1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * numa.c 3 * numa.c
4 * 4 *
5 * numa: Simulate NUMA-sensitive workload and 5 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
6 */ 6 */
7 7
8 #include <inttypes.h> 8 #include <inttypes.h>
>> 9 /* For the CLR_() macros */
>> 10 #include <pthread.h>
9 11
10 #include <subcmd/parse-options.h> 12 #include <subcmd/parse-options.h>
11 #include "../util/cloexec.h" 13 #include "../util/cloexec.h"
12 14
13 #include "bench.h" 15 #include "bench.h"
14 16
15 #include <errno.h> 17 #include <errno.h>
16 #include <sched.h> 18 #include <sched.h>
17 #include <stdio.h> 19 #include <stdio.h>
18 #include <assert.h> 20 #include <assert.h>
19 #include <debug.h> <<
20 #include <malloc.h> 21 #include <malloc.h>
21 #include <signal.h> 22 #include <signal.h>
22 #include <stdlib.h> 23 #include <stdlib.h>
23 #include <string.h> 24 #include <string.h>
24 #include <unistd.h> 25 #include <unistd.h>
25 #include <sys/mman.h> 26 #include <sys/mman.h>
26 #include <sys/time.h> 27 #include <sys/time.h>
27 #include <sys/resource.h> 28 #include <sys/resource.h>
28 #include <sys/wait.h> 29 #include <sys/wait.h>
29 #include <sys/prctl.h> 30 #include <sys/prctl.h>
30 #include <sys/types.h> 31 #include <sys/types.h>
31 #include <linux/kernel.h> 32 #include <linux/kernel.h>
32 #include <linux/time64.h> 33 #include <linux/time64.h>
33 #include <linux/numa.h> 34 #include <linux/numa.h>
34 #include <linux/zalloc.h> 35 #include <linux/zalloc.h>
35 36
36 #include "../util/header.h" 37 #include "../util/header.h"
37 #include "../util/mutex.h" <<
38 #include <numa.h> 38 #include <numa.h>
39 #include <numaif.h> 39 #include <numaif.h>
40 40
41 #ifndef RUSAGE_THREAD 41 #ifndef RUSAGE_THREAD
42 # define RUSAGE_THREAD 1 42 # define RUSAGE_THREAD 1
43 #endif 43 #endif
44 44
45 /* 45 /*
46 * Regular printout to the terminal, suppresse 46 * Regular printout to the terminal, suppressed if -q is specified:
47 */ 47 */
48 #define tprintf(x...) do { if (g && g->p.show_ 48 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
49 49
50 /* 50 /*
51 * Debug printf: 51 * Debug printf:
52 */ 52 */
53 #undef dprintf 53 #undef dprintf
54 #define dprintf(x...) do { if (g && g->p.show_ 54 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
55 55
56 struct thread_data { 56 struct thread_data {
57 int curr_cpu; 57 int curr_cpu;
58 cpu_set_t *bind_cpumask; 58 cpu_set_t *bind_cpumask;
59 int bind_node; 59 int bind_node;
60 u8 *process_data; 60 u8 *process_data;
61 int process_nr; 61 int process_nr;
62 int thread_nr; 62 int thread_nr;
63 int task_nr; 63 int task_nr;
64 unsigned int loops_done; 64 unsigned int loops_done;
65 u64 val; 65 u64 val;
66 u64 runtime_ns; 66 u64 runtime_ns;
67 u64 system_time_ns 67 u64 system_time_ns;
68 u64 user_time_ns; 68 u64 user_time_ns;
69 double speed_gbs; 69 double speed_gbs;
70 struct mutex *process_lock; !! 70 pthread_mutex_t *process_lock;
71 }; 71 };
72 72
73 /* Parameters set by options: */ 73 /* Parameters set by options: */
74 74
75 struct params { 75 struct params {
76 /* Startup synchronization: */ 76 /* Startup synchronization: */
77 bool serialize_star 77 bool serialize_startup;
78 78
79 /* Task hierarchy: */ 79 /* Task hierarchy: */
80 int nr_proc; 80 int nr_proc;
81 int nr_threads; 81 int nr_threads;
82 82
83 /* Working set sizes: */ 83 /* Working set sizes: */
84 const char *mb_global_str 84 const char *mb_global_str;
85 const char *mb_proc_str; 85 const char *mb_proc_str;
86 const char *mb_proc_locke 86 const char *mb_proc_locked_str;
87 const char *mb_thread_str 87 const char *mb_thread_str;
88 88
89 double mb_global; 89 double mb_global;
90 double mb_proc; 90 double mb_proc;
91 double mb_proc_locked 91 double mb_proc_locked;
92 double mb_thread; 92 double mb_thread;
93 93
94 /* Access patterns to the working set: 94 /* Access patterns to the working set: */
95 bool data_reads; 95 bool data_reads;
96 bool data_writes; 96 bool data_writes;
97 bool data_backwards 97 bool data_backwards;
98 bool data_zero_mems 98 bool data_zero_memset;
99 bool data_rand_walk 99 bool data_rand_walk;
100 u32 nr_loops; 100 u32 nr_loops;
101 u32 nr_secs; 101 u32 nr_secs;
102 u32 sleep_usecs; 102 u32 sleep_usecs;
103 103
104 /* Working set initialization: */ 104 /* Working set initialization: */
105 bool init_zero; 105 bool init_zero;
106 bool init_random; 106 bool init_random;
107 bool init_cpu0; 107 bool init_cpu0;
108 108
109 /* Misc options: */ 109 /* Misc options: */
110 int show_details; 110 int show_details;
111 int run_all; 111 int run_all;
112 int thp; 112 int thp;
113 113
114 long bytes_global; 114 long bytes_global;
115 long bytes_process; 115 long bytes_process;
116 long bytes_process_ 116 long bytes_process_locked;
117 long bytes_thread; 117 long bytes_thread;
118 118
119 int nr_tasks; 119 int nr_tasks;
>> 120 bool show_quiet;
120 121
121 bool show_convergen 122 bool show_convergence;
122 bool measure_conver 123 bool measure_convergence;
123 124
124 int perturb_secs; 125 int perturb_secs;
125 int nr_cpus; 126 int nr_cpus;
126 int nr_nodes; 127 int nr_nodes;
127 128
128 /* Affinity options -C and -N: */ 129 /* Affinity options -C and -N: */
129 char *cpu_list_str; 130 char *cpu_list_str;
130 char *node_list_str 131 char *node_list_str;
131 }; 132 };
132 133
133 134
134 /* Global, read-writable area, accessible to a 135 /* Global, read-writable area, accessible to all processes and threads: */
135 136
136 struct global_info { 137 struct global_info {
137 u8 *data; 138 u8 *data;
138 139
139 struct mutex startup_mutex; !! 140 pthread_mutex_t startup_mutex;
140 struct cond startup_cond; !! 141 pthread_cond_t startup_cond;
141 int nr_tasks_start 142 int nr_tasks_started;
142 143
143 struct mutex start_work_mut !! 144 pthread_mutex_t start_work_mutex;
144 struct cond start_work_con !! 145 pthread_cond_t start_work_cond;
145 int nr_tasks_worki 146 int nr_tasks_working;
146 bool start_work; 147 bool start_work;
147 148
148 struct mutex stop_work_mute !! 149 pthread_mutex_t stop_work_mutex;
149 u64 bytes_done; 150 u64 bytes_done;
150 151
151 struct thread_data *threads; 152 struct thread_data *threads;
152 153
153 /* Convergence latency measurement: */ 154 /* Convergence latency measurement: */
154 bool all_converged; 155 bool all_converged;
155 bool stop_work; 156 bool stop_work;
156 157
157 int print_once; 158 int print_once;
158 159
159 struct params p; 160 struct params p;
160 }; 161 };
161 162
162 static struct global_info *g = NULL; 163 static struct global_info *g = NULL;
163 164
164 static int parse_cpus_opt(const struct option 165 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
165 static int parse_nodes_opt(const struct option 166 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
166 167
167 struct params p0; 168 struct params p0;
168 169
169 static const struct option options[] = { 170 static const struct option options[] = {
170 OPT_INTEGER('p', "nr_proc" , &p0. 171 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
171 OPT_INTEGER('t', "nr_threads" , &p0. 172 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
172 173
173 OPT_STRING('G', "mb_global" , &p0. 174 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
174 OPT_STRING('P', "mb_proc" , &p0. 175 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
175 OPT_STRING('L', "mb_proc_locked", &p0. 176 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
176 OPT_STRING('T', "mb_thread" , &p0. 177 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
177 178
178 OPT_UINTEGER('l', "nr_loops" , &p0. 179 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
179 OPT_UINTEGER('s', "nr_secs" , &p0. 180 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
180 OPT_UINTEGER('u', "usleep" , &p0. 181 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
181 182
182 OPT_BOOLEAN('R', "data_reads" , &p0. 183 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
183 OPT_BOOLEAN('W', "data_writes" , &p0. 184 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
184 OPT_BOOLEAN('B', "data_backwards", &p0 185 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
185 OPT_BOOLEAN('Z', "data_zero_memset", & 186 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
186 OPT_BOOLEAN('r', "data_rand_walk", &p0 187 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
187 188
188 189
189 OPT_BOOLEAN('z', "init_zero" , &p0. 190 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
190 OPT_BOOLEAN('I', "init_random" , &p0. 191 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
191 OPT_BOOLEAN('', "init_cpu0" , &p0.i 192 OPT_BOOLEAN('', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
192 OPT_INTEGER('x', "perturb_secs", &p0.p 193 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
193 194
194 OPT_INCR ('d', "show_details" , &p0. 195 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
195 OPT_INCR ('a', "all" , &p0. 196 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
196 OPT_INTEGER('H', "thp" , &p0. 197 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
197 OPT_BOOLEAN('c', "show_convergence", & 198 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
198 "convergence is reached wh 199 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
199 OPT_BOOLEAN('m', "measure_convergence" 200 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
200 OPT_BOOLEAN('q', "quiet" , &qui !! 201 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
201 "quiet mode (do not show a <<
202 OPT_BOOLEAN('S', "serialize-startup", 202 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
203 203
204 /* Special option string parsing callb 204 /* Special option string parsing callbacks: */
205 OPT_CALLBACK('C', "cpus", NULL, "cpu[, 205 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
206 "bind the first N task 206 "bind the first N tasks to these specific cpus (the rest is unbound)",
207 parse_cpus_opt), 207 parse_cpus_opt),
208 OPT_CALLBACK('M', "memnodes", NULL, "n 208 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
209 "bind the first N task 209 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
210 parse_nodes_opt), 210 parse_nodes_opt),
211 OPT_END() 211 OPT_END()
212 }; 212 };
213 213
214 static const char * const bench_numa_usage[] = 214 static const char * const bench_numa_usage[] = {
215 "perf bench numa <options>", 215 "perf bench numa <options>",
216 NULL 216 NULL
217 }; 217 };
218 218
219 static const char * const numa_usage[] = { 219 static const char * const numa_usage[] = {
220 "perf bench numa mem [<options>]", 220 "perf bench numa mem [<options>]",
221 NULL 221 NULL
222 }; 222 };
223 223
224 /* 224 /*
225 * To get number of numa nodes present. 225 * To get number of numa nodes present.
226 */ 226 */
227 static int nr_numa_nodes(void) 227 static int nr_numa_nodes(void)
228 { 228 {
229 int i, nr_nodes = 0; 229 int i, nr_nodes = 0;
230 230
231 for (i = 0; i < g->p.nr_nodes; i++) { 231 for (i = 0; i < g->p.nr_nodes; i++) {
232 if (numa_bitmask_isbitset(numa 232 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
233 nr_nodes++; 233 nr_nodes++;
234 } 234 }
235 235
236 return nr_nodes; 236 return nr_nodes;
237 } 237 }
238 238
239 /* 239 /*
240 * To check if given numa node is present. 240 * To check if given numa node is present.
241 */ 241 */
242 static int is_node_present(int node) 242 static int is_node_present(int node)
243 { 243 {
244 return numa_bitmask_isbitset(numa_node 244 return numa_bitmask_isbitset(numa_nodes_ptr, node);
245 } 245 }
246 246
247 /* 247 /*
248 * To check given numa node has cpus. 248 * To check given numa node has cpus.
249 */ 249 */
250 static bool node_has_cpus(int node) 250 static bool node_has_cpus(int node)
251 { 251 {
252 struct bitmask *cpumask = numa_allocat 252 struct bitmask *cpumask = numa_allocate_cpumask();
253 bool ret = false; /* fall back to nocp 253 bool ret = false; /* fall back to nocpus */
254 int cpu; 254 int cpu;
255 255
256 BUG_ON(!cpumask); 256 BUG_ON(!cpumask);
257 if (!numa_node_to_cpus(node, cpumask)) 257 if (!numa_node_to_cpus(node, cpumask)) {
258 for (cpu = 0; cpu < (int)cpuma 258 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
259 if (numa_bitmask_isbit 259 if (numa_bitmask_isbitset(cpumask, cpu)) {
260 ret = true; 260 ret = true;
261 break; 261 break;
262 } 262 }
263 } 263 }
264 } 264 }
265 numa_free_cpumask(cpumask); 265 numa_free_cpumask(cpumask);
266 266
267 return ret; 267 return ret;
268 } 268 }
269 269
270 static cpu_set_t *bind_to_cpu(int target_cpu) 270 static cpu_set_t *bind_to_cpu(int target_cpu)
271 { 271 {
272 int nrcpus = numa_num_possible_cpus(); 272 int nrcpus = numa_num_possible_cpus();
273 cpu_set_t *orig_mask, *mask; 273 cpu_set_t *orig_mask, *mask;
274 size_t size; 274 size_t size;
275 275
276 orig_mask = CPU_ALLOC(nrcpus); 276 orig_mask = CPU_ALLOC(nrcpus);
277 BUG_ON(!orig_mask); 277 BUG_ON(!orig_mask);
278 size = CPU_ALLOC_SIZE(nrcpus); 278 size = CPU_ALLOC_SIZE(nrcpus);
279 CPU_ZERO_S(size, orig_mask); 279 CPU_ZERO_S(size, orig_mask);
280 280
281 if (sched_getaffinity(0, size, orig_ma 281 if (sched_getaffinity(0, size, orig_mask))
282 goto err_out; 282 goto err_out;
283 283
284 mask = CPU_ALLOC(nrcpus); 284 mask = CPU_ALLOC(nrcpus);
285 if (!mask) 285 if (!mask)
286 goto err_out; 286 goto err_out;
287 287
288 CPU_ZERO_S(size, mask); 288 CPU_ZERO_S(size, mask);
289 289
290 if (target_cpu == -1) { 290 if (target_cpu == -1) {
291 int cpu; 291 int cpu;
292 292
293 for (cpu = 0; cpu < g->p.nr_cp 293 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
294 CPU_SET_S(cpu, size, m 294 CPU_SET_S(cpu, size, mask);
295 } else { 295 } else {
296 if (target_cpu < 0 || target_c 296 if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
297 goto err; 297 goto err;
298 298
299 CPU_SET_S(target_cpu, size, ma 299 CPU_SET_S(target_cpu, size, mask);
300 } 300 }
301 301
302 if (sched_setaffinity(0, size, mask)) 302 if (sched_setaffinity(0, size, mask))
303 goto err; 303 goto err;
304 304
305 return orig_mask; 305 return orig_mask;
306 306
307 err: 307 err:
308 CPU_FREE(mask); 308 CPU_FREE(mask);
309 err_out: 309 err_out:
310 CPU_FREE(orig_mask); 310 CPU_FREE(orig_mask);
311 311
312 /* BUG_ON due to failure in allocation 312 /* BUG_ON due to failure in allocation of orig_mask/mask */
313 BUG_ON(-1); 313 BUG_ON(-1);
314 return NULL; 314 return NULL;
315 } 315 }
316 316
317 static cpu_set_t *bind_to_node(int target_node 317 static cpu_set_t *bind_to_node(int target_node)
318 { 318 {
319 int nrcpus = numa_num_possible_cpus(); 319 int nrcpus = numa_num_possible_cpus();
320 size_t size; 320 size_t size;
321 cpu_set_t *orig_mask, *mask; 321 cpu_set_t *orig_mask, *mask;
322 int cpu; 322 int cpu;
323 323
324 orig_mask = CPU_ALLOC(nrcpus); 324 orig_mask = CPU_ALLOC(nrcpus);
325 BUG_ON(!orig_mask); 325 BUG_ON(!orig_mask);
326 size = CPU_ALLOC_SIZE(nrcpus); 326 size = CPU_ALLOC_SIZE(nrcpus);
327 CPU_ZERO_S(size, orig_mask); 327 CPU_ZERO_S(size, orig_mask);
328 328
329 if (sched_getaffinity(0, size, orig_ma 329 if (sched_getaffinity(0, size, orig_mask))
330 goto err_out; 330 goto err_out;
331 331
332 mask = CPU_ALLOC(nrcpus); 332 mask = CPU_ALLOC(nrcpus);
333 if (!mask) 333 if (!mask)
334 goto err_out; 334 goto err_out;
335 335
336 CPU_ZERO_S(size, mask); 336 CPU_ZERO_S(size, mask);
337 337
338 if (target_node == NUMA_NO_NODE) { 338 if (target_node == NUMA_NO_NODE) {
339 for (cpu = 0; cpu < g->p.nr_cp 339 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
340 CPU_SET_S(cpu, size, m 340 CPU_SET_S(cpu, size, mask);
341 } else { 341 } else {
342 struct bitmask *cpumask = numa 342 struct bitmask *cpumask = numa_allocate_cpumask();
343 343
344 if (!cpumask) 344 if (!cpumask)
345 goto err; 345 goto err;
346 346
347 if (!numa_node_to_cpus(target_ 347 if (!numa_node_to_cpus(target_node, cpumask)) {
348 for (cpu = 0; cpu < (i 348 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
349 if (numa_bitma 349 if (numa_bitmask_isbitset(cpumask, cpu))
350 CPU_SE 350 CPU_SET_S(cpu, size, mask);
351 } 351 }
352 } 352 }
353 numa_free_cpumask(cpumask); 353 numa_free_cpumask(cpumask);
354 } 354 }
355 355
356 if (sched_setaffinity(0, size, mask)) 356 if (sched_setaffinity(0, size, mask))
357 goto err; 357 goto err;
358 358
359 return orig_mask; 359 return orig_mask;
360 360
361 err: 361 err:
362 CPU_FREE(mask); 362 CPU_FREE(mask);
363 err_out: 363 err_out:
364 CPU_FREE(orig_mask); 364 CPU_FREE(orig_mask);
365 365
366 /* BUG_ON due to failure in allocation 366 /* BUG_ON due to failure in allocation of orig_mask/mask */
367 BUG_ON(-1); 367 BUG_ON(-1);
368 return NULL; 368 return NULL;
369 } 369 }
370 370
371 static void bind_to_cpumask(cpu_set_t *mask) 371 static void bind_to_cpumask(cpu_set_t *mask)
372 { 372 {
373 int ret; 373 int ret;
374 size_t size = CPU_ALLOC_SIZE(numa_num_ 374 size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
375 375
376 ret = sched_setaffinity(0, size, mask) 376 ret = sched_setaffinity(0, size, mask);
377 if (ret) { 377 if (ret) {
378 CPU_FREE(mask); 378 CPU_FREE(mask);
379 BUG_ON(ret); 379 BUG_ON(ret);
380 } 380 }
381 } 381 }
382 382
383 static void mempol_restore(void) 383 static void mempol_restore(void)
384 { 384 {
385 int ret; 385 int ret;
386 386
387 ret = set_mempolicy(MPOL_DEFAULT, NULL 387 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
388 388
389 BUG_ON(ret); 389 BUG_ON(ret);
390 } 390 }
391 391
392 static void bind_to_memnode(int node) 392 static void bind_to_memnode(int node)
393 { 393 {
394 struct bitmask *node_mask; 394 struct bitmask *node_mask;
395 int ret; 395 int ret;
396 396
397 if (node == NUMA_NO_NODE) 397 if (node == NUMA_NO_NODE)
398 return; 398 return;
399 399
400 node_mask = numa_allocate_nodemask(); 400 node_mask = numa_allocate_nodemask();
401 BUG_ON(!node_mask); 401 BUG_ON(!node_mask);
402 402
403 numa_bitmask_clearall(node_mask); 403 numa_bitmask_clearall(node_mask);
404 numa_bitmask_setbit(node_mask, node); 404 numa_bitmask_setbit(node_mask, node);
405 405
406 ret = set_mempolicy(MPOL_BIND, node_ma 406 ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
407 dprintf("binding to node %d, mask: %01 407 dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
408 408
409 numa_bitmask_free(node_mask); 409 numa_bitmask_free(node_mask);
410 BUG_ON(ret); 410 BUG_ON(ret);
411 } 411 }
412 412
413 #define HPSIZE (2*1024*1024) 413 #define HPSIZE (2*1024*1024)
414 414
415 #define set_taskname(fmt...) 415 #define set_taskname(fmt...) \
416 do { 416 do { \
417 char name[20]; 417 char name[20]; \
418 418 \
419 snprintf(name, 20, fmt); 419 snprintf(name, 20, fmt); \
420 prctl(PR_SET_NAME, name); 420 prctl(PR_SET_NAME, name); \
421 } while (0) 421 } while (0)
422 422
423 static u8 *alloc_data(ssize_t bytes0, int map_ 423 static u8 *alloc_data(ssize_t bytes0, int map_flags,
424 int init_zero, int init_ 424 int init_zero, int init_cpu0, int thp, int init_random)
425 { 425 {
426 cpu_set_t *orig_mask = NULL; 426 cpu_set_t *orig_mask = NULL;
427 ssize_t bytes; 427 ssize_t bytes;
428 u8 *buf; 428 u8 *buf;
429 int ret; 429 int ret;
430 430
431 if (!bytes0) 431 if (!bytes0)
432 return NULL; 432 return NULL;
433 433
434 /* Allocate and initialize all memory 434 /* Allocate and initialize all memory on CPU#0: */
435 if (init_cpu0) { 435 if (init_cpu0) {
436 int node = numa_node_of_cpu(0) 436 int node = numa_node_of_cpu(0);
437 437
438 orig_mask = bind_to_node(node) 438 orig_mask = bind_to_node(node);
439 bind_to_memnode(node); 439 bind_to_memnode(node);
440 } 440 }
441 441
442 bytes = bytes0 + HPSIZE; 442 bytes = bytes0 + HPSIZE;
443 443
444 buf = (void *)mmap(0, bytes, PROT_READ 444 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
445 BUG_ON(buf == (void *)-1); 445 BUG_ON(buf == (void *)-1);
446 446
447 if (map_flags == MAP_PRIVATE) { 447 if (map_flags == MAP_PRIVATE) {
448 if (thp > 0) { 448 if (thp > 0) {
449 ret = madvise(buf, byt 449 ret = madvise(buf, bytes, MADV_HUGEPAGE);
450 if (ret && !g->print_o 450 if (ret && !g->print_once) {
451 g->print_once 451 g->print_once = 1;
452 printf("WARNIN 452 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
453 } 453 }
454 } 454 }
455 if (thp < 0) { 455 if (thp < 0) {
456 ret = madvise(buf, byt 456 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
457 if (ret && !g->print_o 457 if (ret && !g->print_once) {
458 g->print_once 458 g->print_once = 1;
459 printf("WARNIN 459 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
460 } 460 }
461 } 461 }
462 } 462 }
463 463
464 if (init_zero) { 464 if (init_zero) {
465 bzero(buf, bytes); 465 bzero(buf, bytes);
466 } else { 466 } else {
467 /* Initialize random contents, 467 /* Initialize random contents, different in each word: */
468 if (init_random) { 468 if (init_random) {
469 u64 *wbuf = (void *)bu 469 u64 *wbuf = (void *)buf;
470 long off = rand(); 470 long off = rand();
471 long i; 471 long i;
472 472
473 for (i = 0; i < bytes/ 473 for (i = 0; i < bytes/8; i++)
474 wbuf[i] = i + 474 wbuf[i] = i + off;
475 } 475 }
476 } 476 }
477 477
478 /* Align to 2MB boundary: */ 478 /* Align to 2MB boundary: */
479 buf = (void *)(((unsigned long)buf + H 479 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
480 480
481 /* Restore affinity: */ 481 /* Restore affinity: */
482 if (init_cpu0) { 482 if (init_cpu0) {
483 bind_to_cpumask(orig_mask); 483 bind_to_cpumask(orig_mask);
484 CPU_FREE(orig_mask); 484 CPU_FREE(orig_mask);
485 mempol_restore(); 485 mempol_restore();
486 } 486 }
487 487
488 return buf; 488 return buf;
489 } 489 }
490 490
491 static void free_data(void *data, ssize_t byte 491 static void free_data(void *data, ssize_t bytes)
492 { 492 {
493 int ret; 493 int ret;
494 494
495 if (!data) 495 if (!data)
496 return; 496 return;
497 497
498 ret = munmap(data, bytes); 498 ret = munmap(data, bytes);
499 BUG_ON(ret); 499 BUG_ON(ret);
500 } 500 }
501 501
502 /* 502 /*
503 * Create a shared memory buffer that can be s 503 * Create a shared memory buffer that can be shared between processes, zeroed:
504 */ 504 */
505 static void * zalloc_shared_data(ssize_t bytes 505 static void * zalloc_shared_data(ssize_t bytes)
506 { 506 {
507 return alloc_data(bytes, MAP_SHARED, 1 507 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
508 } 508 }
509 509
510 /* 510 /*
511 * Create a shared memory buffer that can be s 511 * Create a shared memory buffer that can be shared between processes:
512 */ 512 */
513 static void * setup_shared_data(ssize_t bytes) 513 static void * setup_shared_data(ssize_t bytes)
514 { 514 {
515 return alloc_data(bytes, MAP_SHARED, 0 515 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
516 } 516 }
517 517
518 /* 518 /*
519 * Allocate process-local memory - this will e 519 * Allocate process-local memory - this will either be shared between
520 * threads of this process, or only be accesse 520 * threads of this process, or only be accessed by this thread:
521 */ 521 */
522 static void * setup_private_data(ssize_t bytes 522 static void * setup_private_data(ssize_t bytes)
523 { 523 {
524 return alloc_data(bytes, MAP_PRIVATE, 524 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
525 } 525 }
526 526
>> 527 /*
>> 528 * Return a process-shared (global) mutex:
>> 529 */
>> 530 static void init_global_mutex(pthread_mutex_t *mutex)
>> 531 {
>> 532 pthread_mutexattr_t attr;
>> 533
>> 534 pthread_mutexattr_init(&attr);
>> 535 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
>> 536 pthread_mutex_init(mutex, &attr);
>> 537 }
>> 538
>> 539 /*
>> 540 * Return a process-shared (global) condition variable:
>> 541 */
>> 542 static void init_global_cond(pthread_cond_t *cond)
>> 543 {
>> 544 pthread_condattr_t attr;
>> 545
>> 546 pthread_condattr_init(&attr);
>> 547 pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
>> 548 pthread_cond_init(cond, &attr);
>> 549 }
>> 550
527 static int parse_cpu_list(const char *arg) 551 static int parse_cpu_list(const char *arg)
528 { 552 {
529 p0.cpu_list_str = strdup(arg); 553 p0.cpu_list_str = strdup(arg);
530 554
531 dprintf("got CPU list: {%s}\n", p0.cpu 555 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
532 556
533 return 0; 557 return 0;
534 } 558 }
535 559
536 static int parse_setup_cpu_list(void) 560 static int parse_setup_cpu_list(void)
537 { 561 {
538 struct thread_data *td; 562 struct thread_data *td;
539 char *str0, *str; 563 char *str0, *str;
540 int t; 564 int t;
541 565
542 if (!g->p.cpu_list_str) 566 if (!g->p.cpu_list_str)
543 return 0; 567 return 0;
544 568
545 dprintf("g->p.nr_tasks: %d\n", g->p.nr 569 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
546 570
547 str0 = str = strdup(g->p.cpu_list_str) 571 str0 = str = strdup(g->p.cpu_list_str);
548 t = 0; 572 t = 0;
549 573
550 BUG_ON(!str); 574 BUG_ON(!str);
551 575
552 tprintf("# binding tasks to CPUs:\n"); 576 tprintf("# binding tasks to CPUs:\n");
553 tprintf("# "); 577 tprintf("# ");
554 578
555 while (true) { 579 while (true) {
556 int bind_cpu, bind_cpu_0, bind 580 int bind_cpu, bind_cpu_0, bind_cpu_1;
557 char *tok, *tok_end, *tok_step 581 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
558 int bind_len; 582 int bind_len;
559 int step; 583 int step;
560 int mul; 584 int mul;
561 585
562 tok = strsep(&str, ","); 586 tok = strsep(&str, ",");
563 if (!tok) 587 if (!tok)
564 break; 588 break;
565 589
566 tok_end = strstr(tok, "-"); 590 tok_end = strstr(tok, "-");
567 591
568 dprintf("\ntoken: {%s}, end: { 592 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
569 if (!tok_end) { 593 if (!tok_end) {
570 /* Single CPU specifie 594 /* Single CPU specified: */
571 bind_cpu_0 = bind_cpu_ 595 bind_cpu_0 = bind_cpu_1 = atol(tok);
572 } else { 596 } else {
573 /* CPU range specified 597 /* CPU range specified (for example: "5-11"): */
574 bind_cpu_0 = atol(tok) 598 bind_cpu_0 = atol(tok);
575 bind_cpu_1 = atol(tok_ 599 bind_cpu_1 = atol(tok_end + 1);
576 } 600 }
577 601
578 step = 1; 602 step = 1;
579 tok_step = strstr(tok, "#"); 603 tok_step = strstr(tok, "#");
580 if (tok_step) { 604 if (tok_step) {
581 step = atol(tok_step + 605 step = atol(tok_step + 1);
582 BUG_ON(step <= 0 || st 606 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
583 } 607 }
584 608
585 /* 609 /*
586 * Mask length. 610 * Mask length.
587 * Eg: "--cpus 8_4-16#4" means 611 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
588 * where the _4 means the next 612 * where the _4 means the next 4 CPUs are allowed.
589 */ 613 */
590 bind_len = 1; 614 bind_len = 1;
591 tok_len = strstr(tok, "_"); 615 tok_len = strstr(tok, "_");
592 if (tok_len) { 616 if (tok_len) {
593 bind_len = atol(tok_le 617 bind_len = atol(tok_len + 1);
594 BUG_ON(bind_len <= 0 | 618 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
595 } 619 }
596 620
597 /* Multiplicator shortcut, "0x 621 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
598 mul = 1; 622 mul = 1;
599 tok_mul = strstr(tok, "x"); 623 tok_mul = strstr(tok, "x");
600 if (tok_mul) { 624 if (tok_mul) {
601 mul = atol(tok_mul + 1 625 mul = atol(tok_mul + 1);
602 BUG_ON(mul <= 0); 626 BUG_ON(mul <= 0);
603 } 627 }
604 628
605 dprintf("CPUs: %d_%d-%d#%dx%d\ 629 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
606 630
607 if (bind_cpu_0 >= g->p.nr_cpus 631 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
608 printf("\nTest not app 632 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
609 return -1; 633 return -1;
610 } 634 }
611 635
612 if (is_cpu_online(bind_cpu_0) 636 if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
613 printf("\nTest not app 637 printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
614 return -1; 638 return -1;
615 } 639 }
616 640
617 BUG_ON(bind_cpu_0 < 0 || bind_ 641 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
618 BUG_ON(bind_cpu_0 > bind_cpu_1 642 BUG_ON(bind_cpu_0 > bind_cpu_1);
619 643
620 for (bind_cpu = bind_cpu_0; bi 644 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
621 size_t size = CPU_ALLO 645 size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
622 int i; 646 int i;
623 647
624 for (i = 0; i < mul; i 648 for (i = 0; i < mul; i++) {
625 int cpu; 649 int cpu;
626 650
627 if (t >= g->p. 651 if (t >= g->p.nr_tasks) {
628 printf 652 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
629 goto o 653 goto out;
630 } 654 }
631 td = g->thread 655 td = g->threads + t;
632 656
633 if (t) 657 if (t)
634 tprint 658 tprintf(",");
635 if (bind_len > 659 if (bind_len > 1) {
636 tprint 660 tprintf("%2d/%d", bind_cpu, bind_len);
637 } else { 661 } else {
638 tprint 662 tprintf("%2d", bind_cpu);
639 } 663 }
640 664
641 td->bind_cpuma 665 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
642 BUG_ON(!td->bi 666 BUG_ON(!td->bind_cpumask);
643 CPU_ZERO_S(siz 667 CPU_ZERO_S(size, td->bind_cpumask);
644 for (cpu = bin 668 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
645 if (cp 669 if (cpu < 0 || cpu >= g->p.nr_cpus) {
646 670 CPU_FREE(td->bind_cpumask);
647 671 BUG_ON(-1);
648 } 672 }
649 CPU_SE 673 CPU_SET_S(cpu, size, td->bind_cpumask);
650 } 674 }
651 t++; 675 t++;
652 } 676 }
653 } 677 }
654 } 678 }
655 out: 679 out:
656 680
657 tprintf("\n"); 681 tprintf("\n");
658 682
659 if (t < g->p.nr_tasks) 683 if (t < g->p.nr_tasks)
660 printf("# NOTE: %d tasks bound 684 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
661 685
662 free(str0); 686 free(str0);
663 return 0; 687 return 0;
664 } 688 }
665 689
666 static int parse_cpus_opt(const struct option 690 static int parse_cpus_opt(const struct option *opt __maybe_unused,
667 const char *arg, int 691 const char *arg, int unset __maybe_unused)
668 { 692 {
669 if (!arg) 693 if (!arg)
670 return -1; 694 return -1;
671 695
672 return parse_cpu_list(arg); 696 return parse_cpu_list(arg);
673 } 697 }
674 698
675 static int parse_node_list(const char *arg) 699 static int parse_node_list(const char *arg)
676 { 700 {
677 p0.node_list_str = strdup(arg); 701 p0.node_list_str = strdup(arg);
678 702
679 dprintf("got NODE list: {%s}\n", p0.no 703 dprintf("got NODE list: {%s}\n", p0.node_list_str);
680 704
681 return 0; 705 return 0;
682 } 706 }
683 707
684 static int parse_setup_node_list(void) 708 static int parse_setup_node_list(void)
685 { 709 {
686 struct thread_data *td; 710 struct thread_data *td;
687 char *str0, *str; 711 char *str0, *str;
688 int t; 712 int t;
689 713
690 if (!g->p.node_list_str) 714 if (!g->p.node_list_str)
691 return 0; 715 return 0;
692 716
693 dprintf("g->p.nr_tasks: %d\n", g->p.nr 717 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
694 718
695 str0 = str = strdup(g->p.node_list_str 719 str0 = str = strdup(g->p.node_list_str);
696 t = 0; 720 t = 0;
697 721
698 BUG_ON(!str); 722 BUG_ON(!str);
699 723
700 tprintf("# binding tasks to NODEs:\n") 724 tprintf("# binding tasks to NODEs:\n");
701 tprintf("# "); 725 tprintf("# ");
702 726
703 while (true) { 727 while (true) {
704 int bind_node, bind_node_0, bi 728 int bind_node, bind_node_0, bind_node_1;
705 char *tok, *tok_end, *tok_step 729 char *tok, *tok_end, *tok_step, *tok_mul;
706 int step; 730 int step;
707 int mul; 731 int mul;
708 732
709 tok = strsep(&str, ","); 733 tok = strsep(&str, ",");
710 if (!tok) 734 if (!tok)
711 break; 735 break;
712 736
713 tok_end = strstr(tok, "-"); 737 tok_end = strstr(tok, "-");
714 738
715 dprintf("\ntoken: {%s}, end: { 739 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
716 if (!tok_end) { 740 if (!tok_end) {
717 /* Single NODE specifi 741 /* Single NODE specified: */
718 bind_node_0 = bind_nod 742 bind_node_0 = bind_node_1 = atol(tok);
719 } else { 743 } else {
720 /* NODE range specifie 744 /* NODE range specified (for example: "5-11"): */
721 bind_node_0 = atol(tok 745 bind_node_0 = atol(tok);
722 bind_node_1 = atol(tok 746 bind_node_1 = atol(tok_end + 1);
723 } 747 }
724 748
725 step = 1; 749 step = 1;
726 tok_step = strstr(tok, "#"); 750 tok_step = strstr(tok, "#");
727 if (tok_step) { 751 if (tok_step) {
728 step = atol(tok_step + 752 step = atol(tok_step + 1);
729 BUG_ON(step <= 0 || st 753 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
730 } 754 }
731 755
732 /* Multiplicator shortcut, "0x 756 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
733 mul = 1; 757 mul = 1;
734 tok_mul = strstr(tok, "x"); 758 tok_mul = strstr(tok, "x");
735 if (tok_mul) { 759 if (tok_mul) {
736 mul = atol(tok_mul + 1 760 mul = atol(tok_mul + 1);
737 BUG_ON(mul <= 0); 761 BUG_ON(mul <= 0);
738 } 762 }
739 763
740 dprintf("NODEs: %d-%d #%d\n", 764 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
741 765
742 if (bind_node_0 >= g->p.nr_nod 766 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
743 printf("\nTest not app 767 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
744 return -1; 768 return -1;
745 } 769 }
746 770
747 BUG_ON(bind_node_0 < 0 || bind 771 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
748 BUG_ON(bind_node_0 > bind_node 772 BUG_ON(bind_node_0 > bind_node_1);
749 773
750 for (bind_node = bind_node_0; 774 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
751 int i; 775 int i;
752 776
753 for (i = 0; i < mul; i 777 for (i = 0; i < mul; i++) {
754 if (t >= g->p. 778 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
755 printf 779 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
756 goto o 780 goto out;
757 } 781 }
758 td = g->thread 782 td = g->threads + t;
759 783
760 if (!t) 784 if (!t)
761 tprint 785 tprintf(" %2d", bind_node);
762 else 786 else
763 tprint 787 tprintf(",%2d", bind_node);
764 788
765 td->bind_node 789 td->bind_node = bind_node;
766 t++; 790 t++;
767 } 791 }
768 } 792 }
769 } 793 }
770 out: 794 out:
771 795
772 tprintf("\n"); 796 tprintf("\n");
773 797
774 if (t < g->p.nr_tasks) 798 if (t < g->p.nr_tasks)
775 printf("# NOTE: %d tasks mem-b 799 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
776 800
777 free(str0); 801 free(str0);
778 return 0; 802 return 0;
779 } 803 }
780 804
781 static int parse_nodes_opt(const struct option 805 static int parse_nodes_opt(const struct option *opt __maybe_unused,
782 const char *arg, int 806 const char *arg, int unset __maybe_unused)
783 { 807 {
784 if (!arg) 808 if (!arg)
785 return -1; 809 return -1;
786 810
787 return parse_node_list(arg); 811 return parse_node_list(arg);
788 } 812 }
789 813
790 static inline uint32_t lfsr_32(uint32_t lfsr) 814 static inline uint32_t lfsr_32(uint32_t lfsr)
791 { 815 {
792 const uint32_t taps = BIT(1) | BIT(5) 816 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
793 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x 817 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
794 } 818 }
795 819
796 /* 820 /*
797 * Make sure there's real data dependency to R 821 * Make sure there's real data dependency to RAM (when read
798 * accesses are enabled), so the compiler, the 822 * accesses are enabled), so the compiler, the CPU and the
799 * kernel (KSM, zero page, etc.) cannot optimi 823 * kernel (KSM, zero page, etc.) cannot optimize away RAM
800 * accesses: 824 * accesses:
801 */ 825 */
802 static inline u64 access_data(u64 *data, u64 v 826 static inline u64 access_data(u64 *data, u64 val)
803 { 827 {
804 if (g->p.data_reads) 828 if (g->p.data_reads)
805 val += *data; 829 val += *data;
806 if (g->p.data_writes) 830 if (g->p.data_writes)
807 *data = val + 1; 831 *data = val + 1;
808 return val; 832 return val;
809 } 833 }
810 834
811 /* 835 /*
812 * The worker process does two types of work, 836 * The worker process does two types of work, a forwards going
813 * loop and a backwards going loop. 837 * loop and a backwards going loop.
814 * 838 *
815 * We do this so that on multiprocessor system 839 * We do this so that on multiprocessor systems we do not create
816 * a 'train' of processing, with highly synchr 840 * a 'train' of processing, with highly synchronized processes,
817 * skewing the whole benchmark. 841 * skewing the whole benchmark.
818 */ 842 */
819 static u64 do_work(u8 *__data, long bytes, int 843 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
820 { 844 {
821 long words = bytes/sizeof(u64); 845 long words = bytes/sizeof(u64);
822 u64 *data = (void *)__data; 846 u64 *data = (void *)__data;
823 long chunk_0, chunk_1; 847 long chunk_0, chunk_1;
824 u64 *d0, *d, *d1; 848 u64 *d0, *d, *d1;
825 long off; 849 long off;
826 long i; 850 long i;
827 851
828 BUG_ON(!data && words); 852 BUG_ON(!data && words);
829 BUG_ON(data && !words); 853 BUG_ON(data && !words);
830 854
831 if (!data) 855 if (!data)
832 return val; 856 return val;
833 857
834 /* Very simple memset() work variant: 858 /* Very simple memset() work variant: */
835 if (g->p.data_zero_memset && !g->p.dat 859 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
836 bzero(data, bytes); 860 bzero(data, bytes);
837 return val; 861 return val;
838 } 862 }
839 863
840 /* Spread out by PID/TID nr and by loo 864 /* Spread out by PID/TID nr and by loop nr: */
841 chunk_0 = words/nr_max; 865 chunk_0 = words/nr_max;
842 chunk_1 = words/g->p.nr_loops; 866 chunk_1 = words/g->p.nr_loops;
843 off = nr*chunk_0 + loop*chunk_1; 867 off = nr*chunk_0 + loop*chunk_1;
844 868
845 while (off >= words) 869 while (off >= words)
846 off -= words; 870 off -= words;
847 871
848 if (g->p.data_rand_walk) { 872 if (g->p.data_rand_walk) {
849 u32 lfsr = nr + loop + val; 873 u32 lfsr = nr + loop + val;
850 long j; !! 874 int j;
851 875
852 for (i = 0; i < words/1024; i+ 876 for (i = 0; i < words/1024; i++) {
853 long start, end; 877 long start, end;
854 878
855 lfsr = lfsr_32(lfsr); 879 lfsr = lfsr_32(lfsr);
856 880
857 start = lfsr % words; 881 start = lfsr % words;
858 end = min(start + 1024 882 end = min(start + 1024, words-1);
859 883
860 if (g->p.data_zero_mem 884 if (g->p.data_zero_memset) {
861 bzero(data + s 885 bzero(data + start, (end-start) * sizeof(u64));
862 } else { 886 } else {
863 for (j = start 887 for (j = start; j < end; j++)
864 val = 888 val = access_data(data + j, val);
865 } 889 }
866 } 890 }
867 } else if (!g->p.data_backwards || (nr 891 } else if (!g->p.data_backwards || (nr + loop) & 1) {
868 /* Process data forwards: */ 892 /* Process data forwards: */
869 893
870 d0 = data + off; 894 d0 = data + off;
871 d = data + off + 1; 895 d = data + off + 1;
872 d1 = data + words; 896 d1 = data + words;
873 897
874 for (;;) { 898 for (;;) {
875 if (unlikely(d >= d1)) 899 if (unlikely(d >= d1))
876 d = data; 900 d = data;
877 if (unlikely(d == d0)) 901 if (unlikely(d == d0))
878 break; 902 break;
879 903
880 val = access_data(d, v 904 val = access_data(d, val);
881 905
882 d++; 906 d++;
883 } 907 }
884 } else { 908 } else {
885 /* Process data backwards: */ 909 /* Process data backwards: */
886 910
887 d0 = data + off; 911 d0 = data + off;
888 d = data + off - 1; 912 d = data + off - 1;
889 d1 = data + words; 913 d1 = data + words;
890 914
891 for (;;) { 915 for (;;) {
892 if (unlikely(d < data) 916 if (unlikely(d < data))
893 d = data + wor 917 d = data + words-1;
894 if (unlikely(d == d0)) 918 if (unlikely(d == d0))
895 break; 919 break;
896 920
897 val = access_data(d, v 921 val = access_data(d, val);
898 922
899 d--; 923 d--;
900 } 924 }
901 } 925 }
902 926
903 return val; 927 return val;
904 } 928 }
905 929
906 static void update_curr_cpu(int task_nr, unsig 930 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
907 { 931 {
908 unsigned int cpu; 932 unsigned int cpu;
909 933
910 cpu = sched_getcpu(); 934 cpu = sched_getcpu();
911 935
912 g->threads[task_nr].curr_cpu = cpu; 936 g->threads[task_nr].curr_cpu = cpu;
913 prctl(0, bytes_worked); 937 prctl(0, bytes_worked);
914 } 938 }
915 939
916 /* 940 /*
917 * Count the number of nodes a process's threa 941 * Count the number of nodes a process's threads
918 * are spread out on. 942 * are spread out on.
919 * 943 *
920 * A count of 1 means that the process is comp 944 * A count of 1 means that the process is compressed
921 * to a single node. A count of g->p.nr_nodes 945 * to a single node. A count of g->p.nr_nodes means it's
922 * spread out on the whole system. 946 * spread out on the whole system.
923 */ 947 */
924 static int count_process_nodes(int process_nr) 948 static int count_process_nodes(int process_nr)
925 { 949 {
926 char *node_present; 950 char *node_present;
927 int nodes; 951 int nodes;
928 int n, t; 952 int n, t;
929 953
930 node_present = (char *)malloc(g->p.nr_ 954 node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
931 BUG_ON(!node_present); 955 BUG_ON(!node_present);
932 for (nodes = 0; nodes < g->p.nr_nodes; 956 for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
933 node_present[nodes] = 0; 957 node_present[nodes] = 0;
934 958
935 for (t = 0; t < g->p.nr_threads; t++) 959 for (t = 0; t < g->p.nr_threads; t++) {
936 struct thread_data *td; 960 struct thread_data *td;
937 int task_nr; 961 int task_nr;
938 int node; 962 int node;
939 963
940 task_nr = process_nr*g->p.nr_t 964 task_nr = process_nr*g->p.nr_threads + t;
941 td = g->threads + task_nr; 965 td = g->threads + task_nr;
942 966
943 node = numa_node_of_cpu(td->cu 967 node = numa_node_of_cpu(td->curr_cpu);
944 if (node < 0) /* curr_cpu was 968 if (node < 0) /* curr_cpu was likely still -1 */ {
945 free(node_present); 969 free(node_present);
946 return 0; 970 return 0;
947 } 971 }
948 972
949 node_present[node] = 1; 973 node_present[node] = 1;
950 } 974 }
951 975
952 nodes = 0; 976 nodes = 0;
953 977
954 for (n = 0; n < g->p.nr_nodes; n++) 978 for (n = 0; n < g->p.nr_nodes; n++)
955 nodes += node_present[n]; 979 nodes += node_present[n];
956 980
957 free(node_present); 981 free(node_present);
958 return nodes; 982 return nodes;
959 } 983 }
960 984
961 /* 985 /*
962 * Count the number of distinct process-thread 986 * Count the number of distinct process-threads a node contains.
963 * 987 *
964 * A count of 1 means that the node contains o 988 * A count of 1 means that the node contains only a single
965 * process. If all nodes on the system contain 989 * process. If all nodes on the system contain at most one
966 * process then we are well-converged. 990 * process then we are well-converged.
967 */ 991 */
968 static int count_node_processes(int node) 992 static int count_node_processes(int node)
969 { 993 {
970 int processes = 0; 994 int processes = 0;
971 int t, p; 995 int t, p;
972 996
973 for (p = 0; p < g->p.nr_proc; p++) { 997 for (p = 0; p < g->p.nr_proc; p++) {
974 for (t = 0; t < g->p.nr_thread 998 for (t = 0; t < g->p.nr_threads; t++) {
975 struct thread_data *td 999 struct thread_data *td;
976 int task_nr; 1000 int task_nr;
977 int n; 1001 int n;
978 1002
979 task_nr = p*g->p.nr_th 1003 task_nr = p*g->p.nr_threads + t;
980 td = g->threads + task 1004 td = g->threads + task_nr;
981 1005
982 n = numa_node_of_cpu(t 1006 n = numa_node_of_cpu(td->curr_cpu);
983 if (n == node) { 1007 if (n == node) {
984 processes++; 1008 processes++;
985 break; 1009 break;
986 } 1010 }
987 } 1011 }
988 } 1012 }
989 1013
990 return processes; 1014 return processes;
991 } 1015 }
992 1016
993 static void calc_convergence_compression(int * 1017 static void calc_convergence_compression(int *strong)
994 { 1018 {
995 unsigned int nodes_min, nodes_max; 1019 unsigned int nodes_min, nodes_max;
996 int p; 1020 int p;
997 1021
998 nodes_min = -1; 1022 nodes_min = -1;
999 nodes_max = 0; 1023 nodes_max = 0;
1000 1024
1001 for (p = 0; p < g->p.nr_proc; p++) { 1025 for (p = 0; p < g->p.nr_proc; p++) {
1002 unsigned int nodes = count_pr 1026 unsigned int nodes = count_process_nodes(p);
1003 1027
1004 if (!nodes) { 1028 if (!nodes) {
1005 *strong = 0; 1029 *strong = 0;
1006 return; 1030 return;
1007 } 1031 }
1008 1032
1009 nodes_min = min(nodes, nodes_ 1033 nodes_min = min(nodes, nodes_min);
1010 nodes_max = max(nodes, nodes_ 1034 nodes_max = max(nodes, nodes_max);
1011 } 1035 }
1012 1036
1013 /* Strong convergence: all threads co 1037 /* Strong convergence: all threads compress on a single node: */
1014 if (nodes_min == 1 && nodes_max == 1) 1038 if (nodes_min == 1 && nodes_max == 1) {
1015 *strong = 1; 1039 *strong = 1;
1016 } else { 1040 } else {
1017 *strong = 0; 1041 *strong = 0;
1018 tprintf(" {%d-%d}", nodes_min 1042 tprintf(" {%d-%d}", nodes_min, nodes_max);
1019 } 1043 }
1020 } 1044 }
1021 1045
1022 static void calc_convergence(double runtime_n 1046 static void calc_convergence(double runtime_ns_max, double *convergence)
1023 { 1047 {
1024 unsigned int loops_done_min, loops_do 1048 unsigned int loops_done_min, loops_done_max;
1025 int process_groups; 1049 int process_groups;
1026 int *nodes; 1050 int *nodes;
1027 int distance; 1051 int distance;
1028 int nr_min; 1052 int nr_min;
1029 int nr_max; 1053 int nr_max;
1030 int strong; 1054 int strong;
1031 int sum; 1055 int sum;
1032 int nr; 1056 int nr;
1033 int node; 1057 int node;
1034 int cpu; 1058 int cpu;
1035 int t; 1059 int t;
1036 1060
1037 if (!g->p.show_convergence && !g->p.m 1061 if (!g->p.show_convergence && !g->p.measure_convergence)
1038 return; 1062 return;
1039 1063
1040 nodes = (int *)malloc(g->p.nr_nodes * 1064 nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1041 BUG_ON(!nodes); 1065 BUG_ON(!nodes);
1042 for (node = 0; node < g->p.nr_nodes; 1066 for (node = 0; node < g->p.nr_nodes; node++)
1043 nodes[node] = 0; 1067 nodes[node] = 0;
1044 1068
1045 loops_done_min = -1; 1069 loops_done_min = -1;
1046 loops_done_max = 0; 1070 loops_done_max = 0;
1047 1071
1048 for (t = 0; t < g->p.nr_tasks; t++) { 1072 for (t = 0; t < g->p.nr_tasks; t++) {
1049 struct thread_data *td = g->t 1073 struct thread_data *td = g->threads + t;
1050 unsigned int loops_done; 1074 unsigned int loops_done;
1051 1075
1052 cpu = td->curr_cpu; 1076 cpu = td->curr_cpu;
1053 1077
1054 /* Not all threads have writt 1078 /* Not all threads have written it yet: */
1055 if (cpu < 0) 1079 if (cpu < 0)
1056 continue; 1080 continue;
1057 1081
1058 node = numa_node_of_cpu(cpu); 1082 node = numa_node_of_cpu(cpu);
1059 1083
1060 nodes[node]++; 1084 nodes[node]++;
1061 1085
1062 loops_done = td->loops_done; 1086 loops_done = td->loops_done;
1063 loops_done_min = min(loops_do 1087 loops_done_min = min(loops_done, loops_done_min);
1064 loops_done_max = max(loops_do 1088 loops_done_max = max(loops_done, loops_done_max);
1065 } 1089 }
1066 1090
1067 nr_max = 0; 1091 nr_max = 0;
1068 nr_min = g->p.nr_tasks; 1092 nr_min = g->p.nr_tasks;
1069 sum = 0; 1093 sum = 0;
1070 1094
1071 for (node = 0; node < g->p.nr_nodes; 1095 for (node = 0; node < g->p.nr_nodes; node++) {
1072 if (!is_node_present(node)) 1096 if (!is_node_present(node))
1073 continue; 1097 continue;
1074 nr = nodes[node]; 1098 nr = nodes[node];
1075 nr_min = min(nr, nr_min); 1099 nr_min = min(nr, nr_min);
1076 nr_max = max(nr, nr_max); 1100 nr_max = max(nr, nr_max);
1077 sum += nr; 1101 sum += nr;
1078 } 1102 }
1079 BUG_ON(nr_min > nr_max); 1103 BUG_ON(nr_min > nr_max);
1080 1104
1081 BUG_ON(sum > g->p.nr_tasks); 1105 BUG_ON(sum > g->p.nr_tasks);
1082 1106
1083 if (0 && (sum < g->p.nr_tasks)) { 1107 if (0 && (sum < g->p.nr_tasks)) {
1084 free(nodes); 1108 free(nodes);
1085 return; 1109 return;
1086 } 1110 }
1087 1111
1088 /* 1112 /*
1089 * Count the number of distinct proce 1113 * Count the number of distinct process groups present
1090 * on nodes - when we are converged t 1114 * on nodes - when we are converged this will decrease
1091 * to g->p.nr_proc: 1115 * to g->p.nr_proc:
1092 */ 1116 */
1093 process_groups = 0; 1117 process_groups = 0;
1094 1118
1095 for (node = 0; node < g->p.nr_nodes; 1119 for (node = 0; node < g->p.nr_nodes; node++) {
1096 int processes; 1120 int processes;
1097 1121
1098 if (!is_node_present(node)) 1122 if (!is_node_present(node))
1099 continue; 1123 continue;
1100 processes = count_node_proces 1124 processes = count_node_processes(node);
1101 nr = nodes[node]; 1125 nr = nodes[node];
1102 tprintf(" %2d/%-2d", nr, proc 1126 tprintf(" %2d/%-2d", nr, processes);
1103 1127
1104 process_groups += processes; 1128 process_groups += processes;
1105 } 1129 }
1106 1130
1107 distance = nr_max - nr_min; 1131 distance = nr_max - nr_min;
1108 1132
1109 tprintf(" [%2d/%-2d]", distance, proc 1133 tprintf(" [%2d/%-2d]", distance, process_groups);
1110 1134
1111 tprintf(" l:%3d-%-3d (%3d)", 1135 tprintf(" l:%3d-%-3d (%3d)",
1112 loops_done_min, loops_done_ma 1136 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1113 1137
1114 if (loops_done_min && loops_done_max) 1138 if (loops_done_min && loops_done_max) {
1115 double skew = 1.0 - (double)l 1139 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1116 1140
1117 tprintf(" [%4.1f%%]", skew * 1141 tprintf(" [%4.1f%%]", skew * 100.0);
1118 } 1142 }
1119 1143
1120 calc_convergence_compression(&strong) 1144 calc_convergence_compression(&strong);
1121 1145
1122 if (strong && process_groups == g->p. 1146 if (strong && process_groups == g->p.nr_proc) {
1123 if (!*convergence) { 1147 if (!*convergence) {
1124 *convergence = runtim 1148 *convergence = runtime_ns_max;
1125 tprintf(" (%6.1fs con 1149 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1126 if (g->p.measure_conv 1150 if (g->p.measure_convergence) {
1127 g->all_conver 1151 g->all_converged = true;
1128 g->stop_work 1152 g->stop_work = true;
1129 } 1153 }
1130 } 1154 }
1131 } else { 1155 } else {
1132 if (*convergence) { 1156 if (*convergence) {
1133 tprintf(" (%6.1fs de- 1157 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1134 *convergence = 0; 1158 *convergence = 0;
1135 } 1159 }
1136 tprintf("\n"); 1160 tprintf("\n");
1137 } 1161 }
1138 1162
1139 free(nodes); 1163 free(nodes);
1140 } 1164 }
1141 1165
1142 static void show_summary(double runtime_ns_ma 1166 static void show_summary(double runtime_ns_max, int l, double *convergence)
1143 { 1167 {
1144 tprintf("\r # %5.1f%% [%.1f mins]", 1168 tprintf("\r # %5.1f%% [%.1f mins]",
1145 (double)(l+1)/g->p.nr_loops*1 1169 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1146 1170
1147 calc_convergence(runtime_ns_max, conv 1171 calc_convergence(runtime_ns_max, convergence);
1148 1172
1149 if (g->p.show_details >= 0) 1173 if (g->p.show_details >= 0)
1150 fflush(stdout); 1174 fflush(stdout);
1151 } 1175 }
1152 1176
1153 static void *worker_thread(void *__tdata) 1177 static void *worker_thread(void *__tdata)
1154 { 1178 {
1155 struct thread_data *td = __tdata; 1179 struct thread_data *td = __tdata;
1156 struct timeval start0, start, stop, d 1180 struct timeval start0, start, stop, diff;
1157 int process_nr = td->process_nr; 1181 int process_nr = td->process_nr;
1158 int thread_nr = td->thread_nr; 1182 int thread_nr = td->thread_nr;
1159 unsigned long last_perturbance; 1183 unsigned long last_perturbance;
1160 int task_nr = td->task_nr; 1184 int task_nr = td->task_nr;
1161 int details = g->p.show_details; 1185 int details = g->p.show_details;
1162 int first_task, last_task; 1186 int first_task, last_task;
1163 double convergence = 0; 1187 double convergence = 0;
1164 u64 val = td->val; 1188 u64 val = td->val;
1165 double runtime_ns_max; 1189 double runtime_ns_max;
1166 u8 *global_data; 1190 u8 *global_data;
1167 u8 *process_data; 1191 u8 *process_data;
1168 u8 *thread_data; 1192 u8 *thread_data;
1169 u64 bytes_done, secs; 1193 u64 bytes_done, secs;
1170 long work_done; 1194 long work_done;
1171 u32 l; 1195 u32 l;
1172 struct rusage rusage; 1196 struct rusage rusage;
1173 1197
1174 bind_to_cpumask(td->bind_cpumask); 1198 bind_to_cpumask(td->bind_cpumask);
1175 bind_to_memnode(td->bind_node); 1199 bind_to_memnode(td->bind_node);
1176 1200
1177 set_taskname("thread %d/%d", process_ 1201 set_taskname("thread %d/%d", process_nr, thread_nr);
1178 1202
1179 global_data = g->data; 1203 global_data = g->data;
1180 process_data = td->process_data; 1204 process_data = td->process_data;
1181 thread_data = setup_private_data(g->p 1205 thread_data = setup_private_data(g->p.bytes_thread);
1182 1206
1183 bytes_done = 0; 1207 bytes_done = 0;
1184 1208
1185 last_task = 0; 1209 last_task = 0;
1186 if (process_nr == g->p.nr_proc-1 && t 1210 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1187 last_task = 1; 1211 last_task = 1;
1188 1212
1189 first_task = 0; 1213 first_task = 0;
1190 if (process_nr == 0 && thread_nr == 0 1214 if (process_nr == 0 && thread_nr == 0)
1191 first_task = 1; 1215 first_task = 1;
1192 1216
1193 if (details >= 2) { 1217 if (details >= 2) {
1194 printf("# thread %2d / %2d g 1218 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1195 process_nr, thread_nr 1219 process_nr, thread_nr, global_data, process_data, thread_data);
1196 } 1220 }
1197 1221
1198 if (g->p.serialize_startup) { 1222 if (g->p.serialize_startup) {
1199 mutex_lock(&g->startup_mutex) !! 1223 pthread_mutex_lock(&g->startup_mutex);
1200 g->nr_tasks_started++; 1224 g->nr_tasks_started++;
1201 /* The last thread wakes the 1225 /* The last thread wakes the main process. */
1202 if (g->nr_tasks_started == g- 1226 if (g->nr_tasks_started == g->p.nr_tasks)
1203 cond_signal(&g->start !! 1227 pthread_cond_signal(&g->startup_cond);
1204 1228
1205 mutex_unlock(&g->startup_mute !! 1229 pthread_mutex_unlock(&g->startup_mutex);
1206 1230
1207 /* Here we will wait for the 1231 /* Here we will wait for the main process to start us all at once: */
1208 mutex_lock(&g->start_work_mut !! 1232 pthread_mutex_lock(&g->start_work_mutex);
1209 g->start_work = false; 1233 g->start_work = false;
1210 g->nr_tasks_working++; 1234 g->nr_tasks_working++;
1211 while (!g->start_work) 1235 while (!g->start_work)
1212 cond_wait(&g->start_w !! 1236 pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
1213 1237
1214 mutex_unlock(&g->start_work_m !! 1238 pthread_mutex_unlock(&g->start_work_mutex);
1215 } 1239 }
1216 1240
1217 gettimeofday(&start0, NULL); 1241 gettimeofday(&start0, NULL);
1218 1242
1219 start = stop = start0; 1243 start = stop = start0;
1220 last_perturbance = start.tv_sec; 1244 last_perturbance = start.tv_sec;
1221 1245
1222 for (l = 0; l < g->p.nr_loops; l++) { 1246 for (l = 0; l < g->p.nr_loops; l++) {
1223 start = stop; 1247 start = stop;
1224 1248
1225 if (g->stop_work) 1249 if (g->stop_work)
1226 break; 1250 break;
1227 1251
1228 val += do_work(global_data, 1252 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1229 val += do_work(process_data, 1253 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1230 val += do_work(thread_data, 1254 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1231 1255
1232 if (g->p.sleep_usecs) { 1256 if (g->p.sleep_usecs) {
1233 mutex_lock(td->proces !! 1257 pthread_mutex_lock(td->process_lock);
1234 usleep(g->p.sleep_use 1258 usleep(g->p.sleep_usecs);
1235 mutex_unlock(td->proc !! 1259 pthread_mutex_unlock(td->process_lock);
1236 } 1260 }
1237 /* 1261 /*
1238 * Amount of work to be done 1262 * Amount of work to be done under a process-global lock:
1239 */ 1263 */
1240 if (g->p.bytes_process_locked 1264 if (g->p.bytes_process_locked) {
1241 mutex_lock(td->proces !! 1265 pthread_mutex_lock(td->process_lock);
1242 val += do_work(proces 1266 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1243 mutex_unlock(td->proc !! 1267 pthread_mutex_unlock(td->process_lock);
1244 } 1268 }
1245 1269
1246 work_done = g->p.bytes_global 1270 work_done = g->p.bytes_global + g->p.bytes_process +
1247 g->p.bytes_proces 1271 g->p.bytes_process_locked + g->p.bytes_thread;
1248 1272
1249 update_curr_cpu(task_nr, work 1273 update_curr_cpu(task_nr, work_done);
1250 bytes_done += work_done; 1274 bytes_done += work_done;
1251 1275
1252 if (details < 0 && !g->p.pert 1276 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1253 continue; 1277 continue;
1254 1278
1255 td->loops_done = l; 1279 td->loops_done = l;
1256 1280
1257 gettimeofday(&stop, NULL); 1281 gettimeofday(&stop, NULL);
1258 1282
1259 /* Check whether our max runt 1283 /* Check whether our max runtime timed out: */
1260 if (g->p.nr_secs) { 1284 if (g->p.nr_secs) {
1261 timersub(&stop, &star 1285 timersub(&stop, &start0, &diff);
1262 if ((u32)diff.tv_sec 1286 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1263 g->stop_work 1287 g->stop_work = true;
1264 break; 1288 break;
1265 } 1289 }
1266 } 1290 }
1267 1291
1268 /* Update the summary at most 1292 /* Update the summary at most once per second: */
1269 if (start.tv_sec == stop.tv_s 1293 if (start.tv_sec == stop.tv_sec)
1270 continue; 1294 continue;
1271 1295
1272 /* 1296 /*
1273 * Perturb the first task's e 1297 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1274 * by migrating to CPU#0: 1298 * by migrating to CPU#0:
1275 */ 1299 */
1276 if (first_task && g->p.pertur 1300 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1277 cpu_set_t *orig_mask; 1301 cpu_set_t *orig_mask;
1278 int target_cpu; 1302 int target_cpu;
1279 int this_cpu; 1303 int this_cpu;
1280 1304
1281 last_perturbance = st 1305 last_perturbance = stop.tv_sec;
1282 1306
1283 /* 1307 /*
1284 * Depending on where 1308 * Depending on where we are running, move into
1285 * the other half of 1309 * the other half of the system, to create some
1286 * real disturbance: 1310 * real disturbance:
1287 */ 1311 */
1288 this_cpu = g->threads 1312 this_cpu = g->threads[task_nr].curr_cpu;
1289 if (this_cpu < g->p.n 1313 if (this_cpu < g->p.nr_cpus/2)
1290 target_cpu = 1314 target_cpu = g->p.nr_cpus-1;
1291 else 1315 else
1292 target_cpu = 1316 target_cpu = 0;
1293 1317
1294 orig_mask = bind_to_c 1318 orig_mask = bind_to_cpu(target_cpu);
1295 1319
1296 /* Here we are runnin 1320 /* Here we are running on the target CPU already */
1297 if (details >= 1) 1321 if (details >= 1)
1298 printf(" (inj 1322 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1299 1323
1300 bind_to_cpumask(orig_ 1324 bind_to_cpumask(orig_mask);
1301 CPU_FREE(orig_mask); 1325 CPU_FREE(orig_mask);
1302 } 1326 }
1303 1327
1304 if (details >= 3) { 1328 if (details >= 3) {
1305 timersub(&stop, &star 1329 timersub(&stop, &start, &diff);
1306 runtime_ns_max = diff 1330 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1307 runtime_ns_max += dif 1331 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1308 1332
1309 if (details >= 0) { 1333 if (details >= 0) {
1310 printf(" #%2d 1334 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1311 proce 1335 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1312 } 1336 }
1313 fflush(stdout); 1337 fflush(stdout);
1314 } 1338 }
1315 if (!last_task) 1339 if (!last_task)
1316 continue; 1340 continue;
1317 1341
1318 timersub(&stop, &start0, &dif 1342 timersub(&stop, &start0, &diff);
1319 runtime_ns_max = diff.tv_sec 1343 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1320 runtime_ns_max += diff.tv_use 1344 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1321 1345
1322 show_summary(runtime_ns_max, 1346 show_summary(runtime_ns_max, l, &convergence);
1323 } 1347 }
1324 1348
1325 gettimeofday(&stop, NULL); 1349 gettimeofday(&stop, NULL);
1326 timersub(&stop, &start0, &diff); 1350 timersub(&stop, &start0, &diff);
1327 td->runtime_ns = diff.tv_sec * NSEC_P 1351 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1328 td->runtime_ns += diff.tv_usec * NSEC 1352 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1329 secs = td->runtime_ns / NSEC_PER_SEC; 1353 secs = td->runtime_ns / NSEC_PER_SEC;
1330 td->speed_gbs = secs ? bytes_done / s 1354 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1331 1355
1332 getrusage(RUSAGE_THREAD, &rusage); 1356 getrusage(RUSAGE_THREAD, &rusage);
1333 td->system_time_ns = rusage.ru_stime. 1357 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1334 td->system_time_ns += rusage.ru_stime 1358 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1335 td->user_time_ns = rusage.ru_utime.tv 1359 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1336 td->user_time_ns += rusage.ru_utime.t 1360 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1337 1361
1338 free_data(thread_data, g->p.bytes_thr 1362 free_data(thread_data, g->p.bytes_thread);
1339 1363
1340 mutex_lock(&g->stop_work_mutex); !! 1364 pthread_mutex_lock(&g->stop_work_mutex);
1341 g->bytes_done += bytes_done; 1365 g->bytes_done += bytes_done;
1342 mutex_unlock(&g->stop_work_mutex); !! 1366 pthread_mutex_unlock(&g->stop_work_mutex);
1343 1367
1344 return NULL; 1368 return NULL;
1345 } 1369 }
1346 1370
1347 /* 1371 /*
1348 * A worker process starts a couple of thread 1372 * A worker process starts a couple of threads:
1349 */ 1373 */
1350 static void worker_process(int process_nr) 1374 static void worker_process(int process_nr)
1351 { 1375 {
1352 struct mutex process_lock; !! 1376 pthread_mutex_t process_lock;
1353 struct thread_data *td; 1377 struct thread_data *td;
1354 pthread_t *pthreads; 1378 pthread_t *pthreads;
1355 u8 *process_data; 1379 u8 *process_data;
1356 int task_nr; 1380 int task_nr;
1357 int ret; 1381 int ret;
1358 int t; 1382 int t;
1359 1383
1360 mutex_init(&process_lock); !! 1384 pthread_mutex_init(&process_lock, NULL);
1361 set_taskname("process %d", process_nr 1385 set_taskname("process %d", process_nr);
1362 1386
1363 /* 1387 /*
1364 * Pick up the memory policy and the 1388 * Pick up the memory policy and the CPU binding of our first thread,
1365 * so that we initialize memory accor 1389 * so that we initialize memory accordingly:
1366 */ 1390 */
1367 task_nr = process_nr*g->p.nr_threads; 1391 task_nr = process_nr*g->p.nr_threads;
1368 td = g->threads + task_nr; 1392 td = g->threads + task_nr;
1369 1393
1370 bind_to_memnode(td->bind_node); 1394 bind_to_memnode(td->bind_node);
1371 bind_to_cpumask(td->bind_cpumask); 1395 bind_to_cpumask(td->bind_cpumask);
1372 1396
1373 pthreads = zalloc(g->p.nr_threads * s 1397 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1374 process_data = setup_private_data(g-> 1398 process_data = setup_private_data(g->p.bytes_process);
1375 1399
1376 if (g->p.show_details >= 3) { 1400 if (g->p.show_details >= 3) {
1377 printf(" # process %2d global 1401 printf(" # process %2d global mem: %p, process mem: %p\n",
1378 process_nr, g->data, 1402 process_nr, g->data, process_data);
1379 } 1403 }
1380 1404
1381 for (t = 0; t < g->p.nr_threads; t++) 1405 for (t = 0; t < g->p.nr_threads; t++) {
1382 task_nr = process_nr*g->p.nr_ 1406 task_nr = process_nr*g->p.nr_threads + t;
1383 td = g->threads + task_nr; 1407 td = g->threads + task_nr;
1384 1408
1385 td->process_data = process_da 1409 td->process_data = process_data;
1386 td->process_nr = process_nr 1410 td->process_nr = process_nr;
1387 td->thread_nr = t; 1411 td->thread_nr = t;
1388 td->task_nr = task_nr; 1412 td->task_nr = task_nr;
1389 td->val = rand(); 1413 td->val = rand();
1390 td->curr_cpu = -1; 1414 td->curr_cpu = -1;
1391 td->process_lock = &process_l 1415 td->process_lock = &process_lock;
1392 1416
1393 ret = pthread_create(pthreads 1417 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1394 BUG_ON(ret); 1418 BUG_ON(ret);
1395 } 1419 }
1396 1420
1397 for (t = 0; t < g->p.nr_threads; t++) 1421 for (t = 0; t < g->p.nr_threads; t++) {
1398 ret = pthread_join(pthreads[t 1422 ret = pthread_join(pthreads[t], NULL);
1399 BUG_ON(ret); 1423 BUG_ON(ret);
1400 } 1424 }
1401 1425
1402 free_data(process_data, g->p.bytes_pr 1426 free_data(process_data, g->p.bytes_process);
1403 free(pthreads); 1427 free(pthreads);
1404 } 1428 }
1405 1429
1406 static void print_summary(void) 1430 static void print_summary(void)
1407 { 1431 {
1408 if (g->p.show_details < 0) 1432 if (g->p.show_details < 0)
1409 return; 1433 return;
1410 1434
1411 printf("\n ###\n"); 1435 printf("\n ###\n");
1412 printf(" # %d %s will execute (on %d 1436 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1413 g->p.nr_tasks, g->p.nr_tasks 1437 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1414 printf(" # %5dx %5ldMB global s 1438 printf(" # %5dx %5ldMB global shared mem operations\n",
1415 g->p.nr_loops, g->p.b 1439 g->p.nr_loops, g->p.bytes_global/1024/1024);
1416 printf(" # %5dx %5ldMB process s 1440 printf(" # %5dx %5ldMB process shared mem operations\n",
1417 g->p.nr_loops, g->p.b 1441 g->p.nr_loops, g->p.bytes_process/1024/1024);
1418 printf(" # %5dx %5ldMB thread l 1442 printf(" # %5dx %5ldMB thread local mem operations\n",
1419 g->p.nr_loops, g->p.b 1443 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1420 1444
1421 printf(" ###\n"); 1445 printf(" ###\n");
1422 1446
1423 printf("\n ###\n"); fflush(stdout); 1447 printf("\n ###\n"); fflush(stdout);
1424 } 1448 }
1425 1449
1426 static void init_thread_data(void) 1450 static void init_thread_data(void)
1427 { 1451 {
1428 ssize_t size = sizeof(*g->threads)*g- 1452 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1429 int t; 1453 int t;
1430 1454
1431 g->threads = zalloc_shared_data(size) 1455 g->threads = zalloc_shared_data(size);
1432 1456
1433 for (t = 0; t < g->p.nr_tasks; t++) { 1457 for (t = 0; t < g->p.nr_tasks; t++) {
1434 struct thread_data *td = g->t 1458 struct thread_data *td = g->threads + t;
1435 size_t cpuset_size = CPU_ALLO 1459 size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1436 int cpu; 1460 int cpu;
1437 1461
1438 /* Allow all nodes by default 1462 /* Allow all nodes by default: */
1439 td->bind_node = NUMA_NO_NODE; 1463 td->bind_node = NUMA_NO_NODE;
1440 1464
1441 /* Allow all CPUs by default: 1465 /* Allow all CPUs by default: */
1442 td->bind_cpumask = CPU_ALLOC( 1466 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1443 BUG_ON(!td->bind_cpumask); 1467 BUG_ON(!td->bind_cpumask);
1444 CPU_ZERO_S(cpuset_size, td->b 1468 CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1445 for (cpu = 0; cpu < g->p.nr_c 1469 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1446 CPU_SET_S(cpu, cpuset 1470 CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1447 } 1471 }
1448 } 1472 }
1449 1473
1450 static void deinit_thread_data(void) 1474 static void deinit_thread_data(void)
1451 { 1475 {
1452 ssize_t size = sizeof(*g->threads)*g- 1476 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1453 int t; 1477 int t;
1454 1478
1455 /* Free the bind_cpumask allocated fo 1479 /* Free the bind_cpumask allocated for thread_data */
1456 for (t = 0; t < g->p.nr_tasks; t++) { 1480 for (t = 0; t < g->p.nr_tasks; t++) {
1457 struct thread_data *td = g->t 1481 struct thread_data *td = g->threads + t;
1458 CPU_FREE(td->bind_cpumask); 1482 CPU_FREE(td->bind_cpumask);
1459 } 1483 }
1460 1484
1461 free_data(g->threads, size); 1485 free_data(g->threads, size);
1462 } 1486 }
1463 1487
1464 static int init(void) 1488 static int init(void)
1465 { 1489 {
1466 g = (void *)alloc_data(sizeof(*g), MA 1490 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1467 1491
1468 /* Copy over options: */ 1492 /* Copy over options: */
1469 g->p = p0; 1493 g->p = p0;
1470 1494
1471 g->p.nr_cpus = numa_num_configured_cp 1495 g->p.nr_cpus = numa_num_configured_cpus();
1472 1496
1473 g->p.nr_nodes = numa_max_node() + 1; 1497 g->p.nr_nodes = numa_max_node() + 1;
1474 1498
1475 /* char array in count_process_nodes( 1499 /* char array in count_process_nodes(): */
1476 BUG_ON(g->p.nr_nodes < 0); 1500 BUG_ON(g->p.nr_nodes < 0);
1477 1501
1478 if (quiet && !g->p.show_details) !! 1502 if (g->p.show_quiet && !g->p.show_details)
1479 g->p.show_details = -1; 1503 g->p.show_details = -1;
1480 1504
1481 /* Some memory should be specified: * 1505 /* Some memory should be specified: */
1482 if (!g->p.mb_global_str && !g->p.mb_p 1506 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1483 return -1; 1507 return -1;
1484 1508
1485 if (g->p.mb_global_str) { 1509 if (g->p.mb_global_str) {
1486 g->p.mb_global = atof(g->p.mb 1510 g->p.mb_global = atof(g->p.mb_global_str);
1487 BUG_ON(g->p.mb_global < 0); 1511 BUG_ON(g->p.mb_global < 0);
1488 } 1512 }
1489 1513
1490 if (g->p.mb_proc_str) { 1514 if (g->p.mb_proc_str) {
1491 g->p.mb_proc = atof(g->p.mb_p 1515 g->p.mb_proc = atof(g->p.mb_proc_str);
1492 BUG_ON(g->p.mb_proc < 0); 1516 BUG_ON(g->p.mb_proc < 0);
1493 } 1517 }
1494 1518
1495 if (g->p.mb_proc_locked_str) { 1519 if (g->p.mb_proc_locked_str) {
1496 g->p.mb_proc_locked = atof(g- 1520 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1497 BUG_ON(g->p.mb_proc_locked < 1521 BUG_ON(g->p.mb_proc_locked < 0);
1498 BUG_ON(g->p.mb_proc_locked > 1522 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1499 } 1523 }
1500 1524
1501 if (g->p.mb_thread_str) { 1525 if (g->p.mb_thread_str) {
1502 g->p.mb_thread = atof(g->p.mb 1526 g->p.mb_thread = atof(g->p.mb_thread_str);
1503 BUG_ON(g->p.mb_thread < 0); 1527 BUG_ON(g->p.mb_thread < 0);
1504 } 1528 }
1505 1529
1506 BUG_ON(g->p.nr_threads <= 0); 1530 BUG_ON(g->p.nr_threads <= 0);
1507 BUG_ON(g->p.nr_proc <= 0); 1531 BUG_ON(g->p.nr_proc <= 0);
1508 1532
1509 g->p.nr_tasks = g->p.nr_proc*g->p.nr_ 1533 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1510 1534
1511 g->p.bytes_global = g-> 1535 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1512 g->p.bytes_process = g-> 1536 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1513 g->p.bytes_process_locked = g-> 1537 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1514 g->p.bytes_thread = g-> 1538 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1515 1539
1516 g->data = setup_shared_data(g->p.byte 1540 g->data = setup_shared_data(g->p.bytes_global);
1517 1541
1518 /* Startup serialization: */ 1542 /* Startup serialization: */
1519 mutex_init_pshared(&g->start_work_mut !! 1543 init_global_mutex(&g->start_work_mutex);
1520 cond_init_pshared(&g->start_work_cond !! 1544 init_global_cond(&g->start_work_cond);
1521 mutex_init_pshared(&g->startup_mutex) !! 1545 init_global_mutex(&g->startup_mutex);
1522 cond_init_pshared(&g->startup_cond); !! 1546 init_global_cond(&g->startup_cond);
1523 mutex_init_pshared(&g->stop_work_mute !! 1547 init_global_mutex(&g->stop_work_mutex);
1524 1548
1525 init_thread_data(); 1549 init_thread_data();
1526 1550
1527 tprintf("#\n"); 1551 tprintf("#\n");
1528 if (parse_setup_cpu_list() || parse_s 1552 if (parse_setup_cpu_list() || parse_setup_node_list())
1529 return -1; 1553 return -1;
1530 tprintf("#\n"); 1554 tprintf("#\n");
1531 1555
1532 print_summary(); 1556 print_summary();
1533 1557
1534 return 0; 1558 return 0;
1535 } 1559 }
1536 1560
1537 static void deinit(void) 1561 static void deinit(void)
1538 { 1562 {
1539 free_data(g->data, g->p.bytes_global) 1563 free_data(g->data, g->p.bytes_global);
1540 g->data = NULL; 1564 g->data = NULL;
1541 1565
1542 deinit_thread_data(); 1566 deinit_thread_data();
1543 1567
1544 free_data(g, sizeof(*g)); 1568 free_data(g, sizeof(*g));
1545 g = NULL; 1569 g = NULL;
1546 } 1570 }
1547 1571
1548 /* 1572 /*
1549 * Print a short or long result, depending on 1573 * Print a short or long result, depending on the verbosity setting:
1550 */ 1574 */
1551 static void print_res(const char *name, doubl 1575 static void print_res(const char *name, double val,
1552 const char *txt_unit, c 1576 const char *txt_unit, const char *txt_short, const char *txt_long)
1553 { 1577 {
1554 if (!name) 1578 if (!name)
1555 name = "main,"; 1579 name = "main,";
1556 1580
1557 if (!quiet) !! 1581 if (!g->p.show_quiet)
1558 printf(" %-30s %15.3f, %-15s 1582 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1559 else 1583 else
1560 printf(" %14.3f %s\n", val, t 1584 printf(" %14.3f %s\n", val, txt_long);
1561 } 1585 }
1562 1586
1563 static int __bench_numa(const char *name) 1587 static int __bench_numa(const char *name)
1564 { 1588 {
1565 struct timeval start, stop, diff; 1589 struct timeval start, stop, diff;
1566 u64 runtime_ns_min, runtime_ns_sum; 1590 u64 runtime_ns_min, runtime_ns_sum;
1567 pid_t *pids, pid, wpid; 1591 pid_t *pids, pid, wpid;
1568 double delta_runtime; 1592 double delta_runtime;
1569 double runtime_avg; 1593 double runtime_avg;
1570 double runtime_sec_max; 1594 double runtime_sec_max;
1571 double runtime_sec_min; 1595 double runtime_sec_min;
1572 int wait_stat; 1596 int wait_stat;
1573 double bytes; 1597 double bytes;
1574 int i, t, p; 1598 int i, t, p;
1575 1599
1576 if (init()) 1600 if (init())
1577 return -1; 1601 return -1;
1578 1602
1579 pids = zalloc(g->p.nr_proc * sizeof(* 1603 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1580 pid = -1; 1604 pid = -1;
1581 1605
1582 if (g->p.serialize_startup) { 1606 if (g->p.serialize_startup) {
1583 tprintf(" #\n"); 1607 tprintf(" #\n");
1584 tprintf(" # Startup synchroni 1608 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1585 } 1609 }
1586 1610
1587 gettimeofday(&start, NULL); 1611 gettimeofday(&start, NULL);
1588 1612
1589 for (i = 0; i < g->p.nr_proc; i++) { 1613 for (i = 0; i < g->p.nr_proc; i++) {
1590 pid = fork(); 1614 pid = fork();
1591 dprintf(" # process %2d: PID 1615 dprintf(" # process %2d: PID %d\n", i, pid);
1592 1616
1593 BUG_ON(pid < 0); 1617 BUG_ON(pid < 0);
1594 if (!pid) { 1618 if (!pid) {
1595 /* Child process: */ 1619 /* Child process: */
1596 worker_process(i); 1620 worker_process(i);
1597 1621
1598 exit(0); 1622 exit(0);
1599 } 1623 }
1600 pids[i] = pid; 1624 pids[i] = pid;
1601 1625
1602 } 1626 }
1603 1627
1604 if (g->p.serialize_startup) { 1628 if (g->p.serialize_startup) {
1605 bool threads_ready = false; 1629 bool threads_ready = false;
1606 double startup_sec; 1630 double startup_sec;
1607 1631
1608 /* 1632 /*
1609 * Wait for all the threads t 1633 * Wait for all the threads to start up. The last thread will
1610 * signal this process. 1634 * signal this process.
1611 */ 1635 */
1612 mutex_lock(&g->startup_mutex) !! 1636 pthread_mutex_lock(&g->startup_mutex);
1613 while (g->nr_tasks_started != 1637 while (g->nr_tasks_started != g->p.nr_tasks)
1614 cond_wait(&g->startup !! 1638 pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
1615 1639
1616 mutex_unlock(&g->startup_mute !! 1640 pthread_mutex_unlock(&g->startup_mutex);
1617 1641
1618 /* Wait for all threads to be 1642 /* Wait for all threads to be at the start_work_cond. */
1619 while (!threads_ready) { 1643 while (!threads_ready) {
1620 mutex_lock(&g->start_ !! 1644 pthread_mutex_lock(&g->start_work_mutex);
1621 threads_ready = (g->n 1645 threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1622 mutex_unlock(&g->star !! 1646 pthread_mutex_unlock(&g->start_work_mutex);
1623 if (!threads_ready) 1647 if (!threads_ready)
1624 usleep(1); 1648 usleep(1);
1625 } 1649 }
1626 1650
1627 gettimeofday(&stop, NULL); 1651 gettimeofday(&stop, NULL);
1628 1652
1629 timersub(&stop, &start, &diff 1653 timersub(&stop, &start, &diff);
1630 1654
1631 startup_sec = diff.tv_sec * N 1655 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1632 startup_sec += diff.tv_usec * 1656 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1633 startup_sec /= NSEC_PER_SEC; 1657 startup_sec /= NSEC_PER_SEC;
1634 1658
1635 tprintf(" threads initialized 1659 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1636 tprintf(" #\n"); 1660 tprintf(" #\n");
1637 1661
1638 start = stop; 1662 start = stop;
1639 /* Start all threads running. 1663 /* Start all threads running. */
1640 mutex_lock(&g->start_work_mut !! 1664 pthread_mutex_lock(&g->start_work_mutex);
1641 g->start_work = true; 1665 g->start_work = true;
1642 mutex_unlock(&g->start_work_m !! 1666 pthread_mutex_unlock(&g->start_work_mutex);
1643 cond_broadcast(&g->start_work !! 1667 pthread_cond_broadcast(&g->start_work_cond);
1644 } else { 1668 } else {
1645 gettimeofday(&start, NULL); 1669 gettimeofday(&start, NULL);
1646 } 1670 }
1647 1671
1648 /* Parent process: */ 1672 /* Parent process: */
1649 1673
1650 1674
1651 for (i = 0; i < g->p.nr_proc; i++) { 1675 for (i = 0; i < g->p.nr_proc; i++) {
1652 wpid = waitpid(pids[i], &wait 1676 wpid = waitpid(pids[i], &wait_stat, 0);
1653 BUG_ON(wpid < 0); 1677 BUG_ON(wpid < 0);
1654 BUG_ON(!WIFEXITED(wait_stat)) 1678 BUG_ON(!WIFEXITED(wait_stat));
1655 1679
1656 } 1680 }
1657 1681
1658 runtime_ns_sum = 0; 1682 runtime_ns_sum = 0;
1659 runtime_ns_min = -1LL; 1683 runtime_ns_min = -1LL;
1660 1684
1661 for (t = 0; t < g->p.nr_tasks; t++) { 1685 for (t = 0; t < g->p.nr_tasks; t++) {
1662 u64 thread_runtime_ns = g->th 1686 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1663 1687
1664 runtime_ns_sum += thread_runt 1688 runtime_ns_sum += thread_runtime_ns;
1665 runtime_ns_min = min(thread_r 1689 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1666 } 1690 }
1667 1691
1668 gettimeofday(&stop, NULL); 1692 gettimeofday(&stop, NULL);
1669 timersub(&stop, &start, &diff); 1693 timersub(&stop, &start, &diff);
1670 1694
1671 BUG_ON(bench_format != BENCH_FORMAT_D 1695 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1672 1696
1673 tprintf("\n ###\n"); 1697 tprintf("\n ###\n");
1674 tprintf("\n"); 1698 tprintf("\n");
1675 1699
1676 runtime_sec_max = diff.tv_sec * NSEC_ 1700 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1677 runtime_sec_max += diff.tv_usec * NSE 1701 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1678 runtime_sec_max /= NSEC_PER_SEC; 1702 runtime_sec_max /= NSEC_PER_SEC;
1679 1703
1680 runtime_sec_min = runtime_ns_min / NS 1704 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1681 1705
1682 bytes = g->bytes_done; 1706 bytes = g->bytes_done;
1683 runtime_avg = (double)runtime_ns_sum 1707 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1684 1708
1685 if (g->p.measure_convergence) { 1709 if (g->p.measure_convergence) {
1686 print_res(name, runtime_sec_m 1710 print_res(name, runtime_sec_max,
1687 "secs,", "NUMA-conver 1711 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1688 } 1712 }
1689 1713
1690 print_res(name, runtime_sec_max, 1714 print_res(name, runtime_sec_max,
1691 "secs,", "runtime-max/thread" 1715 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1692 1716
1693 print_res(name, runtime_sec_min, 1717 print_res(name, runtime_sec_min,
1694 "secs,", "runtime-min/thread" 1718 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1695 1719
1696 print_res(name, runtime_avg, 1720 print_res(name, runtime_avg,
1697 "secs,", "runtime-avg/thread" 1721 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1698 1722
1699 delta_runtime = (runtime_sec_max - ru 1723 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1700 print_res(name, delta_runtime / runti 1724 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1701 "%,", "spread-runtime/thread" 1725 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1702 1726
1703 print_res(name, bytes / g->p.nr_tasks 1727 print_res(name, bytes / g->p.nr_tasks / 1e9,
1704 "GB,", "data/thread", 1728 "GB,", "data/thread", "GB data processed, per thread");
1705 1729
1706 print_res(name, bytes / 1e9, 1730 print_res(name, bytes / 1e9,
1707 "GB,", "data-total", 1731 "GB,", "data-total", "GB data processed, total");
1708 1732
1709 print_res(name, runtime_sec_max * NSE 1733 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1710 "nsecs,", "runtime/byte/threa 1734 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1711 1735
1712 print_res(name, bytes / g->p.nr_tasks 1736 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1713 "GB/sec,", "thread-speed", 1737 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1714 1738
1715 print_res(name, bytes / runtime_sec_m 1739 print_res(name, bytes / runtime_sec_max / 1e9,
1716 "GB/sec,", "total-speed", 1740 "GB/sec,", "total-speed", "GB/sec total speed");
1717 1741
1718 if (g->p.show_details >= 2) { 1742 if (g->p.show_details >= 2) {
1719 char tname[14 + 2 * 11 + 1]; 1743 char tname[14 + 2 * 11 + 1];
1720 struct thread_data *td; 1744 struct thread_data *td;
1721 for (p = 0; p < g->p.nr_proc; 1745 for (p = 0; p < g->p.nr_proc; p++) {
1722 for (t = 0; t < g->p. 1746 for (t = 0; t < g->p.nr_threads; t++) {
1723 memset(tname, 1747 memset(tname, 0, sizeof(tname));
1724 td = g->threa 1748 td = g->threads + p*g->p.nr_threads + t;
1725 snprintf(tnam 1749 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1726 print_res(tna 1750 print_res(tname, td->speed_gbs,
1727 "GB/s 1751 "GB/sec", "thread-speed", "GB/sec/thread speed");
1728 print_res(tna 1752 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1729 "secs 1753 "secs", "thread-system-time", "system CPU time/thread");
1730 print_res(tna 1754 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1731 "secs 1755 "secs", "thread-user-time", "user CPU time/thread");
1732 } 1756 }
1733 } 1757 }
1734 } 1758 }
1735 1759
1736 free(pids); 1760 free(pids);
1737 1761
1738 deinit(); 1762 deinit();
1739 1763
1740 return 0; 1764 return 0;
1741 } 1765 }
1742 1766
1743 #define MAX_ARGS 50 1767 #define MAX_ARGS 50
1744 1768
1745 static int command_size(const char **argv) 1769 static int command_size(const char **argv)
1746 { 1770 {
1747 int size = 0; 1771 int size = 0;
1748 1772
1749 while (*argv) { 1773 while (*argv) {
1750 size++; 1774 size++;
1751 argv++; 1775 argv++;
1752 } 1776 }
1753 1777
1754 BUG_ON(size >= MAX_ARGS); 1778 BUG_ON(size >= MAX_ARGS);
1755 1779
1756 return size; 1780 return size;
1757 } 1781 }
1758 1782
1759 static void init_params(struct params *p, con 1783 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1760 { 1784 {
1761 int i; 1785 int i;
1762 1786
1763 printf("\n # Running %s \"perf bench 1787 printf("\n # Running %s \"perf bench numa", name);
1764 1788
1765 for (i = 0; i < argc; i++) 1789 for (i = 0; i < argc; i++)
1766 printf(" %s", argv[i]); 1790 printf(" %s", argv[i]);
1767 1791
1768 printf("\"\n"); 1792 printf("\"\n");
1769 1793
1770 memset(p, 0, sizeof(*p)); 1794 memset(p, 0, sizeof(*p));
1771 1795
1772 /* Initialize nonzero defaults: */ 1796 /* Initialize nonzero defaults: */
1773 1797
1774 p->serialize_startup = 1; 1798 p->serialize_startup = 1;
1775 p->data_reads = tru 1799 p->data_reads = true;
1776 p->data_writes = tru 1800 p->data_writes = true;
1777 p->data_backwards = tru 1801 p->data_backwards = true;
1778 p->data_rand_walk = tru 1802 p->data_rand_walk = true;
1779 p->nr_loops = -1; 1803 p->nr_loops = -1;
1780 p->init_random = tru 1804 p->init_random = true;
1781 p->mb_global_str = "1" 1805 p->mb_global_str = "1";
1782 p->nr_proc = 1; 1806 p->nr_proc = 1;
1783 p->nr_threads = 1; 1807 p->nr_threads = 1;
1784 p->nr_secs = 5; 1808 p->nr_secs = 5;
1785 p->run_all = arg 1809 p->run_all = argc == 1;
1786 } 1810 }
1787 1811
1788 static int run_bench_numa(const char *name, c 1812 static int run_bench_numa(const char *name, const char **argv)
1789 { 1813 {
1790 int argc = command_size(argv); 1814 int argc = command_size(argv);
1791 1815
1792 init_params(&p0, name, argc, argv); 1816 init_params(&p0, name, argc, argv);
1793 argc = parse_options(argc, argv, opti 1817 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1794 if (argc) 1818 if (argc)
1795 goto err; 1819 goto err;
1796 1820
1797 if (__bench_numa(name)) 1821 if (__bench_numa(name))
1798 goto err; 1822 goto err;
1799 1823
1800 return 0; 1824 return 0;
1801 1825
1802 err: 1826 err:
1803 return -1; 1827 return -1;
1804 } 1828 }
1805 1829
1806 #define OPT_BW_RAM "-s", "20", 1830 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1807 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, 1831 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1808 1832
1809 #define OPT_CONV "-s", "100", 1833 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1810 #define OPT_CONV_NOTHP OPT_CONV, 1834 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1811 1835
1812 #define OPT_BW "-s", "20", 1836 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1813 #define OPT_BW_NOTHP OPT_BW, 1837 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1814 1838
1815 /* 1839 /*
1816 * The built-in test-suite executed by "perf 1840 * The built-in test-suite executed by "perf bench numa -a".
1817 * 1841 *
1818 * (A minimum of 4 nodes and 16 GB of RAM is 1842 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1819 */ 1843 */
1820 static const char *tests[][MAX_ARGS] = { 1844 static const char *tests[][MAX_ARGS] = {
1821 /* Basic single-stream NUMA bandwidth meas 1845 /* Basic single-stream NUMA bandwidth measurements: */
1822 { "RAM-bw-local,", "mem", "-p", "1", 1846 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1823 "-C" , "", "-M", 1847 "-C" , "", "-M", "", OPT_BW_RAM },
1824 { "RAM-bw-local-NOTHP,", 1848 { "RAM-bw-local-NOTHP,",
1825 "mem", "-p", "1", 1849 "mem", "-p", "1", "-t", "1", "-P", "1024",
1826 "-C" , "", "-M", 1850 "-C" , "", "-M", "", OPT_BW_RAM_NOTHP },
1827 { "RAM-bw-remote,", "mem", "-p", "1", 1851 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1828 "-C" , "", "-M", 1852 "-C" , "", "-M", "1", OPT_BW_RAM },
1829 1853
1830 /* 2-stream NUMA bandwidth measurements: * 1854 /* 2-stream NUMA bandwidth measurements: */
1831 { "RAM-bw-local-2x,", "mem", "-p", "2", 1855 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1832 "-C", "0,2", "-M", 1856 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1833 { "RAM-bw-remote-2x,", "mem", "-p", "2", 1857 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1834 "-C", "0,2", "-M", 1858 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1835 1859
1836 /* Cross-stream NUMA bandwidth measurement 1860 /* Cross-stream NUMA bandwidth measurement: */
1837 { "RAM-bw-cross,", "mem", "-p", "2", 1861 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1838 "-C", "0,8", "-M", 1862 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1839 1863
1840 /* Convergence latency measurements: */ 1864 /* Convergence latency measurements: */
1841 { " 1x3-convergence,", "mem", "-p", "1", 1865 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1842 { " 1x4-convergence,", "mem", "-p", "1", 1866 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1843 { " 1x6-convergence,", "mem", "-p", "1", 1867 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1844 { " 2x3-convergence,", "mem", "-p", "2", 1868 { " 2x3-convergence,", "mem", "-p", "2", "-t", "3", "-P", "1020", OPT_CONV },
1845 { " 3x3-convergence,", "mem", "-p", "3", 1869 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1846 { " 4x4-convergence,", "mem", "-p", "4", 1870 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1847 { " 4x4-convergence-NOTHP,", 1871 { " 4x4-convergence-NOTHP,",
1848 "mem", "-p", "4", 1872 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1849 { " 4x6-convergence,", "mem", "-p", "4", 1873 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1850 { " 4x8-convergence,", "mem", "-p", "4", 1874 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1851 { " 8x4-convergence,", "mem", "-p", "8", 1875 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1852 { " 8x4-convergence-NOTHP,", 1876 { " 8x4-convergence-NOTHP,",
1853 "mem", "-p", "8", 1877 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1854 { " 3x1-convergence,", "mem", "-p", "3", 1878 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1855 { " 4x1-convergence,", "mem", "-p", "4", 1879 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1856 { " 8x1-convergence,", "mem", "-p", "8", 1880 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1857 { "16x1-convergence,", "mem", "-p", "16", 1881 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1858 { "32x1-convergence,", "mem", "-p", "32", 1882 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1859 1883
1860 /* Various NUMA process/thread layout band 1884 /* Various NUMA process/thread layout bandwidth measurements: */
1861 { " 2x1-bw-process,", "mem", "-p", "2", 1885 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1862 { " 3x1-bw-process,", "mem", "-p", "3", 1886 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1863 { " 4x1-bw-process,", "mem", "-p", "4", 1887 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1864 { " 8x1-bw-process,", "mem", "-p", "8", 1888 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1865 { " 8x1-bw-process-NOTHP,", 1889 { " 8x1-bw-process-NOTHP,",
1866 "mem", "-p", "8", 1890 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1867 { "16x1-bw-process,", "mem", "-p", "16", 1891 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1868 1892
1869 { " 1x4-bw-thread,", "mem", "-p", "1", 1893 { " 1x4-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1870 { " 1x8-bw-thread,", "mem", "-p", "1", 1894 { " 1x8-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1871 { "1x16-bw-thread,", "mem", "-p", "1", 1895 { "1x16-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1872 { "1x32-bw-thread,", "mem", "-p", "1", 1896 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1873 1897
1874 { " 2x3-bw-process,", "mem", "-p", "2", 1898 { " 2x3-bw-process,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1875 { " 4x4-bw-process,", "mem", "-p", "4", 1899 { " 4x4-bw-process,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1876 { " 4x6-bw-process,", "mem", "-p", "4", 1900 { " 4x6-bw-process,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1877 { " 4x8-bw-process,", "mem", "-p", "4", 1901 { " 4x8-bw-process,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1878 { " 4x8-bw-process-NOTHP,", 1902 { " 4x8-bw-process-NOTHP,",
1879 "mem", "-p", "4", 1903 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1880 { " 3x3-bw-process,", "mem", "-p", "3", 1904 { " 3x3-bw-process,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1881 { " 5x5-bw-process,", "mem", "-p", "5", 1905 { " 5x5-bw-process,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1882 1906
1883 { "2x16-bw-process,", "mem", "-p", "2", 1907 { "2x16-bw-process,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1884 { "1x32-bw-process,", "mem", "-p", "1", 1908 { "1x32-bw-process,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1885 1909
1886 { "numa02-bw,", "mem", "-p", "1", 1910 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1887 { "numa02-bw-NOTHP,", "mem", "-p", "1", 1911 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1888 { "numa01-bw-thread,", "mem", "-p", "2", 1912 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1889 { "numa01-bw-thread-NOTHP,", 1913 { "numa01-bw-thread-NOTHP,",
1890 "mem", "-p", "2", 1914 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1891 }; 1915 };
1892 1916
1893 static int bench_all(void) 1917 static int bench_all(void)
1894 { 1918 {
1895 int nr = ARRAY_SIZE(tests); 1919 int nr = ARRAY_SIZE(tests);
1896 int ret; 1920 int ret;
1897 int i; 1921 int i;
1898 1922
1899 ret = system("echo ' #'; echo ' # Run 1923 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1900 BUG_ON(ret < 0); 1924 BUG_ON(ret < 0);
1901 1925
1902 for (i = 0; i < nr; i++) { 1926 for (i = 0; i < nr; i++) {
1903 run_bench_numa(tests[i][0], t 1927 run_bench_numa(tests[i][0], tests[i] + 1);
1904 } 1928 }
1905 1929
1906 printf("\n"); 1930 printf("\n");
1907 1931
1908 return 0; 1932 return 0;
1909 } 1933 }
1910 1934
1911 int bench_numa(int argc, const char **argv) 1935 int bench_numa(int argc, const char **argv)
1912 { 1936 {
1913 init_params(&p0, "main,", argc, argv) 1937 init_params(&p0, "main,", argc, argv);
1914 argc = parse_options(argc, argv, opti 1938 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1915 if (argc) 1939 if (argc)
1916 goto err; 1940 goto err;
1917 1941
1918 if (p0.run_all) 1942 if (p0.run_all)
1919 return bench_all(); 1943 return bench_all();
1920 1944
1921 if (__bench_numa(NULL)) 1945 if (__bench_numa(NULL))
1922 goto err; 1946 goto err;
1923 1947
1924 return 0; 1948 return 0;
1925 1949
1926 err: 1950 err:
1927 usage_with_options(numa_usage, option 1951 usage_with_options(numa_usage, options);
1928 return -1; 1952 return -1;
1929 } 1953 }
1930 1954
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.