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