1/*
2 * numa.c
3 *
4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
5 */
6
7#include "../perf.h"
8#include "../builtin.h"
9#include "../util/util.h"
10#include "../util/parse-options.h"
11
12#include "bench.h"
13
14#include <errno.h>
15#include <sched.h>
16#include <stdio.h>
17#include <assert.h>
18#include <malloc.h>
19#include <signal.h>
20#include <stdlib.h>
21#include <string.h>
22#include <unistd.h>
23#include <pthread.h>
24#include <sys/mman.h>
25#include <sys/time.h>
26#include <sys/wait.h>
27#include <sys/prctl.h>
28#include <sys/types.h>
29
30#include <numa.h>
31#include <numaif.h>
32
33/*
34 * Regular printout to the terminal, supressed if -q is specified:
35 */
36#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
37
38/*
39 * Debug printf:
40 */
41#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
42
43struct thread_data {
44	int			curr_cpu;
45	cpu_set_t		bind_cpumask;
46	int			bind_node;
47	u8			*process_data;
48	int			process_nr;
49	int			thread_nr;
50	int			task_nr;
51	unsigned int		loops_done;
52	u64			val;
53	u64			runtime_ns;
54	pthread_mutex_t		*process_lock;
55};
56
57/* Parameters set by options: */
58
59struct params {
60	/* Startup synchronization: */
61	bool			serialize_startup;
62
63	/* Task hierarchy: */
64	int			nr_proc;
65	int			nr_threads;
66
67	/* Working set sizes: */
68	const char		*mb_global_str;
69	const char		*mb_proc_str;
70	const char		*mb_proc_locked_str;
71	const char		*mb_thread_str;
72
73	double			mb_global;
74	double			mb_proc;
75	double			mb_proc_locked;
76	double			mb_thread;
77
78	/* Access patterns to the working set: */
79	bool			data_reads;
80	bool			data_writes;
81	bool			data_backwards;
82	bool			data_zero_memset;
83	bool			data_rand_walk;
84	u32			nr_loops;
85	u32			nr_secs;
86	u32			sleep_usecs;
87
88	/* Working set initialization: */
89	bool			init_zero;
90	bool			init_random;
91	bool			init_cpu0;
92
93	/* Misc options: */
94	int			show_details;
95	int			run_all;
96	int			thp;
97
98	long			bytes_global;
99	long			bytes_process;
100	long			bytes_process_locked;
101	long			bytes_thread;
102
103	int			nr_tasks;
104	bool			show_quiet;
105
106	bool			show_convergence;
107	bool			measure_convergence;
108
109	int			perturb_secs;
110	int			nr_cpus;
111	int			nr_nodes;
112
113	/* Affinity options -C and -N: */
114	char			*cpu_list_str;
115	char			*node_list_str;
116};
117
118
119/* Global, read-writable area, accessible to all processes and threads: */
120
121struct global_info {
122	u8			*data;
123
124	pthread_mutex_t		startup_mutex;
125	int			nr_tasks_started;
126
127	pthread_mutex_t		startup_done_mutex;
128
129	pthread_mutex_t		start_work_mutex;
130	int			nr_tasks_working;
131
132	pthread_mutex_t		stop_work_mutex;
133	u64			bytes_done;
134
135	struct thread_data	*threads;
136
137	/* Convergence latency measurement: */
138	bool			all_converged;
139	bool			stop_work;
140
141	int			print_once;
142
143	struct params		p;
144};
145
146static struct global_info	*g = NULL;
147
148static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
149static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
150
151struct params p0;
152
153static const struct option options[] = {
154	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
155	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
156
157	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
158	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
159	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
160	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
161
162	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run"),
163	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run"),
164	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
165
166	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via writes (can be mixed with -W)"),
167	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
168	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
169	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
170	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
171
172
173	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
174	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
175	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
176	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
177
178	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
179	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
180	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
181	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
182	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
183	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"quiet mode"),
184	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
185
186	/* Special option string parsing callbacks: */
187        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
188			"bind the first N tasks to these specific cpus (the rest is unbound)",
189			parse_cpus_opt),
190        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
191			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
192			parse_nodes_opt),
193	OPT_END()
194};
195
196static const char * const bench_numa_usage[] = {
197	"perf bench numa <options>",
198	NULL
199};
200
201static const char * const numa_usage[] = {
202	"perf bench numa mem [<options>]",
203	NULL
204};
205
206static cpu_set_t bind_to_cpu(int target_cpu)
207{
208	cpu_set_t orig_mask, mask;
209	int ret;
210
211	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
212	BUG_ON(ret);
213
214	CPU_ZERO(&mask);
215
216	if (target_cpu == -1) {
217		int cpu;
218
219		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
220			CPU_SET(cpu, &mask);
221	} else {
222		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
223		CPU_SET(target_cpu, &mask);
224	}
225
226	ret = sched_setaffinity(0, sizeof(mask), &mask);
227	BUG_ON(ret);
228
229	return orig_mask;
230}
231
232static cpu_set_t bind_to_node(int target_node)
233{
234	int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
235	cpu_set_t orig_mask, mask;
236	int cpu;
237	int ret;
238
239	BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
240	BUG_ON(!cpus_per_node);
241
242	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
243	BUG_ON(ret);
244
245	CPU_ZERO(&mask);
246
247	if (target_node == -1) {
248		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
249			CPU_SET(cpu, &mask);
250	} else {
251		int cpu_start = (target_node + 0) * cpus_per_node;
252		int cpu_stop  = (target_node + 1) * cpus_per_node;
253
254		BUG_ON(cpu_stop > g->p.nr_cpus);
255
256		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
257			CPU_SET(cpu, &mask);
258	}
259
260	ret = sched_setaffinity(0, sizeof(mask), &mask);
261	BUG_ON(ret);
262
263	return orig_mask;
264}
265
266static void bind_to_cpumask(cpu_set_t mask)
267{
268	int ret;
269
270	ret = sched_setaffinity(0, sizeof(mask), &mask);
271	BUG_ON(ret);
272}
273
274static void mempol_restore(void)
275{
276	int ret;
277
278	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
279
280	BUG_ON(ret);
281}
282
283static void bind_to_memnode(int node)
284{
285	unsigned long nodemask;
286	int ret;
287
288	if (node == -1)
289		return;
290
291	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask));
292	nodemask = 1L << node;
293
294	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
295	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
296
297	BUG_ON(ret);
298}
299
300#define HPSIZE (2*1024*1024)
301
302#define set_taskname(fmt...)				\
303do {							\
304	char name[20];					\
305							\
306	snprintf(name, 20, fmt);			\
307	prctl(PR_SET_NAME, name);			\
308} while (0)
309
310static u8 *alloc_data(ssize_t bytes0, int map_flags,
311		      int init_zero, int init_cpu0, int thp, int init_random)
312{
313	cpu_set_t orig_mask;
314	ssize_t bytes;
315	u8 *buf;
316	int ret;
317
318	if (!bytes0)
319		return NULL;
320
321	/* Allocate and initialize all memory on CPU#0: */
322	if (init_cpu0) {
323		orig_mask = bind_to_node(0);
324		bind_to_memnode(0);
325	}
326
327	bytes = bytes0 + HPSIZE;
328
329	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
330	BUG_ON(buf == (void *)-1);
331
332	if (map_flags == MAP_PRIVATE) {
333		if (thp > 0) {
334			ret = madvise(buf, bytes, MADV_HUGEPAGE);
335			if (ret && !g->print_once) {
336				g->print_once = 1;
337				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
338			}
339		}
340		if (thp < 0) {
341			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
342			if (ret && !g->print_once) {
343				g->print_once = 1;
344				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
345			}
346		}
347	}
348
349	if (init_zero) {
350		bzero(buf, bytes);
351	} else {
352		/* Initialize random contents, different in each word: */
353		if (init_random) {
354			u64 *wbuf = (void *)buf;
355			long off = rand();
356			long i;
357
358			for (i = 0; i < bytes/8; i++)
359				wbuf[i] = i + off;
360		}
361	}
362
363	/* Align to 2MB boundary: */
364	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
365
366	/* Restore affinity: */
367	if (init_cpu0) {
368		bind_to_cpumask(orig_mask);
369		mempol_restore();
370	}
371
372	return buf;
373}
374
375static void free_data(void *data, ssize_t bytes)
376{
377	int ret;
378
379	if (!data)
380		return;
381
382	ret = munmap(data, bytes);
383	BUG_ON(ret);
384}
385
386/*
387 * Create a shared memory buffer that can be shared between processes, zeroed:
388 */
389static void * zalloc_shared_data(ssize_t bytes)
390{
391	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
392}
393
394/*
395 * Create a shared memory buffer that can be shared between processes:
396 */
397static void * setup_shared_data(ssize_t bytes)
398{
399	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
400}
401
402/*
403 * Allocate process-local memory - this will either be shared between
404 * threads of this process, or only be accessed by this thread:
405 */
406static void * setup_private_data(ssize_t bytes)
407{
408	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
409}
410
411/*
412 * Return a process-shared (global) mutex:
413 */
414static void init_global_mutex(pthread_mutex_t *mutex)
415{
416	pthread_mutexattr_t attr;
417
418	pthread_mutexattr_init(&attr);
419	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
420	pthread_mutex_init(mutex, &attr);
421}
422
423static int parse_cpu_list(const char *arg)
424{
425	p0.cpu_list_str = strdup(arg);
426
427	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
428
429	return 0;
430}
431
432static int parse_setup_cpu_list(void)
433{
434	struct thread_data *td;
435	char *str0, *str;
436	int t;
437
438	if (!g->p.cpu_list_str)
439		return 0;
440
441	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
442
443	str0 = str = strdup(g->p.cpu_list_str);
444	t = 0;
445
446	BUG_ON(!str);
447
448	tprintf("# binding tasks to CPUs:\n");
449	tprintf("#  ");
450
451	while (true) {
452		int bind_cpu, bind_cpu_0, bind_cpu_1;
453		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
454		int bind_len;
455		int step;
456		int mul;
457
458		tok = strsep(&str, ",");
459		if (!tok)
460			break;
461
462		tok_end = strstr(tok, "-");
463
464		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
465		if (!tok_end) {
466			/* Single CPU specified: */
467			bind_cpu_0 = bind_cpu_1 = atol(tok);
468		} else {
469			/* CPU range specified (for example: "5-11"): */
470			bind_cpu_0 = atol(tok);
471			bind_cpu_1 = atol(tok_end + 1);
472		}
473
474		step = 1;
475		tok_step = strstr(tok, "#");
476		if (tok_step) {
477			step = atol(tok_step + 1);
478			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
479		}
480
481		/*
482		 * Mask length.
483		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
484		 * where the _4 means the next 4 CPUs are allowed.
485		 */
486		bind_len = 1;
487		tok_len = strstr(tok, "_");
488		if (tok_len) {
489			bind_len = atol(tok_len + 1);
490			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
491		}
492
493		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
494		mul = 1;
495		tok_mul = strstr(tok, "x");
496		if (tok_mul) {
497			mul = atol(tok_mul + 1);
498			BUG_ON(mul <= 0);
499		}
500
501		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
502
503		if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
504			printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
505			return -1;
506		}
507
508		BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
509		BUG_ON(bind_cpu_0 > bind_cpu_1);
510
511		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
512			int i;
513
514			for (i = 0; i < mul; i++) {
515				int cpu;
516
517				if (t >= g->p.nr_tasks) {
518					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
519					goto out;
520				}
521				td = g->threads + t;
522
523				if (t)
524					tprintf(",");
525				if (bind_len > 1) {
526					tprintf("%2d/%d", bind_cpu, bind_len);
527				} else {
528					tprintf("%2d", bind_cpu);
529				}
530
531				CPU_ZERO(&td->bind_cpumask);
532				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
533					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
534					CPU_SET(cpu, &td->bind_cpumask);
535				}
536				t++;
537			}
538		}
539	}
540out:
541
542	tprintf("\n");
543
544	if (t < g->p.nr_tasks)
545		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
546
547	free(str0);
548	return 0;
549}
550
551static int parse_cpus_opt(const struct option *opt __maybe_unused,
552			  const char *arg, int unset __maybe_unused)
553{
554	if (!arg)
555		return -1;
556
557	return parse_cpu_list(arg);
558}
559
560static int parse_node_list(const char *arg)
561{
562	p0.node_list_str = strdup(arg);
563
564	dprintf("got NODE list: {%s}\n", p0.node_list_str);
565
566	return 0;
567}
568
569static int parse_setup_node_list(void)
570{
571	struct thread_data *td;
572	char *str0, *str;
573	int t;
574
575	if (!g->p.node_list_str)
576		return 0;
577
578	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
579
580	str0 = str = strdup(g->p.node_list_str);
581	t = 0;
582
583	BUG_ON(!str);
584
585	tprintf("# binding tasks to NODEs:\n");
586	tprintf("# ");
587
588	while (true) {
589		int bind_node, bind_node_0, bind_node_1;
590		char *tok, *tok_end, *tok_step, *tok_mul;
591		int step;
592		int mul;
593
594		tok = strsep(&str, ",");
595		if (!tok)
596			break;
597
598		tok_end = strstr(tok, "-");
599
600		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
601		if (!tok_end) {
602			/* Single NODE specified: */
603			bind_node_0 = bind_node_1 = atol(tok);
604		} else {
605			/* NODE range specified (for example: "5-11"): */
606			bind_node_0 = atol(tok);
607			bind_node_1 = atol(tok_end + 1);
608		}
609
610		step = 1;
611		tok_step = strstr(tok, "#");
612		if (tok_step) {
613			step = atol(tok_step + 1);
614			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
615		}
616
617		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
618		mul = 1;
619		tok_mul = strstr(tok, "x");
620		if (tok_mul) {
621			mul = atol(tok_mul + 1);
622			BUG_ON(mul <= 0);
623		}
624
625		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
626
627		if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
628			printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
629			return -1;
630		}
631
632		BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
633		BUG_ON(bind_node_0 > bind_node_1);
634
635		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
636			int i;
637
638			for (i = 0; i < mul; i++) {
639				if (t >= g->p.nr_tasks) {
640					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
641					goto out;
642				}
643				td = g->threads + t;
644
645				if (!t)
646					tprintf(" %2d", bind_node);
647				else
648					tprintf(",%2d", bind_node);
649
650				td->bind_node = bind_node;
651				t++;
652			}
653		}
654	}
655out:
656
657	tprintf("\n");
658
659	if (t < g->p.nr_tasks)
660		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
661
662	free(str0);
663	return 0;
664}
665
666static int parse_nodes_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_node_list(arg);
673
674	return 0;
675}
676
677#define BIT(x) (1ul << x)
678
679static inline uint32_t lfsr_32(uint32_t lfsr)
680{
681	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
682	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
683}
684
685/*
686 * Make sure there's real data dependency to RAM (when read
687 * accesses are enabled), so the compiler, the CPU and the
688 * kernel (KSM, zero page, etc.) cannot optimize away RAM
689 * accesses:
690 */
691static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
692{
693	if (g->p.data_reads)
694		val += *data;
695	if (g->p.data_writes)
696		*data = val + 1;
697	return val;
698}
699
700/*
701 * The worker process does two types of work, a forwards going
702 * loop and a backwards going loop.
703 *
704 * We do this so that on multiprocessor systems we do not create
705 * a 'train' of processing, with highly synchronized processes,
706 * skewing the whole benchmark.
707 */
708static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
709{
710	long words = bytes/sizeof(u64);
711	u64 *data = (void *)__data;
712	long chunk_0, chunk_1;
713	u64 *d0, *d, *d1;
714	long off;
715	long i;
716
717	BUG_ON(!data && words);
718	BUG_ON(data && !words);
719
720	if (!data)
721		return val;
722
723	/* Very simple memset() work variant: */
724	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
725		bzero(data, bytes);
726		return val;
727	}
728
729	/* Spread out by PID/TID nr and by loop nr: */
730	chunk_0 = words/nr_max;
731	chunk_1 = words/g->p.nr_loops;
732	off = nr*chunk_0 + loop*chunk_1;
733
734	while (off >= words)
735		off -= words;
736
737	if (g->p.data_rand_walk) {
738		u32 lfsr = nr + loop + val;
739		int j;
740
741		for (i = 0; i < words/1024; i++) {
742			long start, end;
743
744			lfsr = lfsr_32(lfsr);
745
746			start = lfsr % words;
747			end = min(start + 1024, words-1);
748
749			if (g->p.data_zero_memset) {
750				bzero(data + start, (end-start) * sizeof(u64));
751			} else {
752				for (j = start; j < end; j++)
753					val = access_data(data + j, val);
754			}
755		}
756	} else if (!g->p.data_backwards || (nr + loop) & 1) {
757
758		d0 = data + off;
759		d  = data + off + 1;
760		d1 = data + words;
761
762		/* Process data forwards: */
763		for (;;) {
764			if (unlikely(d >= d1))
765				d = data;
766			if (unlikely(d == d0))
767				break;
768
769			val = access_data(d, val);
770
771			d++;
772		}
773	} else {
774		/* Process data backwards: */
775
776		d0 = data + off;
777		d  = data + off - 1;
778		d1 = data + words;
779
780		/* Process data forwards: */
781		for (;;) {
782			if (unlikely(d < data))
783				d = data + words-1;
784			if (unlikely(d == d0))
785				break;
786
787			val = access_data(d, val);
788
789			d--;
790		}
791	}
792
793	return val;
794}
795
796static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
797{
798	unsigned int cpu;
799
800	cpu = sched_getcpu();
801
802	g->threads[task_nr].curr_cpu = cpu;
803	prctl(0, bytes_worked);
804}
805
806#define MAX_NR_NODES	64
807
808/*
809 * Count the number of nodes a process's threads
810 * are spread out on.
811 *
812 * A count of 1 means that the process is compressed
813 * to a single node. A count of g->p.nr_nodes means it's
814 * spread out on the whole system.
815 */
816static int count_process_nodes(int process_nr)
817{
818	char node_present[MAX_NR_NODES] = { 0, };
819	int nodes;
820	int n, t;
821
822	for (t = 0; t < g->p.nr_threads; t++) {
823		struct thread_data *td;
824		int task_nr;
825		int node;
826
827		task_nr = process_nr*g->p.nr_threads + t;
828		td = g->threads + task_nr;
829
830		node = numa_node_of_cpu(td->curr_cpu);
831		if (node < 0) /* curr_cpu was likely still -1 */
832			return 0;
833
834		node_present[node] = 1;
835	}
836
837	nodes = 0;
838
839	for (n = 0; n < MAX_NR_NODES; n++)
840		nodes += node_present[n];
841
842	return nodes;
843}
844
845/*
846 * Count the number of distinct process-threads a node contains.
847 *
848 * A count of 1 means that the node contains only a single
849 * process. If all nodes on the system contain at most one
850 * process then we are well-converged.
851 */
852static int count_node_processes(int node)
853{
854	int processes = 0;
855	int t, p;
856
857	for (p = 0; p < g->p.nr_proc; p++) {
858		for (t = 0; t < g->p.nr_threads; t++) {
859			struct thread_data *td;
860			int task_nr;
861			int n;
862
863			task_nr = p*g->p.nr_threads + t;
864			td = g->threads + task_nr;
865
866			n = numa_node_of_cpu(td->curr_cpu);
867			if (n == node) {
868				processes++;
869				break;
870			}
871		}
872	}
873
874	return processes;
875}
876
877static void calc_convergence_compression(int *strong)
878{
879	unsigned int nodes_min, nodes_max;
880	int p;
881
882	nodes_min = -1;
883	nodes_max =  0;
884
885	for (p = 0; p < g->p.nr_proc; p++) {
886		unsigned int nodes = count_process_nodes(p);
887
888		if (!nodes) {
889			*strong = 0;
890			return;
891		}
892
893		nodes_min = min(nodes, nodes_min);
894		nodes_max = max(nodes, nodes_max);
895	}
896
897	/* Strong convergence: all threads compress on a single node: */
898	if (nodes_min == 1 && nodes_max == 1) {
899		*strong = 1;
900	} else {
901		*strong = 0;
902		tprintf(" {%d-%d}", nodes_min, nodes_max);
903	}
904}
905
906static void calc_convergence(double runtime_ns_max, double *convergence)
907{
908	unsigned int loops_done_min, loops_done_max;
909	int process_groups;
910	int nodes[MAX_NR_NODES];
911	int distance;
912	int nr_min;
913	int nr_max;
914	int strong;
915	int sum;
916	int nr;
917	int node;
918	int cpu;
919	int t;
920
921	if (!g->p.show_convergence && !g->p.measure_convergence)
922		return;
923
924	for (node = 0; node < g->p.nr_nodes; node++)
925		nodes[node] = 0;
926
927	loops_done_min = -1;
928	loops_done_max = 0;
929
930	for (t = 0; t < g->p.nr_tasks; t++) {
931		struct thread_data *td = g->threads + t;
932		unsigned int loops_done;
933
934		cpu = td->curr_cpu;
935
936		/* Not all threads have written it yet: */
937		if (cpu < 0)
938			continue;
939
940		node = numa_node_of_cpu(cpu);
941
942		nodes[node]++;
943
944		loops_done = td->loops_done;
945		loops_done_min = min(loops_done, loops_done_min);
946		loops_done_max = max(loops_done, loops_done_max);
947	}
948
949	nr_max = 0;
950	nr_min = g->p.nr_tasks;
951	sum = 0;
952
953	for (node = 0; node < g->p.nr_nodes; node++) {
954		nr = nodes[node];
955		nr_min = min(nr, nr_min);
956		nr_max = max(nr, nr_max);
957		sum += nr;
958	}
959	BUG_ON(nr_min > nr_max);
960
961	BUG_ON(sum > g->p.nr_tasks);
962
963	if (0 && (sum < g->p.nr_tasks))
964		return;
965
966	/*
967	 * Count the number of distinct process groups present
968	 * on nodes - when we are converged this will decrease
969	 * to g->p.nr_proc:
970	 */
971	process_groups = 0;
972
973	for (node = 0; node < g->p.nr_nodes; node++) {
974		int processes = count_node_processes(node);
975
976		nr = nodes[node];
977		tprintf(" %2d/%-2d", nr, processes);
978
979		process_groups += processes;
980	}
981
982	distance = nr_max - nr_min;
983
984	tprintf(" [%2d/%-2d]", distance, process_groups);
985
986	tprintf(" l:%3d-%-3d (%3d)",
987		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
988
989	if (loops_done_min && loops_done_max) {
990		double skew = 1.0 - (double)loops_done_min/loops_done_max;
991
992		tprintf(" [%4.1f%%]", skew * 100.0);
993	}
994
995	calc_convergence_compression(&strong);
996
997	if (strong && process_groups == g->p.nr_proc) {
998		if (!*convergence) {
999			*convergence = runtime_ns_max;
1000			tprintf(" (%6.1fs converged)\n", *convergence/1e9);
1001			if (g->p.measure_convergence) {
1002				g->all_converged = true;
1003				g->stop_work = true;
1004			}
1005		}
1006	} else {
1007		if (*convergence) {
1008			tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
1009			*convergence = 0;
1010		}
1011		tprintf("\n");
1012	}
1013}
1014
1015static void show_summary(double runtime_ns_max, int l, double *convergence)
1016{
1017	tprintf("\r #  %5.1f%%  [%.1f mins]",
1018		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1019
1020	calc_convergence(runtime_ns_max, convergence);
1021
1022	if (g->p.show_details >= 0)
1023		fflush(stdout);
1024}
1025
1026static void *worker_thread(void *__tdata)
1027{
1028	struct thread_data *td = __tdata;
1029	struct timeval start0, start, stop, diff;
1030	int process_nr = td->process_nr;
1031	int thread_nr = td->thread_nr;
1032	unsigned long last_perturbance;
1033	int task_nr = td->task_nr;
1034	int details = g->p.show_details;
1035	int first_task, last_task;
1036	double convergence = 0;
1037	u64 val = td->val;
1038	double runtime_ns_max;
1039	u8 *global_data;
1040	u8 *process_data;
1041	u8 *thread_data;
1042	u64 bytes_done;
1043	long work_done;
1044	u32 l;
1045
1046	bind_to_cpumask(td->bind_cpumask);
1047	bind_to_memnode(td->bind_node);
1048
1049	set_taskname("thread %d/%d", process_nr, thread_nr);
1050
1051	global_data = g->data;
1052	process_data = td->process_data;
1053	thread_data = setup_private_data(g->p.bytes_thread);
1054
1055	bytes_done = 0;
1056
1057	last_task = 0;
1058	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1059		last_task = 1;
1060
1061	first_task = 0;
1062	if (process_nr == 0 && thread_nr == 0)
1063		first_task = 1;
1064
1065	if (details >= 2) {
1066		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1067			process_nr, thread_nr, global_data, process_data, thread_data);
1068	}
1069
1070	if (g->p.serialize_startup) {
1071		pthread_mutex_lock(&g->startup_mutex);
1072		g->nr_tasks_started++;
1073		pthread_mutex_unlock(&g->startup_mutex);
1074
1075		/* Here we will wait for the main process to start us all at once: */
1076		pthread_mutex_lock(&g->start_work_mutex);
1077		g->nr_tasks_working++;
1078
1079		/* Last one wake the main process: */
1080		if (g->nr_tasks_working == g->p.nr_tasks)
1081			pthread_mutex_unlock(&g->startup_done_mutex);
1082
1083		pthread_mutex_unlock(&g->start_work_mutex);
1084	}
1085
1086	gettimeofday(&start0, NULL);
1087
1088	start = stop = start0;
1089	last_perturbance = start.tv_sec;
1090
1091	for (l = 0; l < g->p.nr_loops; l++) {
1092		start = stop;
1093
1094		if (g->stop_work)
1095			break;
1096
1097		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
1098		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
1099		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
1100
1101		if (g->p.sleep_usecs) {
1102			pthread_mutex_lock(td->process_lock);
1103			usleep(g->p.sleep_usecs);
1104			pthread_mutex_unlock(td->process_lock);
1105		}
1106		/*
1107		 * Amount of work to be done under a process-global lock:
1108		 */
1109		if (g->p.bytes_process_locked) {
1110			pthread_mutex_lock(td->process_lock);
1111			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
1112			pthread_mutex_unlock(td->process_lock);
1113		}
1114
1115		work_done = g->p.bytes_global + g->p.bytes_process +
1116			    g->p.bytes_process_locked + g->p.bytes_thread;
1117
1118		update_curr_cpu(task_nr, work_done);
1119		bytes_done += work_done;
1120
1121		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1122			continue;
1123
1124		td->loops_done = l;
1125
1126		gettimeofday(&stop, NULL);
1127
1128		/* Check whether our max runtime timed out: */
1129		if (g->p.nr_secs) {
1130			timersub(&stop, &start0, &diff);
1131			if ((u32)diff.tv_sec >= g->p.nr_secs) {
1132				g->stop_work = true;
1133				break;
1134			}
1135		}
1136
1137		/* Update the summary at most once per second: */
1138		if (start.tv_sec == stop.tv_sec)
1139			continue;
1140
1141		/*
1142		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1143		 * by migrating to CPU#0:
1144		 */
1145		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1146			cpu_set_t orig_mask;
1147			int target_cpu;
1148			int this_cpu;
1149
1150			last_perturbance = stop.tv_sec;
1151
1152			/*
1153			 * Depending on where we are running, move into
1154			 * the other half of the system, to create some
1155			 * real disturbance:
1156			 */
1157			this_cpu = g->threads[task_nr].curr_cpu;
1158			if (this_cpu < g->p.nr_cpus/2)
1159				target_cpu = g->p.nr_cpus-1;
1160			else
1161				target_cpu = 0;
1162
1163			orig_mask = bind_to_cpu(target_cpu);
1164
1165			/* Here we are running on the target CPU already */
1166			if (details >= 1)
1167				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1168
1169			bind_to_cpumask(orig_mask);
1170		}
1171
1172		if (details >= 3) {
1173			timersub(&stop, &start, &diff);
1174			runtime_ns_max = diff.tv_sec * 1000000000;
1175			runtime_ns_max += diff.tv_usec * 1000;
1176
1177			if (details >= 0) {
1178				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1179					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1180			}
1181			fflush(stdout);
1182		}
1183		if (!last_task)
1184			continue;
1185
1186		timersub(&stop, &start0, &diff);
1187		runtime_ns_max = diff.tv_sec * 1000000000ULL;
1188		runtime_ns_max += diff.tv_usec * 1000ULL;
1189
1190		show_summary(runtime_ns_max, l, &convergence);
1191	}
1192
1193	gettimeofday(&stop, NULL);
1194	timersub(&stop, &start0, &diff);
1195	td->runtime_ns = diff.tv_sec * 1000000000ULL;
1196	td->runtime_ns += diff.tv_usec * 1000ULL;
1197
1198	free_data(thread_data, g->p.bytes_thread);
1199
1200	pthread_mutex_lock(&g->stop_work_mutex);
1201	g->bytes_done += bytes_done;
1202	pthread_mutex_unlock(&g->stop_work_mutex);
1203
1204	return NULL;
1205}
1206
1207/*
1208 * A worker process starts a couple of threads:
1209 */
1210static void worker_process(int process_nr)
1211{
1212	pthread_mutex_t process_lock;
1213	struct thread_data *td;
1214	pthread_t *pthreads;
1215	u8 *process_data;
1216	int task_nr;
1217	int ret;
1218	int t;
1219
1220	pthread_mutex_init(&process_lock, NULL);
1221	set_taskname("process %d", process_nr);
1222
1223	/*
1224	 * Pick up the memory policy and the CPU binding of our first thread,
1225	 * so that we initialize memory accordingly:
1226	 */
1227	task_nr = process_nr*g->p.nr_threads;
1228	td = g->threads + task_nr;
1229
1230	bind_to_memnode(td->bind_node);
1231	bind_to_cpumask(td->bind_cpumask);
1232
1233	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1234	process_data = setup_private_data(g->p.bytes_process);
1235
1236	if (g->p.show_details >= 3) {
1237		printf(" # process %2d global mem: %p, process mem: %p\n",
1238			process_nr, g->data, process_data);
1239	}
1240
1241	for (t = 0; t < g->p.nr_threads; t++) {
1242		task_nr = process_nr*g->p.nr_threads + t;
1243		td = g->threads + task_nr;
1244
1245		td->process_data = process_data;
1246		td->process_nr   = process_nr;
1247		td->thread_nr    = t;
1248		td->task_nr	 = task_nr;
1249		td->val          = rand();
1250		td->curr_cpu	 = -1;
1251		td->process_lock = &process_lock;
1252
1253		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1254		BUG_ON(ret);
1255	}
1256
1257	for (t = 0; t < g->p.nr_threads; t++) {
1258                ret = pthread_join(pthreads[t], NULL);
1259		BUG_ON(ret);
1260	}
1261
1262	free_data(process_data, g->p.bytes_process);
1263	free(pthreads);
1264}
1265
1266static void print_summary(void)
1267{
1268	if (g->p.show_details < 0)
1269		return;
1270
1271	printf("\n ###\n");
1272	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1273		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1274	printf(" #      %5dx %5ldMB global  shared mem operations\n",
1275			g->p.nr_loops, g->p.bytes_global/1024/1024);
1276	printf(" #      %5dx %5ldMB process shared mem operations\n",
1277			g->p.nr_loops, g->p.bytes_process/1024/1024);
1278	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
1279			g->p.nr_loops, g->p.bytes_thread/1024/1024);
1280
1281	printf(" ###\n");
1282
1283	printf("\n ###\n"); fflush(stdout);
1284}
1285
1286static void init_thread_data(void)
1287{
1288	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1289	int t;
1290
1291	g->threads = zalloc_shared_data(size);
1292
1293	for (t = 0; t < g->p.nr_tasks; t++) {
1294		struct thread_data *td = g->threads + t;
1295		int cpu;
1296
1297		/* Allow all nodes by default: */
1298		td->bind_node = -1;
1299
1300		/* Allow all CPUs by default: */
1301		CPU_ZERO(&td->bind_cpumask);
1302		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1303			CPU_SET(cpu, &td->bind_cpumask);
1304	}
1305}
1306
1307static void deinit_thread_data(void)
1308{
1309	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1310
1311	free_data(g->threads, size);
1312}
1313
1314static int init(void)
1315{
1316	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1317
1318	/* Copy over options: */
1319	g->p = p0;
1320
1321	g->p.nr_cpus = numa_num_configured_cpus();
1322
1323	g->p.nr_nodes = numa_max_node() + 1;
1324
1325	/* char array in count_process_nodes(): */
1326	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1327
1328	if (g->p.show_quiet && !g->p.show_details)
1329		g->p.show_details = -1;
1330
1331	/* Some memory should be specified: */
1332	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1333		return -1;
1334
1335	if (g->p.mb_global_str) {
1336		g->p.mb_global = atof(g->p.mb_global_str);
1337		BUG_ON(g->p.mb_global < 0);
1338	}
1339
1340	if (g->p.mb_proc_str) {
1341		g->p.mb_proc = atof(g->p.mb_proc_str);
1342		BUG_ON(g->p.mb_proc < 0);
1343	}
1344
1345	if (g->p.mb_proc_locked_str) {
1346		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1347		BUG_ON(g->p.mb_proc_locked < 0);
1348		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1349	}
1350
1351	if (g->p.mb_thread_str) {
1352		g->p.mb_thread = atof(g->p.mb_thread_str);
1353		BUG_ON(g->p.mb_thread < 0);
1354	}
1355
1356	BUG_ON(g->p.nr_threads <= 0);
1357	BUG_ON(g->p.nr_proc <= 0);
1358
1359	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1360
1361	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
1362	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
1363	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
1364	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
1365
1366	g->data = setup_shared_data(g->p.bytes_global);
1367
1368	/* Startup serialization: */
1369	init_global_mutex(&g->start_work_mutex);
1370	init_global_mutex(&g->startup_mutex);
1371	init_global_mutex(&g->startup_done_mutex);
1372	init_global_mutex(&g->stop_work_mutex);
1373
1374	init_thread_data();
1375
1376	tprintf("#\n");
1377	if (parse_setup_cpu_list() || parse_setup_node_list())
1378		return -1;
1379	tprintf("#\n");
1380
1381	print_summary();
1382
1383	return 0;
1384}
1385
1386static void deinit(void)
1387{
1388	free_data(g->data, g->p.bytes_global);
1389	g->data = NULL;
1390
1391	deinit_thread_data();
1392
1393	free_data(g, sizeof(*g));
1394	g = NULL;
1395}
1396
1397/*
1398 * Print a short or long result, depending on the verbosity setting:
1399 */
1400static void print_res(const char *name, double val,
1401		      const char *txt_unit, const char *txt_short, const char *txt_long)
1402{
1403	if (!name)
1404		name = "main,";
1405
1406	if (!g->p.show_quiet)
1407		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1408	else
1409		printf(" %14.3f %s\n", val, txt_long);
1410}
1411
1412static int __bench_numa(const char *name)
1413{
1414	struct timeval start, stop, diff;
1415	u64 runtime_ns_min, runtime_ns_sum;
1416	pid_t *pids, pid, wpid;
1417	double delta_runtime;
1418	double runtime_avg;
1419	double runtime_sec_max;
1420	double runtime_sec_min;
1421	int wait_stat;
1422	double bytes;
1423	int i, t;
1424
1425	if (init())
1426		return -1;
1427
1428	pids = zalloc(g->p.nr_proc * sizeof(*pids));
1429	pid = -1;
1430
1431	/* All threads try to acquire it, this way we can wait for them to start up: */
1432	pthread_mutex_lock(&g->start_work_mutex);
1433
1434	if (g->p.serialize_startup) {
1435		tprintf(" #\n");
1436		tprintf(" # Startup synchronization: ..."); fflush(stdout);
1437	}
1438
1439	gettimeofday(&start, NULL);
1440
1441	for (i = 0; i < g->p.nr_proc; i++) {
1442		pid = fork();
1443		dprintf(" # process %2d: PID %d\n", i, pid);
1444
1445		BUG_ON(pid < 0);
1446		if (!pid) {
1447			/* Child process: */
1448			worker_process(i);
1449
1450			exit(0);
1451		}
1452		pids[i] = pid;
1453
1454	}
1455	/* Wait for all the threads to start up: */
1456	while (g->nr_tasks_started != g->p.nr_tasks)
1457		usleep(1000);
1458
1459	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1460
1461	if (g->p.serialize_startup) {
1462		double startup_sec;
1463
1464		pthread_mutex_lock(&g->startup_done_mutex);
1465
1466		/* This will start all threads: */
1467		pthread_mutex_unlock(&g->start_work_mutex);
1468
1469		/* This mutex is locked - the last started thread will wake us: */
1470		pthread_mutex_lock(&g->startup_done_mutex);
1471
1472		gettimeofday(&stop, NULL);
1473
1474		timersub(&stop, &start, &diff);
1475
1476		startup_sec = diff.tv_sec * 1000000000.0;
1477		startup_sec += diff.tv_usec * 1000.0;
1478		startup_sec /= 1e9;
1479
1480		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1481		tprintf(" #\n");
1482
1483		start = stop;
1484		pthread_mutex_unlock(&g->startup_done_mutex);
1485	} else {
1486		gettimeofday(&start, NULL);
1487	}
1488
1489	/* Parent process: */
1490
1491
1492	for (i = 0; i < g->p.nr_proc; i++) {
1493		wpid = waitpid(pids[i], &wait_stat, 0);
1494		BUG_ON(wpid < 0);
1495		BUG_ON(!WIFEXITED(wait_stat));
1496
1497	}
1498
1499	runtime_ns_sum = 0;
1500	runtime_ns_min = -1LL;
1501
1502	for (t = 0; t < g->p.nr_tasks; t++) {
1503		u64 thread_runtime_ns = g->threads[t].runtime_ns;
1504
1505		runtime_ns_sum += thread_runtime_ns;
1506		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1507	}
1508
1509	gettimeofday(&stop, NULL);
1510	timersub(&stop, &start, &diff);
1511
1512	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1513
1514	tprintf("\n ###\n");
1515	tprintf("\n");
1516
1517	runtime_sec_max = diff.tv_sec * 1000000000.0;
1518	runtime_sec_max += diff.tv_usec * 1000.0;
1519	runtime_sec_max /= 1e9;
1520
1521	runtime_sec_min = runtime_ns_min/1e9;
1522
1523	bytes = g->bytes_done;
1524	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1525
1526	if (g->p.measure_convergence) {
1527		print_res(name, runtime_sec_max,
1528			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1529	}
1530
1531	print_res(name, runtime_sec_max,
1532		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
1533
1534	print_res(name, runtime_sec_min,
1535		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
1536
1537	print_res(name, runtime_avg,
1538		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
1539
1540	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1541	print_res(name, delta_runtime / runtime_sec_max * 100.0,
1542		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
1543
1544	print_res(name, bytes / g->p.nr_tasks / 1e9,
1545		"GB,", "data/thread",		"GB data processed, per thread");
1546
1547	print_res(name, bytes / 1e9,
1548		"GB,", "data-total",		"GB data processed, total");
1549
1550	print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1551		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1552
1553	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1554		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
1555
1556	print_res(name, bytes / runtime_sec_max / 1e9,
1557		"GB/sec,", "total-speed",	"GB/sec total speed");
1558
1559	free(pids);
1560
1561	deinit();
1562
1563	return 0;
1564}
1565
1566#define MAX_ARGS 50
1567
1568static int command_size(const char **argv)
1569{
1570	int size = 0;
1571
1572	while (*argv) {
1573		size++;
1574		argv++;
1575	}
1576
1577	BUG_ON(size >= MAX_ARGS);
1578
1579	return size;
1580}
1581
1582static void init_params(struct params *p, const char *name, int argc, const char **argv)
1583{
1584	int i;
1585
1586	printf("\n # Running %s \"perf bench numa", name);
1587
1588	for (i = 0; i < argc; i++)
1589		printf(" %s", argv[i]);
1590
1591	printf("\"\n");
1592
1593	memset(p, 0, sizeof(*p));
1594
1595	/* Initialize nonzero defaults: */
1596
1597	p->serialize_startup		= 1;
1598	p->data_reads			= true;
1599	p->data_writes			= true;
1600	p->data_backwards		= true;
1601	p->data_rand_walk		= true;
1602	p->nr_loops			= -1;
1603	p->init_random			= true;
1604	p->mb_global_str		= "1";
1605	p->nr_proc			= 1;
1606	p->nr_threads			= 1;
1607	p->nr_secs			= 5;
1608	p->run_all			= argc == 1;
1609}
1610
1611static int run_bench_numa(const char *name, const char **argv)
1612{
1613	int argc = command_size(argv);
1614
1615	init_params(&p0, name, argc, argv);
1616	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1617	if (argc)
1618		goto err;
1619
1620	if (__bench_numa(name))
1621		goto err;
1622
1623	return 0;
1624
1625err:
1626	return -1;
1627}
1628
1629#define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
1630#define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
1631
1632#define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
1633#define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
1634
1635#define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
1636#define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
1637
1638/*
1639 * The built-in test-suite executed by "perf bench numa -a".
1640 *
1641 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1642 */
1643static const char *tests[][MAX_ARGS] = {
1644   /* Basic single-stream NUMA bandwidth measurements: */
1645   { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1646			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
1647   { "RAM-bw-local-NOTHP,",
1648			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1649			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
1650   { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
1651			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
1652
1653   /* 2-stream NUMA bandwidth measurements: */
1654   { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1655			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1656   { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1657		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1658
1659   /* Cross-stream NUMA bandwidth measurement: */
1660   { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
1661		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1662
1663   /* Convergence latency measurements: */
1664   { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
1665   { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
1666   { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
1667   { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1668   { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
1669   { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
1670   { " 4x4-convergence-NOTHP,",
1671			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1672   { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
1673   { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
1674   { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
1675   { " 8x4-convergence-NOTHP,",
1676			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
1677   { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
1678   { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
1679   { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
1680   { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
1681   { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
1682
1683   /* Various NUMA process/thread layout bandwidth measurements: */
1684   { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
1685   { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
1686   { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
1687   { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
1688   { " 8x1-bw-process-NOTHP,",
1689			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
1690   { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
1691
1692   { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
1693   { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
1694   { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
1695   { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
1696
1697   { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
1698   { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
1699   { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
1700   { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
1701   { " 4x8-bw-thread-NOTHP,",
1702			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
1703   { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
1704   { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
1705
1706   { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
1707   { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
1708
1709   { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
1710   { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
1711   { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
1712   { "numa01-bw-thread-NOTHP,",
1713			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
1714};
1715
1716static int bench_all(void)
1717{
1718	int nr = ARRAY_SIZE(tests);
1719	int ret;
1720	int i;
1721
1722	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1723	BUG_ON(ret < 0);
1724
1725	for (i = 0; i < nr; i++) {
1726		run_bench_numa(tests[i][0], tests[i] + 1);
1727	}
1728
1729	printf("\n");
1730
1731	return 0;
1732}
1733
1734int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1735{
1736	init_params(&p0, "main,", argc, argv);
1737	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1738	if (argc)
1739		goto err;
1740
1741	if (p0.run_all)
1742		return bench_all();
1743
1744	if (__bench_numa(NULL))
1745		goto err;
1746
1747	return 0;
1748
1749err:
1750	usage_with_options(numa_usage, options);
1751	return -1;
1752}
1753