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 #undef dprintf
  47 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  48
  49 struct thread_data {
  50         int                     curr_cpu;
  51         cpu_set_t               bind_cpumask;
  52         int                     bind_node;
  53         u8                      *process_data;
  54         int                     process_nr;
  55         int                     thread_nr;
  56         int                     task_nr;
  57         unsigned int            loops_done;
  58         u64                     val;
  59         u64                     runtime_ns;
  60         u64                     system_time_ns;
  61         u64                     user_time_ns;
  62         double                  speed_gbs;
  63         pthread_mutex_t         *process_lock;
  64 };
  65
  66 /* Parameters set by options: */
  67
  68 struct params {
  69         /* Startup synchronization: */
  70         bool                    serialize_startup;
  71
  72         /* Task hierarchy: */
  73         int                     nr_proc;
  74         int                     nr_threads;
  75
  76         /* Working set sizes: */
  77         const char              *mb_global_str;
  78         const char              *mb_proc_str;
  79         const char              *mb_proc_locked_str;
  80         const char              *mb_thread_str;
  81
  82         double                  mb_global;
  83         double                  mb_proc;
  84         double                  mb_proc_locked;
  85         double                  mb_thread;
  86
  87         /* Access patterns to the working set: */
  88         bool                    data_reads;
  89         bool                    data_writes;
  90         bool                    data_backwards;
  91         bool                    data_zero_memset;
  92         bool                    data_rand_walk;
  93         u32                     nr_loops;
  94         u32                     nr_secs;
  95         u32                     sleep_usecs;
  96
  97         /* Working set initialization: */
  98         bool                    init_zero;
  99         bool                    init_random;
 100         bool                    init_cpu0;
 101
 102         /* Misc options: */
 103         int                     show_details;
 104         int                     run_all;
 105         int                     thp;
 106
 107         long                    bytes_global;
 108         long                    bytes_process;
 109         long                    bytes_process_locked;
 110         long                    bytes_thread;
 111
 112         int                     nr_tasks;
 113         bool                    show_quiet;
 114
 115         bool                    show_convergence;
 116         bool                    measure_convergence;
 117
 118         int                     perturb_secs;
 119         int                     nr_cpus;
 120         int                     nr_nodes;
 121
 122         /* Affinity options -C and -N: */
 123         char                    *cpu_list_str;
 124         char                    *node_list_str;
 125 };
 126
 127
 128 /* Global, read-writable area, accessible to all processes and threads: */
 129
 130 struct global_info {
 131         u8                      *data;
 132
 133         pthread_mutex_t         startup_mutex;
 134         int                     nr_tasks_started;
 135
 136         pthread_mutex_t         startup_done_mutex;
 137
 138         pthread_mutex_t         start_work_mutex;
 139         int                     nr_tasks_working;
 140
 141         pthread_mutex_t         stop_work_mutex;
 142         u64                     bytes_done;
 143
 144         struct thread_data      *threads;
 145
 146         /* Convergence latency measurement: */
 147         bool                    all_converged;
 148         bool                    stop_work;
 149
 150         int                     print_once;
 151
 152         struct params           p;
 153 };
 154
 155 static struct global_info       *g = NULL;
 156
 157 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 158 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 159
 160 struct params p0;
 161
 162 static const struct option options[] = {
 163         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 164         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 165
 166         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 167         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 168         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 169         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 170
 171         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 172         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 173         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 174
 175         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via writes (can be mixed with -W)"),
 176         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 177         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 178         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 179         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 180
 181
 182         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 183         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 184         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 185         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 186
 187         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 188         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 189         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 190         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
 191                     "convergence is reached when each process (all its threads) is running on a single NUMA node."),
 192         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 193         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 194         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 195
 196         /* Special option string parsing callbacks: */
 197         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 198                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 199                         parse_cpus_opt),
 200         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 201                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 202                         parse_nodes_opt),
 203         OPT_END()
 204 };
 205
 206 static const char * const bench_numa_usage[] = {
 207         "perf bench numa <options>",
 208         NULL
 209 };
 210
 211 static const char * const numa_usage[] = {
 212         "perf bench numa mem [<options>]",
 213         NULL
 214 };
 215
 216 static cpu_set_t bind_to_cpu(int target_cpu)
 217 {
 218         cpu_set_t orig_mask, mask;
 219         int ret;
 220
 221         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 222         BUG_ON(ret);
 223
 224         CPU_ZERO(&mask);
 225
 226         if (target_cpu == -1) {
 227                 int cpu;
 228
 229                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 230                         CPU_SET(cpu, &mask);
 231         } else {
 232                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 233                 CPU_SET(target_cpu, &mask);
 234         }
 235
 236         ret = sched_setaffinity(0, sizeof(mask), &mask);
 237         BUG_ON(ret);
 238
 239         return orig_mask;
 240 }
 241
 242 static cpu_set_t bind_to_node(int target_node)
 243 {
 244         int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 245         cpu_set_t orig_mask, mask;
 246         int cpu;
 247         int ret;
 248
 249         BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 250         BUG_ON(!cpus_per_node);
 251
 252         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 253         BUG_ON(ret);
 254
 255         CPU_ZERO(&mask);
 256
 257         if (target_node == -1) {
 258                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 259                         CPU_SET(cpu, &mask);
 260         } else {
 261                 int cpu_start = (target_node + 0) * cpus_per_node;
 262                 int cpu_stop  = (target_node + 1) * cpus_per_node;
 263
 264                 BUG_ON(cpu_stop > g->p.nr_cpus);
 265
 266                 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 267                         CPU_SET(cpu, &mask);
 268         }
 269
 270         ret = sched_setaffinity(0, sizeof(mask), &mask);
 271         BUG_ON(ret);
 272
 273         return orig_mask;
 274 }
 275
 276 static void bind_to_cpumask(cpu_set_t mask)
 277 {
 278         int ret;
 279
 280         ret = sched_setaffinity(0, sizeof(mask), &mask);
 281         BUG_ON(ret);
 282 }
 283
 284 static void mempol_restore(void)
 285 {
 286         int ret;
 287
 288         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 289
 290         BUG_ON(ret);
 291 }
 292
 293 static void bind_to_memnode(int node)
 294 {
 295         unsigned long nodemask;
 296         int ret;
 297
 298         if (node == -1)
 299                 return;
 300
 301         BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 302         nodemask = 1L << node;
 303
 304         ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 305         dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 306
 307         BUG_ON(ret);
 308 }
 309
 310 #define HPSIZE (2*1024*1024)
 311
 312 #define set_taskname(fmt...)                            \
 313 do {                                                    \
 314         char name[20];                                  \
 315                                                         \
 316         snprintf(name, 20, fmt);                        \
 317         prctl(PR_SET_NAME, name);                       \
 318 } while (0)
 319
 320 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 321                       int init_zero, int init_cpu0, int thp, int init_random)
 322 {
 323         cpu_set_t orig_mask;
 324         ssize_t bytes;
 325         u8 *buf;
 326         int ret;
 327
 328         if (!bytes0)
 329                 return NULL;
 330
 331         /* Allocate and initialize all memory on CPU#0: */
 332         if (init_cpu0) {
 333                 orig_mask = bind_to_node(0);
 334                 bind_to_memnode(0);
 335         }
 336
 337         bytes = bytes0 + HPSIZE;
 338
 339         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 340         BUG_ON(buf == (void *)-1);
 341
 342         if (map_flags == MAP_PRIVATE) {
 343                 if (thp > 0) {
 344                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 345                         if (ret && !g->print_once) {
 346                                 g->print_once = 1;
 347                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 348                         }
 349                 }
 350                 if (thp < 0) {
 351                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 352                         if (ret && !g->print_once) {
 353                                 g->print_once = 1;
 354                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 355                         }
 356                 }
 357         }
 358
 359         if (init_zero) {
 360                 bzero(buf, bytes);
 361         } else {
 362                 /* Initialize random contents, different in each word: */
 363                 if (init_random) {
 364                         u64 *wbuf = (void *)buf;
 365                         long off = rand();
 366                         long i;
 367
 368                         for (i = 0; i < bytes/8; i++)
 369                                 wbuf[i] = i + off;
 370                 }
 371         }
 372
 373         /* Align to 2MB boundary: */
 374         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 375
 376         /* Restore affinity: */
 377         if (init_cpu0) {
 378                 bind_to_cpumask(orig_mask);
 379                 mempol_restore();
 380         }
 381
 382         return buf;
 383 }
 384
 385 static void free_data(void *data, ssize_t bytes)
 386 {
 387         int ret;
 388
 389         if (!data)
 390                 return;
 391
 392         ret = munmap(data, bytes);
 393         BUG_ON(ret);
 394 }
 395
 396 /*
 397  * Create a shared memory buffer that can be shared between processes, zeroed:
 398  */
 399 static void * zalloc_shared_data(ssize_t bytes)
 400 {
 401         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 402 }
 403
 404 /*
 405  * Create a shared memory buffer that can be shared between processes:
 406  */
 407 static void * setup_shared_data(ssize_t bytes)
 408 {
 409         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 410 }
 411
 412 /*
 413  * Allocate process-local memory - this will either be shared between
 414  * threads of this process, or only be accessed by this thread:
 415  */
 416 static void * setup_private_data(ssize_t bytes)
 417 {
 418         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 419 }
 420
 421 /*
 422  * Return a process-shared (global) mutex:
 423  */
 424 static void init_global_mutex(pthread_mutex_t *mutex)
 425 {
 426         pthread_mutexattr_t attr;
 427
 428         pthread_mutexattr_init(&attr);
 429         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 430         pthread_mutex_init(mutex, &attr);
 431 }
 432
 433 static int parse_cpu_list(const char *arg)
 434 {
 435         p0.cpu_list_str = strdup(arg);
 436
 437         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 438
 439         return 0;
 440 }
 441
 442 static int parse_setup_cpu_list(void)
 443 {
 444         struct thread_data *td;
 445         char *str0, *str;
 446         int t;
 447
 448         if (!g->p.cpu_list_str)
 449                 return 0;
 450
 451         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 452
 453         str0 = str = strdup(g->p.cpu_list_str);
 454         t = 0;
 455
 456         BUG_ON(!str);
 457
 458         tprintf("# binding tasks to CPUs:\n");
 459         tprintf("#  ");
 460
 461         while (true) {
 462                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 463                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 464                 int bind_len;
 465                 int step;
 466                 int mul;
 467
 468                 tok = strsep(&str, ",");
 469                 if (!tok)
 470                         break;
 471
 472                 tok_end = strstr(tok, "-");
 473
 474                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 475                 if (!tok_end) {
 476                         /* Single CPU specified: */
 477                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 478                 } else {
 479                         /* CPU range specified (for example: "5-11"): */
 480                         bind_cpu_0 = atol(tok);
 481                         bind_cpu_1 = atol(tok_end + 1);
 482                 }
 483
 484                 step = 1;
 485                 tok_step = strstr(tok, "#");
 486                 if (tok_step) {
 487                         step = atol(tok_step + 1);
 488                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 489                 }
 490
 491                 /*
 492                  * Mask length.
 493                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 494                  * where the _4 means the next 4 CPUs are allowed.
 495                  */
 496                 bind_len = 1;
 497                 tok_len = strstr(tok, "_");
 498                 if (tok_len) {
 499                         bind_len = atol(tok_len + 1);
 500                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 501                 }
 502
 503                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 504                 mul = 1;
 505                 tok_mul = strstr(tok, "x");
 506                 if (tok_mul) {
 507                         mul = atol(tok_mul + 1);
 508                         BUG_ON(mul <= 0);
 509                 }
 510
 511                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 512
 513                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 514                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 515                         return -1;
 516                 }
 517
 518                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 519                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 520
 521                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 522                         int i;
 523
 524                         for (i = 0; i < mul; i++) {
 525                                 int cpu;
 526
 527                                 if (t >= g->p.nr_tasks) {
 528                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 529                                         goto out;
 530                                 }
 531                                 td = g->threads + t;
 532
 533                                 if (t)
 534                                         tprintf(",");
 535                                 if (bind_len > 1) {
 536                                         tprintf("%2d/%d", bind_cpu, bind_len);
 537                                 } else {
 538                                         tprintf("%2d", bind_cpu);
 539                                 }
 540
 541                                 CPU_ZERO(&td->bind_cpumask);
 542                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 543                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 544                                         CPU_SET(cpu, &td->bind_cpumask);
 545                                 }
 546                                 t++;
 547                         }
 548                 }
 549         }
 550 out:
 551
 552         tprintf("\n");
 553
 554         if (t < g->p.nr_tasks)
 555                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 556
 557         free(str0);
 558         return 0;
 559 }
 560
 561 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 562                           const char *arg, int unset __maybe_unused)
 563 {
 564         if (!arg)
 565                 return -1;
 566
 567         return parse_cpu_list(arg);
 568 }
 569
 570 static int parse_node_list(const char *arg)
 571 {
 572         p0.node_list_str = strdup(arg);
 573
 574         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 575
 576         return 0;
 577 }
 578
 579 static int parse_setup_node_list(void)
 580 {
 581         struct thread_data *td;
 582         char *str0, *str;
 583         int t;
 584
 585         if (!g->p.node_list_str)
 586                 return 0;
 587
 588         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 589
 590         str0 = str = strdup(g->p.node_list_str);
 591         t = 0;
 592
 593         BUG_ON(!str);
 594
 595         tprintf("# binding tasks to NODEs:\n");
 596         tprintf("# ");
 597
 598         while (true) {
 599                 int bind_node, bind_node_0, bind_node_1;
 600                 char *tok, *tok_end, *tok_step, *tok_mul;
 601                 int step;
 602                 int mul;
 603
 604                 tok = strsep(&str, ",");
 605                 if (!tok)
 606                         break;
 607
 608                 tok_end = strstr(tok, "-");
 609
 610                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 611                 if (!tok_end) {
 612                         /* Single NODE specified: */
 613                         bind_node_0 = bind_node_1 = atol(tok);
 614                 } else {
 615                         /* NODE range specified (for example: "5-11"): */
 616                         bind_node_0 = atol(tok);
 617                         bind_node_1 = atol(tok_end + 1);
 618                 }
 619
 620                 step = 1;
 621                 tok_step = strstr(tok, "#");
 622                 if (tok_step) {
 623                         step = atol(tok_step + 1);
 624                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 625                 }
 626
 627                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 628                 mul = 1;
 629                 tok_mul = strstr(tok, "x");
 630                 if (tok_mul) {
 631                         mul = atol(tok_mul + 1);
 632                         BUG_ON(mul <= 0);
 633                 }
 634
 635                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 636
 637                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 638                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 639                         return -1;
 640                 }
 641
 642                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 643                 BUG_ON(bind_node_0 > bind_node_1);
 644
 645                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 646                         int i;
 647
 648                         for (i = 0; i < mul; i++) {
 649                                 if (t >= g->p.nr_tasks) {
 650                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 651                                         goto out;
 652                                 }
 653                                 td = g->threads + t;
 654
 655                                 if (!t)
 656                                         tprintf(" %2d", bind_node);
 657                                 else
 658                                         tprintf(",%2d", bind_node);
 659
 660                                 td->bind_node = bind_node;
 661                                 t++;
 662                         }
 663                 }
 664         }
 665 out:
 666
 667         tprintf("\n");
 668
 669         if (t < g->p.nr_tasks)
 670                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 671
 672         free(str0);
 673         return 0;
 674 }
 675
 676 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 677                           const char *arg, int unset __maybe_unused)
 678 {
 679         if (!arg)
 680                 return -1;
 681
 682         return parse_node_list(arg);
 683
 684         return 0;
 685 }
 686
 687 #define BIT(x) (1ul << x)
 688
 689 static inline uint32_t lfsr_32(uint32_t lfsr)
 690 {
 691         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 692         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 693 }
 694
 695 /*
 696  * Make sure there's real data dependency to RAM (when read
 697  * accesses are enabled), so the compiler, the CPU and the
 698  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 699  * accesses:
 700  */
 701 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 702 {
 703         if (g->p.data_reads)
 704                 val += *data;
 705         if (g->p.data_writes)
 706                 *data = val + 1;
 707         return val;
 708 }
 709
 710 /*
 711  * The worker process does two types of work, a forwards going
 712  * loop and a backwards going loop.
 713  *
 714  * We do this so that on multiprocessor systems we do not create
 715  * a 'train' of processing, with highly synchronized processes,
 716  * skewing the whole benchmark.
 717  */
 718 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 719 {
 720         long words = bytes/sizeof(u64);
 721         u64 *data = (void *)__data;
 722         long chunk_0, chunk_1;
 723         u64 *d0, *d, *d1;
 724         long off;
 725         long i;
 726
 727         BUG_ON(!data && words);
 728         BUG_ON(data && !words);
 729
 730         if (!data)
 731                 return val;
 732
 733         /* Very simple memset() work variant: */
 734         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 735                 bzero(data, bytes);
 736                 return val;
 737         }
 738
 739         /* Spread out by PID/TID nr and by loop nr: */
 740         chunk_0 = words/nr_max;
 741         chunk_1 = words/g->p.nr_loops;
 742         off = nr*chunk_0 + loop*chunk_1;
 743
 744         while (off >= words)
 745                 off -= words;
 746
 747         if (g->p.data_rand_walk) {
 748                 u32 lfsr = nr + loop + val;
 749                 int j;
 750
 751                 for (i = 0; i < words/1024; i++) {
 752                         long start, end;
 753
 754                         lfsr = lfsr_32(lfsr);
 755
 756                         start = lfsr % words;
 757                         end = min(start + 1024, words-1);
 758
 759                         if (g->p.data_zero_memset) {
 760                                 bzero(data + start, (end-start) * sizeof(u64));
 761                         } else {
 762                                 for (j = start; j < end; j++)
 763                                         val = access_data(data + j, val);
 764                         }
 765                 }
 766         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 767
 768                 d0 = data + off;
 769                 d  = data + off + 1;
 770                 d1 = data + words;
 771
 772                 /* Process data forwards: */
 773                 for (;;) {
 774                         if (unlikely(d >= d1))
 775                                 d = data;
 776                         if (unlikely(d == d0))
 777                                 break;
 778
 779                         val = access_data(d, val);
 780
 781                         d++;
 782                 }
 783         } else {
 784                 /* Process data backwards: */
 785
 786                 d0 = data + off;
 787                 d  = data + off - 1;
 788                 d1 = data + words;
 789
 790                 /* Process data forwards: */
 791                 for (;;) {
 792                         if (unlikely(d < data))
 793                                 d = data + words-1;
 794                         if (unlikely(d == d0))
 795                                 break;
 796
 797                         val = access_data(d, val);
 798
 799                         d--;
 800                 }
 801         }
 802
 803         return val;
 804 }
 805
 806 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 807 {
 808         unsigned int cpu;
 809
 810         cpu = sched_getcpu();
 811
 812         g->threads[task_nr].curr_cpu = cpu;
 813         prctl(0, bytes_worked);
 814 }
 815
 816 #define MAX_NR_NODES    64
 817
 818 /*
 819  * Count the number of nodes a process's threads
 820  * are spread out on.
 821  *
 822  * A count of 1 means that the process is compressed
 823  * to a single node. A count of g->p.nr_nodes means it's
 824  * spread out on the whole system.
 825  */
 826 static int count_process_nodes(int process_nr)
 827 {
 828         char node_present[MAX_NR_NODES] = { 0, };
 829         int nodes;
 830         int n, t;
 831
 832         for (t = 0; t < g->p.nr_threads; t++) {
 833                 struct thread_data *td;
 834                 int task_nr;
 835                 int node;
 836
 837                 task_nr = process_nr*g->p.nr_threads + t;
 838                 td = g->threads + task_nr;
 839
 840                 node = numa_node_of_cpu(td->curr_cpu);
 841                 if (node < 0) /* curr_cpu was likely still -1 */
 842                         return 0;
 843
 844                 node_present[node] = 1;
 845         }
 846
 847         nodes = 0;
 848
 849         for (n = 0; n < MAX_NR_NODES; n++)
 850                 nodes += node_present[n];
 851
 852         return nodes;
 853 }
 854
 855 /*
 856  * Count the number of distinct process-threads a node contains.
 857  *
 858  * A count of 1 means that the node contains only a single
 859  * process. If all nodes on the system contain at most one
 860  * process then we are well-converged.
 861  */
 862 static int count_node_processes(int node)
 863 {
 864         int processes = 0;
 865         int t, p;
 866
 867         for (p = 0; p < g->p.nr_proc; p++) {
 868                 for (t = 0; t < g->p.nr_threads; t++) {
 869                         struct thread_data *td;
 870                         int task_nr;
 871                         int n;
 872
 873                         task_nr = p*g->p.nr_threads + t;
 874                         td = g->threads + task_nr;
 875
 876                         n = numa_node_of_cpu(td->curr_cpu);
 877                         if (n == node) {
 878                                 processes++;
 879                                 break;
 880                         }
 881                 }
 882         }
 883
 884         return processes;
 885 }
 886
 887 static void calc_convergence_compression(int *strong)
 888 {
 889         unsigned int nodes_min, nodes_max;
 890         int p;
 891
 892         nodes_min = -1;
 893         nodes_max =  0;
 894
 895         for (p = 0; p < g->p.nr_proc; p++) {
 896                 unsigned int nodes = count_process_nodes(p);
 897
 898                 if (!nodes) {
 899                         *strong = 0;
 900                         return;
 901                 }
 902
 903                 nodes_min = min(nodes, nodes_min);
 904                 nodes_max = max(nodes, nodes_max);
 905         }
 906
 907         /* Strong convergence: all threads compress on a single node: */
 908         if (nodes_min == 1 && nodes_max == 1) {
 909                 *strong = 1;
 910         } else {
 911                 *strong = 0;
 912                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 913         }
 914 }
 915
 916 static void calc_convergence(double runtime_ns_max, double *convergence)
 917 {
 918         unsigned int loops_done_min, loops_done_max;
 919         int process_groups;
 920         int nodes[MAX_NR_NODES];
 921         int distance;
 922         int nr_min;
 923         int nr_max;
 924         int strong;
 925         int sum;
 926         int nr;
 927         int node;
 928         int cpu;
 929         int t;
 930
 931         if (!g->p.show_convergence && !g->p.measure_convergence)
 932                 return;
 933
 934         for (node = 0; node < g->p.nr_nodes; node++)
 935                 nodes[node] = 0;
 936
 937         loops_done_min = -1;
 938         loops_done_max = 0;
 939
 940         for (t = 0; t < g->p.nr_tasks; t++) {
 941                 struct thread_data *td = g->threads + t;
 942                 unsigned int loops_done;
 943
 944                 cpu = td->curr_cpu;
 945
 946                 /* Not all threads have written it yet: */
 947                 if (cpu < 0)
 948                         continue;
 949
 950                 node = numa_node_of_cpu(cpu);
 951
 952                 nodes[node]++;
 953
 954                 loops_done = td->loops_done;
 955                 loops_done_min = min(loops_done, loops_done_min);
 956                 loops_done_max = max(loops_done, loops_done_max);
 957         }
 958
 959         nr_max = 0;
 960         nr_min = g->p.nr_tasks;
 961         sum = 0;
 962
 963         for (node = 0; node < g->p.nr_nodes; node++) {
 964                 nr = nodes[node];
 965                 nr_min = min(nr, nr_min);
 966                 nr_max = max(nr, nr_max);
 967                 sum += nr;
 968         }
 969         BUG_ON(nr_min > nr_max);
 970
 971         BUG_ON(sum > g->p.nr_tasks);
 972
 973         if (0 && (sum < g->p.nr_tasks))
 974                 return;
 975
 976         /*
 977          * Count the number of distinct process groups present
 978          * on nodes - when we are converged this will decrease
 979          * to g->p.nr_proc:
 980          */
 981         process_groups = 0;
 982
 983         for (node = 0; node < g->p.nr_nodes; node++) {
 984                 int processes = count_node_processes(node);
 985
 986                 nr = nodes[node];
 987                 tprintf(" %2d/%-2d", nr, processes);
 988
 989                 process_groups += processes;
 990         }
 991
 992         distance = nr_max - nr_min;
 993
 994         tprintf(" [%2d/%-2d]", distance, process_groups);
 995
 996         tprintf(" l:%3d-%-3d (%3d)",
 997                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 998
 999         if (loops_done_min && loops_done_max) {
1000                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1001
1002                 tprintf(" [%4.1f%%]", skew * 100.0);
1003         }
1004
1005         calc_convergence_compression(&strong);
1006
1007         if (strong && process_groups == g->p.nr_proc) {
1008                 if (!*convergence) {
1009                         *convergence = runtime_ns_max;
1010                         tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1011                         if (g->p.measure_convergence) {
1012                                 g->all_converged = true;
1013                                 g->stop_work = true;
1014                         }
1015                 }
1016         } else {
1017                 if (*convergence) {
1018                         tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1019                         *convergence = 0;
1020                 }
1021                 tprintf("\n");
1022         }
1023 }
1024
1025 static void show_summary(double runtime_ns_max, int l, double *convergence)
1026 {
1027         tprintf("\r #  %5.1f%%  [%.1f mins]",
1028                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1029
1030         calc_convergence(runtime_ns_max, convergence);
1031
1032         if (g->p.show_details >= 0)
1033                 fflush(stdout);
1034 }
1035
1036 static void *worker_thread(void *__tdata)
1037 {
1038         struct thread_data *td = __tdata;
1039         struct timeval start0, start, stop, diff;
1040         int process_nr = td->process_nr;
1041         int thread_nr = td->thread_nr;
1042         unsigned long last_perturbance;
1043         int task_nr = td->task_nr;
1044         int details = g->p.show_details;
1045         int first_task, last_task;
1046         double convergence = 0;
1047         u64 val = td->val;
1048         double runtime_ns_max;
1049         u8 *global_data;
1050         u8 *process_data;
1051         u8 *thread_data;
1052         u64 bytes_done;
1053         long work_done;
1054         u32 l;
1055         struct rusage rusage;
1056
1057         bind_to_cpumask(td->bind_cpumask);
1058         bind_to_memnode(td->bind_node);
1059
1060         set_taskname("thread %d/%d", process_nr, thread_nr);
1061
1062         global_data = g->data;
1063         process_data = td->process_data;
1064         thread_data = setup_private_data(g->p.bytes_thread);
1065
1066         bytes_done = 0;
1067
1068         last_task = 0;
1069         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1070                 last_task = 1;
1071
1072         first_task = 0;
1073         if (process_nr == 0 && thread_nr == 0)
1074                 first_task = 1;
1075
1076         if (details >= 2) {
1077                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1078                         process_nr, thread_nr, global_data, process_data, thread_data);
1079         }
1080
1081         if (g->p.serialize_startup) {
1082                 pthread_mutex_lock(&g->startup_mutex);
1083                 g->nr_tasks_started++;
1084                 pthread_mutex_unlock(&g->startup_mutex);
1085
1086                 /* Here we will wait for the main process to start us all at once: */
1087                 pthread_mutex_lock(&g->start_work_mutex);
1088                 g->nr_tasks_working++;
1089
1090                 /* Last one wake the main process: */
1091                 if (g->nr_tasks_working == g->p.nr_tasks)
1092                         pthread_mutex_unlock(&g->startup_done_mutex);
1093
1094                 pthread_mutex_unlock(&g->start_work_mutex);
1095         }
1096
1097         gettimeofday(&start0, NULL);
1098
1099         start = stop = start0;
1100         last_perturbance = start.tv_sec;
1101
1102         for (l = 0; l < g->p.nr_loops; l++) {
1103                 start = stop;
1104
1105                 if (g->stop_work)
1106                         break;
1107
1108                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1109                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1110                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1111
1112                 if (g->p.sleep_usecs) {
1113                         pthread_mutex_lock(td->process_lock);
1114                         usleep(g->p.sleep_usecs);
1115                         pthread_mutex_unlock(td->process_lock);
1116                 }
1117                 /*
1118                  * Amount of work to be done under a process-global lock:
1119                  */
1120                 if (g->p.bytes_process_locked) {
1121                         pthread_mutex_lock(td->process_lock);
1122                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1123                         pthread_mutex_unlock(td->process_lock);
1124                 }
1125
1126                 work_done = g->p.bytes_global + g->p.bytes_process +
1127                             g->p.bytes_process_locked + g->p.bytes_thread;
1128
1129                 update_curr_cpu(task_nr, work_done);
1130                 bytes_done += work_done;
1131
1132                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1133                         continue;
1134
1135                 td->loops_done = l;
1136
1137                 gettimeofday(&stop, NULL);
1138
1139                 /* Check whether our max runtime timed out: */
1140                 if (g->p.nr_secs) {
1141                         timersub(&stop, &start0, &diff);
1142                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1143                                 g->stop_work = true;
1144                                 break;
1145                         }
1146                 }
1147
1148                 /* Update the summary at most once per second: */
1149                 if (start.tv_sec == stop.tv_sec)
1150                         continue;
1151
1152                 /*
1153                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1154                  * by migrating to CPU#0:
1155                  */
1156                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1157                         cpu_set_t orig_mask;
1158                         int target_cpu;
1159                         int this_cpu;
1160
1161                         last_perturbance = stop.tv_sec;
1162
1163                         /*
1164                          * Depending on where we are running, move into
1165                          * the other half of the system, to create some
1166                          * real disturbance:
1167                          */
1168                         this_cpu = g->threads[task_nr].curr_cpu;
1169                         if (this_cpu < g->p.nr_cpus/2)
1170                                 target_cpu = g->p.nr_cpus-1;
1171                         else
1172                                 target_cpu = 0;
1173
1174                         orig_mask = bind_to_cpu(target_cpu);
1175
1176                         /* Here we are running on the target CPU already */
1177                         if (details >= 1)
1178                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1179
1180                         bind_to_cpumask(orig_mask);
1181                 }
1182
1183                 if (details >= 3) {
1184                         timersub(&stop, &start, &diff);
1185                         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1186                         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1187
1188                         if (details >= 0) {
1189                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1190                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1191                         }
1192                         fflush(stdout);
1193                 }
1194                 if (!last_task)
1195                         continue;
1196
1197                 timersub(&stop, &start0, &diff);
1198                 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1199                 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1200
1201                 show_summary(runtime_ns_max, l, &convergence);
1202         }
1203
1204         gettimeofday(&stop, NULL);
1205         timersub(&stop, &start0, &diff);
1206         td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1207         td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1208         td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1209
1210         getrusage(RUSAGE_THREAD, &rusage);
1211         td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1212         td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1213         td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1214         td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1215
1216         free_data(thread_data, g->p.bytes_thread);
1217
1218         pthread_mutex_lock(&g->stop_work_mutex);
1219         g->bytes_done += bytes_done;
1220         pthread_mutex_unlock(&g->stop_work_mutex);
1221
1222         return NULL;
1223 }
1224
1225 /*
1226  * A worker process starts a couple of threads:
1227  */
1228 static void worker_process(int process_nr)
1229 {
1230         pthread_mutex_t process_lock;
1231         struct thread_data *td;
1232         pthread_t *pthreads;
1233         u8 *process_data;
1234         int task_nr;
1235         int ret;
1236         int t;
1237
1238         pthread_mutex_init(&process_lock, NULL);
1239         set_taskname("process %d", process_nr);
1240
1241         /*
1242          * Pick up the memory policy and the CPU binding of our first thread,
1243          * so that we initialize memory accordingly:
1244          */
1245         task_nr = process_nr*g->p.nr_threads;
1246         td = g->threads + task_nr;
1247
1248         bind_to_memnode(td->bind_node);
1249         bind_to_cpumask(td->bind_cpumask);
1250
1251         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1252         process_data = setup_private_data(g->p.bytes_process);
1253
1254         if (g->p.show_details >= 3) {
1255                 printf(" # process %2d global mem: %p, process mem: %p\n",
1256                         process_nr, g->data, process_data);
1257         }
1258
1259         for (t = 0; t < g->p.nr_threads; t++) {
1260                 task_nr = process_nr*g->p.nr_threads + t;
1261                 td = g->threads + task_nr;
1262
1263                 td->process_data = process_data;
1264                 td->process_nr   = process_nr;
1265                 td->thread_nr    = t;
1266                 td->task_nr      = task_nr;
1267                 td->val          = rand();
1268                 td->curr_cpu     = -1;
1269                 td->process_lock = &process_lock;
1270
1271                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1272                 BUG_ON(ret);
1273         }
1274
1275         for (t = 0; t < g->p.nr_threads; t++) {
1276                 ret = pthread_join(pthreads[t], NULL);
1277                 BUG_ON(ret);
1278         }
1279
1280         free_data(process_data, g->p.bytes_process);
1281         free(pthreads);
1282 }
1283
1284 static void print_summary(void)
1285 {
1286         if (g->p.show_details < 0)
1287                 return;
1288
1289         printf("\n ###\n");
1290         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1291                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1292         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1293                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1294         printf(" #      %5dx %5ldMB process shared mem operations\n",
1295                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1296         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1297                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1298
1299         printf(" ###\n");
1300
1301         printf("\n ###\n"); fflush(stdout);
1302 }
1303
1304 static void init_thread_data(void)
1305 {
1306         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1307         int t;
1308
1309         g->threads = zalloc_shared_data(size);
1310
1311         for (t = 0; t < g->p.nr_tasks; t++) {
1312                 struct thread_data *td = g->threads + t;
1313                 int cpu;
1314
1315                 /* Allow all nodes by default: */
1316                 td->bind_node = -1;
1317
1318                 /* Allow all CPUs by default: */
1319                 CPU_ZERO(&td->bind_cpumask);
1320                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1321                         CPU_SET(cpu, &td->bind_cpumask);
1322         }
1323 }
1324
1325 static void deinit_thread_data(void)
1326 {
1327         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1328
1329         free_data(g->threads, size);
1330 }
1331
1332 static int init(void)
1333 {
1334         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1335
1336         /* Copy over options: */
1337         g->p = p0;
1338
1339         g->p.nr_cpus = numa_num_configured_cpus();
1340
1341         g->p.nr_nodes = numa_max_node() + 1;
1342
1343         /* char array in count_process_nodes(): */
1344         BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1345
1346         if (g->p.show_quiet && !g->p.show_details)
1347                 g->p.show_details = -1;
1348
1349         /* Some memory should be specified: */
1350         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1351                 return -1;
1352
1353         if (g->p.mb_global_str) {
1354                 g->p.mb_global = atof(g->p.mb_global_str);
1355                 BUG_ON(g->p.mb_global < 0);
1356         }
1357
1358         if (g->p.mb_proc_str) {
1359                 g->p.mb_proc = atof(g->p.mb_proc_str);
1360                 BUG_ON(g->p.mb_proc < 0);
1361         }
1362
1363         if (g->p.mb_proc_locked_str) {
1364                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1365                 BUG_ON(g->p.mb_proc_locked < 0);
1366                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1367         }
1368
1369         if (g->p.mb_thread_str) {
1370                 g->p.mb_thread = atof(g->p.mb_thread_str);
1371                 BUG_ON(g->p.mb_thread < 0);
1372         }
1373
1374         BUG_ON(g->p.nr_threads <= 0);
1375         BUG_ON(g->p.nr_proc <= 0);
1376
1377         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1378
1379         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1380         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1381         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1382         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1383
1384         g->data = setup_shared_data(g->p.bytes_global);
1385
1386         /* Startup serialization: */
1387         init_global_mutex(&g->start_work_mutex);
1388         init_global_mutex(&g->startup_mutex);
1389         init_global_mutex(&g->startup_done_mutex);
1390         init_global_mutex(&g->stop_work_mutex);
1391
1392         init_thread_data();
1393
1394         tprintf("#\n");
1395         if (parse_setup_cpu_list() || parse_setup_node_list())
1396                 return -1;
1397         tprintf("#\n");
1398
1399         print_summary();
1400
1401         return 0;
1402 }
1403
1404 static void deinit(void)
1405 {
1406         free_data(g->data, g->p.bytes_global);
1407         g->data = NULL;
1408
1409         deinit_thread_data();
1410
1411         free_data(g, sizeof(*g));
1412         g = NULL;
1413 }
1414
1415 /*
1416  * Print a short or long result, depending on the verbosity setting:
1417  */
1418 static void print_res(const char *name, double val,
1419                       const char *txt_unit, const char *txt_short, const char *txt_long)
1420 {
1421         if (!name)
1422                 name = "main,";
1423
1424         if (!g->p.show_quiet)
1425                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1426         else
1427                 printf(" %14.3f %s\n", val, txt_long);
1428 }
1429
1430 static int __bench_numa(const char *name)
1431 {
1432         struct timeval start, stop, diff;
1433         u64 runtime_ns_min, runtime_ns_sum;
1434         pid_t *pids, pid, wpid;
1435         double delta_runtime;
1436         double runtime_avg;
1437         double runtime_sec_max;
1438         double runtime_sec_min;
1439         int wait_stat;
1440         double bytes;
1441         int i, t, p;
1442
1443         if (init())
1444                 return -1;
1445
1446         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1447         pid = -1;
1448
1449         /* All threads try to acquire it, this way we can wait for them to start up: */
1450         pthread_mutex_lock(&g->start_work_mutex);
1451
1452         if (g->p.serialize_startup) {
1453                 tprintf(" #\n");
1454                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1455         }
1456
1457         gettimeofday(&start, NULL);
1458
1459         for (i = 0; i < g->p.nr_proc; i++) {
1460                 pid = fork();
1461                 dprintf(" # process %2d: PID %d\n", i, pid);
1462
1463                 BUG_ON(pid < 0);
1464                 if (!pid) {
1465                         /* Child process: */
1466                         worker_process(i);
1467
1468                         exit(0);
1469                 }
1470                 pids[i] = pid;
1471
1472         }
1473         /* Wait for all the threads to start up: */
1474         while (g->nr_tasks_started != g->p.nr_tasks)
1475                 usleep(USEC_PER_MSEC);
1476
1477         BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1478
1479         if (g->p.serialize_startup) {
1480                 double startup_sec;
1481
1482                 pthread_mutex_lock(&g->startup_done_mutex);
1483
1484                 /* This will start all threads: */
1485                 pthread_mutex_unlock(&g->start_work_mutex);
1486
1487                 /* This mutex is locked - the last started thread will wake us: */
1488                 pthread_mutex_lock(&g->startup_done_mutex);
1489
1490                 gettimeofday(&stop, NULL);
1491
1492                 timersub(&stop, &start, &diff);
1493
1494                 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1495                 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1496                 startup_sec /= NSEC_PER_SEC;
1497
1498                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1499                 tprintf(" #\n");
1500
1501                 start = stop;
1502                 pthread_mutex_unlock(&g->startup_done_mutex);
1503         } else {
1504                 gettimeofday(&start, NULL);
1505         }
1506
1507         /* Parent process: */
1508
1509
1510         for (i = 0; i < g->p.nr_proc; i++) {
1511                 wpid = waitpid(pids[i], &wait_stat, 0);
1512                 BUG_ON(wpid < 0);
1513                 BUG_ON(!WIFEXITED(wait_stat));
1514
1515         }
1516
1517         runtime_ns_sum = 0;
1518         runtime_ns_min = -1LL;
1519
1520         for (t = 0; t < g->p.nr_tasks; t++) {
1521                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1522
1523                 runtime_ns_sum += thread_runtime_ns;
1524                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1525         }
1526
1527         gettimeofday(&stop, NULL);
1528         timersub(&stop, &start, &diff);
1529
1530         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1531
1532         tprintf("\n ###\n");
1533         tprintf("\n");
1534
1535         runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1536         runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1537         runtime_sec_max /= NSEC_PER_SEC;
1538
1539         runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1540
1541         bytes = g->bytes_done;
1542         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1543
1544         if (g->p.measure_convergence) {
1545                 print_res(name, runtime_sec_max,
1546                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1547         }
1548
1549         print_res(name, runtime_sec_max,
1550                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1551
1552         print_res(name, runtime_sec_min,
1553                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1554
1555         print_res(name, runtime_avg,
1556                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1557
1558         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1559         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1560                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1561
1562         print_res(name, bytes / g->p.nr_tasks / 1e9,
1563                 "GB,", "data/thread",           "GB data processed, per thread");
1564
1565         print_res(name, bytes / 1e9,
1566                 "GB,", "data-total",            "GB data processed, total");
1567
1568         print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1569                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1570
1571         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1572                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1573
1574         print_res(name, bytes / runtime_sec_max / 1e9,
1575                 "GB/sec,", "total-speed",       "GB/sec total speed");
1576
1577         if (g->p.show_details >= 2) {
1578                 char tname[14 + 2 * 10 + 1];
1579                 struct thread_data *td;
1580                 for (p = 0; p < g->p.nr_proc; p++) {
1581                         for (t = 0; t < g->p.nr_threads; t++) {
1582                                 memset(tname, 0, sizeof(tname));
1583                                 td = g->threads + p*g->p.nr_threads + t;
1584                                 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1585                                 print_res(tname, td->speed_gbs,
1586                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1587                                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1588                                         "secs", "thread-system-time", "system CPU time/thread");
1589                                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1590                                         "secs", "thread-user-time", "user CPU time/thread");
1591                         }
1592                 }
1593         }
1594
1595         free(pids);
1596
1597         deinit();
1598
1599         return 0;
1600 }
1601
1602 #define MAX_ARGS 50
1603
1604 static int command_size(const char **argv)
1605 {
1606         int size = 0;
1607
1608         while (*argv) {
1609                 size++;
1610                 argv++;
1611         }
1612
1613         BUG_ON(size >= MAX_ARGS);
1614
1615         return size;
1616 }
1617
1618 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1619 {
1620         int i;
1621
1622         printf("\n # Running %s \"perf bench numa", name);
1623
1624         for (i = 0; i < argc; i++)
1625                 printf(" %s", argv[i]);
1626
1627         printf("\"\n");
1628
1629         memset(p, 0, sizeof(*p));
1630
1631         /* Initialize nonzero defaults: */
1632
1633         p->serialize_startup            = 1;
1634         p->data_reads                   = true;
1635         p->data_writes                  = true;
1636         p->data_backwards               = true;
1637         p->data_rand_walk               = true;
1638         p->nr_loops                     = -1;
1639         p->init_random                  = true;
1640         p->mb_global_str                = "1";
1641         p->nr_proc                      = 1;
1642         p->nr_threads                   = 1;
1643         p->nr_secs                      = 5;
1644         p->run_all                      = argc == 1;
1645 }
1646
1647 static int run_bench_numa(const char *name, const char **argv)
1648 {
1649         int argc = command_size(argv);
1650
1651         init_params(&p0, name, argc, argv);
1652         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1653         if (argc)
1654                 goto err;
1655
1656         if (__bench_numa(name))
1657                 goto err;
1658
1659         return 0;
1660
1661 err:
1662         return -1;
1663 }
1664
1665 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1666 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1667
1668 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1669 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1670
1671 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1672 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1673
1674 /*
1675  * The built-in test-suite executed by "perf bench numa -a".
1676  *
1677  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1678  */
1679 static const char *tests[][MAX_ARGS] = {
1680    /* Basic single-stream NUMA bandwidth measurements: */
1681    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1682                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1683    { "RAM-bw-local-NOTHP,",
1684                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1685                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1686    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1687                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1688
1689    /* 2-stream NUMA bandwidth measurements: */
1690    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1691                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1692    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1693                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1694
1695    /* Cross-stream NUMA bandwidth measurement: */
1696    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1697                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1698
1699    /* Convergence latency measurements: */
1700    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1701    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1702    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1703    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1704    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1705    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1706    { " 4x4-convergence-NOTHP,",
1707                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1708    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1709    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1710    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1711    { " 8x4-convergence-NOTHP,",
1712                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1713    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1714    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1715    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1716    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1717    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1718
1719    /* Various NUMA process/thread layout bandwidth measurements: */
1720    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1721    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1722    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1723    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1724    { " 8x1-bw-process-NOTHP,",
1725                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1726    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1727
1728    { " 4x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1729    { " 8x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1730    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1731    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1732
1733    { " 2x3-bw-thread,",   "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1734    { " 4x4-bw-thread,",   "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1735    { " 4x6-bw-thread,",   "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1736    { " 4x8-bw-thread,",   "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1737    { " 4x8-bw-thread-NOTHP,",
1738                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1739    { " 3x3-bw-thread,",   "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1740    { " 5x5-bw-thread,",   "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1741
1742    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1743    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1744
1745    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1746    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1747    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1748    { "numa01-bw-thread-NOTHP,",
1749                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1750 };
1751
1752 static int bench_all(void)
1753 {
1754         int nr = ARRAY_SIZE(tests);
1755         int ret;
1756         int i;
1757
1758         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1759         BUG_ON(ret < 0);
1760
1761         for (i = 0; i < nr; i++) {
1762                 run_bench_numa(tests[i][0], tests[i] + 1);
1763         }
1764
1765         printf("\n");
1766
1767         return 0;
1768 }
1769
1770 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1771 {
1772         init_params(&p0, "main,", argc, argv);
1773         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1774         if (argc)
1775                 goto err;
1776
1777         if (p0.run_all)
1778                 return bench_all();
1779
1780         if (__bench_numa(NULL))
1781                 goto err;
1782
1783         return 0;
1784
1785 err:
1786         usage_with_options(numa_usage, options);
1787         return -1;
1788 }