xref: /linux/tools/perf/bench/numa.c (revision 856e7c4b619af622d56b3b454f7bec32a170ac99)
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 "../perf.h"
13 #include "../builtin.h"
14 #include "../util/util.h"
15 #include <subcmd/parse-options.h>
16 #include "../util/cloexec.h"
17 
18 #include "bench.h"
19 
20 #include <errno.h>
21 #include <sched.h>
22 #include <stdio.h>
23 #include <assert.h>
24 #include <malloc.h>
25 #include <signal.h>
26 #include <stdlib.h>
27 #include <string.h>
28 #include <unistd.h>
29 #include <sys/mman.h>
30 #include <sys/time.h>
31 #include <sys/resource.h>
32 #include <sys/wait.h>
33 #include <sys/prctl.h>
34 #include <sys/types.h>
35 #include <linux/kernel.h>
36 #include <linux/time64.h>
37 
38 #include <numa.h>
39 #include <numaif.h>
40 
41 /*
42  * Regular printout to the terminal, supressed if -q is specified:
43  */
44 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
45 
46 /*
47  * Debug printf:
48  */
49 #undef dprintf
50 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
51 
52 struct thread_data {
53 	int			curr_cpu;
54 	cpu_set_t		bind_cpumask;
55 	int			bind_node;
56 	u8			*process_data;
57 	int			process_nr;
58 	int			thread_nr;
59 	int			task_nr;
60 	unsigned int		loops_done;
61 	u64			val;
62 	u64			runtime_ns;
63 	u64			system_time_ns;
64 	u64			user_time_ns;
65 	double			speed_gbs;
66 	pthread_mutex_t		*process_lock;
67 };
68 
69 /* Parameters set by options: */
70 
71 struct params {
72 	/* Startup synchronization: */
73 	bool			serialize_startup;
74 
75 	/* Task hierarchy: */
76 	int			nr_proc;
77 	int			nr_threads;
78 
79 	/* Working set sizes: */
80 	const char		*mb_global_str;
81 	const char		*mb_proc_str;
82 	const char		*mb_proc_locked_str;
83 	const char		*mb_thread_str;
84 
85 	double			mb_global;
86 	double			mb_proc;
87 	double			mb_proc_locked;
88 	double			mb_thread;
89 
90 	/* Access patterns to the working set: */
91 	bool			data_reads;
92 	bool			data_writes;
93 	bool			data_backwards;
94 	bool			data_zero_memset;
95 	bool			data_rand_walk;
96 	u32			nr_loops;
97 	u32			nr_secs;
98 	u32			sleep_usecs;
99 
100 	/* Working set initialization: */
101 	bool			init_zero;
102 	bool			init_random;
103 	bool			init_cpu0;
104 
105 	/* Misc options: */
106 	int			show_details;
107 	int			run_all;
108 	int			thp;
109 
110 	long			bytes_global;
111 	long			bytes_process;
112 	long			bytes_process_locked;
113 	long			bytes_thread;
114 
115 	int			nr_tasks;
116 	bool			show_quiet;
117 
118 	bool			show_convergence;
119 	bool			measure_convergence;
120 
121 	int			perturb_secs;
122 	int			nr_cpus;
123 	int			nr_nodes;
124 
125 	/* Affinity options -C and -N: */
126 	char			*cpu_list_str;
127 	char			*node_list_str;
128 };
129 
130 
131 /* Global, read-writable area, accessible to all processes and threads: */
132 
133 struct global_info {
134 	u8			*data;
135 
136 	pthread_mutex_t		startup_mutex;
137 	int			nr_tasks_started;
138 
139 	pthread_mutex_t		startup_done_mutex;
140 
141 	pthread_mutex_t		start_work_mutex;
142 	int			nr_tasks_working;
143 
144 	pthread_mutex_t		stop_work_mutex;
145 	u64			bytes_done;
146 
147 	struct thread_data	*threads;
148 
149 	/* Convergence latency measurement: */
150 	bool			all_converged;
151 	bool			stop_work;
152 
153 	int			print_once;
154 
155 	struct params		p;
156 };
157 
158 static struct global_info	*g = NULL;
159 
160 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
161 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
162 
163 struct params p0;
164 
165 static const struct option options[] = {
166 	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
167 	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
168 
169 	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
170 	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
171 	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
172 	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
173 
174 	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
175 	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
176 	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
177 
178 	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via reads (can be mixed with -W)"),
179 	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
180 	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
181 	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
182 	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
183 
184 
185 	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
186 	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
187 	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
188 	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
189 
190 	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
191 	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
192 	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
193 	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
194 		    "convergence is reached when each process (all its threads) is running on a single NUMA node."),
195 	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
196 	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"quiet mode"),
197 	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
198 
199 	/* Special option string parsing callbacks: */
200         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
201 			"bind the first N tasks to these specific cpus (the rest is unbound)",
202 			parse_cpus_opt),
203         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
204 			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
205 			parse_nodes_opt),
206 	OPT_END()
207 };
208 
209 static const char * const bench_numa_usage[] = {
210 	"perf bench numa <options>",
211 	NULL
212 };
213 
214 static const char * const numa_usage[] = {
215 	"perf bench numa mem [<options>]",
216 	NULL
217 };
218 
219 /*
220  * To get number of numa nodes present.
221  */
222 static int nr_numa_nodes(void)
223 {
224 	int i, nr_nodes = 0;
225 
226 	for (i = 0; i < g->p.nr_nodes; i++) {
227 		if (numa_bitmask_isbitset(numa_nodes_ptr, i))
228 			nr_nodes++;
229 	}
230 
231 	return nr_nodes;
232 }
233 
234 /*
235  * To check if given numa node is present.
236  */
237 static int is_node_present(int node)
238 {
239 	return numa_bitmask_isbitset(numa_nodes_ptr, node);
240 }
241 
242 /*
243  * To check given numa node has cpus.
244  */
245 static bool node_has_cpus(int node)
246 {
247 	struct bitmask *cpu = numa_allocate_cpumask();
248 	unsigned int i;
249 
250 	if (cpu && !numa_node_to_cpus(node, cpu)) {
251 		for (i = 0; i < cpu->size; i++) {
252 			if (numa_bitmask_isbitset(cpu, i))
253 				return true;
254 		}
255 	}
256 
257 	return false; /* lets fall back to nocpus safely */
258 }
259 
260 static cpu_set_t bind_to_cpu(int target_cpu)
261 {
262 	cpu_set_t orig_mask, mask;
263 	int ret;
264 
265 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
266 	BUG_ON(ret);
267 
268 	CPU_ZERO(&mask);
269 
270 	if (target_cpu == -1) {
271 		int cpu;
272 
273 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
274 			CPU_SET(cpu, &mask);
275 	} else {
276 		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
277 		CPU_SET(target_cpu, &mask);
278 	}
279 
280 	ret = sched_setaffinity(0, sizeof(mask), &mask);
281 	BUG_ON(ret);
282 
283 	return orig_mask;
284 }
285 
286 static cpu_set_t bind_to_node(int target_node)
287 {
288 	int cpus_per_node = g->p.nr_cpus / nr_numa_nodes();
289 	cpu_set_t orig_mask, mask;
290 	int cpu;
291 	int ret;
292 
293 	BUG_ON(cpus_per_node * nr_numa_nodes() != g->p.nr_cpus);
294 	BUG_ON(!cpus_per_node);
295 
296 	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
297 	BUG_ON(ret);
298 
299 	CPU_ZERO(&mask);
300 
301 	if (target_node == -1) {
302 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
303 			CPU_SET(cpu, &mask);
304 	} else {
305 		int cpu_start = (target_node + 0) * cpus_per_node;
306 		int cpu_stop  = (target_node + 1) * cpus_per_node;
307 
308 		BUG_ON(cpu_stop > g->p.nr_cpus);
309 
310 		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
311 			CPU_SET(cpu, &mask);
312 	}
313 
314 	ret = sched_setaffinity(0, sizeof(mask), &mask);
315 	BUG_ON(ret);
316 
317 	return orig_mask;
318 }
319 
320 static void bind_to_cpumask(cpu_set_t mask)
321 {
322 	int ret;
323 
324 	ret = sched_setaffinity(0, sizeof(mask), &mask);
325 	BUG_ON(ret);
326 }
327 
328 static void mempol_restore(void)
329 {
330 	int ret;
331 
332 	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
333 
334 	BUG_ON(ret);
335 }
336 
337 static void bind_to_memnode(int node)
338 {
339 	unsigned long nodemask;
340 	int ret;
341 
342 	if (node == -1)
343 		return;
344 
345 	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
346 	nodemask = 1L << node;
347 
348 	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
349 	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
350 
351 	BUG_ON(ret);
352 }
353 
354 #define HPSIZE (2*1024*1024)
355 
356 #define set_taskname(fmt...)				\
357 do {							\
358 	char name[20];					\
359 							\
360 	snprintf(name, 20, fmt);			\
361 	prctl(PR_SET_NAME, name);			\
362 } while (0)
363 
364 static u8 *alloc_data(ssize_t bytes0, int map_flags,
365 		      int init_zero, int init_cpu0, int thp, int init_random)
366 {
367 	cpu_set_t orig_mask;
368 	ssize_t bytes;
369 	u8 *buf;
370 	int ret;
371 
372 	if (!bytes0)
373 		return NULL;
374 
375 	/* Allocate and initialize all memory on CPU#0: */
376 	if (init_cpu0) {
377 		orig_mask = bind_to_node(0);
378 		bind_to_memnode(0);
379 	}
380 
381 	bytes = bytes0 + HPSIZE;
382 
383 	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
384 	BUG_ON(buf == (void *)-1);
385 
386 	if (map_flags == MAP_PRIVATE) {
387 		if (thp > 0) {
388 			ret = madvise(buf, bytes, MADV_HUGEPAGE);
389 			if (ret && !g->print_once) {
390 				g->print_once = 1;
391 				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
392 			}
393 		}
394 		if (thp < 0) {
395 			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
396 			if (ret && !g->print_once) {
397 				g->print_once = 1;
398 				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
399 			}
400 		}
401 	}
402 
403 	if (init_zero) {
404 		bzero(buf, bytes);
405 	} else {
406 		/* Initialize random contents, different in each word: */
407 		if (init_random) {
408 			u64 *wbuf = (void *)buf;
409 			long off = rand();
410 			long i;
411 
412 			for (i = 0; i < bytes/8; i++)
413 				wbuf[i] = i + off;
414 		}
415 	}
416 
417 	/* Align to 2MB boundary: */
418 	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
419 
420 	/* Restore affinity: */
421 	if (init_cpu0) {
422 		bind_to_cpumask(orig_mask);
423 		mempol_restore();
424 	}
425 
426 	return buf;
427 }
428 
429 static void free_data(void *data, ssize_t bytes)
430 {
431 	int ret;
432 
433 	if (!data)
434 		return;
435 
436 	ret = munmap(data, bytes);
437 	BUG_ON(ret);
438 }
439 
440 /*
441  * Create a shared memory buffer that can be shared between processes, zeroed:
442  */
443 static void * zalloc_shared_data(ssize_t bytes)
444 {
445 	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
446 }
447 
448 /*
449  * Create a shared memory buffer that can be shared between processes:
450  */
451 static void * setup_shared_data(ssize_t bytes)
452 {
453 	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
454 }
455 
456 /*
457  * Allocate process-local memory - this will either be shared between
458  * threads of this process, or only be accessed by this thread:
459  */
460 static void * setup_private_data(ssize_t bytes)
461 {
462 	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
463 }
464 
465 /*
466  * Return a process-shared (global) mutex:
467  */
468 static void init_global_mutex(pthread_mutex_t *mutex)
469 {
470 	pthread_mutexattr_t attr;
471 
472 	pthread_mutexattr_init(&attr);
473 	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
474 	pthread_mutex_init(mutex, &attr);
475 }
476 
477 static int parse_cpu_list(const char *arg)
478 {
479 	p0.cpu_list_str = strdup(arg);
480 
481 	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
482 
483 	return 0;
484 }
485 
486 static int parse_setup_cpu_list(void)
487 {
488 	struct thread_data *td;
489 	char *str0, *str;
490 	int t;
491 
492 	if (!g->p.cpu_list_str)
493 		return 0;
494 
495 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
496 
497 	str0 = str = strdup(g->p.cpu_list_str);
498 	t = 0;
499 
500 	BUG_ON(!str);
501 
502 	tprintf("# binding tasks to CPUs:\n");
503 	tprintf("#  ");
504 
505 	while (true) {
506 		int bind_cpu, bind_cpu_0, bind_cpu_1;
507 		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
508 		int bind_len;
509 		int step;
510 		int mul;
511 
512 		tok = strsep(&str, ",");
513 		if (!tok)
514 			break;
515 
516 		tok_end = strstr(tok, "-");
517 
518 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
519 		if (!tok_end) {
520 			/* Single CPU specified: */
521 			bind_cpu_0 = bind_cpu_1 = atol(tok);
522 		} else {
523 			/* CPU range specified (for example: "5-11"): */
524 			bind_cpu_0 = atol(tok);
525 			bind_cpu_1 = atol(tok_end + 1);
526 		}
527 
528 		step = 1;
529 		tok_step = strstr(tok, "#");
530 		if (tok_step) {
531 			step = atol(tok_step + 1);
532 			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
533 		}
534 
535 		/*
536 		 * Mask length.
537 		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
538 		 * where the _4 means the next 4 CPUs are allowed.
539 		 */
540 		bind_len = 1;
541 		tok_len = strstr(tok, "_");
542 		if (tok_len) {
543 			bind_len = atol(tok_len + 1);
544 			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
545 		}
546 
547 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
548 		mul = 1;
549 		tok_mul = strstr(tok, "x");
550 		if (tok_mul) {
551 			mul = atol(tok_mul + 1);
552 			BUG_ON(mul <= 0);
553 		}
554 
555 		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
556 
557 		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
558 			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
559 			return -1;
560 		}
561 
562 		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
563 		BUG_ON(bind_cpu_0 > bind_cpu_1);
564 
565 		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
566 			int i;
567 
568 			for (i = 0; i < mul; i++) {
569 				int cpu;
570 
571 				if (t >= g->p.nr_tasks) {
572 					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
573 					goto out;
574 				}
575 				td = g->threads + t;
576 
577 				if (t)
578 					tprintf(",");
579 				if (bind_len > 1) {
580 					tprintf("%2d/%d", bind_cpu, bind_len);
581 				} else {
582 					tprintf("%2d", bind_cpu);
583 				}
584 
585 				CPU_ZERO(&td->bind_cpumask);
586 				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
587 					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
588 					CPU_SET(cpu, &td->bind_cpumask);
589 				}
590 				t++;
591 			}
592 		}
593 	}
594 out:
595 
596 	tprintf("\n");
597 
598 	if (t < g->p.nr_tasks)
599 		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
600 
601 	free(str0);
602 	return 0;
603 }
604 
605 static int parse_cpus_opt(const struct option *opt __maybe_unused,
606 			  const char *arg, int unset __maybe_unused)
607 {
608 	if (!arg)
609 		return -1;
610 
611 	return parse_cpu_list(arg);
612 }
613 
614 static int parse_node_list(const char *arg)
615 {
616 	p0.node_list_str = strdup(arg);
617 
618 	dprintf("got NODE list: {%s}\n", p0.node_list_str);
619 
620 	return 0;
621 }
622 
623 static int parse_setup_node_list(void)
624 {
625 	struct thread_data *td;
626 	char *str0, *str;
627 	int t;
628 
629 	if (!g->p.node_list_str)
630 		return 0;
631 
632 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
633 
634 	str0 = str = strdup(g->p.node_list_str);
635 	t = 0;
636 
637 	BUG_ON(!str);
638 
639 	tprintf("# binding tasks to NODEs:\n");
640 	tprintf("# ");
641 
642 	while (true) {
643 		int bind_node, bind_node_0, bind_node_1;
644 		char *tok, *tok_end, *tok_step, *tok_mul;
645 		int step;
646 		int mul;
647 
648 		tok = strsep(&str, ",");
649 		if (!tok)
650 			break;
651 
652 		tok_end = strstr(tok, "-");
653 
654 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
655 		if (!tok_end) {
656 			/* Single NODE specified: */
657 			bind_node_0 = bind_node_1 = atol(tok);
658 		} else {
659 			/* NODE range specified (for example: "5-11"): */
660 			bind_node_0 = atol(tok);
661 			bind_node_1 = atol(tok_end + 1);
662 		}
663 
664 		step = 1;
665 		tok_step = strstr(tok, "#");
666 		if (tok_step) {
667 			step = atol(tok_step + 1);
668 			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
669 		}
670 
671 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
672 		mul = 1;
673 		tok_mul = strstr(tok, "x");
674 		if (tok_mul) {
675 			mul = atol(tok_mul + 1);
676 			BUG_ON(mul <= 0);
677 		}
678 
679 		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
680 
681 		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
682 			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
683 			return -1;
684 		}
685 
686 		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
687 		BUG_ON(bind_node_0 > bind_node_1);
688 
689 		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
690 			int i;
691 
692 			for (i = 0; i < mul; i++) {
693 				if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
694 					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
695 					goto out;
696 				}
697 				td = g->threads + t;
698 
699 				if (!t)
700 					tprintf(" %2d", bind_node);
701 				else
702 					tprintf(",%2d", bind_node);
703 
704 				td->bind_node = bind_node;
705 				t++;
706 			}
707 		}
708 	}
709 out:
710 
711 	tprintf("\n");
712 
713 	if (t < g->p.nr_tasks)
714 		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
715 
716 	free(str0);
717 	return 0;
718 }
719 
720 static int parse_nodes_opt(const struct option *opt __maybe_unused,
721 			  const char *arg, int unset __maybe_unused)
722 {
723 	if (!arg)
724 		return -1;
725 
726 	return parse_node_list(arg);
727 
728 	return 0;
729 }
730 
731 #define BIT(x) (1ul << x)
732 
733 static inline uint32_t lfsr_32(uint32_t lfsr)
734 {
735 	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
736 	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
737 }
738 
739 /*
740  * Make sure there's real data dependency to RAM (when read
741  * accesses are enabled), so the compiler, the CPU and the
742  * kernel (KSM, zero page, etc.) cannot optimize away RAM
743  * accesses:
744  */
745 static inline u64 access_data(u64 *data, u64 val)
746 {
747 	if (g->p.data_reads)
748 		val += *data;
749 	if (g->p.data_writes)
750 		*data = val + 1;
751 	return val;
752 }
753 
754 /*
755  * The worker process does two types of work, a forwards going
756  * loop and a backwards going loop.
757  *
758  * We do this so that on multiprocessor systems we do not create
759  * a 'train' of processing, with highly synchronized processes,
760  * skewing the whole benchmark.
761  */
762 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
763 {
764 	long words = bytes/sizeof(u64);
765 	u64 *data = (void *)__data;
766 	long chunk_0, chunk_1;
767 	u64 *d0, *d, *d1;
768 	long off;
769 	long i;
770 
771 	BUG_ON(!data && words);
772 	BUG_ON(data && !words);
773 
774 	if (!data)
775 		return val;
776 
777 	/* Very simple memset() work variant: */
778 	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
779 		bzero(data, bytes);
780 		return val;
781 	}
782 
783 	/* Spread out by PID/TID nr and by loop nr: */
784 	chunk_0 = words/nr_max;
785 	chunk_1 = words/g->p.nr_loops;
786 	off = nr*chunk_0 + loop*chunk_1;
787 
788 	while (off >= words)
789 		off -= words;
790 
791 	if (g->p.data_rand_walk) {
792 		u32 lfsr = nr + loop + val;
793 		int j;
794 
795 		for (i = 0; i < words/1024; i++) {
796 			long start, end;
797 
798 			lfsr = lfsr_32(lfsr);
799 
800 			start = lfsr % words;
801 			end = min(start + 1024, words-1);
802 
803 			if (g->p.data_zero_memset) {
804 				bzero(data + start, (end-start) * sizeof(u64));
805 			} else {
806 				for (j = start; j < end; j++)
807 					val = access_data(data + j, val);
808 			}
809 		}
810 	} else if (!g->p.data_backwards || (nr + loop) & 1) {
811 
812 		d0 = data + off;
813 		d  = data + off + 1;
814 		d1 = data + words;
815 
816 		/* Process data forwards: */
817 		for (;;) {
818 			if (unlikely(d >= d1))
819 				d = data;
820 			if (unlikely(d == d0))
821 				break;
822 
823 			val = access_data(d, val);
824 
825 			d++;
826 		}
827 	} else {
828 		/* Process data backwards: */
829 
830 		d0 = data + off;
831 		d  = data + off - 1;
832 		d1 = data + words;
833 
834 		/* Process data forwards: */
835 		for (;;) {
836 			if (unlikely(d < data))
837 				d = data + words-1;
838 			if (unlikely(d == d0))
839 				break;
840 
841 			val = access_data(d, val);
842 
843 			d--;
844 		}
845 	}
846 
847 	return val;
848 }
849 
850 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
851 {
852 	unsigned int cpu;
853 
854 	cpu = sched_getcpu();
855 
856 	g->threads[task_nr].curr_cpu = cpu;
857 	prctl(0, bytes_worked);
858 }
859 
860 #define MAX_NR_NODES	64
861 
862 /*
863  * Count the number of nodes a process's threads
864  * are spread out on.
865  *
866  * A count of 1 means that the process is compressed
867  * to a single node. A count of g->p.nr_nodes means it's
868  * spread out on the whole system.
869  */
870 static int count_process_nodes(int process_nr)
871 {
872 	char node_present[MAX_NR_NODES] = { 0, };
873 	int nodes;
874 	int n, t;
875 
876 	for (t = 0; t < g->p.nr_threads; t++) {
877 		struct thread_data *td;
878 		int task_nr;
879 		int node;
880 
881 		task_nr = process_nr*g->p.nr_threads + t;
882 		td = g->threads + task_nr;
883 
884 		node = numa_node_of_cpu(td->curr_cpu);
885 		if (node < 0) /* curr_cpu was likely still -1 */
886 			return 0;
887 
888 		node_present[node] = 1;
889 	}
890 
891 	nodes = 0;
892 
893 	for (n = 0; n < MAX_NR_NODES; n++)
894 		nodes += node_present[n];
895 
896 	return nodes;
897 }
898 
899 /*
900  * Count the number of distinct process-threads a node contains.
901  *
902  * A count of 1 means that the node contains only a single
903  * process. If all nodes on the system contain at most one
904  * process then we are well-converged.
905  */
906 static int count_node_processes(int node)
907 {
908 	int processes = 0;
909 	int t, p;
910 
911 	for (p = 0; p < g->p.nr_proc; p++) {
912 		for (t = 0; t < g->p.nr_threads; t++) {
913 			struct thread_data *td;
914 			int task_nr;
915 			int n;
916 
917 			task_nr = p*g->p.nr_threads + t;
918 			td = g->threads + task_nr;
919 
920 			n = numa_node_of_cpu(td->curr_cpu);
921 			if (n == node) {
922 				processes++;
923 				break;
924 			}
925 		}
926 	}
927 
928 	return processes;
929 }
930 
931 static void calc_convergence_compression(int *strong)
932 {
933 	unsigned int nodes_min, nodes_max;
934 	int p;
935 
936 	nodes_min = -1;
937 	nodes_max =  0;
938 
939 	for (p = 0; p < g->p.nr_proc; p++) {
940 		unsigned int nodes = count_process_nodes(p);
941 
942 		if (!nodes) {
943 			*strong = 0;
944 			return;
945 		}
946 
947 		nodes_min = min(nodes, nodes_min);
948 		nodes_max = max(nodes, nodes_max);
949 	}
950 
951 	/* Strong convergence: all threads compress on a single node: */
952 	if (nodes_min == 1 && nodes_max == 1) {
953 		*strong = 1;
954 	} else {
955 		*strong = 0;
956 		tprintf(" {%d-%d}", nodes_min, nodes_max);
957 	}
958 }
959 
960 static void calc_convergence(double runtime_ns_max, double *convergence)
961 {
962 	unsigned int loops_done_min, loops_done_max;
963 	int process_groups;
964 	int nodes[MAX_NR_NODES];
965 	int distance;
966 	int nr_min;
967 	int nr_max;
968 	int strong;
969 	int sum;
970 	int nr;
971 	int node;
972 	int cpu;
973 	int t;
974 
975 	if (!g->p.show_convergence && !g->p.measure_convergence)
976 		return;
977 
978 	for (node = 0; node < g->p.nr_nodes; node++)
979 		nodes[node] = 0;
980 
981 	loops_done_min = -1;
982 	loops_done_max = 0;
983 
984 	for (t = 0; t < g->p.nr_tasks; t++) {
985 		struct thread_data *td = g->threads + t;
986 		unsigned int loops_done;
987 
988 		cpu = td->curr_cpu;
989 
990 		/* Not all threads have written it yet: */
991 		if (cpu < 0)
992 			continue;
993 
994 		node = numa_node_of_cpu(cpu);
995 
996 		nodes[node]++;
997 
998 		loops_done = td->loops_done;
999 		loops_done_min = min(loops_done, loops_done_min);
1000 		loops_done_max = max(loops_done, loops_done_max);
1001 	}
1002 
1003 	nr_max = 0;
1004 	nr_min = g->p.nr_tasks;
1005 	sum = 0;
1006 
1007 	for (node = 0; node < g->p.nr_nodes; node++) {
1008 		if (!is_node_present(node))
1009 			continue;
1010 		nr = nodes[node];
1011 		nr_min = min(nr, nr_min);
1012 		nr_max = max(nr, nr_max);
1013 		sum += nr;
1014 	}
1015 	BUG_ON(nr_min > nr_max);
1016 
1017 	BUG_ON(sum > g->p.nr_tasks);
1018 
1019 	if (0 && (sum < g->p.nr_tasks))
1020 		return;
1021 
1022 	/*
1023 	 * Count the number of distinct process groups present
1024 	 * on nodes - when we are converged this will decrease
1025 	 * to g->p.nr_proc:
1026 	 */
1027 	process_groups = 0;
1028 
1029 	for (node = 0; node < g->p.nr_nodes; node++) {
1030 		int processes;
1031 
1032 		if (!is_node_present(node))
1033 			continue;
1034 		processes = count_node_processes(node);
1035 		nr = nodes[node];
1036 		tprintf(" %2d/%-2d", nr, processes);
1037 
1038 		process_groups += processes;
1039 	}
1040 
1041 	distance = nr_max - nr_min;
1042 
1043 	tprintf(" [%2d/%-2d]", distance, process_groups);
1044 
1045 	tprintf(" l:%3d-%-3d (%3d)",
1046 		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1047 
1048 	if (loops_done_min && loops_done_max) {
1049 		double skew = 1.0 - (double)loops_done_min/loops_done_max;
1050 
1051 		tprintf(" [%4.1f%%]", skew * 100.0);
1052 	}
1053 
1054 	calc_convergence_compression(&strong);
1055 
1056 	if (strong && process_groups == g->p.nr_proc) {
1057 		if (!*convergence) {
1058 			*convergence = runtime_ns_max;
1059 			tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1060 			if (g->p.measure_convergence) {
1061 				g->all_converged = true;
1062 				g->stop_work = true;
1063 			}
1064 		}
1065 	} else {
1066 		if (*convergence) {
1067 			tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1068 			*convergence = 0;
1069 		}
1070 		tprintf("\n");
1071 	}
1072 }
1073 
1074 static void show_summary(double runtime_ns_max, int l, double *convergence)
1075 {
1076 	tprintf("\r #  %5.1f%%  [%.1f mins]",
1077 		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1078 
1079 	calc_convergence(runtime_ns_max, convergence);
1080 
1081 	if (g->p.show_details >= 0)
1082 		fflush(stdout);
1083 }
1084 
1085 static void *worker_thread(void *__tdata)
1086 {
1087 	struct thread_data *td = __tdata;
1088 	struct timeval start0, start, stop, diff;
1089 	int process_nr = td->process_nr;
1090 	int thread_nr = td->thread_nr;
1091 	unsigned long last_perturbance;
1092 	int task_nr = td->task_nr;
1093 	int details = g->p.show_details;
1094 	int first_task, last_task;
1095 	double convergence = 0;
1096 	u64 val = td->val;
1097 	double runtime_ns_max;
1098 	u8 *global_data;
1099 	u8 *process_data;
1100 	u8 *thread_data;
1101 	u64 bytes_done;
1102 	long work_done;
1103 	u32 l;
1104 	struct rusage rusage;
1105 
1106 	bind_to_cpumask(td->bind_cpumask);
1107 	bind_to_memnode(td->bind_node);
1108 
1109 	set_taskname("thread %d/%d", process_nr, thread_nr);
1110 
1111 	global_data = g->data;
1112 	process_data = td->process_data;
1113 	thread_data = setup_private_data(g->p.bytes_thread);
1114 
1115 	bytes_done = 0;
1116 
1117 	last_task = 0;
1118 	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1119 		last_task = 1;
1120 
1121 	first_task = 0;
1122 	if (process_nr == 0 && thread_nr == 0)
1123 		first_task = 1;
1124 
1125 	if (details >= 2) {
1126 		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1127 			process_nr, thread_nr, global_data, process_data, thread_data);
1128 	}
1129 
1130 	if (g->p.serialize_startup) {
1131 		pthread_mutex_lock(&g->startup_mutex);
1132 		g->nr_tasks_started++;
1133 		pthread_mutex_unlock(&g->startup_mutex);
1134 
1135 		/* Here we will wait for the main process to start us all at once: */
1136 		pthread_mutex_lock(&g->start_work_mutex);
1137 		g->nr_tasks_working++;
1138 
1139 		/* Last one wake the main process: */
1140 		if (g->nr_tasks_working == g->p.nr_tasks)
1141 			pthread_mutex_unlock(&g->startup_done_mutex);
1142 
1143 		pthread_mutex_unlock(&g->start_work_mutex);
1144 	}
1145 
1146 	gettimeofday(&start0, NULL);
1147 
1148 	start = stop = start0;
1149 	last_perturbance = start.tv_sec;
1150 
1151 	for (l = 0; l < g->p.nr_loops; l++) {
1152 		start = stop;
1153 
1154 		if (g->stop_work)
1155 			break;
1156 
1157 		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1158 		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1159 		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1160 
1161 		if (g->p.sleep_usecs) {
1162 			pthread_mutex_lock(td->process_lock);
1163 			usleep(g->p.sleep_usecs);
1164 			pthread_mutex_unlock(td->process_lock);
1165 		}
1166 		/*
1167 		 * Amount of work to be done under a process-global lock:
1168 		 */
1169 		if (g->p.bytes_process_locked) {
1170 			pthread_mutex_lock(td->process_lock);
1171 			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1172 			pthread_mutex_unlock(td->process_lock);
1173 		}
1174 
1175 		work_done = g->p.bytes_global + g->p.bytes_process +
1176 			    g->p.bytes_process_locked + g->p.bytes_thread;
1177 
1178 		update_curr_cpu(task_nr, work_done);
1179 		bytes_done += work_done;
1180 
1181 		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1182 			continue;
1183 
1184 		td->loops_done = l;
1185 
1186 		gettimeofday(&stop, NULL);
1187 
1188 		/* Check whether our max runtime timed out: */
1189 		if (g->p.nr_secs) {
1190 			timersub(&stop, &start0, &diff);
1191 			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1192 				g->stop_work = true;
1193 				break;
1194 			}
1195 		}
1196 
1197 		/* Update the summary at most once per second: */
1198 		if (start.tv_sec == stop.tv_sec)
1199 			continue;
1200 
1201 		/*
1202 		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1203 		 * by migrating to CPU#0:
1204 		 */
1205 		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1206 			cpu_set_t orig_mask;
1207 			int target_cpu;
1208 			int this_cpu;
1209 
1210 			last_perturbance = stop.tv_sec;
1211 
1212 			/*
1213 			 * Depending on where we are running, move into
1214 			 * the other half of the system, to create some
1215 			 * real disturbance:
1216 			 */
1217 			this_cpu = g->threads[task_nr].curr_cpu;
1218 			if (this_cpu < g->p.nr_cpus/2)
1219 				target_cpu = g->p.nr_cpus-1;
1220 			else
1221 				target_cpu = 0;
1222 
1223 			orig_mask = bind_to_cpu(target_cpu);
1224 
1225 			/* Here we are running on the target CPU already */
1226 			if (details >= 1)
1227 				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1228 
1229 			bind_to_cpumask(orig_mask);
1230 		}
1231 
1232 		if (details >= 3) {
1233 			timersub(&stop, &start, &diff);
1234 			runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1235 			runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1236 
1237 			if (details >= 0) {
1238 				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1239 					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1240 			}
1241 			fflush(stdout);
1242 		}
1243 		if (!last_task)
1244 			continue;
1245 
1246 		timersub(&stop, &start0, &diff);
1247 		runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1248 		runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1249 
1250 		show_summary(runtime_ns_max, l, &convergence);
1251 	}
1252 
1253 	gettimeofday(&stop, NULL);
1254 	timersub(&stop, &start0, &diff);
1255 	td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1256 	td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1257 	td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1258 
1259 	getrusage(RUSAGE_THREAD, &rusage);
1260 	td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1261 	td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1262 	td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1263 	td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1264 
1265 	free_data(thread_data, g->p.bytes_thread);
1266 
1267 	pthread_mutex_lock(&g->stop_work_mutex);
1268 	g->bytes_done += bytes_done;
1269 	pthread_mutex_unlock(&g->stop_work_mutex);
1270 
1271 	return NULL;
1272 }
1273 
1274 /*
1275  * A worker process starts a couple of threads:
1276  */
1277 static void worker_process(int process_nr)
1278 {
1279 	pthread_mutex_t process_lock;
1280 	struct thread_data *td;
1281 	pthread_t *pthreads;
1282 	u8 *process_data;
1283 	int task_nr;
1284 	int ret;
1285 	int t;
1286 
1287 	pthread_mutex_init(&process_lock, NULL);
1288 	set_taskname("process %d", process_nr);
1289 
1290 	/*
1291 	 * Pick up the memory policy and the CPU binding of our first thread,
1292 	 * so that we initialize memory accordingly:
1293 	 */
1294 	task_nr = process_nr*g->p.nr_threads;
1295 	td = g->threads + task_nr;
1296 
1297 	bind_to_memnode(td->bind_node);
1298 	bind_to_cpumask(td->bind_cpumask);
1299 
1300 	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1301 	process_data = setup_private_data(g->p.bytes_process);
1302 
1303 	if (g->p.show_details >= 3) {
1304 		printf(" # process %2d global mem: %p, process mem: %p\n",
1305 			process_nr, g->data, process_data);
1306 	}
1307 
1308 	for (t = 0; t < g->p.nr_threads; t++) {
1309 		task_nr = process_nr*g->p.nr_threads + t;
1310 		td = g->threads + task_nr;
1311 
1312 		td->process_data = process_data;
1313 		td->process_nr   = process_nr;
1314 		td->thread_nr    = t;
1315 		td->task_nr	 = task_nr;
1316 		td->val          = rand();
1317 		td->curr_cpu	 = -1;
1318 		td->process_lock = &process_lock;
1319 
1320 		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1321 		BUG_ON(ret);
1322 	}
1323 
1324 	for (t = 0; t < g->p.nr_threads; t++) {
1325                 ret = pthread_join(pthreads[t], NULL);
1326 		BUG_ON(ret);
1327 	}
1328 
1329 	free_data(process_data, g->p.bytes_process);
1330 	free(pthreads);
1331 }
1332 
1333 static void print_summary(void)
1334 {
1335 	if (g->p.show_details < 0)
1336 		return;
1337 
1338 	printf("\n ###\n");
1339 	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1340 		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1341 	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1342 			g->p.nr_loops, g->p.bytes_global/1024/1024);
1343 	printf(" #      %5dx %5ldMB process shared mem operations\n",
1344 			g->p.nr_loops, g->p.bytes_process/1024/1024);
1345 	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1346 			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1347 
1348 	printf(" ###\n");
1349 
1350 	printf("\n ###\n"); fflush(stdout);
1351 }
1352 
1353 static void init_thread_data(void)
1354 {
1355 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1356 	int t;
1357 
1358 	g->threads = zalloc_shared_data(size);
1359 
1360 	for (t = 0; t < g->p.nr_tasks; t++) {
1361 		struct thread_data *td = g->threads + t;
1362 		int cpu;
1363 
1364 		/* Allow all nodes by default: */
1365 		td->bind_node = -1;
1366 
1367 		/* Allow all CPUs by default: */
1368 		CPU_ZERO(&td->bind_cpumask);
1369 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1370 			CPU_SET(cpu, &td->bind_cpumask);
1371 	}
1372 }
1373 
1374 static void deinit_thread_data(void)
1375 {
1376 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1377 
1378 	free_data(g->threads, size);
1379 }
1380 
1381 static int init(void)
1382 {
1383 	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1384 
1385 	/* Copy over options: */
1386 	g->p = p0;
1387 
1388 	g->p.nr_cpus = numa_num_configured_cpus();
1389 
1390 	g->p.nr_nodes = numa_max_node() + 1;
1391 
1392 	/* char array in count_process_nodes(): */
1393 	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1394 
1395 	if (g->p.show_quiet && !g->p.show_details)
1396 		g->p.show_details = -1;
1397 
1398 	/* Some memory should be specified: */
1399 	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1400 		return -1;
1401 
1402 	if (g->p.mb_global_str) {
1403 		g->p.mb_global = atof(g->p.mb_global_str);
1404 		BUG_ON(g->p.mb_global < 0);
1405 	}
1406 
1407 	if (g->p.mb_proc_str) {
1408 		g->p.mb_proc = atof(g->p.mb_proc_str);
1409 		BUG_ON(g->p.mb_proc < 0);
1410 	}
1411 
1412 	if (g->p.mb_proc_locked_str) {
1413 		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1414 		BUG_ON(g->p.mb_proc_locked < 0);
1415 		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1416 	}
1417 
1418 	if (g->p.mb_thread_str) {
1419 		g->p.mb_thread = atof(g->p.mb_thread_str);
1420 		BUG_ON(g->p.mb_thread < 0);
1421 	}
1422 
1423 	BUG_ON(g->p.nr_threads <= 0);
1424 	BUG_ON(g->p.nr_proc <= 0);
1425 
1426 	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1427 
1428 	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1429 	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1430 	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1431 	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1432 
1433 	g->data = setup_shared_data(g->p.bytes_global);
1434 
1435 	/* Startup serialization: */
1436 	init_global_mutex(&g->start_work_mutex);
1437 	init_global_mutex(&g->startup_mutex);
1438 	init_global_mutex(&g->startup_done_mutex);
1439 	init_global_mutex(&g->stop_work_mutex);
1440 
1441 	init_thread_data();
1442 
1443 	tprintf("#\n");
1444 	if (parse_setup_cpu_list() || parse_setup_node_list())
1445 		return -1;
1446 	tprintf("#\n");
1447 
1448 	print_summary();
1449 
1450 	return 0;
1451 }
1452 
1453 static void deinit(void)
1454 {
1455 	free_data(g->data, g->p.bytes_global);
1456 	g->data = NULL;
1457 
1458 	deinit_thread_data();
1459 
1460 	free_data(g, sizeof(*g));
1461 	g = NULL;
1462 }
1463 
1464 /*
1465  * Print a short or long result, depending on the verbosity setting:
1466  */
1467 static void print_res(const char *name, double val,
1468 		      const char *txt_unit, const char *txt_short, const char *txt_long)
1469 {
1470 	if (!name)
1471 		name = "main,";
1472 
1473 	if (!g->p.show_quiet)
1474 		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1475 	else
1476 		printf(" %14.3f %s\n", val, txt_long);
1477 }
1478 
1479 static int __bench_numa(const char *name)
1480 {
1481 	struct timeval start, stop, diff;
1482 	u64 runtime_ns_min, runtime_ns_sum;
1483 	pid_t *pids, pid, wpid;
1484 	double delta_runtime;
1485 	double runtime_avg;
1486 	double runtime_sec_max;
1487 	double runtime_sec_min;
1488 	int wait_stat;
1489 	double bytes;
1490 	int i, t, p;
1491 
1492 	if (init())
1493 		return -1;
1494 
1495 	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1496 	pid = -1;
1497 
1498 	/* All threads try to acquire it, this way we can wait for them to start up: */
1499 	pthread_mutex_lock(&g->start_work_mutex);
1500 
1501 	if (g->p.serialize_startup) {
1502 		tprintf(" #\n");
1503 		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1504 	}
1505 
1506 	gettimeofday(&start, NULL);
1507 
1508 	for (i = 0; i < g->p.nr_proc; i++) {
1509 		pid = fork();
1510 		dprintf(" # process %2d: PID %d\n", i, pid);
1511 
1512 		BUG_ON(pid < 0);
1513 		if (!pid) {
1514 			/* Child process: */
1515 			worker_process(i);
1516 
1517 			exit(0);
1518 		}
1519 		pids[i] = pid;
1520 
1521 	}
1522 	/* Wait for all the threads to start up: */
1523 	while (g->nr_tasks_started != g->p.nr_tasks)
1524 		usleep(USEC_PER_MSEC);
1525 
1526 	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1527 
1528 	if (g->p.serialize_startup) {
1529 		double startup_sec;
1530 
1531 		pthread_mutex_lock(&g->startup_done_mutex);
1532 
1533 		/* This will start all threads: */
1534 		pthread_mutex_unlock(&g->start_work_mutex);
1535 
1536 		/* This mutex is locked - the last started thread will wake us: */
1537 		pthread_mutex_lock(&g->startup_done_mutex);
1538 
1539 		gettimeofday(&stop, NULL);
1540 
1541 		timersub(&stop, &start, &diff);
1542 
1543 		startup_sec = diff.tv_sec * NSEC_PER_SEC;
1544 		startup_sec += diff.tv_usec * NSEC_PER_USEC;
1545 		startup_sec /= NSEC_PER_SEC;
1546 
1547 		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1548 		tprintf(" #\n");
1549 
1550 		start = stop;
1551 		pthread_mutex_unlock(&g->startup_done_mutex);
1552 	} else {
1553 		gettimeofday(&start, NULL);
1554 	}
1555 
1556 	/* Parent process: */
1557 
1558 
1559 	for (i = 0; i < g->p.nr_proc; i++) {
1560 		wpid = waitpid(pids[i], &wait_stat, 0);
1561 		BUG_ON(wpid < 0);
1562 		BUG_ON(!WIFEXITED(wait_stat));
1563 
1564 	}
1565 
1566 	runtime_ns_sum = 0;
1567 	runtime_ns_min = -1LL;
1568 
1569 	for (t = 0; t < g->p.nr_tasks; t++) {
1570 		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1571 
1572 		runtime_ns_sum += thread_runtime_ns;
1573 		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1574 	}
1575 
1576 	gettimeofday(&stop, NULL);
1577 	timersub(&stop, &start, &diff);
1578 
1579 	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1580 
1581 	tprintf("\n ###\n");
1582 	tprintf("\n");
1583 
1584 	runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1585 	runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1586 	runtime_sec_max /= NSEC_PER_SEC;
1587 
1588 	runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1589 
1590 	bytes = g->bytes_done;
1591 	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1592 
1593 	if (g->p.measure_convergence) {
1594 		print_res(name, runtime_sec_max,
1595 			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1596 	}
1597 
1598 	print_res(name, runtime_sec_max,
1599 		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1600 
1601 	print_res(name, runtime_sec_min,
1602 		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1603 
1604 	print_res(name, runtime_avg,
1605 		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1606 
1607 	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1608 	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1609 		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1610 
1611 	print_res(name, bytes / g->p.nr_tasks / 1e9,
1612 		"GB,", "data/thread",		"GB data processed, per thread");
1613 
1614 	print_res(name, bytes / 1e9,
1615 		"GB,", "data-total",		"GB data processed, total");
1616 
1617 	print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1618 		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1619 
1620 	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1621 		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1622 
1623 	print_res(name, bytes / runtime_sec_max / 1e9,
1624 		"GB/sec,", "total-speed",	"GB/sec total speed");
1625 
1626 	if (g->p.show_details >= 2) {
1627 		char tname[14 + 2 * 10 + 1];
1628 		struct thread_data *td;
1629 		for (p = 0; p < g->p.nr_proc; p++) {
1630 			for (t = 0; t < g->p.nr_threads; t++) {
1631 				memset(tname, 0, sizeof(tname));
1632 				td = g->threads + p*g->p.nr_threads + t;
1633 				snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1634 				print_res(tname, td->speed_gbs,
1635 					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1636 				print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1637 					"secs",	"thread-system-time", "system CPU time/thread");
1638 				print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1639 					"secs",	"thread-user-time", "user CPU time/thread");
1640 			}
1641 		}
1642 	}
1643 
1644 	free(pids);
1645 
1646 	deinit();
1647 
1648 	return 0;
1649 }
1650 
1651 #define MAX_ARGS 50
1652 
1653 static int command_size(const char **argv)
1654 {
1655 	int size = 0;
1656 
1657 	while (*argv) {
1658 		size++;
1659 		argv++;
1660 	}
1661 
1662 	BUG_ON(size >= MAX_ARGS);
1663 
1664 	return size;
1665 }
1666 
1667 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1668 {
1669 	int i;
1670 
1671 	printf("\n # Running %s \"perf bench numa", name);
1672 
1673 	for (i = 0; i < argc; i++)
1674 		printf(" %s", argv[i]);
1675 
1676 	printf("\"\n");
1677 
1678 	memset(p, 0, sizeof(*p));
1679 
1680 	/* Initialize nonzero defaults: */
1681 
1682 	p->serialize_startup		= 1;
1683 	p->data_reads			= true;
1684 	p->data_writes			= true;
1685 	p->data_backwards		= true;
1686 	p->data_rand_walk		= true;
1687 	p->nr_loops			= -1;
1688 	p->init_random			= true;
1689 	p->mb_global_str		= "1";
1690 	p->nr_proc			= 1;
1691 	p->nr_threads			= 1;
1692 	p->nr_secs			= 5;
1693 	p->run_all			= argc == 1;
1694 }
1695 
1696 static int run_bench_numa(const char *name, const char **argv)
1697 {
1698 	int argc = command_size(argv);
1699 
1700 	init_params(&p0, name, argc, argv);
1701 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1702 	if (argc)
1703 		goto err;
1704 
1705 	if (__bench_numa(name))
1706 		goto err;
1707 
1708 	return 0;
1709 
1710 err:
1711 	return -1;
1712 }
1713 
1714 #define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1715 #define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1716 
1717 #define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1718 #define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1719 
1720 #define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1721 #define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1722 
1723 /*
1724  * The built-in test-suite executed by "perf bench numa -a".
1725  *
1726  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1727  */
1728 static const char *tests[][MAX_ARGS] = {
1729    /* Basic single-stream NUMA bandwidth measurements: */
1730    { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1731 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1732    { "RAM-bw-local-NOTHP,",
1733 			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1734 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1735    { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1736 			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1737 
1738    /* 2-stream NUMA bandwidth measurements: */
1739    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1740 			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1741    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1742 		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1743 
1744    /* Cross-stream NUMA bandwidth measurement: */
1745    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1746 		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1747 
1748    /* Convergence latency measurements: */
1749    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1750    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1751    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1752    { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1753    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1754    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1755    { " 4x4-convergence-NOTHP,",
1756 			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1757    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1758    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1759    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1760    { " 8x4-convergence-NOTHP,",
1761 			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1762    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1763    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1764    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1765    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1766    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1767 
1768    /* Various NUMA process/thread layout bandwidth measurements: */
1769    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1770    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1771    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1772    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1773    { " 8x1-bw-process-NOTHP,",
1774 			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1775    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1776 
1777    { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1778    { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1779    { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1780    { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1781 
1782    { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1783    { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1784    { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1785    { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1786    { " 4x8-bw-thread-NOTHP,",
1787 			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1788    { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1789    { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1790 
1791    { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1792    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1793 
1794    { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1795    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1796    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1797    { "numa01-bw-thread-NOTHP,",
1798 			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1799 };
1800 
1801 static int bench_all(void)
1802 {
1803 	int nr = ARRAY_SIZE(tests);
1804 	int ret;
1805 	int i;
1806 
1807 	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1808 	BUG_ON(ret < 0);
1809 
1810 	for (i = 0; i < nr; i++) {
1811 		run_bench_numa(tests[i][0], tests[i] + 1);
1812 	}
1813 
1814 	printf("\n");
1815 
1816 	return 0;
1817 }
1818 
1819 int bench_numa(int argc, const char **argv)
1820 {
1821 	init_params(&p0, "main,", argc, argv);
1822 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1823 	if (argc)
1824 		goto err;
1825 
1826 	if (p0.run_all)
1827 		return bench_all();
1828 
1829 	if (__bench_numa(NULL))
1830 		goto err;
1831 
1832 	return 0;
1833 
1834 err:
1835 	usage_with_options(numa_usage, options);
1836 	return -1;
1837 }
1838