Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Mar 24-27, 2025, special US time zones
Register
Loading...
v4.6
 
   1/*
   2 * numa.c
   3 *
   4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5 */
   6
   7#include "../perf.h"
   8#include "../builtin.h"
   9#include "../util/util.h"
  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 <pthread.h>
  25#include <sys/mman.h>
  26#include <sys/time.h>
  27#include <sys/resource.h>
  28#include <sys/wait.h>
  29#include <sys/prctl.h>
 
  30#include <sys/types.h>
  31
 
 
 
 
 
 
 
  32#include <numa.h>
  33#include <numaif.h>
  34
 
 
 
 
  35/*
  36 * Regular printout to the terminal, supressed if -q is specified:
  37 */
  38#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  39
  40/*
  41 * Debug printf:
  42 */
 
  43#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  44
  45struct thread_data {
  46	int			curr_cpu;
  47	cpu_set_t		bind_cpumask;
  48	int			bind_node;
  49	u8			*process_data;
  50	int			process_nr;
  51	int			thread_nr;
  52	int			task_nr;
  53	unsigned int		loops_done;
  54	u64			val;
  55	u64			runtime_ns;
  56	u64			system_time_ns;
  57	u64			user_time_ns;
  58	double			speed_gbs;
  59	pthread_mutex_t		*process_lock;
  60};
  61
  62/* Parameters set by options: */
  63
  64struct params {
  65	/* Startup synchronization: */
  66	bool			serialize_startup;
  67
  68	/* Task hierarchy: */
  69	int			nr_proc;
  70	int			nr_threads;
  71
  72	/* Working set sizes: */
  73	const char		*mb_global_str;
  74	const char		*mb_proc_str;
  75	const char		*mb_proc_locked_str;
  76	const char		*mb_thread_str;
  77
  78	double			mb_global;
  79	double			mb_proc;
  80	double			mb_proc_locked;
  81	double			mb_thread;
  82
  83	/* Access patterns to the working set: */
  84	bool			data_reads;
  85	bool			data_writes;
  86	bool			data_backwards;
  87	bool			data_zero_memset;
  88	bool			data_rand_walk;
  89	u32			nr_loops;
  90	u32			nr_secs;
  91	u32			sleep_usecs;
  92
  93	/* Working set initialization: */
  94	bool			init_zero;
  95	bool			init_random;
  96	bool			init_cpu0;
  97
  98	/* Misc options: */
  99	int			show_details;
 100	int			run_all;
 101	int			thp;
 102
 103	long			bytes_global;
 104	long			bytes_process;
 105	long			bytes_process_locked;
 106	long			bytes_thread;
 107
 108	int			nr_tasks;
 109	bool			show_quiet;
 110
 111	bool			show_convergence;
 112	bool			measure_convergence;
 113
 114	int			perturb_secs;
 115	int			nr_cpus;
 116	int			nr_nodes;
 117
 118	/* Affinity options -C and -N: */
 119	char			*cpu_list_str;
 120	char			*node_list_str;
 121};
 122
 123
 124/* Global, read-writable area, accessible to all processes and threads: */
 125
 126struct global_info {
 127	u8			*data;
 128
 129	pthread_mutex_t		startup_mutex;
 
 130	int			nr_tasks_started;
 131
 132	pthread_mutex_t		startup_done_mutex;
 133
 134	pthread_mutex_t		start_work_mutex;
 135	int			nr_tasks_working;
 
 136
 137	pthread_mutex_t		stop_work_mutex;
 138	u64			bytes_done;
 139
 140	struct thread_data	*threads;
 141
 142	/* Convergence latency measurement: */
 143	bool			all_converged;
 144	bool			stop_work;
 145
 146	int			print_once;
 147
 148	struct params		p;
 149};
 150
 151static struct global_info	*g = NULL;
 152
 153static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 154static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 155
 156struct params p0;
 157
 158static const struct option options[] = {
 159	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
 160	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
 161
 162	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
 163	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
 164	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 165	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
 166
 167	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
 168	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
 169	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
 170
 171	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via writes (can be mixed with -W)"),
 172	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
 173	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
 174	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 175	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
 176
 177
 178	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
 179	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
 180	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
 181	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
 182
 183	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
 184	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
 185	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 186	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
 
 187	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
 188	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"quiet mode"),
 
 189	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 190
 191	/* Special option string parsing callbacks: */
 192        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 193			"bind the first N tasks to these specific cpus (the rest is unbound)",
 194			parse_cpus_opt),
 195        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 196			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
 197			parse_nodes_opt),
 198	OPT_END()
 199};
 200
 201static const char * const bench_numa_usage[] = {
 202	"perf bench numa <options>",
 203	NULL
 204};
 205
 206static const char * const numa_usage[] = {
 207	"perf bench numa mem [<options>]",
 208	NULL
 209};
 210
 211static cpu_set_t bind_to_cpu(int target_cpu)
 
 
 
 212{
 213	cpu_set_t orig_mask, mask;
 214	int ret;
 215
 216	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 217	BUG_ON(ret);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 218
 219	CPU_ZERO(&mask);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 220
 221	if (target_cpu == -1) {
 222		int cpu;
 223
 224		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 225			CPU_SET(cpu, &mask);
 226	} else {
 227		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 228		CPU_SET(target_cpu, &mask);
 
 
 229	}
 230
 231	ret = sched_setaffinity(0, sizeof(mask), &mask);
 232	BUG_ON(ret);
 233
 234	return orig_mask;
 
 
 
 
 
 
 
 
 
 235}
 236
 237static cpu_set_t bind_to_node(int target_node)
 238{
 239	int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 240	cpu_set_t orig_mask, mask;
 
 241	int cpu;
 242	int ret;
 243
 244	BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 245	BUG_ON(!cpus_per_node);
 
 
 246
 247	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 248	BUG_ON(ret);
 249
 250	CPU_ZERO(&mask);
 
 
 251
 252	if (target_node == -1) {
 
 
 253		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 254			CPU_SET(cpu, &mask);
 255	} else {
 256		int cpu_start = (target_node + 0) * cpus_per_node;
 257		int cpu_stop  = (target_node + 1) * cpus_per_node;
 258
 259		BUG_ON(cpu_stop > g->p.nr_cpus);
 
 260
 261		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 262			CPU_SET(cpu, &mask);
 
 
 
 
 
 263	}
 264
 265	ret = sched_setaffinity(0, sizeof(mask), &mask);
 266	BUG_ON(ret);
 267
 268	return orig_mask;
 
 
 
 
 
 
 
 
 
 269}
 270
 271static void bind_to_cpumask(cpu_set_t mask)
 272{
 273	int ret;
 
 274
 275	ret = sched_setaffinity(0, sizeof(mask), &mask);
 276	BUG_ON(ret);
 
 
 
 277}
 278
 279static void mempol_restore(void)
 280{
 281	int ret;
 282
 283	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 284
 285	BUG_ON(ret);
 286}
 287
 288static void bind_to_memnode(int node)
 289{
 290	unsigned long nodemask;
 291	int ret;
 292
 293	if (node == -1)
 294		return;
 295
 296	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 297	nodemask = 1L << node;
 
 
 
 298
 299	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 300	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 301
 
 302	BUG_ON(ret);
 303}
 304
 305#define HPSIZE (2*1024*1024)
 306
 307#define set_taskname(fmt...)				\
 308do {							\
 309	char name[20];					\
 310							\
 311	snprintf(name, 20, fmt);			\
 312	prctl(PR_SET_NAME, name);			\
 313} while (0)
 314
 315static u8 *alloc_data(ssize_t bytes0, int map_flags,
 316		      int init_zero, int init_cpu0, int thp, int init_random)
 317{
 318	cpu_set_t orig_mask;
 319	ssize_t bytes;
 320	u8 *buf;
 321	int ret;
 322
 323	if (!bytes0)
 324		return NULL;
 325
 326	/* Allocate and initialize all memory on CPU#0: */
 327	if (init_cpu0) {
 328		orig_mask = bind_to_node(0);
 329		bind_to_memnode(0);
 
 
 330	}
 331
 332	bytes = bytes0 + HPSIZE;
 333
 334	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 335	BUG_ON(buf == (void *)-1);
 336
 337	if (map_flags == MAP_PRIVATE) {
 338		if (thp > 0) {
 339			ret = madvise(buf, bytes, MADV_HUGEPAGE);
 340			if (ret && !g->print_once) {
 341				g->print_once = 1;
 342				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 343			}
 344		}
 345		if (thp < 0) {
 346			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 347			if (ret && !g->print_once) {
 348				g->print_once = 1;
 349				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 350			}
 351		}
 352	}
 353
 354	if (init_zero) {
 355		bzero(buf, bytes);
 356	} else {
 357		/* Initialize random contents, different in each word: */
 358		if (init_random) {
 359			u64 *wbuf = (void *)buf;
 360			long off = rand();
 361			long i;
 362
 363			for (i = 0; i < bytes/8; i++)
 364				wbuf[i] = i + off;
 365		}
 366	}
 367
 368	/* Align to 2MB boundary: */
 369	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 370
 371	/* Restore affinity: */
 372	if (init_cpu0) {
 373		bind_to_cpumask(orig_mask);
 
 374		mempol_restore();
 375	}
 376
 377	return buf;
 378}
 379
 380static void free_data(void *data, ssize_t bytes)
 381{
 382	int ret;
 383
 384	if (!data)
 385		return;
 386
 387	ret = munmap(data, bytes);
 388	BUG_ON(ret);
 389}
 390
 391/*
 392 * Create a shared memory buffer that can be shared between processes, zeroed:
 393 */
 394static void * zalloc_shared_data(ssize_t bytes)
 395{
 396	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 397}
 398
 399/*
 400 * Create a shared memory buffer that can be shared between processes:
 401 */
 402static void * setup_shared_data(ssize_t bytes)
 403{
 404	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 405}
 406
 407/*
 408 * Allocate process-local memory - this will either be shared between
 409 * threads of this process, or only be accessed by this thread:
 410 */
 411static void * setup_private_data(ssize_t bytes)
 412{
 413	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 414}
 415
 416/*
 417 * Return a process-shared (global) mutex:
 418 */
 419static void init_global_mutex(pthread_mutex_t *mutex)
 420{
 421	pthread_mutexattr_t attr;
 422
 423	pthread_mutexattr_init(&attr);
 424	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 425	pthread_mutex_init(mutex, &attr);
 426}
 427
 428static int parse_cpu_list(const char *arg)
 429{
 430	p0.cpu_list_str = strdup(arg);
 431
 432	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 433
 434	return 0;
 435}
 436
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 437static int parse_setup_cpu_list(void)
 438{
 439	struct thread_data *td;
 440	char *str0, *str;
 441	int t;
 442
 443	if (!g->p.cpu_list_str)
 444		return 0;
 445
 446	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 447
 448	str0 = str = strdup(g->p.cpu_list_str);
 449	t = 0;
 450
 451	BUG_ON(!str);
 452
 453	tprintf("# binding tasks to CPUs:\n");
 454	tprintf("#  ");
 455
 456	while (true) {
 457		int bind_cpu, bind_cpu_0, bind_cpu_1;
 458		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 459		int bind_len;
 460		int step;
 461		int mul;
 462
 463		tok = strsep(&str, ",");
 464		if (!tok)
 465			break;
 466
 467		tok_end = strstr(tok, "-");
 468
 469		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 470		if (!tok_end) {
 471			/* Single CPU specified: */
 472			bind_cpu_0 = bind_cpu_1 = atol(tok);
 473		} else {
 474			/* CPU range specified (for example: "5-11"): */
 475			bind_cpu_0 = atol(tok);
 476			bind_cpu_1 = atol(tok_end + 1);
 477		}
 478
 479		step = 1;
 480		tok_step = strstr(tok, "#");
 481		if (tok_step) {
 482			step = atol(tok_step + 1);
 483			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 484		}
 485
 486		/*
 487		 * Mask length.
 488		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 489		 * where the _4 means the next 4 CPUs are allowed.
 490		 */
 491		bind_len = 1;
 492		tok_len = strstr(tok, "_");
 493		if (tok_len) {
 494			bind_len = atol(tok_len + 1);
 495			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 496		}
 497
 498		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 499		mul = 1;
 500		tok_mul = strstr(tok, "x");
 501		if (tok_mul) {
 502			mul = atol(tok_mul + 1);
 503			BUG_ON(mul <= 0);
 504		}
 505
 506		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 507
 508		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 509			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 510			return -1;
 511		}
 512
 
 
 
 
 
 513		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 514		BUG_ON(bind_cpu_0 > bind_cpu_1);
 515
 516		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 
 517			int i;
 518
 519			for (i = 0; i < mul; i++) {
 520				int cpu;
 521
 522				if (t >= g->p.nr_tasks) {
 523					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 524					goto out;
 525				}
 526				td = g->threads + t;
 527
 528				if (t)
 529					tprintf(",");
 530				if (bind_len > 1) {
 531					tprintf("%2d/%d", bind_cpu, bind_len);
 532				} else {
 533					tprintf("%2d", bind_cpu);
 534				}
 535
 536				CPU_ZERO(&td->bind_cpumask);
 
 
 537				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 538					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 539					CPU_SET(cpu, &td->bind_cpumask);
 
 
 
 540				}
 541				t++;
 542			}
 543		}
 544	}
 545out:
 546
 547	tprintf("\n");
 548
 549	if (t < g->p.nr_tasks)
 550		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 551
 552	free(str0);
 553	return 0;
 554}
 555
 556static int parse_cpus_opt(const struct option *opt __maybe_unused,
 557			  const char *arg, int unset __maybe_unused)
 558{
 559	if (!arg)
 560		return -1;
 561
 562	return parse_cpu_list(arg);
 563}
 564
 565static int parse_node_list(const char *arg)
 566{
 567	p0.node_list_str = strdup(arg);
 568
 569	dprintf("got NODE list: {%s}\n", p0.node_list_str);
 570
 571	return 0;
 572}
 573
 574static int parse_setup_node_list(void)
 575{
 576	struct thread_data *td;
 577	char *str0, *str;
 578	int t;
 579
 580	if (!g->p.node_list_str)
 581		return 0;
 582
 583	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 584
 585	str0 = str = strdup(g->p.node_list_str);
 586	t = 0;
 587
 588	BUG_ON(!str);
 589
 590	tprintf("# binding tasks to NODEs:\n");
 591	tprintf("# ");
 592
 593	while (true) {
 594		int bind_node, bind_node_0, bind_node_1;
 595		char *tok, *tok_end, *tok_step, *tok_mul;
 596		int step;
 597		int mul;
 598
 599		tok = strsep(&str, ",");
 600		if (!tok)
 601			break;
 602
 603		tok_end = strstr(tok, "-");
 604
 605		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 606		if (!tok_end) {
 607			/* Single NODE specified: */
 608			bind_node_0 = bind_node_1 = atol(tok);
 609		} else {
 610			/* NODE range specified (for example: "5-11"): */
 611			bind_node_0 = atol(tok);
 612			bind_node_1 = atol(tok_end + 1);
 613		}
 614
 615		step = 1;
 616		tok_step = strstr(tok, "#");
 617		if (tok_step) {
 618			step = atol(tok_step + 1);
 619			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 620		}
 621
 622		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 623		mul = 1;
 624		tok_mul = strstr(tok, "x");
 625		if (tok_mul) {
 626			mul = atol(tok_mul + 1);
 627			BUG_ON(mul <= 0);
 628		}
 629
 630		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 631
 632		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 633			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 634			return -1;
 635		}
 636
 637		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 638		BUG_ON(bind_node_0 > bind_node_1);
 639
 640		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 641			int i;
 642
 643			for (i = 0; i < mul; i++) {
 644				if (t >= g->p.nr_tasks) {
 645					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 646					goto out;
 647				}
 648				td = g->threads + t;
 649
 650				if (!t)
 651					tprintf(" %2d", bind_node);
 652				else
 653					tprintf(",%2d", bind_node);
 654
 655				td->bind_node = bind_node;
 656				t++;
 657			}
 658		}
 659	}
 660out:
 661
 662	tprintf("\n");
 663
 664	if (t < g->p.nr_tasks)
 665		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 666
 667	free(str0);
 668	return 0;
 669}
 670
 671static int parse_nodes_opt(const struct option *opt __maybe_unused,
 672			  const char *arg, int unset __maybe_unused)
 673{
 674	if (!arg)
 675		return -1;
 676
 677	return parse_node_list(arg);
 678
 679	return 0;
 680}
 681
 682#define BIT(x) (1ul << x)
 683
 684static inline uint32_t lfsr_32(uint32_t lfsr)
 685{
 686	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 687	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 688}
 689
 690/*
 691 * Make sure there's real data dependency to RAM (when read
 692 * accesses are enabled), so the compiler, the CPU and the
 693 * kernel (KSM, zero page, etc.) cannot optimize away RAM
 694 * accesses:
 695 */
 696static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 697{
 698	if (g->p.data_reads)
 699		val += *data;
 700	if (g->p.data_writes)
 701		*data = val + 1;
 702	return val;
 703}
 704
 705/*
 706 * The worker process does two types of work, a forwards going
 707 * loop and a backwards going loop.
 708 *
 709 * We do this so that on multiprocessor systems we do not create
 710 * a 'train' of processing, with highly synchronized processes,
 711 * skewing the whole benchmark.
 712 */
 713static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 714{
 715	long words = bytes/sizeof(u64);
 716	u64 *data = (void *)__data;
 717	long chunk_0, chunk_1;
 718	u64 *d0, *d, *d1;
 719	long off;
 720	long i;
 721
 722	BUG_ON(!data && words);
 723	BUG_ON(data && !words);
 724
 725	if (!data)
 726		return val;
 727
 728	/* Very simple memset() work variant: */
 729	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 730		bzero(data, bytes);
 731		return val;
 732	}
 733
 734	/* Spread out by PID/TID nr and by loop nr: */
 735	chunk_0 = words/nr_max;
 736	chunk_1 = words/g->p.nr_loops;
 737	off = nr*chunk_0 + loop*chunk_1;
 738
 739	while (off >= words)
 740		off -= words;
 741
 742	if (g->p.data_rand_walk) {
 743		u32 lfsr = nr + loop + val;
 744		int j;
 745
 746		for (i = 0; i < words/1024; i++) {
 747			long start, end;
 748
 749			lfsr = lfsr_32(lfsr);
 750
 751			start = lfsr % words;
 752			end = min(start + 1024, words-1);
 753
 754			if (g->p.data_zero_memset) {
 755				bzero(data + start, (end-start) * sizeof(u64));
 756			} else {
 757				for (j = start; j < end; j++)
 758					val = access_data(data + j, val);
 759			}
 760		}
 761	} else if (!g->p.data_backwards || (nr + loop) & 1) {
 
 762
 763		d0 = data + off;
 764		d  = data + off + 1;
 765		d1 = data + words;
 766
 767		/* Process data forwards: */
 768		for (;;) {
 769			if (unlikely(d >= d1))
 770				d = data;
 771			if (unlikely(d == d0))
 772				break;
 773
 774			val = access_data(d, val);
 775
 776			d++;
 777		}
 778	} else {
 779		/* Process data backwards: */
 780
 781		d0 = data + off;
 782		d  = data + off - 1;
 783		d1 = data + words;
 784
 785		/* Process data forwards: */
 786		for (;;) {
 787			if (unlikely(d < data))
 788				d = data + words-1;
 789			if (unlikely(d == d0))
 790				break;
 791
 792			val = access_data(d, val);
 793
 794			d--;
 795		}
 796	}
 797
 798	return val;
 799}
 800
 801static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 802{
 803	unsigned int cpu;
 804
 805	cpu = sched_getcpu();
 806
 807	g->threads[task_nr].curr_cpu = cpu;
 808	prctl(0, bytes_worked);
 809}
 810
 811#define MAX_NR_NODES	64
 812
 813/*
 814 * Count the number of nodes a process's threads
 815 * are spread out on.
 816 *
 817 * A count of 1 means that the process is compressed
 818 * to a single node. A count of g->p.nr_nodes means it's
 819 * spread out on the whole system.
 820 */
 821static int count_process_nodes(int process_nr)
 822{
 823	char node_present[MAX_NR_NODES] = { 0, };
 824	int nodes;
 825	int n, t;
 826
 
 
 
 
 
 827	for (t = 0; t < g->p.nr_threads; t++) {
 828		struct thread_data *td;
 829		int task_nr;
 830		int node;
 831
 832		task_nr = process_nr*g->p.nr_threads + t;
 833		td = g->threads + task_nr;
 834
 835		node = numa_node_of_cpu(td->curr_cpu);
 836		if (node < 0) /* curr_cpu was likely still -1 */
 
 837			return 0;
 
 838
 839		node_present[node] = 1;
 840	}
 841
 842	nodes = 0;
 843
 844	for (n = 0; n < MAX_NR_NODES; n++)
 845		nodes += node_present[n];
 846
 
 847	return nodes;
 848}
 849
 850/*
 851 * Count the number of distinct process-threads a node contains.
 852 *
 853 * A count of 1 means that the node contains only a single
 854 * process. If all nodes on the system contain at most one
 855 * process then we are well-converged.
 856 */
 857static int count_node_processes(int node)
 858{
 859	int processes = 0;
 860	int t, p;
 861
 862	for (p = 0; p < g->p.nr_proc; p++) {
 863		for (t = 0; t < g->p.nr_threads; t++) {
 864			struct thread_data *td;
 865			int task_nr;
 866			int n;
 867
 868			task_nr = p*g->p.nr_threads + t;
 869			td = g->threads + task_nr;
 870
 871			n = numa_node_of_cpu(td->curr_cpu);
 872			if (n == node) {
 873				processes++;
 874				break;
 875			}
 876		}
 877	}
 878
 879	return processes;
 880}
 881
 882static void calc_convergence_compression(int *strong)
 883{
 884	unsigned int nodes_min, nodes_max;
 885	int p;
 886
 887	nodes_min = -1;
 888	nodes_max =  0;
 889
 890	for (p = 0; p < g->p.nr_proc; p++) {
 891		unsigned int nodes = count_process_nodes(p);
 892
 893		if (!nodes) {
 894			*strong = 0;
 895			return;
 896		}
 897
 898		nodes_min = min(nodes, nodes_min);
 899		nodes_max = max(nodes, nodes_max);
 900	}
 901
 902	/* Strong convergence: all threads compress on a single node: */
 903	if (nodes_min == 1 && nodes_max == 1) {
 904		*strong = 1;
 905	} else {
 906		*strong = 0;
 907		tprintf(" {%d-%d}", nodes_min, nodes_max);
 908	}
 909}
 910
 911static void calc_convergence(double runtime_ns_max, double *convergence)
 912{
 913	unsigned int loops_done_min, loops_done_max;
 914	int process_groups;
 915	int nodes[MAX_NR_NODES];
 916	int distance;
 917	int nr_min;
 918	int nr_max;
 919	int strong;
 920	int sum;
 921	int nr;
 922	int node;
 923	int cpu;
 924	int t;
 925
 926	if (!g->p.show_convergence && !g->p.measure_convergence)
 927		return;
 928
 
 
 929	for (node = 0; node < g->p.nr_nodes; node++)
 930		nodes[node] = 0;
 931
 932	loops_done_min = -1;
 933	loops_done_max = 0;
 934
 935	for (t = 0; t < g->p.nr_tasks; t++) {
 936		struct thread_data *td = g->threads + t;
 937		unsigned int loops_done;
 938
 939		cpu = td->curr_cpu;
 940
 941		/* Not all threads have written it yet: */
 942		if (cpu < 0)
 943			continue;
 944
 945		node = numa_node_of_cpu(cpu);
 946
 947		nodes[node]++;
 948
 949		loops_done = td->loops_done;
 950		loops_done_min = min(loops_done, loops_done_min);
 951		loops_done_max = max(loops_done, loops_done_max);
 952	}
 953
 954	nr_max = 0;
 955	nr_min = g->p.nr_tasks;
 956	sum = 0;
 957
 958	for (node = 0; node < g->p.nr_nodes; node++) {
 
 
 959		nr = nodes[node];
 960		nr_min = min(nr, nr_min);
 961		nr_max = max(nr, nr_max);
 962		sum += nr;
 963	}
 964	BUG_ON(nr_min > nr_max);
 965
 966	BUG_ON(sum > g->p.nr_tasks);
 967
 968	if (0 && (sum < g->p.nr_tasks))
 
 969		return;
 
 970
 971	/*
 972	 * Count the number of distinct process groups present
 973	 * on nodes - when we are converged this will decrease
 974	 * to g->p.nr_proc:
 975	 */
 976	process_groups = 0;
 977
 978	for (node = 0; node < g->p.nr_nodes; node++) {
 979		int processes = count_node_processes(node);
 980
 
 
 
 981		nr = nodes[node];
 982		tprintf(" %2d/%-2d", nr, processes);
 983
 984		process_groups += processes;
 985	}
 986
 987	distance = nr_max - nr_min;
 988
 989	tprintf(" [%2d/%-2d]", distance, process_groups);
 990
 991	tprintf(" l:%3d-%-3d (%3d)",
 992		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 993
 994	if (loops_done_min && loops_done_max) {
 995		double skew = 1.0 - (double)loops_done_min/loops_done_max;
 996
 997		tprintf(" [%4.1f%%]", skew * 100.0);
 998	}
 999
1000	calc_convergence_compression(&strong);
1001
1002	if (strong && process_groups == g->p.nr_proc) {
1003		if (!*convergence) {
1004			*convergence = runtime_ns_max;
1005			tprintf(" (%6.1fs converged)\n", *convergence/1e9);
1006			if (g->p.measure_convergence) {
1007				g->all_converged = true;
1008				g->stop_work = true;
1009			}
1010		}
1011	} else {
1012		if (*convergence) {
1013			tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
1014			*convergence = 0;
1015		}
1016		tprintf("\n");
1017	}
 
 
1018}
1019
1020static void show_summary(double runtime_ns_max, int l, double *convergence)
1021{
1022	tprintf("\r #  %5.1f%%  [%.1f mins]",
1023		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1024
1025	calc_convergence(runtime_ns_max, convergence);
1026
1027	if (g->p.show_details >= 0)
1028		fflush(stdout);
1029}
1030
1031static void *worker_thread(void *__tdata)
1032{
1033	struct thread_data *td = __tdata;
1034	struct timeval start0, start, stop, diff;
1035	int process_nr = td->process_nr;
1036	int thread_nr = td->thread_nr;
1037	unsigned long last_perturbance;
1038	int task_nr = td->task_nr;
1039	int details = g->p.show_details;
1040	int first_task, last_task;
1041	double convergence = 0;
1042	u64 val = td->val;
1043	double runtime_ns_max;
1044	u8 *global_data;
1045	u8 *process_data;
1046	u8 *thread_data;
1047	u64 bytes_done;
1048	long work_done;
1049	u32 l;
1050	struct rusage rusage;
1051
1052	bind_to_cpumask(td->bind_cpumask);
1053	bind_to_memnode(td->bind_node);
1054
1055	set_taskname("thread %d/%d", process_nr, thread_nr);
1056
1057	global_data = g->data;
1058	process_data = td->process_data;
1059	thread_data = setup_private_data(g->p.bytes_thread);
1060
1061	bytes_done = 0;
1062
1063	last_task = 0;
1064	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1065		last_task = 1;
1066
1067	first_task = 0;
1068	if (process_nr == 0 && thread_nr == 0)
1069		first_task = 1;
1070
1071	if (details >= 2) {
1072		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1073			process_nr, thread_nr, global_data, process_data, thread_data);
1074	}
1075
1076	if (g->p.serialize_startup) {
1077		pthread_mutex_lock(&g->startup_mutex);
1078		g->nr_tasks_started++;
1079		pthread_mutex_unlock(&g->startup_mutex);
 
 
 
 
1080
1081		/* Here we will wait for the main process to start us all at once: */
1082		pthread_mutex_lock(&g->start_work_mutex);
 
1083		g->nr_tasks_working++;
 
 
1084
1085		/* Last one wake the main process: */
1086		if (g->nr_tasks_working == g->p.nr_tasks)
1087			pthread_mutex_unlock(&g->startup_done_mutex);
1088
1089		pthread_mutex_unlock(&g->start_work_mutex);
1090	}
1091
1092	gettimeofday(&start0, NULL);
1093
1094	start = stop = start0;
1095	last_perturbance = start.tv_sec;
1096
1097	for (l = 0; l < g->p.nr_loops; l++) {
1098		start = stop;
1099
1100		if (g->stop_work)
1101			break;
1102
1103		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1104		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1105		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1106
1107		if (g->p.sleep_usecs) {
1108			pthread_mutex_lock(td->process_lock);
1109			usleep(g->p.sleep_usecs);
1110			pthread_mutex_unlock(td->process_lock);
1111		}
1112		/*
1113		 * Amount of work to be done under a process-global lock:
1114		 */
1115		if (g->p.bytes_process_locked) {
1116			pthread_mutex_lock(td->process_lock);
1117			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1118			pthread_mutex_unlock(td->process_lock);
1119		}
1120
1121		work_done = g->p.bytes_global + g->p.bytes_process +
1122			    g->p.bytes_process_locked + g->p.bytes_thread;
1123
1124		update_curr_cpu(task_nr, work_done);
1125		bytes_done += work_done;
1126
1127		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1128			continue;
1129
1130		td->loops_done = l;
1131
1132		gettimeofday(&stop, NULL);
1133
1134		/* Check whether our max runtime timed out: */
1135		if (g->p.nr_secs) {
1136			timersub(&stop, &start0, &diff);
1137			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1138				g->stop_work = true;
1139				break;
1140			}
1141		}
1142
1143		/* Update the summary at most once per second: */
1144		if (start.tv_sec == stop.tv_sec)
1145			continue;
1146
1147		/*
1148		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1149		 * by migrating to CPU#0:
1150		 */
1151		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1152			cpu_set_t orig_mask;
1153			int target_cpu;
1154			int this_cpu;
1155
1156			last_perturbance = stop.tv_sec;
1157
1158			/*
1159			 * Depending on where we are running, move into
1160			 * the other half of the system, to create some
1161			 * real disturbance:
1162			 */
1163			this_cpu = g->threads[task_nr].curr_cpu;
1164			if (this_cpu < g->p.nr_cpus/2)
1165				target_cpu = g->p.nr_cpus-1;
1166			else
1167				target_cpu = 0;
1168
1169			orig_mask = bind_to_cpu(target_cpu);
1170
1171			/* Here we are running on the target CPU already */
1172			if (details >= 1)
1173				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1174
1175			bind_to_cpumask(orig_mask);
 
1176		}
1177
1178		if (details >= 3) {
1179			timersub(&stop, &start, &diff);
1180			runtime_ns_max = diff.tv_sec * 1000000000;
1181			runtime_ns_max += diff.tv_usec * 1000;
1182
1183			if (details >= 0) {
1184				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1185					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1186			}
1187			fflush(stdout);
1188		}
1189		if (!last_task)
1190			continue;
1191
1192		timersub(&stop, &start0, &diff);
1193		runtime_ns_max = diff.tv_sec * 1000000000ULL;
1194		runtime_ns_max += diff.tv_usec * 1000ULL;
1195
1196		show_summary(runtime_ns_max, l, &convergence);
1197	}
1198
1199	gettimeofday(&stop, NULL);
1200	timersub(&stop, &start0, &diff);
1201	td->runtime_ns = diff.tv_sec * 1000000000ULL;
1202	td->runtime_ns += diff.tv_usec * 1000ULL;
1203	td->speed_gbs = bytes_done / (td->runtime_ns / 1e9) / 1e9;
 
1204
1205	getrusage(RUSAGE_THREAD, &rusage);
1206	td->system_time_ns = rusage.ru_stime.tv_sec * 1000000000ULL;
1207	td->system_time_ns += rusage.ru_stime.tv_usec * 1000ULL;
1208	td->user_time_ns = rusage.ru_utime.tv_sec * 1000000000ULL;
1209	td->user_time_ns += rusage.ru_utime.tv_usec * 1000ULL;
1210
1211	free_data(thread_data, g->p.bytes_thread);
1212
1213	pthread_mutex_lock(&g->stop_work_mutex);
1214	g->bytes_done += bytes_done;
1215	pthread_mutex_unlock(&g->stop_work_mutex);
1216
1217	return NULL;
1218}
1219
1220/*
1221 * A worker process starts a couple of threads:
1222 */
1223static void worker_process(int process_nr)
1224{
1225	pthread_mutex_t process_lock;
1226	struct thread_data *td;
1227	pthread_t *pthreads;
1228	u8 *process_data;
1229	int task_nr;
1230	int ret;
1231	int t;
1232
1233	pthread_mutex_init(&process_lock, NULL);
1234	set_taskname("process %d", process_nr);
1235
1236	/*
1237	 * Pick up the memory policy and the CPU binding of our first thread,
1238	 * so that we initialize memory accordingly:
1239	 */
1240	task_nr = process_nr*g->p.nr_threads;
1241	td = g->threads + task_nr;
1242
1243	bind_to_memnode(td->bind_node);
1244	bind_to_cpumask(td->bind_cpumask);
1245
1246	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1247	process_data = setup_private_data(g->p.bytes_process);
1248
1249	if (g->p.show_details >= 3) {
1250		printf(" # process %2d global mem: %p, process mem: %p\n",
1251			process_nr, g->data, process_data);
1252	}
1253
1254	for (t = 0; t < g->p.nr_threads; t++) {
1255		task_nr = process_nr*g->p.nr_threads + t;
1256		td = g->threads + task_nr;
1257
1258		td->process_data = process_data;
1259		td->process_nr   = process_nr;
1260		td->thread_nr    = t;
1261		td->task_nr	 = task_nr;
1262		td->val          = rand();
1263		td->curr_cpu	 = -1;
1264		td->process_lock = &process_lock;
1265
1266		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1267		BUG_ON(ret);
1268	}
1269
1270	for (t = 0; t < g->p.nr_threads; t++) {
1271                ret = pthread_join(pthreads[t], NULL);
1272		BUG_ON(ret);
1273	}
1274
1275	free_data(process_data, g->p.bytes_process);
1276	free(pthreads);
1277}
1278
1279static void print_summary(void)
1280{
1281	if (g->p.show_details < 0)
1282		return;
1283
1284	printf("\n ###\n");
1285	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1286		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1287	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1288			g->p.nr_loops, g->p.bytes_global/1024/1024);
1289	printf(" #      %5dx %5ldMB process shared mem operations\n",
1290			g->p.nr_loops, g->p.bytes_process/1024/1024);
1291	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1292			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1293
1294	printf(" ###\n");
1295
1296	printf("\n ###\n"); fflush(stdout);
1297}
1298
1299static void init_thread_data(void)
1300{
1301	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1302	int t;
1303
1304	g->threads = zalloc_shared_data(size);
1305
1306	for (t = 0; t < g->p.nr_tasks; t++) {
1307		struct thread_data *td = g->threads + t;
 
1308		int cpu;
1309
1310		/* Allow all nodes by default: */
1311		td->bind_node = -1;
1312
1313		/* Allow all CPUs by default: */
1314		CPU_ZERO(&td->bind_cpumask);
 
 
1315		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1316			CPU_SET(cpu, &td->bind_cpumask);
1317	}
1318}
1319
1320static void deinit_thread_data(void)
1321{
1322	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
 
 
 
 
 
 
 
1323
1324	free_data(g->threads, size);
1325}
1326
1327static int init(void)
1328{
1329	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1330
1331	/* Copy over options: */
1332	g->p = p0;
1333
1334	g->p.nr_cpus = numa_num_configured_cpus();
1335
1336	g->p.nr_nodes = numa_max_node() + 1;
1337
1338	/* char array in count_process_nodes(): */
1339	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1340
1341	if (g->p.show_quiet && !g->p.show_details)
1342		g->p.show_details = -1;
1343
1344	/* Some memory should be specified: */
1345	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1346		return -1;
1347
1348	if (g->p.mb_global_str) {
1349		g->p.mb_global = atof(g->p.mb_global_str);
1350		BUG_ON(g->p.mb_global < 0);
1351	}
1352
1353	if (g->p.mb_proc_str) {
1354		g->p.mb_proc = atof(g->p.mb_proc_str);
1355		BUG_ON(g->p.mb_proc < 0);
1356	}
1357
1358	if (g->p.mb_proc_locked_str) {
1359		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1360		BUG_ON(g->p.mb_proc_locked < 0);
1361		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1362	}
1363
1364	if (g->p.mb_thread_str) {
1365		g->p.mb_thread = atof(g->p.mb_thread_str);
1366		BUG_ON(g->p.mb_thread < 0);
1367	}
1368
1369	BUG_ON(g->p.nr_threads <= 0);
1370	BUG_ON(g->p.nr_proc <= 0);
1371
1372	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1373
1374	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1375	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1376	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1377	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1378
1379	g->data = setup_shared_data(g->p.bytes_global);
1380
1381	/* Startup serialization: */
1382	init_global_mutex(&g->start_work_mutex);
1383	init_global_mutex(&g->startup_mutex);
1384	init_global_mutex(&g->startup_done_mutex);
1385	init_global_mutex(&g->stop_work_mutex);
 
1386
1387	init_thread_data();
1388
1389	tprintf("#\n");
1390	if (parse_setup_cpu_list() || parse_setup_node_list())
1391		return -1;
1392	tprintf("#\n");
1393
1394	print_summary();
1395
1396	return 0;
1397}
1398
1399static void deinit(void)
1400{
1401	free_data(g->data, g->p.bytes_global);
1402	g->data = NULL;
1403
1404	deinit_thread_data();
1405
1406	free_data(g, sizeof(*g));
1407	g = NULL;
1408}
1409
1410/*
1411 * Print a short or long result, depending on the verbosity setting:
1412 */
1413static void print_res(const char *name, double val,
1414		      const char *txt_unit, const char *txt_short, const char *txt_long)
1415{
1416	if (!name)
1417		name = "main,";
1418
1419	if (!g->p.show_quiet)
1420		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1421	else
1422		printf(" %14.3f %s\n", val, txt_long);
1423}
1424
1425static int __bench_numa(const char *name)
1426{
1427	struct timeval start, stop, diff;
1428	u64 runtime_ns_min, runtime_ns_sum;
1429	pid_t *pids, pid, wpid;
1430	double delta_runtime;
1431	double runtime_avg;
1432	double runtime_sec_max;
1433	double runtime_sec_min;
1434	int wait_stat;
1435	double bytes;
1436	int i, t, p;
1437
1438	if (init())
1439		return -1;
1440
1441	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1442	pid = -1;
1443
1444	/* All threads try to acquire it, this way we can wait for them to start up: */
1445	pthread_mutex_lock(&g->start_work_mutex);
1446
1447	if (g->p.serialize_startup) {
1448		tprintf(" #\n");
1449		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1450	}
1451
1452	gettimeofday(&start, NULL);
1453
1454	for (i = 0; i < g->p.nr_proc; i++) {
1455		pid = fork();
1456		dprintf(" # process %2d: PID %d\n", i, pid);
1457
1458		BUG_ON(pid < 0);
1459		if (!pid) {
1460			/* Child process: */
1461			worker_process(i);
1462
1463			exit(0);
1464		}
1465		pids[i] = pid;
1466
1467	}
1468	/* Wait for all the threads to start up: */
1469	while (g->nr_tasks_started != g->p.nr_tasks)
1470		usleep(1000);
1471
1472	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1473
1474	if (g->p.serialize_startup) {
 
1475		double startup_sec;
1476
1477		pthread_mutex_lock(&g->startup_done_mutex);
1478
1479		/* This will start all threads: */
1480		pthread_mutex_unlock(&g->start_work_mutex);
1481
1482		/* This mutex is locked - the last started thread will wake us: */
1483		pthread_mutex_lock(&g->startup_done_mutex);
 
 
 
 
 
 
 
 
 
 
 
1484
1485		gettimeofday(&stop, NULL);
1486
1487		timersub(&stop, &start, &diff);
1488
1489		startup_sec = diff.tv_sec * 1000000000.0;
1490		startup_sec += diff.tv_usec * 1000.0;
1491		startup_sec /= 1e9;
1492
1493		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1494		tprintf(" #\n");
1495
1496		start = stop;
1497		pthread_mutex_unlock(&g->startup_done_mutex);
 
 
 
 
1498	} else {
1499		gettimeofday(&start, NULL);
1500	}
1501
1502	/* Parent process: */
1503
1504
1505	for (i = 0; i < g->p.nr_proc; i++) {
1506		wpid = waitpid(pids[i], &wait_stat, 0);
1507		BUG_ON(wpid < 0);
1508		BUG_ON(!WIFEXITED(wait_stat));
1509
1510	}
1511
1512	runtime_ns_sum = 0;
1513	runtime_ns_min = -1LL;
1514
1515	for (t = 0; t < g->p.nr_tasks; t++) {
1516		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1517
1518		runtime_ns_sum += thread_runtime_ns;
1519		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1520	}
1521
1522	gettimeofday(&stop, NULL);
1523	timersub(&stop, &start, &diff);
1524
1525	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1526
1527	tprintf("\n ###\n");
1528	tprintf("\n");
1529
1530	runtime_sec_max = diff.tv_sec * 1000000000.0;
1531	runtime_sec_max += diff.tv_usec * 1000.0;
1532	runtime_sec_max /= 1e9;
1533
1534	runtime_sec_min = runtime_ns_min/1e9;
1535
1536	bytes = g->bytes_done;
1537	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1538
1539	if (g->p.measure_convergence) {
1540		print_res(name, runtime_sec_max,
1541			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1542	}
1543
1544	print_res(name, runtime_sec_max,
1545		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1546
1547	print_res(name, runtime_sec_min,
1548		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1549
1550	print_res(name, runtime_avg,
1551		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1552
1553	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1554	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1555		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1556
1557	print_res(name, bytes / g->p.nr_tasks / 1e9,
1558		"GB,", "data/thread",		"GB data processed, per thread");
1559
1560	print_res(name, bytes / 1e9,
1561		"GB,", "data-total",		"GB data processed, total");
1562
1563	print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1564		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1565
1566	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1567		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1568
1569	print_res(name, bytes / runtime_sec_max / 1e9,
1570		"GB/sec,", "total-speed",	"GB/sec total speed");
1571
1572	if (g->p.show_details >= 2) {
1573		char tname[32];
1574		struct thread_data *td;
1575		for (p = 0; p < g->p.nr_proc; p++) {
1576			for (t = 0; t < g->p.nr_threads; t++) {
1577				memset(tname, 0, 32);
1578				td = g->threads + p*g->p.nr_threads + t;
1579				snprintf(tname, 32, "process%d:thread%d", p, t);
1580				print_res(tname, td->speed_gbs,
1581					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1582				print_res(tname, td->system_time_ns / 1e9,
1583					"secs",	"thread-system-time", "system CPU time/thread");
1584				print_res(tname, td->user_time_ns / 1e9,
1585					"secs",	"thread-user-time", "user CPU time/thread");
1586			}
1587		}
1588	}
1589
1590	free(pids);
1591
1592	deinit();
1593
1594	return 0;
1595}
1596
1597#define MAX_ARGS 50
1598
1599static int command_size(const char **argv)
1600{
1601	int size = 0;
1602
1603	while (*argv) {
1604		size++;
1605		argv++;
1606	}
1607
1608	BUG_ON(size >= MAX_ARGS);
1609
1610	return size;
1611}
1612
1613static void init_params(struct params *p, const char *name, int argc, const char **argv)
1614{
1615	int i;
1616
1617	printf("\n # Running %s \"perf bench numa", name);
1618
1619	for (i = 0; i < argc; i++)
1620		printf(" %s", argv[i]);
1621
1622	printf("\"\n");
1623
1624	memset(p, 0, sizeof(*p));
1625
1626	/* Initialize nonzero defaults: */
1627
1628	p->serialize_startup		= 1;
1629	p->data_reads			= true;
1630	p->data_writes			= true;
1631	p->data_backwards		= true;
1632	p->data_rand_walk		= true;
1633	p->nr_loops			= -1;
1634	p->init_random			= true;
1635	p->mb_global_str		= "1";
1636	p->nr_proc			= 1;
1637	p->nr_threads			= 1;
1638	p->nr_secs			= 5;
1639	p->run_all			= argc == 1;
1640}
1641
1642static int run_bench_numa(const char *name, const char **argv)
1643{
1644	int argc = command_size(argv);
1645
1646	init_params(&p0, name, argc, argv);
1647	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1648	if (argc)
1649		goto err;
1650
1651	if (__bench_numa(name))
1652		goto err;
1653
1654	return 0;
1655
1656err:
1657	return -1;
1658}
1659
1660#define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1661#define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1662
1663#define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1664#define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1665
1666#define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1667#define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1668
1669/*
1670 * The built-in test-suite executed by "perf bench numa -a".
1671 *
1672 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1673 */
1674static const char *tests[][MAX_ARGS] = {
1675   /* Basic single-stream NUMA bandwidth measurements: */
1676   { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1677			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1678   { "RAM-bw-local-NOTHP,",
1679			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1680			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1681   { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1682			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1683
1684   /* 2-stream NUMA bandwidth measurements: */
1685   { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1686			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1687   { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1688		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1689
1690   /* Cross-stream NUMA bandwidth measurement: */
1691   { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1692		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1693
1694   /* Convergence latency measurements: */
1695   { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1696   { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1697   { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1698   { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1699   { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1700   { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1701   { " 4x4-convergence-NOTHP,",
1702			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1703   { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1704   { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1705   { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1706   { " 8x4-convergence-NOTHP,",
1707			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1708   { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1709   { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1710   { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1711   { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1712   { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1713
1714   /* Various NUMA process/thread layout bandwidth measurements: */
1715   { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1716   { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1717   { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1718   { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1719   { " 8x1-bw-process-NOTHP,",
1720			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1721   { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1722
1723   { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1724   { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1725   { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1726   { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1727
1728   { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1729   { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1730   { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1731   { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1732   { " 4x8-bw-thread-NOTHP,",
1733			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1734   { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1735   { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1736
1737   { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1738   { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1739
1740   { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1741   { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1742   { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1743   { "numa01-bw-thread-NOTHP,",
1744			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1745};
1746
1747static int bench_all(void)
1748{
1749	int nr = ARRAY_SIZE(tests);
1750	int ret;
1751	int i;
1752
1753	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1754	BUG_ON(ret < 0);
1755
1756	for (i = 0; i < nr; i++) {
1757		run_bench_numa(tests[i][0], tests[i] + 1);
1758	}
1759
1760	printf("\n");
1761
1762	return 0;
1763}
1764
1765int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1766{
1767	init_params(&p0, "main,", argc, argv);
1768	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1769	if (argc)
1770		goto err;
1771
1772	if (p0.run_all)
1773		return bench_all();
1774
1775	if (__bench_numa(NULL))
1776		goto err;
1777
1778	return 0;
1779
1780err:
1781	usage_with_options(numa_usage, options);
1782	return -1;
1783}
v6.13.7
   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 <debug.h>
  20#include <malloc.h>
  21#include <signal.h>
  22#include <stdlib.h>
  23#include <string.h>
  24#include <unistd.h>
 
  25#include <sys/mman.h>
  26#include <sys/time.h>
  27#include <sys/resource.h>
  28#include <sys/wait.h>
  29#include <sys/prctl.h>
  30#include <sys/stat.h>
  31#include <sys/types.h>
  32#include <linux/kernel.h>
  33#include <linux/time64.h>
  34#include <linux/numa.h>
  35#include <linux/zalloc.h>
  36
  37#include "../util/header.h"
  38#include "../util/mutex.h"
  39#include <api/fs/fs.h>
  40#include <numa.h>
  41#include <numaif.h>
  42
  43#ifndef RUSAGE_THREAD
  44# define RUSAGE_THREAD 1
  45#endif
  46
  47/*
  48 * Regular printout to the terminal, suppressed if -q is specified:
  49 */
  50#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  51
  52/*
  53 * Debug printf:
  54 */
  55#undef dprintf
  56#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  57
  58struct thread_data {
  59	int			curr_cpu;
  60	cpu_set_t		*bind_cpumask;
  61	int			bind_node;
  62	u8			*process_data;
  63	int			process_nr;
  64	int			thread_nr;
  65	int			task_nr;
  66	unsigned int		loops_done;
  67	u64			val;
  68	u64			runtime_ns;
  69	u64			system_time_ns;
  70	u64			user_time_ns;
  71	double			speed_gbs;
  72	struct mutex		*process_lock;
  73};
  74
  75/* Parameters set by options: */
  76
  77struct params {
  78	/* Startup synchronization: */
  79	bool			serialize_startup;
  80
  81	/* Task hierarchy: */
  82	int			nr_proc;
  83	int			nr_threads;
  84
  85	/* Working set sizes: */
  86	const char		*mb_global_str;
  87	const char		*mb_proc_str;
  88	const char		*mb_proc_locked_str;
  89	const char		*mb_thread_str;
  90
  91	double			mb_global;
  92	double			mb_proc;
  93	double			mb_proc_locked;
  94	double			mb_thread;
  95
  96	/* Access patterns to the working set: */
  97	bool			data_reads;
  98	bool			data_writes;
  99	bool			data_backwards;
 100	bool			data_zero_memset;
 101	bool			data_rand_walk;
 102	u32			nr_loops;
 103	u32			nr_secs;
 104	u32			sleep_usecs;
 105
 106	/* Working set initialization: */
 107	bool			init_zero;
 108	bool			init_random;
 109	bool			init_cpu0;
 110
 111	/* Misc options: */
 112	int			show_details;
 113	int			run_all;
 114	int			thp;
 115
 116	long			bytes_global;
 117	long			bytes_process;
 118	long			bytes_process_locked;
 119	long			bytes_thread;
 120
 121	int			nr_tasks;
 
 122
 123	bool			show_convergence;
 124	bool			measure_convergence;
 125
 126	int			perturb_secs;
 127	int			nr_cpus;
 128	int			nr_nodes;
 129
 130	/* Affinity options -C and -N: */
 131	char			*cpu_list_str;
 132	char			*node_list_str;
 133};
 134
 135
 136/* Global, read-writable area, accessible to all processes and threads: */
 137
 138struct global_info {
 139	u8			*data;
 140
 141	struct mutex		startup_mutex;
 142	struct cond		startup_cond;
 143	int			nr_tasks_started;
 144
 145	struct mutex		start_work_mutex;
 146	struct cond		start_work_cond;
 
 147	int			nr_tasks_working;
 148	bool			start_work;
 149
 150	struct mutex		stop_work_mutex;
 151	u64			bytes_done;
 152
 153	struct thread_data	*threads;
 154
 155	/* Convergence latency measurement: */
 156	bool			all_converged;
 157	bool			stop_work;
 158
 159	int			print_once;
 160
 161	struct params		p;
 162};
 163
 164static struct global_info	*g = NULL;
 165
 166static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 167static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 168
 169struct params p0;
 170
 171static const struct option options[] = {
 172	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
 173	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
 174
 175	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
 176	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
 177	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 178	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
 179
 180	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
 181	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
 182	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
 183
 184	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via reads (can be mixed with -W)"),
 185	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
 186	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
 187	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 188	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
 189
 190
 191	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
 192	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
 193	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
 194	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
 195
 196	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
 197	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
 198	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 199	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
 200		    "convergence is reached when each process (all its threads) is running on a single NUMA node."),
 201	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
 202	OPT_BOOLEAN('q', "quiet"	, &quiet,
 203		    "quiet mode (do not show any warnings or messages)"),
 204	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 205
 206	/* Special option string parsing callbacks: */
 207        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 208			"bind the first N tasks to these specific cpus (the rest is unbound)",
 209			parse_cpus_opt),
 210        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 211			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
 212			parse_nodes_opt),
 213	OPT_END()
 214};
 215
 216static const char * const bench_numa_usage[] = {
 217	"perf bench numa <options>",
 218	NULL
 219};
 220
 221static const char * const numa_usage[] = {
 222	"perf bench numa mem [<options>]",
 223	NULL
 224};
 225
 226/*
 227 * To get number of numa nodes present.
 228 */
 229static int nr_numa_nodes(void)
 230{
 231	int i, nr_nodes = 0;
 
 232
 233	for (i = 0; i < g->p.nr_nodes; i++) {
 234		if (numa_bitmask_isbitset(numa_nodes_ptr, i))
 235			nr_nodes++;
 236	}
 237
 238	return nr_nodes;
 239}
 240
 241/*
 242 * To check if given numa node is present.
 243 */
 244static int is_node_present(int node)
 245{
 246	return numa_bitmask_isbitset(numa_nodes_ptr, node);
 247}
 248
 249/*
 250 * To check given numa node has cpus.
 251 */
 252static bool node_has_cpus(int node)
 253{
 254	struct bitmask *cpumask = numa_allocate_cpumask();
 255	bool ret = false; /* fall back to nocpus */
 256	int cpu;
 257
 258	BUG_ON(!cpumask);
 259	if (!numa_node_to_cpus(node, cpumask)) {
 260		for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 261			if (numa_bitmask_isbitset(cpumask, cpu)) {
 262				ret = true;
 263				break;
 264			}
 265		}
 266	}
 267	numa_free_cpumask(cpumask);
 268
 269	return ret;
 270}
 271
 272static cpu_set_t *bind_to_cpu(int target_cpu)
 273{
 274	int nrcpus = numa_num_possible_cpus();
 275	cpu_set_t *orig_mask, *mask;
 276	size_t size;
 277
 278	orig_mask = CPU_ALLOC(nrcpus);
 279	BUG_ON(!orig_mask);
 280	size = CPU_ALLOC_SIZE(nrcpus);
 281	CPU_ZERO_S(size, orig_mask);
 282
 283	if (sched_getaffinity(0, size, orig_mask))
 284		goto err_out;
 285
 286	mask = CPU_ALLOC(nrcpus);
 287	if (!mask)
 288		goto err_out;
 289
 290	CPU_ZERO_S(size, mask);
 291
 292	if (target_cpu == -1) {
 293		int cpu;
 294
 295		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 296			CPU_SET_S(cpu, size, mask);
 297	} else {
 298		if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
 299			goto err;
 300
 301		CPU_SET_S(target_cpu, size, mask);
 302	}
 303
 304	if (sched_setaffinity(0, size, mask))
 305		goto err;
 306
 307	return orig_mask;
 308
 309err:
 310	CPU_FREE(mask);
 311err_out:
 312	CPU_FREE(orig_mask);
 313
 314	/* BUG_ON due to failure in allocation of orig_mask/mask */
 315	BUG_ON(-1);
 316	return NULL;
 317}
 318
 319static cpu_set_t *bind_to_node(int target_node)
 320{
 321	int nrcpus = numa_num_possible_cpus();
 322	size_t size;
 323	cpu_set_t *orig_mask, *mask;
 324	int cpu;
 
 325
 326	orig_mask = CPU_ALLOC(nrcpus);
 327	BUG_ON(!orig_mask);
 328	size = CPU_ALLOC_SIZE(nrcpus);
 329	CPU_ZERO_S(size, orig_mask);
 330
 331	if (sched_getaffinity(0, size, orig_mask))
 332		goto err_out;
 333
 334	mask = CPU_ALLOC(nrcpus);
 335	if (!mask)
 336		goto err_out;
 337
 338	CPU_ZERO_S(size, mask);
 339
 340	if (target_node == NUMA_NO_NODE) {
 341		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 342			CPU_SET_S(cpu, size, mask);
 343	} else {
 344		struct bitmask *cpumask = numa_allocate_cpumask();
 
 345
 346		if (!cpumask)
 347			goto err;
 348
 349		if (!numa_node_to_cpus(target_node, cpumask)) {
 350			for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
 351				if (numa_bitmask_isbitset(cpumask, cpu))
 352					CPU_SET_S(cpu, size, mask);
 353			}
 354		}
 355		numa_free_cpumask(cpumask);
 356	}
 357
 358	if (sched_setaffinity(0, size, mask))
 359		goto err;
 360
 361	return orig_mask;
 362
 363err:
 364	CPU_FREE(mask);
 365err_out:
 366	CPU_FREE(orig_mask);
 367
 368	/* BUG_ON due to failure in allocation of orig_mask/mask */
 369	BUG_ON(-1);
 370	return NULL;
 371}
 372
 373static void bind_to_cpumask(cpu_set_t *mask)
 374{
 375	int ret;
 376	size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
 377
 378	ret = sched_setaffinity(0, size, mask);
 379	if (ret) {
 380		CPU_FREE(mask);
 381		BUG_ON(ret);
 382	}
 383}
 384
 385static void mempol_restore(void)
 386{
 387	int ret;
 388
 389	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 390
 391	BUG_ON(ret);
 392}
 393
 394static void bind_to_memnode(int node)
 395{
 396	struct bitmask *node_mask;
 397	int ret;
 398
 399	if (node == NUMA_NO_NODE)
 400		return;
 401
 402	node_mask = numa_allocate_nodemask();
 403	BUG_ON(!node_mask);
 404
 405	numa_bitmask_clearall(node_mask);
 406	numa_bitmask_setbit(node_mask, node);
 407
 408	ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
 409	dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
 410
 411	numa_bitmask_free(node_mask);
 412	BUG_ON(ret);
 413}
 414
 415#define HPSIZE (2*1024*1024)
 416
 417#define set_taskname(fmt...)				\
 418do {							\
 419	char name[20];					\
 420							\
 421	snprintf(name, 20, fmt);			\
 422	prctl(PR_SET_NAME, name);			\
 423} while (0)
 424
 425static u8 *alloc_data(ssize_t bytes0, int map_flags,
 426		      int init_zero, int init_cpu0, int thp, int init_random)
 427{
 428	cpu_set_t *orig_mask = NULL;
 429	ssize_t bytes;
 430	u8 *buf;
 431	int ret;
 432
 433	if (!bytes0)
 434		return NULL;
 435
 436	/* Allocate and initialize all memory on CPU#0: */
 437	if (init_cpu0) {
 438		int node = numa_node_of_cpu(0);
 439
 440		orig_mask = bind_to_node(node);
 441		bind_to_memnode(node);
 442	}
 443
 444	bytes = bytes0 + HPSIZE;
 445
 446	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 447	BUG_ON(buf == (void *)-1);
 448
 449	if (map_flags == MAP_PRIVATE) {
 450		if (thp > 0) {
 451			ret = madvise(buf, bytes, MADV_HUGEPAGE);
 452			if (ret && !g->print_once) {
 453				g->print_once = 1;
 454				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 455			}
 456		}
 457		if (thp < 0) {
 458			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 459			if (ret && !g->print_once) {
 460				g->print_once = 1;
 461				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 462			}
 463		}
 464	}
 465
 466	if (init_zero) {
 467		bzero(buf, bytes);
 468	} else {
 469		/* Initialize random contents, different in each word: */
 470		if (init_random) {
 471			u64 *wbuf = (void *)buf;
 472			long off = rand();
 473			long i;
 474
 475			for (i = 0; i < bytes/8; i++)
 476				wbuf[i] = i + off;
 477		}
 478	}
 479
 480	/* Align to 2MB boundary: */
 481	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 482
 483	/* Restore affinity: */
 484	if (init_cpu0) {
 485		bind_to_cpumask(orig_mask);
 486		CPU_FREE(orig_mask);
 487		mempol_restore();
 488	}
 489
 490	return buf;
 491}
 492
 493static void free_data(void *data, ssize_t bytes)
 494{
 495	int ret;
 496
 497	if (!data)
 498		return;
 499
 500	ret = munmap(data, bytes);
 501	BUG_ON(ret);
 502}
 503
 504/*
 505 * Create a shared memory buffer that can be shared between processes, zeroed:
 506 */
 507static void * zalloc_shared_data(ssize_t bytes)
 508{
 509	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 510}
 511
 512/*
 513 * Create a shared memory buffer that can be shared between processes:
 514 */
 515static void * setup_shared_data(ssize_t bytes)
 516{
 517	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 518}
 519
 520/*
 521 * Allocate process-local memory - this will either be shared between
 522 * threads of this process, or only be accessed by this thread:
 523 */
 524static void * setup_private_data(ssize_t bytes)
 525{
 526	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 527}
 528
 
 
 
 
 
 
 
 
 
 
 
 
 529static int parse_cpu_list(const char *arg)
 530{
 531	p0.cpu_list_str = strdup(arg);
 532
 533	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 534
 535	return 0;
 536}
 537
 538/*
 539 * Check whether a CPU is online
 540 *
 541 * Returns:
 542 *     1 -> if CPU is online
 543 *     0 -> if CPU is offline
 544 *    -1 -> error case
 545 */
 546static int is_cpu_online(unsigned int cpu)
 547{
 548	char *str;
 549	size_t strlen;
 550	char buf[256];
 551	int status = -1;
 552	struct stat statbuf;
 553
 554	snprintf(buf, sizeof(buf),
 555		"/sys/devices/system/cpu/cpu%d", cpu);
 556	if (stat(buf, &statbuf) != 0)
 557		return 0;
 558
 559	/*
 560	 * Check if /sys/devices/system/cpu/cpux/online file
 561	 * exists. Some cases cpu0 won't have online file since
 562	 * it is not expected to be turned off generally.
 563	 * In kernels without CONFIG_HOTPLUG_CPU, this
 564	 * file won't exist
 565	 */
 566	snprintf(buf, sizeof(buf),
 567		"/sys/devices/system/cpu/cpu%d/online", cpu);
 568	if (stat(buf, &statbuf) != 0)
 569		return 1;
 570
 571	/*
 572	 * Read online file using sysfs__read_str.
 573	 * If read or open fails, return -1.
 574	 * If read succeeds, return value from file
 575	 * which gets stored in "str"
 576	 */
 577	snprintf(buf, sizeof(buf),
 578		"devices/system/cpu/cpu%d/online", cpu);
 579
 580	if (sysfs__read_str(buf, &str, &strlen) < 0)
 581		return status;
 582
 583	status = atoi(str);
 584
 585	free(str);
 586	return status;
 587}
 588
 589static int parse_setup_cpu_list(void)
 590{
 591	struct thread_data *td;
 592	char *str0, *str;
 593	int t;
 594
 595	if (!g->p.cpu_list_str)
 596		return 0;
 597
 598	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 599
 600	str0 = str = strdup(g->p.cpu_list_str);
 601	t = 0;
 602
 603	BUG_ON(!str);
 604
 605	tprintf("# binding tasks to CPUs:\n");
 606	tprintf("#  ");
 607
 608	while (true) {
 609		int bind_cpu, bind_cpu_0, bind_cpu_1;
 610		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 611		int bind_len;
 612		int step;
 613		int mul;
 614
 615		tok = strsep(&str, ",");
 616		if (!tok)
 617			break;
 618
 619		tok_end = strstr(tok, "-");
 620
 621		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 622		if (!tok_end) {
 623			/* Single CPU specified: */
 624			bind_cpu_0 = bind_cpu_1 = atol(tok);
 625		} else {
 626			/* CPU range specified (for example: "5-11"): */
 627			bind_cpu_0 = atol(tok);
 628			bind_cpu_1 = atol(tok_end + 1);
 629		}
 630
 631		step = 1;
 632		tok_step = strstr(tok, "#");
 633		if (tok_step) {
 634			step = atol(tok_step + 1);
 635			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 636		}
 637
 638		/*
 639		 * Mask length.
 640		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 641		 * where the _4 means the next 4 CPUs are allowed.
 642		 */
 643		bind_len = 1;
 644		tok_len = strstr(tok, "_");
 645		if (tok_len) {
 646			bind_len = atol(tok_len + 1);
 647			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 648		}
 649
 650		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 651		mul = 1;
 652		tok_mul = strstr(tok, "x");
 653		if (tok_mul) {
 654			mul = atol(tok_mul + 1);
 655			BUG_ON(mul <= 0);
 656		}
 657
 658		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 659
 660		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 661			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 662			return -1;
 663		}
 664
 665		if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
 666			printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
 667			return -1;
 668		}
 669
 670		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 671		BUG_ON(bind_cpu_0 > bind_cpu_1);
 672
 673		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 674			size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
 675			int i;
 676
 677			for (i = 0; i < mul; i++) {
 678				int cpu;
 679
 680				if (t >= g->p.nr_tasks) {
 681					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 682					goto out;
 683				}
 684				td = g->threads + t;
 685
 686				if (t)
 687					tprintf(",");
 688				if (bind_len > 1) {
 689					tprintf("%2d/%d", bind_cpu, bind_len);
 690				} else {
 691					tprintf("%2d", bind_cpu);
 692				}
 693
 694				td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
 695				BUG_ON(!td->bind_cpumask);
 696				CPU_ZERO_S(size, td->bind_cpumask);
 697				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 698					if (cpu < 0 || cpu >= g->p.nr_cpus) {
 699						CPU_FREE(td->bind_cpumask);
 700						BUG_ON(-1);
 701					}
 702					CPU_SET_S(cpu, size, td->bind_cpumask);
 703				}
 704				t++;
 705			}
 706		}
 707	}
 708out:
 709
 710	tprintf("\n");
 711
 712	if (t < g->p.nr_tasks)
 713		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 714
 715	free(str0);
 716	return 0;
 717}
 718
 719static int parse_cpus_opt(const struct option *opt __maybe_unused,
 720			  const char *arg, int unset __maybe_unused)
 721{
 722	if (!arg)
 723		return -1;
 724
 725	return parse_cpu_list(arg);
 726}
 727
 728static int parse_node_list(const char *arg)
 729{
 730	p0.node_list_str = strdup(arg);
 731
 732	dprintf("got NODE list: {%s}\n", p0.node_list_str);
 733
 734	return 0;
 735}
 736
 737static int parse_setup_node_list(void)
 738{
 739	struct thread_data *td;
 740	char *str0, *str;
 741	int t;
 742
 743	if (!g->p.node_list_str)
 744		return 0;
 745
 746	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 747
 748	str0 = str = strdup(g->p.node_list_str);
 749	t = 0;
 750
 751	BUG_ON(!str);
 752
 753	tprintf("# binding tasks to NODEs:\n");
 754	tprintf("# ");
 755
 756	while (true) {
 757		int bind_node, bind_node_0, bind_node_1;
 758		char *tok, *tok_end, *tok_step, *tok_mul;
 759		int step;
 760		int mul;
 761
 762		tok = strsep(&str, ",");
 763		if (!tok)
 764			break;
 765
 766		tok_end = strstr(tok, "-");
 767
 768		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 769		if (!tok_end) {
 770			/* Single NODE specified: */
 771			bind_node_0 = bind_node_1 = atol(tok);
 772		} else {
 773			/* NODE range specified (for example: "5-11"): */
 774			bind_node_0 = atol(tok);
 775			bind_node_1 = atol(tok_end + 1);
 776		}
 777
 778		step = 1;
 779		tok_step = strstr(tok, "#");
 780		if (tok_step) {
 781			step = atol(tok_step + 1);
 782			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 783		}
 784
 785		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 786		mul = 1;
 787		tok_mul = strstr(tok, "x");
 788		if (tok_mul) {
 789			mul = atol(tok_mul + 1);
 790			BUG_ON(mul <= 0);
 791		}
 792
 793		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 794
 795		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 796			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 797			return -1;
 798		}
 799
 800		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 801		BUG_ON(bind_node_0 > bind_node_1);
 802
 803		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 804			int i;
 805
 806			for (i = 0; i < mul; i++) {
 807				if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
 808					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 809					goto out;
 810				}
 811				td = g->threads + t;
 812
 813				if (!t)
 814					tprintf(" %2d", bind_node);
 815				else
 816					tprintf(",%2d", bind_node);
 817
 818				td->bind_node = bind_node;
 819				t++;
 820			}
 821		}
 822	}
 823out:
 824
 825	tprintf("\n");
 826
 827	if (t < g->p.nr_tasks)
 828		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 829
 830	free(str0);
 831	return 0;
 832}
 833
 834static int parse_nodes_opt(const struct option *opt __maybe_unused,
 835			  const char *arg, int unset __maybe_unused)
 836{
 837	if (!arg)
 838		return -1;
 839
 840	return parse_node_list(arg);
 
 
 841}
 842
 
 
 843static inline uint32_t lfsr_32(uint32_t lfsr)
 844{
 845	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 846	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 847}
 848
 849/*
 850 * Make sure there's real data dependency to RAM (when read
 851 * accesses are enabled), so the compiler, the CPU and the
 852 * kernel (KSM, zero page, etc.) cannot optimize away RAM
 853 * accesses:
 854 */
 855static inline u64 access_data(u64 *data, u64 val)
 856{
 857	if (g->p.data_reads)
 858		val += *data;
 859	if (g->p.data_writes)
 860		*data = val + 1;
 861	return val;
 862}
 863
 864/*
 865 * The worker process does two types of work, a forwards going
 866 * loop and a backwards going loop.
 867 *
 868 * We do this so that on multiprocessor systems we do not create
 869 * a 'train' of processing, with highly synchronized processes,
 870 * skewing the whole benchmark.
 871 */
 872static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 873{
 874	long words = bytes/sizeof(u64);
 875	u64 *data = (void *)__data;
 876	long chunk_0, chunk_1;
 877	u64 *d0, *d, *d1;
 878	long off;
 879	long i;
 880
 881	BUG_ON(!data && words);
 882	BUG_ON(data && !words);
 883
 884	if (!data)
 885		return val;
 886
 887	/* Very simple memset() work variant: */
 888	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 889		bzero(data, bytes);
 890		return val;
 891	}
 892
 893	/* Spread out by PID/TID nr and by loop nr: */
 894	chunk_0 = words/nr_max;
 895	chunk_1 = words/g->p.nr_loops;
 896	off = nr*chunk_0 + loop*chunk_1;
 897
 898	while (off >= words)
 899		off -= words;
 900
 901	if (g->p.data_rand_walk) {
 902		u32 lfsr = nr + loop + val;
 903		long j;
 904
 905		for (i = 0; i < words/1024; i++) {
 906			long start, end;
 907
 908			lfsr = lfsr_32(lfsr);
 909
 910			start = lfsr % words;
 911			end = min(start + 1024, words-1);
 912
 913			if (g->p.data_zero_memset) {
 914				bzero(data + start, (end-start) * sizeof(u64));
 915			} else {
 916				for (j = start; j < end; j++)
 917					val = access_data(data + j, val);
 918			}
 919		}
 920	} else if (!g->p.data_backwards || (nr + loop) & 1) {
 921		/* Process data forwards: */
 922
 923		d0 = data + off;
 924		d  = data + off + 1;
 925		d1 = data + words;
 926
 
 927		for (;;) {
 928			if (unlikely(d >= d1))
 929				d = data;
 930			if (unlikely(d == d0))
 931				break;
 932
 933			val = access_data(d, val);
 934
 935			d++;
 936		}
 937	} else {
 938		/* Process data backwards: */
 939
 940		d0 = data + off;
 941		d  = data + off - 1;
 942		d1 = data + words;
 943
 
 944		for (;;) {
 945			if (unlikely(d < data))
 946				d = data + words-1;
 947			if (unlikely(d == d0))
 948				break;
 949
 950			val = access_data(d, val);
 951
 952			d--;
 953		}
 954	}
 955
 956	return val;
 957}
 958
 959static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 960{
 961	unsigned int cpu;
 962
 963	cpu = sched_getcpu();
 964
 965	g->threads[task_nr].curr_cpu = cpu;
 966	prctl(0, bytes_worked);
 967}
 968
 
 
 969/*
 970 * Count the number of nodes a process's threads
 971 * are spread out on.
 972 *
 973 * A count of 1 means that the process is compressed
 974 * to a single node. A count of g->p.nr_nodes means it's
 975 * spread out on the whole system.
 976 */
 977static int count_process_nodes(int process_nr)
 978{
 979	char *node_present;
 980	int nodes;
 981	int n, t;
 982
 983	node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
 984	BUG_ON(!node_present);
 985	for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
 986		node_present[nodes] = 0;
 987
 988	for (t = 0; t < g->p.nr_threads; t++) {
 989		struct thread_data *td;
 990		int task_nr;
 991		int node;
 992
 993		task_nr = process_nr*g->p.nr_threads + t;
 994		td = g->threads + task_nr;
 995
 996		node = numa_node_of_cpu(td->curr_cpu);
 997		if (node < 0) /* curr_cpu was likely still -1 */ {
 998			free(node_present);
 999			return 0;
1000		}
1001
1002		node_present[node] = 1;
1003	}
1004
1005	nodes = 0;
1006
1007	for (n = 0; n < g->p.nr_nodes; n++)
1008		nodes += node_present[n];
1009
1010	free(node_present);
1011	return nodes;
1012}
1013
1014/*
1015 * Count the number of distinct process-threads a node contains.
1016 *
1017 * A count of 1 means that the node contains only a single
1018 * process. If all nodes on the system contain at most one
1019 * process then we are well-converged.
1020 */
1021static int count_node_processes(int node)
1022{
1023	int processes = 0;
1024	int t, p;
1025
1026	for (p = 0; p < g->p.nr_proc; p++) {
1027		for (t = 0; t < g->p.nr_threads; t++) {
1028			struct thread_data *td;
1029			int task_nr;
1030			int n;
1031
1032			task_nr = p*g->p.nr_threads + t;
1033			td = g->threads + task_nr;
1034
1035			n = numa_node_of_cpu(td->curr_cpu);
1036			if (n == node) {
1037				processes++;
1038				break;
1039			}
1040		}
1041	}
1042
1043	return processes;
1044}
1045
1046static void calc_convergence_compression(int *strong)
1047{
1048	unsigned int nodes_min, nodes_max;
1049	int p;
1050
1051	nodes_min = -1;
1052	nodes_max =  0;
1053
1054	for (p = 0; p < g->p.nr_proc; p++) {
1055		unsigned int nodes = count_process_nodes(p);
1056
1057		if (!nodes) {
1058			*strong = 0;
1059			return;
1060		}
1061
1062		nodes_min = min(nodes, nodes_min);
1063		nodes_max = max(nodes, nodes_max);
1064	}
1065
1066	/* Strong convergence: all threads compress on a single node: */
1067	if (nodes_min == 1 && nodes_max == 1) {
1068		*strong = 1;
1069	} else {
1070		*strong = 0;
1071		tprintf(" {%d-%d}", nodes_min, nodes_max);
1072	}
1073}
1074
1075static void calc_convergence(double runtime_ns_max, double *convergence)
1076{
1077	unsigned int loops_done_min, loops_done_max;
1078	int process_groups;
1079	int *nodes;
1080	int distance;
1081	int nr_min;
1082	int nr_max;
1083	int strong;
1084	int sum;
1085	int nr;
1086	int node;
1087	int cpu;
1088	int t;
1089
1090	if (!g->p.show_convergence && !g->p.measure_convergence)
1091		return;
1092
1093	nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1094	BUG_ON(!nodes);
1095	for (node = 0; node < g->p.nr_nodes; node++)
1096		nodes[node] = 0;
1097
1098	loops_done_min = -1;
1099	loops_done_max = 0;
1100
1101	for (t = 0; t < g->p.nr_tasks; t++) {
1102		struct thread_data *td = g->threads + t;
1103		unsigned int loops_done;
1104
1105		cpu = td->curr_cpu;
1106
1107		/* Not all threads have written it yet: */
1108		if (cpu < 0)
1109			continue;
1110
1111		node = numa_node_of_cpu(cpu);
1112
1113		nodes[node]++;
1114
1115		loops_done = td->loops_done;
1116		loops_done_min = min(loops_done, loops_done_min);
1117		loops_done_max = max(loops_done, loops_done_max);
1118	}
1119
1120	nr_max = 0;
1121	nr_min = g->p.nr_tasks;
1122	sum = 0;
1123
1124	for (node = 0; node < g->p.nr_nodes; node++) {
1125		if (!is_node_present(node))
1126			continue;
1127		nr = nodes[node];
1128		nr_min = min(nr, nr_min);
1129		nr_max = max(nr, nr_max);
1130		sum += nr;
1131	}
1132	BUG_ON(nr_min > nr_max);
1133
1134	BUG_ON(sum > g->p.nr_tasks);
1135
1136	if (0 && (sum < g->p.nr_tasks)) {
1137		free(nodes);
1138		return;
1139	}
1140
1141	/*
1142	 * Count the number of distinct process groups present
1143	 * on nodes - when we are converged this will decrease
1144	 * to g->p.nr_proc:
1145	 */
1146	process_groups = 0;
1147
1148	for (node = 0; node < g->p.nr_nodes; node++) {
1149		int processes;
1150
1151		if (!is_node_present(node))
1152			continue;
1153		processes = count_node_processes(node);
1154		nr = nodes[node];
1155		tprintf(" %2d/%-2d", nr, processes);
1156
1157		process_groups += processes;
1158	}
1159
1160	distance = nr_max - nr_min;
1161
1162	tprintf(" [%2d/%-2d]", distance, process_groups);
1163
1164	tprintf(" l:%3d-%-3d (%3d)",
1165		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1166
1167	if (loops_done_min && loops_done_max) {
1168		double skew = 1.0 - (double)loops_done_min/loops_done_max;
1169
1170		tprintf(" [%4.1f%%]", skew * 100.0);
1171	}
1172
1173	calc_convergence_compression(&strong);
1174
1175	if (strong && process_groups == g->p.nr_proc) {
1176		if (!*convergence) {
1177			*convergence = runtime_ns_max;
1178			tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1179			if (g->p.measure_convergence) {
1180				g->all_converged = true;
1181				g->stop_work = true;
1182			}
1183		}
1184	} else {
1185		if (*convergence) {
1186			tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1187			*convergence = 0;
1188		}
1189		tprintf("\n");
1190	}
1191
1192	free(nodes);
1193}
1194
1195static void show_summary(double runtime_ns_max, int l, double *convergence)
1196{
1197	tprintf("\r #  %5.1f%%  [%.1f mins]",
1198		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1199
1200	calc_convergence(runtime_ns_max, convergence);
1201
1202	if (g->p.show_details >= 0)
1203		fflush(stdout);
1204}
1205
1206static void *worker_thread(void *__tdata)
1207{
1208	struct thread_data *td = __tdata;
1209	struct timeval start0, start, stop, diff;
1210	int process_nr = td->process_nr;
1211	int thread_nr = td->thread_nr;
1212	unsigned long last_perturbance;
1213	int task_nr = td->task_nr;
1214	int details = g->p.show_details;
1215	int first_task, last_task;
1216	double convergence = 0;
1217	u64 val = td->val;
1218	double runtime_ns_max;
1219	u8 *global_data;
1220	u8 *process_data;
1221	u8 *thread_data;
1222	u64 bytes_done, secs;
1223	long work_done;
1224	u32 l;
1225	struct rusage rusage;
1226
1227	bind_to_cpumask(td->bind_cpumask);
1228	bind_to_memnode(td->bind_node);
1229
1230	set_taskname("thread %d/%d", process_nr, thread_nr);
1231
1232	global_data = g->data;
1233	process_data = td->process_data;
1234	thread_data = setup_private_data(g->p.bytes_thread);
1235
1236	bytes_done = 0;
1237
1238	last_task = 0;
1239	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1240		last_task = 1;
1241
1242	first_task = 0;
1243	if (process_nr == 0 && thread_nr == 0)
1244		first_task = 1;
1245
1246	if (details >= 2) {
1247		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1248			process_nr, thread_nr, global_data, process_data, thread_data);
1249	}
1250
1251	if (g->p.serialize_startup) {
1252		mutex_lock(&g->startup_mutex);
1253		g->nr_tasks_started++;
1254		/* The last thread wakes the main process. */
1255		if (g->nr_tasks_started == g->p.nr_tasks)
1256			cond_signal(&g->startup_cond);
1257
1258		mutex_unlock(&g->startup_mutex);
1259
1260		/* Here we will wait for the main process to start us all at once: */
1261		mutex_lock(&g->start_work_mutex);
1262		g->start_work = false;
1263		g->nr_tasks_working++;
1264		while (!g->start_work)
1265			cond_wait(&g->start_work_cond, &g->start_work_mutex);
1266
1267		mutex_unlock(&g->start_work_mutex);
 
 
 
 
1268	}
1269
1270	gettimeofday(&start0, NULL);
1271
1272	start = stop = start0;
1273	last_perturbance = start.tv_sec;
1274
1275	for (l = 0; l < g->p.nr_loops; l++) {
1276		start = stop;
1277
1278		if (g->stop_work)
1279			break;
1280
1281		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1282		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1283		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1284
1285		if (g->p.sleep_usecs) {
1286			mutex_lock(td->process_lock);
1287			usleep(g->p.sleep_usecs);
1288			mutex_unlock(td->process_lock);
1289		}
1290		/*
1291		 * Amount of work to be done under a process-global lock:
1292		 */
1293		if (g->p.bytes_process_locked) {
1294			mutex_lock(td->process_lock);
1295			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1296			mutex_unlock(td->process_lock);
1297		}
1298
1299		work_done = g->p.bytes_global + g->p.bytes_process +
1300			    g->p.bytes_process_locked + g->p.bytes_thread;
1301
1302		update_curr_cpu(task_nr, work_done);
1303		bytes_done += work_done;
1304
1305		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1306			continue;
1307
1308		td->loops_done = l;
1309
1310		gettimeofday(&stop, NULL);
1311
1312		/* Check whether our max runtime timed out: */
1313		if (g->p.nr_secs) {
1314			timersub(&stop, &start0, &diff);
1315			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1316				g->stop_work = true;
1317				break;
1318			}
1319		}
1320
1321		/* Update the summary at most once per second: */
1322		if (start.tv_sec == stop.tv_sec)
1323			continue;
1324
1325		/*
1326		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1327		 * by migrating to CPU#0:
1328		 */
1329		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1330			cpu_set_t *orig_mask;
1331			int target_cpu;
1332			int this_cpu;
1333
1334			last_perturbance = stop.tv_sec;
1335
1336			/*
1337			 * Depending on where we are running, move into
1338			 * the other half of the system, to create some
1339			 * real disturbance:
1340			 */
1341			this_cpu = g->threads[task_nr].curr_cpu;
1342			if (this_cpu < g->p.nr_cpus/2)
1343				target_cpu = g->p.nr_cpus-1;
1344			else
1345				target_cpu = 0;
1346
1347			orig_mask = bind_to_cpu(target_cpu);
1348
1349			/* Here we are running on the target CPU already */
1350			if (details >= 1)
1351				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1352
1353			bind_to_cpumask(orig_mask);
1354			CPU_FREE(orig_mask);
1355		}
1356
1357		if (details >= 3) {
1358			timersub(&stop, &start, &diff);
1359			runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1360			runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1361
1362			if (details >= 0) {
1363				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1364					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1365			}
1366			fflush(stdout);
1367		}
1368		if (!last_task)
1369			continue;
1370
1371		timersub(&stop, &start0, &diff);
1372		runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1373		runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1374
1375		show_summary(runtime_ns_max, l, &convergence);
1376	}
1377
1378	gettimeofday(&stop, NULL);
1379	timersub(&stop, &start0, &diff);
1380	td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1381	td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1382	secs = td->runtime_ns / NSEC_PER_SEC;
1383	td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1384
1385	getrusage(RUSAGE_THREAD, &rusage);
1386	td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1387	td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1388	td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1389	td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1390
1391	free_data(thread_data, g->p.bytes_thread);
1392
1393	mutex_lock(&g->stop_work_mutex);
1394	g->bytes_done += bytes_done;
1395	mutex_unlock(&g->stop_work_mutex);
1396
1397	return NULL;
1398}
1399
1400/*
1401 * A worker process starts a couple of threads:
1402 */
1403static void worker_process(int process_nr)
1404{
1405	struct mutex process_lock;
1406	struct thread_data *td;
1407	pthread_t *pthreads;
1408	u8 *process_data;
1409	int task_nr;
1410	int ret;
1411	int t;
1412
1413	mutex_init(&process_lock);
1414	set_taskname("process %d", process_nr);
1415
1416	/*
1417	 * Pick up the memory policy and the CPU binding of our first thread,
1418	 * so that we initialize memory accordingly:
1419	 */
1420	task_nr = process_nr*g->p.nr_threads;
1421	td = g->threads + task_nr;
1422
1423	bind_to_memnode(td->bind_node);
1424	bind_to_cpumask(td->bind_cpumask);
1425
1426	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1427	process_data = setup_private_data(g->p.bytes_process);
1428
1429	if (g->p.show_details >= 3) {
1430		printf(" # process %2d global mem: %p, process mem: %p\n",
1431			process_nr, g->data, process_data);
1432	}
1433
1434	for (t = 0; t < g->p.nr_threads; t++) {
1435		task_nr = process_nr*g->p.nr_threads + t;
1436		td = g->threads + task_nr;
1437
1438		td->process_data = process_data;
1439		td->process_nr   = process_nr;
1440		td->thread_nr    = t;
1441		td->task_nr	 = task_nr;
1442		td->val          = rand();
1443		td->curr_cpu	 = -1;
1444		td->process_lock = &process_lock;
1445
1446		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1447		BUG_ON(ret);
1448	}
1449
1450	for (t = 0; t < g->p.nr_threads; t++) {
1451                ret = pthread_join(pthreads[t], NULL);
1452		BUG_ON(ret);
1453	}
1454
1455	free_data(process_data, g->p.bytes_process);
1456	free(pthreads);
1457}
1458
1459static void print_summary(void)
1460{
1461	if (g->p.show_details < 0)
1462		return;
1463
1464	printf("\n ###\n");
1465	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1466		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1467	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1468			g->p.nr_loops, g->p.bytes_global/1024/1024);
1469	printf(" #      %5dx %5ldMB process shared mem operations\n",
1470			g->p.nr_loops, g->p.bytes_process/1024/1024);
1471	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1472			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1473
1474	printf(" ###\n");
1475
1476	printf("\n ###\n"); fflush(stdout);
1477}
1478
1479static void init_thread_data(void)
1480{
1481	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1482	int t;
1483
1484	g->threads = zalloc_shared_data(size);
1485
1486	for (t = 0; t < g->p.nr_tasks; t++) {
1487		struct thread_data *td = g->threads + t;
1488		size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1489		int cpu;
1490
1491		/* Allow all nodes by default: */
1492		td->bind_node = NUMA_NO_NODE;
1493
1494		/* Allow all CPUs by default: */
1495		td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1496		BUG_ON(!td->bind_cpumask);
1497		CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1498		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1499			CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1500	}
1501}
1502
1503static void deinit_thread_data(void)
1504{
1505	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1506	int t;
1507
1508	/* Free the bind_cpumask allocated for thread_data */
1509	for (t = 0; t < g->p.nr_tasks; t++) {
1510		struct thread_data *td = g->threads + t;
1511		CPU_FREE(td->bind_cpumask);
1512	}
1513
1514	free_data(g->threads, size);
1515}
1516
1517static int init(void)
1518{
1519	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1520
1521	/* Copy over options: */
1522	g->p = p0;
1523
1524	g->p.nr_cpus = numa_num_configured_cpus();
1525
1526	g->p.nr_nodes = numa_max_node() + 1;
1527
1528	/* char array in count_process_nodes(): */
1529	BUG_ON(g->p.nr_nodes < 0);
1530
1531	if (quiet && !g->p.show_details)
1532		g->p.show_details = -1;
1533
1534	/* Some memory should be specified: */
1535	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1536		return -1;
1537
1538	if (g->p.mb_global_str) {
1539		g->p.mb_global = atof(g->p.mb_global_str);
1540		BUG_ON(g->p.mb_global < 0);
1541	}
1542
1543	if (g->p.mb_proc_str) {
1544		g->p.mb_proc = atof(g->p.mb_proc_str);
1545		BUG_ON(g->p.mb_proc < 0);
1546	}
1547
1548	if (g->p.mb_proc_locked_str) {
1549		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1550		BUG_ON(g->p.mb_proc_locked < 0);
1551		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1552	}
1553
1554	if (g->p.mb_thread_str) {
1555		g->p.mb_thread = atof(g->p.mb_thread_str);
1556		BUG_ON(g->p.mb_thread < 0);
1557	}
1558
1559	BUG_ON(g->p.nr_threads <= 0);
1560	BUG_ON(g->p.nr_proc <= 0);
1561
1562	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1563
1564	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1565	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1566	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1567	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1568
1569	g->data = setup_shared_data(g->p.bytes_global);
1570
1571	/* Startup serialization: */
1572	mutex_init_pshared(&g->start_work_mutex);
1573	cond_init_pshared(&g->start_work_cond);
1574	mutex_init_pshared(&g->startup_mutex);
1575	cond_init_pshared(&g->startup_cond);
1576	mutex_init_pshared(&g->stop_work_mutex);
1577
1578	init_thread_data();
1579
1580	tprintf("#\n");
1581	if (parse_setup_cpu_list() || parse_setup_node_list())
1582		return -1;
1583	tprintf("#\n");
1584
1585	print_summary();
1586
1587	return 0;
1588}
1589
1590static void deinit(void)
1591{
1592	free_data(g->data, g->p.bytes_global);
1593	g->data = NULL;
1594
1595	deinit_thread_data();
1596
1597	free_data(g, sizeof(*g));
1598	g = NULL;
1599}
1600
1601/*
1602 * Print a short or long result, depending on the verbosity setting:
1603 */
1604static void print_res(const char *name, double val,
1605		      const char *txt_unit, const char *txt_short, const char *txt_long)
1606{
1607	if (!name)
1608		name = "main,";
1609
1610	if (!quiet)
1611		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1612	else
1613		printf(" %14.3f %s\n", val, txt_long);
1614}
1615
1616static int __bench_numa(const char *name)
1617{
1618	struct timeval start, stop, diff;
1619	u64 runtime_ns_min, runtime_ns_sum;
1620	pid_t *pids, pid, wpid;
1621	double delta_runtime;
1622	double runtime_avg;
1623	double runtime_sec_max;
1624	double runtime_sec_min;
1625	int wait_stat;
1626	double bytes;
1627	int i, t, p;
1628
1629	if (init())
1630		return -1;
1631
1632	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1633	pid = -1;
1634
 
 
 
1635	if (g->p.serialize_startup) {
1636		tprintf(" #\n");
1637		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1638	}
1639
1640	gettimeofday(&start, NULL);
1641
1642	for (i = 0; i < g->p.nr_proc; i++) {
1643		pid = fork();
1644		dprintf(" # process %2d: PID %d\n", i, pid);
1645
1646		BUG_ON(pid < 0);
1647		if (!pid) {
1648			/* Child process: */
1649			worker_process(i);
1650
1651			exit(0);
1652		}
1653		pids[i] = pid;
1654
1655	}
 
 
 
 
 
1656
1657	if (g->p.serialize_startup) {
1658		bool threads_ready = false;
1659		double startup_sec;
1660
1661		/*
1662		 * Wait for all the threads to start up. The last thread will
1663		 * signal this process.
1664		 */
1665		mutex_lock(&g->startup_mutex);
1666		while (g->nr_tasks_started != g->p.nr_tasks)
1667			cond_wait(&g->startup_cond, &g->startup_mutex);
1668
1669		mutex_unlock(&g->startup_mutex);
1670
1671		/* Wait for all threads to be at the start_work_cond. */
1672		while (!threads_ready) {
1673			mutex_lock(&g->start_work_mutex);
1674			threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1675			mutex_unlock(&g->start_work_mutex);
1676			if (!threads_ready)
1677				usleep(1);
1678		}
1679
1680		gettimeofday(&stop, NULL);
1681
1682		timersub(&stop, &start, &diff);
1683
1684		startup_sec = diff.tv_sec * NSEC_PER_SEC;
1685		startup_sec += diff.tv_usec * NSEC_PER_USEC;
1686		startup_sec /= NSEC_PER_SEC;
1687
1688		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1689		tprintf(" #\n");
1690
1691		start = stop;
1692		/* Start all threads running. */
1693		mutex_lock(&g->start_work_mutex);
1694		g->start_work = true;
1695		mutex_unlock(&g->start_work_mutex);
1696		cond_broadcast(&g->start_work_cond);
1697	} else {
1698		gettimeofday(&start, NULL);
1699	}
1700
1701	/* Parent process: */
1702
1703
1704	for (i = 0; i < g->p.nr_proc; i++) {
1705		wpid = waitpid(pids[i], &wait_stat, 0);
1706		BUG_ON(wpid < 0);
1707		BUG_ON(!WIFEXITED(wait_stat));
1708
1709	}
1710
1711	runtime_ns_sum = 0;
1712	runtime_ns_min = -1LL;
1713
1714	for (t = 0; t < g->p.nr_tasks; t++) {
1715		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1716
1717		runtime_ns_sum += thread_runtime_ns;
1718		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1719	}
1720
1721	gettimeofday(&stop, NULL);
1722	timersub(&stop, &start, &diff);
1723
1724	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1725
1726	tprintf("\n ###\n");
1727	tprintf("\n");
1728
1729	runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1730	runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1731	runtime_sec_max /= NSEC_PER_SEC;
1732
1733	runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1734
1735	bytes = g->bytes_done;
1736	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1737
1738	if (g->p.measure_convergence) {
1739		print_res(name, runtime_sec_max,
1740			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1741	}
1742
1743	print_res(name, runtime_sec_max,
1744		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1745
1746	print_res(name, runtime_sec_min,
1747		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1748
1749	print_res(name, runtime_avg,
1750		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1751
1752	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1753	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1754		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1755
1756	print_res(name, bytes / g->p.nr_tasks / 1e9,
1757		"GB,", "data/thread",		"GB data processed, per thread");
1758
1759	print_res(name, bytes / 1e9,
1760		"GB,", "data-total",		"GB data processed, total");
1761
1762	print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1763		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1764
1765	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1766		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1767
1768	print_res(name, bytes / runtime_sec_max / 1e9,
1769		"GB/sec,", "total-speed",	"GB/sec total speed");
1770
1771	if (g->p.show_details >= 2) {
1772		char tname[14 + 2 * 11 + 1];
1773		struct thread_data *td;
1774		for (p = 0; p < g->p.nr_proc; p++) {
1775			for (t = 0; t < g->p.nr_threads; t++) {
1776				memset(tname, 0, sizeof(tname));
1777				td = g->threads + p*g->p.nr_threads + t;
1778				snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1779				print_res(tname, td->speed_gbs,
1780					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1781				print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1782					"secs",	"thread-system-time", "system CPU time/thread");
1783				print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1784					"secs",	"thread-user-time", "user CPU time/thread");
1785			}
1786		}
1787	}
1788
1789	free(pids);
1790
1791	deinit();
1792
1793	return 0;
1794}
1795
1796#define MAX_ARGS 50
1797
1798static int command_size(const char **argv)
1799{
1800	int size = 0;
1801
1802	while (*argv) {
1803		size++;
1804		argv++;
1805	}
1806
1807	BUG_ON(size >= MAX_ARGS);
1808
1809	return size;
1810}
1811
1812static void init_params(struct params *p, const char *name, int argc, const char **argv)
1813{
1814	int i;
1815
1816	printf("\n # Running %s \"perf bench numa", name);
1817
1818	for (i = 0; i < argc; i++)
1819		printf(" %s", argv[i]);
1820
1821	printf("\"\n");
1822
1823	memset(p, 0, sizeof(*p));
1824
1825	/* Initialize nonzero defaults: */
1826
1827	p->serialize_startup		= 1;
1828	p->data_reads			= true;
1829	p->data_writes			= true;
1830	p->data_backwards		= true;
1831	p->data_rand_walk		= true;
1832	p->nr_loops			= -1;
1833	p->init_random			= true;
1834	p->mb_global_str		= "1";
1835	p->nr_proc			= 1;
1836	p->nr_threads			= 1;
1837	p->nr_secs			= 5;
1838	p->run_all			= argc == 1;
1839}
1840
1841static int run_bench_numa(const char *name, const char **argv)
1842{
1843	int argc = command_size(argv);
1844
1845	init_params(&p0, name, argc, argv);
1846	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1847	if (argc)
1848		goto err;
1849
1850	if (__bench_numa(name))
1851		goto err;
1852
1853	return 0;
1854
1855err:
1856	return -1;
1857}
1858
1859#define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1860#define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1861
1862#define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1863#define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1864
1865#define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1866#define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1867
1868/*
1869 * The built-in test-suite executed by "perf bench numa -a".
1870 *
1871 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1872 */
1873static const char *tests[][MAX_ARGS] = {
1874   /* Basic single-stream NUMA bandwidth measurements: */
1875   { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1876			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1877   { "RAM-bw-local-NOTHP,",
1878			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1879			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1880   { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1881			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1882
1883   /* 2-stream NUMA bandwidth measurements: */
1884   { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1885			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1886   { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1887		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1888
1889   /* Cross-stream NUMA bandwidth measurement: */
1890   { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1891		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1892
1893   /* Convergence latency measurements: */
1894   { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1895   { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1896   { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1897   { " 2x3-convergence,", "mem",  "-p",  "2", "-t",  "3", "-P", "1020", OPT_CONV },
1898   { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1899   { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1900   { " 4x4-convergence-NOTHP,",
1901			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1902   { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1903   { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1904   { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1905   { " 8x4-convergence-NOTHP,",
1906			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1907   { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1908   { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1909   { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1910   { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1911   { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1912
1913   /* Various NUMA process/thread layout bandwidth measurements: */
1914   { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1915   { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1916   { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1917   { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1918   { " 8x1-bw-process-NOTHP,",
1919			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1920   { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1921
1922   { " 1x4-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1923   { " 1x8-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1924   { "1x16-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1925   { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1926
1927   { " 2x3-bw-process,",  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1928   { " 4x4-bw-process,",  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1929   { " 4x6-bw-process,",  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1930   { " 4x8-bw-process,",  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1931   { " 4x8-bw-process-NOTHP,",
1932			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1933   { " 3x3-bw-process,",  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1934   { " 5x5-bw-process,",  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1935
1936   { "2x16-bw-process,",  "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1937   { "1x32-bw-process,",  "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1938
1939   { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1940   { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1941   { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1942   { "numa01-bw-thread-NOTHP,",
1943			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1944};
1945
1946static int bench_all(void)
1947{
1948	int nr = ARRAY_SIZE(tests);
1949	int ret;
1950	int i;
1951
1952	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1953	BUG_ON(ret < 0);
1954
1955	for (i = 0; i < nr; i++) {
1956		run_bench_numa(tests[i][0], tests[i] + 1);
1957	}
1958
1959	printf("\n");
1960
1961	return 0;
1962}
1963
1964int bench_numa(int argc, const char **argv)
1965{
1966	init_params(&p0, "main,", argc, argv);
1967	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1968	if (argc)
1969		goto err;
1970
1971	if (p0.run_all)
1972		return bench_all();
1973
1974	if (__bench_numa(NULL))
1975		goto err;
1976
1977	return 0;
1978
1979err:
1980	usage_with_options(numa_usage, options);
1981	return -1;
1982}