tools/perf/bench/numa.c

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