tools/perf/bench/numa.c

   1 /*
   2  * numa.c
   3  *
   4  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5  */
   6
   7 /* For the CLR_() macros */
   8 #include <pthread.h>
   9
  10 #include "../perf.h"
  11 #include "../builtin.h"
  12 #include "../util/util.h"
  13 #include <subcmd/parse-options.h>
  14 #include "../util/cloexec.h"
  15
  16 #include "bench.h"
  17
  18 #include <errno.h>
  19 #include <sched.h>
  20 #include <stdio.h>
  21 #include <assert.h>
  22 #include <malloc.h>
  23 #include <signal.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <sys/mman.h>
  28 #include <sys/time.h>
  29 #include <sys/resource.h>
  30 #include <sys/wait.h>
  31 #include <sys/prctl.h>
  32 #include <sys/types.h>
  33
  34 #include <numa.h>
  35 #include <numaif.h>
  36
  37 /*
  38  * Regular printout to the terminal, supressed if -q is specified:
  39  */
  40 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  41
  42 /*
  43  * Debug printf:
  44  */
  45 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  46
  47 struct thread_data {
  48         int                     curr_cpu;
  49         cpu_set_t               bind_cpumask;
  50         int                     bind_node;
  51         u8                      *process_data;
  52         int                     process_nr;
  53         int                     thread_nr;
  54         int                     task_nr;
  55         unsigned int            loops_done;
  56         u64                     val;
  57         u64                     runtime_ns;
  58         u64                     system_time_ns;
  59         u64                     user_time_ns;
  60         double                  speed_gbs;
  61         pthread_mutex_t         *process_lock;
  62 };
  63
  64 /* Parameters set by options: */
  65
  66 struct params {
  67         /* Startup synchronization: */
  68         bool                    serialize_startup;
  69
  70         /* Task hierarchy: */
  71         int                     nr_proc;
  72         int                     nr_threads;
  73
  74         /* Working set sizes: */
  75         const char              *mb_global_str;
  76         const char              *mb_proc_str;
  77         const char              *mb_proc_locked_str;
  78         const char              *mb_thread_str;
  79
  80         double                  mb_global;
  81         double                  mb_proc;
  82         double                  mb_proc_locked;
  83         double                  mb_thread;
  84
  85         /* Access patterns to the working set: */
  86         bool                    data_reads;
  87         bool                    data_writes;
  88         bool                    data_backwards;
  89         bool                    data_zero_memset;
  90         bool                    data_rand_walk;
  91         u32                     nr_loops;
  92         u32                     nr_secs;
  93         u32                     sleep_usecs;
  94
  95         /* Working set initialization: */
  96         bool                    init_zero;
  97         bool                    init_random;
  98         bool                    init_cpu0;
  99
 100         /* Misc options: */
 101         int                     show_details;
 102         int                     run_all;
 103         int                     thp;
 104
 105         long                    bytes_global;
 106         long                    bytes_process;
 107         long                    bytes_process_locked;
 108         long                    bytes_thread;
 109
 110         int                     nr_tasks;
 111         bool                    show_quiet;
 112
 113         bool                    show_convergence;
 114         bool                    measure_convergence;
 115
 116         int                     perturb_secs;
 117         int                     nr_cpus;
 118         int                     nr_nodes;
 119
 120         /* Affinity options -C and -N: */
 121         char                    *cpu_list_str;
 122         char                    *node_list_str;
 123 };
 124
 125
 126 /* Global, read-writable area, accessible to all processes and threads: */
 127
 128 struct global_info {
 129         u8                      *data;
 130
 131         pthread_mutex_t         startup_mutex;
 132         int                     nr_tasks_started;
 133
 134         pthread_mutex_t         startup_done_mutex;
 135
 136         pthread_mutex_t         start_work_mutex;
 137         int                     nr_tasks_working;
 138
 139         pthread_mutex_t         stop_work_mutex;
 140         u64                     bytes_done;
 141
 142         struct thread_data      *threads;
 143
 144         /* Convergence latency measurement: */
 145         bool                    all_converged;
 146         bool                    stop_work;
 147
 148         int                     print_once;
 149
 150         struct params           p;
 151 };
 152
 153 static struct global_info       *g = NULL;
 154
 155 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 156 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 157
 158 struct params p0;
 159
 160 static const struct option options[] = {
 161         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 162         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 163
 164         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 165         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 166         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 167         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 168
 169         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 170         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 171         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 172
 173         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via writes (can be mixed with -W)"),
 174         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 175         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 176         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 177         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 178
 179
 180         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 181         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 182         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 183         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 184
 185         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 186         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 187         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 188         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
 189         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 190         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 191         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 192
 193         /* Special option string parsing callbacks: */
 194         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 195                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 196                         parse_cpus_opt),
 197         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 198                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 199                         parse_nodes_opt),
 200         OPT_END()
 201 };
 202
 203 static const char * const bench_numa_usage[] = {
 204         "perf bench numa <options>",
 205         NULL
 206 };
 207
 208 static const char * const numa_usage[] = {
 209         "perf bench numa mem [<options>]",
 210         NULL
 211 };
 212
 213 static cpu_set_t bind_to_cpu(int target_cpu)
 214 {
 215         cpu_set_t orig_mask, mask;
 216         int ret;
 217
 218         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 219         BUG_ON(ret);
 220
 221         CPU_ZERO(&mask);
 222
 223         if (target_cpu == -1) {
 224                 int cpu;
 225
 226                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 227                         CPU_SET(cpu, &mask);
 228         } else {
 229                 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 230                 CPU_SET(target_cpu, &mask);
 231         }
 232
 233         ret = sched_setaffinity(0, sizeof(mask), &mask);
 234         BUG_ON(ret);
 235
 236         return orig_mask;
 237 }
 238
 239 static cpu_set_t bind_to_node(int target_node)
 240 {
 241         int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 242         cpu_set_t orig_mask, mask;
 243         int cpu;
 244         int ret;
 245
 246         BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 247         BUG_ON(!cpus_per_node);
 248
 249         ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 250         BUG_ON(ret);
 251
 252         CPU_ZERO(&mask);
 253
 254         if (target_node == -1) {
 255                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 256                         CPU_SET(cpu, &mask);
 257         } else {
 258                 int cpu_start = (target_node + 0) * cpus_per_node;
 259                 int cpu_stop  = (target_node + 1) * cpus_per_node;
 260
 261                 BUG_ON(cpu_stop > g->p.nr_cpus);
 262
 263                 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 264                         CPU_SET(cpu, &mask);
 265         }
 266
 267         ret = sched_setaffinity(0, sizeof(mask), &mask);
 268         BUG_ON(ret);
 269
 270         return orig_mask;
 271 }
 272
 273 static void bind_to_cpumask(cpu_set_t mask)
 274 {
 275         int ret;
 276
 277         ret = sched_setaffinity(0, sizeof(mask), &mask);
 278         BUG_ON(ret);
 279 }
 280
 281 static void mempol_restore(void)
 282 {
 283         int ret;
 284
 285         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 286
 287         BUG_ON(ret);
 288 }
 289
 290 static void bind_to_memnode(int node)
 291 {
 292         unsigned long nodemask;
 293         int ret;
 294
 295         if (node == -1)
 296                 return;
 297
 298         BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 299         nodemask = 1L << node;
 300
 301         ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 302         dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 303
 304         BUG_ON(ret);
 305 }
 306
 307 #define HPSIZE (2*1024*1024)
 308
 309 #define set_taskname(fmt...)                            \
 310 do {                                                    \
 311         char name[20];                                  \
 312                                                         \
 313         snprintf(name, 20, fmt);                        \
 314         prctl(PR_SET_NAME, name);                       \
 315 } while (0)
 316
 317 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 318                       int init_zero, int init_cpu0, int thp, int init_random)
 319 {
 320         cpu_set_t orig_mask;
 321         ssize_t bytes;
 322         u8 *buf;
 323         int ret;
 324
 325         if (!bytes0)
 326                 return NULL;
 327
 328         /* Allocate and initialize all memory on CPU#0: */
 329         if (init_cpu0) {
 330                 orig_mask = bind_to_node(0);
 331                 bind_to_memnode(0);
 332         }
 333
 334         bytes = bytes0 + HPSIZE;
 335
 336         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 337         BUG_ON(buf == (void *)-1);
 338
 339         if (map_flags == MAP_PRIVATE) {
 340                 if (thp > 0) {
 341                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 342                         if (ret && !g->print_once) {
 343                                 g->print_once = 1;
 344                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 345                         }
 346                 }
 347                 if (thp < 0) {
 348                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 349                         if (ret && !g->print_once) {
 350                                 g->print_once = 1;
 351                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 352                         }
 353                 }
 354         }
 355
 356         if (init_zero) {
 357                 bzero(buf, bytes);
 358         } else {
 359                 /* Initialize random contents, different in each word: */
 360                 if (init_random) {
 361                         u64 *wbuf = (void *)buf;
 362                         long off = rand();
 363                         long i;
 364
 365                         for (i = 0; i < bytes/8; i++)
 366                                 wbuf[i] = i + off;
 367                 }
 368         }
 369
 370         /* Align to 2MB boundary: */
 371         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 372
 373         /* Restore affinity: */
 374         if (init_cpu0) {
 375                 bind_to_cpumask(orig_mask);
 376                 mempol_restore();
 377         }
 378
 379         return buf;
 380 }
 381
 382 static void free_data(void *data, ssize_t bytes)
 383 {
 384         int ret;
 385
 386         if (!data)
 387                 return;
 388
 389         ret = munmap(data, bytes);
 390         BUG_ON(ret);
 391 }
 392
 393 /*
 394  * Create a shared memory buffer that can be shared between processes, zeroed:
 395  */
 396 static void * zalloc_shared_data(ssize_t bytes)
 397 {
 398         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 399 }
 400
 401 /*
 402  * Create a shared memory buffer that can be shared between processes:
 403  */
 404 static void * setup_shared_data(ssize_t bytes)
 405 {
 406         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 407 }
 408
 409 /*
 410  * Allocate process-local memory - this will either be shared between
 411  * threads of this process, or only be accessed by this thread:
 412  */
 413 static void * setup_private_data(ssize_t bytes)
 414 {
 415         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 416 }
 417
 418 /*
 419  * Return a process-shared (global) mutex:
 420  */
 421 static void init_global_mutex(pthread_mutex_t *mutex)
 422 {
 423         pthread_mutexattr_t attr;
 424
 425         pthread_mutexattr_init(&attr);
 426         pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 427         pthread_mutex_init(mutex, &attr);
 428 }
 429
 430 static int parse_cpu_list(const char *arg)
 431 {
 432         p0.cpu_list_str = strdup(arg);
 433
 434         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 435
 436         return 0;
 437 }
 438
 439 static int parse_setup_cpu_list(void)
 440 {
 441         struct thread_data *td;
 442         char *str0, *str;
 443         int t;
 444
 445         if (!g->p.cpu_list_str)
 446                 return 0;
 447
 448         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 449
 450         str0 = str = strdup(g->p.cpu_list_str);
 451         t = 0;
 452
 453         BUG_ON(!str);
 454
 455         tprintf("# binding tasks to CPUs:\n");
 456         tprintf("#  ");
 457
 458         while (true) {
 459                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 460                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 461                 int bind_len;
 462                 int step;
 463                 int mul;
 464
 465                 tok = strsep(&str, ",");
 466                 if (!tok)
 467                         break;
 468
 469                 tok_end = strstr(tok, "-");
 470
 471                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 472                 if (!tok_end) {
 473                         /* Single CPU specified: */
 474                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 475                 } else {
 476                         /* CPU range specified (for example: "5-11"): */
 477                         bind_cpu_0 = atol(tok);
 478                         bind_cpu_1 = atol(tok_end + 1);
 479                 }
 480
 481                 step = 1;
 482                 tok_step = strstr(tok, "#");
 483                 if (tok_step) {
 484                         step = atol(tok_step + 1);
 485                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 486                 }
 487
 488                 /*
 489                  * Mask length.
 490                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 491                  * where the _4 means the next 4 CPUs are allowed.
 492                  */
 493                 bind_len = 1;
 494                 tok_len = strstr(tok, "_");
 495                 if (tok_len) {
 496                         bind_len = atol(tok_len + 1);
 497                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 498                 }
 499
 500                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 501                 mul = 1;
 502                 tok_mul = strstr(tok, "x");
 503                 if (tok_mul) {
 504                         mul = atol(tok_mul + 1);
 505                         BUG_ON(mul <= 0);
 506                 }
 507
 508                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 509
 510                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 511                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 512                         return -1;
 513                 }
 514
 515                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 516                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 517
 518                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 519                         int i;
 520
 521                         for (i = 0; i < mul; i++) {
 522                                 int cpu;
 523
 524                                 if (t >= g->p.nr_tasks) {
 525                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 526                                         goto out;
 527                                 }
 528                                 td = g->threads + t;
 529
 530                                 if (t)
 531                                         tprintf(",");
 532                                 if (bind_len > 1) {
 533                                         tprintf("%2d/%d", bind_cpu, bind_len);
 534                                 } else {
 535                                         tprintf("%2d", bind_cpu);
 536                                 }
 537
 538                                 CPU_ZERO(&td->bind_cpumask);
 539                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 540                                         BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 541                                         CPU_SET(cpu, &td->bind_cpumask);
 542                                 }
 543                                 t++;
 544                         }
 545                 }
 546         }
 547 out:
 548
 549         tprintf("\n");
 550
 551         if (t < g->p.nr_tasks)
 552                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 553
 554         free(str0);
 555         return 0;
 556 }
 557
 558 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 559                           const char *arg, int unset __maybe_unused)
 560 {
 561         if (!arg)
 562                 return -1;
 563
 564         return parse_cpu_list(arg);
 565 }
 566
 567 static int parse_node_list(const char *arg)
 568 {
 569         p0.node_list_str = strdup(arg);
 570
 571         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 572
 573         return 0;
 574 }
 575
 576 static int parse_setup_node_list(void)
 577 {
 578         struct thread_data *td;
 579         char *str0, *str;
 580         int t;
 581
 582         if (!g->p.node_list_str)
 583                 return 0;
 584
 585         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 586
 587         str0 = str = strdup(g->p.node_list_str);
 588         t = 0;
 589
 590         BUG_ON(!str);
 591
 592         tprintf("# binding tasks to NODEs:\n");
 593         tprintf("# ");
 594
 595         while (true) {
 596                 int bind_node, bind_node_0, bind_node_1;
 597                 char *tok, *tok_end, *tok_step, *tok_mul;
 598                 int step;
 599                 int mul;
 600
 601                 tok = strsep(&str, ",");
 602                 if (!tok)
 603                         break;
 604
 605                 tok_end = strstr(tok, "-");
 606
 607                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 608                 if (!tok_end) {
 609                         /* Single NODE specified: */
 610                         bind_node_0 = bind_node_1 = atol(tok);
 611                 } else {
 612                         /* NODE range specified (for example: "5-11"): */
 613                         bind_node_0 = atol(tok);
 614                         bind_node_1 = atol(tok_end + 1);
 615                 }
 616
 617                 step = 1;
 618                 tok_step = strstr(tok, "#");
 619                 if (tok_step) {
 620                         step = atol(tok_step + 1);
 621                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 622                 }
 623
 624                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 625                 mul = 1;
 626                 tok_mul = strstr(tok, "x");
 627                 if (tok_mul) {
 628                         mul = atol(tok_mul + 1);
 629                         BUG_ON(mul <= 0);
 630                 }
 631
 632                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 633
 634                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 635                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 636                         return -1;
 637                 }
 638
 639                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 640                 BUG_ON(bind_node_0 > bind_node_1);
 641
 642                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 643                         int i;
 644
 645                         for (i = 0; i < mul; i++) {
 646                                 if (t >= g->p.nr_tasks) {
 647                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 648                                         goto out;
 649                                 }
 650                                 td = g->threads + t;
 651
 652                                 if (!t)
 653                                         tprintf(" %2d", bind_node);
 654                                 else
 655                                         tprintf(",%2d", bind_node);
 656
 657                                 td->bind_node = bind_node;
 658                                 t++;
 659                         }
 660                 }
 661         }
 662 out:
 663
 664         tprintf("\n");
 665
 666         if (t < g->p.nr_tasks)
 667                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 668
 669         free(str0);
 670         return 0;
 671 }
 672
 673 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 674                           const char *arg, int unset __maybe_unused)
 675 {
 676         if (!arg)
 677                 return -1;
 678
 679         return parse_node_list(arg);
 680
 681         return 0;
 682 }
 683
 684 #define BIT(x) (1ul << x)
 685
 686 static inline uint32_t lfsr_32(uint32_t lfsr)
 687 {
 688         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 689         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 690 }
 691
 692 /*
 693  * Make sure there's real data dependency to RAM (when read
 694  * accesses are enabled), so the compiler, the CPU and the
 695  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 696  * accesses:
 697  */
 698 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 699 {
 700         if (g->p.data_reads)
 701                 val += *data;
 702         if (g->p.data_writes)
 703                 *data = val + 1;
 704         return val;
 705 }
 706
 707 /*
 708  * The worker process does two types of work, a forwards going
 709  * loop and a backwards going loop.
 710  *
 711  * We do this so that on multiprocessor systems we do not create
 712  * a 'train' of processing, with highly synchronized processes,
 713  * skewing the whole benchmark.
 714  */
 715 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 716 {
 717         long words = bytes/sizeof(u64);
 718         u64 *data = (void *)__data;
 719         long chunk_0, chunk_1;
 720         u64 *d0, *d, *d1;
 721         long off;
 722         long i;
 723
 724         BUG_ON(!data && words);
 725         BUG_ON(data && !words);
 726
 727         if (!data)
 728                 return val;
 729
 730         /* Very simple memset() work variant: */
 731         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 732                 bzero(data, bytes);
 733                 return val;
 734         }
 735
 736         /* Spread out by PID/TID nr and by loop nr: */
 737         chunk_0 = words/nr_max;
 738         chunk_1 = words/g->p.nr_loops;
 739         off = nr*chunk_0 + loop*chunk_1;
 740
 741         while (off >= words)
 742                 off -= words;
 743
 744         if (g->p.data_rand_walk) {
 745                 u32 lfsr = nr + loop + val;
 746                 int j;
 747
 748                 for (i = 0; i < words/1024; i++) {
 749                         long start, end;
 750
 751                         lfsr = lfsr_32(lfsr);
 752
 753                         start = lfsr % words;
 754                         end = min(start + 1024, words-1);
 755
 756                         if (g->p.data_zero_memset) {
 757                                 bzero(data + start, (end-start) * sizeof(u64));
 758                         } else {
 759                                 for (j = start; j < end; j++)
 760                                         val = access_data(data + j, val);
 761                         }
 762                 }
 763         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 764
 765                 d0 = data + off;
 766                 d  = data + off + 1;
 767                 d1 = data + words;
 768
 769                 /* Process data forwards: */
 770                 for (;;) {
 771                         if (unlikely(d >= d1))
 772                                 d = data;
 773                         if (unlikely(d == d0))
 774                                 break;
 775
 776                         val = access_data(d, val);
 777
 778                         d++;
 779                 }
 780         } else {
 781                 /* Process data backwards: */
 782
 783                 d0 = data + off;
 784                 d  = data + off - 1;
 785                 d1 = data + words;
 786
 787                 /* Process data forwards: */
 788                 for (;;) {
 789                         if (unlikely(d < data))
 790                                 d = data + words-1;
 791                         if (unlikely(d == d0))
 792                                 break;
 793
 794                         val = access_data(d, val);
 795
 796                         d--;
 797                 }
 798         }
 799
 800         return val;
 801 }
 802
 803 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 804 {
 805         unsigned int cpu;
 806
 807         cpu = sched_getcpu();
 808
 809         g->threads[task_nr].curr_cpu = cpu;
 810         prctl(0, bytes_worked);
 811 }
 812
 813 #define MAX_NR_NODES    64
 814
 815 /*
 816  * Count the number of nodes a process's threads
 817  * are spread out on.
 818  *
 819  * A count of 1 means that the process is compressed
 820  * to a single node. A count of g->p.nr_nodes means it's
 821  * spread out on the whole system.
 822  */
 823 static int count_process_nodes(int process_nr)
 824 {
 825         char node_present[MAX_NR_NODES] = { 0, };
 826         int nodes;
 827         int n, t;
 828
 829         for (t = 0; t < g->p.nr_threads; t++) {
 830                 struct thread_data *td;
 831                 int task_nr;
 832                 int node;
 833
 834                 task_nr = process_nr*g->p.nr_threads + t;
 835                 td = g->threads + task_nr;
 836
 837                 node = numa_node_of_cpu(td->curr_cpu);
 838                 if (node < 0) /* curr_cpu was likely still -1 */
 839                         return 0;
 840
 841                 node_present[node] = 1;
 842         }
 843
 844         nodes = 0;
 845
 846         for (n = 0; n < MAX_NR_NODES; n++)
 847                 nodes += node_present[n];
 848
 849         return nodes;
 850 }
 851
 852 /*
 853  * Count the number of distinct process-threads a node contains.
 854  *
 855  * A count of 1 means that the node contains only a single
 856  * process. If all nodes on the system contain at most one
 857  * process then we are well-converged.
 858  */
 859 static int count_node_processes(int node)
 860 {
 861         int processes = 0;
 862         int t, p;
 863
 864         for (p = 0; p < g->p.nr_proc; p++) {
 865                 for (t = 0; t < g->p.nr_threads; t++) {
 866                         struct thread_data *td;
 867                         int task_nr;
 868                         int n;
 869
 870                         task_nr = p*g->p.nr_threads + t;
 871                         td = g->threads + task_nr;
 872
 873                         n = numa_node_of_cpu(td->curr_cpu);
 874                         if (n == node) {
 875                                 processes++;
 876                                 break;
 877                         }
 878                 }
 879         }
 880
 881         return processes;
 882 }
 883
 884 static void calc_convergence_compression(int *strong)
 885 {
 886         unsigned int nodes_min, nodes_max;
 887         int p;
 888
 889         nodes_min = -1;
 890         nodes_max =  0;
 891
 892         for (p = 0; p < g->p.nr_proc; p++) {
 893                 unsigned int nodes = count_process_nodes(p);
 894
 895                 if (!nodes) {
 896                         *strong = 0;
 897                         return;
 898                 }
 899
 900                 nodes_min = min(nodes, nodes_min);
 901                 nodes_max = max(nodes, nodes_max);
 902         }
 903
 904         /* Strong convergence: all threads compress on a single node: */
 905         if (nodes_min == 1 && nodes_max == 1) {
 906                 *strong = 1;
 907         } else {
 908                 *strong = 0;
 909                 tprintf(" {%d-%d}", nodes_min, nodes_max);
 910         }
 911 }
 912
 913 static void calc_convergence(double runtime_ns_max, double *convergence)
 914 {
 915         unsigned int loops_done_min, loops_done_max;
 916         int process_groups;
 917         int nodes[MAX_NR_NODES];
 918         int distance;
 919         int nr_min;
 920         int nr_max;
 921         int strong;
 922         int sum;
 923         int nr;
 924         int node;
 925         int cpu;
 926         int t;
 927
 928         if (!g->p.show_convergence && !g->p.measure_convergence)
 929                 return;
 930
 931         for (node = 0; node < g->p.nr_nodes; node++)
 932                 nodes[node] = 0;
 933
 934         loops_done_min = -1;
 935         loops_done_max = 0;
 936
 937         for (t = 0; t < g->p.nr_tasks; t++) {
 938                 struct thread_data *td = g->threads + t;
 939                 unsigned int loops_done;
 940
 941                 cpu = td->curr_cpu;
 942
 943                 /* Not all threads have written it yet: */
 944                 if (cpu < 0)
 945                         continue;
 946
 947                 node = numa_node_of_cpu(cpu);
 948
 949                 nodes[node]++;
 950
 951                 loops_done = td->loops_done;
 952                 loops_done_min = min(loops_done, loops_done_min);
 953                 loops_done_max = max(loops_done, loops_done_max);
 954         }
 955
 956         nr_max = 0;
 957         nr_min = g->p.nr_tasks;
 958         sum = 0;
 959
 960         for (node = 0; node < g->p.nr_nodes; node++) {
 961                 nr = nodes[node];
 962                 nr_min = min(nr, nr_min);
 963                 nr_max = max(nr, nr_max);
 964                 sum += nr;
 965         }
 966         BUG_ON(nr_min > nr_max);
 967
 968         BUG_ON(sum > g->p.nr_tasks);
 969
 970         if (0 && (sum < g->p.nr_tasks))
 971                 return;
 972
 973         /*
 974          * Count the number of distinct process groups present
 975          * on nodes - when we are converged this will decrease
 976          * to g->p.nr_proc:
 977          */
 978         process_groups = 0;
 979
 980         for (node = 0; node < g->p.nr_nodes; node++) {
 981                 int processes = count_node_processes(node);
 982
 983                 nr = nodes[node];
 984                 tprintf(" %2d/%-2d", nr, processes);
 985
 986                 process_groups += processes;
 987         }
 988
 989         distance = nr_max - nr_min;
 990
 991         tprintf(" [%2d/%-2d]", distance, process_groups);
 992
 993         tprintf(" l:%3d-%-3d (%3d)",
 994                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 995
 996         if (loops_done_min && loops_done_max) {
 997                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
 998
 999                 tprintf(" [%4.1f%%]", skew * 100.0);
1000         }
1001
1002         calc_convergence_compression(&strong);
1003
1004         if (strong && process_groups == g->p.nr_proc) {
1005                 if (!*convergence) {
1006                         *convergence = runtime_ns_max;
1007                         tprintf(" (%6.1fs converged)\n", *convergence/1e9);
1008                         if (g->p.measure_convergence) {
1009                                 g->all_converged = true;
1010                                 g->stop_work = true;
1011                         }
1012                 }
1013         } else {
1014                 if (*convergence) {
1015                         tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
1016                         *convergence = 0;
1017                 }
1018                 tprintf("\n");
1019         }
1020 }
1021
1022 static void show_summary(double runtime_ns_max, int l, double *convergence)
1023 {
1024         tprintf("\r #  %5.1f%%  [%.1f mins]",
1025                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1026
1027         calc_convergence(runtime_ns_max, convergence);
1028
1029         if (g->p.show_details >= 0)
1030                 fflush(stdout);
1031 }
1032
1033 static void *worker_thread(void *__tdata)
1034 {
1035         struct thread_data *td = __tdata;
1036         struct timeval start0, start, stop, diff;
1037         int process_nr = td->process_nr;
1038         int thread_nr = td->thread_nr;
1039         unsigned long last_perturbance;
1040         int task_nr = td->task_nr;
1041         int details = g->p.show_details;
1042         int first_task, last_task;
1043         double convergence = 0;
1044         u64 val = td->val;
1045         double runtime_ns_max;
1046         u8 *global_data;
1047         u8 *process_data;
1048         u8 *thread_data;
1049         u64 bytes_done;
1050         long work_done;
1051         u32 l;
1052         struct rusage rusage;
1053
1054         bind_to_cpumask(td->bind_cpumask);
1055         bind_to_memnode(td->bind_node);
1056
1057         set_taskname("thread %d/%d", process_nr, thread_nr);
1058
1059         global_data = g->data;
1060         process_data = td->process_data;
1061         thread_data = setup_private_data(g->p.bytes_thread);
1062
1063         bytes_done = 0;
1064
1065         last_task = 0;
1066         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1067                 last_task = 1;
1068
1069         first_task = 0;
1070         if (process_nr == 0 && thread_nr == 0)
1071                 first_task = 1;
1072
1073         if (details >= 2) {
1074                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1075                         process_nr, thread_nr, global_data, process_data, thread_data);
1076         }
1077
1078         if (g->p.serialize_startup) {
1079                 pthread_mutex_lock(&g->startup_mutex);
1080                 g->nr_tasks_started++;
1081                 pthread_mutex_unlock(&g->startup_mutex);
1082
1083                 /* Here we will wait for the main process to start us all at once: */
1084                 pthread_mutex_lock(&g->start_work_mutex);
1085                 g->nr_tasks_working++;
1086
1087                 /* Last one wake the main process: */
1088                 if (g->nr_tasks_working == g->p.nr_tasks)
1089                         pthread_mutex_unlock(&g->startup_done_mutex);
1090
1091                 pthread_mutex_unlock(&g->start_work_mutex);
1092         }
1093
1094         gettimeofday(&start0, NULL);
1095
1096         start = stop = start0;
1097         last_perturbance = start.tv_sec;
1098
1099         for (l = 0; l < g->p.nr_loops; l++) {
1100                 start = stop;
1101
1102                 if (g->stop_work)
1103                         break;
1104
1105                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1106                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1107                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1108
1109                 if (g->p.sleep_usecs) {
1110                         pthread_mutex_lock(td->process_lock);
1111                         usleep(g->p.sleep_usecs);
1112                         pthread_mutex_unlock(td->process_lock);
1113                 }
1114                 /*
1115                  * Amount of work to be done under a process-global lock:
1116                  */
1117                 if (g->p.bytes_process_locked) {
1118                         pthread_mutex_lock(td->process_lock);
1119                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1120                         pthread_mutex_unlock(td->process_lock);
1121                 }
1122
1123                 work_done = g->p.bytes_global + g->p.bytes_process +
1124                             g->p.bytes_process_locked + g->p.bytes_thread;
1125
1126                 update_curr_cpu(task_nr, work_done);
1127                 bytes_done += work_done;
1128
1129                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1130                         continue;
1131
1132                 td->loops_done = l;
1133
1134                 gettimeofday(&stop, NULL);
1135
1136                 /* Check whether our max runtime timed out: */
1137                 if (g->p.nr_secs) {
1138                         timersub(&stop, &start0, &diff);
1139                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1140                                 g->stop_work = true;
1141                                 break;
1142                         }
1143                 }
1144
1145                 /* Update the summary at most once per second: */
1146                 if (start.tv_sec == stop.tv_sec)
1147                         continue;
1148
1149                 /*
1150                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1151                  * by migrating to CPU#0:
1152                  */
1153                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1154                         cpu_set_t orig_mask;
1155                         int target_cpu;
1156                         int this_cpu;
1157
1158                         last_perturbance = stop.tv_sec;
1159
1160                         /*
1161                          * Depending on where we are running, move into
1162                          * the other half of the system, to create some
1163                          * real disturbance:
1164                          */
1165                         this_cpu = g->threads[task_nr].curr_cpu;
1166                         if (this_cpu < g->p.nr_cpus/2)
1167                                 target_cpu = g->p.nr_cpus-1;
1168                         else
1169                                 target_cpu = 0;
1170
1171                         orig_mask = bind_to_cpu(target_cpu);
1172
1173                         /* Here we are running on the target CPU already */
1174                         if (details >= 1)
1175                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1176
1177                         bind_to_cpumask(orig_mask);
1178                 }
1179
1180                 if (details >= 3) {
1181                         timersub(&stop, &start, &diff);
1182                         runtime_ns_max = diff.tv_sec * 1000000000;
1183                         runtime_ns_max += diff.tv_usec * 1000;
1184
1185                         if (details >= 0) {
1186                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1187                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1188                         }
1189                         fflush(stdout);
1190                 }
1191                 if (!last_task)
1192                         continue;
1193
1194                 timersub(&stop, &start0, &diff);
1195                 runtime_ns_max = diff.tv_sec * 1000000000ULL;
1196                 runtime_ns_max += diff.tv_usec * 1000ULL;
1197
1198                 show_summary(runtime_ns_max, l, &convergence);
1199         }
1200
1201         gettimeofday(&stop, NULL);
1202         timersub(&stop, &start0, &diff);
1203         td->runtime_ns = diff.tv_sec * 1000000000ULL;
1204         td->runtime_ns += diff.tv_usec * 1000ULL;
1205         td->speed_gbs = bytes_done / (td->runtime_ns / 1e9) / 1e9;
1206
1207         getrusage(RUSAGE_THREAD, &rusage);
1208         td->system_time_ns = rusage.ru_stime.tv_sec * 1000000000ULL;
1209         td->system_time_ns += rusage.ru_stime.tv_usec * 1000ULL;
1210         td->user_time_ns = rusage.ru_utime.tv_sec * 1000000000ULL;
1211         td->user_time_ns += rusage.ru_utime.tv_usec * 1000ULL;
1212
1213         free_data(thread_data, g->p.bytes_thread);
1214
1215         pthread_mutex_lock(&g->stop_work_mutex);
1216         g->bytes_done += bytes_done;
1217         pthread_mutex_unlock(&g->stop_work_mutex);
1218
1219         return NULL;
1220 }
1221
1222 /*
1223  * A worker process starts a couple of threads:
1224  */
1225 static void worker_process(int process_nr)
1226 {
1227         pthread_mutex_t process_lock;
1228         struct thread_data *td;
1229         pthread_t *pthreads;
1230         u8 *process_data;
1231         int task_nr;
1232         int ret;
1233         int t;
1234
1235         pthread_mutex_init(&process_lock, NULL);
1236         set_taskname("process %d", process_nr);
1237
1238         /*
1239          * Pick up the memory policy and the CPU binding of our first thread,
1240          * so that we initialize memory accordingly:
1241          */
1242         task_nr = process_nr*g->p.nr_threads;
1243         td = g->threads + task_nr;
1244
1245         bind_to_memnode(td->bind_node);
1246         bind_to_cpumask(td->bind_cpumask);
1247
1248         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1249         process_data = setup_private_data(g->p.bytes_process);
1250
1251         if (g->p.show_details >= 3) {
1252                 printf(" # process %2d global mem: %p, process mem: %p\n",
1253                         process_nr, g->data, process_data);
1254         }
1255
1256         for (t = 0; t < g->p.nr_threads; t++) {
1257                 task_nr = process_nr*g->p.nr_threads + t;
1258                 td = g->threads + task_nr;
1259
1260                 td->process_data = process_data;
1261                 td->process_nr   = process_nr;
1262                 td->thread_nr    = t;
1263                 td->task_nr      = task_nr;
1264                 td->val          = rand();
1265                 td->curr_cpu     = -1;
1266                 td->process_lock = &process_lock;
1267
1268                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1269                 BUG_ON(ret);
1270         }
1271
1272         for (t = 0; t < g->p.nr_threads; t++) {
1273                 ret = pthread_join(pthreads[t], NULL);
1274                 BUG_ON(ret);
1275         }
1276
1277         free_data(process_data, g->p.bytes_process);
1278         free(pthreads);
1279 }
1280
1281 static void print_summary(void)
1282 {
1283         if (g->p.show_details < 0)
1284                 return;
1285
1286         printf("\n ###\n");
1287         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1288                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1289         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1290                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1291         printf(" #      %5dx %5ldMB process shared mem operations\n",
1292                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1293         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1294                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1295
1296         printf(" ###\n");
1297
1298         printf("\n ###\n"); fflush(stdout);
1299 }
1300
1301 static void init_thread_data(void)
1302 {
1303         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1304         int t;
1305
1306         g->threads = zalloc_shared_data(size);
1307
1308         for (t = 0; t < g->p.nr_tasks; t++) {
1309                 struct thread_data *td = g->threads + t;
1310                 int cpu;
1311
1312                 /* Allow all nodes by default: */
1313                 td->bind_node = -1;
1314
1315                 /* Allow all CPUs by default: */
1316                 CPU_ZERO(&td->bind_cpumask);
1317                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1318                         CPU_SET(cpu, &td->bind_cpumask);
1319         }
1320 }
1321
1322 static void deinit_thread_data(void)
1323 {
1324         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1325
1326         free_data(g->threads, size);
1327 }
1328
1329 static int init(void)
1330 {
1331         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1332
1333         /* Copy over options: */
1334         g->p = p0;
1335
1336         g->p.nr_cpus = numa_num_configured_cpus();
1337
1338         g->p.nr_nodes = numa_max_node() + 1;
1339
1340         /* char array in count_process_nodes(): */
1341         BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1342
1343         if (g->p.show_quiet && !g->p.show_details)
1344                 g->p.show_details = -1;
1345
1346         /* Some memory should be specified: */
1347         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1348                 return -1;
1349
1350         if (g->p.mb_global_str) {
1351                 g->p.mb_global = atof(g->p.mb_global_str);
1352                 BUG_ON(g->p.mb_global < 0);
1353         }
1354
1355         if (g->p.mb_proc_str) {
1356                 g->p.mb_proc = atof(g->p.mb_proc_str);
1357                 BUG_ON(g->p.mb_proc < 0);
1358         }
1359
1360         if (g->p.mb_proc_locked_str) {
1361                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1362                 BUG_ON(g->p.mb_proc_locked < 0);
1363                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1364         }
1365
1366         if (g->p.mb_thread_str) {
1367                 g->p.mb_thread = atof(g->p.mb_thread_str);
1368                 BUG_ON(g->p.mb_thread < 0);
1369         }
1370
1371         BUG_ON(g->p.nr_threads <= 0);
1372         BUG_ON(g->p.nr_proc <= 0);
1373
1374         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1375
1376         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1377         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1378         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1379         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1380
1381         g->data = setup_shared_data(g->p.bytes_global);
1382
1383         /* Startup serialization: */
1384         init_global_mutex(&g->start_work_mutex);
1385         init_global_mutex(&g->startup_mutex);
1386         init_global_mutex(&g->startup_done_mutex);
1387         init_global_mutex(&g->stop_work_mutex);
1388
1389         init_thread_data();
1390
1391         tprintf("#\n");
1392         if (parse_setup_cpu_list() || parse_setup_node_list())
1393                 return -1;
1394         tprintf("#\n");
1395
1396         print_summary();
1397
1398         return 0;
1399 }
1400
1401 static void deinit(void)
1402 {
1403         free_data(g->data, g->p.bytes_global);
1404         g->data = NULL;
1405
1406         deinit_thread_data();
1407
1408         free_data(g, sizeof(*g));
1409         g = NULL;
1410 }
1411
1412 /*
1413  * Print a short or long result, depending on the verbosity setting:
1414  */
1415 static void print_res(const char *name, double val,
1416                       const char *txt_unit, const char *txt_short, const char *txt_long)
1417 {
1418         if (!name)
1419                 name = "main,";
1420
1421         if (!g->p.show_quiet)
1422                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1423         else
1424                 printf(" %14.3f %s\n", val, txt_long);
1425 }
1426
1427 static int __bench_numa(const char *name)
1428 {
1429         struct timeval start, stop, diff;
1430         u64 runtime_ns_min, runtime_ns_sum;
1431         pid_t *pids, pid, wpid;
1432         double delta_runtime;
1433         double runtime_avg;
1434         double runtime_sec_max;
1435         double runtime_sec_min;
1436         int wait_stat;
1437         double bytes;
1438         int i, t, p;
1439
1440         if (init())
1441                 return -1;
1442
1443         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1444         pid = -1;
1445
1446         /* All threads try to acquire it, this way we can wait for them to start up: */
1447         pthread_mutex_lock(&g->start_work_mutex);
1448
1449         if (g->p.serialize_startup) {
1450                 tprintf(" #\n");
1451                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1452         }
1453
1454         gettimeofday(&start, NULL);
1455
1456         for (i = 0; i < g->p.nr_proc; i++) {
1457                 pid = fork();
1458                 dprintf(" # process %2d: PID %d\n", i, pid);
1459
1460                 BUG_ON(pid < 0);
1461                 if (!pid) {
1462                         /* Child process: */
1463                         worker_process(i);
1464
1465                         exit(0);
1466                 }
1467                 pids[i] = pid;
1468
1469         }
1470         /* Wait for all the threads to start up: */
1471         while (g->nr_tasks_started != g->p.nr_tasks)
1472                 usleep(1000);
1473
1474         BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1475
1476         if (g->p.serialize_startup) {
1477                 double startup_sec;
1478
1479                 pthread_mutex_lock(&g->startup_done_mutex);
1480
1481                 /* This will start all threads: */
1482                 pthread_mutex_unlock(&g->start_work_mutex);
1483
1484                 /* This mutex is locked - the last started thread will wake us: */
1485                 pthread_mutex_lock(&g->startup_done_mutex);
1486
1487                 gettimeofday(&stop, NULL);
1488
1489                 timersub(&stop, &start, &diff);
1490
1491                 startup_sec = diff.tv_sec * 1000000000.0;
1492                 startup_sec += diff.tv_usec * 1000.0;
1493                 startup_sec /= 1e9;
1494
1495                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1496                 tprintf(" #\n");
1497
1498                 start = stop;
1499                 pthread_mutex_unlock(&g->startup_done_mutex);
1500         } else {
1501                 gettimeofday(&start, NULL);
1502         }
1503
1504         /* Parent process: */
1505
1506
1507         for (i = 0; i < g->p.nr_proc; i++) {
1508                 wpid = waitpid(pids[i], &wait_stat, 0);
1509                 BUG_ON(wpid < 0);
1510                 BUG_ON(!WIFEXITED(wait_stat));
1511
1512         }
1513
1514         runtime_ns_sum = 0;
1515         runtime_ns_min = -1LL;
1516
1517         for (t = 0; t < g->p.nr_tasks; t++) {
1518                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1519
1520                 runtime_ns_sum += thread_runtime_ns;
1521                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1522         }
1523
1524         gettimeofday(&stop, NULL);
1525         timersub(&stop, &start, &diff);
1526
1527         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1528
1529         tprintf("\n ###\n");
1530         tprintf("\n");
1531
1532         runtime_sec_max = diff.tv_sec * 1000000000.0;
1533         runtime_sec_max += diff.tv_usec * 1000.0;
1534         runtime_sec_max /= 1e9;
1535
1536         runtime_sec_min = runtime_ns_min/1e9;
1537
1538         bytes = g->bytes_done;
1539         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1540
1541         if (g->p.measure_convergence) {
1542                 print_res(name, runtime_sec_max,
1543                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1544         }
1545
1546         print_res(name, runtime_sec_max,
1547                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1548
1549         print_res(name, runtime_sec_min,
1550                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1551
1552         print_res(name, runtime_avg,
1553                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1554
1555         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1556         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1557                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1558
1559         print_res(name, bytes / g->p.nr_tasks / 1e9,
1560                 "GB,", "data/thread",           "GB data processed, per thread");
1561
1562         print_res(name, bytes / 1e9,
1563                 "GB,", "data-total",            "GB data processed, total");
1564
1565         print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1566                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1567
1568         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1569                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1570
1571         print_res(name, bytes / runtime_sec_max / 1e9,
1572                 "GB/sec,", "total-speed",       "GB/sec total speed");
1573
1574         if (g->p.show_details >= 2) {
1575                 char tname[32];
1576                 struct thread_data *td;
1577                 for (p = 0; p < g->p.nr_proc; p++) {
1578                         for (t = 0; t < g->p.nr_threads; t++) {
1579                                 memset(tname, 0, 32);
1580                                 td = g->threads + p*g->p.nr_threads + t;
1581                                 snprintf(tname, 32, "process%d:thread%d", p, t);
1582                                 print_res(tname, td->speed_gbs,
1583                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1584                                 print_res(tname, td->system_time_ns / 1e9,
1585                                         "secs", "thread-system-time", "system CPU time/thread");
1586                                 print_res(tname, td->user_time_ns / 1e9,
1587                                         "secs", "thread-user-time", "user CPU time/thread");
1588                         }
1589                 }
1590         }
1591
1592         free(pids);
1593
1594         deinit();
1595
1596         return 0;
1597 }
1598
1599 #define MAX_ARGS 50
1600
1601 static int command_size(const char **argv)
1602 {
1603         int size = 0;
1604
1605         while (*argv) {
1606                 size++;
1607                 argv++;
1608         }
1609
1610         BUG_ON(size >= MAX_ARGS);
1611
1612         return size;
1613 }
1614
1615 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1616 {
1617         int i;
1618
1619         printf("\n # Running %s \"perf bench numa", name);
1620
1621         for (i = 0; i < argc; i++)
1622                 printf(" %s", argv[i]);
1623
1624         printf("\"\n");
1625
1626         memset(p, 0, sizeof(*p));
1627
1628         /* Initialize nonzero defaults: */
1629
1630         p->serialize_startup            = 1;
1631         p->data_reads                   = true;
1632         p->data_writes                  = true;
1633         p->data_backwards               = true;
1634         p->data_rand_walk               = true;
1635         p->nr_loops                     = -1;
1636         p->init_random                  = true;
1637         p->mb_global_str                = "1";
1638         p->nr_proc                      = 1;
1639         p->nr_threads                   = 1;
1640         p->nr_secs                      = 5;
1641         p->run_all                      = argc == 1;
1642 }
1643
1644 static int run_bench_numa(const char *name, const char **argv)
1645 {
1646         int argc = command_size(argv);
1647
1648         init_params(&p0, name, argc, argv);
1649         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1650         if (argc)
1651                 goto err;
1652
1653         if (__bench_numa(name))
1654                 goto err;
1655
1656         return 0;
1657
1658 err:
1659         return -1;
1660 }
1661
1662 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1663 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1664
1665 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1666 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1667
1668 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1669 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1670
1671 /*
1672  * The built-in test-suite executed by "perf bench numa -a".
1673  *
1674  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1675  */
1676 static const char *tests[][MAX_ARGS] = {
1677    /* Basic single-stream NUMA bandwidth measurements: */
1678    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1679                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1680    { "RAM-bw-local-NOTHP,",
1681                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1682                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1683    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1684                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1685
1686    /* 2-stream NUMA bandwidth measurements: */
1687    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1688                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1689    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1690                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1691
1692    /* Cross-stream NUMA bandwidth measurement: */
1693    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1694                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1695
1696    /* Convergence latency measurements: */
1697    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1698    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1699    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1700    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1701    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1702    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1703    { " 4x4-convergence-NOTHP,",
1704                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1705    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1706    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1707    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1708    { " 8x4-convergence-NOTHP,",
1709                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1710    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1711    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1712    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1713    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1714    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1715
1716    /* Various NUMA process/thread layout bandwidth measurements: */
1717    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1718    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1719    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1720    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1721    { " 8x1-bw-process-NOTHP,",
1722                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1723    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1724
1725    { " 4x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1726    { " 8x1-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1727    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1728    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1729
1730    { " 2x3-bw-thread,",   "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1731    { " 4x4-bw-thread,",   "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1732    { " 4x6-bw-thread,",   "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1733    { " 4x8-bw-thread,",   "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1734    { " 4x8-bw-thread-NOTHP,",
1735                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1736    { " 3x3-bw-thread,",   "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1737    { " 5x5-bw-thread,",   "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1738
1739    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1740    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1741
1742    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1743    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1744    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1745    { "numa01-bw-thread-NOTHP,",
1746                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1747 };
1748
1749 static int bench_all(void)
1750 {
1751         int nr = ARRAY_SIZE(tests);
1752         int ret;
1753         int i;
1754
1755         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1756         BUG_ON(ret < 0);
1757
1758         for (i = 0; i < nr; i++) {
1759                 run_bench_numa(tests[i][0], tests[i] + 1);
1760         }
1761
1762         printf("\n");
1763
1764         return 0;
1765 }
1766
1767 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1768 {
1769         init_params(&p0, "main,", argc, argv);
1770         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1771         if (argc)
1772                 goto err;
1773
1774         if (p0.run_all)
1775                 return bench_all();
1776
1777         if (__bench_numa(NULL))
1778                 goto err;
1779
1780         return 0;
1781
1782 err:
1783         usage_with_options(numa_usage, options);
1784         return -1;
1785 }