projects / librseq.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Skip the tests if the rseq syscall is unavailable
[librseq.git] / tests / param_test.c
1 // SPDX-License-Identifier: LGPL-2.1-only
2 #ifndef _GNU_SOURCE
3 #define _GNU_SOURCE
4 #endif
5 #include <assert.h>
6 #include <pthread.h>
7 #include <sched.h>
8 #include <stdint.h>
9 #include <stdio.h>
10 #include <stdlib.h>
11 #include <string.h>
12 #include <syscall.h>
13 #include <unistd.h>
14 #include <poll.h>
15 #include <sys/types.h>
16 #include <signal.h>
17 #include <errno.h>
18 #include <stddef.h>
19
20 static inline pid_t rseq_gettid(void)
21 {
22 return syscall(__NR_gettid);
23 }
24
25 #define NR_INJECT 9
26 static int loop_cnt[NR_INJECT + 1];
27
28 static int loop_cnt_1 asm("asm_loop_cnt_1") __attribute__((used));
29 static int loop_cnt_2 asm("asm_loop_cnt_2") __attribute__((used));
30 static int loop_cnt_3 asm("asm_loop_cnt_3") __attribute__((used));
31 static int loop_cnt_4 asm("asm_loop_cnt_4") __attribute__((used));
32 static int loop_cnt_5 asm("asm_loop_cnt_5") __attribute__((used));
33 static int loop_cnt_6 asm("asm_loop_cnt_6") __attribute__((used));
34
35 static int opt_modulo, verbose;
36
37 static int opt_yield, opt_signal, opt_sleep,
38 opt_disable_rseq, opt_threads = 200,
39 opt_disable_mod = 0, opt_test = 's', opt_mb = 0;
40
41 #ifndef RSEQ_SKIP_FASTPATH
42 static long long opt_reps = 5000;
43 #else
44 static long long opt_reps = 100;
45 #endif
46
47 static __thread __attribute__((tls_model("initial-exec")))
48 unsigned int signals_delivered;
49
50 #ifndef BENCHMARK
51
52 static __thread __attribute__((tls_model("initial-exec"), unused))
53 int yield_mod_cnt, nr_abort;
54
55 #define printf_verbose(fmt, ...) \
56 do { \
57 if (verbose) \
58 printf(fmt, ## __VA_ARGS__); \
59 } while (0)
60
61 #ifdef __i386__
62
63 #define INJECT_ASM_REG "eax"
64
65 #define RSEQ_INJECT_CLOBBER \
66 , INJECT_ASM_REG
67
68 #define RSEQ_INJECT_ASM(n) \
69 "mov asm_loop_cnt_" #n ", %%" INJECT_ASM_REG "\n\t" \
70 "test %%" INJECT_ASM_REG ",%%" INJECT_ASM_REG "\n\t" \
71 "jz 333f\n\t" \
72 "222:\n\t" \
73 "dec %%" INJECT_ASM_REG "\n\t" \
74 "jnz 222b\n\t" \
75 "333:\n\t"
76
77 #elif defined(__x86_64__)
78
79 #define INJECT_ASM_REG_P "rax"
80 #define INJECT_ASM_REG "eax"
81
82 #define RSEQ_INJECT_CLOBBER \
83 , INJECT_ASM_REG_P \
84 , INJECT_ASM_REG
85
86 #define RSEQ_INJECT_ASM(n) \
87 "lea asm_loop_cnt_" #n "(%%rip), %%" INJECT_ASM_REG_P "\n\t" \
88 "mov (%%" INJECT_ASM_REG_P "), %%" INJECT_ASM_REG "\n\t" \
89 "test %%" INJECT_ASM_REG ",%%" INJECT_ASM_REG "\n\t" \
90 "jz 333f\n\t" \
91 "222:\n\t" \
92 "dec %%" INJECT_ASM_REG "\n\t" \
93 "jnz 222b\n\t" \
94 "333:\n\t"
95
96 #elif defined(__s390__)
97
98 #define RSEQ_INJECT_INPUT \
99 , [loop_cnt_1]"m"(loop_cnt[1]) \
100 , [loop_cnt_2]"m"(loop_cnt[2]) \
101 , [loop_cnt_3]"m"(loop_cnt[3]) \
102 , [loop_cnt_4]"m"(loop_cnt[4]) \
103 , [loop_cnt_5]"m"(loop_cnt[5]) \
104 , [loop_cnt_6]"m"(loop_cnt[6])
105
106 #define INJECT_ASM_REG "r12"
107
108 #define RSEQ_INJECT_CLOBBER \
109 , INJECT_ASM_REG
110
111 #define RSEQ_INJECT_ASM(n) \
112 "l %%" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
113 "ltr %%" INJECT_ASM_REG ", %%" INJECT_ASM_REG "\n\t" \
114 "je 333f\n\t" \
115 "222:\n\t" \
116 "ahi %%" INJECT_ASM_REG ", -1\n\t" \
117 "jnz 222b\n\t" \
118 "333:\n\t"
119
120 #elif defined(__ARMEL__)
121
122 #define RSEQ_INJECT_INPUT \
123 , [loop_cnt_1]"m"(loop_cnt[1]) \
124 , [loop_cnt_2]"m"(loop_cnt[2]) \
125 , [loop_cnt_3]"m"(loop_cnt[3]) \
126 , [loop_cnt_4]"m"(loop_cnt[4]) \
127 , [loop_cnt_5]"m"(loop_cnt[5]) \
128 , [loop_cnt_6]"m"(loop_cnt[6])
129
130 #define INJECT_ASM_REG "r4"
131
132 #define RSEQ_INJECT_CLOBBER \
133 , INJECT_ASM_REG
134
135 #define RSEQ_INJECT_ASM(n) \
136 "ldr " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
137 "cmp " INJECT_ASM_REG ", #0\n\t" \
138 "beq 333f\n\t" \
139 "222:\n\t" \
140 "subs " INJECT_ASM_REG ", #1\n\t" \
141 "bne 222b\n\t" \
142 "333:\n\t"
143
144 #elif defined(__AARCH64EL__)
145
146 #define RSEQ_INJECT_INPUT \
147 , [loop_cnt_1] "Qo" (loop_cnt[1]) \
148 , [loop_cnt_2] "Qo" (loop_cnt[2]) \
149 , [loop_cnt_3] "Qo" (loop_cnt[3]) \
150 , [loop_cnt_4] "Qo" (loop_cnt[4]) \
151 , [loop_cnt_5] "Qo" (loop_cnt[5]) \
152 , [loop_cnt_6] "Qo" (loop_cnt[6])
153
154 #define INJECT_ASM_REG RSEQ_ASM_TMP_REG32
155
156 #define RSEQ_INJECT_ASM(n) \
157 " ldr " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n" \
158 " cbz " INJECT_ASM_REG ", 333f\n" \
159 "222:\n" \
160 " sub " INJECT_ASM_REG ", " INJECT_ASM_REG ", #1\n" \
161 " cbnz " INJECT_ASM_REG ", 222b\n" \
162 "333:\n"
163
164 #elif __PPC__
165
166 #define RSEQ_INJECT_INPUT \
167 , [loop_cnt_1]"m"(loop_cnt[1]) \
168 , [loop_cnt_2]"m"(loop_cnt[2]) \
169 , [loop_cnt_3]"m"(loop_cnt[3]) \
170 , [loop_cnt_4]"m"(loop_cnt[4]) \
171 , [loop_cnt_5]"m"(loop_cnt[5]) \
172 , [loop_cnt_6]"m"(loop_cnt[6])
173
174 #define INJECT_ASM_REG "r18"
175
176 #define RSEQ_INJECT_CLOBBER \
177 , INJECT_ASM_REG
178
179 #define RSEQ_INJECT_ASM(n) \
180 "lwz %%" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
181 "cmpwi %%" INJECT_ASM_REG ", 0\n\t" \
182 "beq 333f\n\t" \
183 "222:\n\t" \
184 "subic. %%" INJECT_ASM_REG ", %%" INJECT_ASM_REG ", 1\n\t" \
185 "bne 222b\n\t" \
186 "333:\n\t"
187
188 #elif defined(__mips__)
189
190 #define RSEQ_INJECT_INPUT \
191 , [loop_cnt_1]"m"(loop_cnt[1]) \
192 , [loop_cnt_2]"m"(loop_cnt[2]) \
193 , [loop_cnt_3]"m"(loop_cnt[3]) \
194 , [loop_cnt_4]"m"(loop_cnt[4]) \
195 , [loop_cnt_5]"m"(loop_cnt[5]) \
196 , [loop_cnt_6]"m"(loop_cnt[6])
197
198 #define INJECT_ASM_REG "$5"
199
200 #define RSEQ_INJECT_CLOBBER \
201 , INJECT_ASM_REG
202
203 #define RSEQ_INJECT_ASM(n) \
204 "lw " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
205 "beqz " INJECT_ASM_REG ", 333f\n\t" \
206 "222:\n\t" \
207 "addiu " INJECT_ASM_REG ", -1\n\t" \
208 "bnez " INJECT_ASM_REG ", 222b\n\t" \
209 "333:\n\t"
210
211 #else
212 #error unsupported target
213 #endif
214
215 #define RSEQ_INJECT_FAILED \
216 nr_abort++;
217
218 #define RSEQ_INJECT_C(n) \
219 { \
220 int loc_i, loc_nr_loops = loop_cnt[n]; \
221 \
222 for (loc_i = 0; loc_i < loc_nr_loops; loc_i++) { \
223 rseq_barrier(); \
224 } \
225 if (loc_nr_loops == -1 && opt_modulo) { \
226 if (yield_mod_cnt == opt_modulo - 1) { \
227 if (opt_sleep > 0) \
228 poll(NULL, 0, opt_sleep); \
229 if (opt_yield) \
230 sched_yield(); \
231 if (opt_signal) \
232 raise(SIGUSR1); \
233 yield_mod_cnt = 0; \
234 } else { \
235 yield_mod_cnt++; \
236 } \
237 } \
238 }
239
240 #else
241
242 #define printf_verbose(fmt, ...)
243
244 #endif /* BENCHMARK */
245
246 #include <rseq/rseq.h>
247
248 struct percpu_lock_entry {
249 intptr_t v;
250 } __attribute__((aligned(128)));
251
252 struct percpu_lock {
253 struct percpu_lock_entry c[CPU_SETSIZE];
254 };
255
256 struct test_data_entry {
257 intptr_t count;
258 } __attribute__((aligned(128)));
259
260 struct spinlock_test_data {
261 struct percpu_lock lock;
262 struct test_data_entry c[CPU_SETSIZE];
263 };
264
265 struct spinlock_thread_test_data {
266 struct spinlock_test_data *data;
267 long long reps;
268 int reg;
269 };
270
271 struct inc_test_data {
272 struct test_data_entry c[CPU_SETSIZE];
273 };
274
275 struct inc_thread_test_data {
276 struct inc_test_data *data;
277 long long reps;
278 int reg;
279 };
280
281 struct percpu_list_node {
282 intptr_t data;
283 struct percpu_list_node *next;
284 };
285
286 struct percpu_list_entry {
287 struct percpu_list_node *head;
288 } __attribute__((aligned(128)));
289
290 struct percpu_list {
291 struct percpu_list_entry c[CPU_SETSIZE];
292 };
293
294 #define BUFFER_ITEM_PER_CPU 100
295
296 struct percpu_buffer_node {
297 intptr_t data;
298 };
299
300 struct percpu_buffer_entry {
301 intptr_t offset;
302 intptr_t buflen;
303 struct percpu_buffer_node **array;
304 } __attribute__((aligned(128)));
305
306 struct percpu_buffer {
307 struct percpu_buffer_entry c[CPU_SETSIZE];
308 };
309
310 #define MEMCPY_BUFFER_ITEM_PER_CPU 100
311
312 struct percpu_memcpy_buffer_node {
313 intptr_t data1;
314 uint64_t data2;
315 };
316
317 struct percpu_memcpy_buffer_entry {
318 intptr_t offset;
319 intptr_t buflen;
320 struct percpu_memcpy_buffer_node *array;
321 } __attribute__((aligned(128)));
322
323 struct percpu_memcpy_buffer {
324 struct percpu_memcpy_buffer_entry c[CPU_SETSIZE];
325 };
326
327 /* A simple percpu spinlock. Grabs lock on current cpu. */
328 static int rseq_this_cpu_lock(struct percpu_lock *lock)
329 {
330 int cpu;
331
332 for (;;) {
333 int ret;
334
335 cpu = rseq_cpu_start();
336 ret = rseq_cmpeqv_storev(&lock->c[cpu].v,
337 0, 1, cpu);
338 if (rseq_likely(!ret))
339 break;
340 /* Retry if comparison fails or rseq aborts. */
341 }
342 /*
343 * Acquire semantic when taking lock after control dependency.
344 * Matches rseq_smp_store_release().
345 */
346 rseq_smp_acquire__after_ctrl_dep();
347 return cpu;
348 }
349
350 static void rseq_percpu_unlock(struct percpu_lock *lock, int cpu)
351 {
352 assert(lock->c[cpu].v == 1);
353 /*
354 * Release lock, with release semantic. Matches
355 * rseq_smp_acquire__after_ctrl_dep().
356 */
357 rseq_smp_store_release(&lock->c[cpu].v, 0);
358 }
359
360 void *test_percpu_spinlock_thread(void *arg)
361 {
362 struct spinlock_thread_test_data *thread_data = arg;
363 struct spinlock_test_data *data = thread_data->data;
364 long long i, reps;
365
366 if (!opt_disable_rseq && thread_data->reg &&
367 rseq_register_current_thread())
368 abort();
369 reps = thread_data->reps;
370 for (i = 0; i < reps; i++) {
371 int cpu = rseq_cpu_start();
372
373 cpu = rseq_this_cpu_lock(&data->lock);
374 data->c[cpu].count++;
375 rseq_percpu_unlock(&data->lock, cpu);
376 #ifndef BENCHMARK
377 if (i != 0 && !(i % (reps / 10)))
378 printf_verbose("tid %d: count %lld\n",
379 (int) rseq_gettid(), i);
380 #endif
381 }
382 printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
383 (int) rseq_gettid(), nr_abort, signals_delivered);
384 if (!opt_disable_rseq && thread_data->reg &&
385 rseq_unregister_current_thread())
386 abort();
387 return NULL;
388 }
389
390 /*
391 * A simple test which implements a sharded counter using a per-cpu
392 * lock. Obviously real applications might prefer to simply use a
393 * per-cpu increment; however, this is reasonable for a test and the
394 * lock can be extended to synchronize more complicated operations.
395 */
396 void test_percpu_spinlock(void)
397 {
398 const int num_threads = opt_threads;
399 int i, ret;
400 uint64_t sum;
401 pthread_t test_threads[num_threads];
402 struct spinlock_test_data data;
403 struct spinlock_thread_test_data thread_data[num_threads];
404
405 memset(&data, 0, sizeof(data));
406 for (i = 0; i < num_threads; i++) {
407 thread_data[i].reps = opt_reps;
408 if (opt_disable_mod <= 0 || (i % opt_disable_mod))
409 thread_data[i].reg = 1;
410 else
411 thread_data[i].reg = 0;
412 thread_data[i].data = &data;
413 ret = pthread_create(&test_threads[i], NULL,
414 test_percpu_spinlock_thread,
415 &thread_data[i]);
416 if (ret) {
417 errno = ret;
418 perror("pthread_create");
419 abort();
420 }
421 }
422
423 for (i = 0; i < num_threads; i++) {
424 ret = pthread_join(test_threads[i], NULL);
425 if (ret) {
426 errno = ret;
427 perror("pthread_join");
428 abort();
429 }
430 }
431
432 sum = 0;
433 for (i = 0; i < CPU_SETSIZE; i++)
434 sum += data.c[i].count;
435
436 assert(sum == (uint64_t)opt_reps * num_threads);
437 }
438
439 void *test_percpu_inc_thread(void *arg)
440 {
441 struct inc_thread_test_data *thread_data = arg;
442 struct inc_test_data *data = thread_data->data;
443 long long i, reps;
444
445 if (!opt_disable_rseq && thread_data->reg &&
446 rseq_register_current_thread())
447 abort();
448 reps = thread_data->reps;
449 for (i = 0; i < reps; i++) {
450 int ret;
451
452 do {
453 int cpu;
454
455 cpu = rseq_cpu_start();
456 ret = rseq_addv(&data->c[cpu].count, 1, cpu);
457 } while (rseq_unlikely(ret));
458 #ifndef BENCHMARK
459 if (i != 0 && !(i % (reps / 10)))
460 printf_verbose("tid %d: count %lld\n",
461 (int) rseq_gettid(), i);
462 #endif
463 }
464 printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
465 (int) rseq_gettid(), nr_abort, signals_delivered);
466 if (!opt_disable_rseq && thread_data->reg &&
467 rseq_unregister_current_thread())
468 abort();
469 return NULL;
470 }
471
472 void test_percpu_inc(void)
473 {
474 const int num_threads = opt_threads;
475 int i, ret;
476 uint64_t sum;
477 pthread_t test_threads[num_threads];
478 struct inc_test_data data;
479 struct inc_thread_test_data thread_data[num_threads];
480
481 memset(&data, 0, sizeof(data));
482 for (i = 0; i < num_threads; i++) {
483 thread_data[i].reps = opt_reps;
484 if (opt_disable_mod <= 0 || (i % opt_disable_mod))
485 thread_data[i].reg = 1;
486 else
487 thread_data[i].reg = 0;
488 thread_data[i].data = &data;
489 ret = pthread_create(&test_threads[i], NULL,
490 test_percpu_inc_thread,
491 &thread_data[i]);
492 if (ret) {
493 errno = ret;
494 perror("pthread_create");
495 abort();
496 }
497 }
498
499 for (i = 0; i < num_threads; i++) {
500 ret = pthread_join(test_threads[i], NULL);
501 if (ret) {
502 errno = ret;
503 perror("pthread_join");
504 abort();
505 }
506 }
507
508 sum = 0;
509 for (i = 0; i < CPU_SETSIZE; i++)
510 sum += data.c[i].count;
511
512 assert(sum == (uint64_t)opt_reps * num_threads);
513 }
514
515 void this_cpu_list_push(struct percpu_list *list,
516 struct percpu_list_node *node,
517 int *_cpu)
518 {
519 int cpu;
520
521 for (;;) {
522 intptr_t *targetptr, newval, expect;
523 int ret;
524
525 cpu = rseq_cpu_start();
526 /* Load list->c[cpu].head with single-copy atomicity. */
527 expect = (intptr_t)RSEQ_READ_ONCE(list->c[cpu].head);
528 newval = (intptr_t)node;
529 targetptr = (intptr_t *)&list->c[cpu].head;
530 node->next = (struct percpu_list_node *)expect;
531 ret = rseq_cmpeqv_storev(targetptr, expect, newval, cpu);
532 if (rseq_likely(!ret))
533 break;
534 /* Retry if comparison fails or rseq aborts. */
535 }
536 if (_cpu)
537 *_cpu = cpu;
538 }
539
540 /*
541 * Unlike a traditional lock-less linked list; the availability of a
542 * rseq primitive allows us to implement pop without concerns over
543 * ABA-type races.
544 */
545 struct percpu_list_node *this_cpu_list_pop(struct percpu_list *list,
546 int *_cpu)
547 {
548 struct percpu_list_node *node = NULL;
549 int cpu;
550
551 for (;;) {
552 struct percpu_list_node *head;
553 intptr_t *targetptr, expectnot, *load;
554 off_t offset;
555 int ret;
556
557 cpu = rseq_cpu_start();
558 targetptr = (intptr_t *)&list->c[cpu].head;
559 expectnot = (intptr_t)NULL;
560 offset = offsetof(struct percpu_list_node, next);
561 load = (intptr_t *)&head;
562 ret = rseq_cmpnev_storeoffp_load(targetptr, expectnot,
563 offset, load, cpu);
564 if (rseq_likely(!ret)) {
565 node = head;
566 break;
567 }
568 if (ret > 0)
569 break;
570 /* Retry if rseq aborts. */
571 }
572 if (_cpu)
573 *_cpu = cpu;
574 return node;
575 }
576
577 /*
578 * __percpu_list_pop is not safe against concurrent accesses. Should
579 * only be used on lists that are not concurrently modified.
580 */
581 struct percpu_list_node *__percpu_list_pop(struct percpu_list *list, int cpu)
582 {
583 struct percpu_list_node *node;
584
585 node = list->c[cpu].head;
586 if (!node)
587 return NULL;
588 list->c[cpu].head = node->next;
589 return node;
590 }
591
592 void *test_percpu_list_thread(void *arg)
593 {
594 long long i, reps;
595 struct percpu_list *list = (struct percpu_list *)arg;
596
597 if (!opt_disable_rseq && rseq_register_current_thread())
598 abort();
599
600 reps = opt_reps;
601 for (i = 0; i < reps; i++) {
602 struct percpu_list_node *node;
603
604 node = this_cpu_list_pop(list, NULL);
605 if (opt_yield)
606 sched_yield(); /* encourage shuffling */
607 if (node)
608 this_cpu_list_push(list, node, NULL);
609 }
610
611 printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
612 (int) rseq_gettid(), nr_abort, signals_delivered);
613 if (!opt_disable_rseq && rseq_unregister_current_thread())
614 abort();
615
616 return NULL;
617 }
618
619 /* Simultaneous modification to a per-cpu linked list from many threads. */
620 void test_percpu_list(void)
621 {
622 const int num_threads = opt_threads;
623 int i, j, ret;
624 uint64_t sum = 0, expected_sum = 0;
625 struct percpu_list list;
626 pthread_t test_threads[num_threads];
627 cpu_set_t allowed_cpus;
628
629 memset(&list, 0, sizeof(list));
630
631 /* Generate list entries for every usable cpu. */
632 sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
633 for (i = 0; i < CPU_SETSIZE; i++) {
634 if (!CPU_ISSET(i, &allowed_cpus))
635 continue;
636 for (j = 1; j <= 100; j++) {
637 struct percpu_list_node *node;
638
639 expected_sum += j;
640
641 node = malloc(sizeof(*node));
642 assert(node);
643 node->data = j;
644 node->next = list.c[i].head;
645 list.c[i].head = node;
646 }
647 }
648
649 for (i = 0; i < num_threads; i++) {
650 ret = pthread_create(&test_threads[i], NULL,
651 test_percpu_list_thread, &list);
652 if (ret) {
653 errno = ret;
654 perror("pthread_create");
655 abort();
656 }
657 }
658
659 for (i = 0; i < num_threads; i++) {
660 ret = pthread_join(test_threads[i], NULL);
661 if (ret) {
662 errno = ret;
663 perror("pthread_join");
664 abort();
665 }
666 }
667
668 for (i = 0; i < CPU_SETSIZE; i++) {
669 struct percpu_list_node *node;
670
671 if (!CPU_ISSET(i, &allowed_cpus))
672 continue;
673
674 while ((node = __percpu_list_pop(&list, i))) {
675 sum += node->data;
676 free(node);
677 }
678 }
679
680 /*
681 * All entries should now be accounted for (unless some external
682 * actor is interfering with our allowed affinity while this
683 * test is running).
684 */
685 assert(sum == expected_sum);
686 }
687
688 bool this_cpu_buffer_push(struct percpu_buffer *buffer,
689 struct percpu_buffer_node *node,
690 int *_cpu)
691 {
692 bool result = false;
693 int cpu;
694
695 for (;;) {
696 intptr_t *targetptr_spec, newval_spec;
697 intptr_t *targetptr_final, newval_final;
698 intptr_t offset;
699 int ret;
700
701 cpu = rseq_cpu_start();
702 offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
703 if (offset == buffer->c[cpu].buflen)
704 break;
705 newval_spec = (intptr_t)node;
706 targetptr_spec = (intptr_t *)&buffer->c[cpu].array[offset];
707 newval_final = offset + 1;
708 targetptr_final = &buffer->c[cpu].offset;
709 if (opt_mb)
710 ret = rseq_cmpeqv_trystorev_storev_release(
711 targetptr_final, offset, targetptr_spec,
712 newval_spec, newval_final, cpu);
713 else
714 ret = rseq_cmpeqv_trystorev_storev(targetptr_final,
715 offset, targetptr_spec, newval_spec,
716 newval_final, cpu);
717 if (rseq_likely(!ret)) {
718 result = true;
719 break;
720 }
721 /* Retry if comparison fails or rseq aborts. */
722 }
723 if (_cpu)
724 *_cpu = cpu;
725 return result;
726 }
727
728 struct percpu_buffer_node *this_cpu_buffer_pop(struct percpu_buffer *buffer,
729 int *_cpu)
730 {
731 struct percpu_buffer_node *head;
732 int cpu;
733
734 for (;;) {
735 intptr_t *targetptr, newval;
736 intptr_t offset;
737 int ret;
738
739 cpu = rseq_cpu_start();
740 /* Load offset with single-copy atomicity. */
741 offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
742 if (offset == 0) {
743 head = NULL;
744 break;
745 }
746 head = RSEQ_READ_ONCE(buffer->c[cpu].array[offset - 1]);
747 newval = offset - 1;
748 targetptr = (intptr_t *)&buffer->c[cpu].offset;
749 ret = rseq_cmpeqv_cmpeqv_storev(targetptr, offset,
750 (intptr_t *)&buffer->c[cpu].array[offset - 1],
751 (intptr_t)head, newval, cpu);
752 if (rseq_likely(!ret))
753 break;
754 /* Retry if comparison fails or rseq aborts. */
755 }
756 if (_cpu)
757 *_cpu = cpu;
758 return head;
759 }
760
761 /*
762 * __percpu_buffer_pop is not safe against concurrent accesses. Should
763 * only be used on buffers that are not concurrently modified.
764 */
765 struct percpu_buffer_node *__percpu_buffer_pop(struct percpu_buffer *buffer,
766 int cpu)
767 {
768 struct percpu_buffer_node *head;
769 intptr_t offset;
770
771 offset = buffer->c[cpu].offset;
772 if (offset == 0)
773 return NULL;
774 head = buffer->c[cpu].array[offset - 1];
775 buffer->c[cpu].offset = offset - 1;
776 return head;
777 }
778
779 void *test_percpu_buffer_thread(void *arg)
780 {
781 long long i, reps;
782 struct percpu_buffer *buffer = (struct percpu_buffer *)arg;
783
784 if (!opt_disable_rseq && rseq_register_current_thread())
785 abort();
786
787 reps = opt_reps;
788 for (i = 0; i < reps; i++) {
789 struct percpu_buffer_node *node;
790
791 node = this_cpu_buffer_pop(buffer, NULL);
792 if (opt_yield)
793 sched_yield(); /* encourage shuffling */
794 if (node) {
795 if (!this_cpu_buffer_push(buffer, node, NULL)) {
796 /* Should increase buffer size. */
797 abort();
798 }
799 }
800 }
801
802 printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
803 (int) rseq_gettid(), nr_abort, signals_delivered);
804 if (!opt_disable_rseq && rseq_unregister_current_thread())
805 abort();
806
807 return NULL;
808 }
809
810 /* Simultaneous modification to a per-cpu buffer from many threads. */
811 void test_percpu_buffer(void)
812 {
813 const int num_threads = opt_threads;
814 int i, j, ret;
815 uint64_t sum = 0, expected_sum = 0;
816 struct percpu_buffer buffer;
817 pthread_t test_threads[num_threads];
818 cpu_set_t allowed_cpus;
819
820 memset(&buffer, 0, sizeof(buffer));
821
822 /* Generate list entries for every usable cpu. */
823 sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
824 for (i = 0; i < CPU_SETSIZE; i++) {
825 if (!CPU_ISSET(i, &allowed_cpus))
826 continue;
827 /* Worse-case is every item in same CPU. */
828 buffer.c[i].array =
829 malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE *
830 BUFFER_ITEM_PER_CPU);
831 assert(buffer.c[i].array);
832 buffer.c[i].buflen = CPU_SETSIZE * BUFFER_ITEM_PER_CPU;
833 for (j = 1; j <= BUFFER_ITEM_PER_CPU; j++) {
834 struct percpu_buffer_node *node;
835
836 expected_sum += j;
837
838 /*
839 * We could theoretically put the word-sized
840 * "data" directly in the buffer. However, we
841 * want to model objects that would not fit
842 * within a single word, so allocate an object
843 * for each node.
844 */
845 node = malloc(sizeof(*node));
846 assert(node);
847 node->data = j;
848 buffer.c[i].array[j - 1] = node;
849 buffer.c[i].offset++;
850 }
851 }
852
853 for (i = 0; i < num_threads; i++) {
854 ret = pthread_create(&test_threads[i], NULL,
855 test_percpu_buffer_thread, &buffer);
856 if (ret) {
857 errno = ret;
858 perror("pthread_create");
859 abort();
860 }
861 }
862
863 for (i = 0; i < num_threads; i++) {
864 ret = pthread_join(test_threads[i], NULL);
865 if (ret) {
866 errno = ret;
867 perror("pthread_join");
868 abort();
869 }
870 }
871
872 for (i = 0; i < CPU_SETSIZE; i++) {
873 struct percpu_buffer_node *node;
874
875 if (!CPU_ISSET(i, &allowed_cpus))
876 continue;
877
878 while ((node = __percpu_buffer_pop(&buffer, i))) {
879 sum += node->data;
880 free(node);
881 }
882 free(buffer.c[i].array);
883 }
884
885 /*
886 * All entries should now be accounted for (unless some external
887 * actor is interfering with our allowed affinity while this
888 * test is running).
889 */
890 assert(sum == expected_sum);
891 }
892
893 bool this_cpu_memcpy_buffer_push(struct percpu_memcpy_buffer *buffer,
894 struct percpu_memcpy_buffer_node item,
895 int *_cpu)
896 {
897 bool result = false;
898 int cpu;
899
900 for (;;) {
901 intptr_t *targetptr_final, newval_final, offset;
902 char *destptr, *srcptr;
903 size_t copylen;
904 int ret;
905
906 cpu = rseq_cpu_start();
907 /* Load offset with single-copy atomicity. */
908 offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
909 if (offset == buffer->c[cpu].buflen)
910 break;
911 destptr = (char *)&buffer->c[cpu].array[offset];
912 srcptr = (char *)&item;
913 /* copylen must be <= 4kB. */
914 copylen = sizeof(item);
915 newval_final = offset + 1;
916 targetptr_final = &buffer->c[cpu].offset;
917 if (opt_mb)
918 ret = rseq_cmpeqv_trymemcpy_storev_release(
919 targetptr_final, offset,
920 destptr, srcptr, copylen,
921 newval_final, cpu);
922 else
923 ret = rseq_cmpeqv_trymemcpy_storev(targetptr_final,
924 offset, destptr, srcptr, copylen,
925 newval_final, cpu);
926 if (rseq_likely(!ret)) {
927 result = true;
928 break;
929 }
930 /* Retry if comparison fails or rseq aborts. */
931 }
932 if (_cpu)
933 *_cpu = cpu;
934 return result;
935 }
936
937 bool this_cpu_memcpy_buffer_pop(struct percpu_memcpy_buffer *buffer,
938 struct percpu_memcpy_buffer_node *item,
939 int *_cpu)
940 {
941 bool result = false;
942 int cpu;
943
944 for (;;) {
945 intptr_t *targetptr_final, newval_final, offset;
946 char *destptr, *srcptr;
947 size_t copylen;
948 int ret;
949
950 cpu = rseq_cpu_start();
951 /* Load offset with single-copy atomicity. */
952 offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
953 if (offset == 0)
954 break;
955 destptr = (char *)item;
956 srcptr = (char *)&buffer->c[cpu].array[offset - 1];
957 /* copylen must be <= 4kB. */
958 copylen = sizeof(*item);
959 newval_final = offset - 1;
960 targetptr_final = &buffer->c[cpu].offset;
961 ret = rseq_cmpeqv_trymemcpy_storev(targetptr_final,
962 offset, destptr, srcptr, copylen,
963 newval_final, cpu);
964 if (rseq_likely(!ret)) {
965 result = true;
966 break;
967 }
968 /* Retry if comparison fails or rseq aborts. */
969 }
970 if (_cpu)
971 *_cpu = cpu;
972 return result;
973 }
974
975 /*
976 * __percpu_memcpy_buffer_pop is not safe against concurrent accesses. Should
977 * only be used on buffers that are not concurrently modified.
978 */
979 bool __percpu_memcpy_buffer_pop(struct percpu_memcpy_buffer *buffer,
980 struct percpu_memcpy_buffer_node *item,
981 int cpu)
982 {
983 intptr_t offset;
984
985 offset = buffer->c[cpu].offset;
986 if (offset == 0)
987 return false;
988 memcpy(item, &buffer->c[cpu].array[offset - 1], sizeof(*item));
989 buffer->c[cpu].offset = offset - 1;
990 return true;
991 }
992
993 void *test_percpu_memcpy_buffer_thread(void *arg)
994 {
995 long long i, reps;
996 struct percpu_memcpy_buffer *buffer = (struct percpu_memcpy_buffer *)arg;
997
998 if (!opt_disable_rseq && rseq_register_current_thread())
999 abort();
1000
1001 reps = opt_reps;
1002 for (i = 0; i < reps; i++) {
1003 struct percpu_memcpy_buffer_node item;
1004 bool result;
1005
1006 result = this_cpu_memcpy_buffer_pop(buffer, &item, NULL);
1007 if (opt_yield)
1008 sched_yield(); /* encourage shuffling */
1009 if (result) {
1010 if (!this_cpu_memcpy_buffer_push(buffer, item, NULL)) {
1011 /* Should increase buffer size. */
1012 abort();
1013 }
1014 }
1015 }
1016
1017 printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
1018 (int) rseq_gettid(), nr_abort, signals_delivered);
1019 if (!opt_disable_rseq && rseq_unregister_current_thread())
1020 abort();
1021
1022 return NULL;
1023 }
1024
1025 /* Simultaneous modification to a per-cpu buffer from many threads. */
1026 void test_percpu_memcpy_buffer(void)
1027 {
1028 const int num_threads = opt_threads;
1029 int i, j, ret;
1030 uint64_t sum = 0, expected_sum = 0;
1031 struct percpu_memcpy_buffer buffer;
1032 pthread_t test_threads[num_threads];
1033 cpu_set_t allowed_cpus;
1034
1035 memset(&buffer, 0, sizeof(buffer));
1036
1037 /* Generate list entries for every usable cpu. */
1038 sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
1039 for (i = 0; i < CPU_SETSIZE; i++) {
1040 if (!CPU_ISSET(i, &allowed_cpus))
1041 continue;
1042 /* Worse-case is every item in same CPU. */
1043 buffer.c[i].array =
1044 malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE *
1045 MEMCPY_BUFFER_ITEM_PER_CPU);
1046 assert(buffer.c[i].array);
1047 buffer.c[i].buflen = CPU_SETSIZE * MEMCPY_BUFFER_ITEM_PER_CPU;
1048 for (j = 1; j <= MEMCPY_BUFFER_ITEM_PER_CPU; j++) {
1049 expected_sum += 2 * j + 1;
1050
1051 /*
1052 * We could theoretically put the word-sized
1053 * "data" directly in the buffer. However, we
1054 * want to model objects that would not fit
1055 * within a single word, so allocate an object
1056 * for each node.
1057 */
1058 buffer.c[i].array[j - 1].data1 = j;
1059 buffer.c[i].array[j - 1].data2 = j + 1;
1060 buffer.c[i].offset++;
1061 }
1062 }
1063
1064 for (i = 0; i < num_threads; i++) {
1065 ret = pthread_create(&test_threads[i], NULL,
1066 test_percpu_memcpy_buffer_thread,
1067 &buffer);
1068 if (ret) {
1069 errno = ret;
1070 perror("pthread_create");
1071 abort();
1072 }
1073 }
1074
1075 for (i = 0; i < num_threads; i++) {
1076 ret = pthread_join(test_threads[i], NULL);
1077 if (ret) {
1078 errno = ret;
1079 perror("pthread_join");
1080 abort();
1081 }
1082 }
1083
1084 for (i = 0; i < CPU_SETSIZE; i++) {
1085 struct percpu_memcpy_buffer_node item;
1086
1087 if (!CPU_ISSET(i, &allowed_cpus))
1088 continue;
1089
1090 while (__percpu_memcpy_buffer_pop(&buffer, &item, i)) {
1091 sum += item.data1;
1092 sum += item.data2;
1093 }
1094 free(buffer.c[i].array);
1095 }
1096
1097 /*
1098 * All entries should now be accounted for (unless some external
1099 * actor is interfering with our allowed affinity while this
1100 * test is running).
1101 */
1102 assert(sum == expected_sum);
1103 }
1104
1105
1106 static void test_signal_interrupt_handler(__attribute__ ((unused)) int signo)
1107 {
1108 signals_delivered++;
1109 }
1110
1111 static int set_signal_handler(void)
1112 {
1113 int ret = 0;
1114 struct sigaction sa;
1115 sigset_t sigset;
1116
1117 ret = sigemptyset(&sigset);
1118 if (ret < 0) {
1119 perror("sigemptyset");
1120 return ret;
1121 }
1122
1123 sa.sa_handler = test_signal_interrupt_handler;
1124 sa.sa_mask = sigset;
1125 sa.sa_flags = 0;
1126 ret = sigaction(SIGUSR1, &sa, NULL);
1127 if (ret < 0) {
1128 perror("sigaction");
1129 return ret;
1130 }
1131
1132 printf_verbose("Signal handler set for SIGUSR1\n");
1133
1134 return ret;
1135 }
1136
1137 static void show_usage(char **argv)
1138 {
1139 printf("Usage : %s <OPTIONS>\n",
1140 argv[0]);
1141 printf("OPTIONS:\n");
1142 printf(" [-1 loops] Number of loops for delay injection 1\n");
1143 printf(" [-2 loops] Number of loops for delay injection 2\n");
1144 printf(" [-3 loops] Number of loops for delay injection 3\n");
1145 printf(" [-4 loops] Number of loops for delay injection 4\n");
1146 printf(" [-5 loops] Number of loops for delay injection 5\n");
1147 printf(" [-6 loops] Number of loops for delay injection 6\n");
1148 printf(" [-7 loops] Number of loops for delay injection 7 (-1 to enable -m)\n");
1149 printf(" [-8 loops] Number of loops for delay injection 8 (-1 to enable -m)\n");
1150 printf(" [-9 loops] Number of loops for delay injection 9 (-1 to enable -m)\n");
1151 printf(" [-m N] Yield/sleep/kill every modulo N (default 0: disabled) (>= 0)\n");
1152 printf(" [-y] Yield\n");
1153 printf(" [-k] Kill thread with signal\n");
1154 printf(" [-s S] S: =0: disabled (default), >0: sleep time (ms)\n");
1155 printf(" [-t N] Number of threads (default 200)\n");
1156 printf(" [-r N] Number of repetitions per thread (default 5000)\n");
1157 printf(" [-d] Disable rseq system call (no initialization)\n");
1158 printf(" [-D M] Disable rseq for each M threads\n");
1159 printf(" [-T test] Choose test: (s)pinlock, (l)ist, (b)uffer, (m)emcpy, (i)ncrement\n");
1160 printf(" [-M] Push into buffer and memcpy buffer with memory barriers.\n");
1161 printf(" [-c] Check if the rseq syscall is available.\n");
1162 printf(" [-v] Verbose output.\n");
1163 printf(" [-h] Show this help.\n");
1164 printf("\n");
1165 }
1166
1167 int main(int argc, char **argv)
1168 {
1169 int i;
1170
1171 for (i = 1; i < argc; i++) {
1172 if (argv[i][0] != '-')
1173 continue;
1174 switch (argv[i][1]) {
1175 case '1':
1176 case '2':
1177 case '3':
1178 case '4':
1179 case '5':
1180 case '6':
1181 case '7':
1182 case '8':
1183 case '9':
1184 if (argc < i + 2) {
1185 show_usage(argv);
1186 goto error;
1187 }
1188 loop_cnt[argv[i][1] - '0'] = atol(argv[i + 1]);
1189 i++;
1190 break;
1191 case 'm':
1192 if (argc < i + 2) {
1193 show_usage(argv);
1194 goto error;
1195 }
1196 opt_modulo = atol(argv[i + 1]);
1197 if (opt_modulo < 0) {
1198 show_usage(argv);
1199 goto error;
1200 }
1201 i++;
1202 break;
1203 case 's':
1204 if (argc < i + 2) {
1205 show_usage(argv);
1206 goto error;
1207 }
1208 opt_sleep = atol(argv[i + 1]);
1209 if (opt_sleep < 0) {
1210 show_usage(argv);
1211 goto error;
1212 }
1213 i++;
1214 break;
1215 case 'y':
1216 opt_yield = 1;
1217 break;
1218 case 'k':
1219 opt_signal = 1;
1220 break;
1221 case 'd':
1222 opt_disable_rseq = 1;
1223 break;
1224 case 'D':
1225 if (argc < i + 2) {
1226 show_usage(argv);
1227 goto error;
1228 }
1229 opt_disable_mod = atol(argv[i + 1]);
1230 if (opt_disable_mod < 0) {
1231 show_usage(argv);
1232 goto error;
1233 }
1234 i++;
1235 break;
1236 case 't':
1237 if (argc < i + 2) {
1238 show_usage(argv);
1239 goto error;
1240 }
1241 opt_threads = atol(argv[i + 1]);
1242 if (opt_threads < 0) {
1243 show_usage(argv);
1244 goto error;
1245 }
1246 i++;
1247 break;
1248 case 'r':
1249 if (argc < i + 2) {
1250 show_usage(argv);
1251 goto error;
1252 }
1253 opt_reps = atoll(argv[i + 1]);
1254 if (opt_reps < 0) {
1255 show_usage(argv);
1256 goto error;
1257 }
1258 i++;
1259 break;
1260 case 'h':
1261 show_usage(argv);
1262 goto end;
1263 case 'T':
1264 if (argc < i + 2) {
1265 show_usage(argv);
1266 goto error;
1267 }
1268 opt_test = *argv[i + 1];
1269 switch (opt_test) {
1270 case 's':
1271 case 'l':
1272 case 'i':
1273 case 'b':
1274 case 'm':
1275 break;
1276 default:
1277 show_usage(argv);
1278 goto error;
1279 }
1280 i++;
1281 break;
1282 case 'v':
1283 verbose = 1;
1284 break;
1285 case 'M':
1286 opt_mb = 1;
1287 break;
1288 case 'c':
1289 if (rseq_available()) {
1290 printf_verbose("The rseq syscall is available.\n");
1291 goto end;
1292 } else {
1293 printf_verbose("The rseq syscall is unavailable.\n");
1294 goto no_rseq;
1295 }
1296 default:
1297 show_usage(argv);
1298 goto error;
1299 }
1300 }
1301
1302 loop_cnt_1 = loop_cnt[1];
1303 loop_cnt_2 = loop_cnt[2];
1304 loop_cnt_3 = loop_cnt[3];
1305 loop_cnt_4 = loop_cnt[4];
1306 loop_cnt_5 = loop_cnt[5];
1307 loop_cnt_6 = loop_cnt[6];
1308
1309 if (set_signal_handler())
1310 goto error;
1311
1312 if (!opt_disable_rseq && rseq_register_current_thread())
1313 goto error;
1314 switch (opt_test) {
1315 case 's':
1316 printf_verbose("spinlock\n");
1317 test_percpu_spinlock();
1318 break;
1319 case 'l':
1320 printf_verbose("linked list\n");
1321 test_percpu_list();
1322 break;
1323 case 'b':
1324 printf_verbose("buffer\n");
1325 test_percpu_buffer();
1326 break;
1327 case 'm':
1328 printf_verbose("memcpy buffer\n");
1329 test_percpu_memcpy_buffer();
1330 break;
1331 case 'i':
1332 printf_verbose("counter increment\n");
1333 test_percpu_inc();
1334 break;
1335 }
1336 if (!opt_disable_rseq && rseq_unregister_current_thread())
1337 abort();
1338 end:
1339 return 0;
1340
1341 error:
1342 return -1;
1343
1344 no_rseq:
1345 return 2;
1346 }
Librseq
This page took 0.057004 seconds and 4 git commands to generate.