1 | #ifndef LIBOMP_TEST_TOPOLOGY_H |
2 | #define LIBOMP_TEST_TOPOLOGY_H |
3 | |
4 | #include "libomp_test_affinity.h" |
5 | #include <stdio.h> |
6 | #include <stdlib.h> |
7 | #include <dirent.h> |
8 | #include <errno.h> |
9 | #include <ctype.h> |
10 | #include <omp.h> |
11 | #include <stdarg.h> |
12 | |
13 | typedef enum topology_obj_type_t { |
14 | TOPOLOGY_OBJ_THREAD, |
15 | TOPOLOGY_OBJ_CORE, |
16 | TOPOLOGY_OBJ_SOCKET, |
17 | TOPOLOGY_OBJ_MAX |
18 | } topology_obj_type_t; |
19 | |
20 | typedef struct place_list_t { |
21 | int num_places; |
22 | int current_place; |
23 | int *place_nums; |
24 | affinity_mask_t **masks; |
25 | } place_list_t; |
26 | |
27 | // Return the first character in file 'f' that is not a whitespace character |
28 | // including newlines and carriage returns |
29 | static int get_first_nonspace_from_file(FILE *f) { |
30 | int c; |
31 | do { |
32 | c = fgetc(stream: f); |
33 | } while (c != EOF && (isspace(c) || c == '\n' || c == '\r')); |
34 | return c; |
35 | } |
36 | |
37 | // Read an integer from file 'f' into 'number' |
38 | // Return 1 on successful read of integer, |
39 | // 0 on unsuccessful read of integer, |
40 | // EOF on end of file. |
41 | static int get_integer_from_file(FILE *f, int *number) { |
42 | int n; |
43 | n = fscanf(stream: f, format: "%d" , number); |
44 | if (feof(stream: f)) |
45 | return EOF; |
46 | if (n != 1) |
47 | return 0; |
48 | return 1; |
49 | } |
50 | |
51 | // Read a siblings list file from Linux /sys/devices/system/cpu/cpu?/topology/* |
52 | static affinity_mask_t *topology_get_mask_from_file(const char *filename) { |
53 | int status = EXIT_SUCCESS; |
54 | FILE *f = fopen(filename: filename, modes: "r" ); |
55 | if (!f) { |
56 | perror(s: filename); |
57 | exit(EXIT_FAILURE); |
58 | } |
59 | affinity_mask_t *mask = affinity_mask_alloc(); |
60 | while (1) { |
61 | int c, i, n, lower, upper; |
62 | // Read the first integer |
63 | n = get_integer_from_file(f, number: &lower); |
64 | if (n == EOF) { |
65 | break; |
66 | } else if (n == 0) { |
67 | fprintf(stderr, format: "syntax error: expected integer\n" ); |
68 | status = EXIT_FAILURE; |
69 | break; |
70 | } |
71 | |
72 | // Now either a , or - |
73 | c = get_first_nonspace_from_file(f); |
74 | if (c == EOF || c == ',') { |
75 | affinity_mask_set(mask, cpu: lower); |
76 | if (c == EOF) |
77 | break; |
78 | } else if (c == '-') { |
79 | n = get_integer_from_file(f, number: &upper); |
80 | if (n == EOF || n == 0) { |
81 | fprintf(stderr, format: "syntax error: expected integer\n" ); |
82 | status = EXIT_FAILURE; |
83 | break; |
84 | } |
85 | for (i = lower; i <= upper; ++i) |
86 | affinity_mask_set(mask, cpu: i); |
87 | c = get_first_nonspace_from_file(f); |
88 | if (c == EOF) { |
89 | break; |
90 | } else if (c == ',') { |
91 | continue; |
92 | } else { |
93 | fprintf(stderr, format: "syntax error: unexpected character: '%c (%d)'\n" , c, |
94 | c); |
95 | status = EXIT_FAILURE; |
96 | break; |
97 | } |
98 | } else { |
99 | fprintf(stderr, format: "syntax error: unexpected character: '%c (%d)'\n" , c, c); |
100 | status = EXIT_FAILURE; |
101 | break; |
102 | } |
103 | } |
104 | fclose(stream: f); |
105 | if (status == EXIT_FAILURE) { |
106 | affinity_mask_free(mask); |
107 | mask = NULL; |
108 | } |
109 | return mask; |
110 | } |
111 | |
112 | static int topology_get_num_cpus() { |
113 | char buf[1024]; |
114 | // Count the number of cpus |
115 | int cpu = 0; |
116 | while (1) { |
117 | snprintf(s: buf, maxlen: sizeof(buf), format: "/sys/devices/system/cpu/cpu%d" , cpu); |
118 | DIR *dir = opendir(name: buf); |
119 | if (dir) { |
120 | closedir(dirp: dir); |
121 | cpu++; |
122 | } else { |
123 | break; |
124 | } |
125 | } |
126 | if (cpu == 0) |
127 | cpu = 1; |
128 | return cpu; |
129 | } |
130 | |
131 | // Return whether the current thread has access to all logical processors |
132 | static int topology_using_full_mask() { |
133 | int cpu; |
134 | int has_all = 1; |
135 | int num_cpus = topology_get_num_cpus(); |
136 | affinity_mask_t *mask = affinity_mask_alloc(); |
137 | get_thread_affinity(mask); |
138 | for (cpu = 0; cpu < num_cpus; ++cpu) { |
139 | if (!affinity_mask_isset(mask, cpu)) { |
140 | has_all = 0; |
141 | break; |
142 | } |
143 | } |
144 | affinity_mask_free(mask); |
145 | return has_all; |
146 | } |
147 | |
148 | // Return array of masks representing OMP_PLACES keyword (e.g., sockets, cores, |
149 | // threads) |
150 | static place_list_t *topology_alloc_type_places(topology_obj_type_t type) { |
151 | char buf[1024]; |
152 | int i, cpu, num_places, num_unique; |
153 | int *place_nums; |
154 | int num_cpus = topology_get_num_cpus(); |
155 | place_list_t *places = (place_list_t *)malloc(size: sizeof(place_list_t)); |
156 | affinity_mask_t **masks = |
157 | (affinity_mask_t **)malloc(size: sizeof(affinity_mask_t *) * num_cpus); |
158 | num_unique = 0; |
159 | for (cpu = 0; cpu < num_cpus; ++cpu) { |
160 | affinity_mask_t *mask; |
161 | if (type == TOPOLOGY_OBJ_CORE) { |
162 | snprintf(s: buf, maxlen: sizeof(buf), |
163 | format: "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list" , |
164 | cpu); |
165 | mask = topology_get_mask_from_file(filename: buf); |
166 | } else if (type == TOPOLOGY_OBJ_SOCKET) { |
167 | snprintf(s: buf, maxlen: sizeof(buf), |
168 | format: "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list" , |
169 | cpu); |
170 | mask = topology_get_mask_from_file(filename: buf); |
171 | } else if (type == TOPOLOGY_OBJ_THREAD) { |
172 | mask = affinity_mask_alloc(); |
173 | affinity_mask_set(mask, cpu); |
174 | } else { |
175 | fprintf(stderr, format: "Unknown topology type (%d)\n" , (int)type); |
176 | exit(EXIT_FAILURE); |
177 | } |
178 | // Check for unique topology objects above the thread level |
179 | if (type != TOPOLOGY_OBJ_THREAD) { |
180 | for (i = 0; i < num_unique; ++i) { |
181 | if (affinity_mask_equal(mask1: masks[i], mask2: mask)) { |
182 | affinity_mask_free(mask); |
183 | mask = NULL; |
184 | break; |
185 | } |
186 | } |
187 | } |
188 | if (mask) |
189 | masks[num_unique++] = mask; |
190 | } |
191 | place_nums = (int *)malloc(size: sizeof(int) * num_unique); |
192 | for (i = 0; i < num_unique; ++i) |
193 | place_nums[i] = i; |
194 | places->num_places = num_unique; |
195 | places->masks = masks; |
196 | places->place_nums = place_nums; |
197 | places->current_place = -1; |
198 | return places; |
199 | } |
200 | |
201 | static place_list_t *topology_alloc_openmp_places() { |
202 | int place, i; |
203 | int num_places = omp_get_num_places(); |
204 | place_list_t *places = (place_list_t *)malloc(size: sizeof(place_list_t)); |
205 | affinity_mask_t **masks = |
206 | (affinity_mask_t **)malloc(size: sizeof(affinity_mask_t *) * num_places); |
207 | int *place_nums = (int *)malloc(size: sizeof(int) * num_places); |
208 | for (place = 0; place < num_places; ++place) { |
209 | int num_procs = omp_get_place_num_procs(place); |
210 | int *ids = (int *)malloc(size: sizeof(int) * num_procs); |
211 | omp_get_place_proc_ids(place, ids); |
212 | affinity_mask_t *mask = affinity_mask_alloc(); |
213 | for (i = 0; i < num_procs; ++i) |
214 | affinity_mask_set(mask, cpu: ids[i]); |
215 | masks[place] = mask; |
216 | place_nums[place] = place; |
217 | } |
218 | places->num_places = num_places; |
219 | places->place_nums = place_nums; |
220 | places->masks = masks; |
221 | places->current_place = omp_get_place_num(); |
222 | return places; |
223 | } |
224 | |
225 | static place_list_t *topology_alloc_openmp_partition() { |
226 | int p, i; |
227 | int num_places = omp_get_partition_num_places(); |
228 | place_list_t *places = (place_list_t *)malloc(size: sizeof(place_list_t)); |
229 | int *place_nums = (int *)malloc(size: sizeof(int) * num_places); |
230 | affinity_mask_t **masks = |
231 | (affinity_mask_t **)malloc(size: sizeof(affinity_mask_t *) * num_places); |
232 | omp_get_partition_place_nums(place_nums); |
233 | for (p = 0; p < num_places; ++p) { |
234 | int place = place_nums[p]; |
235 | int num_procs = omp_get_place_num_procs(place); |
236 | int *ids = (int *)malloc(size: sizeof(int) * num_procs); |
237 | if (num_procs == 0) { |
238 | fprintf(stderr, format: "place %d has 0 procs?\n" , place); |
239 | exit(EXIT_FAILURE); |
240 | } |
241 | omp_get_place_proc_ids(place, ids); |
242 | affinity_mask_t *mask = affinity_mask_alloc(); |
243 | for (i = 0; i < num_procs; ++i) |
244 | affinity_mask_set(mask, cpu: ids[i]); |
245 | if (affinity_mask_count(mask) == 0) { |
246 | fprintf(stderr, format: "place %d has 0 procs set?\n" , place); |
247 | exit(EXIT_FAILURE); |
248 | } |
249 | masks[p] = mask; |
250 | } |
251 | places->num_places = num_places; |
252 | places->place_nums = place_nums; |
253 | places->masks = masks; |
254 | places->current_place = omp_get_place_num(); |
255 | return places; |
256 | } |
257 | |
258 | // Free the array of masks from one of: topology_alloc_type_masks() |
259 | // or topology_alloc_openmp_masks() |
260 | static void topology_free_places(place_list_t *places) { |
261 | int i; |
262 | for (i = 0; i < places->num_places; ++i) |
263 | affinity_mask_free(mask: places->masks[i]); |
264 | free(ptr: places->masks); |
265 | free(ptr: places->place_nums); |
266 | free(ptr: places); |
267 | } |
268 | |
269 | static void topology_print_places(const place_list_t *p) { |
270 | int i; |
271 | char buf[1024]; |
272 | for (i = 0; i < p->num_places; ++i) { |
273 | affinity_mask_snprintf(buf, bufsize: sizeof(buf), mask: p->masks[i]); |
274 | printf(format: "Place %d: %s\n" , p->place_nums[i], buf); |
275 | } |
276 | } |
277 | |
278 | // Print out an error message, possibly with two problem place lists, |
279 | // and then exit with failure |
280 | static void proc_bind_die(omp_proc_bind_t proc_bind, int T, int P, |
281 | const char *format, ...) { |
282 | va_list args; |
283 | va_start(args, format); |
284 | const char *pb; |
285 | switch (proc_bind) { |
286 | case omp_proc_bind_false: |
287 | pb = "False" ; |
288 | break; |
289 | case omp_proc_bind_true: |
290 | pb = "True" ; |
291 | break; |
292 | case omp_proc_bind_master: |
293 | pb = "Master (Primary)" ; |
294 | break; |
295 | case omp_proc_bind_close: |
296 | pb = "Close" ; |
297 | break; |
298 | case omp_proc_bind_spread: |
299 | pb = "Spread" ; |
300 | break; |
301 | default: |
302 | pb = "(Unknown Proc Bind Type)" ; |
303 | break; |
304 | } |
305 | if (proc_bind == omp_proc_bind_spread || proc_bind == omp_proc_bind_close) { |
306 | if (T <= P) { |
307 | fprintf(stderr, format: "%s : (T(%d) <= P(%d)) : " , pb, T, P); |
308 | } else { |
309 | fprintf(stderr, format: "%s : (T(%d) > P(%d)) : " , pb, T, P); |
310 | } |
311 | } else { |
312 | fprintf(stderr, format: "%s : T = %d, P = %d : " , pb, T, P); |
313 | } |
314 | vfprintf(stderr, format: format, arg: args); |
315 | va_end(args); |
316 | |
317 | exit(EXIT_FAILURE); |
318 | } |
319 | |
320 | // Return 1 on failure, 0 on success. |
321 | static void proc_bind_check(omp_proc_bind_t proc_bind, |
322 | const place_list_t *parent, place_list_t **children, |
323 | int nchildren) { |
324 | place_list_t *partition; |
325 | int T, i, j, place, low, high, first, last, count, current_place, num_places; |
326 | const int *place_nums; |
327 | int P = parent->num_places; |
328 | |
329 | // Find the correct T (there could be null entries in children) |
330 | place_list_t **partitions = |
331 | (place_list_t **)malloc(size: sizeof(place_list_t *) * nchildren); |
332 | T = 0; |
333 | for (i = 0; i < nchildren; ++i) |
334 | if (children[i]) |
335 | partitions[T++] = children[i]; |
336 | // Only able to check spread, close, master (primary) |
337 | if (proc_bind != omp_proc_bind_spread && proc_bind != omp_proc_bind_close && |
338 | proc_bind != omp_proc_bind_master) |
339 | proc_bind_die(proc_bind, T, P, NULL, NULL, |
340 | "Cannot check this proc bind type\n" ); |
341 | |
342 | if (proc_bind == omp_proc_bind_spread) { |
343 | if (T <= P) { |
344 | // Run through each subpartition |
345 | for (i = 0; i < T; ++i) { |
346 | partition = partitions[i]; |
347 | place_nums = partition->place_nums; |
348 | num_places = partition->num_places; |
349 | current_place = partition->current_place; |
350 | // Correct count? |
351 | low = P / T; |
352 | high = P / T + (P % T ? 1 : 0); |
353 | if (num_places != low && num_places != high) { |
354 | proc_bind_die(proc_bind, T, P, |
355 | "Incorrect number of places for thread %d: %d. " |
356 | "Expecting between %d and %d\n" , |
357 | i, num_places, low, high); |
358 | } |
359 | // Consecutive places? |
360 | for (j = 1; j < num_places; ++j) { |
361 | if (place_nums[j] != (place_nums[j - 1] + 1) % P) { |
362 | proc_bind_die(proc_bind, T, P, |
363 | "Not consecutive places: %d, %d in partition\n" , |
364 | place_nums[j - 1], place_nums[j]); |
365 | } |
366 | } |
367 | first = place_nums[0]; |
368 | last = place_nums[num_places - 1]; |
369 | // Primary thread executes on place of the parent thread? |
370 | if (i == 0) { |
371 | if (current_place != parent->current_place) { |
372 | proc_bind_die( |
373 | proc_bind, T, P, |
374 | "Primary thread not on same place (%d) as parent thread (%d)\n" , |
375 | current_place, parent->current_place); |
376 | } |
377 | } else { |
378 | // Thread's current place is first place within it's partition? |
379 | if (current_place != first) { |
380 | proc_bind_die(proc_bind, T, P, |
381 | "Thread's current place (%d) is not the first place " |
382 | "in its partition [%d, %d]\n" , |
383 | current_place, first, last); |
384 | } |
385 | } |
386 | // Partitions don't have intersections? |
387 | int f1 = first; |
388 | int l1 = last; |
389 | for (j = 0; j < i; ++j) { |
390 | int f2 = partitions[j]->place_nums[0]; |
391 | int l2 = partitions[j]->place_nums[partitions[j]->num_places - 1]; |
392 | if (f1 > l1 && f2 > l2) { |
393 | proc_bind_die(proc_bind, T, P, |
394 | "partitions intersect. [%d, %d] and [%d, %d]\n" , f1, |
395 | l1, f2, l2); |
396 | } |
397 | if (f1 > l1 && f2 <= l2) |
398 | if (f1 < l2 || l1 > f2) { |
399 | proc_bind_die(proc_bind, T, P, |
400 | "partitions intersect. [%d, %d] and [%d, %d]\n" , f1, |
401 | l1, f2, l2); |
402 | } |
403 | if (f1 <= l1 && f2 > l2) |
404 | if (f2 < l1 || l2 > f1) { |
405 | proc_bind_die(proc_bind, T, P, |
406 | "partitions intersect. [%d, %d] and [%d, %d]\n" , f1, |
407 | l1, f2, l2); |
408 | } |
409 | if (f1 <= l1 && f2 <= l2) |
410 | if (!(f2 > l1 || l2 < f1)) { |
411 | proc_bind_die(proc_bind, T, P, |
412 | "partitions intersect. [%d, %d] and [%d, %d]\n" , f1, |
413 | l1, f2, l2); |
414 | } |
415 | } |
416 | } |
417 | } else { |
418 | // T > P |
419 | // Each partition has only one place? |
420 | for (i = 0; i < T; ++i) { |
421 | if (partitions[i]->num_places != 1) { |
422 | proc_bind_die( |
423 | proc_bind, T, P, |
424 | "Incorrect number of places for thread %d: %d. Expecting 1\n" , i, |
425 | partitions[i]->num_places); |
426 | } |
427 | } |
428 | // Correct number of consecutive threads per partition? |
429 | low = T / P; |
430 | high = T / P + (T % P ? 1 : 0); |
431 | for (i = 1, count = 1; i < T; ++i) { |
432 | if (partitions[i]->place_nums[0] == partitions[i - 1]->place_nums[0]) { |
433 | count++; |
434 | if (count > high) { |
435 | proc_bind_die( |
436 | proc_bind, T, P, |
437 | "Too many threads have place %d for their partition\n" , |
438 | partitions[i]->place_nums[0]); |
439 | } |
440 | } else { |
441 | if (count < low) { |
442 | proc_bind_die( |
443 | proc_bind, T, P, |
444 | "Not enough threads have place %d for their partition\n" , |
445 | partitions[i]->place_nums[0]); |
446 | } |
447 | count = 1; |
448 | } |
449 | } |
450 | // Primary thread executes on place of the parent thread? |
451 | current_place = partitions[0]->place_nums[0]; |
452 | if (parent->current_place != -1 && |
453 | current_place != parent->current_place) { |
454 | proc_bind_die( |
455 | proc_bind, T, P, |
456 | "Primary thread not on same place (%d) as parent thread (%d)\n" , |
457 | current_place, parent->current_place); |
458 | } |
459 | } |
460 | } else if (proc_bind == omp_proc_bind_close || |
461 | proc_bind == omp_proc_bind_master) { |
462 | // Check that each subpartition is the same as the parent |
463 | for (i = 0; i < T; ++i) { |
464 | partition = partitions[i]; |
465 | place_nums = partition->place_nums; |
466 | num_places = partition->num_places; |
467 | current_place = partition->current_place; |
468 | if (parent->num_places != num_places) { |
469 | proc_bind_die(proc_bind, T, P, |
470 | "Number of places in subpartition (%d) does not match " |
471 | "parent (%d)\n" , |
472 | num_places, parent->num_places); |
473 | } |
474 | for (j = 0; j < num_places; ++j) { |
475 | if (parent->place_nums[j] != place_nums[j]) { |
476 | proc_bind_die(proc_bind, T, P, |
477 | "Subpartition place (%d) does not match " |
478 | "parent partition place (%d)\n" , |
479 | place_nums[j], parent->place_nums[j]); |
480 | } |
481 | } |
482 | } |
483 | // Find index into place_nums of current place for parent |
484 | for (j = 0; j < parent->num_places; ++j) |
485 | if (parent->place_nums[j] == parent->current_place) |
486 | break; |
487 | if (proc_bind == omp_proc_bind_close) { |
488 | if (T <= P) { |
489 | // close T <= P |
490 | // check place assignment for each thread |
491 | for (i = 0; i < T; ++i) { |
492 | partition = partitions[i]; |
493 | current_place = partition->current_place; |
494 | if (current_place != parent->place_nums[j]) { |
495 | proc_bind_die( |
496 | proc_bind, T, P, |
497 | "Thread %d's current place (%d) is incorrect. expected %d\n" , i, |
498 | current_place, parent->place_nums[j]); |
499 | } |
500 | j = (j + 1) % parent->num_places; |
501 | } |
502 | } else { |
503 | // close T > P |
504 | // check place assignment for each thread |
505 | low = T / P; |
506 | high = T / P + (T % P ? 1 : 0); |
507 | count = 1; |
508 | if (partitions[0]->current_place != parent->current_place) { |
509 | proc_bind_die( |
510 | proc_bind, T, P, |
511 | "Primary thread's place (%d) is not parent thread's place (%d)\n" , |
512 | partitions[0]->current_place, parent->current_place); |
513 | } |
514 | for (i = 1; i < T; ++i) { |
515 | current_place = partitions[i]->current_place; |
516 | if (current_place == parent->place_nums[j]) { |
517 | count++; |
518 | if (count > high) { |
519 | proc_bind_die( |
520 | proc_bind, T, P, |
521 | "Too many threads have place %d for their current place\n" , |
522 | current_place); |
523 | } |
524 | } else { |
525 | if (count < low) { |
526 | proc_bind_die( |
527 | proc_bind, T, P, |
528 | "Not enough threads have place %d for their current place\n" , |
529 | parent->place_nums[j]); |
530 | } |
531 | j = (j + 1) % parent->num_places; |
532 | if (current_place != parent->place_nums[j]) { |
533 | proc_bind_die( |
534 | proc_bind, T, P, |
535 | "Thread %d's place (%d) is not corret. Expected %d\n" , i, |
536 | partitions[i]->current_place, parent->place_nums[j]); |
537 | } |
538 | count = 1; |
539 | } |
540 | } |
541 | } |
542 | } else { |
543 | // proc_bind_primary |
544 | // Every thread should be assigned to the primary thread's place |
545 | for (i = 0; i < T; ++i) { |
546 | if (partitions[i]->current_place != parent->current_place) { |
547 | proc_bind_die( |
548 | proc_bind, T, P, |
549 | "Thread %d's place (%d) is not the primary thread's place (%d)\n" , |
550 | i, partitions[i]->current_place, parent->current_place); |
551 | } |
552 | } |
553 | } |
554 | } |
555 | |
556 | // Check that each partition's current place is within the partition |
557 | for (i = 0; i < T; ++i) { |
558 | current_place = partitions[i]->current_place; |
559 | num_places = partitions[i]->num_places; |
560 | first = partitions[i]->place_nums[0]; |
561 | last = partitions[i]->place_nums[num_places - 1]; |
562 | for (j = 0; j < num_places; ++j) |
563 | if (partitions[i]->place_nums[j] == current_place) |
564 | break; |
565 | if (j == num_places) { |
566 | proc_bind_die(proc_bind, T, P, |
567 | "Thread %d's current place (%d) is not within its " |
568 | "partition [%d, %d]\n" , |
569 | i, current_place, first, last); |
570 | } |
571 | } |
572 | |
573 | free(ptr: partitions); |
574 | } |
575 | |
576 | #endif |
577 | |