Linux Audio

Check our new training course

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}
v4.10.11
   1/*
   2 * numa.c
   3 *
   4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
   5 */
   6
   7/* For the CLR_() macros */
   8#include <pthread.h>
   9
  10#include "../perf.h"
  11#include "../builtin.h"
  12#include "../util/util.h"
  13#include <subcmd/parse-options.h>
  14#include "../util/cloexec.h"
  15
  16#include "bench.h"
  17
  18#include <errno.h>
  19#include <sched.h>
  20#include <stdio.h>
  21#include <assert.h>
  22#include <malloc.h>
  23#include <signal.h>
  24#include <stdlib.h>
  25#include <string.h>
  26#include <unistd.h>
 
  27#include <sys/mman.h>
  28#include <sys/time.h>
  29#include <sys/resource.h>
  30#include <sys/wait.h>
  31#include <sys/prctl.h>
  32#include <sys/types.h>
  33#include <linux/time64.h>
  34
  35#include <numa.h>
  36#include <numaif.h>
  37
  38/*
  39 * Regular printout to the terminal, supressed if -q is specified:
  40 */
  41#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
  42
  43/*
  44 * Debug printf:
  45 */
  46#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
  47
  48struct thread_data {
  49	int			curr_cpu;
  50	cpu_set_t		bind_cpumask;
  51	int			bind_node;
  52	u8			*process_data;
  53	int			process_nr;
  54	int			thread_nr;
  55	int			task_nr;
  56	unsigned int		loops_done;
  57	u64			val;
  58	u64			runtime_ns;
  59	u64			system_time_ns;
  60	u64			user_time_ns;
  61	double			speed_gbs;
  62	pthread_mutex_t		*process_lock;
  63};
  64
  65/* Parameters set by options: */
  66
  67struct params {
  68	/* Startup synchronization: */
  69	bool			serialize_startup;
  70
  71	/* Task hierarchy: */
  72	int			nr_proc;
  73	int			nr_threads;
  74
  75	/* Working set sizes: */
  76	const char		*mb_global_str;
  77	const char		*mb_proc_str;
  78	const char		*mb_proc_locked_str;
  79	const char		*mb_thread_str;
  80
  81	double			mb_global;
  82	double			mb_proc;
  83	double			mb_proc_locked;
  84	double			mb_thread;
  85
  86	/* Access patterns to the working set: */
  87	bool			data_reads;
  88	bool			data_writes;
  89	bool			data_backwards;
  90	bool			data_zero_memset;
  91	bool			data_rand_walk;
  92	u32			nr_loops;
  93	u32			nr_secs;
  94	u32			sleep_usecs;
  95
  96	/* Working set initialization: */
  97	bool			init_zero;
  98	bool			init_random;
  99	bool			init_cpu0;
 100
 101	/* Misc options: */
 102	int			show_details;
 103	int			run_all;
 104	int			thp;
 105
 106	long			bytes_global;
 107	long			bytes_process;
 108	long			bytes_process_locked;
 109	long			bytes_thread;
 110
 111	int			nr_tasks;
 112	bool			show_quiet;
 113
 114	bool			show_convergence;
 115	bool			measure_convergence;
 116
 117	int			perturb_secs;
 118	int			nr_cpus;
 119	int			nr_nodes;
 120
 121	/* Affinity options -C and -N: */
 122	char			*cpu_list_str;
 123	char			*node_list_str;
 124};
 125
 126
 127/* Global, read-writable area, accessible to all processes and threads: */
 128
 129struct global_info {
 130	u8			*data;
 131
 132	pthread_mutex_t		startup_mutex;
 133	int			nr_tasks_started;
 134
 135	pthread_mutex_t		startup_done_mutex;
 136
 137	pthread_mutex_t		start_work_mutex;
 138	int			nr_tasks_working;
 139
 140	pthread_mutex_t		stop_work_mutex;
 141	u64			bytes_done;
 142
 143	struct thread_data	*threads;
 144
 145	/* Convergence latency measurement: */
 146	bool			all_converged;
 147	bool			stop_work;
 148
 149	int			print_once;
 150
 151	struct params		p;
 152};
 153
 154static struct global_info	*g = NULL;
 155
 156static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
 157static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
 158
 159struct params p0;
 160
 161static const struct option options[] = {
 162	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
 163	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
 164
 165	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
 166	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
 167	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
 168	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
 169
 170	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
 171	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
 172	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
 173
 174	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via writes (can be mixed with -W)"),
 175	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
 176	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
 177	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
 178	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
 179
 180
 181	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
 182	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
 183	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
 184	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
 185
 186	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
 187	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
 188	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
 189	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
 190	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
 191	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"quiet mode"),
 192	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
 193
 194	/* Special option string parsing callbacks: */
 195        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
 196			"bind the first N tasks to these specific cpus (the rest is unbound)",
 197			parse_cpus_opt),
 198        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
 199			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
 200			parse_nodes_opt),
 201	OPT_END()
 202};
 203
 204static const char * const bench_numa_usage[] = {
 205	"perf bench numa <options>",
 206	NULL
 207};
 208
 209static const char * const numa_usage[] = {
 210	"perf bench numa mem [<options>]",
 211	NULL
 212};
 213
 214static cpu_set_t bind_to_cpu(int target_cpu)
 215{
 216	cpu_set_t orig_mask, mask;
 217	int ret;
 218
 219	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 220	BUG_ON(ret);
 221
 222	CPU_ZERO(&mask);
 223
 224	if (target_cpu == -1) {
 225		int cpu;
 226
 227		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 228			CPU_SET(cpu, &mask);
 229	} else {
 230		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
 231		CPU_SET(target_cpu, &mask);
 232	}
 233
 234	ret = sched_setaffinity(0, sizeof(mask), &mask);
 235	BUG_ON(ret);
 236
 237	return orig_mask;
 238}
 239
 240static cpu_set_t bind_to_node(int target_node)
 241{
 242	int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
 243	cpu_set_t orig_mask, mask;
 244	int cpu;
 245	int ret;
 246
 247	BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
 248	BUG_ON(!cpus_per_node);
 249
 250	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
 251	BUG_ON(ret);
 252
 253	CPU_ZERO(&mask);
 254
 255	if (target_node == -1) {
 256		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
 257			CPU_SET(cpu, &mask);
 258	} else {
 259		int cpu_start = (target_node + 0) * cpus_per_node;
 260		int cpu_stop  = (target_node + 1) * cpus_per_node;
 261
 262		BUG_ON(cpu_stop > g->p.nr_cpus);
 263
 264		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
 265			CPU_SET(cpu, &mask);
 266	}
 267
 268	ret = sched_setaffinity(0, sizeof(mask), &mask);
 269	BUG_ON(ret);
 270
 271	return orig_mask;
 272}
 273
 274static void bind_to_cpumask(cpu_set_t mask)
 275{
 276	int ret;
 277
 278	ret = sched_setaffinity(0, sizeof(mask), &mask);
 279	BUG_ON(ret);
 280}
 281
 282static void mempol_restore(void)
 283{
 284	int ret;
 285
 286	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
 287
 288	BUG_ON(ret);
 289}
 290
 291static void bind_to_memnode(int node)
 292{
 293	unsigned long nodemask;
 294	int ret;
 295
 296	if (node == -1)
 297		return;
 298
 299	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
 300	nodemask = 1L << node;
 301
 302	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
 303	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
 304
 305	BUG_ON(ret);
 306}
 307
 308#define HPSIZE (2*1024*1024)
 309
 310#define set_taskname(fmt...)				\
 311do {							\
 312	char name[20];					\
 313							\
 314	snprintf(name, 20, fmt);			\
 315	prctl(PR_SET_NAME, name);			\
 316} while (0)
 317
 318static u8 *alloc_data(ssize_t bytes0, int map_flags,
 319		      int init_zero, int init_cpu0, int thp, int init_random)
 320{
 321	cpu_set_t orig_mask;
 322	ssize_t bytes;
 323	u8 *buf;
 324	int ret;
 325
 326	if (!bytes0)
 327		return NULL;
 328
 329	/* Allocate and initialize all memory on CPU#0: */
 330	if (init_cpu0) {
 331		orig_mask = bind_to_node(0);
 332		bind_to_memnode(0);
 333	}
 334
 335	bytes = bytes0 + HPSIZE;
 336
 337	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
 338	BUG_ON(buf == (void *)-1);
 339
 340	if (map_flags == MAP_PRIVATE) {
 341		if (thp > 0) {
 342			ret = madvise(buf, bytes, MADV_HUGEPAGE);
 343			if (ret && !g->print_once) {
 344				g->print_once = 1;
 345				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
 346			}
 347		}
 348		if (thp < 0) {
 349			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
 350			if (ret && !g->print_once) {
 351				g->print_once = 1;
 352				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
 353			}
 354		}
 355	}
 356
 357	if (init_zero) {
 358		bzero(buf, bytes);
 359	} else {
 360		/* Initialize random contents, different in each word: */
 361		if (init_random) {
 362			u64 *wbuf = (void *)buf;
 363			long off = rand();
 364			long i;
 365
 366			for (i = 0; i < bytes/8; i++)
 367				wbuf[i] = i + off;
 368		}
 369	}
 370
 371	/* Align to 2MB boundary: */
 372	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
 373
 374	/* Restore affinity: */
 375	if (init_cpu0) {
 376		bind_to_cpumask(orig_mask);
 377		mempol_restore();
 378	}
 379
 380	return buf;
 381}
 382
 383static void free_data(void *data, ssize_t bytes)
 384{
 385	int ret;
 386
 387	if (!data)
 388		return;
 389
 390	ret = munmap(data, bytes);
 391	BUG_ON(ret);
 392}
 393
 394/*
 395 * Create a shared memory buffer that can be shared between processes, zeroed:
 396 */
 397static void * zalloc_shared_data(ssize_t bytes)
 398{
 399	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 400}
 401
 402/*
 403 * Create a shared memory buffer that can be shared between processes:
 404 */
 405static void * setup_shared_data(ssize_t bytes)
 406{
 407	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 408}
 409
 410/*
 411 * Allocate process-local memory - this will either be shared between
 412 * threads of this process, or only be accessed by this thread:
 413 */
 414static void * setup_private_data(ssize_t bytes)
 415{
 416	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
 417}
 418
 419/*
 420 * Return a process-shared (global) mutex:
 421 */
 422static void init_global_mutex(pthread_mutex_t *mutex)
 423{
 424	pthread_mutexattr_t attr;
 425
 426	pthread_mutexattr_init(&attr);
 427	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
 428	pthread_mutex_init(mutex, &attr);
 429}
 430
 431static int parse_cpu_list(const char *arg)
 432{
 433	p0.cpu_list_str = strdup(arg);
 434
 435	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
 436
 437	return 0;
 438}
 439
 440static int parse_setup_cpu_list(void)
 441{
 442	struct thread_data *td;
 443	char *str0, *str;
 444	int t;
 445
 446	if (!g->p.cpu_list_str)
 447		return 0;
 448
 449	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 450
 451	str0 = str = strdup(g->p.cpu_list_str);
 452	t = 0;
 453
 454	BUG_ON(!str);
 455
 456	tprintf("# binding tasks to CPUs:\n");
 457	tprintf("#  ");
 458
 459	while (true) {
 460		int bind_cpu, bind_cpu_0, bind_cpu_1;
 461		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
 462		int bind_len;
 463		int step;
 464		int mul;
 465
 466		tok = strsep(&str, ",");
 467		if (!tok)
 468			break;
 469
 470		tok_end = strstr(tok, "-");
 471
 472		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 473		if (!tok_end) {
 474			/* Single CPU specified: */
 475			bind_cpu_0 = bind_cpu_1 = atol(tok);
 476		} else {
 477			/* CPU range specified (for example: "5-11"): */
 478			bind_cpu_0 = atol(tok);
 479			bind_cpu_1 = atol(tok_end + 1);
 480		}
 481
 482		step = 1;
 483		tok_step = strstr(tok, "#");
 484		if (tok_step) {
 485			step = atol(tok_step + 1);
 486			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
 487		}
 488
 489		/*
 490		 * Mask length.
 491		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
 492		 * where the _4 means the next 4 CPUs are allowed.
 493		 */
 494		bind_len = 1;
 495		tok_len = strstr(tok, "_");
 496		if (tok_len) {
 497			bind_len = atol(tok_len + 1);
 498			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
 499		}
 500
 501		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 502		mul = 1;
 503		tok_mul = strstr(tok, "x");
 504		if (tok_mul) {
 505			mul = atol(tok_mul + 1);
 506			BUG_ON(mul <= 0);
 507		}
 508
 509		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
 510
 511		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
 512			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
 513			return -1;
 514		}
 515
 516		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
 517		BUG_ON(bind_cpu_0 > bind_cpu_1);
 518
 519		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
 520			int i;
 521
 522			for (i = 0; i < mul; i++) {
 523				int cpu;
 524
 525				if (t >= g->p.nr_tasks) {
 526					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
 527					goto out;
 528				}
 529				td = g->threads + t;
 530
 531				if (t)
 532					tprintf(",");
 533				if (bind_len > 1) {
 534					tprintf("%2d/%d", bind_cpu, bind_len);
 535				} else {
 536					tprintf("%2d", bind_cpu);
 537				}
 538
 539				CPU_ZERO(&td->bind_cpumask);
 540				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
 541					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
 542					CPU_SET(cpu, &td->bind_cpumask);
 543				}
 544				t++;
 545			}
 546		}
 547	}
 548out:
 549
 550	tprintf("\n");
 551
 552	if (t < g->p.nr_tasks)
 553		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 554
 555	free(str0);
 556	return 0;
 557}
 558
 559static int parse_cpus_opt(const struct option *opt __maybe_unused,
 560			  const char *arg, int unset __maybe_unused)
 561{
 562	if (!arg)
 563		return -1;
 564
 565	return parse_cpu_list(arg);
 566}
 567
 568static int parse_node_list(const char *arg)
 569{
 570	p0.node_list_str = strdup(arg);
 571
 572	dprintf("got NODE list: {%s}\n", p0.node_list_str);
 573
 574	return 0;
 575}
 576
 577static int parse_setup_node_list(void)
 578{
 579	struct thread_data *td;
 580	char *str0, *str;
 581	int t;
 582
 583	if (!g->p.node_list_str)
 584		return 0;
 585
 586	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
 587
 588	str0 = str = strdup(g->p.node_list_str);
 589	t = 0;
 590
 591	BUG_ON(!str);
 592
 593	tprintf("# binding tasks to NODEs:\n");
 594	tprintf("# ");
 595
 596	while (true) {
 597		int bind_node, bind_node_0, bind_node_1;
 598		char *tok, *tok_end, *tok_step, *tok_mul;
 599		int step;
 600		int mul;
 601
 602		tok = strsep(&str, ",");
 603		if (!tok)
 604			break;
 605
 606		tok_end = strstr(tok, "-");
 607
 608		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
 609		if (!tok_end) {
 610			/* Single NODE specified: */
 611			bind_node_0 = bind_node_1 = atol(tok);
 612		} else {
 613			/* NODE range specified (for example: "5-11"): */
 614			bind_node_0 = atol(tok);
 615			bind_node_1 = atol(tok_end + 1);
 616		}
 617
 618		step = 1;
 619		tok_step = strstr(tok, "#");
 620		if (tok_step) {
 621			step = atol(tok_step + 1);
 622			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
 623		}
 624
 625		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
 626		mul = 1;
 627		tok_mul = strstr(tok, "x");
 628		if (tok_mul) {
 629			mul = atol(tok_mul + 1);
 630			BUG_ON(mul <= 0);
 631		}
 632
 633		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
 634
 635		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
 636			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
 637			return -1;
 638		}
 639
 640		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
 641		BUG_ON(bind_node_0 > bind_node_1);
 642
 643		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
 644			int i;
 645
 646			for (i = 0; i < mul; i++) {
 647				if (t >= g->p.nr_tasks) {
 648					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
 649					goto out;
 650				}
 651				td = g->threads + t;
 652
 653				if (!t)
 654					tprintf(" %2d", bind_node);
 655				else
 656					tprintf(",%2d", bind_node);
 657
 658				td->bind_node = bind_node;
 659				t++;
 660			}
 661		}
 662	}
 663out:
 664
 665	tprintf("\n");
 666
 667	if (t < g->p.nr_tasks)
 668		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
 669
 670	free(str0);
 671	return 0;
 672}
 673
 674static int parse_nodes_opt(const struct option *opt __maybe_unused,
 675			  const char *arg, int unset __maybe_unused)
 676{
 677	if (!arg)
 678		return -1;
 679
 680	return parse_node_list(arg);
 681
 682	return 0;
 683}
 684
 685#define BIT(x) (1ul << x)
 686
 687static inline uint32_t lfsr_32(uint32_t lfsr)
 688{
 689	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
 690	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
 691}
 692
 693/*
 694 * Make sure there's real data dependency to RAM (when read
 695 * accesses are enabled), so the compiler, the CPU and the
 696 * kernel (KSM, zero page, etc.) cannot optimize away RAM
 697 * accesses:
 698 */
 699static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
 700{
 701	if (g->p.data_reads)
 702		val += *data;
 703	if (g->p.data_writes)
 704		*data = val + 1;
 705	return val;
 706}
 707
 708/*
 709 * The worker process does two types of work, a forwards going
 710 * loop and a backwards going loop.
 711 *
 712 * We do this so that on multiprocessor systems we do not create
 713 * a 'train' of processing, with highly synchronized processes,
 714 * skewing the whole benchmark.
 715 */
 716static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
 717{
 718	long words = bytes/sizeof(u64);
 719	u64 *data = (void *)__data;
 720	long chunk_0, chunk_1;
 721	u64 *d0, *d, *d1;
 722	long off;
 723	long i;
 724
 725	BUG_ON(!data && words);
 726	BUG_ON(data && !words);
 727
 728	if (!data)
 729		return val;
 730
 731	/* Very simple memset() work variant: */
 732	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
 733		bzero(data, bytes);
 734		return val;
 735	}
 736
 737	/* Spread out by PID/TID nr and by loop nr: */
 738	chunk_0 = words/nr_max;
 739	chunk_1 = words/g->p.nr_loops;
 740	off = nr*chunk_0 + loop*chunk_1;
 741
 742	while (off >= words)
 743		off -= words;
 744
 745	if (g->p.data_rand_walk) {
 746		u32 lfsr = nr + loop + val;
 747		int j;
 748
 749		for (i = 0; i < words/1024; i++) {
 750			long start, end;
 751
 752			lfsr = lfsr_32(lfsr);
 753
 754			start = lfsr % words;
 755			end = min(start + 1024, words-1);
 756
 757			if (g->p.data_zero_memset) {
 758				bzero(data + start, (end-start) * sizeof(u64));
 759			} else {
 760				for (j = start; j < end; j++)
 761					val = access_data(data + j, val);
 762			}
 763		}
 764	} else if (!g->p.data_backwards || (nr + loop) & 1) {
 765
 766		d0 = data + off;
 767		d  = data + off + 1;
 768		d1 = data + words;
 769
 770		/* Process data forwards: */
 771		for (;;) {
 772			if (unlikely(d >= d1))
 773				d = data;
 774			if (unlikely(d == d0))
 775				break;
 776
 777			val = access_data(d, val);
 778
 779			d++;
 780		}
 781	} else {
 782		/* Process data backwards: */
 783
 784		d0 = data + off;
 785		d  = data + off - 1;
 786		d1 = data + words;
 787
 788		/* Process data forwards: */
 789		for (;;) {
 790			if (unlikely(d < data))
 791				d = data + words-1;
 792			if (unlikely(d == d0))
 793				break;
 794
 795			val = access_data(d, val);
 796
 797			d--;
 798		}
 799	}
 800
 801	return val;
 802}
 803
 804static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
 805{
 806	unsigned int cpu;
 807
 808	cpu = sched_getcpu();
 809
 810	g->threads[task_nr].curr_cpu = cpu;
 811	prctl(0, bytes_worked);
 812}
 813
 814#define MAX_NR_NODES	64
 815
 816/*
 817 * Count the number of nodes a process's threads
 818 * are spread out on.
 819 *
 820 * A count of 1 means that the process is compressed
 821 * to a single node. A count of g->p.nr_nodes means it's
 822 * spread out on the whole system.
 823 */
 824static int count_process_nodes(int process_nr)
 825{
 826	char node_present[MAX_NR_NODES] = { 0, };
 827	int nodes;
 828	int n, t;
 829
 830	for (t = 0; t < g->p.nr_threads; t++) {
 831		struct thread_data *td;
 832		int task_nr;
 833		int node;
 834
 835		task_nr = process_nr*g->p.nr_threads + t;
 836		td = g->threads + task_nr;
 837
 838		node = numa_node_of_cpu(td->curr_cpu);
 839		if (node < 0) /* curr_cpu was likely still -1 */
 840			return 0;
 841
 842		node_present[node] = 1;
 843	}
 844
 845	nodes = 0;
 846
 847	for (n = 0; n < MAX_NR_NODES; n++)
 848		nodes += node_present[n];
 849
 850	return nodes;
 851}
 852
 853/*
 854 * Count the number of distinct process-threads a node contains.
 855 *
 856 * A count of 1 means that the node contains only a single
 857 * process. If all nodes on the system contain at most one
 858 * process then we are well-converged.
 859 */
 860static int count_node_processes(int node)
 861{
 862	int processes = 0;
 863	int t, p;
 864
 865	for (p = 0; p < g->p.nr_proc; p++) {
 866		for (t = 0; t < g->p.nr_threads; t++) {
 867			struct thread_data *td;
 868			int task_nr;
 869			int n;
 870
 871			task_nr = p*g->p.nr_threads + t;
 872			td = g->threads + task_nr;
 873
 874			n = numa_node_of_cpu(td->curr_cpu);
 875			if (n == node) {
 876				processes++;
 877				break;
 878			}
 879		}
 880	}
 881
 882	return processes;
 883}
 884
 885static void calc_convergence_compression(int *strong)
 886{
 887	unsigned int nodes_min, nodes_max;
 888	int p;
 889
 890	nodes_min = -1;
 891	nodes_max =  0;
 892
 893	for (p = 0; p < g->p.nr_proc; p++) {
 894		unsigned int nodes = count_process_nodes(p);
 895
 896		if (!nodes) {
 897			*strong = 0;
 898			return;
 899		}
 900
 901		nodes_min = min(nodes, nodes_min);
 902		nodes_max = max(nodes, nodes_max);
 903	}
 904
 905	/* Strong convergence: all threads compress on a single node: */
 906	if (nodes_min == 1 && nodes_max == 1) {
 907		*strong = 1;
 908	} else {
 909		*strong = 0;
 910		tprintf(" {%d-%d}", nodes_min, nodes_max);
 911	}
 912}
 913
 914static void calc_convergence(double runtime_ns_max, double *convergence)
 915{
 916	unsigned int loops_done_min, loops_done_max;
 917	int process_groups;
 918	int nodes[MAX_NR_NODES];
 919	int distance;
 920	int nr_min;
 921	int nr_max;
 922	int strong;
 923	int sum;
 924	int nr;
 925	int node;
 926	int cpu;
 927	int t;
 928
 929	if (!g->p.show_convergence && !g->p.measure_convergence)
 930		return;
 931
 932	for (node = 0; node < g->p.nr_nodes; node++)
 933		nodes[node] = 0;
 934
 935	loops_done_min = -1;
 936	loops_done_max = 0;
 937
 938	for (t = 0; t < g->p.nr_tasks; t++) {
 939		struct thread_data *td = g->threads + t;
 940		unsigned int loops_done;
 941
 942		cpu = td->curr_cpu;
 943
 944		/* Not all threads have written it yet: */
 945		if (cpu < 0)
 946			continue;
 947
 948		node = numa_node_of_cpu(cpu);
 949
 950		nodes[node]++;
 951
 952		loops_done = td->loops_done;
 953		loops_done_min = min(loops_done, loops_done_min);
 954		loops_done_max = max(loops_done, loops_done_max);
 955	}
 956
 957	nr_max = 0;
 958	nr_min = g->p.nr_tasks;
 959	sum = 0;
 960
 961	for (node = 0; node < g->p.nr_nodes; node++) {
 962		nr = nodes[node];
 963		nr_min = min(nr, nr_min);
 964		nr_max = max(nr, nr_max);
 965		sum += nr;
 966	}
 967	BUG_ON(nr_min > nr_max);
 968
 969	BUG_ON(sum > g->p.nr_tasks);
 970
 971	if (0 && (sum < g->p.nr_tasks))
 972		return;
 973
 974	/*
 975	 * Count the number of distinct process groups present
 976	 * on nodes - when we are converged this will decrease
 977	 * to g->p.nr_proc:
 978	 */
 979	process_groups = 0;
 980
 981	for (node = 0; node < g->p.nr_nodes; node++) {
 982		int processes = count_node_processes(node);
 983
 984		nr = nodes[node];
 985		tprintf(" %2d/%-2d", nr, processes);
 986
 987		process_groups += processes;
 988	}
 989
 990	distance = nr_max - nr_min;
 991
 992	tprintf(" [%2d/%-2d]", distance, process_groups);
 993
 994	tprintf(" l:%3d-%-3d (%3d)",
 995		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
 996
 997	if (loops_done_min && loops_done_max) {
 998		double skew = 1.0 - (double)loops_done_min/loops_done_max;
 999
1000		tprintf(" [%4.1f%%]", skew * 100.0);
1001	}
1002
1003	calc_convergence_compression(&strong);
1004
1005	if (strong && process_groups == g->p.nr_proc) {
1006		if (!*convergence) {
1007			*convergence = runtime_ns_max;
1008			tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1009			if (g->p.measure_convergence) {
1010				g->all_converged = true;
1011				g->stop_work = true;
1012			}
1013		}
1014	} else {
1015		if (*convergence) {
1016			tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1017			*convergence = 0;
1018		}
1019		tprintf("\n");
1020	}
1021}
1022
1023static void show_summary(double runtime_ns_max, int l, double *convergence)
1024{
1025	tprintf("\r #  %5.1f%%  [%.1f mins]",
1026		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1027
1028	calc_convergence(runtime_ns_max, convergence);
1029
1030	if (g->p.show_details >= 0)
1031		fflush(stdout);
1032}
1033
1034static void *worker_thread(void *__tdata)
1035{
1036	struct thread_data *td = __tdata;
1037	struct timeval start0, start, stop, diff;
1038	int process_nr = td->process_nr;
1039	int thread_nr = td->thread_nr;
1040	unsigned long last_perturbance;
1041	int task_nr = td->task_nr;
1042	int details = g->p.show_details;
1043	int first_task, last_task;
1044	double convergence = 0;
1045	u64 val = td->val;
1046	double runtime_ns_max;
1047	u8 *global_data;
1048	u8 *process_data;
1049	u8 *thread_data;
1050	u64 bytes_done;
1051	long work_done;
1052	u32 l;
1053	struct rusage rusage;
1054
1055	bind_to_cpumask(td->bind_cpumask);
1056	bind_to_memnode(td->bind_node);
1057
1058	set_taskname("thread %d/%d", process_nr, thread_nr);
1059
1060	global_data = g->data;
1061	process_data = td->process_data;
1062	thread_data = setup_private_data(g->p.bytes_thread);
1063
1064	bytes_done = 0;
1065
1066	last_task = 0;
1067	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1068		last_task = 1;
1069
1070	first_task = 0;
1071	if (process_nr == 0 && thread_nr == 0)
1072		first_task = 1;
1073
1074	if (details >= 2) {
1075		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1076			process_nr, thread_nr, global_data, process_data, thread_data);
1077	}
1078
1079	if (g->p.serialize_startup) {
1080		pthread_mutex_lock(&g->startup_mutex);
1081		g->nr_tasks_started++;
1082		pthread_mutex_unlock(&g->startup_mutex);
1083
1084		/* Here we will wait for the main process to start us all at once: */
1085		pthread_mutex_lock(&g->start_work_mutex);
1086		g->nr_tasks_working++;
1087
1088		/* Last one wake the main process: */
1089		if (g->nr_tasks_working == g->p.nr_tasks)
1090			pthread_mutex_unlock(&g->startup_done_mutex);
1091
1092		pthread_mutex_unlock(&g->start_work_mutex);
1093	}
1094
1095	gettimeofday(&start0, NULL);
1096
1097	start = stop = start0;
1098	last_perturbance = start.tv_sec;
1099
1100	for (l = 0; l < g->p.nr_loops; l++) {
1101		start = stop;
1102
1103		if (g->stop_work)
1104			break;
1105
1106		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1107		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1108		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1109
1110		if (g->p.sleep_usecs) {
1111			pthread_mutex_lock(td->process_lock);
1112			usleep(g->p.sleep_usecs);
1113			pthread_mutex_unlock(td->process_lock);
1114		}
1115		/*
1116		 * Amount of work to be done under a process-global lock:
1117		 */
1118		if (g->p.bytes_process_locked) {
1119			pthread_mutex_lock(td->process_lock);
1120			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1121			pthread_mutex_unlock(td->process_lock);
1122		}
1123
1124		work_done = g->p.bytes_global + g->p.bytes_process +
1125			    g->p.bytes_process_locked + g->p.bytes_thread;
1126
1127		update_curr_cpu(task_nr, work_done);
1128		bytes_done += work_done;
1129
1130		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1131			continue;
1132
1133		td->loops_done = l;
1134
1135		gettimeofday(&stop, NULL);
1136
1137		/* Check whether our max runtime timed out: */
1138		if (g->p.nr_secs) {
1139			timersub(&stop, &start0, &diff);
1140			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1141				g->stop_work = true;
1142				break;
1143			}
1144		}
1145
1146		/* Update the summary at most once per second: */
1147		if (start.tv_sec == stop.tv_sec)
1148			continue;
1149
1150		/*
1151		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1152		 * by migrating to CPU#0:
1153		 */
1154		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1155			cpu_set_t orig_mask;
1156			int target_cpu;
1157			int this_cpu;
1158
1159			last_perturbance = stop.tv_sec;
1160
1161			/*
1162			 * Depending on where we are running, move into
1163			 * the other half of the system, to create some
1164			 * real disturbance:
1165			 */
1166			this_cpu = g->threads[task_nr].curr_cpu;
1167			if (this_cpu < g->p.nr_cpus/2)
1168				target_cpu = g->p.nr_cpus-1;
1169			else
1170				target_cpu = 0;
1171
1172			orig_mask = bind_to_cpu(target_cpu);
1173
1174			/* Here we are running on the target CPU already */
1175			if (details >= 1)
1176				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1177
1178			bind_to_cpumask(orig_mask);
1179		}
1180
1181		if (details >= 3) {
1182			timersub(&stop, &start, &diff);
1183			runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1184			runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1185
1186			if (details >= 0) {
1187				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1188					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1189			}
1190			fflush(stdout);
1191		}
1192		if (!last_task)
1193			continue;
1194
1195		timersub(&stop, &start0, &diff);
1196		runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1197		runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1198
1199		show_summary(runtime_ns_max, l, &convergence);
1200	}
1201
1202	gettimeofday(&stop, NULL);
1203	timersub(&stop, &start0, &diff);
1204	td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1205	td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1206	td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1207
1208	getrusage(RUSAGE_THREAD, &rusage);
1209	td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1210	td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1211	td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1212	td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1213
1214	free_data(thread_data, g->p.bytes_thread);
1215
1216	pthread_mutex_lock(&g->stop_work_mutex);
1217	g->bytes_done += bytes_done;
1218	pthread_mutex_unlock(&g->stop_work_mutex);
1219
1220	return NULL;
1221}
1222
1223/*
1224 * A worker process starts a couple of threads:
1225 */
1226static void worker_process(int process_nr)
1227{
1228	pthread_mutex_t process_lock;
1229	struct thread_data *td;
1230	pthread_t *pthreads;
1231	u8 *process_data;
1232	int task_nr;
1233	int ret;
1234	int t;
1235
1236	pthread_mutex_init(&process_lock, NULL);
1237	set_taskname("process %d", process_nr);
1238
1239	/*
1240	 * Pick up the memory policy and the CPU binding of our first thread,
1241	 * so that we initialize memory accordingly:
1242	 */
1243	task_nr = process_nr*g->p.nr_threads;
1244	td = g->threads + task_nr;
1245
1246	bind_to_memnode(td->bind_node);
1247	bind_to_cpumask(td->bind_cpumask);
1248
1249	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1250	process_data = setup_private_data(g->p.bytes_process);
1251
1252	if (g->p.show_details >= 3) {
1253		printf(" # process %2d global mem: %p, process mem: %p\n",
1254			process_nr, g->data, process_data);
1255	}
1256
1257	for (t = 0; t < g->p.nr_threads; t++) {
1258		task_nr = process_nr*g->p.nr_threads + t;
1259		td = g->threads + task_nr;
1260
1261		td->process_data = process_data;
1262		td->process_nr   = process_nr;
1263		td->thread_nr    = t;
1264		td->task_nr	 = task_nr;
1265		td->val          = rand();
1266		td->curr_cpu	 = -1;
1267		td->process_lock = &process_lock;
1268
1269		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1270		BUG_ON(ret);
1271	}
1272
1273	for (t = 0; t < g->p.nr_threads; t++) {
1274                ret = pthread_join(pthreads[t], NULL);
1275		BUG_ON(ret);
1276	}
1277
1278	free_data(process_data, g->p.bytes_process);
1279	free(pthreads);
1280}
1281
1282static void print_summary(void)
1283{
1284	if (g->p.show_details < 0)
1285		return;
1286
1287	printf("\n ###\n");
1288	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1289		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1290	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1291			g->p.nr_loops, g->p.bytes_global/1024/1024);
1292	printf(" #      %5dx %5ldMB process shared mem operations\n",
1293			g->p.nr_loops, g->p.bytes_process/1024/1024);
1294	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1295			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1296
1297	printf(" ###\n");
1298
1299	printf("\n ###\n"); fflush(stdout);
1300}
1301
1302static void init_thread_data(void)
1303{
1304	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1305	int t;
1306
1307	g->threads = zalloc_shared_data(size);
1308
1309	for (t = 0; t < g->p.nr_tasks; t++) {
1310		struct thread_data *td = g->threads + t;
1311		int cpu;
1312
1313		/* Allow all nodes by default: */
1314		td->bind_node = -1;
1315
1316		/* Allow all CPUs by default: */
1317		CPU_ZERO(&td->bind_cpumask);
1318		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1319			CPU_SET(cpu, &td->bind_cpumask);
1320	}
1321}
1322
1323static void deinit_thread_data(void)
1324{
1325	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1326
1327	free_data(g->threads, size);
1328}
1329
1330static int init(void)
1331{
1332	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1333
1334	/* Copy over options: */
1335	g->p = p0;
1336
1337	g->p.nr_cpus = numa_num_configured_cpus();
1338
1339	g->p.nr_nodes = numa_max_node() + 1;
1340
1341	/* char array in count_process_nodes(): */
1342	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1343
1344	if (g->p.show_quiet && !g->p.show_details)
1345		g->p.show_details = -1;
1346
1347	/* Some memory should be specified: */
1348	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1349		return -1;
1350
1351	if (g->p.mb_global_str) {
1352		g->p.mb_global = atof(g->p.mb_global_str);
1353		BUG_ON(g->p.mb_global < 0);
1354	}
1355
1356	if (g->p.mb_proc_str) {
1357		g->p.mb_proc = atof(g->p.mb_proc_str);
1358		BUG_ON(g->p.mb_proc < 0);
1359	}
1360
1361	if (g->p.mb_proc_locked_str) {
1362		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1363		BUG_ON(g->p.mb_proc_locked < 0);
1364		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1365	}
1366
1367	if (g->p.mb_thread_str) {
1368		g->p.mb_thread = atof(g->p.mb_thread_str);
1369		BUG_ON(g->p.mb_thread < 0);
1370	}
1371
1372	BUG_ON(g->p.nr_threads <= 0);
1373	BUG_ON(g->p.nr_proc <= 0);
1374
1375	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1376
1377	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1378	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1379	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1380	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1381
1382	g->data = setup_shared_data(g->p.bytes_global);
1383
1384	/* Startup serialization: */
1385	init_global_mutex(&g->start_work_mutex);
1386	init_global_mutex(&g->startup_mutex);
1387	init_global_mutex(&g->startup_done_mutex);
1388	init_global_mutex(&g->stop_work_mutex);
1389
1390	init_thread_data();
1391
1392	tprintf("#\n");
1393	if (parse_setup_cpu_list() || parse_setup_node_list())
1394		return -1;
1395	tprintf("#\n");
1396
1397	print_summary();
1398
1399	return 0;
1400}
1401
1402static void deinit(void)
1403{
1404	free_data(g->data, g->p.bytes_global);
1405	g->data = NULL;
1406
1407	deinit_thread_data();
1408
1409	free_data(g, sizeof(*g));
1410	g = NULL;
1411}
1412
1413/*
1414 * Print a short or long result, depending on the verbosity setting:
1415 */
1416static void print_res(const char *name, double val,
1417		      const char *txt_unit, const char *txt_short, const char *txt_long)
1418{
1419	if (!name)
1420		name = "main,";
1421
1422	if (!g->p.show_quiet)
1423		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1424	else
1425		printf(" %14.3f %s\n", val, txt_long);
1426}
1427
1428static int __bench_numa(const char *name)
1429{
1430	struct timeval start, stop, diff;
1431	u64 runtime_ns_min, runtime_ns_sum;
1432	pid_t *pids, pid, wpid;
1433	double delta_runtime;
1434	double runtime_avg;
1435	double runtime_sec_max;
1436	double runtime_sec_min;
1437	int wait_stat;
1438	double bytes;
1439	int i, t, p;
1440
1441	if (init())
1442		return -1;
1443
1444	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1445	pid = -1;
1446
1447	/* All threads try to acquire it, this way we can wait for them to start up: */
1448	pthread_mutex_lock(&g->start_work_mutex);
1449
1450	if (g->p.serialize_startup) {
1451		tprintf(" #\n");
1452		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1453	}
1454
1455	gettimeofday(&start, NULL);
1456
1457	for (i = 0; i < g->p.nr_proc; i++) {
1458		pid = fork();
1459		dprintf(" # process %2d: PID %d\n", i, pid);
1460
1461		BUG_ON(pid < 0);
1462		if (!pid) {
1463			/* Child process: */
1464			worker_process(i);
1465
1466			exit(0);
1467		}
1468		pids[i] = pid;
1469
1470	}
1471	/* Wait for all the threads to start up: */
1472	while (g->nr_tasks_started != g->p.nr_tasks)
1473		usleep(USEC_PER_MSEC);
1474
1475	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1476
1477	if (g->p.serialize_startup) {
1478		double startup_sec;
1479
1480		pthread_mutex_lock(&g->startup_done_mutex);
1481
1482		/* This will start all threads: */
1483		pthread_mutex_unlock(&g->start_work_mutex);
1484
1485		/* This mutex is locked - the last started thread will wake us: */
1486		pthread_mutex_lock(&g->startup_done_mutex);
1487
1488		gettimeofday(&stop, NULL);
1489
1490		timersub(&stop, &start, &diff);
1491
1492		startup_sec = diff.tv_sec * NSEC_PER_SEC;
1493		startup_sec += diff.tv_usec * NSEC_PER_USEC;
1494		startup_sec /= NSEC_PER_SEC;
1495
1496		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1497		tprintf(" #\n");
1498
1499		start = stop;
1500		pthread_mutex_unlock(&g->startup_done_mutex);
1501	} else {
1502		gettimeofday(&start, NULL);
1503	}
1504
1505	/* Parent process: */
1506
1507
1508	for (i = 0; i < g->p.nr_proc; i++) {
1509		wpid = waitpid(pids[i], &wait_stat, 0);
1510		BUG_ON(wpid < 0);
1511		BUG_ON(!WIFEXITED(wait_stat));
1512
1513	}
1514
1515	runtime_ns_sum = 0;
1516	runtime_ns_min = -1LL;
1517
1518	for (t = 0; t < g->p.nr_tasks; t++) {
1519		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1520
1521		runtime_ns_sum += thread_runtime_ns;
1522		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1523	}
1524
1525	gettimeofday(&stop, NULL);
1526	timersub(&stop, &start, &diff);
1527
1528	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1529
1530	tprintf("\n ###\n");
1531	tprintf("\n");
1532
1533	runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1534	runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1535	runtime_sec_max /= NSEC_PER_SEC;
1536
1537	runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1538
1539	bytes = g->bytes_done;
1540	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1541
1542	if (g->p.measure_convergence) {
1543		print_res(name, runtime_sec_max,
1544			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1545	}
1546
1547	print_res(name, runtime_sec_max,
1548		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1549
1550	print_res(name, runtime_sec_min,
1551		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1552
1553	print_res(name, runtime_avg,
1554		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1555
1556	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1557	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1558		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1559
1560	print_res(name, bytes / g->p.nr_tasks / 1e9,
1561		"GB,", "data/thread",		"GB data processed, per thread");
1562
1563	print_res(name, bytes / 1e9,
1564		"GB,", "data-total",		"GB data processed, total");
1565
1566	print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1567		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1568
1569	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1570		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1571
1572	print_res(name, bytes / runtime_sec_max / 1e9,
1573		"GB/sec,", "total-speed",	"GB/sec total speed");
1574
1575	if (g->p.show_details >= 2) {
1576		char tname[32];
1577		struct thread_data *td;
1578		for (p = 0; p < g->p.nr_proc; p++) {
1579			for (t = 0; t < g->p.nr_threads; t++) {
1580				memset(tname, 0, 32);
1581				td = g->threads + p*g->p.nr_threads + t;
1582				snprintf(tname, 32, "process%d:thread%d", p, t);
1583				print_res(tname, td->speed_gbs,
1584					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1585				print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1586					"secs",	"thread-system-time", "system CPU time/thread");
1587				print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1588					"secs",	"thread-user-time", "user CPU time/thread");
1589			}
1590		}
1591	}
1592
1593	free(pids);
1594
1595	deinit();
1596
1597	return 0;
1598}
1599
1600#define MAX_ARGS 50
1601
1602static int command_size(const char **argv)
1603{
1604	int size = 0;
1605
1606	while (*argv) {
1607		size++;
1608		argv++;
1609	}
1610
1611	BUG_ON(size >= MAX_ARGS);
1612
1613	return size;
1614}
1615
1616static void init_params(struct params *p, const char *name, int argc, const char **argv)
1617{
1618	int i;
1619
1620	printf("\n # Running %s \"perf bench numa", name);
1621
1622	for (i = 0; i < argc; i++)
1623		printf(" %s", argv[i]);
1624
1625	printf("\"\n");
1626
1627	memset(p, 0, sizeof(*p));
1628
1629	/* Initialize nonzero defaults: */
1630
1631	p->serialize_startup		= 1;
1632	p->data_reads			= true;
1633	p->data_writes			= true;
1634	p->data_backwards		= true;
1635	p->data_rand_walk		= true;
1636	p->nr_loops			= -1;
1637	p->init_random			= true;
1638	p->mb_global_str		= "1";
1639	p->nr_proc			= 1;
1640	p->nr_threads			= 1;
1641	p->nr_secs			= 5;
1642	p->run_all			= argc == 1;
1643}
1644
1645static int run_bench_numa(const char *name, const char **argv)
1646{
1647	int argc = command_size(argv);
1648
1649	init_params(&p0, name, argc, argv);
1650	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1651	if (argc)
1652		goto err;
1653
1654	if (__bench_numa(name))
1655		goto err;
1656
1657	return 0;
1658
1659err:
1660	return -1;
1661}
1662
1663#define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1664#define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1665
1666#define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1667#define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1668
1669#define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1670#define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1671
1672/*
1673 * The built-in test-suite executed by "perf bench numa -a".
1674 *
1675 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1676 */
1677static const char *tests[][MAX_ARGS] = {
1678   /* Basic single-stream NUMA bandwidth measurements: */
1679   { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1680			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1681   { "RAM-bw-local-NOTHP,",
1682			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1683			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1684   { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1685			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1686
1687   /* 2-stream NUMA bandwidth measurements: */
1688   { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1689			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1690   { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1691		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1692
1693   /* Cross-stream NUMA bandwidth measurement: */
1694   { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1695		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1696
1697   /* Convergence latency measurements: */
1698   { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1699   { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1700   { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1701   { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1702   { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1703   { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1704   { " 4x4-convergence-NOTHP,",
1705			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1706   { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1707   { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1708   { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1709   { " 8x4-convergence-NOTHP,",
1710			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1711   { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1712   { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1713   { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1714   { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1715   { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1716
1717   /* Various NUMA process/thread layout bandwidth measurements: */
1718   { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1719   { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1720   { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1721   { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1722   { " 8x1-bw-process-NOTHP,",
1723			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1724   { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1725
1726   { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1727   { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1728   { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1729   { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1730
1731   { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1732   { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1733   { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1734   { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1735   { " 4x8-bw-thread-NOTHP,",
1736			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1737   { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1738   { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1739
1740   { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1741   { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1742
1743   { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1744   { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1745   { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1746   { "numa01-bw-thread-NOTHP,",
1747			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1748};
1749
1750static int bench_all(void)
1751{
1752	int nr = ARRAY_SIZE(tests);
1753	int ret;
1754	int i;
1755
1756	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1757	BUG_ON(ret < 0);
1758
1759	for (i = 0; i < nr; i++) {
1760		run_bench_numa(tests[i][0], tests[i] + 1);
1761	}
1762
1763	printf("\n");
1764
1765	return 0;
1766}
1767
1768int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1769{
1770	init_params(&p0, "main,", argc, argv);
1771	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1772	if (argc)
1773		goto err;
1774
1775	if (p0.run_all)
1776		return bench_all();
1777
1778	if (__bench_numa(NULL))
1779		goto err;
1780
1781	return 0;
1782
1783err:
1784	usage_with_options(numa_usage, options);
1785	return -1;
1786}