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