tools/perf/bench/numa.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * numa.c
   4  *
   5  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   6  */
   7
   8 #include <inttypes.h>
   9
  10 #include <subcmd/parse-options.h>
  11 #include "../util/cloexec.h"
  12
  13 #include "bench.h"
  14
  15 #include <errno.h>
  16 #include <sched.h>
  17 #include <stdio.h>
  18 #include <assert.h>
  19 #include <malloc.h>
  20 #include <signal.h>
  21 #include <stdlib.h>
  22 #include <string.h>
  23 #include <unistd.h>
  24 #include <sys/mman.h>
  25 #include <sys/time.h>
  26 #include <sys/resource.h>
  27 #include <sys/wait.h>
  28 #include <sys/prctl.h>
  29 #include <sys/types.h>
  30 #include <linux/kernel.h>
  31 #include <linux/time64.h>
  32 #include <linux/numa.h>
  33 #include <linux/zalloc.h>
  34
  35 #include "../util/header.h"
  36 #include "../util/mutex.h"
  37 #include <numa.h>
  38 #include <numaif.h>
  39
  40 #ifndef RUSAGE_THREAD
  41 # define RUSAGE_THREAD 1
  42 #endif
  43
  44 /*
  45  * Regular printout to the terminal, suppressed if -q is specified:
  46  */
  47 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  48
  49 /*
  50  * Debug printf:
  51  */
  52 #undef dprintf
  53 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  54
  55 struct thread_data {
  56         int                     curr_cpu;
  57         cpu_set_t               *bind_cpumask;
  58         int                     bind_node;
  59         u8                      *process_data;
  60         int                     process_nr;
  61         int                     thread_nr;
  62         int                     task_nr;
  63         unsigned int            loops_done;
  64         u64                     val;
  65         u64                     runtime_ns;
  66         u64                     system_time_ns;
  67         u64                     user_time_ns;
  68         double                  speed_gbs;
  69         struct mutex            *process_lock;
  70 };
  71
  72 /* Parameters set by options: */
  73
  74 struct params {
  75         /* Startup synchronization: */
  76         bool                    serialize_startup;
  77
  78         /* Task hierarchy: */
  79         int                     nr_proc;
  80         int                     nr_threads;
  81
  82         /* Working set sizes: */
  83         const char              *mb_global_str;
  84         const char              *mb_proc_str;
  85         const char              *mb_proc_locked_str;
  86         const char              *mb_thread_str;
  87
  88         double                  mb_global;
  89         double                  mb_proc;
  90         double                  mb_proc_locked;
  91         double                  mb_thread;
  92
  93         /* Access patterns to the working set: */
  94         bool                    data_reads;
  95         bool                    data_writes;
  96         bool                    data_backwards;
  97         bool                    data_zero_memset;
  98         bool                    data_rand_walk;
  99         u32                     nr_loops;
 100         u32                     nr_secs;
 101         u32                     sleep_usecs;
 102
 103         /* Working set initialization: */
 104         bool                    init_zero;
 105         bool                    init_random;
 106         bool                    init_cpu0;
 107
 108         /* Misc options: */
 109         int                     show_details;
 110         int                     run_all;
 111         int                     thp;
 112
 113         long                    bytes_global;
 114         long                    bytes_process;
 115         long                    bytes_process_locked;
 116         long                    bytes_thread;
 117
 118         int                     nr_tasks;
 119         bool                    show_quiet;
 120
 121         bool                    show_convergence;
 122         bool                    measure_convergence;
 123
 124         int                     perturb_secs;
 125         int                     nr_cpus;
 126         int                     nr_nodes;
 127
 128         /* Affinity options -C and -N: */
 129         char                    *cpu_list_str;
 130         char                    *node_list_str;
 131 };
 132
 133
 134 /* Global, read-writable area, accessible to all processes and threads: */
 135
 136 struct global_info {
 137         u8                      *data;
 138
 139         struct mutex            startup_mutex;
 140         struct cond             startup_cond;
 141         int                     nr_tasks_started;
 142
 143         struct mutex            start_work_mutex;
 144         struct cond             start_work_cond;
 145         int                     nr_tasks_working;
 146         bool                    start_work;
 147
 148         struct mutex            stop_work_mutex;
 149         u64                     bytes_done;
 150
 151         struct thread_data      *threads;
 152
 153         /* Convergence latency measurement: */
 154         bool                    all_converged;
 155         bool                    stop_work;
 156
 157         int                     print_once;
 158
 159         struct params           p;
 160 };
 161
 162 static struct global_info       *g = NULL;
 163
 164 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 165 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 166
 167 struct params p0;
 168
 169 static const struct option options[] = {
 170         OPT_INTEGER('p', "nr_proc"      , &p0.nr_proc,          "number of processes"),
 171         OPT_INTEGER('t', "nr_threads"   , &p0.nr_threads,       "number of threads per process"),
 172
 173         OPT_STRING('G', "mb_global"     , &p0.mb_global_str,    "MB", "global  memory (MBs)"),
 174         OPT_STRING('P', "mb_proc"       , &p0.mb_proc_str,      "MB", "process memory (MBs)"),
 175         OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 176         OPT_STRING('T', "mb_thread"     , &p0.mb_thread_str,    "MB", "thread  memory (MBs)"),
 177
 178         OPT_UINTEGER('l', "nr_loops"    , &p0.nr_loops,         "max number of loops to run (default: unlimited)"),
 179         OPT_UINTEGER('s', "nr_secs"     , &p0.nr_secs,          "max number of seconds to run (default: 5 secs)"),
 180         OPT_UINTEGER('u', "usleep"      , &p0.sleep_usecs,      "usecs to sleep per loop iteration"),
 181
 182         OPT_BOOLEAN('R', "data_reads"   , &p0.data_reads,       "access the data via reads (can be mixed with -W)"),
 183         OPT_BOOLEAN('W', "data_writes"  , &p0.data_writes,      "access the data via writes (can be mixed with -R)"),
 184         OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,  "access the data backwards as well"),
 185         OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 186         OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,  "access the data with random (32bit LFSR) walk"),
 187
 188
 189         OPT_BOOLEAN('z', "init_zero"    , &p0.init_zero,        "bzero the initial allocations"),
 190         OPT_BOOLEAN('I', "init_random"  , &p0.init_random,      "randomize the contents of the initial allocations"),
 191         OPT_BOOLEAN('0', "init_cpu0"    , &p0.init_cpu0,        "do the initial allocations on CPU#0"),
 192         OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,      "perturb thread 0/0 every X secs, to test convergence stability"),
 193
 194         OPT_INCR   ('d', "show_details" , &p0.show_details,     "Show details"),
 195         OPT_INCR   ('a', "all"          , &p0.run_all,          "Run all tests in the suite"),
 196         OPT_INTEGER('H', "thp"          , &p0.thp,              "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 197         OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
 198                     "convergence is reached when each process (all its threads) is running on a single NUMA node."),
 199         OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
 200         OPT_BOOLEAN('q', "quiet"        , &p0.show_quiet,       "quiet mode"),
 201         OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 202
 203         /* Special option string parsing callbacks: */
 204         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 205                         "bind the first N tasks to these specific cpus (the rest is unbound)",
 206                         parse_cpus_opt),
 207         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 208                         "bind the first N tasks to these specific memory nodes (the rest is unbound)",
 209                         parse_nodes_opt),
 210         OPT_END()
 211 };
 212
 213 static const char * const bench_numa_usage[] = {
 214         "perf bench numa <options>",
 215         NULL
 216 };
 217
 218 static const char * const numa_usage[] = {
 219         "perf bench numa mem [<options>]",
 220         NULL
 221 };
 222
 223 /*
 224  * To get number of numa nodes present.
 225  */
 226 static int nr_numa_nodes(void)
 227 {
 228         int i, nr_nodes = 0;
 229
 230         for (i = 0; i < g->p.nr_nodes; i++) {
 231                 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
 232                         nr_nodes++;
 233         }
 234
 235         return nr_nodes;
 236 }
 237
 238 /*
 239  * To check if given numa node is present.
 240  */
 241 static int is_node_present(int node)
 242 {
 243         return numa_bitmask_isbitset(numa_nodes_ptr, node);
 244 }
 245
 246 /*
 247  * To check given numa node has cpus.
 248  */
 249 static bool node_has_cpus(int node)
 250 {
 251         struct bitmask *cpumask = numa_allocate_cpumask();
 252         bool ret = false; /* fall back to nocpus */
 253         int cpu;
 254
 255         BUG_ON(!cpumask);
 256         if (!numa_node_to_cpus(node, cpumask)) {
 257                 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 258                         if (numa_bitmask_isbitset(cpumask, cpu)) {
 259                                 ret = true;
 260                                 break;
 261                         }
 262                 }
 263         }
 264         numa_free_cpumask(cpumask);
 265
 266         return ret;
 267 }
 268
 269 static cpu_set_t *bind_to_cpu(int target_cpu)
 270 {
 271         int nrcpus = numa_num_possible_cpus();
 272         cpu_set_t *orig_mask, *mask;
 273         size_t size;
 274
 275         orig_mask = CPU_ALLOC(nrcpus);
 276         BUG_ON(!orig_mask);
 277         size = CPU_ALLOC_SIZE(nrcpus);
 278         CPU_ZERO_S(size, orig_mask);
 279
 280         if (sched_getaffinity(0, size, orig_mask))
 281                 goto err_out;
 282
 283         mask = CPU_ALLOC(nrcpus);
 284         if (!mask)
 285                 goto err_out;
 286
 287         CPU_ZERO_S(size, mask);
 288
 289         if (target_cpu == -1) {
 290                 int cpu;
 291
 292                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 293                         CPU_SET_S(cpu, size, mask);
 294         } else {
 295                 if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
 296                         goto err;
 297
 298                 CPU_SET_S(target_cpu, size, mask);
 299         }
 300
 301         if (sched_setaffinity(0, size, mask))
 302                 goto err;
 303
 304         return orig_mask;
 305
 306 err:
 307         CPU_FREE(mask);
 308 err_out:
 309         CPU_FREE(orig_mask);
 310
 311         /* BUG_ON due to failure in allocation of orig_mask/mask */
 312         BUG_ON(-1);
 313         return NULL;
 314 }
 315
 316 static cpu_set_t *bind_to_node(int target_node)
 317 {
 318         int nrcpus = numa_num_possible_cpus();
 319         size_t size;
 320         cpu_set_t *orig_mask, *mask;
 321         int cpu;
 322
 323         orig_mask = CPU_ALLOC(nrcpus);
 324         BUG_ON(!orig_mask);
 325         size = CPU_ALLOC_SIZE(nrcpus);
 326         CPU_ZERO_S(size, orig_mask);
 327
 328         if (sched_getaffinity(0, size, orig_mask))
 329                 goto err_out;
 330
 331         mask = CPU_ALLOC(nrcpus);
 332         if (!mask)
 333                 goto err_out;
 334
 335         CPU_ZERO_S(size, mask);
 336
 337         if (target_node == NUMA_NO_NODE) {
 338                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 339                         CPU_SET_S(cpu, size, mask);
 340         } else {
 341                 struct bitmask *cpumask = numa_allocate_cpumask();
 342
 343                 if (!cpumask)
 344                         goto err;
 345
 346                 if (!numa_node_to_cpus(target_node, cpumask)) {
 347                         for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 348                                 if (numa_bitmask_isbitset(cpumask, cpu))
 349                                         CPU_SET_S(cpu, size, mask);
 350                         }
 351                 }
 352                 numa_free_cpumask(cpumask);
 353         }
 354
 355         if (sched_setaffinity(0, size, mask))
 356                 goto err;
 357
 358         return orig_mask;
 359
 360 err:
 361         CPU_FREE(mask);
 362 err_out:
 363         CPU_FREE(orig_mask);
 364
 365         /* BUG_ON due to failure in allocation of orig_mask/mask */
 366         BUG_ON(-1);
 367         return NULL;
 368 }
 369
 370 static void bind_to_cpumask(cpu_set_t *mask)
 371 {
 372         int ret;
 373         size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
 374
 375         ret = sched_setaffinity(0, size, mask);
 376         if (ret) {
 377                 CPU_FREE(mask);
 378                 BUG_ON(ret);
 379         }
 380 }
 381
 382 static void mempol_restore(void)
 383 {
 384         int ret;
 385
 386         ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 387
 388         BUG_ON(ret);
 389 }
 390
 391 static void bind_to_memnode(int node)
 392 {
 393         struct bitmask *node_mask;
 394         int ret;
 395
 396         if (node == NUMA_NO_NODE)
 397                 return;
 398
 399         node_mask = numa_allocate_nodemask();
 400         BUG_ON(!node_mask);
 401
 402         numa_bitmask_clearall(node_mask);
 403         numa_bitmask_setbit(node_mask, node);
 404
 405         ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
 406         dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
 407
 408         numa_bitmask_free(node_mask);
 409         BUG_ON(ret);
 410 }
 411
 412 #define HPSIZE (2*1024*1024)
 413
 414 #define set_taskname(fmt...)                            \
 415 do {                                                    \
 416         char name[20];                                  \
 417                                                         \
 418         snprintf(name, 20, fmt);                        \
 419         prctl(PR_SET_NAME, name);                       \
 420 } while (0)
 421
 422 static u8 *alloc_data(ssize_t bytes0, int map_flags,
 423                       int init_zero, int init_cpu0, int thp, int init_random)
 424 {
 425         cpu_set_t *orig_mask = NULL;
 426         ssize_t bytes;
 427         u8 *buf;
 428         int ret;
 429
 430         if (!bytes0)
 431                 return NULL;
 432
 433         /* Allocate and initialize all memory on CPU#0: */
 434         if (init_cpu0) {
 435                 int node = numa_node_of_cpu(0);
 436
 437                 orig_mask = bind_to_node(node);
 438                 bind_to_memnode(node);
 439         }
 440
 441         bytes = bytes0 + HPSIZE;
 442
 443         buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 444         BUG_ON(buf == (void *)-1);
 445
 446         if (map_flags == MAP_PRIVATE) {
 447                 if (thp > 0) {
 448                         ret = madvise(buf, bytes, MADV_HUGEPAGE);
 449                         if (ret && !g->print_once) {
 450                                 g->print_once = 1;
 451                                 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 452                         }
 453                 }
 454                 if (thp < 0) {
 455                         ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 456                         if (ret && !g->print_once) {
 457                                 g->print_once = 1;
 458                                 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 459                         }
 460                 }
 461         }
 462
 463         if (init_zero) {
 464                 bzero(buf, bytes);
 465         } else {
 466                 /* Initialize random contents, different in each word: */
 467                 if (init_random) {
 468                         u64 *wbuf = (void *)buf;
 469                         long off = rand();
 470                         long i;
 471
 472                         for (i = 0; i < bytes/8; i++)
 473                                 wbuf[i] = i + off;
 474                 }
 475         }
 476
 477         /* Align to 2MB boundary: */
 478         buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 479
 480         /* Restore affinity: */
 481         if (init_cpu0) {
 482                 bind_to_cpumask(orig_mask);
 483                 CPU_FREE(orig_mask);
 484                 mempol_restore();
 485         }
 486
 487         return buf;
 488 }
 489
 490 static void free_data(void *data, ssize_t bytes)
 491 {
 492         int ret;
 493
 494         if (!data)
 495                 return;
 496
 497         ret = munmap(data, bytes);
 498         BUG_ON(ret);
 499 }
 500
 501 /*
 502  * Create a shared memory buffer that can be shared between processes, zeroed:
 503  */
 504 static void * zalloc_shared_data(ssize_t bytes)
 505 {
 506         return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 507 }
 508
 509 /*
 510  * Create a shared memory buffer that can be shared between processes:
 511  */
 512 static void * setup_shared_data(ssize_t bytes)
 513 {
 514         return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 515 }
 516
 517 /*
 518  * Allocate process-local memory - this will either be shared between
 519  * threads of this process, or only be accessed by this thread:
 520  */
 521 static void * setup_private_data(ssize_t bytes)
 522 {
 523         return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 524 }
 525
 526 static int parse_cpu_list(const char *arg)
 527 {
 528         p0.cpu_list_str = strdup(arg);
 529
 530         dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 531
 532         return 0;
 533 }
 534
 535 static int parse_setup_cpu_list(void)
 536 {
 537         struct thread_data *td;
 538         char *str0, *str;
 539         int t;
 540
 541         if (!g->p.cpu_list_str)
 542                 return 0;
 543
 544         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 545
 546         str0 = str = strdup(g->p.cpu_list_str);
 547         t = 0;
 548
 549         BUG_ON(!str);
 550
 551         tprintf("# binding tasks to CPUs:\n");
 552         tprintf("#  ");
 553
 554         while (true) {
 555                 int bind_cpu, bind_cpu_0, bind_cpu_1;
 556                 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 557                 int bind_len;
 558                 int step;
 559                 int mul;
 560
 561                 tok = strsep(&str, ",");
 562                 if (!tok)
 563                         break;
 564
 565                 tok_end = strstr(tok, "-");
 566
 567                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 568                 if (!tok_end) {
 569                         /* Single CPU specified: */
 570                         bind_cpu_0 = bind_cpu_1 = atol(tok);
 571                 } else {
 572                         /* CPU range specified (for example: "5-11"): */
 573                         bind_cpu_0 = atol(tok);
 574                         bind_cpu_1 = atol(tok_end + 1);
 575                 }
 576
 577                 step = 1;
 578                 tok_step = strstr(tok, "#");
 579                 if (tok_step) {
 580                         step = atol(tok_step + 1);
 581                         BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 582                 }
 583
 584                 /*
 585                  * Mask length.
 586                  * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 587                  * where the _4 means the next 4 CPUs are allowed.
 588                  */
 589                 bind_len = 1;
 590                 tok_len = strstr(tok, "_");
 591                 if (tok_len) {
 592                         bind_len = atol(tok_len + 1);
 593                         BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 594                 }
 595
 596                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 597                 mul = 1;
 598                 tok_mul = strstr(tok, "x");
 599                 if (tok_mul) {
 600                         mul = atol(tok_mul + 1);
 601                         BUG_ON(mul <= 0);
 602                 }
 603
 604                 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 605
 606                 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 607                         printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 608                         return -1;
 609                 }
 610
 611                 if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
 612                         printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
 613                         return -1;
 614                 }
 615
 616                 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 617                 BUG_ON(bind_cpu_0 > bind_cpu_1);
 618
 619                 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 620                         size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
 621                         int i;
 622
 623                         for (i = 0; i < mul; i++) {
 624                                 int cpu;
 625
 626                                 if (t >= g->p.nr_tasks) {
 627                                         printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 628                                         goto out;
 629                                 }
 630                                 td = g->threads + t;
 631
 632                                 if (t)
 633                                         tprintf(",");
 634                                 if (bind_len > 1) {
 635                                         tprintf("%2d/%d", bind_cpu, bind_len);
 636                                 } else {
 637                                         tprintf("%2d", bind_cpu);
 638                                 }
 639
 640                                 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
 641                                 BUG_ON(!td->bind_cpumask);
 642                                 CPU_ZERO_S(size, td->bind_cpumask);
 643                                 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 644                                         if (cpu < 0 || cpu >= g->p.nr_cpus) {
 645                                                 CPU_FREE(td->bind_cpumask);
 646                                                 BUG_ON(-1);
 647                                         }
 648                                         CPU_SET_S(cpu, size, td->bind_cpumask);
 649                                 }
 650                                 t++;
 651                         }
 652                 }
 653         }
 654 out:
 655
 656         tprintf("\n");
 657
 658         if (t < g->p.nr_tasks)
 659                 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 660
 661         free(str0);
 662         return 0;
 663 }
 664
 665 static int parse_cpus_opt(const struct option *opt __maybe_unused,
 666                           const char *arg, int unset __maybe_unused)
 667 {
 668         if (!arg)
 669                 return -1;
 670
 671         return parse_cpu_list(arg);
 672 }
 673
 674 static int parse_node_list(const char *arg)
 675 {
 676         p0.node_list_str = strdup(arg);
 677
 678         dprintf("got NODE list: {%s}\n", p0.node_list_str);
 679
 680         return 0;
 681 }
 682
 683 static int parse_setup_node_list(void)
 684 {
 685         struct thread_data *td;
 686         char *str0, *str;
 687         int t;
 688
 689         if (!g->p.node_list_str)
 690                 return 0;
 691
 692         dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 693
 694         str0 = str = strdup(g->p.node_list_str);
 695         t = 0;
 696
 697         BUG_ON(!str);
 698
 699         tprintf("# binding tasks to NODEs:\n");
 700         tprintf("# ");
 701
 702         while (true) {
 703                 int bind_node, bind_node_0, bind_node_1;
 704                 char *tok, *tok_end, *tok_step, *tok_mul;
 705                 int step;
 706                 int mul;
 707
 708                 tok = strsep(&str, ",");
 709                 if (!tok)
 710                         break;
 711
 712                 tok_end = strstr(tok, "-");
 713
 714                 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 715                 if (!tok_end) {
 716                         /* Single NODE specified: */
 717                         bind_node_0 = bind_node_1 = atol(tok);
 718                 } else {
 719                         /* NODE range specified (for example: "5-11"): */
 720                         bind_node_0 = atol(tok);
 721                         bind_node_1 = atol(tok_end + 1);
 722                 }
 723
 724                 step = 1;
 725                 tok_step = strstr(tok, "#");
 726                 if (tok_step) {
 727                         step = atol(tok_step + 1);
 728                         BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 729                 }
 730
 731                 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 732                 mul = 1;
 733                 tok_mul = strstr(tok, "x");
 734                 if (tok_mul) {
 735                         mul = atol(tok_mul + 1);
 736                         BUG_ON(mul <= 0);
 737                 }
 738
 739                 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 740
 741                 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 742                         printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 743                         return -1;
 744                 }
 745
 746                 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 747                 BUG_ON(bind_node_0 > bind_node_1);
 748
 749                 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 750                         int i;
 751
 752                         for (i = 0; i < mul; i++) {
 753                                 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
 754                                         printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 755                                         goto out;
 756                                 }
 757                                 td = g->threads + t;
 758
 759                                 if (!t)
 760                                         tprintf(" %2d", bind_node);
 761                                 else
 762                                         tprintf(",%2d", bind_node);
 763
 764                                 td->bind_node = bind_node;
 765                                 t++;
 766                         }
 767                 }
 768         }
 769 out:
 770
 771         tprintf("\n");
 772
 773         if (t < g->p.nr_tasks)
 774                 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 775
 776         free(str0);
 777         return 0;
 778 }
 779
 780 static int parse_nodes_opt(const struct option *opt __maybe_unused,
 781                           const char *arg, int unset __maybe_unused)
 782 {
 783         if (!arg)
 784                 return -1;
 785
 786         return parse_node_list(arg);
 787 }
 788
 789 static inline uint32_t lfsr_32(uint32_t lfsr)
 790 {
 791         const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 792         return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 793 }
 794
 795 /*
 796  * Make sure there's real data dependency to RAM (when read
 797  * accesses are enabled), so the compiler, the CPU and the
 798  * kernel (KSM, zero page, etc.) cannot optimize away RAM
 799  * accesses:
 800  */
 801 static inline u64 access_data(u64 *data, u64 val)
 802 {
 803         if (g->p.data_reads)
 804                 val += *data;
 805         if (g->p.data_writes)
 806                 *data = val + 1;
 807         return val;
 808 }
 809
 810 /*
 811  * The worker process does two types of work, a forwards going
 812  * loop and a backwards going loop.
 813  *
 814  * We do this so that on multiprocessor systems we do not create
 815  * a 'train' of processing, with highly synchronized processes,
 816  * skewing the whole benchmark.
 817  */
 818 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 819 {
 820         long words = bytes/sizeof(u64);
 821         u64 *data = (void *)__data;
 822         long chunk_0, chunk_1;
 823         u64 *d0, *d, *d1;
 824         long off;
 825         long i;
 826
 827         BUG_ON(!data && words);
 828         BUG_ON(data && !words);
 829
 830         if (!data)
 831                 return val;
 832
 833         /* Very simple memset() work variant: */
 834         if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 835                 bzero(data, bytes);
 836                 return val;
 837         }
 838
 839         /* Spread out by PID/TID nr and by loop nr: */
 840         chunk_0 = words/nr_max;
 841         chunk_1 = words/g->p.nr_loops;
 842         off = nr*chunk_0 + loop*chunk_1;
 843
 844         while (off >= words)
 845                 off -= words;
 846
 847         if (g->p.data_rand_walk) {
 848                 u32 lfsr = nr + loop + val;
 849                 int j;
 850
 851                 for (i = 0; i < words/1024; i++) {
 852                         long start, end;
 853
 854                         lfsr = lfsr_32(lfsr);
 855
 856                         start = lfsr % words;
 857                         end = min(start + 1024, words-1);
 858
 859                         if (g->p.data_zero_memset) {
 860                                 bzero(data + start, (end-start) * sizeof(u64));
 861                         } else {
 862                                 for (j = start; j < end; j++)
 863                                         val = access_data(data + j, val);
 864                         }
 865                 }
 866         } else if (!g->p.data_backwards || (nr + loop) & 1) {
 867                 /* Process data forwards: */
 868
 869                 d0 = data + off;
 870                 d  = data + off + 1;
 871                 d1 = data + words;
 872
 873                 for (;;) {
 874                         if (unlikely(d >= d1))
 875                                 d = data;
 876                         if (unlikely(d == d0))
 877                                 break;
 878
 879                         val = access_data(d, val);
 880
 881                         d++;
 882                 }
 883         } else {
 884                 /* Process data backwards: */
 885
 886                 d0 = data + off;
 887                 d  = data + off - 1;
 888                 d1 = data + words;
 889
 890                 for (;;) {
 891                         if (unlikely(d < data))
 892                                 d = data + words-1;
 893                         if (unlikely(d == d0))
 894                                 break;
 895
 896                         val = access_data(d, val);
 897
 898                         d--;
 899                 }
 900         }
 901
 902         return val;
 903 }
 904
 905 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 906 {
 907         unsigned int cpu;
 908
 909         cpu = sched_getcpu();
 910
 911         g->threads[task_nr].curr_cpu = cpu;
 912         prctl(0, bytes_worked);
 913 }
 914
 915 /*
 916  * Count the number of nodes a process's threads
 917  * are spread out on.
 918  *
 919  * A count of 1 means that the process is compressed
 920  * to a single node. A count of g->p.nr_nodes means it's
 921  * spread out on the whole system.
 922  */
 923 static int count_process_nodes(int process_nr)
 924 {
 925         char *node_present;
 926         int nodes;
 927         int n, t;
 928
 929         node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
 930         BUG_ON(!node_present);
 931         for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
 932                 node_present[nodes] = 0;
 933
 934         for (t = 0; t < g->p.nr_threads; t++) {
 935                 struct thread_data *td;
 936                 int task_nr;
 937                 int node;
 938
 939                 task_nr = process_nr*g->p.nr_threads + t;
 940                 td = g->threads + task_nr;
 941
 942                 node = numa_node_of_cpu(td->curr_cpu);
 943                 if (node < 0) /* curr_cpu was likely still -1 */ {
 944                         free(node_present);
 945                         return 0;
 946                 }
 947
 948                 node_present[node] = 1;
 949         }
 950
 951         nodes = 0;
 952
 953         for (n = 0; n < g->p.nr_nodes; n++)
 954                 nodes += node_present[n];
 955
 956         free(node_present);
 957         return nodes;
 958 }
 959
 960 /*
 961  * Count the number of distinct process-threads a node contains.
 962  *
 963  * A count of 1 means that the node contains only a single
 964  * process. If all nodes on the system contain at most one
 965  * process then we are well-converged.
 966  */
 967 static int count_node_processes(int node)
 968 {
 969         int processes = 0;
 970         int t, p;
 971
 972         for (p = 0; p < g->p.nr_proc; p++) {
 973                 for (t = 0; t < g->p.nr_threads; t++) {
 974                         struct thread_data *td;
 975                         int task_nr;
 976                         int n;
 977
 978                         task_nr = p*g->p.nr_threads + t;
 979                         td = g->threads + task_nr;
 980
 981                         n = numa_node_of_cpu(td->curr_cpu);
 982                         if (n == node) {
 983                                 processes++;
 984                                 break;
 985                         }
 986                 }
 987         }
 988
 989         return processes;
 990 }
 991
 992 static void calc_convergence_compression(int *strong)
 993 {
 994         unsigned int nodes_min, nodes_max;
 995         int p;
 996
 997         nodes_min = -1;
 998         nodes_max =  0;
 999
1000         for (p = 0; p < g->p.nr_proc; p++) {
1001                 unsigned int nodes = count_process_nodes(p);
1002
1003                 if (!nodes) {
1004                         *strong = 0;
1005                         return;
1006                 }
1007
1008                 nodes_min = min(nodes, nodes_min);
1009                 nodes_max = max(nodes, nodes_max);
1010         }
1011
1012         /* Strong convergence: all threads compress on a single node: */
1013         if (nodes_min == 1 && nodes_max == 1) {
1014                 *strong = 1;
1015         } else {
1016                 *strong = 0;
1017                 tprintf(" {%d-%d}", nodes_min, nodes_max);
1018         }
1019 }
1020
1021 static void calc_convergence(double runtime_ns_max, double *convergence)
1022 {
1023         unsigned int loops_done_min, loops_done_max;
1024         int process_groups;
1025         int *nodes;
1026         int distance;
1027         int nr_min;
1028         int nr_max;
1029         int strong;
1030         int sum;
1031         int nr;
1032         int node;
1033         int cpu;
1034         int t;
1035
1036         if (!g->p.show_convergence && !g->p.measure_convergence)
1037                 return;
1038
1039         nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1040         BUG_ON(!nodes);
1041         for (node = 0; node < g->p.nr_nodes; node++)
1042                 nodes[node] = 0;
1043
1044         loops_done_min = -1;
1045         loops_done_max = 0;
1046
1047         for (t = 0; t < g->p.nr_tasks; t++) {
1048                 struct thread_data *td = g->threads + t;
1049                 unsigned int loops_done;
1050
1051                 cpu = td->curr_cpu;
1052
1053                 /* Not all threads have written it yet: */
1054                 if (cpu < 0)
1055                         continue;
1056
1057                 node = numa_node_of_cpu(cpu);
1058
1059                 nodes[node]++;
1060
1061                 loops_done = td->loops_done;
1062                 loops_done_min = min(loops_done, loops_done_min);
1063                 loops_done_max = max(loops_done, loops_done_max);
1064         }
1065
1066         nr_max = 0;
1067         nr_min = g->p.nr_tasks;
1068         sum = 0;
1069
1070         for (node = 0; node < g->p.nr_nodes; node++) {
1071                 if (!is_node_present(node))
1072                         continue;
1073                 nr = nodes[node];
1074                 nr_min = min(nr, nr_min);
1075                 nr_max = max(nr, nr_max);
1076                 sum += nr;
1077         }
1078         BUG_ON(nr_min > nr_max);
1079
1080         BUG_ON(sum > g->p.nr_tasks);
1081
1082         if (0 && (sum < g->p.nr_tasks)) {
1083                 free(nodes);
1084                 return;
1085         }
1086
1087         /*
1088          * Count the number of distinct process groups present
1089          * on nodes - when we are converged this will decrease
1090          * to g->p.nr_proc:
1091          */
1092         process_groups = 0;
1093
1094         for (node = 0; node < g->p.nr_nodes; node++) {
1095                 int processes;
1096
1097                 if (!is_node_present(node))
1098                         continue;
1099                 processes = count_node_processes(node);
1100                 nr = nodes[node];
1101                 tprintf(" %2d/%-2d", nr, processes);
1102
1103                 process_groups += processes;
1104         }
1105
1106         distance = nr_max - nr_min;
1107
1108         tprintf(" [%2d/%-2d]", distance, process_groups);
1109
1110         tprintf(" l:%3d-%-3d (%3d)",
1111                 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1112
1113         if (loops_done_min && loops_done_max) {
1114                 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1115
1116                 tprintf(" [%4.1f%%]", skew * 100.0);
1117         }
1118
1119         calc_convergence_compression(&strong);
1120
1121         if (strong && process_groups == g->p.nr_proc) {
1122                 if (!*convergence) {
1123                         *convergence = runtime_ns_max;
1124                         tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1125                         if (g->p.measure_convergence) {
1126                                 g->all_converged = true;
1127                                 g->stop_work = true;
1128                         }
1129                 }
1130         } else {
1131                 if (*convergence) {
1132                         tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1133                         *convergence = 0;
1134                 }
1135                 tprintf("\n");
1136         }
1137
1138         free(nodes);
1139 }
1140
1141 static void show_summary(double runtime_ns_max, int l, double *convergence)
1142 {
1143         tprintf("\r #  %5.1f%%  [%.1f mins]",
1144                 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1145
1146         calc_convergence(runtime_ns_max, convergence);
1147
1148         if (g->p.show_details >= 0)
1149                 fflush(stdout);
1150 }
1151
1152 static void *worker_thread(void *__tdata)
1153 {
1154         struct thread_data *td = __tdata;
1155         struct timeval start0, start, stop, diff;
1156         int process_nr = td->process_nr;
1157         int thread_nr = td->thread_nr;
1158         unsigned long last_perturbance;
1159         int task_nr = td->task_nr;
1160         int details = g->p.show_details;
1161         int first_task, last_task;
1162         double convergence = 0;
1163         u64 val = td->val;
1164         double runtime_ns_max;
1165         u8 *global_data;
1166         u8 *process_data;
1167         u8 *thread_data;
1168         u64 bytes_done, secs;
1169         long work_done;
1170         u32 l;
1171         struct rusage rusage;
1172
1173         bind_to_cpumask(td->bind_cpumask);
1174         bind_to_memnode(td->bind_node);
1175
1176         set_taskname("thread %d/%d", process_nr, thread_nr);
1177
1178         global_data = g->data;
1179         process_data = td->process_data;
1180         thread_data = setup_private_data(g->p.bytes_thread);
1181
1182         bytes_done = 0;
1183
1184         last_task = 0;
1185         if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1186                 last_task = 1;
1187
1188         first_task = 0;
1189         if (process_nr == 0 && thread_nr == 0)
1190                 first_task = 1;
1191
1192         if (details >= 2) {
1193                 printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1194                         process_nr, thread_nr, global_data, process_data, thread_data);
1195         }
1196
1197         if (g->p.serialize_startup) {
1198                 mutex_lock(&g->startup_mutex);
1199                 g->nr_tasks_started++;
1200                 /* The last thread wakes the main process. */
1201                 if (g->nr_tasks_started == g->p.nr_tasks)
1202                         cond_signal(&g->startup_cond);
1203
1204                 mutex_unlock(&g->startup_mutex);
1205
1206                 /* Here we will wait for the main process to start us all at once: */
1207                 mutex_lock(&g->start_work_mutex);
1208                 g->start_work = false;
1209                 g->nr_tasks_working++;
1210                 while (!g->start_work)
1211                         cond_wait(&g->start_work_cond, &g->start_work_mutex);
1212
1213                 mutex_unlock(&g->start_work_mutex);
1214         }
1215
1216         gettimeofday(&start0, NULL);
1217
1218         start = stop = start0;
1219         last_perturbance = start.tv_sec;
1220
1221         for (l = 0; l < g->p.nr_loops; l++) {
1222                 start = stop;
1223
1224                 if (g->stop_work)
1225                         break;
1226
1227                 val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,      l, val);
1228                 val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,   l, val);
1229                 val += do_work(thread_data,  g->p.bytes_thread,  0,          1,         l, val);
1230
1231                 if (g->p.sleep_usecs) {
1232                         mutex_lock(td->process_lock);
1233                         usleep(g->p.sleep_usecs);
1234                         mutex_unlock(td->process_lock);
1235                 }
1236                 /*
1237                  * Amount of work to be done under a process-global lock:
1238                  */
1239                 if (g->p.bytes_process_locked) {
1240                         mutex_lock(td->process_lock);
1241                         val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,    l, val);
1242                         mutex_unlock(td->process_lock);
1243                 }
1244
1245                 work_done = g->p.bytes_global + g->p.bytes_process +
1246                             g->p.bytes_process_locked + g->p.bytes_thread;
1247
1248                 update_curr_cpu(task_nr, work_done);
1249                 bytes_done += work_done;
1250
1251                 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1252                         continue;
1253
1254                 td->loops_done = l;
1255
1256                 gettimeofday(&stop, NULL);
1257
1258                 /* Check whether our max runtime timed out: */
1259                 if (g->p.nr_secs) {
1260                         timersub(&stop, &start0, &diff);
1261                         if ((u32)diff.tv_sec >= g->p.nr_secs) {
1262                                 g->stop_work = true;
1263                                 break;
1264                         }
1265                 }
1266
1267                 /* Update the summary at most once per second: */
1268                 if (start.tv_sec == stop.tv_sec)
1269                         continue;
1270
1271                 /*
1272                  * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1273                  * by migrating to CPU#0:
1274                  */
1275                 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1276                         cpu_set_t *orig_mask;
1277                         int target_cpu;
1278                         int this_cpu;
1279
1280                         last_perturbance = stop.tv_sec;
1281
1282                         /*
1283                          * Depending on where we are running, move into
1284                          * the other half of the system, to create some
1285                          * real disturbance:
1286                          */
1287                         this_cpu = g->threads[task_nr].curr_cpu;
1288                         if (this_cpu < g->p.nr_cpus/2)
1289                                 target_cpu = g->p.nr_cpus-1;
1290                         else
1291                                 target_cpu = 0;
1292
1293                         orig_mask = bind_to_cpu(target_cpu);
1294
1295                         /* Here we are running on the target CPU already */
1296                         if (details >= 1)
1297                                 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1298
1299                         bind_to_cpumask(orig_mask);
1300                         CPU_FREE(orig_mask);
1301                 }
1302
1303                 if (details >= 3) {
1304                         timersub(&stop, &start, &diff);
1305                         runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1306                         runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1307
1308                         if (details >= 0) {
1309                                 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1310                                         process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1311                         }
1312                         fflush(stdout);
1313                 }
1314                 if (!last_task)
1315                         continue;
1316
1317                 timersub(&stop, &start0, &diff);
1318                 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1319                 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1320
1321                 show_summary(runtime_ns_max, l, &convergence);
1322         }
1323
1324         gettimeofday(&stop, NULL);
1325         timersub(&stop, &start0, &diff);
1326         td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1327         td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1328         secs = td->runtime_ns / NSEC_PER_SEC;
1329         td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1330
1331         getrusage(RUSAGE_THREAD, &rusage);
1332         td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1333         td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1334         td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1335         td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1336
1337         free_data(thread_data, g->p.bytes_thread);
1338
1339         mutex_lock(&g->stop_work_mutex);
1340         g->bytes_done += bytes_done;
1341         mutex_unlock(&g->stop_work_mutex);
1342
1343         return NULL;
1344 }
1345
1346 /*
1347  * A worker process starts a couple of threads:
1348  */
1349 static void worker_process(int process_nr)
1350 {
1351         struct mutex process_lock;
1352         struct thread_data *td;
1353         pthread_t *pthreads;
1354         u8 *process_data;
1355         int task_nr;
1356         int ret;
1357         int t;
1358
1359         mutex_init(&process_lock);
1360         set_taskname("process %d", process_nr);
1361
1362         /*
1363          * Pick up the memory policy and the CPU binding of our first thread,
1364          * so that we initialize memory accordingly:
1365          */
1366         task_nr = process_nr*g->p.nr_threads;
1367         td = g->threads + task_nr;
1368
1369         bind_to_memnode(td->bind_node);
1370         bind_to_cpumask(td->bind_cpumask);
1371
1372         pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1373         process_data = setup_private_data(g->p.bytes_process);
1374
1375         if (g->p.show_details >= 3) {
1376                 printf(" # process %2d global mem: %p, process mem: %p\n",
1377                         process_nr, g->data, process_data);
1378         }
1379
1380         for (t = 0; t < g->p.nr_threads; t++) {
1381                 task_nr = process_nr*g->p.nr_threads + t;
1382                 td = g->threads + task_nr;
1383
1384                 td->process_data = process_data;
1385                 td->process_nr   = process_nr;
1386                 td->thread_nr    = t;
1387                 td->task_nr      = task_nr;
1388                 td->val          = rand();
1389                 td->curr_cpu     = -1;
1390                 td->process_lock = &process_lock;
1391
1392                 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1393                 BUG_ON(ret);
1394         }
1395
1396         for (t = 0; t < g->p.nr_threads; t++) {
1397                 ret = pthread_join(pthreads[t], NULL);
1398                 BUG_ON(ret);
1399         }
1400
1401         free_data(process_data, g->p.bytes_process);
1402         free(pthreads);
1403 }
1404
1405 static void print_summary(void)
1406 {
1407         if (g->p.show_details < 0)
1408                 return;
1409
1410         printf("\n ###\n");
1411         printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1412                 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1413         printf(" #      %5dx %5ldMB global  shared mem operations\n",
1414                         g->p.nr_loops, g->p.bytes_global/1024/1024);
1415         printf(" #      %5dx %5ldMB process shared mem operations\n",
1416                         g->p.nr_loops, g->p.bytes_process/1024/1024);
1417         printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1418                         g->p.nr_loops, g->p.bytes_thread/1024/1024);
1419
1420         printf(" ###\n");
1421
1422         printf("\n ###\n"); fflush(stdout);
1423 }
1424
1425 static void init_thread_data(void)
1426 {
1427         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1428         int t;
1429
1430         g->threads = zalloc_shared_data(size);
1431
1432         for (t = 0; t < g->p.nr_tasks; t++) {
1433                 struct thread_data *td = g->threads + t;
1434                 size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1435                 int cpu;
1436
1437                 /* Allow all nodes by default: */
1438                 td->bind_node = NUMA_NO_NODE;
1439
1440                 /* Allow all CPUs by default: */
1441                 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1442                 BUG_ON(!td->bind_cpumask);
1443                 CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1444                 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1445                         CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1446         }
1447 }
1448
1449 static void deinit_thread_data(void)
1450 {
1451         ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1452         int t;
1453
1454         /* Free the bind_cpumask allocated for thread_data */
1455         for (t = 0; t < g->p.nr_tasks; t++) {
1456                 struct thread_data *td = g->threads + t;
1457                 CPU_FREE(td->bind_cpumask);
1458         }
1459
1460         free_data(g->threads, size);
1461 }
1462
1463 static int init(void)
1464 {
1465         g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1466
1467         /* Copy over options: */
1468         g->p = p0;
1469
1470         g->p.nr_cpus = numa_num_configured_cpus();
1471
1472         g->p.nr_nodes = numa_max_node() + 1;
1473
1474         /* char array in count_process_nodes(): */
1475         BUG_ON(g->p.nr_nodes < 0);
1476
1477         if (g->p.show_quiet && !g->p.show_details)
1478                 g->p.show_details = -1;
1479
1480         /* Some memory should be specified: */
1481         if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1482                 return -1;
1483
1484         if (g->p.mb_global_str) {
1485                 g->p.mb_global = atof(g->p.mb_global_str);
1486                 BUG_ON(g->p.mb_global < 0);
1487         }
1488
1489         if (g->p.mb_proc_str) {
1490                 g->p.mb_proc = atof(g->p.mb_proc_str);
1491                 BUG_ON(g->p.mb_proc < 0);
1492         }
1493
1494         if (g->p.mb_proc_locked_str) {
1495                 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1496                 BUG_ON(g->p.mb_proc_locked < 0);
1497                 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1498         }
1499
1500         if (g->p.mb_thread_str) {
1501                 g->p.mb_thread = atof(g->p.mb_thread_str);
1502                 BUG_ON(g->p.mb_thread < 0);
1503         }
1504
1505         BUG_ON(g->p.nr_threads <= 0);
1506         BUG_ON(g->p.nr_proc <= 0);
1507
1508         g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1509
1510         g->p.bytes_global               = g->p.mb_global        *1024L*1024L;
1511         g->p.bytes_process              = g->p.mb_proc          *1024L*1024L;
1512         g->p.bytes_process_locked       = g->p.mb_proc_locked   *1024L*1024L;
1513         g->p.bytes_thread               = g->p.mb_thread        *1024L*1024L;
1514
1515         g->data = setup_shared_data(g->p.bytes_global);
1516
1517         /* Startup serialization: */
1518         mutex_init_pshared(&g->start_work_mutex);
1519         cond_init_pshared(&g->start_work_cond);
1520         mutex_init_pshared(&g->startup_mutex);
1521         cond_init_pshared(&g->startup_cond);
1522         mutex_init_pshared(&g->stop_work_mutex);
1523
1524         init_thread_data();
1525
1526         tprintf("#\n");
1527         if (parse_setup_cpu_list() || parse_setup_node_list())
1528                 return -1;
1529         tprintf("#\n");
1530
1531         print_summary();
1532
1533         return 0;
1534 }
1535
1536 static void deinit(void)
1537 {
1538         free_data(g->data, g->p.bytes_global);
1539         g->data = NULL;
1540
1541         deinit_thread_data();
1542
1543         free_data(g, sizeof(*g));
1544         g = NULL;
1545 }
1546
1547 /*
1548  * Print a short or long result, depending on the verbosity setting:
1549  */
1550 static void print_res(const char *name, double val,
1551                       const char *txt_unit, const char *txt_short, const char *txt_long)
1552 {
1553         if (!name)
1554                 name = "main,";
1555
1556         if (!g->p.show_quiet)
1557                 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1558         else
1559                 printf(" %14.3f %s\n", val, txt_long);
1560 }
1561
1562 static int __bench_numa(const char *name)
1563 {
1564         struct timeval start, stop, diff;
1565         u64 runtime_ns_min, runtime_ns_sum;
1566         pid_t *pids, pid, wpid;
1567         double delta_runtime;
1568         double runtime_avg;
1569         double runtime_sec_max;
1570         double runtime_sec_min;
1571         int wait_stat;
1572         double bytes;
1573         int i, t, p;
1574
1575         if (init())
1576                 return -1;
1577
1578         pids = zalloc(g->p.nr_proc * sizeof(*pids));
1579         pid = -1;
1580
1581         if (g->p.serialize_startup) {
1582                 tprintf(" #\n");
1583                 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1584         }
1585
1586         gettimeofday(&start, NULL);
1587
1588         for (i = 0; i < g->p.nr_proc; i++) {
1589                 pid = fork();
1590                 dprintf(" # process %2d: PID %d\n", i, pid);
1591
1592                 BUG_ON(pid < 0);
1593                 if (!pid) {
1594                         /* Child process: */
1595                         worker_process(i);
1596
1597                         exit(0);
1598                 }
1599                 pids[i] = pid;
1600
1601         }
1602
1603         if (g->p.serialize_startup) {
1604                 bool threads_ready = false;
1605                 double startup_sec;
1606
1607                 /*
1608                  * Wait for all the threads to start up. The last thread will
1609                  * signal this process.
1610                  */
1611                 mutex_lock(&g->startup_mutex);
1612                 while (g->nr_tasks_started != g->p.nr_tasks)
1613                         cond_wait(&g->startup_cond, &g->startup_mutex);
1614
1615                 mutex_unlock(&g->startup_mutex);
1616
1617                 /* Wait for all threads to be at the start_work_cond. */
1618                 while (!threads_ready) {
1619                         mutex_lock(&g->start_work_mutex);
1620                         threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1621                         mutex_unlock(&g->start_work_mutex);
1622                         if (!threads_ready)
1623                                 usleep(1);
1624                 }
1625
1626                 gettimeofday(&stop, NULL);
1627
1628                 timersub(&stop, &start, &diff);
1629
1630                 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1631                 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1632                 startup_sec /= NSEC_PER_SEC;
1633
1634                 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1635                 tprintf(" #\n");
1636
1637                 start = stop;
1638                 /* Start all threads running. */
1639                 mutex_lock(&g->start_work_mutex);
1640                 g->start_work = true;
1641                 mutex_unlock(&g->start_work_mutex);
1642                 cond_broadcast(&g->start_work_cond);
1643         } else {
1644                 gettimeofday(&start, NULL);
1645         }
1646
1647         /* Parent process: */
1648
1649
1650         for (i = 0; i < g->p.nr_proc; i++) {
1651                 wpid = waitpid(pids[i], &wait_stat, 0);
1652                 BUG_ON(wpid < 0);
1653                 BUG_ON(!WIFEXITED(wait_stat));
1654
1655         }
1656
1657         runtime_ns_sum = 0;
1658         runtime_ns_min = -1LL;
1659
1660         for (t = 0; t < g->p.nr_tasks; t++) {
1661                 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1662
1663                 runtime_ns_sum += thread_runtime_ns;
1664                 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1665         }
1666
1667         gettimeofday(&stop, NULL);
1668         timersub(&stop, &start, &diff);
1669
1670         BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1671
1672         tprintf("\n ###\n");
1673         tprintf("\n");
1674
1675         runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1676         runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1677         runtime_sec_max /= NSEC_PER_SEC;
1678
1679         runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1680
1681         bytes = g->bytes_done;
1682         runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1683
1684         if (g->p.measure_convergence) {
1685                 print_res(name, runtime_sec_max,
1686                         "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1687         }
1688
1689         print_res(name, runtime_sec_max,
1690                 "secs,", "runtime-max/thread",  "secs slowest (max) thread-runtime");
1691
1692         print_res(name, runtime_sec_min,
1693                 "secs,", "runtime-min/thread",  "secs fastest (min) thread-runtime");
1694
1695         print_res(name, runtime_avg,
1696                 "secs,", "runtime-avg/thread",  "secs average thread-runtime");
1697
1698         delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1699         print_res(name, delta_runtime / runtime_sec_max * 100.0,
1700                 "%,", "spread-runtime/thread",  "% difference between max/avg runtime");
1701
1702         print_res(name, bytes / g->p.nr_tasks / 1e9,
1703                 "GB,", "data/thread",           "GB data processed, per thread");
1704
1705         print_res(name, bytes / 1e9,
1706                 "GB,", "data-total",            "GB data processed, total");
1707
1708         print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1709                 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1710
1711         print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1712                 "GB/sec,", "thread-speed",      "GB/sec/thread speed");
1713
1714         print_res(name, bytes / runtime_sec_max / 1e9,
1715                 "GB/sec,", "total-speed",       "GB/sec total speed");
1716
1717         if (g->p.show_details >= 2) {
1718                 char tname[14 + 2 * 11 + 1];
1719                 struct thread_data *td;
1720                 for (p = 0; p < g->p.nr_proc; p++) {
1721                         for (t = 0; t < g->p.nr_threads; t++) {
1722                                 memset(tname, 0, sizeof(tname));
1723                                 td = g->threads + p*g->p.nr_threads + t;
1724                                 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1725                                 print_res(tname, td->speed_gbs,
1726                                         "GB/sec",       "thread-speed", "GB/sec/thread speed");
1727                                 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1728                                         "secs", "thread-system-time", "system CPU time/thread");
1729                                 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1730                                         "secs", "thread-user-time", "user CPU time/thread");
1731                         }
1732                 }
1733         }
1734
1735         free(pids);
1736
1737         deinit();
1738
1739         return 0;
1740 }
1741
1742 #define MAX_ARGS 50
1743
1744 static int command_size(const char **argv)
1745 {
1746         int size = 0;
1747
1748         while (*argv) {
1749                 size++;
1750                 argv++;
1751         }
1752
1753         BUG_ON(size >= MAX_ARGS);
1754
1755         return size;
1756 }
1757
1758 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1759 {
1760         int i;
1761
1762         printf("\n # Running %s \"perf bench numa", name);
1763
1764         for (i = 0; i < argc; i++)
1765                 printf(" %s", argv[i]);
1766
1767         printf("\"\n");
1768
1769         memset(p, 0, sizeof(*p));
1770
1771         /* Initialize nonzero defaults: */
1772
1773         p->serialize_startup            = 1;
1774         p->data_reads                   = true;
1775         p->data_writes                  = true;
1776         p->data_backwards               = true;
1777         p->data_rand_walk               = true;
1778         p->nr_loops                     = -1;
1779         p->init_random                  = true;
1780         p->mb_global_str                = "1";
1781         p->nr_proc                      = 1;
1782         p->nr_threads                   = 1;
1783         p->nr_secs                      = 5;
1784         p->run_all                      = argc == 1;
1785 }
1786
1787 static int run_bench_numa(const char *name, const char **argv)
1788 {
1789         int argc = command_size(argv);
1790
1791         init_params(&p0, name, argc, argv);
1792         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1793         if (argc)
1794                 goto err;
1795
1796         if (__bench_numa(name))
1797                 goto err;
1798
1799         return 0;
1800
1801 err:
1802         return -1;
1803 }
1804
1805 #define OPT_BW_RAM              "-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1806 #define OPT_BW_RAM_NOTHP        OPT_BW_RAM,             "--thp", "-1"
1807
1808 #define OPT_CONV                "-s", "100", "-zZ0qcm", "--thp", " 1"
1809 #define OPT_CONV_NOTHP          OPT_CONV,               "--thp", "-1"
1810
1811 #define OPT_BW                  "-s",  "20", "-zZ0q",   "--thp", " 1"
1812 #define OPT_BW_NOTHP            OPT_BW,                 "--thp", "-1"
1813
1814 /*
1815  * The built-in test-suite executed by "perf bench numa -a".
1816  *
1817  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1818  */
1819 static const char *tests[][MAX_ARGS] = {
1820    /* Basic single-stream NUMA bandwidth measurements: */
1821    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1822                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1823    { "RAM-bw-local-NOTHP,",
1824                           "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1825                           "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1826    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1827                           "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1828
1829    /* 2-stream NUMA bandwidth measurements: */
1830    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1831                            "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1832    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1833                            "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1834
1835    /* Cross-stream NUMA bandwidth measurement: */
1836    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1837                            "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1838
1839    /* Convergence latency measurements: */
1840    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1841    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1842    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1843    { " 2x3-convergence,", "mem",  "-p",  "2", "-t",  "3", "-P", "1020", OPT_CONV },
1844    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1845    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1846    { " 4x4-convergence-NOTHP,",
1847                           "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1848    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1849    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1850    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1851    { " 8x4-convergence-NOTHP,",
1852                           "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1853    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1854    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1855    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1856    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1857    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1858
1859    /* Various NUMA process/thread layout bandwidth measurements: */
1860    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1861    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1862    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1863    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1864    { " 8x1-bw-process-NOTHP,",
1865                           "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1866    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1867
1868    { " 1x4-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1869    { " 1x8-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1870    { "1x16-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1871    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1872
1873    { " 2x3-bw-process,",  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1874    { " 4x4-bw-process,",  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1875    { " 4x6-bw-process,",  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1876    { " 4x8-bw-process,",  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1877    { " 4x8-bw-process-NOTHP,",
1878                           "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1879    { " 3x3-bw-process,",  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1880    { " 5x5-bw-process,",  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1881
1882    { "2x16-bw-process,",  "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1883    { "1x32-bw-process,",  "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1884
1885    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1886    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1887    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1888    { "numa01-bw-thread-NOTHP,",
1889                           "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1890 };
1891
1892 static int bench_all(void)
1893 {
1894         int nr = ARRAY_SIZE(tests);
1895         int ret;
1896         int i;
1897
1898         ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1899         BUG_ON(ret < 0);
1900
1901         for (i = 0; i < nr; i++) {
1902                 run_bench_numa(tests[i][0], tests[i] + 1);
1903         }
1904
1905         printf("\n");
1906
1907         return 0;
1908 }
1909
1910 int bench_numa(int argc, const char **argv)
1911 {
1912         init_params(&p0, "main,", argc, argv);
1913         argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1914         if (argc)
1915                 goto err;
1916
1917         if (p0.run_all)
1918                 return bench_all();
1919
1920         if (__bench_numa(NULL))
1921                 goto err;
1922
1923         return 0;
1924
1925 err:
1926         usage_with_options(numa_usage, options);
1927         return -1;
1928 }