xref: /openbmc/linux/samples/bpf/test_lru_dist.c (revision 3b23dc52)
1 /*
2  * Copyright (c) 2016 Facebook
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of version 2 of the GNU General Public
6  * License as published by the Free Software Foundation.
7  */
8 #define _GNU_SOURCE
9 #include <linux/types.h>
10 #include <stdio.h>
11 #include <unistd.h>
12 #include <linux/bpf.h>
13 #include <errno.h>
14 #include <string.h>
15 #include <assert.h>
16 #include <sched.h>
17 #include <sys/wait.h>
18 #include <sys/stat.h>
19 #include <sys/resource.h>
20 #include <fcntl.h>
21 #include <stdlib.h>
22 #include <time.h>
23 
24 #include <bpf/bpf.h>
25 #include "bpf_util.h"
26 
27 #define min(a, b) ((a) < (b) ? (a) : (b))
28 #ifndef offsetof
29 # define offsetof(TYPE, MEMBER)	((size_t)&((TYPE *)0)->MEMBER)
30 #endif
31 #define container_of(ptr, type, member) ({			\
32 	const typeof( ((type *)0)->member ) *__mptr = (ptr);	\
33 	(type *)( (char *)__mptr - offsetof(type,member) );})
34 
35 static int nr_cpus;
36 static unsigned long long *dist_keys;
37 static unsigned int dist_key_counts;
38 
39 struct list_head {
40 	struct list_head *next, *prev;
41 };
42 
43 static inline void INIT_LIST_HEAD(struct list_head *list)
44 {
45 	list->next = list;
46 	list->prev = list;
47 }
48 
49 static inline int list_empty(const struct list_head *head)
50 {
51 	return head->next == head;
52 }
53 
54 static inline void __list_add(struct list_head *new,
55 			      struct list_head *prev,
56 			      struct list_head *next)
57 {
58 	next->prev = new;
59 	new->next = next;
60 	new->prev = prev;
61 	prev->next = new;
62 }
63 
64 static inline void list_add(struct list_head *new, struct list_head *head)
65 {
66 	__list_add(new, head, head->next);
67 }
68 
69 static inline void __list_del(struct list_head *prev, struct list_head *next)
70 {
71 	next->prev = prev;
72 	prev->next = next;
73 }
74 
75 static inline void __list_del_entry(struct list_head *entry)
76 {
77 	__list_del(entry->prev, entry->next);
78 }
79 
80 static inline void list_move(struct list_head *list, struct list_head *head)
81 {
82 	__list_del_entry(list);
83 	list_add(list, head);
84 }
85 
86 #define list_entry(ptr, type, member) \
87 	container_of(ptr, type, member)
88 
89 #define list_last_entry(ptr, type, member) \
90 	list_entry((ptr)->prev, type, member)
91 
92 struct pfect_lru_node {
93 	struct list_head list;
94 	unsigned long long key;
95 };
96 
97 struct pfect_lru {
98 	struct list_head list;
99 	struct pfect_lru_node *free_nodes;
100 	unsigned int cur_size;
101 	unsigned int lru_size;
102 	unsigned int nr_unique;
103 	unsigned int nr_misses;
104 	unsigned int total;
105 	int map_fd;
106 };
107 
108 static void pfect_lru_init(struct pfect_lru *lru, unsigned int lru_size,
109 			   unsigned int nr_possible_elems)
110 {
111 	lru->map_fd = bpf_create_map(BPF_MAP_TYPE_HASH,
112 				     sizeof(unsigned long long),
113 				     sizeof(struct pfect_lru_node *),
114 				     nr_possible_elems, 0);
115 	assert(lru->map_fd != -1);
116 
117 	lru->free_nodes = malloc(lru_size * sizeof(struct pfect_lru_node));
118 	assert(lru->free_nodes);
119 
120 	INIT_LIST_HEAD(&lru->list);
121 	lru->cur_size = 0;
122 	lru->lru_size = lru_size;
123 	lru->nr_unique = lru->nr_misses = lru->total = 0;
124 }
125 
126 static void pfect_lru_destroy(struct pfect_lru *lru)
127 {
128 	close(lru->map_fd);
129 	free(lru->free_nodes);
130 }
131 
132 static int pfect_lru_lookup_or_insert(struct pfect_lru *lru,
133 				      unsigned long long key)
134 {
135 	struct pfect_lru_node *node = NULL;
136 	int seen = 0;
137 
138 	lru->total++;
139 	if (!bpf_map_lookup_elem(lru->map_fd, &key, &node)) {
140 		if (node) {
141 			list_move(&node->list, &lru->list);
142 			return 1;
143 		}
144 		seen = 1;
145 	}
146 
147 	if (lru->cur_size < lru->lru_size) {
148 		node =  &lru->free_nodes[lru->cur_size++];
149 		INIT_LIST_HEAD(&node->list);
150 	} else {
151 		struct pfect_lru_node *null_node = NULL;
152 
153 		node = list_last_entry(&lru->list,
154 				       struct pfect_lru_node,
155 				       list);
156 		bpf_map_update_elem(lru->map_fd, &node->key, &null_node, BPF_EXIST);
157 	}
158 
159 	node->key = key;
160 	list_move(&node->list, &lru->list);
161 
162 	lru->nr_misses++;
163 	if (seen) {
164 		assert(!bpf_map_update_elem(lru->map_fd, &key, &node, BPF_EXIST));
165 	} else {
166 		lru->nr_unique++;
167 		assert(!bpf_map_update_elem(lru->map_fd, &key, &node, BPF_NOEXIST));
168 	}
169 
170 	return seen;
171 }
172 
173 static unsigned int read_keys(const char *dist_file,
174 			      unsigned long long **keys)
175 {
176 	struct stat fst;
177 	unsigned long long *retkeys;
178 	unsigned int counts = 0;
179 	int dist_fd;
180 	char *b, *l;
181 	int i;
182 
183 	dist_fd = open(dist_file, 0);
184 	assert(dist_fd != -1);
185 
186 	assert(fstat(dist_fd, &fst) == 0);
187 	b = malloc(fst.st_size);
188 	assert(b);
189 
190 	assert(read(dist_fd, b, fst.st_size) == fst.st_size);
191 	close(dist_fd);
192 	for (i = 0; i < fst.st_size; i++) {
193 		if (b[i] == '\n')
194 			counts++;
195 	}
196 	counts++; /* in case the last line has no \n */
197 
198 	retkeys = malloc(counts * sizeof(unsigned long long));
199 	assert(retkeys);
200 
201 	counts = 0;
202 	for (l = strtok(b, "\n"); l; l = strtok(NULL, "\n"))
203 		retkeys[counts++] = strtoull(l, NULL, 10);
204 	free(b);
205 
206 	*keys = retkeys;
207 
208 	return counts;
209 }
210 
211 static int create_map(int map_type, int map_flags, unsigned int size)
212 {
213 	int map_fd;
214 
215 	map_fd = bpf_create_map(map_type, sizeof(unsigned long long),
216 				sizeof(unsigned long long), size, map_flags);
217 
218 	if (map_fd == -1)
219 		perror("bpf_create_map");
220 
221 	return map_fd;
222 }
223 
224 static int sched_next_online(int pid, int next_to_try)
225 {
226 	cpu_set_t cpuset;
227 
228 	if (next_to_try == nr_cpus)
229 		return -1;
230 
231 	while (next_to_try < nr_cpus) {
232 		CPU_ZERO(&cpuset);
233 		CPU_SET(next_to_try++, &cpuset);
234 		if (!sched_setaffinity(pid, sizeof(cpuset), &cpuset))
235 			break;
236 	}
237 
238 	return next_to_try;
239 }
240 
241 static void run_parallel(unsigned int tasks, void (*fn)(int i, void *data),
242 			 void *data)
243 {
244 	int next_sched_cpu = 0;
245 	pid_t pid[tasks];
246 	int i;
247 
248 	for (i = 0; i < tasks; i++) {
249 		pid[i] = fork();
250 		if (pid[i] == 0) {
251 			next_sched_cpu = sched_next_online(0, next_sched_cpu);
252 			fn(i, data);
253 			exit(0);
254 		} else if (pid[i] == -1) {
255 			printf("couldn't spawn #%d process\n", i);
256 			exit(1);
257 		}
258 		/* It is mostly redundant and just allow the parent
259 		 * process to update next_shced_cpu for the next child
260 		 * process
261 		 */
262 		next_sched_cpu = sched_next_online(pid[i], next_sched_cpu);
263 	}
264 	for (i = 0; i < tasks; i++) {
265 		int status;
266 
267 		assert(waitpid(pid[i], &status, 0) == pid[i]);
268 		assert(status == 0);
269 	}
270 }
271 
272 static void do_test_lru_dist(int task, void *data)
273 {
274 	unsigned int nr_misses = 0;
275 	struct pfect_lru pfect_lru;
276 	unsigned long long key, value = 1234;
277 	unsigned int i;
278 
279 	unsigned int lru_map_fd = ((unsigned int *)data)[0];
280 	unsigned int lru_size = ((unsigned int *)data)[1];
281 	unsigned long long key_offset = task * dist_key_counts;
282 
283 	pfect_lru_init(&pfect_lru, lru_size, dist_key_counts);
284 
285 	for (i = 0; i < dist_key_counts; i++) {
286 		key = dist_keys[i] + key_offset;
287 
288 		pfect_lru_lookup_or_insert(&pfect_lru, key);
289 
290 		if (!bpf_map_lookup_elem(lru_map_fd, &key, &value))
291 			continue;
292 
293 		if (bpf_map_update_elem(lru_map_fd, &key, &value, BPF_NOEXIST)) {
294 			printf("bpf_map_update_elem(lru_map_fd, %llu): errno:%d\n",
295 			       key, errno);
296 			assert(0);
297 		}
298 
299 		nr_misses++;
300 	}
301 
302 	printf("    task:%d BPF LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
303 	       task, pfect_lru.nr_unique, dist_key_counts, nr_misses,
304 	       dist_key_counts);
305 	printf("    task:%d Perfect LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
306 	       task, pfect_lru.nr_unique, pfect_lru.total,
307 	       pfect_lru.nr_misses, pfect_lru.total);
308 
309 	pfect_lru_destroy(&pfect_lru);
310 	close(lru_map_fd);
311 }
312 
313 static void test_parallel_lru_dist(int map_type, int map_flags,
314 				   int nr_tasks, unsigned int lru_size)
315 {
316 	int child_data[2];
317 	int lru_map_fd;
318 
319 	printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
320 	       map_flags);
321 
322 	if (map_flags & BPF_F_NO_COMMON_LRU)
323 		lru_map_fd = create_map(map_type, map_flags,
324 					nr_cpus * lru_size);
325 	else
326 		lru_map_fd = create_map(map_type, map_flags,
327 					nr_tasks * lru_size);
328 	assert(lru_map_fd != -1);
329 
330 	child_data[0] = lru_map_fd;
331 	child_data[1] = lru_size;
332 
333 	run_parallel(nr_tasks, do_test_lru_dist, child_data);
334 
335 	close(lru_map_fd);
336 }
337 
338 static void test_lru_loss0(int map_type, int map_flags)
339 {
340 	unsigned long long key, value[nr_cpus];
341 	unsigned int old_unused_losses = 0;
342 	unsigned int new_unused_losses = 0;
343 	unsigned int used_losses = 0;
344 	int map_fd;
345 
346 	printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
347 	       map_flags);
348 
349 	assert(sched_next_online(0, 0) != -1);
350 
351 	if (map_flags & BPF_F_NO_COMMON_LRU)
352 		map_fd = create_map(map_type, map_flags, 900 * nr_cpus);
353 	else
354 		map_fd = create_map(map_type, map_flags, 900);
355 
356 	assert(map_fd != -1);
357 
358 	value[0] = 1234;
359 
360 	for (key = 1; key <= 1000; key++) {
361 		int start_key, end_key;
362 
363 		assert(bpf_map_update_elem(map_fd, &key, value, BPF_NOEXIST) == 0);
364 
365 		start_key = 101;
366 		end_key = min(key, 900);
367 
368 		while (start_key <= end_key) {
369 			bpf_map_lookup_elem(map_fd, &start_key, value);
370 			start_key++;
371 		}
372 	}
373 
374 	for (key = 1; key <= 1000; key++) {
375 		if (bpf_map_lookup_elem(map_fd, &key, value)) {
376 			if (key <= 100)
377 				old_unused_losses++;
378 			else if (key <= 900)
379 				used_losses++;
380 			else
381 				new_unused_losses++;
382 		}
383 	}
384 
385 	close(map_fd);
386 
387 	printf("older-elem-losses:%d(/100) active-elem-losses:%d(/800) "
388 	       "newer-elem-losses:%d(/100)\n",
389 	       old_unused_losses, used_losses, new_unused_losses);
390 }
391 
392 static void test_lru_loss1(int map_type, int map_flags)
393 {
394 	unsigned long long key, value[nr_cpus];
395 	int map_fd;
396 	unsigned int nr_losses = 0;
397 
398 	printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
399 	       map_flags);
400 
401 	assert(sched_next_online(0, 0) != -1);
402 
403 	if (map_flags & BPF_F_NO_COMMON_LRU)
404 		map_fd = create_map(map_type, map_flags, 1000 * nr_cpus);
405 	else
406 		map_fd = create_map(map_type, map_flags, 1000);
407 
408 	assert(map_fd != -1);
409 
410 	value[0] = 1234;
411 
412 	for (key = 1; key <= 1000; key++)
413 		assert(!bpf_map_update_elem(map_fd, &key, value, BPF_NOEXIST));
414 
415 	for (key = 1; key <= 1000; key++) {
416 		if (bpf_map_lookup_elem(map_fd, &key, value))
417 			nr_losses++;
418 	}
419 
420 	close(map_fd);
421 
422 	printf("nr_losses:%d(/1000)\n", nr_losses);
423 }
424 
425 static void do_test_parallel_lru_loss(int task, void *data)
426 {
427 	const unsigned int nr_stable_elems = 1000;
428 	const unsigned int nr_repeats = 100000;
429 
430 	int map_fd = *(int *)data;
431 	unsigned long long stable_base;
432 	unsigned long long key, value[nr_cpus];
433 	unsigned long long next_ins_key;
434 	unsigned int nr_losses = 0;
435 	unsigned int i;
436 
437 	stable_base = task * nr_repeats * 2 + 1;
438 	next_ins_key = stable_base;
439 	value[0] = 1234;
440 	for (i = 0; i < nr_stable_elems; i++) {
441 		assert(bpf_map_update_elem(map_fd, &next_ins_key, value,
442 				       BPF_NOEXIST) == 0);
443 		next_ins_key++;
444 	}
445 
446 	for (i = 0; i < nr_repeats; i++) {
447 		int rn;
448 
449 		rn = rand();
450 
451 		if (rn % 10) {
452 			key = rn % nr_stable_elems + stable_base;
453 			bpf_map_lookup_elem(map_fd, &key, value);
454 		} else {
455 			bpf_map_update_elem(map_fd, &next_ins_key, value,
456 					BPF_NOEXIST);
457 			next_ins_key++;
458 		}
459 	}
460 
461 	key = stable_base;
462 	for (i = 0; i < nr_stable_elems; i++) {
463 		if (bpf_map_lookup_elem(map_fd, &key, value))
464 			nr_losses++;
465 		key++;
466 	}
467 
468 	printf("    task:%d nr_losses:%u\n", task, nr_losses);
469 }
470 
471 static void test_parallel_lru_loss(int map_type, int map_flags, int nr_tasks)
472 {
473 	int map_fd;
474 
475 	printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
476 	       map_flags);
477 
478 	/* Give 20% more than the active working set */
479 	if (map_flags & BPF_F_NO_COMMON_LRU)
480 		map_fd = create_map(map_type, map_flags,
481 				    nr_cpus * (1000 + 200));
482 	else
483 		map_fd = create_map(map_type, map_flags,
484 				    nr_tasks * (1000 + 200));
485 
486 	assert(map_fd != -1);
487 
488 	run_parallel(nr_tasks, do_test_parallel_lru_loss, &map_fd);
489 
490 	close(map_fd);
491 }
492 
493 int main(int argc, char **argv)
494 {
495 	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
496 	int map_flags[] = {0, BPF_F_NO_COMMON_LRU};
497 	const char *dist_file;
498 	int nr_tasks = 1;
499 	int lru_size;
500 	int f;
501 
502 	if (argc < 4) {
503 		printf("Usage: %s <dist-file> <lru-size> <nr-tasks>\n",
504 		       argv[0]);
505 		return -1;
506 	}
507 
508 	dist_file = argv[1];
509 	lru_size = atoi(argv[2]);
510 	nr_tasks = atoi(argv[3]);
511 
512 	setbuf(stdout, NULL);
513 
514 	assert(!setrlimit(RLIMIT_MEMLOCK, &r));
515 
516 	srand(time(NULL));
517 
518 	nr_cpus = bpf_num_possible_cpus();
519 	assert(nr_cpus != -1);
520 	printf("nr_cpus:%d\n\n", nr_cpus);
521 
522 	nr_tasks = min(nr_tasks, nr_cpus);
523 
524 	dist_key_counts = read_keys(dist_file, &dist_keys);
525 	if (!dist_key_counts) {
526 		printf("%s has no key\n", dist_file);
527 		return -1;
528 	}
529 
530 	for (f = 0; f < sizeof(map_flags) / sizeof(*map_flags); f++) {
531 		test_lru_loss0(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
532 		test_lru_loss1(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
533 		test_parallel_lru_loss(BPF_MAP_TYPE_LRU_HASH, map_flags[f],
534 				       nr_tasks);
535 		test_parallel_lru_dist(BPF_MAP_TYPE_LRU_HASH, map_flags[f],
536 				       nr_tasks, lru_size);
537 		printf("\n");
538 	}
539 
540 	free(dist_keys);
541 
542 	return 0;
543 }
544