tools/perf/bench/numa.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * numa.c
   4  *
   5  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   6  */
   7
   8 #include <inttypes.h>
   9 /* For the CLR_() macros */
  10 #include <pthread.h>
  11
  12 #include <subcmd/parse-options.h>
  13 #include "../util/cloexec.h"
  14
  15 #include "bench.h"
  16
  17 #include <errno.h>
  18 #include <sched.h>
  19 #include <stdio.h>
  20 #include <assert.h>
  21 #include <malloc.h>
  22 #include <signal.h>
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <unistd.h>
  26 #include <sys/mman.h>
  27 #include <sys/time.h>
  28 #include <sys/resource.h>
  29 #include <sys/wait.h>
  30 #include <sys/prctl.h>
  31 #include <sys/types.h>
  32 #include <linux/kernel.h>
  33 #include <linux/time64.h>
  34 #include <linux/numa.h>
  35 #include <linux/zalloc.h>
  36
  37 #include <numa.h>
  38 #include <numaif.h>
  39
  40 #ifndef RUSAGE_THREAD
  41 # define RUSAGE_THREAD 1
  42 #endif
  43
  44 /*
  45  * Regular printout to the terminal, suppressed if -q is specified:
  46  */
  47 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  48
  49 /*
  50  * Debug printf:
  51  */
  52 #undef dprintf
  53 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  54
  55 struct thread_data {
  56         int                     curr_cpu;
  57         cpu_set_t               bind_cpumask;
  58         int                     bind_node;
  59         u8                      *process_data;
  60         int                     process_nr;
  61         int                     thread_nr;
  62         int                     task_nr;
  63         unsigned int            loops_done;
  64         u64                     val;
  65         u64                     runtime_ns;
  66         u64                     system_time_ns;
  67         u64                     user_time_ns;
  68         double                  speed_gbs;
  69         pthread_mutex_t         *process_lock;
  70 };
  71
  72 /* Parameters set by options: */
  73
  74 struct params {
  75         /* Startup synchronization: */
  76         bool                    serialize_startup;
  77
  78         /* Task hierarchy: */
  79         int                     nr_proc;
  80         int                     nr_threads;
  81
  82         /* Working set sizes: */
  83         const char              *mb_global_str;
  84         const char              *mb_proc_str;
  85         const char              *mb_proc_locked_str;
  86         const char              *mb_thread_str;
  87
  88         double                  mb_global;
  89         double                  mb_proc;
  90         double                  mb_proc_locked;
  91         double                  mb_thread;
  92
  93         /* Access patterns to the working set: */
  94         bool                    data_reads;
  95         bool                    data_writes;
  96         bool                    data_backwards;
  97         bool                    data_zero_memset;
  98         bool                    data_rand_walk;
  99         u32                     nr_loops;
 100         u32                     nr_secs;
 101         u32                     sleep_usecs;
 102
 103         /* Working set initialization: */
 104         bool                    init_zero;
 105         bool                    init_random;
 106         bool                    init_cpu0;
 107
 108         /* Misc options: */
 109         int                     show_details;
 110         int                     run_all;
 111         int                     thp;
 112
 113         long                    bytes_global;
 114         long                    bytes_process;
 115         long                    bytes_process_locked;
 116         long                    bytes_thread;
 117
 118         int                     nr_tasks;
 119         bool                    show_quiet;
 120
 121         bool                    show_convergence;
 122         bool                    measure_convergence;
 123
 124         int                     perturb_secs;
 125         int                     nr_cpus;
 126         int                     nr_nodes;
 127
 128         /* Affinity options -C and -N: */
 129         char                    *cpu_list_str;
 130         char                    *node_list_str;
 131 };
 132
 133
 134 /* Global, read-writable area, accessible to all processes and threads: */
 135
 136 struct global_info {
 137         u8                      *data;
 138
 139         pthread_mutex_t         startup_mutex;
 140         pthread_cond_t          startup_cond;
 141         int                     nr_tasks_started;
 142
 143         pthread_mutex_t         start_work_mutex;
 144         pthread_cond_t          start_work_cond;
 145         int                     nr_tasks_working;
 146         bool                    start_work;
 147
 148         pthread_mutex_t         stop_work_mutex;
 149         u64                     bytes_done;
 150
 151         struct thread_data      *threads;
 152
 153         /* Convergence latency measurement: */
 154         bool                    all_converged;
 155         bool                    stop_work;
 156
 157         int                     print_once;
 158
 159         struct params           p;
 160 };
 161
 162 static struct global_info       *g = NULL;
 163
 164 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 165 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 166
 167 struct params p0;
 168
 169 static const struct option options[] = {
 170         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 171         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 172
 173         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 174         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 175         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 176         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 177
 178         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 179         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 180         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 181
 182         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via reads (can be mixed with -W)"),
 183         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 184         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 185         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 186         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 187
 188
 189         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 190         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 191         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 192         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 193
 194         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 195         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 196         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 197         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
 198                     "convergence is reached when each process (all its threads) is running on a single NUMA node."),
 199         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 200         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 201         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 202
 203         /* Special option string parsing callbacks: */
 204         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 205                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 206                         parse_cpus_opt),
 207         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 208                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 209                         parse_nodes_opt),
 210         OPT_END()
 211 };
 212
 213 static const char * const bench_numa_usage[] = {
 214         "perf bench numa <options>",
 215         NULL
 216 };
 217
 218 static const char * const numa_usage[] = {
 219         "perf bench numa mem [<options>]",
 220         NULL
 221 };
 222
 223 /*
 224  * To get number of numa nodes present.
 225  */
 226 static int nr_numa_nodes(void)
 227 {
 228         int i, nr_nodes = 0;
 229
 230         for (i = 0; i < g->p.nr_nodes; i++) {
 231                 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
 232                         nr_nodes++;
 233         }
 234
 235         return nr_nodes;
 236 }
 237
 238 /*
 239  * To check if given numa node is present.
 240  */
 241 static int is_node_present(int node)
 242 {
 243         return numa_bitmask_isbitset(numa_nodes_ptr, node);
 244 }
 245
 246 /*
 247  * To check given numa node has cpus.
 248  */
 249 static bool node_has_cpus(int node)
 250 {
 251         struct bitmask *cpumask = numa_allocate_cpumask();
 252         bool ret = false; /* fall back to nocpus */
 253         int cpu;
 254
 255         BUG_ON(!cpumask);
 256         if (!numa_node_to_cpus(node, cpumask)) {
 257                 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 258                         if (numa_bitmask_isbitset(cpumask, cpu)) {
 259                                 ret = true;
 260                                 break;
 261                         }
 262                 }
 263         }
 264         numa_free_cpumask(cpumask);
 265
 266         return ret;
 267 }
 268
 269 static cpu_set_t bind_to_cpu(int target_cpu)
 270 {
 271         cpu_set_t orig_mask, mask;
 272         int ret;
 273
 274         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 275         BUG_ON(ret);
 276
 277         CPU_ZERO(&mask);
 278
 279         if (target_cpu == -1) {
 280                 int cpu;
 281
 282                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 283                         CPU_SET(cpu, &mask);
 284         } else {
 285                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 286                 CPU_SET(target_cpu, &mask);
 287         }
 288
 289         ret = sched_setaffinity(0, sizeof(mask), &mask);
 290         BUG_ON(ret);
 291
 292         return orig_mask;
 293 }
 294
 295 static cpu_set_t bind_to_node(int target_node)
 296 {
 297         cpu_set_t orig_mask, mask;
 298         int cpu;
 299         int ret;
 300
 301         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 302         BUG_ON(ret);
 303
 304         CPU_ZERO(&mask);
 305
 306         if (target_node == NUMA_NO_NODE) {
 307                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 308                         CPU_SET(cpu, &mask);
 309         } else {
 310                 struct bitmask *cpumask = numa_allocate_cpumask();
 311
 312                 BUG_ON(!cpumask);
 313                 if (!numa_node_to_cpus(target_node, cpumask)) {
 314                         for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 315                                 if (numa_bitmask_isbitset(cpumask, cpu))
 316                                         CPU_SET(cpu, &mask);
 317                         }
 318                 }
 319                 numa_free_cpumask(cpumask);
 320         }
 321
 322         ret = sched_setaffinity(0, sizeof(mask), &mask);
 323         BUG_ON(ret);
 324
 325         return orig_mask;
 326 }
 327
 328 static void bind_to_cpumask(cpu_set_t mask)
 329 {
 330         int ret;
 331
 332         ret = sched_setaffinity(0, sizeof(mask), &mask);
 333         BUG_ON(ret);
 334 }
 335
 336 static void mempol_restore(void)
 337 {
 338         int ret;
 339
 340         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 341
 342         BUG_ON(ret);
 343 }
 344
 345 static void bind_to_memnode(int node)
 346 {
 347         struct bitmask *node_mask;
 348         int ret;
 349
 350         if (node == NUMA_NO_NODE)
 351                 return;
 352
 353         node_mask = numa_allocate_nodemask();
 354         BUG_ON(!node_mask);
 355
 356         numa_bitmask_clearall(node_mask);
 357         numa_bitmask_setbit(node_mask, node);
 358
 359         ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
 360         dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
 361
 362         numa_bitmask_free(node_mask);
 363         BUG_ON(ret);
 364 }
 365
 366 #define HPSIZE (2*1024*1024)
 367
 368 #define set_taskname(fmt...)                            \
 369 do {                                                    \
 370         char name[20];                                  \
 371                                                         \
 372         snprintf(name, 20, fmt);                        \
 373         prctl(PR_SET_NAME, name);                       \
 374 } while (0)
 375
 376 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 377                       int init_zero, int init_cpu0, int thp, int init_random)
 378 {
 379         cpu_set_t orig_mask;
 380         ssize_t bytes;
 381         u8 *buf;
 382         int ret;
 383
 384         if (!bytes0)
 385                 return NULL;
 386
 387         /* Allocate and initialize all memory on CPU#0: */
 388         if (init_cpu0) {
 389                 int node = numa_node_of_cpu(0);
 390
 391                 orig_mask = bind_to_node(node);
 392                 bind_to_memnode(node);
 393         }
 394
 395         bytes = bytes0 + HPSIZE;
 396
 397         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 398         BUG_ON(buf == (void *)-1);
 399
 400         if (map_flags == MAP_PRIVATE) {
 401                 if (thp > 0) {
 402                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 403                         if (ret && !g->print_once) {
 404                                 g->print_once = 1;
 405                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 406                         }
 407                 }
 408                 if (thp < 0) {
 409                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 410                         if (ret && !g->print_once) {
 411                                 g->print_once = 1;
 412                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 413                         }
 414                 }
 415         }
 416
 417         if (init_zero) {
 418                 bzero(buf, bytes);
 419         } else {
 420                 /* Initialize random contents, different in each word: */
 421                 if (init_random) {
 422                         u64 *wbuf = (void *)buf;
 423                         long off = rand();
 424                         long i;
 425
 426                         for (i = 0; i < bytes/8; i++)
 427                                 wbuf[i] = i + off;
 428                 }
 429         }
 430
 431         /* Align to 2MB boundary: */
 432         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 433
 434         /* Restore affinity: */
 435         if (init_cpu0) {
 436                 bind_to_cpumask(orig_mask);
 437                 mempol_restore();
 438         }
 439
 440         return buf;
 441 }
 442
 443 static void free_data(void *data, ssize_t bytes)
 444 {
 445         int ret;
 446
 447         if (!data)
 448                 return;
 449
 450         ret = munmap(data, bytes);
 451         BUG_ON(ret);
 452 }
 453
 454 /*
 455  * Create a shared memory buffer that can be shared between processes, zeroed:
 456  */
 457 static void * zalloc_shared_data(ssize_t bytes)
 458 {
 459         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 460 }
 461
 462 /*
 463  * Create a shared memory buffer that can be shared between processes:
 464  */
 465 static void * setup_shared_data(ssize_t bytes)
 466 {
 467         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 468 }
 469
 470 /*
 471  * Allocate process-local memory - this will either be shared between
 472  * threads of this process, or only be accessed by this thread:
 473  */
 474 static void * setup_private_data(ssize_t bytes)
 475 {
 476         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 477 }
 478
 479 /*
 480  * Return a process-shared (global) mutex:
 481  */
 482 static void init_global_mutex(pthread_mutex_t *mutex)
 483 {
 484         pthread_mutexattr_t attr;
 485
 486         pthread_mutexattr_init(&attr);
 487         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 488         pthread_mutex_init(mutex, &attr);
 489 }
 490
 491 /*
 492  * Return a process-shared (global) condition variable:
 493  */
 494 static void init_global_cond(pthread_cond_t *cond)
 495 {
 496         pthread_condattr_t attr;
 497
 498         pthread_condattr_init(&attr);
 499         pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 500         pthread_cond_init(cond, &attr);
 501 }
 502
 503 static int parse_cpu_list(const char *arg)
 504 {
 505         p0.cpu_list_str = strdup(arg);
 506
 507         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 508
 509         return 0;
 510 }
 511
 512 static int parse_setup_cpu_list(void)
 513 {
 514         struct thread_data *td;
 515         char *str0, *str;
 516         int t;
 517
 518         if (!g->p.cpu_list_str)
 519                 return 0;
 520
 521         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 522
 523         str0 = str = strdup(g->p.cpu_list_str);
 524         t = 0;
 525
 526         BUG_ON(!str);
 527
 528         tprintf("# binding tasks to CPUs:\n");
 529         tprintf("#  ");
 530
 531         while (true) {
 532                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 533                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 534                 int bind_len;
 535                 int step;
 536                 int mul;
 537
 538                 tok = strsep(&str, ",");
 539                 if (!tok)
 540                         break;
 541
 542                 tok_end = strstr(tok, "-");
 543
 544                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 545                 if (!tok_end) {
 546                         /* Single CPU specified: */
 547                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 548                 } else {
 549                         /* CPU range specified (for example: "5-11"): */
 550                         bind_cpu_0 = atol(tok);
 551                         bind_cpu_1 = atol(tok_end + 1);
 552                 }
 553
 554                 step = 1;
 555                 tok_step = strstr(tok, "#");
 556                 if (tok_step) {
 557                         step = atol(tok_step + 1);
 558                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 559                 }
 560
 561                 /*
 562                  * Mask length.
 563                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 564                  * where the _4 means the next 4 CPUs are allowed.
 565                  */
 566                 bind_len = 1;
 567                 tok_len = strstr(tok, "_");
 568                 if (tok_len) {
 569                         bind_len = atol(tok_len + 1);
 570                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 571                 }
 572
 573                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 574                 mul = 1;
 575                 tok_mul = strstr(tok, "x");
 576                 if (tok_mul) {
 577                         mul = atol(tok_mul + 1);
 578                         BUG_ON(mul <= 0);
 579                 }
 580
 581                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 582
 583                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 584                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 585                         return -1;
 586                 }
 587
 588                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 589                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 590
 591                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 592                         int i;
 593
 594                         for (i = 0; i < mul; i++) {
 595                                 int cpu;
 596
 597                                 if (t >= g->p.nr_tasks) {
 598                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 599                                         goto out;
 600                                 }
 601                                 td = g->threads + t;
 602
 603                                 if (t)
 604                                         tprintf(",");
 605                                 if (bind_len > 1) {
 606                                         tprintf("%2d/%d", bind_cpu, bind_len);
 607                                 } else {
 608                                         tprintf("%2d", bind_cpu);
 609                                 }
 610
 611                                 CPU_ZERO(&td->bind_cpumask);
 612                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 613                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 614                                         CPU_SET(cpu, &td->bind_cpumask);
 615                                 }
 616                                 t++;
 617                         }
 618                 }
 619         }
 620 out:
 621
 622         tprintf("\n");
 623
 624         if (t < g->p.nr_tasks)
 625                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 626
 627         free(str0);
 628         return 0;
 629 }
 630
 631 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 632                           const char *arg, int unset __maybe_unused)
 633 {
 634         if (!arg)
 635                 return -1;
 636
 637         return parse_cpu_list(arg);
 638 }
 639
 640 static int parse_node_list(const char *arg)
 641 {
 642         p0.node_list_str = strdup(arg);
 643
 644         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 645
 646         return 0;
 647 }
 648
 649 static int parse_setup_node_list(void)
 650 {
 651         struct thread_data *td;
 652         char *str0, *str;
 653         int t;
 654
 655         if (!g->p.node_list_str)
 656                 return 0;
 657
 658         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 659
 660         str0 = str = strdup(g->p.node_list_str);
 661         t = 0;
 662
 663         BUG_ON(!str);
 664
 665         tprintf("# binding tasks to NODEs:\n");
 666         tprintf("# ");
 667
 668         while (true) {
 669                 int bind_node, bind_node_0, bind_node_1;
 670                 char *tok, *tok_end, *tok_step, *tok_mul;
 671                 int step;
 672                 int mul;
 673
 674                 tok = strsep(&str, ",");
 675                 if (!tok)
 676                         break;
 677
 678                 tok_end = strstr(tok, "-");
 679
 680                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 681                 if (!tok_end) {
 682                         /* Single NODE specified: */
 683                         bind_node_0 = bind_node_1 = atol(tok);
 684                 } else {
 685                         /* NODE range specified (for example: "5-11"): */
 686                         bind_node_0 = atol(tok);
 687                         bind_node_1 = atol(tok_end + 1);
 688                 }
 689
 690                 step = 1;
 691                 tok_step = strstr(tok, "#");
 692                 if (tok_step) {
 693                         step = atol(tok_step + 1);
 694                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 695                 }
 696
 697                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 698                 mul = 1;
 699                 tok_mul = strstr(tok, "x");
 700                 if (tok_mul) {
 701                         mul = atol(tok_mul + 1);
 702                         BUG_ON(mul <= 0);
 703                 }
 704
 705                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 706
 707                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 708                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 709                         return -1;
 710                 }
 711
 712                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 713                 BUG_ON(bind_node_0 > bind_node_1);
 714
 715                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 716                         int i;
 717
 718                         for (i = 0; i < mul; i++) {
 719                                 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
 720                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 721                                         goto out;
 722                                 }
 723                                 td = g->threads + t;
 724
 725                                 if (!t)
 726                                         tprintf(" %2d", bind_node);
 727                                 else
 728                                         tprintf(",%2d", bind_node);
 729
 730                                 td->bind_node = bind_node;
 731                                 t++;
 732                         }
 733                 }
 734         }
 735 out:
 736
 737         tprintf("\n");
 738
 739         if (t < g->p.nr_tasks)
 740                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 741
 742         free(str0);
 743         return 0;
 744 }
 745
 746 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 747                           const char *arg, int unset __maybe_unused)
 748 {
 749         if (!arg)
 750                 return -1;
 751
 752         return parse_node_list(arg);
 753 }
 754
 755 #define BIT(x) (1ul << x)
 756
 757 static inline uint32_t lfsr_32(uint32_t lfsr)
 758 {
 759         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 760         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 761 }
 762
 763 /*
 764  * Make sure there's real data dependency to RAM (when read
 765  * accesses are enabled), so the compiler, the CPU and the
 766  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 767  * accesses:
 768  */
 769 static inline u64 access_data(u64 *data, u64 val)
 770 {
 771         if (g->p.data_reads)
 772                 val += *data;
 773         if (g->p.data_writes)
 774                 *data = val + 1;
 775         return val;
 776 }
 777
 778 /*
 779  * The worker process does two types of work, a forwards going
 780  * loop and a backwards going loop.
 781  *
 782  * We do this so that on multiprocessor systems we do not create
 783  * a 'train' of processing, with highly synchronized processes,
 784  * skewing the whole benchmark.
 785  */
 786 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 787 {
 788         long words = bytes/sizeof(u64);
 789         u64 *data = (void *)__data;
 790         long chunk_0, chunk_1;
 791         u64 *d0, *d, *d1;
 792         long off;
 793         long i;
 794
 795         BUG_ON(!data && words);
 796         BUG_ON(data && !words);
 797
 798         if (!data)
 799                 return val;
 800
 801         /* Very simple memset() work variant: */
 802         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 803                 bzero(data, bytes);
 804                 return val;
 805         }
 806
 807         /* Spread out by PID/TID nr and by loop nr: */
 808         chunk_0 = words/nr_max;
 809         chunk_1 = words/g->p.nr_loops;
 810         off = nr*chunk_0 + loop*chunk_1;
 811
 812         while (off >= words)
 813                 off -= words;
 814
 815         if (g->p.data_rand_walk) {
 816                 u32 lfsr = nr + loop + val;
 817                 int j;
 818
 819                 for (i = 0; i < words/1024; i++) {
 820                         long start, end;
 821
 822                         lfsr = lfsr_32(lfsr);
 823
 824                         start = lfsr % words;
 825                         end = min(start + 1024, words-1);
 826
 827                         if (g->p.data_zero_memset) {
 828                                 bzero(data + start, (end-start) * sizeof(u64));
 829                         } else {
 830                                 for (j = start; j < end; j++)
 831                                         val = access_data(data + j, val);
 832                         }
 833                 }
 834         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 835                 /* Process data forwards: */
 836
 837                 d0 = data + off;
 838                 d  = data + off + 1;
 839                 d1 = data + words;
 840
 841                 for (;;) {
 842                         if (unlikely(d >= d1))
 843                                 d = data;
 844                         if (unlikely(d == d0))
 845                                 break;
 846
 847                         val = access_data(d, val);
 848
 849                         d++;
 850                 }
 851         } else {
 852                 /* Process data backwards: */
 853
 854                 d0 = data + off;
 855                 d  = data + off - 1;
 856                 d1 = data + words;
 857
 858                 for (;;) {
 859                         if (unlikely(d < data))
 860                                 d = data + words-1;
 861                         if (unlikely(d == d0))
 862                                 break;
 863
 864                         val = access_data(d, val);
 865
 866                         d--;
 867                 }
 868         }
 869
 870         return val;
 871 }
 872
 873 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 874 {
 875         unsigned int cpu;
 876
 877         cpu = sched_getcpu();
 878
 879         g->threads[task_nr].curr_cpu = cpu;
 880         prctl(0, bytes_worked);
 881 }
 882
 883 /*
 884  * Count the number of nodes a process's threads
 885  * are spread out on.
 886  *
 887  * A count of 1 means that the process is compressed
 888  * to a single node. A count of g->p.nr_nodes means it's
 889  * spread out on the whole system.
 890  */
 891 static int count_process_nodes(int process_nr)
 892 {
 893         char *node_present;
 894         int nodes;
 895         int n, t;
 896
 897         node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
 898         BUG_ON(!node_present);
 899         for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
 900                 node_present[nodes] = 0;
 901
 902         for (t = 0; t < g->p.nr_threads; t++) {
 903                 struct thread_data *td;
 904                 int task_nr;
 905                 int node;
 906
 907                 task_nr = process_nr*g->p.nr_threads + t;
 908                 td = g->threads + task_nr;
 909
 910                 node = numa_node_of_cpu(td->curr_cpu);
 911                 if (node < 0) /* curr_cpu was likely still -1 */ {
 912                         free(node_present);
 913                         return 0;
 914                 }
 915
 916                 node_present[node] = 1;
 917         }
 918
 919         nodes = 0;
 920
 921         for (n = 0; n < g->p.nr_nodes; n++)
 922                 nodes += node_present[n];
 923
 924         free(node_present);
 925         return nodes;
 926 }
 927
 928 /*
 929  * Count the number of distinct process-threads a node contains.
 930  *
 931  * A count of 1 means that the node contains only a single
 932  * process. If all nodes on the system contain at most one
 933  * process then we are well-converged.
 934  */
 935 static int count_node_processes(int node)
 936 {
 937         int processes = 0;
 938         int t, p;
 939
 940         for (p = 0; p < g->p.nr_proc; p++) {
 941                 for (t = 0; t < g->p.nr_threads; t++) {
 942                         struct thread_data *td;
 943                         int task_nr;
 944                         int n;
 945
 946                         task_nr = p*g->p.nr_threads + t;
 947                         td = g->threads + task_nr;
 948
 949                         n = numa_node_of_cpu(td->curr_cpu);
 950                         if (n == node) {
 951                                 processes++;
 952                                 break;
 953                         }
 954                 }
 955         }
 956
 957         return processes;
 958 }
 959
 960 static void calc_convergence_compression(int *strong)
 961 {
 962         unsigned int nodes_min, nodes_max;
 963         int p;
 964
 965         nodes_min = -1;
 966         nodes_max =  0;
 967
 968         for (p = 0; p < g->p.nr_proc; p++) {
 969                 unsigned int nodes = count_process_nodes(p);
 970
 971                 if (!nodes) {
 972                         *strong = 0;
 973                         return;
 974                 }
 975
 976                 nodes_min = min(nodes, nodes_min);
 977                 nodes_max = max(nodes, nodes_max);
 978         }
 979
 980         /* Strong convergence: all threads compress on a single node: */
 981         if (nodes_min == 1 && nodes_max == 1) {
 982                 *strong = 1;
 983         } else {
 984                 *strong = 0;
 985                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 986         }
 987 }
 988
 989 static void calc_convergence(double runtime_ns_max, double *convergence)
 990 {
 991         unsigned int loops_done_min, loops_done_max;
 992         int process_groups;
 993         int *nodes;
 994         int distance;
 995         int nr_min;
 996         int nr_max;
 997         int strong;
 998         int sum;
 999         int nr;
1000         int node;
1001         int cpu;
1002         int t;
1003
1004         if (!g->p.show_convergence && !g->p.measure_convergence)
1005                 return;
1006
1007         nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1008         BUG_ON(!nodes);
1009         for (node = 0; node < g->p.nr_nodes; node++)
1010                 nodes[node] = 0;
1011
1012         loops_done_min = -1;
1013         loops_done_max = 0;
1014
1015         for (t = 0; t < g->p.nr_tasks; t++) {
1016                 struct thread_data *td = g->threads + t;
1017                 unsigned int loops_done;
1018
1019                 cpu = td->curr_cpu;
1020
1021                 /* Not all threads have written it yet: */
1022                 if (cpu < 0)
1023                         continue;
1024
1025                 node = numa_node_of_cpu(cpu);
1026
1027                 nodes[node]++;
1028
1029                 loops_done = td->loops_done;
1030                 loops_done_min = min(loops_done, loops_done_min);
1031                 loops_done_max = max(loops_done, loops_done_max);
1032         }
1033
1034         nr_max = 0;
1035         nr_min = g->p.nr_tasks;
1036         sum = 0;
1037
1038         for (node = 0; node < g->p.nr_nodes; node++) {
1039                 if (!is_node_present(node))
1040                         continue;
1041                 nr = nodes[node];
1042                 nr_min = min(nr, nr_min);
1043                 nr_max = max(nr, nr_max);
1044                 sum += nr;
1045         }
1046         BUG_ON(nr_min > nr_max);
1047
1048         BUG_ON(sum > g->p.nr_tasks);
1049
1050         if (0 && (sum < g->p.nr_tasks)) {
1051                 free(nodes);
1052                 return;
1053         }
1054
1055         /*
1056          * Count the number of distinct process groups present
1057          * on nodes - when we are converged this will decrease
1058          * to g->p.nr_proc:
1059          */
1060         process_groups = 0;
1061
1062         for (node = 0; node < g->p.nr_nodes; node++) {
1063                 int processes;
1064
1065                 if (!is_node_present(node))
1066                         continue;
1067                 processes = count_node_processes(node);
1068                 nr = nodes[node];
1069                 tprintf(" %2d/%-2d", nr, processes);
1070
1071                 process_groups += processes;
1072         }
1073
1074         distance = nr_max - nr_min;
1075
1076         tprintf(" [%2d/%-2d]", distance, process_groups);
1077
1078         tprintf(" l:%3d-%-3d (%3d)",
1079                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1080
1081         if (loops_done_min && loops_done_max) {
1082                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1083
1084                 tprintf(" [%4.1f%%]", skew * 100.0);
1085         }
1086
1087         calc_convergence_compression(&strong);
1088
1089         if (strong && process_groups == g->p.nr_proc) {
1090                 if (!*convergence) {
1091                         *convergence = runtime_ns_max;
1092                         tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1093                         if (g->p.measure_convergence) {
1094                                 g->all_converged = true;
1095                                 g->stop_work = true;
1096                         }
1097                 }
1098         } else {
1099                 if (*convergence) {
1100                         tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1101                         *convergence = 0;
1102                 }
1103                 tprintf("\n");
1104         }
1105
1106         free(nodes);
1107 }
1108
1109 static void show_summary(double runtime_ns_max, int l, double *convergence)
1110 {
1111         tprintf("\r #  %5.1f%%  [%.1f mins]",
1112                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1113
1114         calc_convergence(runtime_ns_max, convergence);
1115
1116         if (g->p.show_details >= 0)
1117                 fflush(stdout);
1118 }
1119
1120 static void *worker_thread(void *__tdata)
1121 {
1122         struct thread_data *td = __tdata;
1123         struct timeval start0, start, stop, diff;
1124         int process_nr = td->process_nr;
1125         int thread_nr = td->thread_nr;
1126         unsigned long last_perturbance;
1127         int task_nr = td->task_nr;
1128         int details = g->p.show_details;
1129         int first_task, last_task;
1130         double convergence = 0;
1131         u64 val = td->val;
1132         double runtime_ns_max;
1133         u8 *global_data;
1134         u8 *process_data;
1135         u8 *thread_data;
1136         u64 bytes_done, secs;
1137         long work_done;
1138         u32 l;
1139         struct rusage rusage;
1140
1141         bind_to_cpumask(td->bind_cpumask);
1142         bind_to_memnode(td->bind_node);
1143
1144         set_taskname("thread %d/%d", process_nr, thread_nr);
1145
1146         global_data = g->data;
1147         process_data = td->process_data;
1148         thread_data = setup_private_data(g->p.bytes_thread);
1149
1150         bytes_done = 0;
1151
1152         last_task = 0;
1153         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1154                 last_task = 1;
1155
1156         first_task = 0;
1157         if (process_nr == 0 && thread_nr == 0)
1158                 first_task = 1;
1159
1160         if (details >= 2) {
1161                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1162                         process_nr, thread_nr, global_data, process_data, thread_data);
1163         }
1164
1165         if (g->p.serialize_startup) {
1166                 pthread_mutex_lock(&g->startup_mutex);
1167                 g->nr_tasks_started++;
1168                 /* The last thread wakes the main process. */
1169                 if (g->nr_tasks_started == g->p.nr_tasks)
1170                         pthread_cond_signal(&g->startup_cond);
1171
1172                 pthread_mutex_unlock(&g->startup_mutex);
1173
1174                 /* Here we will wait for the main process to start us all at once: */
1175                 pthread_mutex_lock(&g->start_work_mutex);
1176                 g->start_work = false;
1177                 g->nr_tasks_working++;
1178                 while (!g->start_work)
1179                         pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
1180
1181                 pthread_mutex_unlock(&g->start_work_mutex);
1182         }
1183
1184         gettimeofday(&start0, NULL);
1185
1186         start = stop = start0;
1187         last_perturbance = start.tv_sec;
1188
1189         for (l = 0; l < g->p.nr_loops; l++) {
1190                 start = stop;
1191
1192                 if (g->stop_work)
1193                         break;
1194
1195                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1196                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1197                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1198
1199                 if (g->p.sleep_usecs) {
1200                         pthread_mutex_lock(td->process_lock);
1201                         usleep(g->p.sleep_usecs);
1202                         pthread_mutex_unlock(td->process_lock);
1203                 }
1204                 /*
1205                  * Amount of work to be done under a process-global lock:
1206                  */
1207                 if (g->p.bytes_process_locked) {
1208                         pthread_mutex_lock(td->process_lock);
1209                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1210                         pthread_mutex_unlock(td->process_lock);
1211                 }
1212
1213                 work_done = g->p.bytes_global + g->p.bytes_process +
1214                             g->p.bytes_process_locked + g->p.bytes_thread;
1215
1216                 update_curr_cpu(task_nr, work_done);
1217                 bytes_done += work_done;
1218
1219                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1220                         continue;
1221
1222                 td->loops_done = l;
1223
1224                 gettimeofday(&stop, NULL);
1225
1226                 /* Check whether our max runtime timed out: */
1227                 if (g->p.nr_secs) {
1228                         timersub(&stop, &start0, &diff);
1229                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1230                                 g->stop_work = true;
1231                                 break;
1232                         }
1233                 }
1234
1235                 /* Update the summary at most once per second: */
1236                 if (start.tv_sec == stop.tv_sec)
1237                         continue;
1238
1239                 /*
1240                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1241                  * by migrating to CPU#0:
1242                  */
1243                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1244                         cpu_set_t orig_mask;
1245                         int target_cpu;
1246                         int this_cpu;
1247
1248                         last_perturbance = stop.tv_sec;
1249
1250                         /*
1251                          * Depending on where we are running, move into
1252                          * the other half of the system, to create some
1253                          * real disturbance:
1254                          */
1255                         this_cpu = g->threads[task_nr].curr_cpu;
1256                         if (this_cpu < g->p.nr_cpus/2)
1257                                 target_cpu = g->p.nr_cpus-1;
1258                         else
1259                                 target_cpu = 0;
1260
1261                         orig_mask = bind_to_cpu(target_cpu);
1262
1263                         /* Here we are running on the target CPU already */
1264                         if (details >= 1)
1265                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1266
1267                         bind_to_cpumask(orig_mask);
1268                 }
1269
1270                 if (details >= 3) {
1271                         timersub(&stop, &start, &diff);
1272                         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1273                         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1274
1275                         if (details >= 0) {
1276                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1277                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1278                         }
1279                         fflush(stdout);
1280                 }
1281                 if (!last_task)
1282                         continue;
1283
1284                 timersub(&stop, &start0, &diff);
1285                 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1286                 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1287
1288                 show_summary(runtime_ns_max, l, &convergence);
1289         }
1290
1291         gettimeofday(&stop, NULL);
1292         timersub(&stop, &start0, &diff);
1293         td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1294         td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1295         secs = td->runtime_ns / NSEC_PER_SEC;
1296         td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1297
1298         getrusage(RUSAGE_THREAD, &rusage);
1299         td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1300         td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1301         td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1302         td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1303
1304         free_data(thread_data, g->p.bytes_thread);
1305
1306         pthread_mutex_lock(&g->stop_work_mutex);
1307         g->bytes_done += bytes_done;
1308         pthread_mutex_unlock(&g->stop_work_mutex);
1309
1310         return NULL;
1311 }
1312
1313 /*
1314  * A worker process starts a couple of threads:
1315  */
1316 static void worker_process(int process_nr)
1317 {
1318         pthread_mutex_t process_lock;
1319         struct thread_data *td;
1320         pthread_t *pthreads;
1321         u8 *process_data;
1322         int task_nr;
1323         int ret;
1324         int t;
1325
1326         pthread_mutex_init(&process_lock, NULL);
1327         set_taskname("process %d", process_nr);
1328
1329         /*
1330          * Pick up the memory policy and the CPU binding of our first thread,
1331          * so that we initialize memory accordingly:
1332          */
1333         task_nr = process_nr*g->p.nr_threads;
1334         td = g->threads + task_nr;
1335
1336         bind_to_memnode(td->bind_node);
1337         bind_to_cpumask(td->bind_cpumask);
1338
1339         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1340         process_data = setup_private_data(g->p.bytes_process);
1341
1342         if (g->p.show_details >= 3) {
1343                 printf(" # process %2d global mem: %p, process mem: %p\n",
1344                         process_nr, g->data, process_data);
1345         }
1346
1347         for (t = 0; t < g->p.nr_threads; t++) {
1348                 task_nr = process_nr*g->p.nr_threads + t;
1349                 td = g->threads + task_nr;
1350
1351                 td->process_data = process_data;
1352                 td->process_nr   = process_nr;
1353                 td->thread_nr    = t;
1354                 td->task_nr      = task_nr;
1355                 td->val          = rand();
1356                 td->curr_cpu     = -1;
1357                 td->process_lock = &process_lock;
1358
1359                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1360                 BUG_ON(ret);
1361         }
1362
1363         for (t = 0; t < g->p.nr_threads; t++) {
1364                 ret = pthread_join(pthreads[t], NULL);
1365                 BUG_ON(ret);
1366         }
1367
1368         free_data(process_data, g->p.bytes_process);
1369         free(pthreads);
1370 }
1371
1372 static void print_summary(void)
1373 {
1374         if (g->p.show_details < 0)
1375                 return;
1376
1377         printf("\n ###\n");
1378         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1379                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1380         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1381                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1382         printf(" #      %5dx %5ldMB process shared mem operations\n",
1383                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1384         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1385                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1386
1387         printf(" ###\n");
1388
1389         printf("\n ###\n"); fflush(stdout);
1390 }
1391
1392 static void init_thread_data(void)
1393 {
1394         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1395         int t;
1396
1397         g->threads = zalloc_shared_data(size);
1398
1399         for (t = 0; t < g->p.nr_tasks; t++) {
1400                 struct thread_data *td = g->threads + t;
1401                 int cpu;
1402
1403                 /* Allow all nodes by default: */
1404                 td->bind_node = NUMA_NO_NODE;
1405
1406                 /* Allow all CPUs by default: */
1407                 CPU_ZERO(&td->bind_cpumask);
1408                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1409                         CPU_SET(cpu, &td->bind_cpumask);
1410         }
1411 }
1412
1413 static void deinit_thread_data(void)
1414 {
1415         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1416
1417         free_data(g->threads, size);
1418 }
1419
1420 static int init(void)
1421 {
1422         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1423
1424         /* Copy over options: */
1425         g->p = p0;
1426
1427         g->p.nr_cpus = numa_num_configured_cpus();
1428
1429         g->p.nr_nodes = numa_max_node() + 1;
1430
1431         /* char array in count_process_nodes(): */
1432         BUG_ON(g->p.nr_nodes < 0);
1433
1434         if (g->p.show_quiet && !g->p.show_details)
1435                 g->p.show_details = -1;
1436
1437         /* Some memory should be specified: */
1438         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1439                 return -1;
1440
1441         if (g->p.mb_global_str) {
1442                 g->p.mb_global = atof(g->p.mb_global_str);
1443                 BUG_ON(g->p.mb_global < 0);
1444         }
1445
1446         if (g->p.mb_proc_str) {
1447                 g->p.mb_proc = atof(g->p.mb_proc_str);
1448                 BUG_ON(g->p.mb_proc < 0);
1449         }
1450
1451         if (g->p.mb_proc_locked_str) {
1452                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1453                 BUG_ON(g->p.mb_proc_locked < 0);
1454                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1455         }
1456
1457         if (g->p.mb_thread_str) {
1458                 g->p.mb_thread = atof(g->p.mb_thread_str);
1459                 BUG_ON(g->p.mb_thread < 0);
1460         }
1461
1462         BUG_ON(g->p.nr_threads <= 0);
1463         BUG_ON(g->p.nr_proc <= 0);
1464
1465         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1466
1467         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1468         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1469         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1470         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1471
1472         g->data = setup_shared_data(g->p.bytes_global);
1473
1474         /* Startup serialization: */
1475         init_global_mutex(&g->start_work_mutex);
1476         init_global_cond(&g->start_work_cond);
1477         init_global_mutex(&g->startup_mutex);
1478         init_global_cond(&g->startup_cond);
1479         init_global_mutex(&g->stop_work_mutex);
1480
1481         init_thread_data();
1482
1483         tprintf("#\n");
1484         if (parse_setup_cpu_list() || parse_setup_node_list())
1485                 return -1;
1486         tprintf("#\n");
1487
1488         print_summary();
1489
1490         return 0;
1491 }
1492
1493 static void deinit(void)
1494 {
1495         free_data(g->data, g->p.bytes_global);
1496         g->data = NULL;
1497
1498         deinit_thread_data();
1499
1500         free_data(g, sizeof(*g));
1501         g = NULL;
1502 }
1503
1504 /*
1505  * Print a short or long result, depending on the verbosity setting:
1506  */
1507 static void print_res(const char *name, double val,
1508                       const char *txt_unit, const char *txt_short, const char *txt_long)
1509 {
1510         if (!name)
1511                 name = "main,";
1512
1513         if (!g->p.show_quiet)
1514                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1515         else
1516                 printf(" %14.3f %s\n", val, txt_long);
1517 }
1518
1519 static int __bench_numa(const char *name)
1520 {
1521         struct timeval start, stop, diff;
1522         u64 runtime_ns_min, runtime_ns_sum;
1523         pid_t *pids, pid, wpid;
1524         double delta_runtime;
1525         double runtime_avg;
1526         double runtime_sec_max;
1527         double runtime_sec_min;
1528         int wait_stat;
1529         double bytes;
1530         int i, t, p;
1531
1532         if (init())
1533                 return -1;
1534
1535         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1536         pid = -1;
1537
1538         if (g->p.serialize_startup) {
1539                 tprintf(" #\n");
1540                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1541         }
1542
1543         gettimeofday(&start, NULL);
1544
1545         for (i = 0; i < g->p.nr_proc; i++) {
1546                 pid = fork();
1547                 dprintf(" # process %2d: PID %d\n", i, pid);
1548
1549                 BUG_ON(pid < 0);
1550                 if (!pid) {
1551                         /* Child process: */
1552                         worker_process(i);
1553
1554                         exit(0);
1555                 }
1556                 pids[i] = pid;
1557
1558         }
1559
1560         if (g->p.serialize_startup) {
1561                 bool threads_ready = false;
1562                 double startup_sec;
1563
1564                 /*
1565                  * Wait for all the threads to start up. The last thread will
1566                  * signal this process.
1567                  */
1568                 pthread_mutex_lock(&g->startup_mutex);
1569                 while (g->nr_tasks_started != g->p.nr_tasks)
1570                         pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
1571
1572                 pthread_mutex_unlock(&g->startup_mutex);
1573
1574                 /* Wait for all threads to be at the start_work_cond. */
1575                 while (!threads_ready) {
1576                         pthread_mutex_lock(&g->start_work_mutex);
1577                         threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1578                         pthread_mutex_unlock(&g->start_work_mutex);
1579                         if (!threads_ready)
1580                                 usleep(1);
1581                 }
1582
1583                 gettimeofday(&stop, NULL);
1584
1585                 timersub(&stop, &start, &diff);
1586
1587                 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1588                 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1589                 startup_sec /= NSEC_PER_SEC;
1590
1591                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1592                 tprintf(" #\n");
1593
1594                 start = stop;
1595                 /* Start all threads running. */
1596                 pthread_mutex_lock(&g->start_work_mutex);
1597                 g->start_work = true;
1598                 pthread_mutex_unlock(&g->start_work_mutex);
1599                 pthread_cond_broadcast(&g->start_work_cond);
1600         } else {
1601                 gettimeofday(&start, NULL);
1602         }
1603
1604         /* Parent process: */
1605
1606
1607         for (i = 0; i < g->p.nr_proc; i++) {
1608                 wpid = waitpid(pids[i], &wait_stat, 0);
1609                 BUG_ON(wpid < 0);
1610                 BUG_ON(!WIFEXITED(wait_stat));
1611
1612         }
1613
1614         runtime_ns_sum = 0;
1615         runtime_ns_min = -1LL;
1616
1617         for (t = 0; t < g->p.nr_tasks; t++) {
1618                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1619
1620                 runtime_ns_sum += thread_runtime_ns;
1621                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1622         }
1623
1624         gettimeofday(&stop, NULL);
1625         timersub(&stop, &start, &diff);
1626
1627         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1628
1629         tprintf("\n ###\n");
1630         tprintf("\n");
1631
1632         runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1633         runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1634         runtime_sec_max /= NSEC_PER_SEC;
1635
1636         runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1637
1638         bytes = g->bytes_done;
1639         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1640
1641         if (g->p.measure_convergence) {
1642                 print_res(name, runtime_sec_max,
1643                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1644         }
1645
1646         print_res(name, runtime_sec_max,
1647                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1648
1649         print_res(name, runtime_sec_min,
1650                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1651
1652         print_res(name, runtime_avg,
1653                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1654
1655         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1656         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1657                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1658
1659         print_res(name, bytes / g->p.nr_tasks / 1e9,
1660                 "GB,", "data/thread",           "GB data processed, per thread");
1661
1662         print_res(name, bytes / 1e9,
1663                 "GB,", "data-total",            "GB data processed, total");
1664
1665         print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1666                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1667
1668         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1669                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1670
1671         print_res(name, bytes / runtime_sec_max / 1e9,
1672                 "GB/sec,", "total-speed",       "GB/sec total speed");
1673
1674         if (g->p.show_details >= 2) {
1675                 char tname[14 + 2 * 10 + 1];
1676                 struct thread_data *td;
1677                 for (p = 0; p < g->p.nr_proc; p++) {
1678                         for (t = 0; t < g->p.nr_threads; t++) {
1679                                 memset(tname, 0, sizeof(tname));
1680                                 td = g->threads + p*g->p.nr_threads + t;
1681                                 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1682                                 print_res(tname, td->speed_gbs,
1683                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1684                                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1685                                         "secs", "thread-system-time", "system CPU time/thread");
1686                                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1687                                         "secs", "thread-user-time", "user CPU time/thread");
1688                         }
1689                 }
1690         }
1691
1692         free(pids);
1693
1694         deinit();
1695
1696         return 0;
1697 }
1698
1699 #define MAX_ARGS 50
1700
1701 static int command_size(const char **argv)
1702 {
1703         int size = 0;
1704
1705         while (*argv) {
1706                 size++;
1707                 argv++;
1708         }
1709
1710         BUG_ON(size >= MAX_ARGS);
1711
1712         return size;
1713 }
1714
1715 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1716 {
1717         int i;
1718
1719         printf("\n # Running %s \"perf bench numa", name);
1720
1721         for (i = 0; i < argc; i++)
1722                 printf(" %s", argv[i]);
1723
1724         printf("\"\n");
1725
1726         memset(p, 0, sizeof(*p));
1727
1728         /* Initialize nonzero defaults: */
1729
1730         p->serialize_startup            = 1;
1731         p->data_reads                   = true;
1732         p->data_writes                  = true;
1733         p->data_backwards               = true;
1734         p->data_rand_walk               = true;
1735         p->nr_loops                     = -1;
1736         p->init_random                  = true;
1737         p->mb_global_str                = "1";
1738         p->nr_proc                      = 1;
1739         p->nr_threads                   = 1;
1740         p->nr_secs                      = 5;
1741         p->run_all                      = argc == 1;
1742 }
1743
1744 static int run_bench_numa(const char *name, const char **argv)
1745 {
1746         int argc = command_size(argv);
1747
1748         init_params(&p0, name, argc, argv);
1749         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1750         if (argc)
1751                 goto err;
1752
1753         if (__bench_numa(name))
1754                 goto err;
1755
1756         return 0;
1757
1758 err:
1759         return -1;
1760 }
1761
1762 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1763 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1764
1765 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1766 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1767
1768 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1769 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1770
1771 /*
1772  * The built-in test-suite executed by "perf bench numa -a".
1773  *
1774  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1775  */
1776 static const char *tests[][MAX_ARGS] = {
1777    /* Basic single-stream NUMA bandwidth measurements: */
1778    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1779                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1780    { "RAM-bw-local-NOTHP,",
1781                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1782                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1783    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1784                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1785
1786    /* 2-stream NUMA bandwidth measurements: */
1787    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1788                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1789    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1790                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1791
1792    /* Cross-stream NUMA bandwidth measurement: */
1793    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1794                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1795
1796    /* Convergence latency measurements: */
1797    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1798    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1799    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1800    { " 2x3-convergence,", "mem",  "-p",  "2", "-t",  "3", "-P", "1020", OPT_CONV },
1801    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1802    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1803    { " 4x4-convergence-NOTHP,",
1804                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1805    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1806    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1807    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1808    { " 8x4-convergence-NOTHP,",
1809                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1810    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1811    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1812    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1813    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1814    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1815
1816    /* Various NUMA process/thread layout bandwidth measurements: */
1817    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1818    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1819    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1820    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1821    { " 8x1-bw-process-NOTHP,",
1822                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1823    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1824
1825    { " 1x4-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1826    { " 1x8-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1827    { "1x16-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1828    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1829
1830    { " 2x3-bw-process,",  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1831    { " 4x4-bw-process,",  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1832    { " 4x6-bw-process,",  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1833    { " 4x8-bw-process,",  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1834    { " 4x8-bw-process-NOTHP,",
1835                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1836    { " 3x3-bw-process,",  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1837    { " 5x5-bw-process,",  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1838
1839    { "2x16-bw-process,",  "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1840    { "1x32-bw-process,",  "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1841
1842    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1843    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1844    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1845    { "numa01-bw-thread-NOTHP,",
1846                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1847 };
1848
1849 static int bench_all(void)
1850 {
1851         int nr = ARRAY_SIZE(tests);
1852         int ret;
1853         int i;
1854
1855         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1856         BUG_ON(ret < 0);
1857
1858         for (i = 0; i < nr; i++) {
1859                 run_bench_numa(tests[i][0], tests[i] + 1);
1860         }
1861
1862         printf("\n");
1863
1864         return 0;
1865 }
1866
1867 int bench_numa(int argc, const char **argv)
1868 {
1869         init_params(&p0, "main,", argc, argv);
1870         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1871         if (argc)
1872                 goto err;
1873
1874         if (p0.run_all)
1875                 return bench_all();
1876
1877         if (__bench_numa(NULL))
1878                 goto err;
1879
1880         return 0;
1881
1882 err:
1883         usage_with_options(numa_usage, options);
1884         return -1;
1885 }