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