xref: /linux/tools/perf/bench/numa.c (revision d198b34f3855eee2571dda03eea75a09c7c31480)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * numa.c
4  *
5  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
6  */
7 
8 #include <inttypes.h>
9 /* For the CLR_() macros */
10 #include <pthread.h>
11 
12 #include <subcmd/parse-options.h>
13 #include "../util/cloexec.h"
14 
15 #include "bench.h"
16 
17 #include <errno.h>
18 #include <sched.h>
19 #include <stdio.h>
20 #include <assert.h>
21 #include <malloc.h>
22 #include <signal.h>
23 #include <stdlib.h>
24 #include <string.h>
25 #include <unistd.h>
26 #include <sys/mman.h>
27 #include <sys/time.h>
28 #include <sys/resource.h>
29 #include <sys/wait.h>
30 #include <sys/prctl.h>
31 #include <sys/types.h>
32 #include <linux/kernel.h>
33 #include <linux/time64.h>
34 #include <linux/numa.h>
35 #include <linux/zalloc.h>
36 
37 #include <numa.h>
38 #include <numaif.h>
39 
40 #ifndef RUSAGE_THREAD
41 # define RUSAGE_THREAD 1
42 #endif
43 
44 /*
45  * Regular printout to the terminal, supressed if -q is specified:
46  */
47 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
48 
49 /*
50  * Debug printf:
51  */
52 #undef dprintf
53 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
54 
55 struct thread_data {
56 	int			curr_cpu;
57 	cpu_set_t		bind_cpumask;
58 	int			bind_node;
59 	u8			*process_data;
60 	int			process_nr;
61 	int			thread_nr;
62 	int			task_nr;
63 	unsigned int		loops_done;
64 	u64			val;
65 	u64			runtime_ns;
66 	u64			system_time_ns;
67 	u64			user_time_ns;
68 	double			speed_gbs;
69 	pthread_mutex_t		*process_lock;
70 };
71 
72 /* Parameters set by options: */
73 
74 struct params {
75 	/* Startup synchronization: */
76 	bool			serialize_startup;
77 
78 	/* Task hierarchy: */
79 	int			nr_proc;
80 	int			nr_threads;
81 
82 	/* Working set sizes: */
83 	const char		*mb_global_str;
84 	const char		*mb_proc_str;
85 	const char		*mb_proc_locked_str;
86 	const char		*mb_thread_str;
87 
88 	double			mb_global;
89 	double			mb_proc;
90 	double			mb_proc_locked;
91 	double			mb_thread;
92 
93 	/* Access patterns to the working set: */
94 	bool			data_reads;
95 	bool			data_writes;
96 	bool			data_backwards;
97 	bool			data_zero_memset;
98 	bool			data_rand_walk;
99 	u32			nr_loops;
100 	u32			nr_secs;
101 	u32			sleep_usecs;
102 
103 	/* Working set initialization: */
104 	bool			init_zero;
105 	bool			init_random;
106 	bool			init_cpu0;
107 
108 	/* Misc options: */
109 	int			show_details;
110 	int			run_all;
111 	int			thp;
112 
113 	long			bytes_global;
114 	long			bytes_process;
115 	long			bytes_process_locked;
116 	long			bytes_thread;
117 
118 	int			nr_tasks;
119 	bool			show_quiet;
120 
121 	bool			show_convergence;
122 	bool			measure_convergence;
123 
124 	int			perturb_secs;
125 	int			nr_cpus;
126 	int			nr_nodes;
127 
128 	/* Affinity options -C and -N: */
129 	char			*cpu_list_str;
130 	char			*node_list_str;
131 };
132 
133 
134 /* Global, read-writable area, accessible to all processes and threads: */
135 
136 struct global_info {
137 	u8			*data;
138 
139 	pthread_mutex_t		startup_mutex;
140 	int			nr_tasks_started;
141 
142 	pthread_mutex_t		startup_done_mutex;
143 
144 	pthread_mutex_t		start_work_mutex;
145 	int			nr_tasks_working;
146 
147 	pthread_mutex_t		stop_work_mutex;
148 	u64			bytes_done;
149 
150 	struct thread_data	*threads;
151 
152 	/* Convergence latency measurement: */
153 	bool			all_converged;
154 	bool			stop_work;
155 
156 	int			print_once;
157 
158 	struct params		p;
159 };
160 
161 static struct global_info	*g = NULL;
162 
163 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
164 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
165 
166 struct params p0;
167 
168 static const struct option options[] = {
169 	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
170 	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
171 
172 	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
173 	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
174 	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
175 	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
176 
177 	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
178 	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
179 	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
180 
181 	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via reads (can be mixed with -W)"),
182 	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
183 	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
184 	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
185 	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
186 
187 
188 	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
189 	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
190 	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
191 	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
192 
193 	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
194 	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
195 	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
196 	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
197 		    "convergence is reached when each process (all its threads) is running on a single NUMA node."),
198 	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
199 	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"quiet mode"),
200 	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
201 
202 	/* Special option string parsing callbacks: */
203         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
204 			"bind the first N tasks to these specific cpus (the rest is unbound)",
205 			parse_cpus_opt),
206         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
207 			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
208 			parse_nodes_opt),
209 	OPT_END()
210 };
211 
212 static const char * const bench_numa_usage[] = {
213 	"perf bench numa <options>",
214 	NULL
215 };
216 
217 static const char * const numa_usage[] = {
218 	"perf bench numa mem [<options>]",
219 	NULL
220 };
221 
222 /*
223  * To get number of numa nodes present.
224  */
225 static int nr_numa_nodes(void)
226 {
227 	int i, nr_nodes = 0;
228 
229 	for (i = 0; i < g->p.nr_nodes; i++) {
230 		if (numa_bitmask_isbitset(numa_nodes_ptr, i))
231 			nr_nodes++;
232 	}
233 
234 	return nr_nodes;
235 }
236 
237 /*
238  * To check if given numa node is present.
239  */
240 static int is_node_present(int node)
241 {
242 	return numa_bitmask_isbitset(numa_nodes_ptr, node);
243 }
244 
245 /*
246  * To check given numa node has cpus.
247  */
248 static bool node_has_cpus(int node)
249 {
250 	struct bitmask *cpu = numa_allocate_cpumask();
251 	unsigned int i;
252 
253 	if (cpu && !numa_node_to_cpus(node, cpu)) {
254 		for (i = 0; i < cpu->size; i++) {
255 			if (numa_bitmask_isbitset(cpu, i))
256 				return true;
257 		}
258 	}
259 
260 	return false; /* lets fall back to nocpus safely */
261 }
262 
263 static cpu_set_t bind_to_cpu(int target_cpu)
264 {
265 	cpu_set_t orig_mask, mask;
266 	int ret;
267 
268 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
269 	BUG_ON(ret);
270 
271 	CPU_ZERO(&mask);
272 
273 	if (target_cpu == -1) {
274 		int cpu;
275 
276 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
277 			CPU_SET(cpu, &mask);
278 	} else {
279 		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
280 		CPU_SET(target_cpu, &mask);
281 	}
282 
283 	ret = sched_setaffinity(0, sizeof(mask), &mask);
284 	BUG_ON(ret);
285 
286 	return orig_mask;
287 }
288 
289 static cpu_set_t bind_to_node(int target_node)
290 {
291 	int cpus_per_node = g->p.nr_cpus / nr_numa_nodes();
292 	cpu_set_t orig_mask, mask;
293 	int cpu;
294 	int ret;
295 
296 	BUG_ON(cpus_per_node * nr_numa_nodes() != g->p.nr_cpus);
297 	BUG_ON(!cpus_per_node);
298 
299 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
300 	BUG_ON(ret);
301 
302 	CPU_ZERO(&mask);
303 
304 	if (target_node == NUMA_NO_NODE) {
305 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
306 			CPU_SET(cpu, &mask);
307 	} else {
308 		int cpu_start = (target_node + 0) * cpus_per_node;
309 		int cpu_stop  = (target_node + 1) * cpus_per_node;
310 
311 		BUG_ON(cpu_stop > g->p.nr_cpus);
312 
313 		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
314 			CPU_SET(cpu, &mask);
315 	}
316 
317 	ret = sched_setaffinity(0, sizeof(mask), &mask);
318 	BUG_ON(ret);
319 
320 	return orig_mask;
321 }
322 
323 static void bind_to_cpumask(cpu_set_t mask)
324 {
325 	int ret;
326 
327 	ret = sched_setaffinity(0, sizeof(mask), &mask);
328 	BUG_ON(ret);
329 }
330 
331 static void mempol_restore(void)
332 {
333 	int ret;
334 
335 	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
336 
337 	BUG_ON(ret);
338 }
339 
340 static void bind_to_memnode(int node)
341 {
342 	unsigned long nodemask;
343 	int ret;
344 
345 	if (node == NUMA_NO_NODE)
346 		return;
347 
348 	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
349 	nodemask = 1L << node;
350 
351 	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
352 	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
353 
354 	BUG_ON(ret);
355 }
356 
357 #define HPSIZE (2*1024*1024)
358 
359 #define set_taskname(fmt...)				\
360 do {							\
361 	char name[20];					\
362 							\
363 	snprintf(name, 20, fmt);			\
364 	prctl(PR_SET_NAME, name);			\
365 } while (0)
366 
367 static u8 *alloc_data(ssize_t bytes0, int map_flags,
368 		      int init_zero, int init_cpu0, int thp, int init_random)
369 {
370 	cpu_set_t orig_mask;
371 	ssize_t bytes;
372 	u8 *buf;
373 	int ret;
374 
375 	if (!bytes0)
376 		return NULL;
377 
378 	/* Allocate and initialize all memory on CPU#0: */
379 	if (init_cpu0) {
380 		int node = numa_node_of_cpu(0);
381 
382 		orig_mask = bind_to_node(node);
383 		bind_to_memnode(node);
384 	}
385 
386 	bytes = bytes0 + HPSIZE;
387 
388 	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
389 	BUG_ON(buf == (void *)-1);
390 
391 	if (map_flags == MAP_PRIVATE) {
392 		if (thp > 0) {
393 			ret = madvise(buf, bytes, MADV_HUGEPAGE);
394 			if (ret && !g->print_once) {
395 				g->print_once = 1;
396 				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
397 			}
398 		}
399 		if (thp < 0) {
400 			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
401 			if (ret && !g->print_once) {
402 				g->print_once = 1;
403 				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
404 			}
405 		}
406 	}
407 
408 	if (init_zero) {
409 		bzero(buf, bytes);
410 	} else {
411 		/* Initialize random contents, different in each word: */
412 		if (init_random) {
413 			u64 *wbuf = (void *)buf;
414 			long off = rand();
415 			long i;
416 
417 			for (i = 0; i < bytes/8; i++)
418 				wbuf[i] = i + off;
419 		}
420 	}
421 
422 	/* Align to 2MB boundary: */
423 	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
424 
425 	/* Restore affinity: */
426 	if (init_cpu0) {
427 		bind_to_cpumask(orig_mask);
428 		mempol_restore();
429 	}
430 
431 	return buf;
432 }
433 
434 static void free_data(void *data, ssize_t bytes)
435 {
436 	int ret;
437 
438 	if (!data)
439 		return;
440 
441 	ret = munmap(data, bytes);
442 	BUG_ON(ret);
443 }
444 
445 /*
446  * Create a shared memory buffer that can be shared between processes, zeroed:
447  */
448 static void * zalloc_shared_data(ssize_t bytes)
449 {
450 	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
451 }
452 
453 /*
454  * Create a shared memory buffer that can be shared between processes:
455  */
456 static void * setup_shared_data(ssize_t bytes)
457 {
458 	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
459 }
460 
461 /*
462  * Allocate process-local memory - this will either be shared between
463  * threads of this process, or only be accessed by this thread:
464  */
465 static void * setup_private_data(ssize_t bytes)
466 {
467 	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
468 }
469 
470 /*
471  * Return a process-shared (global) mutex:
472  */
473 static void init_global_mutex(pthread_mutex_t *mutex)
474 {
475 	pthread_mutexattr_t attr;
476 
477 	pthread_mutexattr_init(&attr);
478 	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
479 	pthread_mutex_init(mutex, &attr);
480 }
481 
482 static int parse_cpu_list(const char *arg)
483 {
484 	p0.cpu_list_str = strdup(arg);
485 
486 	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
487 
488 	return 0;
489 }
490 
491 static int parse_setup_cpu_list(void)
492 {
493 	struct thread_data *td;
494 	char *str0, *str;
495 	int t;
496 
497 	if (!g->p.cpu_list_str)
498 		return 0;
499 
500 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
501 
502 	str0 = str = strdup(g->p.cpu_list_str);
503 	t = 0;
504 
505 	BUG_ON(!str);
506 
507 	tprintf("# binding tasks to CPUs:\n");
508 	tprintf("#  ");
509 
510 	while (true) {
511 		int bind_cpu, bind_cpu_0, bind_cpu_1;
512 		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
513 		int bind_len;
514 		int step;
515 		int mul;
516 
517 		tok = strsep(&str, ",");
518 		if (!tok)
519 			break;
520 
521 		tok_end = strstr(tok, "-");
522 
523 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
524 		if (!tok_end) {
525 			/* Single CPU specified: */
526 			bind_cpu_0 = bind_cpu_1 = atol(tok);
527 		} else {
528 			/* CPU range specified (for example: "5-11"): */
529 			bind_cpu_0 = atol(tok);
530 			bind_cpu_1 = atol(tok_end + 1);
531 		}
532 
533 		step = 1;
534 		tok_step = strstr(tok, "#");
535 		if (tok_step) {
536 			step = atol(tok_step + 1);
537 			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
538 		}
539 
540 		/*
541 		 * Mask length.
542 		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
543 		 * where the _4 means the next 4 CPUs are allowed.
544 		 */
545 		bind_len = 1;
546 		tok_len = strstr(tok, "_");
547 		if (tok_len) {
548 			bind_len = atol(tok_len + 1);
549 			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
550 		}
551 
552 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
553 		mul = 1;
554 		tok_mul = strstr(tok, "x");
555 		if (tok_mul) {
556 			mul = atol(tok_mul + 1);
557 			BUG_ON(mul <= 0);
558 		}
559 
560 		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
561 
562 		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
563 			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
564 			return -1;
565 		}
566 
567 		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
568 		BUG_ON(bind_cpu_0 > bind_cpu_1);
569 
570 		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
571 			int i;
572 
573 			for (i = 0; i < mul; i++) {
574 				int cpu;
575 
576 				if (t >= g->p.nr_tasks) {
577 					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
578 					goto out;
579 				}
580 				td = g->threads + t;
581 
582 				if (t)
583 					tprintf(",");
584 				if (bind_len > 1) {
585 					tprintf("%2d/%d", bind_cpu, bind_len);
586 				} else {
587 					tprintf("%2d", bind_cpu);
588 				}
589 
590 				CPU_ZERO(&td->bind_cpumask);
591 				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
592 					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
593 					CPU_SET(cpu, &td->bind_cpumask);
594 				}
595 				t++;
596 			}
597 		}
598 	}
599 out:
600 
601 	tprintf("\n");
602 
603 	if (t < g->p.nr_tasks)
604 		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
605 
606 	free(str0);
607 	return 0;
608 }
609 
610 static int parse_cpus_opt(const struct option *opt __maybe_unused,
611 			  const char *arg, int unset __maybe_unused)
612 {
613 	if (!arg)
614 		return -1;
615 
616 	return parse_cpu_list(arg);
617 }
618 
619 static int parse_node_list(const char *arg)
620 {
621 	p0.node_list_str = strdup(arg);
622 
623 	dprintf("got NODE list: {%s}\n", p0.node_list_str);
624 
625 	return 0;
626 }
627 
628 static int parse_setup_node_list(void)
629 {
630 	struct thread_data *td;
631 	char *str0, *str;
632 	int t;
633 
634 	if (!g->p.node_list_str)
635 		return 0;
636 
637 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
638 
639 	str0 = str = strdup(g->p.node_list_str);
640 	t = 0;
641 
642 	BUG_ON(!str);
643 
644 	tprintf("# binding tasks to NODEs:\n");
645 	tprintf("# ");
646 
647 	while (true) {
648 		int bind_node, bind_node_0, bind_node_1;
649 		char *tok, *tok_end, *tok_step, *tok_mul;
650 		int step;
651 		int mul;
652 
653 		tok = strsep(&str, ",");
654 		if (!tok)
655 			break;
656 
657 		tok_end = strstr(tok, "-");
658 
659 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
660 		if (!tok_end) {
661 			/* Single NODE specified: */
662 			bind_node_0 = bind_node_1 = atol(tok);
663 		} else {
664 			/* NODE range specified (for example: "5-11"): */
665 			bind_node_0 = atol(tok);
666 			bind_node_1 = atol(tok_end + 1);
667 		}
668 
669 		step = 1;
670 		tok_step = strstr(tok, "#");
671 		if (tok_step) {
672 			step = atol(tok_step + 1);
673 			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
674 		}
675 
676 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
677 		mul = 1;
678 		tok_mul = strstr(tok, "x");
679 		if (tok_mul) {
680 			mul = atol(tok_mul + 1);
681 			BUG_ON(mul <= 0);
682 		}
683 
684 		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
685 
686 		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
687 			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
688 			return -1;
689 		}
690 
691 		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
692 		BUG_ON(bind_node_0 > bind_node_1);
693 
694 		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
695 			int i;
696 
697 			for (i = 0; i < mul; i++) {
698 				if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
699 					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
700 					goto out;
701 				}
702 				td = g->threads + t;
703 
704 				if (!t)
705 					tprintf(" %2d", bind_node);
706 				else
707 					tprintf(",%2d", bind_node);
708 
709 				td->bind_node = bind_node;
710 				t++;
711 			}
712 		}
713 	}
714 out:
715 
716 	tprintf("\n");
717 
718 	if (t < g->p.nr_tasks)
719 		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
720 
721 	free(str0);
722 	return 0;
723 }
724 
725 static int parse_nodes_opt(const struct option *opt __maybe_unused,
726 			  const char *arg, int unset __maybe_unused)
727 {
728 	if (!arg)
729 		return -1;
730 
731 	return parse_node_list(arg);
732 
733 	return 0;
734 }
735 
736 #define BIT(x) (1ul << x)
737 
738 static inline uint32_t lfsr_32(uint32_t lfsr)
739 {
740 	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
741 	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
742 }
743 
744 /*
745  * Make sure there's real data dependency to RAM (when read
746  * accesses are enabled), so the compiler, the CPU and the
747  * kernel (KSM, zero page, etc.) cannot optimize away RAM
748  * accesses:
749  */
750 static inline u64 access_data(u64 *data, u64 val)
751 {
752 	if (g->p.data_reads)
753 		val += *data;
754 	if (g->p.data_writes)
755 		*data = val + 1;
756 	return val;
757 }
758 
759 /*
760  * The worker process does two types of work, a forwards going
761  * loop and a backwards going loop.
762  *
763  * We do this so that on multiprocessor systems we do not create
764  * a 'train' of processing, with highly synchronized processes,
765  * skewing the whole benchmark.
766  */
767 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
768 {
769 	long words = bytes/sizeof(u64);
770 	u64 *data = (void *)__data;
771 	long chunk_0, chunk_1;
772 	u64 *d0, *d, *d1;
773 	long off;
774 	long i;
775 
776 	BUG_ON(!data && words);
777 	BUG_ON(data && !words);
778 
779 	if (!data)
780 		return val;
781 
782 	/* Very simple memset() work variant: */
783 	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
784 		bzero(data, bytes);
785 		return val;
786 	}
787 
788 	/* Spread out by PID/TID nr and by loop nr: */
789 	chunk_0 = words/nr_max;
790 	chunk_1 = words/g->p.nr_loops;
791 	off = nr*chunk_0 + loop*chunk_1;
792 
793 	while (off >= words)
794 		off -= words;
795 
796 	if (g->p.data_rand_walk) {
797 		u32 lfsr = nr + loop + val;
798 		int j;
799 
800 		for (i = 0; i < words/1024; i++) {
801 			long start, end;
802 
803 			lfsr = lfsr_32(lfsr);
804 
805 			start = lfsr % words;
806 			end = min(start + 1024, words-1);
807 
808 			if (g->p.data_zero_memset) {
809 				bzero(data + start, (end-start) * sizeof(u64));
810 			} else {
811 				for (j = start; j < end; j++)
812 					val = access_data(data + j, val);
813 			}
814 		}
815 	} else if (!g->p.data_backwards || (nr + loop) & 1) {
816 
817 		d0 = data + off;
818 		d  = data + off + 1;
819 		d1 = data + words;
820 
821 		/* Process data forwards: */
822 		for (;;) {
823 			if (unlikely(d >= d1))
824 				d = data;
825 			if (unlikely(d == d0))
826 				break;
827 
828 			val = access_data(d, val);
829 
830 			d++;
831 		}
832 	} else {
833 		/* Process data backwards: */
834 
835 		d0 = data + off;
836 		d  = data + off - 1;
837 		d1 = data + words;
838 
839 		/* Process data forwards: */
840 		for (;;) {
841 			if (unlikely(d < data))
842 				d = data + words-1;
843 			if (unlikely(d == d0))
844 				break;
845 
846 			val = access_data(d, val);
847 
848 			d--;
849 		}
850 	}
851 
852 	return val;
853 }
854 
855 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
856 {
857 	unsigned int cpu;
858 
859 	cpu = sched_getcpu();
860 
861 	g->threads[task_nr].curr_cpu = cpu;
862 	prctl(0, bytes_worked);
863 }
864 
865 #define MAX_NR_NODES	64
866 
867 /*
868  * Count the number of nodes a process's threads
869  * are spread out on.
870  *
871  * A count of 1 means that the process is compressed
872  * to a single node. A count of g->p.nr_nodes means it's
873  * spread out on the whole system.
874  */
875 static int count_process_nodes(int process_nr)
876 {
877 	char node_present[MAX_NR_NODES] = { 0, };
878 	int nodes;
879 	int n, t;
880 
881 	for (t = 0; t < g->p.nr_threads; t++) {
882 		struct thread_data *td;
883 		int task_nr;
884 		int node;
885 
886 		task_nr = process_nr*g->p.nr_threads + t;
887 		td = g->threads + task_nr;
888 
889 		node = numa_node_of_cpu(td->curr_cpu);
890 		if (node < 0) /* curr_cpu was likely still -1 */
891 			return 0;
892 
893 		node_present[node] = 1;
894 	}
895 
896 	nodes = 0;
897 
898 	for (n = 0; n < MAX_NR_NODES; n++)
899 		nodes += node_present[n];
900 
901 	return nodes;
902 }
903 
904 /*
905  * Count the number of distinct process-threads a node contains.
906  *
907  * A count of 1 means that the node contains only a single
908  * process. If all nodes on the system contain at most one
909  * process then we are well-converged.
910  */
911 static int count_node_processes(int node)
912 {
913 	int processes = 0;
914 	int t, p;
915 
916 	for (p = 0; p < g->p.nr_proc; p++) {
917 		for (t = 0; t < g->p.nr_threads; t++) {
918 			struct thread_data *td;
919 			int task_nr;
920 			int n;
921 
922 			task_nr = p*g->p.nr_threads + t;
923 			td = g->threads + task_nr;
924 
925 			n = numa_node_of_cpu(td->curr_cpu);
926 			if (n == node) {
927 				processes++;
928 				break;
929 			}
930 		}
931 	}
932 
933 	return processes;
934 }
935 
936 static void calc_convergence_compression(int *strong)
937 {
938 	unsigned int nodes_min, nodes_max;
939 	int p;
940 
941 	nodes_min = -1;
942 	nodes_max =  0;
943 
944 	for (p = 0; p < g->p.nr_proc; p++) {
945 		unsigned int nodes = count_process_nodes(p);
946 
947 		if (!nodes) {
948 			*strong = 0;
949 			return;
950 		}
951 
952 		nodes_min = min(nodes, nodes_min);
953 		nodes_max = max(nodes, nodes_max);
954 	}
955 
956 	/* Strong convergence: all threads compress on a single node: */
957 	if (nodes_min == 1 && nodes_max == 1) {
958 		*strong = 1;
959 	} else {
960 		*strong = 0;
961 		tprintf(" {%d-%d}", nodes_min, nodes_max);
962 	}
963 }
964 
965 static void calc_convergence(double runtime_ns_max, double *convergence)
966 {
967 	unsigned int loops_done_min, loops_done_max;
968 	int process_groups;
969 	int nodes[MAX_NR_NODES];
970 	int distance;
971 	int nr_min;
972 	int nr_max;
973 	int strong;
974 	int sum;
975 	int nr;
976 	int node;
977 	int cpu;
978 	int t;
979 
980 	if (!g->p.show_convergence && !g->p.measure_convergence)
981 		return;
982 
983 	for (node = 0; node < g->p.nr_nodes; node++)
984 		nodes[node] = 0;
985 
986 	loops_done_min = -1;
987 	loops_done_max = 0;
988 
989 	for (t = 0; t < g->p.nr_tasks; t++) {
990 		struct thread_data *td = g->threads + t;
991 		unsigned int loops_done;
992 
993 		cpu = td->curr_cpu;
994 
995 		/* Not all threads have written it yet: */
996 		if (cpu < 0)
997 			continue;
998 
999 		node = numa_node_of_cpu(cpu);
1000 
1001 		nodes[node]++;
1002 
1003 		loops_done = td->loops_done;
1004 		loops_done_min = min(loops_done, loops_done_min);
1005 		loops_done_max = max(loops_done, loops_done_max);
1006 	}
1007 
1008 	nr_max = 0;
1009 	nr_min = g->p.nr_tasks;
1010 	sum = 0;
1011 
1012 	for (node = 0; node < g->p.nr_nodes; node++) {
1013 		if (!is_node_present(node))
1014 			continue;
1015 		nr = nodes[node];
1016 		nr_min = min(nr, nr_min);
1017 		nr_max = max(nr, nr_max);
1018 		sum += nr;
1019 	}
1020 	BUG_ON(nr_min > nr_max);
1021 
1022 	BUG_ON(sum > g->p.nr_tasks);
1023 
1024 	if (0 && (sum < g->p.nr_tasks))
1025 		return;
1026 
1027 	/*
1028 	 * Count the number of distinct process groups present
1029 	 * on nodes - when we are converged this will decrease
1030 	 * to g->p.nr_proc:
1031 	 */
1032 	process_groups = 0;
1033 
1034 	for (node = 0; node < g->p.nr_nodes; node++) {
1035 		int processes;
1036 
1037 		if (!is_node_present(node))
1038 			continue;
1039 		processes = count_node_processes(node);
1040 		nr = nodes[node];
1041 		tprintf(" %2d/%-2d", nr, processes);
1042 
1043 		process_groups += processes;
1044 	}
1045 
1046 	distance = nr_max - nr_min;
1047 
1048 	tprintf(" [%2d/%-2d]", distance, process_groups);
1049 
1050 	tprintf(" l:%3d-%-3d (%3d)",
1051 		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1052 
1053 	if (loops_done_min && loops_done_max) {
1054 		double skew = 1.0 - (double)loops_done_min/loops_done_max;
1055 
1056 		tprintf(" [%4.1f%%]", skew * 100.0);
1057 	}
1058 
1059 	calc_convergence_compression(&strong);
1060 
1061 	if (strong && process_groups == g->p.nr_proc) {
1062 		if (!*convergence) {
1063 			*convergence = runtime_ns_max;
1064 			tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1065 			if (g->p.measure_convergence) {
1066 				g->all_converged = true;
1067 				g->stop_work = true;
1068 			}
1069 		}
1070 	} else {
1071 		if (*convergence) {
1072 			tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1073 			*convergence = 0;
1074 		}
1075 		tprintf("\n");
1076 	}
1077 }
1078 
1079 static void show_summary(double runtime_ns_max, int l, double *convergence)
1080 {
1081 	tprintf("\r #  %5.1f%%  [%.1f mins]",
1082 		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1083 
1084 	calc_convergence(runtime_ns_max, convergence);
1085 
1086 	if (g->p.show_details >= 0)
1087 		fflush(stdout);
1088 }
1089 
1090 static void *worker_thread(void *__tdata)
1091 {
1092 	struct thread_data *td = __tdata;
1093 	struct timeval start0, start, stop, diff;
1094 	int process_nr = td->process_nr;
1095 	int thread_nr = td->thread_nr;
1096 	unsigned long last_perturbance;
1097 	int task_nr = td->task_nr;
1098 	int details = g->p.show_details;
1099 	int first_task, last_task;
1100 	double convergence = 0;
1101 	u64 val = td->val;
1102 	double runtime_ns_max;
1103 	u8 *global_data;
1104 	u8 *process_data;
1105 	u8 *thread_data;
1106 	u64 bytes_done, secs;
1107 	long work_done;
1108 	u32 l;
1109 	struct rusage rusage;
1110 
1111 	bind_to_cpumask(td->bind_cpumask);
1112 	bind_to_memnode(td->bind_node);
1113 
1114 	set_taskname("thread %d/%d", process_nr, thread_nr);
1115 
1116 	global_data = g->data;
1117 	process_data = td->process_data;
1118 	thread_data = setup_private_data(g->p.bytes_thread);
1119 
1120 	bytes_done = 0;
1121 
1122 	last_task = 0;
1123 	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1124 		last_task = 1;
1125 
1126 	first_task = 0;
1127 	if (process_nr == 0 && thread_nr == 0)
1128 		first_task = 1;
1129 
1130 	if (details >= 2) {
1131 		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1132 			process_nr, thread_nr, global_data, process_data, thread_data);
1133 	}
1134 
1135 	if (g->p.serialize_startup) {
1136 		pthread_mutex_lock(&g->startup_mutex);
1137 		g->nr_tasks_started++;
1138 		pthread_mutex_unlock(&g->startup_mutex);
1139 
1140 		/* Here we will wait for the main process to start us all at once: */
1141 		pthread_mutex_lock(&g->start_work_mutex);
1142 		g->nr_tasks_working++;
1143 
1144 		/* Last one wake the main process: */
1145 		if (g->nr_tasks_working == g->p.nr_tasks)
1146 			pthread_mutex_unlock(&g->startup_done_mutex);
1147 
1148 		pthread_mutex_unlock(&g->start_work_mutex);
1149 	}
1150 
1151 	gettimeofday(&start0, NULL);
1152 
1153 	start = stop = start0;
1154 	last_perturbance = start.tv_sec;
1155 
1156 	for (l = 0; l < g->p.nr_loops; l++) {
1157 		start = stop;
1158 
1159 		if (g->stop_work)
1160 			break;
1161 
1162 		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1163 		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1164 		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1165 
1166 		if (g->p.sleep_usecs) {
1167 			pthread_mutex_lock(td->process_lock);
1168 			usleep(g->p.sleep_usecs);
1169 			pthread_mutex_unlock(td->process_lock);
1170 		}
1171 		/*
1172 		 * Amount of work to be done under a process-global lock:
1173 		 */
1174 		if (g->p.bytes_process_locked) {
1175 			pthread_mutex_lock(td->process_lock);
1176 			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1177 			pthread_mutex_unlock(td->process_lock);
1178 		}
1179 
1180 		work_done = g->p.bytes_global + g->p.bytes_process +
1181 			    g->p.bytes_process_locked + g->p.bytes_thread;
1182 
1183 		update_curr_cpu(task_nr, work_done);
1184 		bytes_done += work_done;
1185 
1186 		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1187 			continue;
1188 
1189 		td->loops_done = l;
1190 
1191 		gettimeofday(&stop, NULL);
1192 
1193 		/* Check whether our max runtime timed out: */
1194 		if (g->p.nr_secs) {
1195 			timersub(&stop, &start0, &diff);
1196 			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1197 				g->stop_work = true;
1198 				break;
1199 			}
1200 		}
1201 
1202 		/* Update the summary at most once per second: */
1203 		if (start.tv_sec == stop.tv_sec)
1204 			continue;
1205 
1206 		/*
1207 		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1208 		 * by migrating to CPU#0:
1209 		 */
1210 		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1211 			cpu_set_t orig_mask;
1212 			int target_cpu;
1213 			int this_cpu;
1214 
1215 			last_perturbance = stop.tv_sec;
1216 
1217 			/*
1218 			 * Depending on where we are running, move into
1219 			 * the other half of the system, to create some
1220 			 * real disturbance:
1221 			 */
1222 			this_cpu = g->threads[task_nr].curr_cpu;
1223 			if (this_cpu < g->p.nr_cpus/2)
1224 				target_cpu = g->p.nr_cpus-1;
1225 			else
1226 				target_cpu = 0;
1227 
1228 			orig_mask = bind_to_cpu(target_cpu);
1229 
1230 			/* Here we are running on the target CPU already */
1231 			if (details >= 1)
1232 				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1233 
1234 			bind_to_cpumask(orig_mask);
1235 		}
1236 
1237 		if (details >= 3) {
1238 			timersub(&stop, &start, &diff);
1239 			runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1240 			runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1241 
1242 			if (details >= 0) {
1243 				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1244 					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1245 			}
1246 			fflush(stdout);
1247 		}
1248 		if (!last_task)
1249 			continue;
1250 
1251 		timersub(&stop, &start0, &diff);
1252 		runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1253 		runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1254 
1255 		show_summary(runtime_ns_max, l, &convergence);
1256 	}
1257 
1258 	gettimeofday(&stop, NULL);
1259 	timersub(&stop, &start0, &diff);
1260 	td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1261 	td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1262 	secs = td->runtime_ns / NSEC_PER_SEC;
1263 	td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1264 
1265 	getrusage(RUSAGE_THREAD, &rusage);
1266 	td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1267 	td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1268 	td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1269 	td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1270 
1271 	free_data(thread_data, g->p.bytes_thread);
1272 
1273 	pthread_mutex_lock(&g->stop_work_mutex);
1274 	g->bytes_done += bytes_done;
1275 	pthread_mutex_unlock(&g->stop_work_mutex);
1276 
1277 	return NULL;
1278 }
1279 
1280 /*
1281  * A worker process starts a couple of threads:
1282  */
1283 static void worker_process(int process_nr)
1284 {
1285 	pthread_mutex_t process_lock;
1286 	struct thread_data *td;
1287 	pthread_t *pthreads;
1288 	u8 *process_data;
1289 	int task_nr;
1290 	int ret;
1291 	int t;
1292 
1293 	pthread_mutex_init(&process_lock, NULL);
1294 	set_taskname("process %d", process_nr);
1295 
1296 	/*
1297 	 * Pick up the memory policy and the CPU binding of our first thread,
1298 	 * so that we initialize memory accordingly:
1299 	 */
1300 	task_nr = process_nr*g->p.nr_threads;
1301 	td = g->threads + task_nr;
1302 
1303 	bind_to_memnode(td->bind_node);
1304 	bind_to_cpumask(td->bind_cpumask);
1305 
1306 	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1307 	process_data = setup_private_data(g->p.bytes_process);
1308 
1309 	if (g->p.show_details >= 3) {
1310 		printf(" # process %2d global mem: %p, process mem: %p\n",
1311 			process_nr, g->data, process_data);
1312 	}
1313 
1314 	for (t = 0; t < g->p.nr_threads; t++) {
1315 		task_nr = process_nr*g->p.nr_threads + t;
1316 		td = g->threads + task_nr;
1317 
1318 		td->process_data = process_data;
1319 		td->process_nr   = process_nr;
1320 		td->thread_nr    = t;
1321 		td->task_nr	 = task_nr;
1322 		td->val          = rand();
1323 		td->curr_cpu	 = -1;
1324 		td->process_lock = &process_lock;
1325 
1326 		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1327 		BUG_ON(ret);
1328 	}
1329 
1330 	for (t = 0; t < g->p.nr_threads; t++) {
1331                 ret = pthread_join(pthreads[t], NULL);
1332 		BUG_ON(ret);
1333 	}
1334 
1335 	free_data(process_data, g->p.bytes_process);
1336 	free(pthreads);
1337 }
1338 
1339 static void print_summary(void)
1340 {
1341 	if (g->p.show_details < 0)
1342 		return;
1343 
1344 	printf("\n ###\n");
1345 	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1346 		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1347 	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1348 			g->p.nr_loops, g->p.bytes_global/1024/1024);
1349 	printf(" #      %5dx %5ldMB process shared mem operations\n",
1350 			g->p.nr_loops, g->p.bytes_process/1024/1024);
1351 	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1352 			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1353 
1354 	printf(" ###\n");
1355 
1356 	printf("\n ###\n"); fflush(stdout);
1357 }
1358 
1359 static void init_thread_data(void)
1360 {
1361 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1362 	int t;
1363 
1364 	g->threads = zalloc_shared_data(size);
1365 
1366 	for (t = 0; t < g->p.nr_tasks; t++) {
1367 		struct thread_data *td = g->threads + t;
1368 		int cpu;
1369 
1370 		/* Allow all nodes by default: */
1371 		td->bind_node = NUMA_NO_NODE;
1372 
1373 		/* Allow all CPUs by default: */
1374 		CPU_ZERO(&td->bind_cpumask);
1375 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1376 			CPU_SET(cpu, &td->bind_cpumask);
1377 	}
1378 }
1379 
1380 static void deinit_thread_data(void)
1381 {
1382 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1383 
1384 	free_data(g->threads, size);
1385 }
1386 
1387 static int init(void)
1388 {
1389 	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1390 
1391 	/* Copy over options: */
1392 	g->p = p0;
1393 
1394 	g->p.nr_cpus = numa_num_configured_cpus();
1395 
1396 	g->p.nr_nodes = numa_max_node() + 1;
1397 
1398 	/* char array in count_process_nodes(): */
1399 	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1400 
1401 	if (g->p.show_quiet && !g->p.show_details)
1402 		g->p.show_details = -1;
1403 
1404 	/* Some memory should be specified: */
1405 	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1406 		return -1;
1407 
1408 	if (g->p.mb_global_str) {
1409 		g->p.mb_global = atof(g->p.mb_global_str);
1410 		BUG_ON(g->p.mb_global < 0);
1411 	}
1412 
1413 	if (g->p.mb_proc_str) {
1414 		g->p.mb_proc = atof(g->p.mb_proc_str);
1415 		BUG_ON(g->p.mb_proc < 0);
1416 	}
1417 
1418 	if (g->p.mb_proc_locked_str) {
1419 		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1420 		BUG_ON(g->p.mb_proc_locked < 0);
1421 		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1422 	}
1423 
1424 	if (g->p.mb_thread_str) {
1425 		g->p.mb_thread = atof(g->p.mb_thread_str);
1426 		BUG_ON(g->p.mb_thread < 0);
1427 	}
1428 
1429 	BUG_ON(g->p.nr_threads <= 0);
1430 	BUG_ON(g->p.nr_proc <= 0);
1431 
1432 	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1433 
1434 	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1435 	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1436 	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1437 	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1438 
1439 	g->data = setup_shared_data(g->p.bytes_global);
1440 
1441 	/* Startup serialization: */
1442 	init_global_mutex(&g->start_work_mutex);
1443 	init_global_mutex(&g->startup_mutex);
1444 	init_global_mutex(&g->startup_done_mutex);
1445 	init_global_mutex(&g->stop_work_mutex);
1446 
1447 	init_thread_data();
1448 
1449 	tprintf("#\n");
1450 	if (parse_setup_cpu_list() || parse_setup_node_list())
1451 		return -1;
1452 	tprintf("#\n");
1453 
1454 	print_summary();
1455 
1456 	return 0;
1457 }
1458 
1459 static void deinit(void)
1460 {
1461 	free_data(g->data, g->p.bytes_global);
1462 	g->data = NULL;
1463 
1464 	deinit_thread_data();
1465 
1466 	free_data(g, sizeof(*g));
1467 	g = NULL;
1468 }
1469 
1470 /*
1471  * Print a short or long result, depending on the verbosity setting:
1472  */
1473 static void print_res(const char *name, double val,
1474 		      const char *txt_unit, const char *txt_short, const char *txt_long)
1475 {
1476 	if (!name)
1477 		name = "main,";
1478 
1479 	if (!g->p.show_quiet)
1480 		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1481 	else
1482 		printf(" %14.3f %s\n", val, txt_long);
1483 }
1484 
1485 static int __bench_numa(const char *name)
1486 {
1487 	struct timeval start, stop, diff;
1488 	u64 runtime_ns_min, runtime_ns_sum;
1489 	pid_t *pids, pid, wpid;
1490 	double delta_runtime;
1491 	double runtime_avg;
1492 	double runtime_sec_max;
1493 	double runtime_sec_min;
1494 	int wait_stat;
1495 	double bytes;
1496 	int i, t, p;
1497 
1498 	if (init())
1499 		return -1;
1500 
1501 	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1502 	pid = -1;
1503 
1504 	/* All threads try to acquire it, this way we can wait for them to start up: */
1505 	pthread_mutex_lock(&g->start_work_mutex);
1506 
1507 	if (g->p.serialize_startup) {
1508 		tprintf(" #\n");
1509 		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1510 	}
1511 
1512 	gettimeofday(&start, NULL);
1513 
1514 	for (i = 0; i < g->p.nr_proc; i++) {
1515 		pid = fork();
1516 		dprintf(" # process %2d: PID %d\n", i, pid);
1517 
1518 		BUG_ON(pid < 0);
1519 		if (!pid) {
1520 			/* Child process: */
1521 			worker_process(i);
1522 
1523 			exit(0);
1524 		}
1525 		pids[i] = pid;
1526 
1527 	}
1528 	/* Wait for all the threads to start up: */
1529 	while (g->nr_tasks_started != g->p.nr_tasks)
1530 		usleep(USEC_PER_MSEC);
1531 
1532 	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1533 
1534 	if (g->p.serialize_startup) {
1535 		double startup_sec;
1536 
1537 		pthread_mutex_lock(&g->startup_done_mutex);
1538 
1539 		/* This will start all threads: */
1540 		pthread_mutex_unlock(&g->start_work_mutex);
1541 
1542 		/* This mutex is locked - the last started thread will wake us: */
1543 		pthread_mutex_lock(&g->startup_done_mutex);
1544 
1545 		gettimeofday(&stop, NULL);
1546 
1547 		timersub(&stop, &start, &diff);
1548 
1549 		startup_sec = diff.tv_sec * NSEC_PER_SEC;
1550 		startup_sec += diff.tv_usec * NSEC_PER_USEC;
1551 		startup_sec /= NSEC_PER_SEC;
1552 
1553 		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1554 		tprintf(" #\n");
1555 
1556 		start = stop;
1557 		pthread_mutex_unlock(&g->startup_done_mutex);
1558 	} else {
1559 		gettimeofday(&start, NULL);
1560 	}
1561 
1562 	/* Parent process: */
1563 
1564 
1565 	for (i = 0; i < g->p.nr_proc; i++) {
1566 		wpid = waitpid(pids[i], &wait_stat, 0);
1567 		BUG_ON(wpid < 0);
1568 		BUG_ON(!WIFEXITED(wait_stat));
1569 
1570 	}
1571 
1572 	runtime_ns_sum = 0;
1573 	runtime_ns_min = -1LL;
1574 
1575 	for (t = 0; t < g->p.nr_tasks; t++) {
1576 		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1577 
1578 		runtime_ns_sum += thread_runtime_ns;
1579 		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1580 	}
1581 
1582 	gettimeofday(&stop, NULL);
1583 	timersub(&stop, &start, &diff);
1584 
1585 	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1586 
1587 	tprintf("\n ###\n");
1588 	tprintf("\n");
1589 
1590 	runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1591 	runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1592 	runtime_sec_max /= NSEC_PER_SEC;
1593 
1594 	runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1595 
1596 	bytes = g->bytes_done;
1597 	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1598 
1599 	if (g->p.measure_convergence) {
1600 		print_res(name, runtime_sec_max,
1601 			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1602 	}
1603 
1604 	print_res(name, runtime_sec_max,
1605 		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1606 
1607 	print_res(name, runtime_sec_min,
1608 		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1609 
1610 	print_res(name, runtime_avg,
1611 		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1612 
1613 	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1614 	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1615 		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1616 
1617 	print_res(name, bytes / g->p.nr_tasks / 1e9,
1618 		"GB,", "data/thread",		"GB data processed, per thread");
1619 
1620 	print_res(name, bytes / 1e9,
1621 		"GB,", "data-total",		"GB data processed, total");
1622 
1623 	print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1624 		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1625 
1626 	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1627 		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1628 
1629 	print_res(name, bytes / runtime_sec_max / 1e9,
1630 		"GB/sec,", "total-speed",	"GB/sec total speed");
1631 
1632 	if (g->p.show_details >= 2) {
1633 		char tname[14 + 2 * 10 + 1];
1634 		struct thread_data *td;
1635 		for (p = 0; p < g->p.nr_proc; p++) {
1636 			for (t = 0; t < g->p.nr_threads; t++) {
1637 				memset(tname, 0, sizeof(tname));
1638 				td = g->threads + p*g->p.nr_threads + t;
1639 				snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1640 				print_res(tname, td->speed_gbs,
1641 					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1642 				print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1643 					"secs",	"thread-system-time", "system CPU time/thread");
1644 				print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1645 					"secs",	"thread-user-time", "user CPU time/thread");
1646 			}
1647 		}
1648 	}
1649 
1650 	free(pids);
1651 
1652 	deinit();
1653 
1654 	return 0;
1655 }
1656 
1657 #define MAX_ARGS 50
1658 
1659 static int command_size(const char **argv)
1660 {
1661 	int size = 0;
1662 
1663 	while (*argv) {
1664 		size++;
1665 		argv++;
1666 	}
1667 
1668 	BUG_ON(size >= MAX_ARGS);
1669 
1670 	return size;
1671 }
1672 
1673 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1674 {
1675 	int i;
1676 
1677 	printf("\n # Running %s \"perf bench numa", name);
1678 
1679 	for (i = 0; i < argc; i++)
1680 		printf(" %s", argv[i]);
1681 
1682 	printf("\"\n");
1683 
1684 	memset(p, 0, sizeof(*p));
1685 
1686 	/* Initialize nonzero defaults: */
1687 
1688 	p->serialize_startup		= 1;
1689 	p->data_reads			= true;
1690 	p->data_writes			= true;
1691 	p->data_backwards		= true;
1692 	p->data_rand_walk		= true;
1693 	p->nr_loops			= -1;
1694 	p->init_random			= true;
1695 	p->mb_global_str		= "1";
1696 	p->nr_proc			= 1;
1697 	p->nr_threads			= 1;
1698 	p->nr_secs			= 5;
1699 	p->run_all			= argc == 1;
1700 }
1701 
1702 static int run_bench_numa(const char *name, const char **argv)
1703 {
1704 	int argc = command_size(argv);
1705 
1706 	init_params(&p0, name, argc, argv);
1707 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1708 	if (argc)
1709 		goto err;
1710 
1711 	if (__bench_numa(name))
1712 		goto err;
1713 
1714 	return 0;
1715 
1716 err:
1717 	return -1;
1718 }
1719 
1720 #define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1721 #define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1722 
1723 #define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1724 #define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1725 
1726 #define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1727 #define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1728 
1729 /*
1730  * The built-in test-suite executed by "perf bench numa -a".
1731  *
1732  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1733  */
1734 static const char *tests[][MAX_ARGS] = {
1735    /* Basic single-stream NUMA bandwidth measurements: */
1736    { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1737 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1738    { "RAM-bw-local-NOTHP,",
1739 			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1740 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1741    { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1742 			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1743 
1744    /* 2-stream NUMA bandwidth measurements: */
1745    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1746 			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1747    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1748 		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1749 
1750    /* Cross-stream NUMA bandwidth measurement: */
1751    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1752 		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1753 
1754    /* Convergence latency measurements: */
1755    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1756    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1757    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1758    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1759    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1760    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1761    { " 4x4-convergence-NOTHP,",
1762 			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1763    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1764    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1765    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1766    { " 8x4-convergence-NOTHP,",
1767 			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1768    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1769    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1770    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1771    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1772    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1773 
1774    /* Various NUMA process/thread layout bandwidth measurements: */
1775    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1776    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1777    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1778    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1779    { " 8x1-bw-process-NOTHP,",
1780 			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1781    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1782 
1783    { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1784    { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1785    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1786    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1787 
1788    { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1789    { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1790    { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1791    { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1792    { " 4x8-bw-thread-NOTHP,",
1793 			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1794    { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1795    { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1796 
1797    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1798    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1799 
1800    { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1801    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1802    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1803    { "numa01-bw-thread-NOTHP,",
1804 			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1805 };
1806 
1807 static int bench_all(void)
1808 {
1809 	int nr = ARRAY_SIZE(tests);
1810 	int ret;
1811 	int i;
1812 
1813 	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1814 	BUG_ON(ret < 0);
1815 
1816 	for (i = 0; i < nr; i++) {
1817 		run_bench_numa(tests[i][0], tests[i] + 1);
1818 	}
1819 
1820 	printf("\n");
1821 
1822 	return 0;
1823 }
1824 
1825 int bench_numa(int argc, const char **argv)
1826 {
1827 	init_params(&p0, "main,", argc, argv);
1828 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1829 	if (argc)
1830 		goto err;
1831 
1832 	if (p0.run_all)
1833 		return bench_all();
1834 
1835 	if (__bench_numa(NULL))
1836 		goto err;
1837 
1838 	return 0;
1839 
1840 err:
1841 	usage_with_options(numa_usage, options);
1842 	return -1;
1843 }
1844