xref: /openbmc/linux/tools/perf/bench/numa.c (revision e65e175b)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * numa.c
4  *
5  * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
6  */
7 
8 #include <inttypes.h>
9 
10 #include <subcmd/parse-options.h>
11 #include "../util/cloexec.h"
12 
13 #include "bench.h"
14 
15 #include <errno.h>
16 #include <sched.h>
17 #include <stdio.h>
18 #include <assert.h>
19 #include <debug.h>
20 #include <malloc.h>
21 #include <signal.h>
22 #include <stdlib.h>
23 #include <string.h>
24 #include <unistd.h>
25 #include <sys/mman.h>
26 #include <sys/time.h>
27 #include <sys/resource.h>
28 #include <sys/wait.h>
29 #include <sys/prctl.h>
30 #include <sys/types.h>
31 #include <linux/kernel.h>
32 #include <linux/time64.h>
33 #include <linux/numa.h>
34 #include <linux/zalloc.h>
35 
36 #include "../util/header.h"
37 #include "../util/mutex.h"
38 #include <numa.h>
39 #include <numaif.h>
40 
41 #ifndef RUSAGE_THREAD
42 # define RUSAGE_THREAD 1
43 #endif
44 
45 /*
46  * Regular printout to the terminal, suppressed if -q is specified:
47  */
48 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
49 
50 /*
51  * Debug printf:
52  */
53 #undef dprintf
54 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
55 
56 struct thread_data {
57 	int			curr_cpu;
58 	cpu_set_t		*bind_cpumask;
59 	int			bind_node;
60 	u8			*process_data;
61 	int			process_nr;
62 	int			thread_nr;
63 	int			task_nr;
64 	unsigned int		loops_done;
65 	u64			val;
66 	u64			runtime_ns;
67 	u64			system_time_ns;
68 	u64			user_time_ns;
69 	double			speed_gbs;
70 	struct mutex		*process_lock;
71 };
72 
73 /* Parameters set by options: */
74 
75 struct params {
76 	/* Startup synchronization: */
77 	bool			serialize_startup;
78 
79 	/* Task hierarchy: */
80 	int			nr_proc;
81 	int			nr_threads;
82 
83 	/* Working set sizes: */
84 	const char		*mb_global_str;
85 	const char		*mb_proc_str;
86 	const char		*mb_proc_locked_str;
87 	const char		*mb_thread_str;
88 
89 	double			mb_global;
90 	double			mb_proc;
91 	double			mb_proc_locked;
92 	double			mb_thread;
93 
94 	/* Access patterns to the working set: */
95 	bool			data_reads;
96 	bool			data_writes;
97 	bool			data_backwards;
98 	bool			data_zero_memset;
99 	bool			data_rand_walk;
100 	u32			nr_loops;
101 	u32			nr_secs;
102 	u32			sleep_usecs;
103 
104 	/* Working set initialization: */
105 	bool			init_zero;
106 	bool			init_random;
107 	bool			init_cpu0;
108 
109 	/* Misc options: */
110 	int			show_details;
111 	int			run_all;
112 	int			thp;
113 
114 	long			bytes_global;
115 	long			bytes_process;
116 	long			bytes_process_locked;
117 	long			bytes_thread;
118 
119 	int			nr_tasks;
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 	struct mutex		startup_mutex;
140 	struct cond		startup_cond;
141 	int			nr_tasks_started;
142 
143 	struct mutex		start_work_mutex;
144 	struct cond		start_work_cond;
145 	int			nr_tasks_working;
146 	bool			start_work;
147 
148 	struct mutex		stop_work_mutex;
149 	u64			bytes_done;
150 
151 	struct thread_data	*threads;
152 
153 	/* Convergence latency measurement: */
154 	bool			all_converged;
155 	bool			stop_work;
156 
157 	int			print_once;
158 
159 	struct params		p;
160 };
161 
162 static struct global_info	*g = NULL;
163 
164 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
165 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
166 
167 struct params p0;
168 
169 static const struct option options[] = {
170 	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
171 	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
172 
173 	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
174 	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
175 	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
176 	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
177 
178 	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run (default: unlimited)"),
179 	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run (default: 5 secs)"),
180 	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
181 
182 	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via reads (can be mixed with -W)"),
183 	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
184 	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
185 	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
186 	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
187 
188 
189 	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
190 	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
191 	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
192 	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
193 
194 	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
195 	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
196 	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
197 	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
198 		    "convergence is reached when each process (all its threads) is running on a single NUMA node."),
199 	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
200 	OPT_BOOLEAN('q', "quiet"	, &quiet,
201 		    "quiet mode (do not show any warnings or messages)"),
202 	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
203 
204 	/* Special option string parsing callbacks: */
205         OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
206 			"bind the first N tasks to these specific cpus (the rest is unbound)",
207 			parse_cpus_opt),
208         OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
209 			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
210 			parse_nodes_opt),
211 	OPT_END()
212 };
213 
214 static const char * const bench_numa_usage[] = {
215 	"perf bench numa <options>",
216 	NULL
217 };
218 
219 static const char * const numa_usage[] = {
220 	"perf bench numa mem [<options>]",
221 	NULL
222 };
223 
224 /*
225  * To get number of numa nodes present.
226  */
227 static int nr_numa_nodes(void)
228 {
229 	int i, nr_nodes = 0;
230 
231 	for (i = 0; i < g->p.nr_nodes; i++) {
232 		if (numa_bitmask_isbitset(numa_nodes_ptr, i))
233 			nr_nodes++;
234 	}
235 
236 	return nr_nodes;
237 }
238 
239 /*
240  * To check if given numa node is present.
241  */
242 static int is_node_present(int node)
243 {
244 	return numa_bitmask_isbitset(numa_nodes_ptr, node);
245 }
246 
247 /*
248  * To check given numa node has cpus.
249  */
250 static bool node_has_cpus(int node)
251 {
252 	struct bitmask *cpumask = numa_allocate_cpumask();
253 	bool ret = false; /* fall back to nocpus */
254 	int cpu;
255 
256 	BUG_ON(!cpumask);
257 	if (!numa_node_to_cpus(node, cpumask)) {
258 		for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
259 			if (numa_bitmask_isbitset(cpumask, cpu)) {
260 				ret = true;
261 				break;
262 			}
263 		}
264 	}
265 	numa_free_cpumask(cpumask);
266 
267 	return ret;
268 }
269 
270 static cpu_set_t *bind_to_cpu(int target_cpu)
271 {
272 	int nrcpus = numa_num_possible_cpus();
273 	cpu_set_t *orig_mask, *mask;
274 	size_t size;
275 
276 	orig_mask = CPU_ALLOC(nrcpus);
277 	BUG_ON(!orig_mask);
278 	size = CPU_ALLOC_SIZE(nrcpus);
279 	CPU_ZERO_S(size, orig_mask);
280 
281 	if (sched_getaffinity(0, size, orig_mask))
282 		goto err_out;
283 
284 	mask = CPU_ALLOC(nrcpus);
285 	if (!mask)
286 		goto err_out;
287 
288 	CPU_ZERO_S(size, mask);
289 
290 	if (target_cpu == -1) {
291 		int cpu;
292 
293 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
294 			CPU_SET_S(cpu, size, mask);
295 	} else {
296 		if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
297 			goto err;
298 
299 		CPU_SET_S(target_cpu, size, mask);
300 	}
301 
302 	if (sched_setaffinity(0, size, mask))
303 		goto err;
304 
305 	return orig_mask;
306 
307 err:
308 	CPU_FREE(mask);
309 err_out:
310 	CPU_FREE(orig_mask);
311 
312 	/* BUG_ON due to failure in allocation of orig_mask/mask */
313 	BUG_ON(-1);
314 	return NULL;
315 }
316 
317 static cpu_set_t *bind_to_node(int target_node)
318 {
319 	int nrcpus = numa_num_possible_cpus();
320 	size_t size;
321 	cpu_set_t *orig_mask, *mask;
322 	int cpu;
323 
324 	orig_mask = CPU_ALLOC(nrcpus);
325 	BUG_ON(!orig_mask);
326 	size = CPU_ALLOC_SIZE(nrcpus);
327 	CPU_ZERO_S(size, orig_mask);
328 
329 	if (sched_getaffinity(0, size, orig_mask))
330 		goto err_out;
331 
332 	mask = CPU_ALLOC(nrcpus);
333 	if (!mask)
334 		goto err_out;
335 
336 	CPU_ZERO_S(size, mask);
337 
338 	if (target_node == NUMA_NO_NODE) {
339 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
340 			CPU_SET_S(cpu, size, mask);
341 	} else {
342 		struct bitmask *cpumask = numa_allocate_cpumask();
343 
344 		if (!cpumask)
345 			goto err;
346 
347 		if (!numa_node_to_cpus(target_node, cpumask)) {
348 			for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
349 				if (numa_bitmask_isbitset(cpumask, cpu))
350 					CPU_SET_S(cpu, size, mask);
351 			}
352 		}
353 		numa_free_cpumask(cpumask);
354 	}
355 
356 	if (sched_setaffinity(0, size, mask))
357 		goto err;
358 
359 	return orig_mask;
360 
361 err:
362 	CPU_FREE(mask);
363 err_out:
364 	CPU_FREE(orig_mask);
365 
366 	/* BUG_ON due to failure in allocation of orig_mask/mask */
367 	BUG_ON(-1);
368 	return NULL;
369 }
370 
371 static void bind_to_cpumask(cpu_set_t *mask)
372 {
373 	int ret;
374 	size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
375 
376 	ret = sched_setaffinity(0, size, mask);
377 	if (ret) {
378 		CPU_FREE(mask);
379 		BUG_ON(ret);
380 	}
381 }
382 
383 static void mempol_restore(void)
384 {
385 	int ret;
386 
387 	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
388 
389 	BUG_ON(ret);
390 }
391 
392 static void bind_to_memnode(int node)
393 {
394 	struct bitmask *node_mask;
395 	int ret;
396 
397 	if (node == NUMA_NO_NODE)
398 		return;
399 
400 	node_mask = numa_allocate_nodemask();
401 	BUG_ON(!node_mask);
402 
403 	numa_bitmask_clearall(node_mask);
404 	numa_bitmask_setbit(node_mask, node);
405 
406 	ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
407 	dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
408 
409 	numa_bitmask_free(node_mask);
410 	BUG_ON(ret);
411 }
412 
413 #define HPSIZE (2*1024*1024)
414 
415 #define set_taskname(fmt...)				\
416 do {							\
417 	char name[20];					\
418 							\
419 	snprintf(name, 20, fmt);			\
420 	prctl(PR_SET_NAME, name);			\
421 } while (0)
422 
423 static u8 *alloc_data(ssize_t bytes0, int map_flags,
424 		      int init_zero, int init_cpu0, int thp, int init_random)
425 {
426 	cpu_set_t *orig_mask = NULL;
427 	ssize_t bytes;
428 	u8 *buf;
429 	int ret;
430 
431 	if (!bytes0)
432 		return NULL;
433 
434 	/* Allocate and initialize all memory on CPU#0: */
435 	if (init_cpu0) {
436 		int node = numa_node_of_cpu(0);
437 
438 		orig_mask = bind_to_node(node);
439 		bind_to_memnode(node);
440 	}
441 
442 	bytes = bytes0 + HPSIZE;
443 
444 	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
445 	BUG_ON(buf == (void *)-1);
446 
447 	if (map_flags == MAP_PRIVATE) {
448 		if (thp > 0) {
449 			ret = madvise(buf, bytes, MADV_HUGEPAGE);
450 			if (ret && !g->print_once) {
451 				g->print_once = 1;
452 				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
453 			}
454 		}
455 		if (thp < 0) {
456 			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
457 			if (ret && !g->print_once) {
458 				g->print_once = 1;
459 				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
460 			}
461 		}
462 	}
463 
464 	if (init_zero) {
465 		bzero(buf, bytes);
466 	} else {
467 		/* Initialize random contents, different in each word: */
468 		if (init_random) {
469 			u64 *wbuf = (void *)buf;
470 			long off = rand();
471 			long i;
472 
473 			for (i = 0; i < bytes/8; i++)
474 				wbuf[i] = i + off;
475 		}
476 	}
477 
478 	/* Align to 2MB boundary: */
479 	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
480 
481 	/* Restore affinity: */
482 	if (init_cpu0) {
483 		bind_to_cpumask(orig_mask);
484 		CPU_FREE(orig_mask);
485 		mempol_restore();
486 	}
487 
488 	return buf;
489 }
490 
491 static void free_data(void *data, ssize_t bytes)
492 {
493 	int ret;
494 
495 	if (!data)
496 		return;
497 
498 	ret = munmap(data, bytes);
499 	BUG_ON(ret);
500 }
501 
502 /*
503  * Create a shared memory buffer that can be shared between processes, zeroed:
504  */
505 static void * zalloc_shared_data(ssize_t bytes)
506 {
507 	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
508 }
509 
510 /*
511  * Create a shared memory buffer that can be shared between processes:
512  */
513 static void * setup_shared_data(ssize_t bytes)
514 {
515 	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
516 }
517 
518 /*
519  * Allocate process-local memory - this will either be shared between
520  * threads of this process, or only be accessed by this thread:
521  */
522 static void * setup_private_data(ssize_t bytes)
523 {
524 	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
525 }
526 
527 static int parse_cpu_list(const char *arg)
528 {
529 	p0.cpu_list_str = strdup(arg);
530 
531 	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
532 
533 	return 0;
534 }
535 
536 static int parse_setup_cpu_list(void)
537 {
538 	struct thread_data *td;
539 	char *str0, *str;
540 	int t;
541 
542 	if (!g->p.cpu_list_str)
543 		return 0;
544 
545 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
546 
547 	str0 = str = strdup(g->p.cpu_list_str);
548 	t = 0;
549 
550 	BUG_ON(!str);
551 
552 	tprintf("# binding tasks to CPUs:\n");
553 	tprintf("#  ");
554 
555 	while (true) {
556 		int bind_cpu, bind_cpu_0, bind_cpu_1;
557 		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
558 		int bind_len;
559 		int step;
560 		int mul;
561 
562 		tok = strsep(&str, ",");
563 		if (!tok)
564 			break;
565 
566 		tok_end = strstr(tok, "-");
567 
568 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
569 		if (!tok_end) {
570 			/* Single CPU specified: */
571 			bind_cpu_0 = bind_cpu_1 = atol(tok);
572 		} else {
573 			/* CPU range specified (for example: "5-11"): */
574 			bind_cpu_0 = atol(tok);
575 			bind_cpu_1 = atol(tok_end + 1);
576 		}
577 
578 		step = 1;
579 		tok_step = strstr(tok, "#");
580 		if (tok_step) {
581 			step = atol(tok_step + 1);
582 			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
583 		}
584 
585 		/*
586 		 * Mask length.
587 		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
588 		 * where the _4 means the next 4 CPUs are allowed.
589 		 */
590 		bind_len = 1;
591 		tok_len = strstr(tok, "_");
592 		if (tok_len) {
593 			bind_len = atol(tok_len + 1);
594 			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
595 		}
596 
597 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
598 		mul = 1;
599 		tok_mul = strstr(tok, "x");
600 		if (tok_mul) {
601 			mul = atol(tok_mul + 1);
602 			BUG_ON(mul <= 0);
603 		}
604 
605 		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
606 
607 		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
608 			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
609 			return -1;
610 		}
611 
612 		if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
613 			printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
614 			return -1;
615 		}
616 
617 		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
618 		BUG_ON(bind_cpu_0 > bind_cpu_1);
619 
620 		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
621 			size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
622 			int i;
623 
624 			for (i = 0; i < mul; i++) {
625 				int cpu;
626 
627 				if (t >= g->p.nr_tasks) {
628 					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
629 					goto out;
630 				}
631 				td = g->threads + t;
632 
633 				if (t)
634 					tprintf(",");
635 				if (bind_len > 1) {
636 					tprintf("%2d/%d", bind_cpu, bind_len);
637 				} else {
638 					tprintf("%2d", bind_cpu);
639 				}
640 
641 				td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
642 				BUG_ON(!td->bind_cpumask);
643 				CPU_ZERO_S(size, td->bind_cpumask);
644 				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
645 					if (cpu < 0 || cpu >= g->p.nr_cpus) {
646 						CPU_FREE(td->bind_cpumask);
647 						BUG_ON(-1);
648 					}
649 					CPU_SET_S(cpu, size, td->bind_cpumask);
650 				}
651 				t++;
652 			}
653 		}
654 	}
655 out:
656 
657 	tprintf("\n");
658 
659 	if (t < g->p.nr_tasks)
660 		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
661 
662 	free(str0);
663 	return 0;
664 }
665 
666 static int parse_cpus_opt(const struct option *opt __maybe_unused,
667 			  const char *arg, int unset __maybe_unused)
668 {
669 	if (!arg)
670 		return -1;
671 
672 	return parse_cpu_list(arg);
673 }
674 
675 static int parse_node_list(const char *arg)
676 {
677 	p0.node_list_str = strdup(arg);
678 
679 	dprintf("got NODE list: {%s}\n", p0.node_list_str);
680 
681 	return 0;
682 }
683 
684 static int parse_setup_node_list(void)
685 {
686 	struct thread_data *td;
687 	char *str0, *str;
688 	int t;
689 
690 	if (!g->p.node_list_str)
691 		return 0;
692 
693 	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
694 
695 	str0 = str = strdup(g->p.node_list_str);
696 	t = 0;
697 
698 	BUG_ON(!str);
699 
700 	tprintf("# binding tasks to NODEs:\n");
701 	tprintf("# ");
702 
703 	while (true) {
704 		int bind_node, bind_node_0, bind_node_1;
705 		char *tok, *tok_end, *tok_step, *tok_mul;
706 		int step;
707 		int mul;
708 
709 		tok = strsep(&str, ",");
710 		if (!tok)
711 			break;
712 
713 		tok_end = strstr(tok, "-");
714 
715 		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
716 		if (!tok_end) {
717 			/* Single NODE specified: */
718 			bind_node_0 = bind_node_1 = atol(tok);
719 		} else {
720 			/* NODE range specified (for example: "5-11"): */
721 			bind_node_0 = atol(tok);
722 			bind_node_1 = atol(tok_end + 1);
723 		}
724 
725 		step = 1;
726 		tok_step = strstr(tok, "#");
727 		if (tok_step) {
728 			step = atol(tok_step + 1);
729 			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
730 		}
731 
732 		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
733 		mul = 1;
734 		tok_mul = strstr(tok, "x");
735 		if (tok_mul) {
736 			mul = atol(tok_mul + 1);
737 			BUG_ON(mul <= 0);
738 		}
739 
740 		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
741 
742 		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
743 			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
744 			return -1;
745 		}
746 
747 		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
748 		BUG_ON(bind_node_0 > bind_node_1);
749 
750 		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
751 			int i;
752 
753 			for (i = 0; i < mul; i++) {
754 				if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
755 					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
756 					goto out;
757 				}
758 				td = g->threads + t;
759 
760 				if (!t)
761 					tprintf(" %2d", bind_node);
762 				else
763 					tprintf(",%2d", bind_node);
764 
765 				td->bind_node = bind_node;
766 				t++;
767 			}
768 		}
769 	}
770 out:
771 
772 	tprintf("\n");
773 
774 	if (t < g->p.nr_tasks)
775 		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
776 
777 	free(str0);
778 	return 0;
779 }
780 
781 static int parse_nodes_opt(const struct option *opt __maybe_unused,
782 			  const char *arg, int unset __maybe_unused)
783 {
784 	if (!arg)
785 		return -1;
786 
787 	return parse_node_list(arg);
788 }
789 
790 static inline uint32_t lfsr_32(uint32_t lfsr)
791 {
792 	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
793 	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
794 }
795 
796 /*
797  * Make sure there's real data dependency to RAM (when read
798  * accesses are enabled), so the compiler, the CPU and the
799  * kernel (KSM, zero page, etc.) cannot optimize away RAM
800  * accesses:
801  */
802 static inline u64 access_data(u64 *data, u64 val)
803 {
804 	if (g->p.data_reads)
805 		val += *data;
806 	if (g->p.data_writes)
807 		*data = val + 1;
808 	return val;
809 }
810 
811 /*
812  * The worker process does two types of work, a forwards going
813  * loop and a backwards going loop.
814  *
815  * We do this so that on multiprocessor systems we do not create
816  * a 'train' of processing, with highly synchronized processes,
817  * skewing the whole benchmark.
818  */
819 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
820 {
821 	long words = bytes/sizeof(u64);
822 	u64 *data = (void *)__data;
823 	long chunk_0, chunk_1;
824 	u64 *d0, *d, *d1;
825 	long off;
826 	long i;
827 
828 	BUG_ON(!data && words);
829 	BUG_ON(data && !words);
830 
831 	if (!data)
832 		return val;
833 
834 	/* Very simple memset() work variant: */
835 	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
836 		bzero(data, bytes);
837 		return val;
838 	}
839 
840 	/* Spread out by PID/TID nr and by loop nr: */
841 	chunk_0 = words/nr_max;
842 	chunk_1 = words/g->p.nr_loops;
843 	off = nr*chunk_0 + loop*chunk_1;
844 
845 	while (off >= words)
846 		off -= words;
847 
848 	if (g->p.data_rand_walk) {
849 		u32 lfsr = nr + loop + val;
850 		int j;
851 
852 		for (i = 0; i < words/1024; i++) {
853 			long start, end;
854 
855 			lfsr = lfsr_32(lfsr);
856 
857 			start = lfsr % words;
858 			end = min(start + 1024, words-1);
859 
860 			if (g->p.data_zero_memset) {
861 				bzero(data + start, (end-start) * sizeof(u64));
862 			} else {
863 				for (j = start; j < end; j++)
864 					val = access_data(data + j, val);
865 			}
866 		}
867 	} else if (!g->p.data_backwards || (nr + loop) & 1) {
868 		/* Process data forwards: */
869 
870 		d0 = data + off;
871 		d  = data + off + 1;
872 		d1 = data + words;
873 
874 		for (;;) {
875 			if (unlikely(d >= d1))
876 				d = data;
877 			if (unlikely(d == d0))
878 				break;
879 
880 			val = access_data(d, val);
881 
882 			d++;
883 		}
884 	} else {
885 		/* Process data backwards: */
886 
887 		d0 = data + off;
888 		d  = data + off - 1;
889 		d1 = data + words;
890 
891 		for (;;) {
892 			if (unlikely(d < data))
893 				d = data + words-1;
894 			if (unlikely(d == d0))
895 				break;
896 
897 			val = access_data(d, val);
898 
899 			d--;
900 		}
901 	}
902 
903 	return val;
904 }
905 
906 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
907 {
908 	unsigned int cpu;
909 
910 	cpu = sched_getcpu();
911 
912 	g->threads[task_nr].curr_cpu = cpu;
913 	prctl(0, bytes_worked);
914 }
915 
916 /*
917  * Count the number of nodes a process's threads
918  * are spread out on.
919  *
920  * A count of 1 means that the process is compressed
921  * to a single node. A count of g->p.nr_nodes means it's
922  * spread out on the whole system.
923  */
924 static int count_process_nodes(int process_nr)
925 {
926 	char *node_present;
927 	int nodes;
928 	int n, t;
929 
930 	node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
931 	BUG_ON(!node_present);
932 	for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
933 		node_present[nodes] = 0;
934 
935 	for (t = 0; t < g->p.nr_threads; t++) {
936 		struct thread_data *td;
937 		int task_nr;
938 		int node;
939 
940 		task_nr = process_nr*g->p.nr_threads + t;
941 		td = g->threads + task_nr;
942 
943 		node = numa_node_of_cpu(td->curr_cpu);
944 		if (node < 0) /* curr_cpu was likely still -1 */ {
945 			free(node_present);
946 			return 0;
947 		}
948 
949 		node_present[node] = 1;
950 	}
951 
952 	nodes = 0;
953 
954 	for (n = 0; n < g->p.nr_nodes; n++)
955 		nodes += node_present[n];
956 
957 	free(node_present);
958 	return nodes;
959 }
960 
961 /*
962  * Count the number of distinct process-threads a node contains.
963  *
964  * A count of 1 means that the node contains only a single
965  * process. If all nodes on the system contain at most one
966  * process then we are well-converged.
967  */
968 static int count_node_processes(int node)
969 {
970 	int processes = 0;
971 	int t, p;
972 
973 	for (p = 0; p < g->p.nr_proc; p++) {
974 		for (t = 0; t < g->p.nr_threads; t++) {
975 			struct thread_data *td;
976 			int task_nr;
977 			int n;
978 
979 			task_nr = p*g->p.nr_threads + t;
980 			td = g->threads + task_nr;
981 
982 			n = numa_node_of_cpu(td->curr_cpu);
983 			if (n == node) {
984 				processes++;
985 				break;
986 			}
987 		}
988 	}
989 
990 	return processes;
991 }
992 
993 static void calc_convergence_compression(int *strong)
994 {
995 	unsigned int nodes_min, nodes_max;
996 	int p;
997 
998 	nodes_min = -1;
999 	nodes_max =  0;
1000 
1001 	for (p = 0; p < g->p.nr_proc; p++) {
1002 		unsigned int nodes = count_process_nodes(p);
1003 
1004 		if (!nodes) {
1005 			*strong = 0;
1006 			return;
1007 		}
1008 
1009 		nodes_min = min(nodes, nodes_min);
1010 		nodes_max = max(nodes, nodes_max);
1011 	}
1012 
1013 	/* Strong convergence: all threads compress on a single node: */
1014 	if (nodes_min == 1 && nodes_max == 1) {
1015 		*strong = 1;
1016 	} else {
1017 		*strong = 0;
1018 		tprintf(" {%d-%d}", nodes_min, nodes_max);
1019 	}
1020 }
1021 
1022 static void calc_convergence(double runtime_ns_max, double *convergence)
1023 {
1024 	unsigned int loops_done_min, loops_done_max;
1025 	int process_groups;
1026 	int *nodes;
1027 	int distance;
1028 	int nr_min;
1029 	int nr_max;
1030 	int strong;
1031 	int sum;
1032 	int nr;
1033 	int node;
1034 	int cpu;
1035 	int t;
1036 
1037 	if (!g->p.show_convergence && !g->p.measure_convergence)
1038 		return;
1039 
1040 	nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1041 	BUG_ON(!nodes);
1042 	for (node = 0; node < g->p.nr_nodes; node++)
1043 		nodes[node] = 0;
1044 
1045 	loops_done_min = -1;
1046 	loops_done_max = 0;
1047 
1048 	for (t = 0; t < g->p.nr_tasks; t++) {
1049 		struct thread_data *td = g->threads + t;
1050 		unsigned int loops_done;
1051 
1052 		cpu = td->curr_cpu;
1053 
1054 		/* Not all threads have written it yet: */
1055 		if (cpu < 0)
1056 			continue;
1057 
1058 		node = numa_node_of_cpu(cpu);
1059 
1060 		nodes[node]++;
1061 
1062 		loops_done = td->loops_done;
1063 		loops_done_min = min(loops_done, loops_done_min);
1064 		loops_done_max = max(loops_done, loops_done_max);
1065 	}
1066 
1067 	nr_max = 0;
1068 	nr_min = g->p.nr_tasks;
1069 	sum = 0;
1070 
1071 	for (node = 0; node < g->p.nr_nodes; node++) {
1072 		if (!is_node_present(node))
1073 			continue;
1074 		nr = nodes[node];
1075 		nr_min = min(nr, nr_min);
1076 		nr_max = max(nr, nr_max);
1077 		sum += nr;
1078 	}
1079 	BUG_ON(nr_min > nr_max);
1080 
1081 	BUG_ON(sum > g->p.nr_tasks);
1082 
1083 	if (0 && (sum < g->p.nr_tasks)) {
1084 		free(nodes);
1085 		return;
1086 	}
1087 
1088 	/*
1089 	 * Count the number of distinct process groups present
1090 	 * on nodes - when we are converged this will decrease
1091 	 * to g->p.nr_proc:
1092 	 */
1093 	process_groups = 0;
1094 
1095 	for (node = 0; node < g->p.nr_nodes; node++) {
1096 		int processes;
1097 
1098 		if (!is_node_present(node))
1099 			continue;
1100 		processes = count_node_processes(node);
1101 		nr = nodes[node];
1102 		tprintf(" %2d/%-2d", nr, processes);
1103 
1104 		process_groups += processes;
1105 	}
1106 
1107 	distance = nr_max - nr_min;
1108 
1109 	tprintf(" [%2d/%-2d]", distance, process_groups);
1110 
1111 	tprintf(" l:%3d-%-3d (%3d)",
1112 		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1113 
1114 	if (loops_done_min && loops_done_max) {
1115 		double skew = 1.0 - (double)loops_done_min/loops_done_max;
1116 
1117 		tprintf(" [%4.1f%%]", skew * 100.0);
1118 	}
1119 
1120 	calc_convergence_compression(&strong);
1121 
1122 	if (strong && process_groups == g->p.nr_proc) {
1123 		if (!*convergence) {
1124 			*convergence = runtime_ns_max;
1125 			tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1126 			if (g->p.measure_convergence) {
1127 				g->all_converged = true;
1128 				g->stop_work = true;
1129 			}
1130 		}
1131 	} else {
1132 		if (*convergence) {
1133 			tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1134 			*convergence = 0;
1135 		}
1136 		tprintf("\n");
1137 	}
1138 
1139 	free(nodes);
1140 }
1141 
1142 static void show_summary(double runtime_ns_max, int l, double *convergence)
1143 {
1144 	tprintf("\r #  %5.1f%%  [%.1f mins]",
1145 		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1146 
1147 	calc_convergence(runtime_ns_max, convergence);
1148 
1149 	if (g->p.show_details >= 0)
1150 		fflush(stdout);
1151 }
1152 
1153 static void *worker_thread(void *__tdata)
1154 {
1155 	struct thread_data *td = __tdata;
1156 	struct timeval start0, start, stop, diff;
1157 	int process_nr = td->process_nr;
1158 	int thread_nr = td->thread_nr;
1159 	unsigned long last_perturbance;
1160 	int task_nr = td->task_nr;
1161 	int details = g->p.show_details;
1162 	int first_task, last_task;
1163 	double convergence = 0;
1164 	u64 val = td->val;
1165 	double runtime_ns_max;
1166 	u8 *global_data;
1167 	u8 *process_data;
1168 	u8 *thread_data;
1169 	u64 bytes_done, secs;
1170 	long work_done;
1171 	u32 l;
1172 	struct rusage rusage;
1173 
1174 	bind_to_cpumask(td->bind_cpumask);
1175 	bind_to_memnode(td->bind_node);
1176 
1177 	set_taskname("thread %d/%d", process_nr, thread_nr);
1178 
1179 	global_data = g->data;
1180 	process_data = td->process_data;
1181 	thread_data = setup_private_data(g->p.bytes_thread);
1182 
1183 	bytes_done = 0;
1184 
1185 	last_task = 0;
1186 	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1187 		last_task = 1;
1188 
1189 	first_task = 0;
1190 	if (process_nr == 0 && thread_nr == 0)
1191 		first_task = 1;
1192 
1193 	if (details >= 2) {
1194 		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1195 			process_nr, thread_nr, global_data, process_data, thread_data);
1196 	}
1197 
1198 	if (g->p.serialize_startup) {
1199 		mutex_lock(&g->startup_mutex);
1200 		g->nr_tasks_started++;
1201 		/* The last thread wakes the main process. */
1202 		if (g->nr_tasks_started == g->p.nr_tasks)
1203 			cond_signal(&g->startup_cond);
1204 
1205 		mutex_unlock(&g->startup_mutex);
1206 
1207 		/* Here we will wait for the main process to start us all at once: */
1208 		mutex_lock(&g->start_work_mutex);
1209 		g->start_work = false;
1210 		g->nr_tasks_working++;
1211 		while (!g->start_work)
1212 			cond_wait(&g->start_work_cond, &g->start_work_mutex);
1213 
1214 		mutex_unlock(&g->start_work_mutex);
1215 	}
1216 
1217 	gettimeofday(&start0, NULL);
1218 
1219 	start = stop = start0;
1220 	last_perturbance = start.tv_sec;
1221 
1222 	for (l = 0; l < g->p.nr_loops; l++) {
1223 		start = stop;
1224 
1225 		if (g->stop_work)
1226 			break;
1227 
1228 		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1229 		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1230 		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1231 
1232 		if (g->p.sleep_usecs) {
1233 			mutex_lock(td->process_lock);
1234 			usleep(g->p.sleep_usecs);
1235 			mutex_unlock(td->process_lock);
1236 		}
1237 		/*
1238 		 * Amount of work to be done under a process-global lock:
1239 		 */
1240 		if (g->p.bytes_process_locked) {
1241 			mutex_lock(td->process_lock);
1242 			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1243 			mutex_unlock(td->process_lock);
1244 		}
1245 
1246 		work_done = g->p.bytes_global + g->p.bytes_process +
1247 			    g->p.bytes_process_locked + g->p.bytes_thread;
1248 
1249 		update_curr_cpu(task_nr, work_done);
1250 		bytes_done += work_done;
1251 
1252 		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1253 			continue;
1254 
1255 		td->loops_done = l;
1256 
1257 		gettimeofday(&stop, NULL);
1258 
1259 		/* Check whether our max runtime timed out: */
1260 		if (g->p.nr_secs) {
1261 			timersub(&stop, &start0, &diff);
1262 			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1263 				g->stop_work = true;
1264 				break;
1265 			}
1266 		}
1267 
1268 		/* Update the summary at most once per second: */
1269 		if (start.tv_sec == stop.tv_sec)
1270 			continue;
1271 
1272 		/*
1273 		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1274 		 * by migrating to CPU#0:
1275 		 */
1276 		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1277 			cpu_set_t *orig_mask;
1278 			int target_cpu;
1279 			int this_cpu;
1280 
1281 			last_perturbance = stop.tv_sec;
1282 
1283 			/*
1284 			 * Depending on where we are running, move into
1285 			 * the other half of the system, to create some
1286 			 * real disturbance:
1287 			 */
1288 			this_cpu = g->threads[task_nr].curr_cpu;
1289 			if (this_cpu < g->p.nr_cpus/2)
1290 				target_cpu = g->p.nr_cpus-1;
1291 			else
1292 				target_cpu = 0;
1293 
1294 			orig_mask = bind_to_cpu(target_cpu);
1295 
1296 			/* Here we are running on the target CPU already */
1297 			if (details >= 1)
1298 				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1299 
1300 			bind_to_cpumask(orig_mask);
1301 			CPU_FREE(orig_mask);
1302 		}
1303 
1304 		if (details >= 3) {
1305 			timersub(&stop, &start, &diff);
1306 			runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1307 			runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1308 
1309 			if (details >= 0) {
1310 				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1311 					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1312 			}
1313 			fflush(stdout);
1314 		}
1315 		if (!last_task)
1316 			continue;
1317 
1318 		timersub(&stop, &start0, &diff);
1319 		runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1320 		runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1321 
1322 		show_summary(runtime_ns_max, l, &convergence);
1323 	}
1324 
1325 	gettimeofday(&stop, NULL);
1326 	timersub(&stop, &start0, &diff);
1327 	td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1328 	td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1329 	secs = td->runtime_ns / NSEC_PER_SEC;
1330 	td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1331 
1332 	getrusage(RUSAGE_THREAD, &rusage);
1333 	td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1334 	td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1335 	td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1336 	td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1337 
1338 	free_data(thread_data, g->p.bytes_thread);
1339 
1340 	mutex_lock(&g->stop_work_mutex);
1341 	g->bytes_done += bytes_done;
1342 	mutex_unlock(&g->stop_work_mutex);
1343 
1344 	return NULL;
1345 }
1346 
1347 /*
1348  * A worker process starts a couple of threads:
1349  */
1350 static void worker_process(int process_nr)
1351 {
1352 	struct mutex process_lock;
1353 	struct thread_data *td;
1354 	pthread_t *pthreads;
1355 	u8 *process_data;
1356 	int task_nr;
1357 	int ret;
1358 	int t;
1359 
1360 	mutex_init(&process_lock);
1361 	set_taskname("process %d", process_nr);
1362 
1363 	/*
1364 	 * Pick up the memory policy and the CPU binding of our first thread,
1365 	 * so that we initialize memory accordingly:
1366 	 */
1367 	task_nr = process_nr*g->p.nr_threads;
1368 	td = g->threads + task_nr;
1369 
1370 	bind_to_memnode(td->bind_node);
1371 	bind_to_cpumask(td->bind_cpumask);
1372 
1373 	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1374 	process_data = setup_private_data(g->p.bytes_process);
1375 
1376 	if (g->p.show_details >= 3) {
1377 		printf(" # process %2d global mem: %p, process mem: %p\n",
1378 			process_nr, g->data, process_data);
1379 	}
1380 
1381 	for (t = 0; t < g->p.nr_threads; t++) {
1382 		task_nr = process_nr*g->p.nr_threads + t;
1383 		td = g->threads + task_nr;
1384 
1385 		td->process_data = process_data;
1386 		td->process_nr   = process_nr;
1387 		td->thread_nr    = t;
1388 		td->task_nr	 = task_nr;
1389 		td->val          = rand();
1390 		td->curr_cpu	 = -1;
1391 		td->process_lock = &process_lock;
1392 
1393 		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1394 		BUG_ON(ret);
1395 	}
1396 
1397 	for (t = 0; t < g->p.nr_threads; t++) {
1398                 ret = pthread_join(pthreads[t], NULL);
1399 		BUG_ON(ret);
1400 	}
1401 
1402 	free_data(process_data, g->p.bytes_process);
1403 	free(pthreads);
1404 }
1405 
1406 static void print_summary(void)
1407 {
1408 	if (g->p.show_details < 0)
1409 		return;
1410 
1411 	printf("\n ###\n");
1412 	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1413 		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1414 	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1415 			g->p.nr_loops, g->p.bytes_global/1024/1024);
1416 	printf(" #      %5dx %5ldMB process shared mem operations\n",
1417 			g->p.nr_loops, g->p.bytes_process/1024/1024);
1418 	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1419 			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1420 
1421 	printf(" ###\n");
1422 
1423 	printf("\n ###\n"); fflush(stdout);
1424 }
1425 
1426 static void init_thread_data(void)
1427 {
1428 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1429 	int t;
1430 
1431 	g->threads = zalloc_shared_data(size);
1432 
1433 	for (t = 0; t < g->p.nr_tasks; t++) {
1434 		struct thread_data *td = g->threads + t;
1435 		size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1436 		int cpu;
1437 
1438 		/* Allow all nodes by default: */
1439 		td->bind_node = NUMA_NO_NODE;
1440 
1441 		/* Allow all CPUs by default: */
1442 		td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1443 		BUG_ON(!td->bind_cpumask);
1444 		CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1445 		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1446 			CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1447 	}
1448 }
1449 
1450 static void deinit_thread_data(void)
1451 {
1452 	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1453 	int t;
1454 
1455 	/* Free the bind_cpumask allocated for thread_data */
1456 	for (t = 0; t < g->p.nr_tasks; t++) {
1457 		struct thread_data *td = g->threads + t;
1458 		CPU_FREE(td->bind_cpumask);
1459 	}
1460 
1461 	free_data(g->threads, size);
1462 }
1463 
1464 static int init(void)
1465 {
1466 	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1467 
1468 	/* Copy over options: */
1469 	g->p = p0;
1470 
1471 	g->p.nr_cpus = numa_num_configured_cpus();
1472 
1473 	g->p.nr_nodes = numa_max_node() + 1;
1474 
1475 	/* char array in count_process_nodes(): */
1476 	BUG_ON(g->p.nr_nodes < 0);
1477 
1478 	if (quiet && !g->p.show_details)
1479 		g->p.show_details = -1;
1480 
1481 	/* Some memory should be specified: */
1482 	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1483 		return -1;
1484 
1485 	if (g->p.mb_global_str) {
1486 		g->p.mb_global = atof(g->p.mb_global_str);
1487 		BUG_ON(g->p.mb_global < 0);
1488 	}
1489 
1490 	if (g->p.mb_proc_str) {
1491 		g->p.mb_proc = atof(g->p.mb_proc_str);
1492 		BUG_ON(g->p.mb_proc < 0);
1493 	}
1494 
1495 	if (g->p.mb_proc_locked_str) {
1496 		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1497 		BUG_ON(g->p.mb_proc_locked < 0);
1498 		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1499 	}
1500 
1501 	if (g->p.mb_thread_str) {
1502 		g->p.mb_thread = atof(g->p.mb_thread_str);
1503 		BUG_ON(g->p.mb_thread < 0);
1504 	}
1505 
1506 	BUG_ON(g->p.nr_threads <= 0);
1507 	BUG_ON(g->p.nr_proc <= 0);
1508 
1509 	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1510 
1511 	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1512 	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1513 	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1514 	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1515 
1516 	g->data = setup_shared_data(g->p.bytes_global);
1517 
1518 	/* Startup serialization: */
1519 	mutex_init_pshared(&g->start_work_mutex);
1520 	cond_init_pshared(&g->start_work_cond);
1521 	mutex_init_pshared(&g->startup_mutex);
1522 	cond_init_pshared(&g->startup_cond);
1523 	mutex_init_pshared(&g->stop_work_mutex);
1524 
1525 	init_thread_data();
1526 
1527 	tprintf("#\n");
1528 	if (parse_setup_cpu_list() || parse_setup_node_list())
1529 		return -1;
1530 	tprintf("#\n");
1531 
1532 	print_summary();
1533 
1534 	return 0;
1535 }
1536 
1537 static void deinit(void)
1538 {
1539 	free_data(g->data, g->p.bytes_global);
1540 	g->data = NULL;
1541 
1542 	deinit_thread_data();
1543 
1544 	free_data(g, sizeof(*g));
1545 	g = NULL;
1546 }
1547 
1548 /*
1549  * Print a short or long result, depending on the verbosity setting:
1550  */
1551 static void print_res(const char *name, double val,
1552 		      const char *txt_unit, const char *txt_short, const char *txt_long)
1553 {
1554 	if (!name)
1555 		name = "main,";
1556 
1557 	if (!quiet)
1558 		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1559 	else
1560 		printf(" %14.3f %s\n", val, txt_long);
1561 }
1562 
1563 static int __bench_numa(const char *name)
1564 {
1565 	struct timeval start, stop, diff;
1566 	u64 runtime_ns_min, runtime_ns_sum;
1567 	pid_t *pids, pid, wpid;
1568 	double delta_runtime;
1569 	double runtime_avg;
1570 	double runtime_sec_max;
1571 	double runtime_sec_min;
1572 	int wait_stat;
1573 	double bytes;
1574 	int i, t, p;
1575 
1576 	if (init())
1577 		return -1;
1578 
1579 	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1580 	pid = -1;
1581 
1582 	if (g->p.serialize_startup) {
1583 		tprintf(" #\n");
1584 		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1585 	}
1586 
1587 	gettimeofday(&start, NULL);
1588 
1589 	for (i = 0; i < g->p.nr_proc; i++) {
1590 		pid = fork();
1591 		dprintf(" # process %2d: PID %d\n", i, pid);
1592 
1593 		BUG_ON(pid < 0);
1594 		if (!pid) {
1595 			/* Child process: */
1596 			worker_process(i);
1597 
1598 			exit(0);
1599 		}
1600 		pids[i] = pid;
1601 
1602 	}
1603 
1604 	if (g->p.serialize_startup) {
1605 		bool threads_ready = false;
1606 		double startup_sec;
1607 
1608 		/*
1609 		 * Wait for all the threads to start up. The last thread will
1610 		 * signal this process.
1611 		 */
1612 		mutex_lock(&g->startup_mutex);
1613 		while (g->nr_tasks_started != g->p.nr_tasks)
1614 			cond_wait(&g->startup_cond, &g->startup_mutex);
1615 
1616 		mutex_unlock(&g->startup_mutex);
1617 
1618 		/* Wait for all threads to be at the start_work_cond. */
1619 		while (!threads_ready) {
1620 			mutex_lock(&g->start_work_mutex);
1621 			threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1622 			mutex_unlock(&g->start_work_mutex);
1623 			if (!threads_ready)
1624 				usleep(1);
1625 		}
1626 
1627 		gettimeofday(&stop, NULL);
1628 
1629 		timersub(&stop, &start, &diff);
1630 
1631 		startup_sec = diff.tv_sec * NSEC_PER_SEC;
1632 		startup_sec += diff.tv_usec * NSEC_PER_USEC;
1633 		startup_sec /= NSEC_PER_SEC;
1634 
1635 		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1636 		tprintf(" #\n");
1637 
1638 		start = stop;
1639 		/* Start all threads running. */
1640 		mutex_lock(&g->start_work_mutex);
1641 		g->start_work = true;
1642 		mutex_unlock(&g->start_work_mutex);
1643 		cond_broadcast(&g->start_work_cond);
1644 	} else {
1645 		gettimeofday(&start, NULL);
1646 	}
1647 
1648 	/* Parent process: */
1649 
1650 
1651 	for (i = 0; i < g->p.nr_proc; i++) {
1652 		wpid = waitpid(pids[i], &wait_stat, 0);
1653 		BUG_ON(wpid < 0);
1654 		BUG_ON(!WIFEXITED(wait_stat));
1655 
1656 	}
1657 
1658 	runtime_ns_sum = 0;
1659 	runtime_ns_min = -1LL;
1660 
1661 	for (t = 0; t < g->p.nr_tasks; t++) {
1662 		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1663 
1664 		runtime_ns_sum += thread_runtime_ns;
1665 		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1666 	}
1667 
1668 	gettimeofday(&stop, NULL);
1669 	timersub(&stop, &start, &diff);
1670 
1671 	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1672 
1673 	tprintf("\n ###\n");
1674 	tprintf("\n");
1675 
1676 	runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1677 	runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1678 	runtime_sec_max /= NSEC_PER_SEC;
1679 
1680 	runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1681 
1682 	bytes = g->bytes_done;
1683 	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1684 
1685 	if (g->p.measure_convergence) {
1686 		print_res(name, runtime_sec_max,
1687 			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1688 	}
1689 
1690 	print_res(name, runtime_sec_max,
1691 		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1692 
1693 	print_res(name, runtime_sec_min,
1694 		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1695 
1696 	print_res(name, runtime_avg,
1697 		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1698 
1699 	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1700 	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1701 		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1702 
1703 	print_res(name, bytes / g->p.nr_tasks / 1e9,
1704 		"GB,", "data/thread",		"GB data processed, per thread");
1705 
1706 	print_res(name, bytes / 1e9,
1707 		"GB,", "data-total",		"GB data processed, total");
1708 
1709 	print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1710 		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1711 
1712 	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1713 		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1714 
1715 	print_res(name, bytes / runtime_sec_max / 1e9,
1716 		"GB/sec,", "total-speed",	"GB/sec total speed");
1717 
1718 	if (g->p.show_details >= 2) {
1719 		char tname[14 + 2 * 11 + 1];
1720 		struct thread_data *td;
1721 		for (p = 0; p < g->p.nr_proc; p++) {
1722 			for (t = 0; t < g->p.nr_threads; t++) {
1723 				memset(tname, 0, sizeof(tname));
1724 				td = g->threads + p*g->p.nr_threads + t;
1725 				snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1726 				print_res(tname, td->speed_gbs,
1727 					"GB/sec",	"thread-speed", "GB/sec/thread speed");
1728 				print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1729 					"secs",	"thread-system-time", "system CPU time/thread");
1730 				print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1731 					"secs",	"thread-user-time", "user CPU time/thread");
1732 			}
1733 		}
1734 	}
1735 
1736 	free(pids);
1737 
1738 	deinit();
1739 
1740 	return 0;
1741 }
1742 
1743 #define MAX_ARGS 50
1744 
1745 static int command_size(const char **argv)
1746 {
1747 	int size = 0;
1748 
1749 	while (*argv) {
1750 		size++;
1751 		argv++;
1752 	}
1753 
1754 	BUG_ON(size >= MAX_ARGS);
1755 
1756 	return size;
1757 }
1758 
1759 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1760 {
1761 	int i;
1762 
1763 	printf("\n # Running %s \"perf bench numa", name);
1764 
1765 	for (i = 0; i < argc; i++)
1766 		printf(" %s", argv[i]);
1767 
1768 	printf("\"\n");
1769 
1770 	memset(p, 0, sizeof(*p));
1771 
1772 	/* Initialize nonzero defaults: */
1773 
1774 	p->serialize_startup		= 1;
1775 	p->data_reads			= true;
1776 	p->data_writes			= true;
1777 	p->data_backwards		= true;
1778 	p->data_rand_walk		= true;
1779 	p->nr_loops			= -1;
1780 	p->init_random			= true;
1781 	p->mb_global_str		= "1";
1782 	p->nr_proc			= 1;
1783 	p->nr_threads			= 1;
1784 	p->nr_secs			= 5;
1785 	p->run_all			= argc == 1;
1786 }
1787 
1788 static int run_bench_numa(const char *name, const char **argv)
1789 {
1790 	int argc = command_size(argv);
1791 
1792 	init_params(&p0, name, argc, argv);
1793 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1794 	if (argc)
1795 		goto err;
1796 
1797 	if (__bench_numa(name))
1798 		goto err;
1799 
1800 	return 0;
1801 
1802 err:
1803 	return -1;
1804 }
1805 
1806 #define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1807 #define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1808 
1809 #define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1810 #define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1811 
1812 #define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1813 #define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1814 
1815 /*
1816  * The built-in test-suite executed by "perf bench numa -a".
1817  *
1818  * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1819  */
1820 static const char *tests[][MAX_ARGS] = {
1821    /* Basic single-stream NUMA bandwidth measurements: */
1822    { "RAM-bw-local,",     "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1823 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1824    { "RAM-bw-local-NOTHP,",
1825 			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1826 			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1827    { "RAM-bw-remote,",    "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1828 			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1829 
1830    /* 2-stream NUMA bandwidth measurements: */
1831    { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1832 			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1833    { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1834 		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1835 
1836    /* Cross-stream NUMA bandwidth measurement: */
1837    { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1838 		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1839 
1840    /* Convergence latency measurements: */
1841    { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1842    { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1843    { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1844    { " 2x3-convergence,", "mem",  "-p",  "2", "-t",  "3", "-P", "1020", OPT_CONV },
1845    { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1846    { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1847    { " 4x4-convergence-NOTHP,",
1848 			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1849    { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1850    { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1851    { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1852    { " 8x4-convergence-NOTHP,",
1853 			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1854    { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1855    { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1856    { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1857    { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1858    { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1859 
1860    /* Various NUMA process/thread layout bandwidth measurements: */
1861    { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1862    { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1863    { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1864    { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1865    { " 8x1-bw-process-NOTHP,",
1866 			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1867    { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1868 
1869    { " 1x4-bw-thread,",   "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1870    { " 1x8-bw-thread,",   "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1871    { "1x16-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1872    { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1873 
1874    { " 2x3-bw-process,",  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1875    { " 4x4-bw-process,",  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1876    { " 4x6-bw-process,",  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1877    { " 4x8-bw-process,",  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1878    { " 4x8-bw-process-NOTHP,",
1879 			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1880    { " 3x3-bw-process,",  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1881    { " 5x5-bw-process,",  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1882 
1883    { "2x16-bw-process,",  "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1884    { "1x32-bw-process,",  "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1885 
1886    { "numa02-bw,",        "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1887    { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1888    { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1889    { "numa01-bw-thread-NOTHP,",
1890 			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1891 };
1892 
1893 static int bench_all(void)
1894 {
1895 	int nr = ARRAY_SIZE(tests);
1896 	int ret;
1897 	int i;
1898 
1899 	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1900 	BUG_ON(ret < 0);
1901 
1902 	for (i = 0; i < nr; i++) {
1903 		run_bench_numa(tests[i][0], tests[i] + 1);
1904 	}
1905 
1906 	printf("\n");
1907 
1908 	return 0;
1909 }
1910 
1911 int bench_numa(int argc, const char **argv)
1912 {
1913 	init_params(&p0, "main,", argc, argv);
1914 	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1915 	if (argc)
1916 		goto err;
1917 
1918 	if (p0.run_all)
1919 		return bench_all();
1920 
1921 	if (__bench_numa(NULL))
1922 		goto err;
1923 
1924 	return 0;
1925 
1926 err:
1927 	usage_with_options(numa_usage, options);
1928 	return -1;
1929 }
1930