xref: /openbmc/linux/tools/perf/bench/numa.c (revision e23feb16)
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 "../util/parse-options.h"
11 
12 #include "bench.h"
13 
14 #include <errno.h>
15 #include <sched.h>
16 #include <stdio.h>
17 #include <assert.h>
18 #include <malloc.h>
19 #include <signal.h>
20 #include <stdlib.h>
21 #include <string.h>
22 #include <unistd.h>
23 #include <pthread.h>
24 #include <sys/mman.h>
25 #include <sys/time.h>
26 #include <sys/wait.h>
27 #include <sys/prctl.h>
28 #include <sys/types.h>
29 
30 #include <numa.h>
31 #include <numaif.h>
32 
33 /*
34  * Regular printout to the terminal, supressed if -q is specified:
35  */
36 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
37 
38 /*
39  * Debug printf:
40  */
41 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
42 
43 struct thread_data {
44 	int			curr_cpu;
45 	cpu_set_t		bind_cpumask;
46 	int			bind_node;
47 	u8			*process_data;
48 	int			process_nr;
49 	int			thread_nr;
50 	int			task_nr;
51 	unsigned int		loops_done;
52 	u64			val;
53 	u64			runtime_ns;
54 	pthread_mutex_t		*process_lock;
55 };
56 
57 /* Parameters set by options: */
58 
59 struct params {
60 	/* Startup synchronization: */
61 	bool			serialize_startup;
62 
63 	/* Task hierarchy: */
64 	int			nr_proc;
65 	int			nr_threads;
66 
67 	/* Working set sizes: */
68 	const char		*mb_global_str;
69 	const char		*mb_proc_str;
70 	const char		*mb_proc_locked_str;
71 	const char		*mb_thread_str;
72 
73 	double			mb_global;
74 	double			mb_proc;
75 	double			mb_proc_locked;
76 	double			mb_thread;
77 
78 	/* Access patterns to the working set: */
79 	bool			data_reads;
80 	bool			data_writes;
81 	bool			data_backwards;
82 	bool			data_zero_memset;
83 	bool			data_rand_walk;
84 	u32			nr_loops;
85 	u32			nr_secs;
86 	u32			sleep_usecs;
87 
88 	/* Working set initialization: */
89 	bool			init_zero;
90 	bool			init_random;
91 	bool			init_cpu0;
92 
93 	/* Misc options: */
94 	int			show_details;
95 	int			run_all;
96 	int			thp;
97 
98 	long			bytes_global;
99 	long			bytes_process;
100 	long			bytes_process_locked;
101 	long			bytes_thread;
102 
103 	int			nr_tasks;
104 	bool			show_quiet;
105 
106 	bool			show_convergence;
107 	bool			measure_convergence;
108 
109 	int			perturb_secs;
110 	int			nr_cpus;
111 	int			nr_nodes;
112 
113 	/* Affinity options -C and -N: */
114 	char			*cpu_list_str;
115 	char			*node_list_str;
116 };
117 
118 
119 /* Global, read-writable area, accessible to all processes and threads: */
120 
121 struct global_info {
122 	u8			*data;
123 
124 	pthread_mutex_t		startup_mutex;
125 	int			nr_tasks_started;
126 
127 	pthread_mutex_t		startup_done_mutex;
128 
129 	pthread_mutex_t		start_work_mutex;
130 	int			nr_tasks_working;
131 
132 	pthread_mutex_t		stop_work_mutex;
133 	u64			bytes_done;
134 
135 	struct thread_data	*threads;
136 
137 	/* Convergence latency measurement: */
138 	bool			all_converged;
139 	bool			stop_work;
140 
141 	int			print_once;
142 
143 	struct params		p;
144 };
145 
146 static struct global_info	*g = NULL;
147 
148 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
149 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
150 
151 struct params p0;
152 
153 static const struct option options[] = {
154 	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
155 	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
156 
157 	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
158 	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
159 	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
160 	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
161 
162 	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run"),
163 	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run"),
164 	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
165 
166 	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via writes (can be mixed with -W)"),
167 	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
168 	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
169 	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
170 	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
171 
172 
173 	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
174 	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
175 	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
176 	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
177 
178 	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
179 	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
180 	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
181 	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
182 	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
183 	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"bzero the initial allocations"),
184 	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
185 
186 	/* Special option string parsing callbacks: */
187         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
188 			"bind the first N tasks to these specific cpus (the rest is unbound)",
189 			parse_cpus_opt),
190         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
191 			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
192 			parse_nodes_opt),
193 	OPT_END()
194 };
195 
196 static const char * const bench_numa_usage[] = {
197 	"perf bench numa <options>",
198 	NULL
199 };
200 
201 static const char * const numa_usage[] = {
202 	"perf bench numa mem [<options>]",
203 	NULL
204 };
205 
206 static cpu_set_t bind_to_cpu(int target_cpu)
207 {
208 	cpu_set_t orig_mask, mask;
209 	int ret;
210 
211 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
212 	BUG_ON(ret);
213 
214 	CPU_ZERO(&mask);
215 
216 	if (target_cpu == -1) {
217 		int cpu;
218 
219 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
220 			CPU_SET(cpu, &mask);
221 	} else {
222 		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
223 		CPU_SET(target_cpu, &mask);
224 	}
225 
226 	ret = sched_setaffinity(0, sizeof(mask), &mask);
227 	BUG_ON(ret);
228 
229 	return orig_mask;
230 }
231 
232 static cpu_set_t bind_to_node(int target_node)
233 {
234 	int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
235 	cpu_set_t orig_mask, mask;
236 	int cpu;
237 	int ret;
238 
239 	BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
240 	BUG_ON(!cpus_per_node);
241 
242 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
243 	BUG_ON(ret);
244 
245 	CPU_ZERO(&mask);
246 
247 	if (target_node == -1) {
248 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
249 			CPU_SET(cpu, &mask);
250 	} else {
251 		int cpu_start = (target_node + 0) * cpus_per_node;
252 		int cpu_stop  = (target_node + 1) * cpus_per_node;
253 
254 		BUG_ON(cpu_stop > g->p.nr_cpus);
255 
256 		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
257 			CPU_SET(cpu, &mask);
258 	}
259 
260 	ret = sched_setaffinity(0, sizeof(mask), &mask);
261 	BUG_ON(ret);
262 
263 	return orig_mask;
264 }
265 
266 static void bind_to_cpumask(cpu_set_t mask)
267 {
268 	int ret;
269 
270 	ret = sched_setaffinity(0, sizeof(mask), &mask);
271 	BUG_ON(ret);
272 }
273 
274 static void mempol_restore(void)
275 {
276 	int ret;
277 
278 	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
279 
280 	BUG_ON(ret);
281 }
282 
283 static void bind_to_memnode(int node)
284 {
285 	unsigned long nodemask;
286 	int ret;
287 
288 	if (node == -1)
289 		return;
290 
291 	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask));
292 	nodemask = 1L << node;
293 
294 	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
295 	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
296 
297 	BUG_ON(ret);
298 }
299 
300 #define HPSIZE (2*1024*1024)
301 
302 #define set_taskname(fmt...)				\
303 do {							\
304 	char name[20];					\
305 							\
306 	snprintf(name, 20, fmt);			\
307 	prctl(PR_SET_NAME, name);			\
308 } while (0)
309 
310 static u8 *alloc_data(ssize_t bytes0, int map_flags,
311 		      int init_zero, int init_cpu0, int thp, int init_random)
312 {
313 	cpu_set_t orig_mask;
314 	ssize_t bytes;
315 	u8 *buf;
316 	int ret;
317 
318 	if (!bytes0)
319 		return NULL;
320 
321 	/* Allocate and initialize all memory on CPU#0: */
322 	if (init_cpu0) {
323 		orig_mask = bind_to_node(0);
324 		bind_to_memnode(0);
325 	}
326 
327 	bytes = bytes0 + HPSIZE;
328 
329 	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
330 	BUG_ON(buf == (void *)-1);
331 
332 	if (map_flags == MAP_PRIVATE) {
333 		if (thp > 0) {
334 			ret = madvise(buf, bytes, MADV_HUGEPAGE);
335 			if (ret && !g->print_once) {
336 				g->print_once = 1;
337 				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
338 			}
339 		}
340 		if (thp < 0) {
341 			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
342 			if (ret && !g->print_once) {
343 				g->print_once = 1;
344 				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
345 			}
346 		}
347 	}
348 
349 	if (init_zero) {
350 		bzero(buf, bytes);
351 	} else {
352 		/* Initialize random contents, different in each word: */
353 		if (init_random) {
354 			u64 *wbuf = (void *)buf;
355 			long off = rand();
356 			long i;
357 
358 			for (i = 0; i < bytes/8; i++)
359 				wbuf[i] = i + off;
360 		}
361 	}
362 
363 	/* Align to 2MB boundary: */
364 	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
365 
366 	/* Restore affinity: */
367 	if (init_cpu0) {
368 		bind_to_cpumask(orig_mask);
369 		mempol_restore();
370 	}
371 
372 	return buf;
373 }
374 
375 static void free_data(void *data, ssize_t bytes)
376 {
377 	int ret;
378 
379 	if (!data)
380 		return;
381 
382 	ret = munmap(data, bytes);
383 	BUG_ON(ret);
384 }
385 
386 /*
387  * Create a shared memory buffer that can be shared between processes, zeroed:
388  */
389 static void * zalloc_shared_data(ssize_t bytes)
390 {
391 	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
392 }
393 
394 /*
395  * Create a shared memory buffer that can be shared between processes:
396  */
397 static void * setup_shared_data(ssize_t bytes)
398 {
399 	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
400 }
401 
402 /*
403  * Allocate process-local memory - this will either be shared between
404  * threads of this process, or only be accessed by this thread:
405  */
406 static void * setup_private_data(ssize_t bytes)
407 {
408 	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
409 }
410 
411 /*
412  * Return a process-shared (global) mutex:
413  */
414 static void init_global_mutex(pthread_mutex_t *mutex)
415 {
416 	pthread_mutexattr_t attr;
417 
418 	pthread_mutexattr_init(&attr);
419 	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
420 	pthread_mutex_init(mutex, &attr);
421 }
422 
423 static int parse_cpu_list(const char *arg)
424 {
425 	p0.cpu_list_str = strdup(arg);
426 
427 	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
428 
429 	return 0;
430 }
431 
432 static void parse_setup_cpu_list(void)
433 {
434 	struct thread_data *td;
435 	char *str0, *str;
436 	int t;
437 
438 	if (!g->p.cpu_list_str)
439 		return;
440 
441 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
442 
443 	str0 = str = strdup(g->p.cpu_list_str);
444 	t = 0;
445 
446 	BUG_ON(!str);
447 
448 	tprintf("# binding tasks to CPUs:\n");
449 	tprintf("#  ");
450 
451 	while (true) {
452 		int bind_cpu, bind_cpu_0, bind_cpu_1;
453 		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
454 		int bind_len;
455 		int step;
456 		int mul;
457 
458 		tok = strsep(&str, ",");
459 		if (!tok)
460 			break;
461 
462 		tok_end = strstr(tok, "-");
463 
464 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
465 		if (!tok_end) {
466 			/* Single CPU specified: */
467 			bind_cpu_0 = bind_cpu_1 = atol(tok);
468 		} else {
469 			/* CPU range specified (for example: "5-11"): */
470 			bind_cpu_0 = atol(tok);
471 			bind_cpu_1 = atol(tok_end + 1);
472 		}
473 
474 		step = 1;
475 		tok_step = strstr(tok, "#");
476 		if (tok_step) {
477 			step = atol(tok_step + 1);
478 			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
479 		}
480 
481 		/*
482 		 * Mask length.
483 		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
484 		 * where the _4 means the next 4 CPUs are allowed.
485 		 */
486 		bind_len = 1;
487 		tok_len = strstr(tok, "_");
488 		if (tok_len) {
489 			bind_len = atol(tok_len + 1);
490 			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
491 		}
492 
493 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
494 		mul = 1;
495 		tok_mul = strstr(tok, "x");
496 		if (tok_mul) {
497 			mul = atol(tok_mul + 1);
498 			BUG_ON(mul <= 0);
499 		}
500 
501 		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
502 
503 		BUG_ON(bind_cpu_0 < 0 || bind_cpu_0 >= g->p.nr_cpus);
504 		BUG_ON(bind_cpu_1 < 0 || bind_cpu_1 >= g->p.nr_cpus);
505 		BUG_ON(bind_cpu_0 > bind_cpu_1);
506 
507 		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
508 			int i;
509 
510 			for (i = 0; i < mul; i++) {
511 				int cpu;
512 
513 				if (t >= g->p.nr_tasks) {
514 					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
515 					goto out;
516 				}
517 				td = g->threads + t;
518 
519 				if (t)
520 					tprintf(",");
521 				if (bind_len > 1) {
522 					tprintf("%2d/%d", bind_cpu, bind_len);
523 				} else {
524 					tprintf("%2d", bind_cpu);
525 				}
526 
527 				CPU_ZERO(&td->bind_cpumask);
528 				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
529 					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
530 					CPU_SET(cpu, &td->bind_cpumask);
531 				}
532 				t++;
533 			}
534 		}
535 	}
536 out:
537 
538 	tprintf("\n");
539 
540 	if (t < g->p.nr_tasks)
541 		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
542 
543 	free(str0);
544 }
545 
546 static int parse_cpus_opt(const struct option *opt __maybe_unused,
547 			  const char *arg, int unset __maybe_unused)
548 {
549 	if (!arg)
550 		return -1;
551 
552 	return parse_cpu_list(arg);
553 }
554 
555 static int parse_node_list(const char *arg)
556 {
557 	p0.node_list_str = strdup(arg);
558 
559 	dprintf("got NODE list: {%s}\n", p0.node_list_str);
560 
561 	return 0;
562 }
563 
564 static void parse_setup_node_list(void)
565 {
566 	struct thread_data *td;
567 	char *str0, *str;
568 	int t;
569 
570 	if (!g->p.node_list_str)
571 		return;
572 
573 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
574 
575 	str0 = str = strdup(g->p.node_list_str);
576 	t = 0;
577 
578 	BUG_ON(!str);
579 
580 	tprintf("# binding tasks to NODEs:\n");
581 	tprintf("# ");
582 
583 	while (true) {
584 		int bind_node, bind_node_0, bind_node_1;
585 		char *tok, *tok_end, *tok_step, *tok_mul;
586 		int step;
587 		int mul;
588 
589 		tok = strsep(&str, ",");
590 		if (!tok)
591 			break;
592 
593 		tok_end = strstr(tok, "-");
594 
595 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
596 		if (!tok_end) {
597 			/* Single NODE specified: */
598 			bind_node_0 = bind_node_1 = atol(tok);
599 		} else {
600 			/* NODE range specified (for example: "5-11"): */
601 			bind_node_0 = atol(tok);
602 			bind_node_1 = atol(tok_end + 1);
603 		}
604 
605 		step = 1;
606 		tok_step = strstr(tok, "#");
607 		if (tok_step) {
608 			step = atol(tok_step + 1);
609 			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
610 		}
611 
612 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
613 		mul = 1;
614 		tok_mul = strstr(tok, "x");
615 		if (tok_mul) {
616 			mul = atol(tok_mul + 1);
617 			BUG_ON(mul <= 0);
618 		}
619 
620 		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
621 
622 		BUG_ON(bind_node_0 < 0 || bind_node_0 >= g->p.nr_nodes);
623 		BUG_ON(bind_node_1 < 0 || bind_node_1 >= g->p.nr_nodes);
624 		BUG_ON(bind_node_0 > bind_node_1);
625 
626 		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
627 			int i;
628 
629 			for (i = 0; i < mul; i++) {
630 				if (t >= g->p.nr_tasks) {
631 					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
632 					goto out;
633 				}
634 				td = g->threads + t;
635 
636 				if (!t)
637 					tprintf(" %2d", bind_node);
638 				else
639 					tprintf(",%2d", bind_node);
640 
641 				td->bind_node = bind_node;
642 				t++;
643 			}
644 		}
645 	}
646 out:
647 
648 	tprintf("\n");
649 
650 	if (t < g->p.nr_tasks)
651 		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
652 
653 	free(str0);
654 }
655 
656 static int parse_nodes_opt(const struct option *opt __maybe_unused,
657 			  const char *arg, int unset __maybe_unused)
658 {
659 	if (!arg)
660 		return -1;
661 
662 	return parse_node_list(arg);
663 
664 	return 0;
665 }
666 
667 #define BIT(x) (1ul << x)
668 
669 static inline uint32_t lfsr_32(uint32_t lfsr)
670 {
671 	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
672 	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
673 }
674 
675 /*
676  * Make sure there's real data dependency to RAM (when read
677  * accesses are enabled), so the compiler, the CPU and the
678  * kernel (KSM, zero page, etc.) cannot optimize away RAM
679  * accesses:
680  */
681 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
682 {
683 	if (g->p.data_reads)
684 		val += *data;
685 	if (g->p.data_writes)
686 		*data = val + 1;
687 	return val;
688 }
689 
690 /*
691  * The worker process does two types of work, a forwards going
692  * loop and a backwards going loop.
693  *
694  * We do this so that on multiprocessor systems we do not create
695  * a 'train' of processing, with highly synchronized processes,
696  * skewing the whole benchmark.
697  */
698 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
699 {
700 	long words = bytes/sizeof(u64);
701 	u64 *data = (void *)__data;
702 	long chunk_0, chunk_1;
703 	u64 *d0, *d, *d1;
704 	long off;
705 	long i;
706 
707 	BUG_ON(!data && words);
708 	BUG_ON(data && !words);
709 
710 	if (!data)
711 		return val;
712 
713 	/* Very simple memset() work variant: */
714 	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
715 		bzero(data, bytes);
716 		return val;
717 	}
718 
719 	/* Spread out by PID/TID nr and by loop nr: */
720 	chunk_0 = words/nr_max;
721 	chunk_1 = words/g->p.nr_loops;
722 	off = nr*chunk_0 + loop*chunk_1;
723 
724 	while (off >= words)
725 		off -= words;
726 
727 	if (g->p.data_rand_walk) {
728 		u32 lfsr = nr + loop + val;
729 		int j;
730 
731 		for (i = 0; i < words/1024; i++) {
732 			long start, end;
733 
734 			lfsr = lfsr_32(lfsr);
735 
736 			start = lfsr % words;
737 			end = min(start + 1024, words-1);
738 
739 			if (g->p.data_zero_memset) {
740 				bzero(data + start, (end-start) * sizeof(u64));
741 			} else {
742 				for (j = start; j < end; j++)
743 					val = access_data(data + j, val);
744 			}
745 		}
746 	} else if (!g->p.data_backwards || (nr + loop) & 1) {
747 
748 		d0 = data + off;
749 		d  = data + off + 1;
750 		d1 = data + words;
751 
752 		/* Process data forwards: */
753 		for (;;) {
754 			if (unlikely(d >= d1))
755 				d = data;
756 			if (unlikely(d == d0))
757 				break;
758 
759 			val = access_data(d, val);
760 
761 			d++;
762 		}
763 	} else {
764 		/* Process data backwards: */
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 < data))
773 				d = data + words-1;
774 			if (unlikely(d == d0))
775 				break;
776 
777 			val = access_data(d, val);
778 
779 			d--;
780 		}
781 	}
782 
783 	return val;
784 }
785 
786 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
787 {
788 	unsigned int cpu;
789 
790 	cpu = sched_getcpu();
791 
792 	g->threads[task_nr].curr_cpu = cpu;
793 	prctl(0, bytes_worked);
794 }
795 
796 #define MAX_NR_NODES	64
797 
798 /*
799  * Count the number of nodes a process's threads
800  * are spread out on.
801  *
802  * A count of 1 means that the process is compressed
803  * to a single node. A count of g->p.nr_nodes means it's
804  * spread out on the whole system.
805  */
806 static int count_process_nodes(int process_nr)
807 {
808 	char node_present[MAX_NR_NODES] = { 0, };
809 	int nodes;
810 	int n, t;
811 
812 	for (t = 0; t < g->p.nr_threads; t++) {
813 		struct thread_data *td;
814 		int task_nr;
815 		int node;
816 
817 		task_nr = process_nr*g->p.nr_threads + t;
818 		td = g->threads + task_nr;
819 
820 		node = numa_node_of_cpu(td->curr_cpu);
821 		node_present[node] = 1;
822 	}
823 
824 	nodes = 0;
825 
826 	for (n = 0; n < MAX_NR_NODES; n++)
827 		nodes += node_present[n];
828 
829 	return nodes;
830 }
831 
832 /*
833  * Count the number of distinct process-threads a node contains.
834  *
835  * A count of 1 means that the node contains only a single
836  * process. If all nodes on the system contain at most one
837  * process then we are well-converged.
838  */
839 static int count_node_processes(int node)
840 {
841 	int processes = 0;
842 	int t, p;
843 
844 	for (p = 0; p < g->p.nr_proc; p++) {
845 		for (t = 0; t < g->p.nr_threads; t++) {
846 			struct thread_data *td;
847 			int task_nr;
848 			int n;
849 
850 			task_nr = p*g->p.nr_threads + t;
851 			td = g->threads + task_nr;
852 
853 			n = numa_node_of_cpu(td->curr_cpu);
854 			if (n == node) {
855 				processes++;
856 				break;
857 			}
858 		}
859 	}
860 
861 	return processes;
862 }
863 
864 static void calc_convergence_compression(int *strong)
865 {
866 	unsigned int nodes_min, nodes_max;
867 	int p;
868 
869 	nodes_min = -1;
870 	nodes_max =  0;
871 
872 	for (p = 0; p < g->p.nr_proc; p++) {
873 		unsigned int nodes = count_process_nodes(p);
874 
875 		nodes_min = min(nodes, nodes_min);
876 		nodes_max = max(nodes, nodes_max);
877 	}
878 
879 	/* Strong convergence: all threads compress on a single node: */
880 	if (nodes_min == 1 && nodes_max == 1) {
881 		*strong = 1;
882 	} else {
883 		*strong = 0;
884 		tprintf(" {%d-%d}", nodes_min, nodes_max);
885 	}
886 }
887 
888 static void calc_convergence(double runtime_ns_max, double *convergence)
889 {
890 	unsigned int loops_done_min, loops_done_max;
891 	int process_groups;
892 	int nodes[MAX_NR_NODES];
893 	int distance;
894 	int nr_min;
895 	int nr_max;
896 	int strong;
897 	int sum;
898 	int nr;
899 	int node;
900 	int cpu;
901 	int t;
902 
903 	if (!g->p.show_convergence && !g->p.measure_convergence)
904 		return;
905 
906 	for (node = 0; node < g->p.nr_nodes; node++)
907 		nodes[node] = 0;
908 
909 	loops_done_min = -1;
910 	loops_done_max = 0;
911 
912 	for (t = 0; t < g->p.nr_tasks; t++) {
913 		struct thread_data *td = g->threads + t;
914 		unsigned int loops_done;
915 
916 		cpu = td->curr_cpu;
917 
918 		/* Not all threads have written it yet: */
919 		if (cpu < 0)
920 			continue;
921 
922 		node = numa_node_of_cpu(cpu);
923 
924 		nodes[node]++;
925 
926 		loops_done = td->loops_done;
927 		loops_done_min = min(loops_done, loops_done_min);
928 		loops_done_max = max(loops_done, loops_done_max);
929 	}
930 
931 	nr_max = 0;
932 	nr_min = g->p.nr_tasks;
933 	sum = 0;
934 
935 	for (node = 0; node < g->p.nr_nodes; node++) {
936 		nr = nodes[node];
937 		nr_min = min(nr, nr_min);
938 		nr_max = max(nr, nr_max);
939 		sum += nr;
940 	}
941 	BUG_ON(nr_min > nr_max);
942 
943 	BUG_ON(sum > g->p.nr_tasks);
944 
945 	if (0 && (sum < g->p.nr_tasks))
946 		return;
947 
948 	/*
949 	 * Count the number of distinct process groups present
950 	 * on nodes - when we are converged this will decrease
951 	 * to g->p.nr_proc:
952 	 */
953 	process_groups = 0;
954 
955 	for (node = 0; node < g->p.nr_nodes; node++) {
956 		int processes = count_node_processes(node);
957 
958 		nr = nodes[node];
959 		tprintf(" %2d/%-2d", nr, processes);
960 
961 		process_groups += processes;
962 	}
963 
964 	distance = nr_max - nr_min;
965 
966 	tprintf(" [%2d/%-2d]", distance, process_groups);
967 
968 	tprintf(" l:%3d-%-3d (%3d)",
969 		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
970 
971 	if (loops_done_min && loops_done_max) {
972 		double skew = 1.0 - (double)loops_done_min/loops_done_max;
973 
974 		tprintf(" [%4.1f%%]", skew * 100.0);
975 	}
976 
977 	calc_convergence_compression(&strong);
978 
979 	if (strong && process_groups == g->p.nr_proc) {
980 		if (!*convergence) {
981 			*convergence = runtime_ns_max;
982 			tprintf(" (%6.1fs converged)\n", *convergence/1e9);
983 			if (g->p.measure_convergence) {
984 				g->all_converged = true;
985 				g->stop_work = true;
986 			}
987 		}
988 	} else {
989 		if (*convergence) {
990 			tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
991 			*convergence = 0;
992 		}
993 		tprintf("\n");
994 	}
995 }
996 
997 static void show_summary(double runtime_ns_max, int l, double *convergence)
998 {
999 	tprintf("\r #  %5.1f%%  [%.1f mins]",
1000 		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1001 
1002 	calc_convergence(runtime_ns_max, convergence);
1003 
1004 	if (g->p.show_details >= 0)
1005 		fflush(stdout);
1006 }
1007 
1008 static void *worker_thread(void *__tdata)
1009 {
1010 	struct thread_data *td = __tdata;
1011 	struct timeval start0, start, stop, diff;
1012 	int process_nr = td->process_nr;
1013 	int thread_nr = td->thread_nr;
1014 	unsigned long last_perturbance;
1015 	int task_nr = td->task_nr;
1016 	int details = g->p.show_details;
1017 	int first_task, last_task;
1018 	double convergence = 0;
1019 	u64 val = td->val;
1020 	double runtime_ns_max;
1021 	u8 *global_data;
1022 	u8 *process_data;
1023 	u8 *thread_data;
1024 	u64 bytes_done;
1025 	long work_done;
1026 	u32 l;
1027 
1028 	bind_to_cpumask(td->bind_cpumask);
1029 	bind_to_memnode(td->bind_node);
1030 
1031 	set_taskname("thread %d/%d", process_nr, thread_nr);
1032 
1033 	global_data = g->data;
1034 	process_data = td->process_data;
1035 	thread_data = setup_private_data(g->p.bytes_thread);
1036 
1037 	bytes_done = 0;
1038 
1039 	last_task = 0;
1040 	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1041 		last_task = 1;
1042 
1043 	first_task = 0;
1044 	if (process_nr == 0 && thread_nr == 0)
1045 		first_task = 1;
1046 
1047 	if (details >= 2) {
1048 		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1049 			process_nr, thread_nr, global_data, process_data, thread_data);
1050 	}
1051 
1052 	if (g->p.serialize_startup) {
1053 		pthread_mutex_lock(&g->startup_mutex);
1054 		g->nr_tasks_started++;
1055 		pthread_mutex_unlock(&g->startup_mutex);
1056 
1057 		/* Here we will wait for the main process to start us all at once: */
1058 		pthread_mutex_lock(&g->start_work_mutex);
1059 		g->nr_tasks_working++;
1060 
1061 		/* Last one wake the main process: */
1062 		if (g->nr_tasks_working == g->p.nr_tasks)
1063 			pthread_mutex_unlock(&g->startup_done_mutex);
1064 
1065 		pthread_mutex_unlock(&g->start_work_mutex);
1066 	}
1067 
1068 	gettimeofday(&start0, NULL);
1069 
1070 	start = stop = start0;
1071 	last_perturbance = start.tv_sec;
1072 
1073 	for (l = 0; l < g->p.nr_loops; l++) {
1074 		start = stop;
1075 
1076 		if (g->stop_work)
1077 			break;
1078 
1079 		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1080 		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1081 		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1082 
1083 		if (g->p.sleep_usecs) {
1084 			pthread_mutex_lock(td->process_lock);
1085 			usleep(g->p.sleep_usecs);
1086 			pthread_mutex_unlock(td->process_lock);
1087 		}
1088 		/*
1089 		 * Amount of work to be done under a process-global lock:
1090 		 */
1091 		if (g->p.bytes_process_locked) {
1092 			pthread_mutex_lock(td->process_lock);
1093 			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1094 			pthread_mutex_unlock(td->process_lock);
1095 		}
1096 
1097 		work_done = g->p.bytes_global + g->p.bytes_process +
1098 			    g->p.bytes_process_locked + g->p.bytes_thread;
1099 
1100 		update_curr_cpu(task_nr, work_done);
1101 		bytes_done += work_done;
1102 
1103 		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1104 			continue;
1105 
1106 		td->loops_done = l;
1107 
1108 		gettimeofday(&stop, NULL);
1109 
1110 		/* Check whether our max runtime timed out: */
1111 		if (g->p.nr_secs) {
1112 			timersub(&stop, &start0, &diff);
1113 			if (diff.tv_sec >= g->p.nr_secs) {
1114 				g->stop_work = true;
1115 				break;
1116 			}
1117 		}
1118 
1119 		/* Update the summary at most once per second: */
1120 		if (start.tv_sec == stop.tv_sec)
1121 			continue;
1122 
1123 		/*
1124 		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1125 		 * by migrating to CPU#0:
1126 		 */
1127 		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1128 			cpu_set_t orig_mask;
1129 			int target_cpu;
1130 			int this_cpu;
1131 
1132 			last_perturbance = stop.tv_sec;
1133 
1134 			/*
1135 			 * Depending on where we are running, move into
1136 			 * the other half of the system, to create some
1137 			 * real disturbance:
1138 			 */
1139 			this_cpu = g->threads[task_nr].curr_cpu;
1140 			if (this_cpu < g->p.nr_cpus/2)
1141 				target_cpu = g->p.nr_cpus-1;
1142 			else
1143 				target_cpu = 0;
1144 
1145 			orig_mask = bind_to_cpu(target_cpu);
1146 
1147 			/* Here we are running on the target CPU already */
1148 			if (details >= 1)
1149 				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1150 
1151 			bind_to_cpumask(orig_mask);
1152 		}
1153 
1154 		if (details >= 3) {
1155 			timersub(&stop, &start, &diff);
1156 			runtime_ns_max = diff.tv_sec * 1000000000;
1157 			runtime_ns_max += diff.tv_usec * 1000;
1158 
1159 			if (details >= 0) {
1160 				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016lx]\n",
1161 					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1162 			}
1163 			fflush(stdout);
1164 		}
1165 		if (!last_task)
1166 			continue;
1167 
1168 		timersub(&stop, &start0, &diff);
1169 		runtime_ns_max = diff.tv_sec * 1000000000ULL;
1170 		runtime_ns_max += diff.tv_usec * 1000ULL;
1171 
1172 		show_summary(runtime_ns_max, l, &convergence);
1173 	}
1174 
1175 	gettimeofday(&stop, NULL);
1176 	timersub(&stop, &start0, &diff);
1177 	td->runtime_ns = diff.tv_sec * 1000000000ULL;
1178 	td->runtime_ns += diff.tv_usec * 1000ULL;
1179 
1180 	free_data(thread_data, g->p.bytes_thread);
1181 
1182 	pthread_mutex_lock(&g->stop_work_mutex);
1183 	g->bytes_done += bytes_done;
1184 	pthread_mutex_unlock(&g->stop_work_mutex);
1185 
1186 	return NULL;
1187 }
1188 
1189 /*
1190  * A worker process starts a couple of threads:
1191  */
1192 static void worker_process(int process_nr)
1193 {
1194 	pthread_mutex_t process_lock;
1195 	struct thread_data *td;
1196 	pthread_t *pthreads;
1197 	u8 *process_data;
1198 	int task_nr;
1199 	int ret;
1200 	int t;
1201 
1202 	pthread_mutex_init(&process_lock, NULL);
1203 	set_taskname("process %d", process_nr);
1204 
1205 	/*
1206 	 * Pick up the memory policy and the CPU binding of our first thread,
1207 	 * so that we initialize memory accordingly:
1208 	 */
1209 	task_nr = process_nr*g->p.nr_threads;
1210 	td = g->threads + task_nr;
1211 
1212 	bind_to_memnode(td->bind_node);
1213 	bind_to_cpumask(td->bind_cpumask);
1214 
1215 	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1216 	process_data = setup_private_data(g->p.bytes_process);
1217 
1218 	if (g->p.show_details >= 3) {
1219 		printf(" # process %2d global mem: %p, process mem: %p\n",
1220 			process_nr, g->data, process_data);
1221 	}
1222 
1223 	for (t = 0; t < g->p.nr_threads; t++) {
1224 		task_nr = process_nr*g->p.nr_threads + t;
1225 		td = g->threads + task_nr;
1226 
1227 		td->process_data = process_data;
1228 		td->process_nr   = process_nr;
1229 		td->thread_nr    = t;
1230 		td->task_nr	 = task_nr;
1231 		td->val          = rand();
1232 		td->curr_cpu	 = -1;
1233 		td->process_lock = &process_lock;
1234 
1235 		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1236 		BUG_ON(ret);
1237 	}
1238 
1239 	for (t = 0; t < g->p.nr_threads; t++) {
1240                 ret = pthread_join(pthreads[t], NULL);
1241 		BUG_ON(ret);
1242 	}
1243 
1244 	free_data(process_data, g->p.bytes_process);
1245 	free(pthreads);
1246 }
1247 
1248 static void print_summary(void)
1249 {
1250 	if (g->p.show_details < 0)
1251 		return;
1252 
1253 	printf("\n ###\n");
1254 	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1255 		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1256 	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1257 			g->p.nr_loops, g->p.bytes_global/1024/1024);
1258 	printf(" #      %5dx %5ldMB process shared mem operations\n",
1259 			g->p.nr_loops, g->p.bytes_process/1024/1024);
1260 	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1261 			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1262 
1263 	printf(" ###\n");
1264 
1265 	printf("\n ###\n"); fflush(stdout);
1266 }
1267 
1268 static void init_thread_data(void)
1269 {
1270 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1271 	int t;
1272 
1273 	g->threads = zalloc_shared_data(size);
1274 
1275 	for (t = 0; t < g->p.nr_tasks; t++) {
1276 		struct thread_data *td = g->threads + t;
1277 		int cpu;
1278 
1279 		/* Allow all nodes by default: */
1280 		td->bind_node = -1;
1281 
1282 		/* Allow all CPUs by default: */
1283 		CPU_ZERO(&td->bind_cpumask);
1284 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1285 			CPU_SET(cpu, &td->bind_cpumask);
1286 	}
1287 }
1288 
1289 static void deinit_thread_data(void)
1290 {
1291 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1292 
1293 	free_data(g->threads, size);
1294 }
1295 
1296 static int init(void)
1297 {
1298 	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1299 
1300 	/* Copy over options: */
1301 	g->p = p0;
1302 
1303 	g->p.nr_cpus = numa_num_configured_cpus();
1304 
1305 	g->p.nr_nodes = numa_max_node() + 1;
1306 
1307 	/* char array in count_process_nodes(): */
1308 	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1309 
1310 	if (g->p.show_quiet && !g->p.show_details)
1311 		g->p.show_details = -1;
1312 
1313 	/* Some memory should be specified: */
1314 	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1315 		return -1;
1316 
1317 	if (g->p.mb_global_str) {
1318 		g->p.mb_global = atof(g->p.mb_global_str);
1319 		BUG_ON(g->p.mb_global < 0);
1320 	}
1321 
1322 	if (g->p.mb_proc_str) {
1323 		g->p.mb_proc = atof(g->p.mb_proc_str);
1324 		BUG_ON(g->p.mb_proc < 0);
1325 	}
1326 
1327 	if (g->p.mb_proc_locked_str) {
1328 		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1329 		BUG_ON(g->p.mb_proc_locked < 0);
1330 		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1331 	}
1332 
1333 	if (g->p.mb_thread_str) {
1334 		g->p.mb_thread = atof(g->p.mb_thread_str);
1335 		BUG_ON(g->p.mb_thread < 0);
1336 	}
1337 
1338 	BUG_ON(g->p.nr_threads <= 0);
1339 	BUG_ON(g->p.nr_proc <= 0);
1340 
1341 	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1342 
1343 	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1344 	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1345 	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1346 	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1347 
1348 	g->data = setup_shared_data(g->p.bytes_global);
1349 
1350 	/* Startup serialization: */
1351 	init_global_mutex(&g->start_work_mutex);
1352 	init_global_mutex(&g->startup_mutex);
1353 	init_global_mutex(&g->startup_done_mutex);
1354 	init_global_mutex(&g->stop_work_mutex);
1355 
1356 	init_thread_data();
1357 
1358 	tprintf("#\n");
1359 	parse_setup_cpu_list();
1360 	parse_setup_node_list();
1361 	tprintf("#\n");
1362 
1363 	print_summary();
1364 
1365 	return 0;
1366 }
1367 
1368 static void deinit(void)
1369 {
1370 	free_data(g->data, g->p.bytes_global);
1371 	g->data = NULL;
1372 
1373 	deinit_thread_data();
1374 
1375 	free_data(g, sizeof(*g));
1376 	g = NULL;
1377 }
1378 
1379 /*
1380  * Print a short or long result, depending on the verbosity setting:
1381  */
1382 static void print_res(const char *name, double val,
1383 		      const char *txt_unit, const char *txt_short, const char *txt_long)
1384 {
1385 	if (!name)
1386 		name = "main,";
1387 
1388 	if (g->p.show_quiet)
1389 		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1390 	else
1391 		printf(" %14.3f %s\n", val, txt_long);
1392 }
1393 
1394 static int __bench_numa(const char *name)
1395 {
1396 	struct timeval start, stop, diff;
1397 	u64 runtime_ns_min, runtime_ns_sum;
1398 	pid_t *pids, pid, wpid;
1399 	double delta_runtime;
1400 	double runtime_avg;
1401 	double runtime_sec_max;
1402 	double runtime_sec_min;
1403 	int wait_stat;
1404 	double bytes;
1405 	int i, t;
1406 
1407 	if (init())
1408 		return -1;
1409 
1410 	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1411 	pid = -1;
1412 
1413 	/* All threads try to acquire it, this way we can wait for them to start up: */
1414 	pthread_mutex_lock(&g->start_work_mutex);
1415 
1416 	if (g->p.serialize_startup) {
1417 		tprintf(" #\n");
1418 		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1419 	}
1420 
1421 	gettimeofday(&start, NULL);
1422 
1423 	for (i = 0; i < g->p.nr_proc; i++) {
1424 		pid = fork();
1425 		dprintf(" # process %2d: PID %d\n", i, pid);
1426 
1427 		BUG_ON(pid < 0);
1428 		if (!pid) {
1429 			/* Child process: */
1430 			worker_process(i);
1431 
1432 			exit(0);
1433 		}
1434 		pids[i] = pid;
1435 
1436 	}
1437 	/* Wait for all the threads to start up: */
1438 	while (g->nr_tasks_started != g->p.nr_tasks)
1439 		usleep(1000);
1440 
1441 	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1442 
1443 	if (g->p.serialize_startup) {
1444 		double startup_sec;
1445 
1446 		pthread_mutex_lock(&g->startup_done_mutex);
1447 
1448 		/* This will start all threads: */
1449 		pthread_mutex_unlock(&g->start_work_mutex);
1450 
1451 		/* This mutex is locked - the last started thread will wake us: */
1452 		pthread_mutex_lock(&g->startup_done_mutex);
1453 
1454 		gettimeofday(&stop, NULL);
1455 
1456 		timersub(&stop, &start, &diff);
1457 
1458 		startup_sec = diff.tv_sec * 1000000000.0;
1459 		startup_sec += diff.tv_usec * 1000.0;
1460 		startup_sec /= 1e9;
1461 
1462 		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1463 		tprintf(" #\n");
1464 
1465 		start = stop;
1466 		pthread_mutex_unlock(&g->startup_done_mutex);
1467 	} else {
1468 		gettimeofday(&start, NULL);
1469 	}
1470 
1471 	/* Parent process: */
1472 
1473 
1474 	for (i = 0; i < g->p.nr_proc; i++) {
1475 		wpid = waitpid(pids[i], &wait_stat, 0);
1476 		BUG_ON(wpid < 0);
1477 		BUG_ON(!WIFEXITED(wait_stat));
1478 
1479 	}
1480 
1481 	runtime_ns_sum = 0;
1482 	runtime_ns_min = -1LL;
1483 
1484 	for (t = 0; t < g->p.nr_tasks; t++) {
1485 		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1486 
1487 		runtime_ns_sum += thread_runtime_ns;
1488 		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1489 	}
1490 
1491 	gettimeofday(&stop, NULL);
1492 	timersub(&stop, &start, &diff);
1493 
1494 	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1495 
1496 	tprintf("\n ###\n");
1497 	tprintf("\n");
1498 
1499 	runtime_sec_max = diff.tv_sec * 1000000000.0;
1500 	runtime_sec_max += diff.tv_usec * 1000.0;
1501 	runtime_sec_max /= 1e9;
1502 
1503 	runtime_sec_min = runtime_ns_min/1e9;
1504 
1505 	bytes = g->bytes_done;
1506 	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1507 
1508 	if (g->p.measure_convergence) {
1509 		print_res(name, runtime_sec_max,
1510 			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1511 	}
1512 
1513 	print_res(name, runtime_sec_max,
1514 		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1515 
1516 	print_res(name, runtime_sec_min,
1517 		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1518 
1519 	print_res(name, runtime_avg,
1520 		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1521 
1522 	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1523 	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1524 		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1525 
1526 	print_res(name, bytes / g->p.nr_tasks / 1e9,
1527 		"GB,", "data/thread",		"GB data processed, per thread");
1528 
1529 	print_res(name, bytes / 1e9,
1530 		"GB,", "data-total",		"GB data processed, total");
1531 
1532 	print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1533 		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1534 
1535 	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1536 		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1537 
1538 	print_res(name, bytes / runtime_sec_max / 1e9,
1539 		"GB/sec,", "total-speed",	"GB/sec total speed");
1540 
1541 	free(pids);
1542 
1543 	deinit();
1544 
1545 	return 0;
1546 }
1547 
1548 #define MAX_ARGS 50
1549 
1550 static int command_size(const char **argv)
1551 {
1552 	int size = 0;
1553 
1554 	while (*argv) {
1555 		size++;
1556 		argv++;
1557 	}
1558 
1559 	BUG_ON(size >= MAX_ARGS);
1560 
1561 	return size;
1562 }
1563 
1564 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1565 {
1566 	int i;
1567 
1568 	printf("\n # Running %s \"perf bench numa", name);
1569 
1570 	for (i = 0; i < argc; i++)
1571 		printf(" %s", argv[i]);
1572 
1573 	printf("\"\n");
1574 
1575 	memset(p, 0, sizeof(*p));
1576 
1577 	/* Initialize nonzero defaults: */
1578 
1579 	p->serialize_startup		= 1;
1580 	p->data_reads			= true;
1581 	p->data_writes			= true;
1582 	p->data_backwards		= true;
1583 	p->data_rand_walk		= true;
1584 	p->nr_loops			= -1;
1585 	p->init_random			= true;
1586 }
1587 
1588 static int run_bench_numa(const char *name, const char **argv)
1589 {
1590 	int argc = command_size(argv);
1591 
1592 	init_params(&p0, name, argc, argv);
1593 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1594 	if (argc)
1595 		goto err;
1596 
1597 	if (__bench_numa(name))
1598 		goto err;
1599 
1600 	return 0;
1601 
1602 err:
1603 	usage_with_options(numa_usage, options);
1604 	return -1;
1605 }
1606 
1607 #define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1608 #define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1609 
1610 #define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1611 #define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1612 
1613 #define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1614 #define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1615 
1616 /*
1617  * The built-in test-suite executed by "perf bench numa -a".
1618  *
1619  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1620  */
1621 static const char *tests[][MAX_ARGS] = {
1622    /* Basic single-stream NUMA bandwidth measurements: */
1623    { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1624 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1625    { "RAM-bw-local-NOTHP,",
1626 			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1627 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1628    { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1629 			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1630 
1631    /* 2-stream NUMA bandwidth measurements: */
1632    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1633 			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1634    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1635 		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1636 
1637    /* Cross-stream NUMA bandwidth measurement: */
1638    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1639 		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1640 
1641    /* Convergence latency measurements: */
1642    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1643    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1644    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1645    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1646    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1647    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1648    { " 4x4-convergence-NOTHP,",
1649 			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1650    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1651    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1652    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1653    { " 8x4-convergence-NOTHP,",
1654 			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1655    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1656    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1657    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1658    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1659    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1660 
1661    /* Various NUMA process/thread layout bandwidth measurements: */
1662    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1663    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1664    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1665    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1666    { " 8x1-bw-process-NOTHP,",
1667 			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1668    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1669 
1670    { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1671    { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1672    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1673    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1674 
1675    { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1676    { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1677    { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1678    { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1679    { " 4x8-bw-thread-NOTHP,",
1680 			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1681    { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1682    { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1683 
1684    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1685    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1686 
1687    { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1688    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1689    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1690    { "numa01-bw-thread-NOTHP,",
1691 			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1692 };
1693 
1694 static int bench_all(void)
1695 {
1696 	int nr = ARRAY_SIZE(tests);
1697 	int ret;
1698 	int i;
1699 
1700 	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1701 	BUG_ON(ret < 0);
1702 
1703 	for (i = 0; i < nr; i++) {
1704 		if (run_bench_numa(tests[i][0], tests[i] + 1))
1705 			return -1;
1706 	}
1707 
1708 	printf("\n");
1709 
1710 	return 0;
1711 }
1712 
1713 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1714 {
1715 	init_params(&p0, "main,", argc, argv);
1716 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1717 	if (argc)
1718 		goto err;
1719 
1720 	if (p0.run_all)
1721 		return bench_all();
1722 
1723 	if (__bench_numa(NULL))
1724 		goto err;
1725 
1726 	return 0;
1727 
1728 err:
1729 	usage_with_options(numa_usage, options);
1730 	return -1;
1731 }
1732