tools/perf/bench/numa.c

   1 /*
   2  * numa.c
   3  *
   4  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5  */
   6
   7 #include <inttypes.h>
   8 /* For the CLR_() macros */
   9 #include <pthread.h>
  10
  11 #include "../perf.h"
  12 #include "../builtin.h"
  13 #include "../util/util.h"
  14 #include <subcmd/parse-options.h>
  15 #include "../util/cloexec.h"
  16
  17 #include "bench.h"
  18
  19 #include <errno.h>
  20 #include <sched.h>
  21 #include <stdio.h>
  22 #include <assert.h>
  23 #include <malloc.h>
  24 #include <signal.h>
  25 #include <stdlib.h>
  26 #include <string.h>
  27 #include <unistd.h>
  28 #include <sys/mman.h>
  29 #include <sys/time.h>
  30 #include <sys/resource.h>
  31 #include <sys/wait.h>
  32 #include <sys/prctl.h>
  33 #include <sys/types.h>
  34 #include <linux/kernel.h>
  35 #include <linux/time64.h>
  36
  37 #include <numa.h>
  38 #include <numaif.h>
  39
  40 /*
  41  * Regular printout to the terminal, supressed if -q is specified:
  42  */
  43 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  44
  45 /*
  46  * Debug printf:
  47  */
  48 #undef dprintf
  49 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  50
  51 struct thread_data {
  52         int                     curr_cpu;
  53         cpu_set_t               bind_cpumask;
  54         int                     bind_node;
  55         u8                      *process_data;
  56         int                     process_nr;
  57         int                     thread_nr;
  58         int                     task_nr;
  59         unsigned int            loops_done;
  60         u64                     val;
  61         u64                     runtime_ns;
  62         u64                     system_time_ns;
  63         u64                     user_time_ns;
  64         double                  speed_gbs;
  65         pthread_mutex_t         *process_lock;
  66 };
  67
  68 /* Parameters set by options: */
  69
  70 struct params {
  71         /* Startup synchronization: */
  72         bool                    serialize_startup;
  73
  74         /* Task hierarchy: */
  75         int                     nr_proc;
  76         int                     nr_threads;
  77
  78         /* Working set sizes: */
  79         const char              *mb_global_str;
  80         const char              *mb_proc_str;
  81         const char              *mb_proc_locked_str;
  82         const char              *mb_thread_str;
  83
  84         double                  mb_global;
  85         double                  mb_proc;
  86         double                  mb_proc_locked;
  87         double                  mb_thread;
  88
  89         /* Access patterns to the working set: */
  90         bool                    data_reads;
  91         bool                    data_writes;
  92         bool                    data_backwards;
  93         bool                    data_zero_memset;
  94         bool                    data_rand_walk;
  95         u32                     nr_loops;
  96         u32                     nr_secs;
  97         u32                     sleep_usecs;
  98
  99         /* Working set initialization: */
 100         bool                    init_zero;
 101         bool                    init_random;
 102         bool                    init_cpu0;
 103
 104         /* Misc options: */
 105         int                     show_details;
 106         int                     run_all;
 107         int                     thp;
 108
 109         long                    bytes_global;
 110         long                    bytes_process;
 111         long                    bytes_process_locked;
 112         long                    bytes_thread;
 113
 114         int                     nr_tasks;
 115         bool                    show_quiet;
 116
 117         bool                    show_convergence;
 118         bool                    measure_convergence;
 119
 120         int                     perturb_secs;
 121         int                     nr_cpus;
 122         int                     nr_nodes;
 123
 124         /* Affinity options -C and -N: */
 125         char                    *cpu_list_str;
 126         char                    *node_list_str;
 127 };
 128
 129
 130 /* Global, read-writable area, accessible to all processes and threads: */
 131
 132 struct global_info {
 133         u8                      *data;
 134
 135         pthread_mutex_t         startup_mutex;
 136         int                     nr_tasks_started;
 137
 138         pthread_mutex_t         startup_done_mutex;
 139
 140         pthread_mutex_t         start_work_mutex;
 141         int                     nr_tasks_working;
 142
 143         pthread_mutex_t         stop_work_mutex;
 144         u64                     bytes_done;
 145
 146         struct thread_data      *threads;
 147
 148         /* Convergence latency measurement: */
 149         bool                    all_converged;
 150         bool                    stop_work;
 151
 152         int                     print_once;
 153
 154         struct params           p;
 155 };
 156
 157 static struct global_info       *g = NULL;
 158
 159 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 160 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 161
 162 struct params p0;
 163
 164 static const struct option options[] = {
 165         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 166         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 167
 168         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 169         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 170         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 171         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 172
 173         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 174         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 175         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 176
 177         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via writes (can be mixed with -W)"),
 178         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 179         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 180         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 181         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 182
 183
 184         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 185         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 186         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 187         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 188
 189         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 190         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 191         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 192         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
 193                     "convergence is reached when each process (all its threads) is running on a single NUMA node."),
 194         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 195         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 196         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 197
 198         /* Special option string parsing callbacks: */
 199         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 200                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 201                         parse_cpus_opt),
 202         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 203                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 204                         parse_nodes_opt),
 205         OPT_END()
 206 };
 207
 208 static const char * const bench_numa_usage[] = {
 209         "perf bench numa <options>",
 210         NULL
 211 };
 212
 213 static const char * const numa_usage[] = {
 214         "perf bench numa mem [<options>]",
 215         NULL
 216 };
 217
 218 static cpu_set_t bind_to_cpu(int target_cpu)
 219 {
 220         cpu_set_t orig_mask, mask;
 221         int ret;
 222
 223         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 224         BUG_ON(ret);
 225
 226         CPU_ZERO(&mask);
 227
 228         if (target_cpu == -1) {
 229                 int cpu;
 230
 231                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 232                         CPU_SET(cpu, &mask);
 233         } else {
 234                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 235                 CPU_SET(target_cpu, &mask);
 236         }
 237
 238         ret = sched_setaffinity(0, sizeof(mask), &mask);
 239         BUG_ON(ret);
 240
 241         return orig_mask;
 242 }
 243
 244 static cpu_set_t bind_to_node(int target_node)
 245 {
 246         int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 247         cpu_set_t orig_mask, mask;
 248         int cpu;
 249         int ret;
 250
 251         BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 252         BUG_ON(!cpus_per_node);
 253
 254         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 255         BUG_ON(ret);
 256
 257         CPU_ZERO(&mask);
 258
 259         if (target_node == -1) {
 260                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 261                         CPU_SET(cpu, &mask);
 262         } else {
 263                 int cpu_start = (target_node + 0) * cpus_per_node;
 264                 int cpu_stop  = (target_node + 1) * cpus_per_node;
 265
 266                 BUG_ON(cpu_stop > g->p.nr_cpus);
 267
 268                 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 269                         CPU_SET(cpu, &mask);
 270         }
 271
 272         ret = sched_setaffinity(0, sizeof(mask), &mask);
 273         BUG_ON(ret);
 274
 275         return orig_mask;
 276 }
 277
 278 static void bind_to_cpumask(cpu_set_t mask)
 279 {
 280         int ret;
 281
 282         ret = sched_setaffinity(0, sizeof(mask), &mask);
 283         BUG_ON(ret);
 284 }
 285
 286 static void mempol_restore(void)
 287 {
 288         int ret;
 289
 290         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 291
 292         BUG_ON(ret);
 293 }
 294
 295 static void bind_to_memnode(int node)
 296 {
 297         unsigned long nodemask;
 298         int ret;
 299
 300         if (node == -1)
 301                 return;
 302
 303         BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 304         nodemask = 1L << node;
 305
 306         ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 307         dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 308
 309         BUG_ON(ret);
 310 }
 311
 312 #define HPSIZE (2*1024*1024)
 313
 314 #define set_taskname(fmt...)                            \
 315 do {                                                    \
 316         char name[20];                                  \
 317                                                         \
 318         snprintf(name, 20, fmt);                        \
 319         prctl(PR_SET_NAME, name);                       \
 320 } while (0)
 321
 322 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 323                       int init_zero, int init_cpu0, int thp, int init_random)
 324 {
 325         cpu_set_t orig_mask;
 326         ssize_t bytes;
 327         u8 *buf;
 328         int ret;
 329
 330         if (!bytes0)
 331                 return NULL;
 332
 333         /* Allocate and initialize all memory on CPU#0: */
 334         if (init_cpu0) {
 335                 orig_mask = bind_to_node(0);
 336                 bind_to_memnode(0);
 337         }
 338
 339         bytes = bytes0 + HPSIZE;
 340
 341         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 342         BUG_ON(buf == (void *)-1);
 343
 344         if (map_flags == MAP_PRIVATE) {
 345                 if (thp > 0) {
 346                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 347                         if (ret && !g->print_once) {
 348                                 g->print_once = 1;
 349                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 350                         }
 351                 }
 352                 if (thp < 0) {
 353                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 354                         if (ret && !g->print_once) {
 355                                 g->print_once = 1;
 356                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 357                         }
 358                 }
 359         }
 360
 361         if (init_zero) {
 362                 bzero(buf, bytes);
 363         } else {
 364                 /* Initialize random contents, different in each word: */
 365                 if (init_random) {
 366                         u64 *wbuf = (void *)buf;
 367                         long off = rand();
 368                         long i;
 369
 370                         for (i = 0; i < bytes/8; i++)
 371                                 wbuf[i] = i + off;
 372                 }
 373         }
 374
 375         /* Align to 2MB boundary: */
 376         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 377
 378         /* Restore affinity: */
 379         if (init_cpu0) {
 380                 bind_to_cpumask(orig_mask);
 381                 mempol_restore();
 382         }
 383
 384         return buf;
 385 }
 386
 387 static void free_data(void *data, ssize_t bytes)
 388 {
 389         int ret;
 390
 391         if (!data)
 392                 return;
 393
 394         ret = munmap(data, bytes);
 395         BUG_ON(ret);
 396 }
 397
 398 /*
 399  * Create a shared memory buffer that can be shared between processes, zeroed:
 400  */
 401 static void * zalloc_shared_data(ssize_t bytes)
 402 {
 403         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 404 }
 405
 406 /*
 407  * Create a shared memory buffer that can be shared between processes:
 408  */
 409 static void * setup_shared_data(ssize_t bytes)
 410 {
 411         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 412 }
 413
 414 /*
 415  * Allocate process-local memory - this will either be shared between
 416  * threads of this process, or only be accessed by this thread:
 417  */
 418 static void * setup_private_data(ssize_t bytes)
 419 {
 420         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 421 }
 422
 423 /*
 424  * Return a process-shared (global) mutex:
 425  */
 426 static void init_global_mutex(pthread_mutex_t *mutex)
 427 {
 428         pthread_mutexattr_t attr;
 429
 430         pthread_mutexattr_init(&attr);
 431         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 432         pthread_mutex_init(mutex, &attr);
 433 }
 434
 435 static int parse_cpu_list(const char *arg)
 436 {
 437         p0.cpu_list_str = strdup(arg);
 438
 439         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 440
 441         return 0;
 442 }
 443
 444 static int parse_setup_cpu_list(void)
 445 {
 446         struct thread_data *td;
 447         char *str0, *str;
 448         int t;
 449
 450         if (!g->p.cpu_list_str)
 451                 return 0;
 452
 453         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 454
 455         str0 = str = strdup(g->p.cpu_list_str);
 456         t = 0;
 457
 458         BUG_ON(!str);
 459
 460         tprintf("# binding tasks to CPUs:\n");
 461         tprintf("#  ");
 462
 463         while (true) {
 464                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 465                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 466                 int bind_len;
 467                 int step;
 468                 int mul;
 469
 470                 tok = strsep(&str, ",");
 471                 if (!tok)
 472                         break;
 473
 474                 tok_end = strstr(tok, "-");
 475
 476                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 477                 if (!tok_end) {
 478                         /* Single CPU specified: */
 479                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 480                 } else {
 481                         /* CPU range specified (for example: "5-11"): */
 482                         bind_cpu_0 = atol(tok);
 483                         bind_cpu_1 = atol(tok_end + 1);
 484                 }
 485
 486                 step = 1;
 487                 tok_step = strstr(tok, "#");
 488                 if (tok_step) {
 489                         step = atol(tok_step + 1);
 490                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 491                 }
 492
 493                 /*
 494                  * Mask length.
 495                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 496                  * where the _4 means the next 4 CPUs are allowed.
 497                  */
 498                 bind_len = 1;
 499                 tok_len = strstr(tok, "_");
 500                 if (tok_len) {
 501                         bind_len = atol(tok_len + 1);
 502                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 503                 }
 504
 505                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 506                 mul = 1;
 507                 tok_mul = strstr(tok, "x");
 508                 if (tok_mul) {
 509                         mul = atol(tok_mul + 1);
 510                         BUG_ON(mul <= 0);
 511                 }
 512
 513                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 514
 515                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 516                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 517                         return -1;
 518                 }
 519
 520                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 521                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 522
 523                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 524                         int i;
 525
 526                         for (i = 0; i < mul; i++) {
 527                                 int cpu;
 528
 529                                 if (t >= g->p.nr_tasks) {
 530                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 531                                         goto out;
 532                                 }
 533                                 td = g->threads + t;
 534
 535                                 if (t)
 536                                         tprintf(",");
 537                                 if (bind_len > 1) {
 538                                         tprintf("%2d/%d", bind_cpu, bind_len);
 539                                 } else {
 540                                         tprintf("%2d", bind_cpu);
 541                                 }
 542
 543                                 CPU_ZERO(&td->bind_cpumask);
 544                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 545                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 546                                         CPU_SET(cpu, &td->bind_cpumask);
 547                                 }
 548                                 t++;
 549                         }
 550                 }
 551         }
 552 out:
 553
 554         tprintf("\n");
 555
 556         if (t < g->p.nr_tasks)
 557                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 558
 559         free(str0);
 560         return 0;
 561 }
 562
 563 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 564                           const char *arg, int unset __maybe_unused)
 565 {
 566         if (!arg)
 567                 return -1;
 568
 569         return parse_cpu_list(arg);
 570 }
 571
 572 static int parse_node_list(const char *arg)
 573 {
 574         p0.node_list_str = strdup(arg);
 575
 576         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 577
 578         return 0;
 579 }
 580
 581 static int parse_setup_node_list(void)
 582 {
 583         struct thread_data *td;
 584         char *str0, *str;
 585         int t;
 586
 587         if (!g->p.node_list_str)
 588                 return 0;
 589
 590         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 591
 592         str0 = str = strdup(g->p.node_list_str);
 593         t = 0;
 594
 595         BUG_ON(!str);
 596
 597         tprintf("# binding tasks to NODEs:\n");
 598         tprintf("# ");
 599
 600         while (true) {
 601                 int bind_node, bind_node_0, bind_node_1;
 602                 char *tok, *tok_end, *tok_step, *tok_mul;
 603                 int step;
 604                 int mul;
 605
 606                 tok = strsep(&str, ",");
 607                 if (!tok)
 608                         break;
 609
 610                 tok_end = strstr(tok, "-");
 611
 612                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 613                 if (!tok_end) {
 614                         /* Single NODE specified: */
 615                         bind_node_0 = bind_node_1 = atol(tok);
 616                 } else {
 617                         /* NODE range specified (for example: "5-11"): */
 618                         bind_node_0 = atol(tok);
 619                         bind_node_1 = atol(tok_end + 1);
 620                 }
 621
 622                 step = 1;
 623                 tok_step = strstr(tok, "#");
 624                 if (tok_step) {
 625                         step = atol(tok_step + 1);
 626                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 627                 }
 628
 629                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 630                 mul = 1;
 631                 tok_mul = strstr(tok, "x");
 632                 if (tok_mul) {
 633                         mul = atol(tok_mul + 1);
 634                         BUG_ON(mul <= 0);
 635                 }
 636
 637                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 638
 639                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 640                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 641                         return -1;
 642                 }
 643
 644                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 645                 BUG_ON(bind_node_0 > bind_node_1);
 646
 647                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 648                         int i;
 649
 650                         for (i = 0; i < mul; i++) {
 651                                 if (t >= g->p.nr_tasks) {
 652                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 653                                         goto out;
 654                                 }
 655                                 td = g->threads + t;
 656
 657                                 if (!t)
 658                                         tprintf(" %2d", bind_node);
 659                                 else
 660                                         tprintf(",%2d", bind_node);
 661
 662                                 td->bind_node = bind_node;
 663                                 t++;
 664                         }
 665                 }
 666         }
 667 out:
 668
 669         tprintf("\n");
 670
 671         if (t < g->p.nr_tasks)
 672                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 673
 674         free(str0);
 675         return 0;
 676 }
 677
 678 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 679                           const char *arg, int unset __maybe_unused)
 680 {
 681         if (!arg)
 682                 return -1;
 683
 684         return parse_node_list(arg);
 685
 686         return 0;
 687 }
 688
 689 #define BIT(x) (1ul << x)
 690
 691 static inline uint32_t lfsr_32(uint32_t lfsr)
 692 {
 693         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 694         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 695 }
 696
 697 /*
 698  * Make sure there's real data dependency to RAM (when read
 699  * accesses are enabled), so the compiler, the CPU and the
 700  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 701  * accesses:
 702  */
 703 static inline u64 access_data(u64 *data, u64 val)
 704 {
 705         if (g->p.data_reads)
 706                 val += *data;
 707         if (g->p.data_writes)
 708                 *data = val + 1;
 709         return val;
 710 }
 711
 712 /*
 713  * The worker process does two types of work, a forwards going
 714  * loop and a backwards going loop.
 715  *
 716  * We do this so that on multiprocessor systems we do not create
 717  * a 'train' of processing, with highly synchronized processes,
 718  * skewing the whole benchmark.
 719  */
 720 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 721 {
 722         long words = bytes/sizeof(u64);
 723         u64 *data = (void *)__data;
 724         long chunk_0, chunk_1;
 725         u64 *d0, *d, *d1;
 726         long off;
 727         long i;
 728
 729         BUG_ON(!data && words);
 730         BUG_ON(data && !words);
 731
 732         if (!data)
 733                 return val;
 734
 735         /* Very simple memset() work variant: */
 736         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 737                 bzero(data, bytes);
 738                 return val;
 739         }
 740
 741         /* Spread out by PID/TID nr and by loop nr: */
 742         chunk_0 = words/nr_max;
 743         chunk_1 = words/g->p.nr_loops;
 744         off = nr*chunk_0 + loop*chunk_1;
 745
 746         while (off >= words)
 747                 off -= words;
 748
 749         if (g->p.data_rand_walk) {
 750                 u32 lfsr = nr + loop + val;
 751                 int j;
 752
 753                 for (i = 0; i < words/1024; i++) {
 754                         long start, end;
 755
 756                         lfsr = lfsr_32(lfsr);
 757
 758                         start = lfsr % words;
 759                         end = min(start + 1024, words-1);
 760
 761                         if (g->p.data_zero_memset) {
 762                                 bzero(data + start, (end-start) * sizeof(u64));
 763                         } else {
 764                                 for (j = start; j < end; j++)
 765                                         val = access_data(data + j, val);
 766                         }
 767                 }
 768         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 769
 770                 d0 = data + off;
 771                 d  = data + off + 1;
 772                 d1 = data + words;
 773
 774                 /* Process data forwards: */
 775                 for (;;) {
 776                         if (unlikely(d >= d1))
 777                                 d = data;
 778                         if (unlikely(d == d0))
 779                                 break;
 780
 781                         val = access_data(d, val);
 782
 783                         d++;
 784                 }
 785         } else {
 786                 /* Process data backwards: */
 787
 788                 d0 = data + off;
 789                 d  = data + off - 1;
 790                 d1 = data + words;
 791
 792                 /* Process data forwards: */
 793                 for (;;) {
 794                         if (unlikely(d < data))
 795                                 d = data + words-1;
 796                         if (unlikely(d == d0))
 797                                 break;
 798
 799                         val = access_data(d, val);
 800
 801                         d--;
 802                 }
 803         }
 804
 805         return val;
 806 }
 807
 808 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 809 {
 810         unsigned int cpu;
 811
 812         cpu = sched_getcpu();
 813
 814         g->threads[task_nr].curr_cpu = cpu;
 815         prctl(0, bytes_worked);
 816 }
 817
 818 #define MAX_NR_NODES    64
 819
 820 /*
 821  * Count the number of nodes a process's threads
 822  * are spread out on.
 823  *
 824  * A count of 1 means that the process is compressed
 825  * to a single node. A count of g->p.nr_nodes means it's
 826  * spread out on the whole system.
 827  */
 828 static int count_process_nodes(int process_nr)
 829 {
 830         char node_present[MAX_NR_NODES] = { 0, };
 831         int nodes;
 832         int n, t;
 833
 834         for (t = 0; t < g->p.nr_threads; t++) {
 835                 struct thread_data *td;
 836                 int task_nr;
 837                 int node;
 838
 839                 task_nr = process_nr*g->p.nr_threads + t;
 840                 td = g->threads + task_nr;
 841
 842                 node = numa_node_of_cpu(td->curr_cpu);
 843                 if (node < 0) /* curr_cpu was likely still -1 */
 844                         return 0;
 845
 846                 node_present[node] = 1;
 847         }
 848
 849         nodes = 0;
 850
 851         for (n = 0; n < MAX_NR_NODES; n++)
 852                 nodes += node_present[n];
 853
 854         return nodes;
 855 }
 856
 857 /*
 858  * Count the number of distinct process-threads a node contains.
 859  *
 860  * A count of 1 means that the node contains only a single
 861  * process. If all nodes on the system contain at most one
 862  * process then we are well-converged.
 863  */
 864 static int count_node_processes(int node)
 865 {
 866         int processes = 0;
 867         int t, p;
 868
 869         for (p = 0; p < g->p.nr_proc; p++) {
 870                 for (t = 0; t < g->p.nr_threads; t++) {
 871                         struct thread_data *td;
 872                         int task_nr;
 873                         int n;
 874
 875                         task_nr = p*g->p.nr_threads + t;
 876                         td = g->threads + task_nr;
 877
 878                         n = numa_node_of_cpu(td->curr_cpu);
 879                         if (n == node) {
 880                                 processes++;
 881                                 break;
 882                         }
 883                 }
 884         }
 885
 886         return processes;
 887 }
 888
 889 static void calc_convergence_compression(int *strong)
 890 {
 891         unsigned int nodes_min, nodes_max;
 892         int p;
 893
 894         nodes_min = -1;
 895         nodes_max =  0;
 896
 897         for (p = 0; p < g->p.nr_proc; p++) {
 898                 unsigned int nodes = count_process_nodes(p);
 899
 900                 if (!nodes) {
 901                         *strong = 0;
 902                         return;
 903                 }
 904
 905                 nodes_min = min(nodes, nodes_min);
 906                 nodes_max = max(nodes, nodes_max);
 907         }
 908
 909         /* Strong convergence: all threads compress on a single node: */
 910         if (nodes_min == 1 && nodes_max == 1) {
 911                 *strong = 1;
 912         } else {
 913                 *strong = 0;
 914                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 915         }
 916 }
 917
 918 static void calc_convergence(double runtime_ns_max, double *convergence)
 919 {
 920         unsigned int loops_done_min, loops_done_max;
 921         int process_groups;
 922         int nodes[MAX_NR_NODES];
 923         int distance;
 924         int nr_min;
 925         int nr_max;
 926         int strong;
 927         int sum;
 928         int nr;
 929         int node;
 930         int cpu;
 931         int t;
 932
 933         if (!g->p.show_convergence && !g->p.measure_convergence)
 934                 return;
 935
 936         for (node = 0; node < g->p.nr_nodes; node++)
 937                 nodes[node] = 0;
 938
 939         loops_done_min = -1;
 940         loops_done_max = 0;
 941
 942         for (t = 0; t < g->p.nr_tasks; t++) {
 943                 struct thread_data *td = g->threads + t;
 944                 unsigned int loops_done;
 945
 946                 cpu = td->curr_cpu;
 947
 948                 /* Not all threads have written it yet: */
 949                 if (cpu < 0)
 950                         continue;
 951
 952                 node = numa_node_of_cpu(cpu);
 953
 954                 nodes[node]++;
 955
 956                 loops_done = td->loops_done;
 957                 loops_done_min = min(loops_done, loops_done_min);
 958                 loops_done_max = max(loops_done, loops_done_max);
 959         }
 960
 961         nr_max = 0;
 962         nr_min = g->p.nr_tasks;
 963         sum = 0;
 964
 965         for (node = 0; node < g->p.nr_nodes; node++) {
 966                 nr = nodes[node];
 967                 nr_min = min(nr, nr_min);
 968                 nr_max = max(nr, nr_max);
 969                 sum += nr;
 970         }
 971         BUG_ON(nr_min > nr_max);
 972
 973         BUG_ON(sum > g->p.nr_tasks);
 974
 975         if (0 && (sum < g->p.nr_tasks))
 976                 return;
 977
 978         /*
 979          * Count the number of distinct process groups present
 980          * on nodes - when we are converged this will decrease
 981          * to g->p.nr_proc:
 982          */
 983         process_groups = 0;
 984
 985         for (node = 0; node < g->p.nr_nodes; node++) {
 986                 int processes = count_node_processes(node);
 987
 988                 nr = nodes[node];
 989                 tprintf(" %2d/%-2d", nr, processes);
 990
 991                 process_groups += processes;
 992         }
 993
 994         distance = nr_max - nr_min;
 995
 996         tprintf(" [%2d/%-2d]", distance, process_groups);
 997
 998         tprintf(" l:%3d-%-3d (%3d)",
 999                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1000
1001         if (loops_done_min && loops_done_max) {
1002                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1003
1004                 tprintf(" [%4.1f%%]", skew * 100.0);
1005         }
1006
1007         calc_convergence_compression(&strong);
1008
1009         if (strong && process_groups == g->p.nr_proc) {
1010                 if (!*convergence) {
1011                         *convergence = runtime_ns_max;
1012                         tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1013                         if (g->p.measure_convergence) {
1014                                 g->all_converged = true;
1015                                 g->stop_work = true;
1016                         }
1017                 }
1018         } else {
1019                 if (*convergence) {
1020                         tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1021                         *convergence = 0;
1022                 }
1023                 tprintf("\n");
1024         }
1025 }
1026
1027 static void show_summary(double runtime_ns_max, int l, double *convergence)
1028 {
1029         tprintf("\r #  %5.1f%%  [%.1f mins]",
1030                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1031
1032         calc_convergence(runtime_ns_max, convergence);
1033
1034         if (g->p.show_details >= 0)
1035                 fflush(stdout);
1036 }
1037
1038 static void *worker_thread(void *__tdata)
1039 {
1040         struct thread_data *td = __tdata;
1041         struct timeval start0, start, stop, diff;
1042         int process_nr = td->process_nr;
1043         int thread_nr = td->thread_nr;
1044         unsigned long last_perturbance;
1045         int task_nr = td->task_nr;
1046         int details = g->p.show_details;
1047         int first_task, last_task;
1048         double convergence = 0;
1049         u64 val = td->val;
1050         double runtime_ns_max;
1051         u8 *global_data;
1052         u8 *process_data;
1053         u8 *thread_data;
1054         u64 bytes_done;
1055         long work_done;
1056         u32 l;
1057         struct rusage rusage;
1058
1059         bind_to_cpumask(td->bind_cpumask);
1060         bind_to_memnode(td->bind_node);
1061
1062         set_taskname("thread %d/%d", process_nr, thread_nr);
1063
1064         global_data = g->data;
1065         process_data = td->process_data;
1066         thread_data = setup_private_data(g->p.bytes_thread);
1067
1068         bytes_done = 0;
1069
1070         last_task = 0;
1071         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1072                 last_task = 1;
1073
1074         first_task = 0;
1075         if (process_nr == 0 && thread_nr == 0)
1076                 first_task = 1;
1077
1078         if (details >= 2) {
1079                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1080                         process_nr, thread_nr, global_data, process_data, thread_data);
1081         }
1082
1083         if (g->p.serialize_startup) {
1084                 pthread_mutex_lock(&g->startup_mutex);
1085                 g->nr_tasks_started++;
1086                 pthread_mutex_unlock(&g->startup_mutex);
1087
1088                 /* Here we will wait for the main process to start us all at once: */
1089                 pthread_mutex_lock(&g->start_work_mutex);
1090                 g->nr_tasks_working++;
1091
1092                 /* Last one wake the main process: */
1093                 if (g->nr_tasks_working == g->p.nr_tasks)
1094                         pthread_mutex_unlock(&g->startup_done_mutex);
1095
1096                 pthread_mutex_unlock(&g->start_work_mutex);
1097         }
1098
1099         gettimeofday(&start0, NULL);
1100
1101         start = stop = start0;
1102         last_perturbance = start.tv_sec;
1103
1104         for (l = 0; l < g->p.nr_loops; l++) {
1105                 start = stop;
1106
1107                 if (g->stop_work)
1108                         break;
1109
1110                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1111                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1112                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1113
1114                 if (g->p.sleep_usecs) {
1115                         pthread_mutex_lock(td->process_lock);
1116                         usleep(g->p.sleep_usecs);
1117                         pthread_mutex_unlock(td->process_lock);
1118                 }
1119                 /*
1120                  * Amount of work to be done under a process-global lock:
1121                  */
1122                 if (g->p.bytes_process_locked) {
1123                         pthread_mutex_lock(td->process_lock);
1124                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1125                         pthread_mutex_unlock(td->process_lock);
1126                 }
1127
1128                 work_done = g->p.bytes_global + g->p.bytes_process +
1129                             g->p.bytes_process_locked + g->p.bytes_thread;
1130
1131                 update_curr_cpu(task_nr, work_done);
1132                 bytes_done += work_done;
1133
1134                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1135                         continue;
1136
1137                 td->loops_done = l;
1138
1139                 gettimeofday(&stop, NULL);
1140
1141                 /* Check whether our max runtime timed out: */
1142                 if (g->p.nr_secs) {
1143                         timersub(&stop, &start0, &diff);
1144                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1145                                 g->stop_work = true;
1146                                 break;
1147                         }
1148                 }
1149
1150                 /* Update the summary at most once per second: */
1151                 if (start.tv_sec == stop.tv_sec)
1152                         continue;
1153
1154                 /*
1155                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1156                  * by migrating to CPU#0:
1157                  */
1158                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1159                         cpu_set_t orig_mask;
1160                         int target_cpu;
1161                         int this_cpu;
1162
1163                         last_perturbance = stop.tv_sec;
1164
1165                         /*
1166                          * Depending on where we are running, move into
1167                          * the other half of the system, to create some
1168                          * real disturbance:
1169                          */
1170                         this_cpu = g->threads[task_nr].curr_cpu;
1171                         if (this_cpu < g->p.nr_cpus/2)
1172                                 target_cpu = g->p.nr_cpus-1;
1173                         else
1174                                 target_cpu = 0;
1175
1176                         orig_mask = bind_to_cpu(target_cpu);
1177
1178                         /* Here we are running on the target CPU already */
1179                         if (details >= 1)
1180                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1181
1182                         bind_to_cpumask(orig_mask);
1183                 }
1184
1185                 if (details >= 3) {
1186                         timersub(&stop, &start, &diff);
1187                         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1188                         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1189
1190                         if (details >= 0) {
1191                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1192                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1193                         }
1194                         fflush(stdout);
1195                 }
1196                 if (!last_task)
1197                         continue;
1198
1199                 timersub(&stop, &start0, &diff);
1200                 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1201                 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1202
1203                 show_summary(runtime_ns_max, l, &convergence);
1204         }
1205
1206         gettimeofday(&stop, NULL);
1207         timersub(&stop, &start0, &diff);
1208         td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1209         td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1210         td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1211
1212         getrusage(RUSAGE_THREAD, &rusage);
1213         td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1214         td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1215         td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1216         td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1217
1218         free_data(thread_data, g->p.bytes_thread);
1219
1220         pthread_mutex_lock(&g->stop_work_mutex);
1221         g->bytes_done += bytes_done;
1222         pthread_mutex_unlock(&g->stop_work_mutex);
1223
1224         return NULL;
1225 }
1226
1227 /*
1228  * A worker process starts a couple of threads:
1229  */
1230 static void worker_process(int process_nr)
1231 {
1232         pthread_mutex_t process_lock;
1233         struct thread_data *td;
1234         pthread_t *pthreads;
1235         u8 *process_data;
1236         int task_nr;
1237         int ret;
1238         int t;
1239
1240         pthread_mutex_init(&process_lock, NULL);
1241         set_taskname("process %d", process_nr);
1242
1243         /*
1244          * Pick up the memory policy and the CPU binding of our first thread,
1245          * so that we initialize memory accordingly:
1246          */
1247         task_nr = process_nr*g->p.nr_threads;
1248         td = g->threads + task_nr;
1249
1250         bind_to_memnode(td->bind_node);
1251         bind_to_cpumask(td->bind_cpumask);
1252
1253         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1254         process_data = setup_private_data(g->p.bytes_process);
1255
1256         if (g->p.show_details >= 3) {
1257                 printf(" # process %2d global mem: %p, process mem: %p\n",
1258                         process_nr, g->data, process_data);
1259         }
1260
1261         for (t = 0; t < g->p.nr_threads; t++) {
1262                 task_nr = process_nr*g->p.nr_threads + t;
1263                 td = g->threads + task_nr;
1264
1265                 td->process_data = process_data;
1266                 td->process_nr   = process_nr;
1267                 td->thread_nr    = t;
1268                 td->task_nr      = task_nr;
1269                 td->val          = rand();
1270                 td->curr_cpu     = -1;
1271                 td->process_lock = &process_lock;
1272
1273                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1274                 BUG_ON(ret);
1275         }
1276
1277         for (t = 0; t < g->p.nr_threads; t++) {
1278                 ret = pthread_join(pthreads[t], NULL);
1279                 BUG_ON(ret);
1280         }
1281
1282         free_data(process_data, g->p.bytes_process);
1283         free(pthreads);
1284 }
1285
1286 static void print_summary(void)
1287 {
1288         if (g->p.show_details < 0)
1289                 return;
1290
1291         printf("\n ###\n");
1292         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1293                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1294         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1295                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1296         printf(" #      %5dx %5ldMB process shared mem operations\n",
1297                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1298         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1299                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1300
1301         printf(" ###\n");
1302
1303         printf("\n ###\n"); fflush(stdout);
1304 }
1305
1306 static void init_thread_data(void)
1307 {
1308         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1309         int t;
1310
1311         g->threads = zalloc_shared_data(size);
1312
1313         for (t = 0; t < g->p.nr_tasks; t++) {
1314                 struct thread_data *td = g->threads + t;
1315                 int cpu;
1316
1317                 /* Allow all nodes by default: */
1318                 td->bind_node = -1;
1319
1320                 /* Allow all CPUs by default: */
1321                 CPU_ZERO(&td->bind_cpumask);
1322                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1323                         CPU_SET(cpu, &td->bind_cpumask);
1324         }
1325 }
1326
1327 static void deinit_thread_data(void)
1328 {
1329         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1330
1331         free_data(g->threads, size);
1332 }
1333
1334 static int init(void)
1335 {
1336         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1337
1338         /* Copy over options: */
1339         g->p = p0;
1340
1341         g->p.nr_cpus = numa_num_configured_cpus();
1342
1343         g->p.nr_nodes = numa_max_node() + 1;
1344
1345         /* char array in count_process_nodes(): */
1346         BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1347
1348         if (g->p.show_quiet && !g->p.show_details)
1349                 g->p.show_details = -1;
1350
1351         /* Some memory should be specified: */
1352         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1353                 return -1;
1354
1355         if (g->p.mb_global_str) {
1356                 g->p.mb_global = atof(g->p.mb_global_str);
1357                 BUG_ON(g->p.mb_global < 0);
1358         }
1359
1360         if (g->p.mb_proc_str) {
1361                 g->p.mb_proc = atof(g->p.mb_proc_str);
1362                 BUG_ON(g->p.mb_proc < 0);
1363         }
1364
1365         if (g->p.mb_proc_locked_str) {
1366                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1367                 BUG_ON(g->p.mb_proc_locked < 0);
1368                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1369         }
1370
1371         if (g->p.mb_thread_str) {
1372                 g->p.mb_thread = atof(g->p.mb_thread_str);
1373                 BUG_ON(g->p.mb_thread < 0);
1374         }
1375
1376         BUG_ON(g->p.nr_threads <= 0);
1377         BUG_ON(g->p.nr_proc <= 0);
1378
1379         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1380
1381         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1382         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1383         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1384         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1385
1386         g->data = setup_shared_data(g->p.bytes_global);
1387
1388         /* Startup serialization: */
1389         init_global_mutex(&g->start_work_mutex);
1390         init_global_mutex(&g->startup_mutex);
1391         init_global_mutex(&g->startup_done_mutex);
1392         init_global_mutex(&g->stop_work_mutex);
1393
1394         init_thread_data();
1395
1396         tprintf("#\n");
1397         if (parse_setup_cpu_list() || parse_setup_node_list())
1398                 return -1;
1399         tprintf("#\n");
1400
1401         print_summary();
1402
1403         return 0;
1404 }
1405
1406 static void deinit(void)
1407 {
1408         free_data(g->data, g->p.bytes_global);
1409         g->data = NULL;
1410
1411         deinit_thread_data();
1412
1413         free_data(g, sizeof(*g));
1414         g = NULL;
1415 }
1416
1417 /*
1418  * Print a short or long result, depending on the verbosity setting:
1419  */
1420 static void print_res(const char *name, double val,
1421                       const char *txt_unit, const char *txt_short, const char *txt_long)
1422 {
1423         if (!name)
1424                 name = "main,";
1425
1426         if (!g->p.show_quiet)
1427                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1428         else
1429                 printf(" %14.3f %s\n", val, txt_long);
1430 }
1431
1432 static int __bench_numa(const char *name)
1433 {
1434         struct timeval start, stop, diff;
1435         u64 runtime_ns_min, runtime_ns_sum;
1436         pid_t *pids, pid, wpid;
1437         double delta_runtime;
1438         double runtime_avg;
1439         double runtime_sec_max;
1440         double runtime_sec_min;
1441         int wait_stat;
1442         double bytes;
1443         int i, t, p;
1444
1445         if (init())
1446                 return -1;
1447
1448         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1449         pid = -1;
1450
1451         /* All threads try to acquire it, this way we can wait for them to start up: */
1452         pthread_mutex_lock(&g->start_work_mutex);
1453
1454         if (g->p.serialize_startup) {
1455                 tprintf(" #\n");
1456                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1457         }
1458
1459         gettimeofday(&start, NULL);
1460
1461         for (i = 0; i < g->p.nr_proc; i++) {
1462                 pid = fork();
1463                 dprintf(" # process %2d: PID %d\n", i, pid);
1464
1465                 BUG_ON(pid < 0);
1466                 if (!pid) {
1467                         /* Child process: */
1468                         worker_process(i);
1469
1470                         exit(0);
1471                 }
1472                 pids[i] = pid;
1473
1474         }
1475         /* Wait for all the threads to start up: */
1476         while (g->nr_tasks_started != g->p.nr_tasks)
1477                 usleep(USEC_PER_MSEC);
1478
1479         BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1480
1481         if (g->p.serialize_startup) {
1482                 double startup_sec;
1483
1484                 pthread_mutex_lock(&g->startup_done_mutex);
1485
1486                 /* This will start all threads: */
1487                 pthread_mutex_unlock(&g->start_work_mutex);
1488
1489                 /* This mutex is locked - the last started thread will wake us: */
1490                 pthread_mutex_lock(&g->startup_done_mutex);
1491
1492                 gettimeofday(&stop, NULL);
1493
1494                 timersub(&stop, &start, &diff);
1495
1496                 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1497                 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1498                 startup_sec /= NSEC_PER_SEC;
1499
1500                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1501                 tprintf(" #\n");
1502
1503                 start = stop;
1504                 pthread_mutex_unlock(&g->startup_done_mutex);
1505         } else {
1506                 gettimeofday(&start, NULL);
1507         }
1508
1509         /* Parent process: */
1510
1511
1512         for (i = 0; i < g->p.nr_proc; i++) {
1513                 wpid = waitpid(pids[i], &wait_stat, 0);
1514                 BUG_ON(wpid < 0);
1515                 BUG_ON(!WIFEXITED(wait_stat));
1516
1517         }
1518
1519         runtime_ns_sum = 0;
1520         runtime_ns_min = -1LL;
1521
1522         for (t = 0; t < g->p.nr_tasks; t++) {
1523                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1524
1525                 runtime_ns_sum += thread_runtime_ns;
1526                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1527         }
1528
1529         gettimeofday(&stop, NULL);
1530         timersub(&stop, &start, &diff);
1531
1532         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1533
1534         tprintf("\n ###\n");
1535         tprintf("\n");
1536
1537         runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1538         runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1539         runtime_sec_max /= NSEC_PER_SEC;
1540
1541         runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1542
1543         bytes = g->bytes_done;
1544         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1545
1546         if (g->p.measure_convergence) {
1547                 print_res(name, runtime_sec_max,
1548                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1549         }
1550
1551         print_res(name, runtime_sec_max,
1552                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1553
1554         print_res(name, runtime_sec_min,
1555                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1556
1557         print_res(name, runtime_avg,
1558                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1559
1560         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1561         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1562                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1563
1564         print_res(name, bytes / g->p.nr_tasks / 1e9,
1565                 "GB,", "data/thread",           "GB data processed, per thread");
1566
1567         print_res(name, bytes / 1e9,
1568                 "GB,", "data-total",            "GB data processed, total");
1569
1570         print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1571                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1572
1573         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1574                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1575
1576         print_res(name, bytes / runtime_sec_max / 1e9,
1577                 "GB/sec,", "total-speed",       "GB/sec total speed");
1578
1579         if (g->p.show_details >= 2) {
1580                 char tname[14 + 2 * 10 + 1];
1581                 struct thread_data *td;
1582                 for (p = 0; p < g->p.nr_proc; p++) {
1583                         for (t = 0; t < g->p.nr_threads; t++) {
1584                                 memset(tname, 0, sizeof(tname));
1585                                 td = g->threads + p*g->p.nr_threads + t;
1586                                 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1587                                 print_res(tname, td->speed_gbs,
1588                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1589                                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1590                                         "secs", "thread-system-time", "system CPU time/thread");
1591                                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1592                                         "secs", "thread-user-time", "user CPU time/thread");
1593                         }
1594                 }
1595         }
1596
1597         free(pids);
1598
1599         deinit();
1600
1601         return 0;
1602 }
1603
1604 #define MAX_ARGS 50
1605
1606 static int command_size(const char **argv)
1607 {
1608         int size = 0;
1609
1610         while (*argv) {
1611                 size++;
1612                 argv++;
1613         }
1614
1615         BUG_ON(size >= MAX_ARGS);
1616
1617         return size;
1618 }
1619
1620 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1621 {
1622         int i;
1623
1624         printf("\n # Running %s \"perf bench numa", name);
1625
1626         for (i = 0; i < argc; i++)
1627                 printf(" %s", argv[i]);
1628
1629         printf("\"\n");
1630
1631         memset(p, 0, sizeof(*p));
1632
1633         /* Initialize nonzero defaults: */
1634
1635         p->serialize_startup            = 1;
1636         p->data_reads                   = true;
1637         p->data_writes                  = true;
1638         p->data_backwards               = true;
1639         p->data_rand_walk               = true;
1640         p->nr_loops                     = -1;
1641         p->init_random                  = true;
1642         p->mb_global_str                = "1";
1643         p->nr_proc                      = 1;
1644         p->nr_threads                   = 1;
1645         p->nr_secs                      = 5;
1646         p->run_all                      = argc == 1;
1647 }
1648
1649 static int run_bench_numa(const char *name, const char **argv)
1650 {
1651         int argc = command_size(argv);
1652
1653         init_params(&p0, name, argc, argv);
1654         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1655         if (argc)
1656                 goto err;
1657
1658         if (__bench_numa(name))
1659                 goto err;
1660
1661         return 0;
1662
1663 err:
1664         return -1;
1665 }
1666
1667 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1668 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1669
1670 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1671 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1672
1673 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1674 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1675
1676 /*
1677  * The built-in test-suite executed by "perf bench numa -a".
1678  *
1679  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1680  */
1681 static const char *tests[][MAX_ARGS] = {
1682    /* Basic single-stream NUMA bandwidth measurements: */
1683    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1684                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1685    { "RAM-bw-local-NOTHP,",
1686                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1687                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1688    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1689                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1690
1691    /* 2-stream NUMA bandwidth measurements: */
1692    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1693                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1694    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1695                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1696
1697    /* Cross-stream NUMA bandwidth measurement: */
1698    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1699                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1700
1701    /* Convergence latency measurements: */
1702    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1703    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1704    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1705    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1706    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1707    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1708    { " 4x4-convergence-NOTHP,",
1709                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1710    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1711    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1712    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1713    { " 8x4-convergence-NOTHP,",
1714                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1715    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1716    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1717    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1718    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1719    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1720
1721    /* Various NUMA process/thread layout bandwidth measurements: */
1722    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1723    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1724    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1725    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1726    { " 8x1-bw-process-NOTHP,",
1727                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1728    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1729
1730    { " 4x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1731    { " 8x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1732    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1733    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1734
1735    { " 2x3-bw-thread,",   "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1736    { " 4x4-bw-thread,",   "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1737    { " 4x6-bw-thread,",   "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1738    { " 4x8-bw-thread,",   "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1739    { " 4x8-bw-thread-NOTHP,",
1740                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1741    { " 3x3-bw-thread,",   "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1742    { " 5x5-bw-thread,",   "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1743
1744    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1745    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1746
1747    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1748    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1749    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1750    { "numa01-bw-thread-NOTHP,",
1751                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1752 };
1753
1754 static int bench_all(void)
1755 {
1756         int nr = ARRAY_SIZE(tests);
1757         int ret;
1758         int i;
1759
1760         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1761         BUG_ON(ret < 0);
1762
1763         for (i = 0; i < nr; i++) {
1764                 run_bench_numa(tests[i][0], tests[i] + 1);
1765         }
1766
1767         printf("\n");
1768
1769         return 0;
1770 }
1771
1772 int bench_numa(int argc, const char **argv)
1773 {
1774         init_params(&p0, "main,", argc, argv);
1775         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1776         if (argc)
1777                 goto err;
1778
1779         if (p0.run_all)
1780                 return bench_all();
1781
1782         if (__bench_numa(NULL))
1783                 goto err;
1784
1785         return 0;
1786
1787 err:
1788         usage_with_options(numa_usage, options);
1789         return -1;
1790 }