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