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