137                 /* Check whether our max runtime timed out: */
1138                 if (g->p.nr_secs) {
1139                         timersub(&stop, &start0, &diff);
1140                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1141                                 g->stop_work = true;
1142                                 break;
1143                         }
1144                 }
1145
1146                 /* Update the summary at most once per second: */
1147                 if (start.tv_sec == stop.tv_sec)
1148                         continue;
1149
1150                 /*
1151                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1152                  * by migrating to CPU#0:
1153                  */
1154                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1155                         cpu_set_t orig_mask;
1156                         int target_cpu;
1157                         int this_cpu;
1158
1159                         last_perturbance = stop.tv_sec;
1160
1161                         /*
1162                          * Depending on where we are running, move into
1163                          * the other half of the system, to create some
1164                          * real disturbance:
1165                          */
1166                         this_cpu = g->threads[task_nr].curr_cpu;
1167                         if (this_cpu < g->p.nr_cpus/2)
1168                                 target_cpu = g->p.nr_cpus-1;
1169                         else
1170                                 target_cpu = 0;
1171
1172                         orig_mask = bind_to_cpu(target_cpu);
1173
1174                         /* Here we are running on the target CPU already */
1175                         if (details >= 1)
1176                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1177
1178                         bind_to_cpumask(orig_mask);
1179                 }
1180
1181                 if (details >= 3) {
1182                         timersub(&stop, &start, &diff);
1183                         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1184                         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1185
1186                         if (details >= 0) {
1187                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1188                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1189                         }
1190                         fflush(stdout);
1191                 }
1192                 if (!last_task)
1193                         continue;
1194
1195                 timersub(&stop, &start0, &diff);
1196                 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1197                 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1198
1199                 show_summary(runtime_ns_max, l, &convergence);
1200         }
1201
1202         gettimeofday(&stop, NULL);
1203         timersub(&stop, &start0, &diff);
1204         td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1205         td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1206         td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1207
1208         getrusage(RUSAGE_THREAD, &rusage);
1209         td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1210         td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1211         td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1212         td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1213
1214         free_data(thread_data, g->p.bytes_thread);
1215
1216         pthread_mutex_lock(&g->stop_work_mutex);
1217         g->bytes_done += bytes_done;
1218         pthread_mutex_unlock(&g->stop_work_mutex);
1219
1220         return NULL;
1221 }
1222
1223 /*
1224  * A worker process starts a couple of threads:
1225  */
1226 static void worker_process(int process_nr)
1227 {
1228         pthread_mutex_t process_lock;
1229         struct thread_data *td;
1230         pthread_t *pthreads;
1231         u8 *process_data;
1232         int task_nr;
1233         int ret;
1234         int t;
1235
1236         pthread_mutex_init(&process_lock, NULL);
1237         set_taskname("process %d", process_nr);
1238
1239         /*
1240          * Pick up the memory policy and the CPU binding of our first thread,
1241          * so that we initialize memory accordingly:
1242          */
1243         task_nr = process_nr*g->p.nr_threads;
1244         td = g->threads + task_nr;
1245
1246         bind_to_memnode(td->bind_node);
1247         bind_to_cpumask(td->bind_cpumask);
1248
1249         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1250         process_data = setup_private_data(g->p.bytes_process);
1251
1252         if (g->p.show_details >= 3) {
1253                 printf(" # process %2d global mem: %p, process mem: %p\n",
1254                         process_nr, g->data, process_data);
1255         }
1256
1257         for (t = 0; t < g->p.nr_threads; t++) {
1258                 task_nr = process_nr*g->p.nr_threads + t;
1259                 td = g->threads + task_nr;
1260
1261                 td->process_data = process_data;
1262                 td->process_nr   = process_nr;
1263                 td->thread_nr    = t;
1264                 td->task_nr      = task_nr;
1265                 td->val          = rand();
1266                 td->curr_cpu     = -1;
1267                 td->process_lock = &process_lock;
1268
1269                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1270                 BUG_ON(ret);
1271         }
1272
1273         for (t = 0; t < g->p.nr_threads; t++) {
1274                 ret = pthread_join(pthreads[t], NULL);
1275                 BUG_ON(ret);
1276         }
1277
1278         free_data(process_data, g->p.bytes_process);
1279         free(pthreads);
1280 }
1281
1282 static void print_summary(void)
1283 {
1284         if (g->p.show_details < 0)
1285                 return;
1286
1287         printf("\n ###\n");
1288         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1289                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1290         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1291                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1292         printf(" #      %5dx %5ldMB process shared mem operations\n",
1293                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1294         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1295                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1296
1297         printf(" ###\n");
1298
1299         printf("\n ###\n"); fflush(stdout);
1300 }
1301
1302 static void init_thread_data(void)
1303 {
1304         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1305         int t;
1306
1307         g->threads = zalloc_shared_data(size);
1308
1309         for (t = 0; t < g->p.nr_tasks; t++) {
1310                 struct thread_data *td = g->threads + t;
1311                 int cpu;
1312
1313                 /* Allow all nodes by default: */
1314                 td->bind_node = -1;
1315
1316                 /* Allow all CPUs by default: */
1317                 CPU_ZERO(&td->bind_cpumask);
1318                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1319                         CPU_SET(cpu, &td->bind_cpumask);
1320         }
1321 }
1322
1323 static void deinit_thread_data(void)
1324 {
1325         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1326
1327         free_data(g->threads, size);
1328 }
1329
1330 static int init(void)
1331 {
1332         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1333
1334         /* Copy over options: */
1335         g->p = p0;
1336
1337         g->p.nr_cpus = numa_num_configured_cpus();
1338
1339         g->p.nr_nodes = numa_max_node() + 1;
1340
1341         /* char array in count_process_nodes(): */
1342         BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1343
1344         if (g->p.show_quiet && !g->p.show_details)
1345                 g->p.show_details = -1;
1346
1347         /* Some memory should be specified: */
1348         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1349                 return -1;
1350
1351         if (g->p.mb_global_str) {
1352                 g->p.mb_global = atof(g->p.mb_global_str);
1353                 BUG_ON(g->p.mb_global < 0);
1354         }
1355
1356         if (g->p.mb_proc_str) {
1357                 g->p.mb_proc = atof(g->p.mb_proc_str);
1358                 BUG_ON(g->p.mb_proc < 0);
1359         }
1360
1361         if (g->p.mb_proc_locked_str) {
1362                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1363                 BUG_ON(g->p.mb_proc_locked < 0);
1364                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1365         }
1366
1367         if (g->p.mb_thread_str) {
1368                 g->p.mb_thread = atof(g->p.mb_thread_str);
1369                 BUG_ON(g->p.mb_thread < 0);
1370         }
1371
1372         BUG_ON(g->p.nr_threads <= 0);
1373         BUG_ON(g->p.nr_proc <= 0);
1374
1375         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1376
1377         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1378         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1379         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1380         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1381
1382         g->data = setup_shared_data(g->p.bytes_global);
1383
1384         /* Startup serialization: */
1385         init_global_mutex(&g->start_work_mutex);
1386         init_global_mutex(&g->startup_mutex);
1387         init_global_mutex(&g->startup_done_mutex);
1388         init_global_mutex(&g->stop_work_mutex);
1389
1390         init_thread_data();
1391
1392         tprintf("#\n");
1393         if (parse_setup_cpu_list() || parse_setup_node_list())
1394                 return -1;
1395         tprintf("#\n");
1396
1397         print_summary();
1398
1399         return 0;
1400 }
1401
1402 static void deinit(void)
1403 {
1404         free_data(g->data, g->p.bytes_global);
1405         g->data = NULL;
1406
1407         deinit_thread_data();
1408
1409         free_data(g, sizeof(*g));
1410         g = NULL;
1411 }
1412
1413 /*
1414  * Print a short or long result, depending on the verbosity setting:
1415  */
1416 static void print_res(const char *name, double val,
1417                       const char *txt_unit, const char *txt_short, const char *txt_long)
1418 {
1419         if (!name)
1420                 name = "main,";
1421
1422         if (!g->p.show_quiet)
1423                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1424         else
1425                 printf(" %14.3f %s\n", val, txt_long);
1426 }
1427
1428 static int __bench_numa(const char *name)
1429 {
1430         struct timeval start, stop, diff;
1431         u64 runtime_ns_min, runtime_ns_sum;
1432         pid_t *pids, pid, wpid;
1433         double delta_runtime;
1434         double runtime_avg;
1435         double runtime_sec_max;
1436         double runtime_sec_min;
1437         int wait_stat;
1438         double bytes;
1439         int i, t, p;
1440
1441         if (init())
1442                 return -1;
1443
1444         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1445         pid = -1;
1446
1447         /* All threads try to acquire it, this way we can wait for them to start up: */
1448         pthread_mutex_lock(&g->start_work_mutex);
1449
1450         if (g->p.serialize_startup) {
1451                 tprintf(" #\n");
1452                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1453         }
1454
1455         gettimeofday(&start, NULL);
1456
1457         for (i = 0; i < g->p.nr_proc; i++) {
1458                 pid = fork();
1459                 dprintf(" # process %2d: PID %d\n", i, pid);
1460
1461                 BUG_ON(pid < 0);
1462                 if (!pid) {
1463                         /* Child process: */
1464                         worker_process(i);
1465
1466                         exit(0);
1467                 }
1468                 pids[i] = pid;
1469
1470         }
1471         /* Wait for all the threads to start up: */
1472         while (g->nr_tasks_started != g->p.nr_tasks)
1473                 usleep(USEC_PER_MSEC);
1474
1475         BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1476
1477         if (g->p.serialize_startup) {
1478                 double startup_sec;
1479
1480                 pthread_mutex_lock(&g->startup_done_mutex);
1481
1482                 /* This will start all threads: */
1483                 pthread_mutex_unlock(&g->start_work_mutex);
1484
1485                 /* This mutex is locked - the last started thread will wake us: */
1486                 pthread_mutex_lock(&g->startup_done_mutex);
1487
1488                 gettimeofday(&stop, NULL);
1489
1490                 timersub(&stop, &start, &diff);
1491
1492                 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1493                 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1494                 startup_sec /= NSEC_PER_SEC;
1495
1496                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1497                 tprintf(" #\n");
1498
1499                 start = stop;
1500                 pthread_mutex_unlock(&g->startup_done_mutex);
1501         } else {
1502                 gettimeofday(&start, NULL);
1503         }
1504
1505         /* Parent process: */
1506
1507
1508         for (i = 0; i < g->p.nr_proc; i++) {
1509                 wpid = waitpid(pids[i], &wait_stat, 0);
1510                 BUG_ON(wpid < 0);
1511                 BUG_ON(!WIFEXITED(wait_stat));
1512
1513         }
1514
1515         runtime_ns_sum = 0;
1516         runtime_ns_min = -1LL;
1517
1518         for (t = 0; t < g->p.nr_tasks; t++) {
1519                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1520
1521                 runtime_ns_sum += thread_runtime_ns;
1522                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1523         }
1524
1525         gettimeofday(&stop, NULL);
1526         timersub(&stop, &start, &diff);
1527
1528         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1529
1530         tprintf("\n ###\n");
1531         tprintf("\n");
1532
1533         runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1534         runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1535         runtime_sec_max /= NSEC_PER_SEC;
1536
1537         runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1538
1539         bytes = g->bytes_done;
1540         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1541
1542         if (g->p.measure_convergence) {
1543                 print_res(name, runtime_sec_max,
1544                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1545         }
1546
1547         print_res(name, runtime_sec_max,
1548                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1549
1550         print_res(name, runtime_sec_min,
1551                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1552
1553         print_res(name, runtime_avg,
1554                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1555
1556         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1557         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1558                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1559
1560         print_res(name, bytes / g->p.nr_tasks / 1e9,
1561                 "GB,", "data/thread",           "GB data processed, per thread");
1562
1563         print_res(name, bytes / 1e9,
1564                 "GB,", "data-total",            "GB data processed, total");
1565
1566         print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1567                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1568
1569         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1570                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1571
1572         print_res(name, bytes / runtime_sec_max / 1e9,
1573                 "GB/sec,", "total-speed",       "GB/sec total speed");
1574
1575         if (g->p.show_details >= 2) {
1576                 char tname[32];
1577                 struct thread_data *td;
1578                 for (p = 0; p < g->p.nr_proc; p++) {
1579                         for (t = 0; t < g->p.nr_threads; t++) {
1580                                 memset(tname, 0, 32);
1581                                 td = g->threads + p*g->p.nr_threads + t;
1582                                 snprintf(tname, 32, "process%d:thread%d", p, t);
1583                                 print_res(tname, td->speed_gbs,
1584                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1585                                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1586                                         "secs", "thread-system-time", "system CPU time/thread");
1587                                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1588                                         "secs", "thread-user-time", "user CPU time/thread");
1589                         }
1590                 }
1591         }
1592
1593         free(pids);
1594
1595         deinit();
1596
1597         return 0;
1598 }
1599
1600 #define MAX_ARGS 50
1601
1602 static int command_size(const char **argv)
1603 {
1604         int size = 0;
1605
1606         while (*argv) {
1607                 size++;
1608                 argv++;
1609         }
1610
1611         BUG_ON(size >= MAX_ARGS);
1612
1613         return size;
1614 }
1615
1616 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1617 {
1618         int i;
1619
1620         printf("\n # Running %s \"perf bench numa", name);
1621
1622         for (i = 0; i < argc; i++)
1623                 printf(" %s", argv[i]);
1624
1625         printf("\"\n");
1626
1627         memset(p, 0, sizeof(*p));
1628
1629         /* Initialize nonzero defaults: */
1630
1631         p->serialize_startup            = 1;
1632         p->data_reads                   = true;
1633         p->data_writes                  = true;
1634         p->data_backwards               = true;
1635         p->data_rand_walk               = true;
1636         p->nr_loops                     = -1;
1637         p->init_random                  = true;
1638         p->mb_global_str                = "1";
1639         p->nr_proc                      = 1;
1640         p->nr_threads                   = 1;
1641         p->nr_secs                      = 5;
1642         p->run_all                      = argc == 1;
1643 }
1644
1645 static int run_bench_numa(const char *name, const char **argv)
1646 {
1647         int argc = command_size(argv);
1648
1649         init_params(&p0, name, argc, argv);
1650         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1651         if (argc)
1652                 goto err;
1653
1654         if (__bench_numa(name))
1655                 goto err;
1656
1657         return 0;
1658
1659 err:
1660         return -1;
1661 }
1662
1663 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1664 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1665
1666 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1667 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1668
1669 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1670 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1671
1672 /*
1673  * The built-in test-suite executed by "perf bench numa -a".
1674  *
1675  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1676  */
1677 static const char *tests[][MAX_ARGS] = {
1678    /* Basic single-stream NUMA bandwidth measurements: */
1679    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1680                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1681    { "RAM-bw-local-NOTHP,",
1682                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1683                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1684    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1685                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1686
1687    /* 2-stream NUMA bandwidth measurements: */
1688    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1689                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1690    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1691                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1692
1693    /* Cross-stream NUMA bandwidth measurement: */
1694    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1695                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1696
1697    /* Convergence latency measurements: */
1698    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1699    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1700    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1701    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1702    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1703    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1704    { " 4x4-convergence-NOTHP,",
1705                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1706    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1707    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1708    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1709    { " 8x4-convergence-NOTHP,",
1710                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1711    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1712    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1713    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1714    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1715    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1716
1717    /* Various NUMA process/thread layout bandwidth measurements: */
1718    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1719    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1720    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1721    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1722    { " 8x1-bw-process-NOTHP,",
1723                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1724    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1725
1726    { " 4x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1727    { " 8x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1728    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1729    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1730
1731    { " 2x3-bw-thread,",   "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1732    { " 4x4-bw-thread,",   "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1733    { " 4x6-bw-thread,",   "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1734    { " 4x8-bw-thread,",   "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1735    { " 4x8-bw-thread-NOTHP,",
1736                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1737    { " 3x3-bw-thread,",   "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1738    { " 5x5-bw-thread,",   "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1739
1740    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1741    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1742
1743    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1744    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1745    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1746    { "numa01-bw-thread-NOTHP,",
1747                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1748 };
1749
1750 static int bench_all(void)
1751 {
1752         int nr = ARRAY_SIZE(tests);
1753         int ret;
1754         int i;
1755
1756         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1757         BUG_ON(ret < 0);
1758
1759         for (i = 0; i < nr; i++) {
1760                 run_bench_numa(tests[i][0], tests[i] + 1);
1761         }
1762
1763         printf("\n");
1764
1765         return 0;
1766 }
1767
1768 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1769 {
1770         init_params(&p0, "main,", argc, argv);
1771         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1772         if (argc)
1773                 goto err;
1774
1775         if (p0.run_all)
1776                 return bench_all();
1777
1778         if (__bench_numa(NULL))
1779                 goto err;
1780
1781         return 0;
1782
1783 err:
1784         usage_with_options(numa_usage, options);
1785         return -1;
1786 }