xref: /linux/tools/testing/selftests/rseq/param_test.c (revision fd56e5104a37f96e1b1ca42b4fd64fd49257fdce)
1 // SPDX-License-Identifier: LGPL-2.1
2 #define _GNU_SOURCE
3 #include <assert.h>
4 #include <linux/membarrier.h>
5 #include <pthread.h>
6 #include <sched.h>
7 #include <stdatomic.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 #include <stdbool.h>
20 
21 static inline pid_t rseq_gettid(void)
22 {
23 	return syscall(__NR_gettid);
24 }
25 
26 #define NR_INJECT	9
27 static int loop_cnt[NR_INJECT + 1];
28 
29 static int loop_cnt_1 asm("asm_loop_cnt_1") __attribute__((used));
30 static int loop_cnt_2 asm("asm_loop_cnt_2") __attribute__((used));
31 static int loop_cnt_3 asm("asm_loop_cnt_3") __attribute__((used));
32 static int loop_cnt_4 asm("asm_loop_cnt_4") __attribute__((used));
33 static int loop_cnt_5 asm("asm_loop_cnt_5") __attribute__((used));
34 static int loop_cnt_6 asm("asm_loop_cnt_6") __attribute__((used));
35 
36 static int opt_modulo, verbose;
37 
38 static int opt_yield, opt_signal, opt_sleep,
39 		opt_disable_rseq, opt_threads = 200,
40 		opt_disable_mod = 0, opt_test = 's';
41 
42 static long long opt_reps = 5000;
43 
44 static __thread __attribute__((tls_model("initial-exec")))
45 unsigned int signals_delivered;
46 
47 #ifndef BENCHMARK
48 
49 static __thread __attribute__((tls_model("initial-exec"), unused))
50 unsigned int yield_mod_cnt, nr_abort;
51 
52 #define printf_verbose(fmt, ...)			\
53 	do {						\
54 		if (verbose)				\
55 			printf(fmt, ## __VA_ARGS__);	\
56 	} while (0)
57 
58 #ifdef __i386__
59 
60 #define INJECT_ASM_REG	"eax"
61 
62 #define RSEQ_INJECT_CLOBBER \
63 	, INJECT_ASM_REG
64 
65 #define RSEQ_INJECT_ASM(n) \
66 	"mov asm_loop_cnt_" #n ", %%" INJECT_ASM_REG "\n\t" \
67 	"test %%" INJECT_ASM_REG ",%%" INJECT_ASM_REG "\n\t" \
68 	"jz 333f\n\t" \
69 	"222:\n\t" \
70 	"dec %%" INJECT_ASM_REG "\n\t" \
71 	"jnz 222b\n\t" \
72 	"333:\n\t"
73 
74 #elif defined(__x86_64__)
75 
76 #define INJECT_ASM_REG_P	"rax"
77 #define INJECT_ASM_REG		"eax"
78 
79 #define RSEQ_INJECT_CLOBBER \
80 	, INJECT_ASM_REG_P \
81 	, INJECT_ASM_REG
82 
83 #define RSEQ_INJECT_ASM(n) \
84 	"lea asm_loop_cnt_" #n "(%%rip), %%" INJECT_ASM_REG_P "\n\t" \
85 	"mov (%%" INJECT_ASM_REG_P "), %%" INJECT_ASM_REG "\n\t" \
86 	"test %%" INJECT_ASM_REG ",%%" INJECT_ASM_REG "\n\t" \
87 	"jz 333f\n\t" \
88 	"222:\n\t" \
89 	"dec %%" INJECT_ASM_REG "\n\t" \
90 	"jnz 222b\n\t" \
91 	"333:\n\t"
92 
93 #elif defined(__s390__)
94 
95 #define RSEQ_INJECT_INPUT \
96 	, [loop_cnt_1]"m"(loop_cnt[1]) \
97 	, [loop_cnt_2]"m"(loop_cnt[2]) \
98 	, [loop_cnt_3]"m"(loop_cnt[3]) \
99 	, [loop_cnt_4]"m"(loop_cnt[4]) \
100 	, [loop_cnt_5]"m"(loop_cnt[5]) \
101 	, [loop_cnt_6]"m"(loop_cnt[6])
102 
103 #define INJECT_ASM_REG	"r12"
104 
105 #define RSEQ_INJECT_CLOBBER \
106 	, INJECT_ASM_REG
107 
108 #define RSEQ_INJECT_ASM(n) \
109 	"l %%" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
110 	"ltr %%" INJECT_ASM_REG ", %%" INJECT_ASM_REG "\n\t" \
111 	"je 333f\n\t" \
112 	"222:\n\t" \
113 	"ahi %%" INJECT_ASM_REG ", -1\n\t" \
114 	"jnz 222b\n\t" \
115 	"333:\n\t"
116 
117 #elif defined(__ARMEL__)
118 
119 #define RSEQ_INJECT_INPUT \
120 	, [loop_cnt_1]"m"(loop_cnt[1]) \
121 	, [loop_cnt_2]"m"(loop_cnt[2]) \
122 	, [loop_cnt_3]"m"(loop_cnt[3]) \
123 	, [loop_cnt_4]"m"(loop_cnt[4]) \
124 	, [loop_cnt_5]"m"(loop_cnt[5]) \
125 	, [loop_cnt_6]"m"(loop_cnt[6])
126 
127 #define INJECT_ASM_REG	"r4"
128 
129 #define RSEQ_INJECT_CLOBBER \
130 	, INJECT_ASM_REG
131 
132 #define RSEQ_INJECT_ASM(n) \
133 	"ldr " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
134 	"cmp " INJECT_ASM_REG ", #0\n\t" \
135 	"beq 333f\n\t" \
136 	"222:\n\t" \
137 	"subs " INJECT_ASM_REG ", #1\n\t" \
138 	"bne 222b\n\t" \
139 	"333:\n\t"
140 
141 #elif defined(__AARCH64EL__)
142 
143 #define RSEQ_INJECT_INPUT \
144 	, [loop_cnt_1] "Qo" (loop_cnt[1]) \
145 	, [loop_cnt_2] "Qo" (loop_cnt[2]) \
146 	, [loop_cnt_3] "Qo" (loop_cnt[3]) \
147 	, [loop_cnt_4] "Qo" (loop_cnt[4]) \
148 	, [loop_cnt_5] "Qo" (loop_cnt[5]) \
149 	, [loop_cnt_6] "Qo" (loop_cnt[6])
150 
151 #define INJECT_ASM_REG	RSEQ_ASM_TMP_REG32
152 
153 #define RSEQ_INJECT_ASM(n) \
154 	"	ldr	" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n"	\
155 	"	cbz	" INJECT_ASM_REG ", 333f\n"			\
156 	"222:\n"							\
157 	"	sub	" INJECT_ASM_REG ", " INJECT_ASM_REG ", #1\n"	\
158 	"	cbnz	" INJECT_ASM_REG ", 222b\n"			\
159 	"333:\n"
160 
161 #elif defined(__PPC__)
162 
163 #define RSEQ_INJECT_INPUT \
164 	, [loop_cnt_1]"m"(loop_cnt[1]) \
165 	, [loop_cnt_2]"m"(loop_cnt[2]) \
166 	, [loop_cnt_3]"m"(loop_cnt[3]) \
167 	, [loop_cnt_4]"m"(loop_cnt[4]) \
168 	, [loop_cnt_5]"m"(loop_cnt[5]) \
169 	, [loop_cnt_6]"m"(loop_cnt[6])
170 
171 #define INJECT_ASM_REG	"r18"
172 
173 #define RSEQ_INJECT_CLOBBER \
174 	, INJECT_ASM_REG
175 
176 #define RSEQ_INJECT_ASM(n) \
177 	"lwz %%" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
178 	"cmpwi %%" INJECT_ASM_REG ", 0\n\t" \
179 	"beq 333f\n\t" \
180 	"222:\n\t" \
181 	"subic. %%" INJECT_ASM_REG ", %%" INJECT_ASM_REG ", 1\n\t" \
182 	"bne 222b\n\t" \
183 	"333:\n\t"
184 
185 #elif defined(__mips__)
186 
187 #define RSEQ_INJECT_INPUT \
188 	, [loop_cnt_1]"m"(loop_cnt[1]) \
189 	, [loop_cnt_2]"m"(loop_cnt[2]) \
190 	, [loop_cnt_3]"m"(loop_cnt[3]) \
191 	, [loop_cnt_4]"m"(loop_cnt[4]) \
192 	, [loop_cnt_5]"m"(loop_cnt[5]) \
193 	, [loop_cnt_6]"m"(loop_cnt[6])
194 
195 #define INJECT_ASM_REG	"$5"
196 
197 #define RSEQ_INJECT_CLOBBER \
198 	, INJECT_ASM_REG
199 
200 #define RSEQ_INJECT_ASM(n) \
201 	"lw " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
202 	"beqz " INJECT_ASM_REG ", 333f\n\t" \
203 	"222:\n\t" \
204 	"addiu " INJECT_ASM_REG ", -1\n\t" \
205 	"bnez " INJECT_ASM_REG ", 222b\n\t" \
206 	"333:\n\t"
207 #elif defined(__riscv)
208 
209 #define RSEQ_INJECT_INPUT \
210 	, [loop_cnt_1]"m"(loop_cnt[1]) \
211 	, [loop_cnt_2]"m"(loop_cnt[2]) \
212 	, [loop_cnt_3]"m"(loop_cnt[3]) \
213 	, [loop_cnt_4]"m"(loop_cnt[4]) \
214 	, [loop_cnt_5]"m"(loop_cnt[5]) \
215 	, [loop_cnt_6]"m"(loop_cnt[6])
216 
217 #define INJECT_ASM_REG	"t1"
218 
219 #define RSEQ_INJECT_CLOBBER \
220 	, INJECT_ASM_REG
221 
222 #define RSEQ_INJECT_ASM(n)					\
223 	"lw " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t"		\
224 	"beqz " INJECT_ASM_REG ", 333f\n\t"			\
225 	"222:\n\t"						\
226 	"addi  " INJECT_ASM_REG "," INJECT_ASM_REG ", -1\n\t"	\
227 	"bnez " INJECT_ASM_REG ", 222b\n\t"			\
228 	"333:\n\t"
229 #elif defined(__or1k__)
230 
231 #define RSEQ_INJECT_INPUT \
232 	, [loop_cnt_1]"m"(loop_cnt[1]) \
233 	, [loop_cnt_2]"m"(loop_cnt[2]) \
234 	, [loop_cnt_3]"m"(loop_cnt[3]) \
235 	, [loop_cnt_4]"m"(loop_cnt[4]) \
236 	, [loop_cnt_5]"m"(loop_cnt[5]) \
237 	, [loop_cnt_6]"m"(loop_cnt[6])
238 
239 #define INJECT_ASM_REG	"r31"
240 
241 #define RSEQ_INJECT_CLOBBER \
242 	, INJECT_ASM_REG
243 
244 #define RSEQ_INJECT_ASM(n)					\
245 	"l.lwz   " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t"	\
246 	"l.sfeqi " INJECT_ASM_REG ", 0\n\t"			\
247 	"l.bf 333f\n\t"						\
248 	" l.nop\n\t"						\
249 	"222:\n\t"						\
250 	"l.addi  " INJECT_ASM_REG "," INJECT_ASM_REG ", -1\n\t"	\
251 	"l.sfeqi " INJECT_ASM_REG ", 0\n\t"			\
252 	"l.bf 222f\n\t"						\
253 	" l.nop\n\t"						\
254 	"333:\n\t"
255 #else
256 #error unsupported target
257 #endif
258 
259 #define RSEQ_INJECT_FAILED \
260 	nr_abort++;
261 
262 #define RSEQ_INJECT_C(n) \
263 { \
264 	int loc_i, loc_nr_loops = loop_cnt[n]; \
265 	\
266 	for (loc_i = 0; loc_i < loc_nr_loops; loc_i++) { \
267 		rseq_barrier(); \
268 	} \
269 	if (loc_nr_loops == -1 && opt_modulo) { \
270 		if (yield_mod_cnt == opt_modulo - 1) { \
271 			if (opt_sleep > 0) \
272 				poll(NULL, 0, opt_sleep); \
273 			if (opt_yield) \
274 				sched_yield(); \
275 			if (opt_signal) \
276 				raise(SIGUSR1); \
277 			yield_mod_cnt = 0; \
278 		} else { \
279 			yield_mod_cnt++; \
280 		} \
281 	} \
282 }
283 
284 #else
285 
286 #define printf_verbose(fmt, ...)
287 
288 #endif /* BENCHMARK */
289 
290 #include "rseq.h"
291 
292 static enum rseq_mo opt_mo = RSEQ_MO_RELAXED;
293 
294 #ifdef RSEQ_ARCH_HAS_OFFSET_DEREF_ADDV
295 #define TEST_MEMBARRIER
296 
297 static int sys_membarrier(int cmd, int flags, int cpu_id)
298 {
299 	return syscall(__NR_membarrier, cmd, flags, cpu_id);
300 }
301 #endif
302 
303 #ifdef BUILDOPT_RSEQ_PERCPU_MM_CID
304 # define RSEQ_PERCPU	RSEQ_PERCPU_MM_CID
305 static
306 int get_current_cpu_id(void)
307 {
308 	return rseq_current_mm_cid();
309 }
310 static
311 bool rseq_validate_cpu_id(void)
312 {
313 	return rseq_mm_cid_available();
314 }
315 static
316 bool rseq_use_cpu_index(void)
317 {
318 	return false;	/* Use mm_cid */
319 }
320 # ifdef TEST_MEMBARRIER
321 /*
322  * Membarrier does not currently support targeting a mm_cid, so
323  * issue the barrier on all cpus.
324  */
325 static
326 int rseq_membarrier_expedited(int cpu)
327 {
328 	return sys_membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ,
329 			      0, 0);
330 }
331 # endif /* TEST_MEMBARRIER */
332 #else
333 # define RSEQ_PERCPU	RSEQ_PERCPU_CPU_ID
334 static
335 int get_current_cpu_id(void)
336 {
337 	return rseq_cpu_start();
338 }
339 static
340 bool rseq_validate_cpu_id(void)
341 {
342 	return rseq_current_cpu_raw() >= 0;
343 }
344 static
345 bool rseq_use_cpu_index(void)
346 {
347 	return true;	/* Use cpu_id as index. */
348 }
349 # ifdef TEST_MEMBARRIER
350 static
351 int rseq_membarrier_expedited(int cpu)
352 {
353 	return sys_membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ,
354 			      MEMBARRIER_CMD_FLAG_CPU, cpu);
355 }
356 # endif /* TEST_MEMBARRIER */
357 #endif
358 
359 struct percpu_lock_entry {
360 	intptr_t v;
361 } __attribute__((aligned(128)));
362 
363 struct percpu_lock {
364 	struct percpu_lock_entry c[CPU_SETSIZE];
365 };
366 
367 struct test_data_entry {
368 	intptr_t count;
369 } __attribute__((aligned(128)));
370 
371 struct spinlock_test_data {
372 	struct percpu_lock lock;
373 	struct test_data_entry c[CPU_SETSIZE];
374 };
375 
376 struct spinlock_thread_test_data {
377 	struct spinlock_test_data *data;
378 	long long reps;
379 	int reg;
380 };
381 
382 struct inc_test_data {
383 	struct test_data_entry c[CPU_SETSIZE];
384 };
385 
386 struct inc_thread_test_data {
387 	struct inc_test_data *data;
388 	long long reps;
389 	int reg;
390 };
391 
392 struct percpu_list_node {
393 	intptr_t data;
394 	struct percpu_list_node *next;
395 };
396 
397 struct percpu_list_entry {
398 	struct percpu_list_node *head;
399 } __attribute__((aligned(128)));
400 
401 struct percpu_list {
402 	struct percpu_list_entry c[CPU_SETSIZE];
403 };
404 
405 #define BUFFER_ITEM_PER_CPU	100
406 
407 struct percpu_buffer_node {
408 	intptr_t data;
409 };
410 
411 struct percpu_buffer_entry {
412 	intptr_t offset;
413 	intptr_t buflen;
414 	struct percpu_buffer_node **array;
415 } __attribute__((aligned(128)));
416 
417 struct percpu_buffer {
418 	struct percpu_buffer_entry c[CPU_SETSIZE];
419 };
420 
421 #define MEMCPY_BUFFER_ITEM_PER_CPU	100
422 
423 struct percpu_memcpy_buffer_node {
424 	intptr_t data1;
425 	uint64_t data2;
426 };
427 
428 struct percpu_memcpy_buffer_entry {
429 	intptr_t offset;
430 	intptr_t buflen;
431 	struct percpu_memcpy_buffer_node *array;
432 } __attribute__((aligned(128)));
433 
434 struct percpu_memcpy_buffer {
435 	struct percpu_memcpy_buffer_entry c[CPU_SETSIZE];
436 };
437 
438 /* A simple percpu spinlock. Grabs lock on current cpu. */
439 static int rseq_this_cpu_lock(struct percpu_lock *lock)
440 {
441 	int cpu;
442 
443 	for (;;) {
444 		int ret;
445 
446 		cpu = get_current_cpu_id();
447 		if (cpu < 0) {
448 			fprintf(stderr, "pid: %d: tid: %d, cpu: %d: cid: %d\n",
449 					getpid(), (int) rseq_gettid(), rseq_current_cpu_raw(), cpu);
450 			abort();
451 		}
452 		ret = rseq_cmpeqv_storev(RSEQ_MO_RELAXED, RSEQ_PERCPU,
453 					 &lock->c[cpu].v,
454 					 0, 1, cpu);
455 		if (rseq_likely(!ret))
456 			break;
457 		/* Retry if comparison fails or rseq aborts. */
458 	}
459 	/*
460 	 * Acquire semantic when taking lock after control dependency.
461 	 * Matches rseq_smp_store_release().
462 	 */
463 	rseq_smp_acquire__after_ctrl_dep();
464 	return cpu;
465 }
466 
467 static void rseq_percpu_unlock(struct percpu_lock *lock, int cpu)
468 {
469 	assert(lock->c[cpu].v == 1);
470 	/*
471 	 * Release lock, with release semantic. Matches
472 	 * rseq_smp_acquire__after_ctrl_dep().
473 	 */
474 	rseq_smp_store_release(&lock->c[cpu].v, 0);
475 }
476 
477 void *test_percpu_spinlock_thread(void *arg)
478 {
479 	struct spinlock_thread_test_data *thread_data = arg;
480 	struct spinlock_test_data *data = thread_data->data;
481 	long long i, reps;
482 
483 	if (!opt_disable_rseq && thread_data->reg &&
484 	    rseq_register_current_thread())
485 		abort();
486 	reps = thread_data->reps;
487 	for (i = 0; i < reps; i++) {
488 		int cpu = rseq_this_cpu_lock(&data->lock);
489 		data->c[cpu].count++;
490 		rseq_percpu_unlock(&data->lock, cpu);
491 #ifndef BENCHMARK
492 		if (i != 0 && !(i % (reps / 10)))
493 			printf_verbose("tid %d: count %lld\n",
494 				       (int) rseq_gettid(), i);
495 #endif
496 	}
497 	printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
498 		       (int) rseq_gettid(), nr_abort, signals_delivered);
499 	if (!opt_disable_rseq && thread_data->reg &&
500 	    rseq_unregister_current_thread())
501 		abort();
502 	return NULL;
503 }
504 
505 /*
506  * A simple test which implements a sharded counter using a per-cpu
507  * lock.  Obviously real applications might prefer to simply use a
508  * per-cpu increment; however, this is reasonable for a test and the
509  * lock can be extended to synchronize more complicated operations.
510  */
511 void test_percpu_spinlock(void)
512 {
513 	const int num_threads = opt_threads;
514 	int i, ret;
515 	uint64_t sum;
516 	pthread_t test_threads[num_threads];
517 	struct spinlock_test_data data;
518 	struct spinlock_thread_test_data thread_data[num_threads];
519 
520 	memset(&data, 0, sizeof(data));
521 	for (i = 0; i < num_threads; i++) {
522 		thread_data[i].reps = opt_reps;
523 		if (opt_disable_mod <= 0 || (i % opt_disable_mod))
524 			thread_data[i].reg = 1;
525 		else
526 			thread_data[i].reg = 0;
527 		thread_data[i].data = &data;
528 		ret = pthread_create(&test_threads[i], NULL,
529 				     test_percpu_spinlock_thread,
530 				     &thread_data[i]);
531 		if (ret) {
532 			errno = ret;
533 			perror("pthread_create");
534 			abort();
535 		}
536 	}
537 
538 	for (i = 0; i < num_threads; i++) {
539 		ret = pthread_join(test_threads[i], NULL);
540 		if (ret) {
541 			errno = ret;
542 			perror("pthread_join");
543 			abort();
544 		}
545 	}
546 
547 	sum = 0;
548 	for (i = 0; i < CPU_SETSIZE; i++)
549 		sum += data.c[i].count;
550 
551 	assert(sum == (uint64_t)opt_reps * num_threads);
552 }
553 
554 void *test_percpu_inc_thread(void *arg)
555 {
556 	struct inc_thread_test_data *thread_data = arg;
557 	struct inc_test_data *data = thread_data->data;
558 	long long i, reps;
559 
560 	if (!opt_disable_rseq && thread_data->reg &&
561 	    rseq_register_current_thread())
562 		abort();
563 	reps = thread_data->reps;
564 	for (i = 0; i < reps; i++) {
565 		int ret;
566 
567 		do {
568 			int cpu;
569 
570 			cpu = get_current_cpu_id();
571 			ret = rseq_addv(RSEQ_MO_RELAXED, RSEQ_PERCPU,
572 					&data->c[cpu].count, 1, cpu);
573 		} while (rseq_unlikely(ret));
574 #ifndef BENCHMARK
575 		if (i != 0 && !(i % (reps / 10)))
576 			printf_verbose("tid %d: count %lld\n",
577 				       (int) rseq_gettid(), i);
578 #endif
579 	}
580 	printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
581 		       (int) rseq_gettid(), nr_abort, signals_delivered);
582 	if (!opt_disable_rseq && thread_data->reg &&
583 	    rseq_unregister_current_thread())
584 		abort();
585 	return NULL;
586 }
587 
588 void test_percpu_inc(void)
589 {
590 	const int num_threads = opt_threads;
591 	int i, ret;
592 	uint64_t sum;
593 	pthread_t test_threads[num_threads];
594 	struct inc_test_data data;
595 	struct inc_thread_test_data thread_data[num_threads];
596 
597 	memset(&data, 0, sizeof(data));
598 	for (i = 0; i < num_threads; i++) {
599 		thread_data[i].reps = opt_reps;
600 		if (opt_disable_mod <= 0 || (i % opt_disable_mod))
601 			thread_data[i].reg = 1;
602 		else
603 			thread_data[i].reg = 0;
604 		thread_data[i].data = &data;
605 		ret = pthread_create(&test_threads[i], NULL,
606 				     test_percpu_inc_thread,
607 				     &thread_data[i]);
608 		if (ret) {
609 			errno = ret;
610 			perror("pthread_create");
611 			abort();
612 		}
613 	}
614 
615 	for (i = 0; i < num_threads; i++) {
616 		ret = pthread_join(test_threads[i], NULL);
617 		if (ret) {
618 			errno = ret;
619 			perror("pthread_join");
620 			abort();
621 		}
622 	}
623 
624 	sum = 0;
625 	for (i = 0; i < CPU_SETSIZE; i++)
626 		sum += data.c[i].count;
627 
628 	assert(sum == (uint64_t)opt_reps * num_threads);
629 }
630 
631 void this_cpu_list_push(struct percpu_list *list,
632 			struct percpu_list_node *node,
633 			int *_cpu)
634 {
635 	int cpu;
636 
637 	for (;;) {
638 		intptr_t *targetptr, newval, expect;
639 		int ret;
640 
641 		cpu = get_current_cpu_id();
642 		/* Load list->c[cpu].head with single-copy atomicity. */
643 		expect = (intptr_t)RSEQ_READ_ONCE(list->c[cpu].head);
644 		newval = (intptr_t)node;
645 		targetptr = (intptr_t *)&list->c[cpu].head;
646 		node->next = (struct percpu_list_node *)expect;
647 		ret = rseq_cmpeqv_storev(RSEQ_MO_RELAXED, RSEQ_PERCPU,
648 					 targetptr, expect, newval, cpu);
649 		if (rseq_likely(!ret))
650 			break;
651 		/* Retry if comparison fails or rseq aborts. */
652 	}
653 	if (_cpu)
654 		*_cpu = cpu;
655 }
656 
657 /*
658  * Unlike a traditional lock-less linked list; the availability of a
659  * rseq primitive allows us to implement pop without concerns over
660  * ABA-type races.
661  */
662 struct percpu_list_node *this_cpu_list_pop(struct percpu_list *list,
663 					   int *_cpu)
664 {
665 	struct percpu_list_node *node = NULL;
666 	int cpu;
667 
668 	for (;;) {
669 		struct percpu_list_node *head;
670 		intptr_t *targetptr, expectnot, *load;
671 		long offset;
672 		int ret;
673 
674 		cpu = get_current_cpu_id();
675 		targetptr = (intptr_t *)&list->c[cpu].head;
676 		expectnot = (intptr_t)NULL;
677 		offset = offsetof(struct percpu_list_node, next);
678 		load = (intptr_t *)&head;
679 		ret = rseq_cmpnev_storeoffp_load(RSEQ_MO_RELAXED, RSEQ_PERCPU,
680 						 targetptr, expectnot,
681 						 offset, load, cpu);
682 		if (rseq_likely(!ret)) {
683 			node = head;
684 			break;
685 		}
686 		if (ret > 0)
687 			break;
688 		/* Retry if rseq aborts. */
689 	}
690 	if (_cpu)
691 		*_cpu = cpu;
692 	return node;
693 }
694 
695 /*
696  * __percpu_list_pop is not safe against concurrent accesses. Should
697  * only be used on lists that are not concurrently modified.
698  */
699 struct percpu_list_node *__percpu_list_pop(struct percpu_list *list, int cpu)
700 {
701 	struct percpu_list_node *node;
702 
703 	node = list->c[cpu].head;
704 	if (!node)
705 		return NULL;
706 	list->c[cpu].head = node->next;
707 	return node;
708 }
709 
710 void *test_percpu_list_thread(void *arg)
711 {
712 	long long i, reps;
713 	struct percpu_list *list = (struct percpu_list *)arg;
714 
715 	if (!opt_disable_rseq && rseq_register_current_thread())
716 		abort();
717 
718 	reps = opt_reps;
719 	for (i = 0; i < reps; i++) {
720 		struct percpu_list_node *node;
721 
722 		node = this_cpu_list_pop(list, NULL);
723 		if (opt_yield)
724 			sched_yield();  /* encourage shuffling */
725 		if (node)
726 			this_cpu_list_push(list, node, NULL);
727 	}
728 
729 	printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
730 		       (int) rseq_gettid(), nr_abort, signals_delivered);
731 	if (!opt_disable_rseq && rseq_unregister_current_thread())
732 		abort();
733 
734 	return NULL;
735 }
736 
737 /* Simultaneous modification to a per-cpu linked list from many threads.  */
738 void test_percpu_list(void)
739 {
740 	const int num_threads = opt_threads;
741 	int i, j, ret;
742 	uint64_t sum = 0, expected_sum = 0;
743 	struct percpu_list list;
744 	pthread_t test_threads[num_threads];
745 	cpu_set_t allowed_cpus;
746 
747 	memset(&list, 0, sizeof(list));
748 
749 	/* Generate list entries for every usable cpu. */
750 	sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
751 	for (i = 0; i < CPU_SETSIZE; i++) {
752 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
753 			continue;
754 		for (j = 1; j <= 100; j++) {
755 			struct percpu_list_node *node;
756 
757 			expected_sum += j;
758 
759 			node = malloc(sizeof(*node));
760 			assert(node);
761 			node->data = j;
762 			node->next = list.c[i].head;
763 			list.c[i].head = node;
764 		}
765 	}
766 
767 	for (i = 0; i < num_threads; i++) {
768 		ret = pthread_create(&test_threads[i], NULL,
769 				     test_percpu_list_thread, &list);
770 		if (ret) {
771 			errno = ret;
772 			perror("pthread_create");
773 			abort();
774 		}
775 	}
776 
777 	for (i = 0; i < num_threads; i++) {
778 		ret = pthread_join(test_threads[i], NULL);
779 		if (ret) {
780 			errno = ret;
781 			perror("pthread_join");
782 			abort();
783 		}
784 	}
785 
786 	for (i = 0; i < CPU_SETSIZE; i++) {
787 		struct percpu_list_node *node;
788 
789 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
790 			continue;
791 
792 		while ((node = __percpu_list_pop(&list, i))) {
793 			sum += node->data;
794 			free(node);
795 		}
796 	}
797 
798 	/*
799 	 * All entries should now be accounted for (unless some external
800 	 * actor is interfering with our allowed affinity while this
801 	 * test is running).
802 	 */
803 	assert(sum == expected_sum);
804 }
805 
806 bool this_cpu_buffer_push(struct percpu_buffer *buffer,
807 			  struct percpu_buffer_node *node,
808 			  int *_cpu)
809 {
810 	bool result = false;
811 	int cpu;
812 
813 	for (;;) {
814 		intptr_t *targetptr_spec, newval_spec;
815 		intptr_t *targetptr_final, newval_final;
816 		intptr_t offset;
817 		int ret;
818 
819 		cpu = get_current_cpu_id();
820 		offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
821 		if (offset == buffer->c[cpu].buflen)
822 			break;
823 		newval_spec = (intptr_t)node;
824 		targetptr_spec = (intptr_t *)&buffer->c[cpu].array[offset];
825 		newval_final = offset + 1;
826 		targetptr_final = &buffer->c[cpu].offset;
827 		ret = rseq_cmpeqv_trystorev_storev(opt_mo, RSEQ_PERCPU,
828 			targetptr_final, offset, targetptr_spec,
829 			newval_spec, newval_final, cpu);
830 		if (rseq_likely(!ret)) {
831 			result = true;
832 			break;
833 		}
834 		/* Retry if comparison fails or rseq aborts. */
835 	}
836 	if (_cpu)
837 		*_cpu = cpu;
838 	return result;
839 }
840 
841 struct percpu_buffer_node *this_cpu_buffer_pop(struct percpu_buffer *buffer,
842 					       int *_cpu)
843 {
844 	struct percpu_buffer_node *head;
845 	int cpu;
846 
847 	for (;;) {
848 		intptr_t *targetptr, newval;
849 		intptr_t offset;
850 		int ret;
851 
852 		cpu = get_current_cpu_id();
853 		/* Load offset with single-copy atomicity. */
854 		offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
855 		if (offset == 0) {
856 			head = NULL;
857 			break;
858 		}
859 		head = RSEQ_READ_ONCE(buffer->c[cpu].array[offset - 1]);
860 		newval = offset - 1;
861 		targetptr = (intptr_t *)&buffer->c[cpu].offset;
862 		ret = rseq_cmpeqv_cmpeqv_storev(RSEQ_MO_RELAXED, RSEQ_PERCPU,
863 			targetptr, offset,
864 			(intptr_t *)&buffer->c[cpu].array[offset - 1],
865 			(intptr_t)head, newval, cpu);
866 		if (rseq_likely(!ret))
867 			break;
868 		/* Retry if comparison fails or rseq aborts. */
869 	}
870 	if (_cpu)
871 		*_cpu = cpu;
872 	return head;
873 }
874 
875 /*
876  * __percpu_buffer_pop is not safe against concurrent accesses. Should
877  * only be used on buffers that are not concurrently modified.
878  */
879 struct percpu_buffer_node *__percpu_buffer_pop(struct percpu_buffer *buffer,
880 					       int cpu)
881 {
882 	struct percpu_buffer_node *head;
883 	intptr_t offset;
884 
885 	offset = buffer->c[cpu].offset;
886 	if (offset == 0)
887 		return NULL;
888 	head = buffer->c[cpu].array[offset - 1];
889 	buffer->c[cpu].offset = offset - 1;
890 	return head;
891 }
892 
893 void *test_percpu_buffer_thread(void *arg)
894 {
895 	long long i, reps;
896 	struct percpu_buffer *buffer = (struct percpu_buffer *)arg;
897 
898 	if (!opt_disable_rseq && rseq_register_current_thread())
899 		abort();
900 
901 	reps = opt_reps;
902 	for (i = 0; i < reps; i++) {
903 		struct percpu_buffer_node *node;
904 
905 		node = this_cpu_buffer_pop(buffer, NULL);
906 		if (opt_yield)
907 			sched_yield();  /* encourage shuffling */
908 		if (node) {
909 			if (!this_cpu_buffer_push(buffer, node, NULL)) {
910 				/* Should increase buffer size. */
911 				abort();
912 			}
913 		}
914 	}
915 
916 	printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
917 		       (int) rseq_gettid(), nr_abort, signals_delivered);
918 	if (!opt_disable_rseq && rseq_unregister_current_thread())
919 		abort();
920 
921 	return NULL;
922 }
923 
924 /* Simultaneous modification to a per-cpu buffer from many threads.  */
925 void test_percpu_buffer(void)
926 {
927 	const int num_threads = opt_threads;
928 	int i, j, ret;
929 	uint64_t sum = 0, expected_sum = 0;
930 	struct percpu_buffer buffer;
931 	pthread_t test_threads[num_threads];
932 	cpu_set_t allowed_cpus;
933 
934 	memset(&buffer, 0, sizeof(buffer));
935 
936 	/* Generate list entries for every usable cpu. */
937 	sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
938 	for (i = 0; i < CPU_SETSIZE; i++) {
939 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
940 			continue;
941 		/* Worse-case is every item in same CPU. */
942 		buffer.c[i].array =
943 			malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE *
944 			       BUFFER_ITEM_PER_CPU);
945 		assert(buffer.c[i].array);
946 		buffer.c[i].buflen = CPU_SETSIZE * BUFFER_ITEM_PER_CPU;
947 		for (j = 1; j <= BUFFER_ITEM_PER_CPU; j++) {
948 			struct percpu_buffer_node *node;
949 
950 			expected_sum += j;
951 
952 			/*
953 			 * We could theoretically put the word-sized
954 			 * "data" directly in the buffer. However, we
955 			 * want to model objects that would not fit
956 			 * within a single word, so allocate an object
957 			 * for each node.
958 			 */
959 			node = malloc(sizeof(*node));
960 			assert(node);
961 			node->data = j;
962 			buffer.c[i].array[j - 1] = node;
963 			buffer.c[i].offset++;
964 		}
965 	}
966 
967 	for (i = 0; i < num_threads; i++) {
968 		ret = pthread_create(&test_threads[i], NULL,
969 				     test_percpu_buffer_thread, &buffer);
970 		if (ret) {
971 			errno = ret;
972 			perror("pthread_create");
973 			abort();
974 		}
975 	}
976 
977 	for (i = 0; i < num_threads; i++) {
978 		ret = pthread_join(test_threads[i], NULL);
979 		if (ret) {
980 			errno = ret;
981 			perror("pthread_join");
982 			abort();
983 		}
984 	}
985 
986 	for (i = 0; i < CPU_SETSIZE; i++) {
987 		struct percpu_buffer_node *node;
988 
989 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
990 			continue;
991 
992 		while ((node = __percpu_buffer_pop(&buffer, i))) {
993 			sum += node->data;
994 			free(node);
995 		}
996 		free(buffer.c[i].array);
997 	}
998 
999 	/*
1000 	 * All entries should now be accounted for (unless some external
1001 	 * actor is interfering with our allowed affinity while this
1002 	 * test is running).
1003 	 */
1004 	assert(sum == expected_sum);
1005 }
1006 
1007 bool this_cpu_memcpy_buffer_push(struct percpu_memcpy_buffer *buffer,
1008 				 struct percpu_memcpy_buffer_node item,
1009 				 int *_cpu)
1010 {
1011 	bool result = false;
1012 	int cpu;
1013 
1014 	for (;;) {
1015 		intptr_t *targetptr_final, newval_final, offset;
1016 		char *destptr, *srcptr;
1017 		size_t copylen;
1018 		int ret;
1019 
1020 		cpu = get_current_cpu_id();
1021 		/* Load offset with single-copy atomicity. */
1022 		offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
1023 		if (offset == buffer->c[cpu].buflen)
1024 			break;
1025 		destptr = (char *)&buffer->c[cpu].array[offset];
1026 		srcptr = (char *)&item;
1027 		/* copylen must be <= 4kB. */
1028 		copylen = sizeof(item);
1029 		newval_final = offset + 1;
1030 		targetptr_final = &buffer->c[cpu].offset;
1031 		ret = rseq_cmpeqv_trymemcpy_storev(
1032 			opt_mo, RSEQ_PERCPU,
1033 			targetptr_final, offset,
1034 			destptr, srcptr, copylen,
1035 			newval_final, cpu);
1036 		if (rseq_likely(!ret)) {
1037 			result = true;
1038 			break;
1039 		}
1040 		/* Retry if comparison fails or rseq aborts. */
1041 	}
1042 	if (_cpu)
1043 		*_cpu = cpu;
1044 	return result;
1045 }
1046 
1047 bool this_cpu_memcpy_buffer_pop(struct percpu_memcpy_buffer *buffer,
1048 				struct percpu_memcpy_buffer_node *item,
1049 				int *_cpu)
1050 {
1051 	bool result = false;
1052 	int cpu;
1053 
1054 	for (;;) {
1055 		intptr_t *targetptr_final, newval_final, offset;
1056 		char *destptr, *srcptr;
1057 		size_t copylen;
1058 		int ret;
1059 
1060 		cpu = get_current_cpu_id();
1061 		/* Load offset with single-copy atomicity. */
1062 		offset = RSEQ_READ_ONCE(buffer->c[cpu].offset);
1063 		if (offset == 0)
1064 			break;
1065 		destptr = (char *)item;
1066 		srcptr = (char *)&buffer->c[cpu].array[offset - 1];
1067 		/* copylen must be <= 4kB. */
1068 		copylen = sizeof(*item);
1069 		newval_final = offset - 1;
1070 		targetptr_final = &buffer->c[cpu].offset;
1071 		ret = rseq_cmpeqv_trymemcpy_storev(RSEQ_MO_RELAXED, RSEQ_PERCPU,
1072 			targetptr_final, offset, destptr, srcptr, copylen,
1073 			newval_final, cpu);
1074 		if (rseq_likely(!ret)) {
1075 			result = true;
1076 			break;
1077 		}
1078 		/* Retry if comparison fails or rseq aborts. */
1079 	}
1080 	if (_cpu)
1081 		*_cpu = cpu;
1082 	return result;
1083 }
1084 
1085 /*
1086  * __percpu_memcpy_buffer_pop is not safe against concurrent accesses. Should
1087  * only be used on buffers that are not concurrently modified.
1088  */
1089 bool __percpu_memcpy_buffer_pop(struct percpu_memcpy_buffer *buffer,
1090 				struct percpu_memcpy_buffer_node *item,
1091 				int cpu)
1092 {
1093 	intptr_t offset;
1094 
1095 	offset = buffer->c[cpu].offset;
1096 	if (offset == 0)
1097 		return false;
1098 	memcpy(item, &buffer->c[cpu].array[offset - 1], sizeof(*item));
1099 	buffer->c[cpu].offset = offset - 1;
1100 	return true;
1101 }
1102 
1103 void *test_percpu_memcpy_buffer_thread(void *arg)
1104 {
1105 	long long i, reps;
1106 	struct percpu_memcpy_buffer *buffer = (struct percpu_memcpy_buffer *)arg;
1107 
1108 	if (!opt_disable_rseq && rseq_register_current_thread())
1109 		abort();
1110 
1111 	reps = opt_reps;
1112 	for (i = 0; i < reps; i++) {
1113 		struct percpu_memcpy_buffer_node item;
1114 		bool result;
1115 
1116 		result = this_cpu_memcpy_buffer_pop(buffer, &item, NULL);
1117 		if (opt_yield)
1118 			sched_yield();  /* encourage shuffling */
1119 		if (result) {
1120 			if (!this_cpu_memcpy_buffer_push(buffer, item, NULL)) {
1121 				/* Should increase buffer size. */
1122 				abort();
1123 			}
1124 		}
1125 	}
1126 
1127 	printf_verbose("tid %d: number of rseq abort: %d, signals delivered: %u\n",
1128 		       (int) rseq_gettid(), nr_abort, signals_delivered);
1129 	if (!opt_disable_rseq && rseq_unregister_current_thread())
1130 		abort();
1131 
1132 	return NULL;
1133 }
1134 
1135 /* Simultaneous modification to a per-cpu buffer from many threads.  */
1136 void test_percpu_memcpy_buffer(void)
1137 {
1138 	const int num_threads = opt_threads;
1139 	int i, j, ret;
1140 	uint64_t sum = 0, expected_sum = 0;
1141 	struct percpu_memcpy_buffer buffer;
1142 	pthread_t test_threads[num_threads];
1143 	cpu_set_t allowed_cpus;
1144 
1145 	memset(&buffer, 0, sizeof(buffer));
1146 
1147 	/* Generate list entries for every usable cpu. */
1148 	sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
1149 	for (i = 0; i < CPU_SETSIZE; i++) {
1150 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
1151 			continue;
1152 		/* Worse-case is every item in same CPU. */
1153 		buffer.c[i].array =
1154 			malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE *
1155 			       MEMCPY_BUFFER_ITEM_PER_CPU);
1156 		assert(buffer.c[i].array);
1157 		buffer.c[i].buflen = CPU_SETSIZE * MEMCPY_BUFFER_ITEM_PER_CPU;
1158 		for (j = 1; j <= MEMCPY_BUFFER_ITEM_PER_CPU; j++) {
1159 			expected_sum += 2 * j + 1;
1160 
1161 			/*
1162 			 * We could theoretically put the word-sized
1163 			 * "data" directly in the buffer. However, we
1164 			 * want to model objects that would not fit
1165 			 * within a single word, so allocate an object
1166 			 * for each node.
1167 			 */
1168 			buffer.c[i].array[j - 1].data1 = j;
1169 			buffer.c[i].array[j - 1].data2 = j + 1;
1170 			buffer.c[i].offset++;
1171 		}
1172 	}
1173 
1174 	for (i = 0; i < num_threads; i++) {
1175 		ret = pthread_create(&test_threads[i], NULL,
1176 				     test_percpu_memcpy_buffer_thread,
1177 				     &buffer);
1178 		if (ret) {
1179 			errno = ret;
1180 			perror("pthread_create");
1181 			abort();
1182 		}
1183 	}
1184 
1185 	for (i = 0; i < num_threads; i++) {
1186 		ret = pthread_join(test_threads[i], NULL);
1187 		if (ret) {
1188 			errno = ret;
1189 			perror("pthread_join");
1190 			abort();
1191 		}
1192 	}
1193 
1194 	for (i = 0; i < CPU_SETSIZE; i++) {
1195 		struct percpu_memcpy_buffer_node item;
1196 
1197 		if (rseq_use_cpu_index() && !CPU_ISSET(i, &allowed_cpus))
1198 			continue;
1199 
1200 		while (__percpu_memcpy_buffer_pop(&buffer, &item, i)) {
1201 			sum += item.data1;
1202 			sum += item.data2;
1203 		}
1204 		free(buffer.c[i].array);
1205 	}
1206 
1207 	/*
1208 	 * All entries should now be accounted for (unless some external
1209 	 * actor is interfering with our allowed affinity while this
1210 	 * test is running).
1211 	 */
1212 	assert(sum == expected_sum);
1213 }
1214 
1215 static void test_signal_interrupt_handler(int signo)
1216 {
1217 	signals_delivered++;
1218 }
1219 
1220 static int set_signal_handler(void)
1221 {
1222 	int ret = 0;
1223 	struct sigaction sa;
1224 	sigset_t sigset;
1225 
1226 	ret = sigemptyset(&sigset);
1227 	if (ret < 0) {
1228 		perror("sigemptyset");
1229 		return ret;
1230 	}
1231 
1232 	sa.sa_handler = test_signal_interrupt_handler;
1233 	sa.sa_mask = sigset;
1234 	sa.sa_flags = 0;
1235 	ret = sigaction(SIGUSR1, &sa, NULL);
1236 	if (ret < 0) {
1237 		perror("sigaction");
1238 		return ret;
1239 	}
1240 
1241 	printf_verbose("Signal handler set for SIGUSR1\n");
1242 
1243 	return ret;
1244 }
1245 
1246 /* Test MEMBARRIER_CMD_PRIVATE_RESTART_RSEQ_ON_CPU membarrier command. */
1247 #ifdef TEST_MEMBARRIER
1248 struct test_membarrier_thread_args {
1249 	int stop;
1250 	intptr_t percpu_list_ptr;
1251 };
1252 
1253 /* Worker threads modify data in their "active" percpu lists. */
1254 void *test_membarrier_worker_thread(void *arg)
1255 {
1256 	struct test_membarrier_thread_args *args =
1257 		(struct test_membarrier_thread_args *)arg;
1258 	const int iters = opt_reps;
1259 	int i;
1260 
1261 	if (rseq_register_current_thread()) {
1262 		fprintf(stderr, "Error: rseq_register_current_thread(...) failed(%d): %s\n",
1263 			errno, strerror(errno));
1264 		abort();
1265 	}
1266 
1267 	/* Wait for initialization. */
1268 	while (!__atomic_load_n(&args->percpu_list_ptr, __ATOMIC_ACQUIRE)) {}
1269 
1270 	for (i = 0; i < iters; ++i) {
1271 		int ret;
1272 
1273 		do {
1274 			int cpu = get_current_cpu_id();
1275 
1276 			ret = rseq_offset_deref_addv(RSEQ_MO_RELAXED, RSEQ_PERCPU,
1277 				&args->percpu_list_ptr,
1278 				sizeof(struct percpu_list_entry) * cpu, 1, cpu);
1279 		} while (rseq_unlikely(ret));
1280 	}
1281 
1282 	if (rseq_unregister_current_thread()) {
1283 		fprintf(stderr, "Error: rseq_unregister_current_thread(...) failed(%d): %s\n",
1284 			errno, strerror(errno));
1285 		abort();
1286 	}
1287 	return NULL;
1288 }
1289 
1290 void test_membarrier_init_percpu_list(struct percpu_list *list)
1291 {
1292 	int i;
1293 
1294 	memset(list, 0, sizeof(*list));
1295 	for (i = 0; i < CPU_SETSIZE; i++) {
1296 		struct percpu_list_node *node;
1297 
1298 		node = malloc(sizeof(*node));
1299 		assert(node);
1300 		node->data = 0;
1301 		node->next = NULL;
1302 		list->c[i].head = node;
1303 	}
1304 }
1305 
1306 void test_membarrier_free_percpu_list(struct percpu_list *list)
1307 {
1308 	int i;
1309 
1310 	for (i = 0; i < CPU_SETSIZE; i++)
1311 		free(list->c[i].head);
1312 }
1313 
1314 /*
1315  * The manager thread swaps per-cpu lists that worker threads see,
1316  * and validates that there are no unexpected modifications.
1317  */
1318 void *test_membarrier_manager_thread(void *arg)
1319 {
1320 	struct test_membarrier_thread_args *args =
1321 		(struct test_membarrier_thread_args *)arg;
1322 	struct percpu_list list_a, list_b;
1323 	intptr_t expect_a = 0, expect_b = 0;
1324 	int cpu_a = 0, cpu_b = 0;
1325 
1326 	if (rseq_register_current_thread()) {
1327 		fprintf(stderr, "Error: rseq_register_current_thread(...) failed(%d): %s\n",
1328 			errno, strerror(errno));
1329 		abort();
1330 	}
1331 
1332 	/* Init lists. */
1333 	test_membarrier_init_percpu_list(&list_a);
1334 	test_membarrier_init_percpu_list(&list_b);
1335 
1336 	__atomic_store_n(&args->percpu_list_ptr, (intptr_t)&list_a, __ATOMIC_RELEASE);
1337 
1338 	while (!__atomic_load_n(&args->stop, __ATOMIC_ACQUIRE)) {
1339 		/* list_a is "active". */
1340 		cpu_a = rand() % CPU_SETSIZE;
1341 		/*
1342 		 * As list_b is "inactive", we should never see changes
1343 		 * to list_b.
1344 		 */
1345 		if (expect_b != __atomic_load_n(&list_b.c[cpu_b].head->data, __ATOMIC_ACQUIRE)) {
1346 			fprintf(stderr, "Membarrier test failed\n");
1347 			abort();
1348 		}
1349 
1350 		/* Make list_b "active". */
1351 		__atomic_store_n(&args->percpu_list_ptr, (intptr_t)&list_b, __ATOMIC_RELEASE);
1352 		if (rseq_membarrier_expedited(cpu_a) &&
1353 				errno != ENXIO /* missing CPU */) {
1354 			perror("sys_membarrier");
1355 			abort();
1356 		}
1357 		/*
1358 		 * Cpu A should now only modify list_b, so the values
1359 		 * in list_a should be stable.
1360 		 */
1361 		expect_a = __atomic_load_n(&list_a.c[cpu_a].head->data, __ATOMIC_ACQUIRE);
1362 
1363 		cpu_b = rand() % CPU_SETSIZE;
1364 		/*
1365 		 * As list_a is "inactive", we should never see changes
1366 		 * to list_a.
1367 		 */
1368 		if (expect_a != __atomic_load_n(&list_a.c[cpu_a].head->data, __ATOMIC_ACQUIRE)) {
1369 			fprintf(stderr, "Membarrier test failed\n");
1370 			abort();
1371 		}
1372 
1373 		/* Make list_a "active". */
1374 		__atomic_store_n(&args->percpu_list_ptr, (intptr_t)&list_a, __ATOMIC_RELEASE);
1375 		if (rseq_membarrier_expedited(cpu_b) &&
1376 				errno != ENXIO /* missing CPU*/) {
1377 			perror("sys_membarrier");
1378 			abort();
1379 		}
1380 		/* Remember a value from list_b. */
1381 		expect_b = __atomic_load_n(&list_b.c[cpu_b].head->data, __ATOMIC_ACQUIRE);
1382 	}
1383 
1384 	test_membarrier_free_percpu_list(&list_a);
1385 	test_membarrier_free_percpu_list(&list_b);
1386 
1387 	if (rseq_unregister_current_thread()) {
1388 		fprintf(stderr, "Error: rseq_unregister_current_thread(...) failed(%d): %s\n",
1389 			errno, strerror(errno));
1390 		abort();
1391 	}
1392 	return NULL;
1393 }
1394 
1395 void test_membarrier(void)
1396 {
1397 	const int num_threads = opt_threads;
1398 	struct test_membarrier_thread_args thread_args;
1399 	pthread_t worker_threads[num_threads];
1400 	pthread_t manager_thread;
1401 	int i, ret;
1402 
1403 	if (sys_membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ, 0, 0)) {
1404 		perror("sys_membarrier");
1405 		abort();
1406 	}
1407 
1408 	thread_args.stop = 0;
1409 	thread_args.percpu_list_ptr = 0;
1410 	ret = pthread_create(&manager_thread, NULL,
1411 			test_membarrier_manager_thread, &thread_args);
1412 	if (ret) {
1413 		errno = ret;
1414 		perror("pthread_create");
1415 		abort();
1416 	}
1417 
1418 	for (i = 0; i < num_threads; i++) {
1419 		ret = pthread_create(&worker_threads[i], NULL,
1420 				test_membarrier_worker_thread, &thread_args);
1421 		if (ret) {
1422 			errno = ret;
1423 			perror("pthread_create");
1424 			abort();
1425 		}
1426 	}
1427 
1428 
1429 	for (i = 0; i < num_threads; i++) {
1430 		ret = pthread_join(worker_threads[i], NULL);
1431 		if (ret) {
1432 			errno = ret;
1433 			perror("pthread_join");
1434 			abort();
1435 		}
1436 	}
1437 
1438 	__atomic_store_n(&thread_args.stop, 1, __ATOMIC_RELEASE);
1439 	ret = pthread_join(manager_thread, NULL);
1440 	if (ret) {
1441 		errno = ret;
1442 		perror("pthread_join");
1443 		abort();
1444 	}
1445 }
1446 #else /* TEST_MEMBARRIER */
1447 void test_membarrier(void)
1448 {
1449 	fprintf(stderr, "rseq_offset_deref_addv is not implemented on this architecture. "
1450 			"Skipping membarrier test.\n");
1451 }
1452 #endif
1453 
1454 static void show_usage(int argc, char **argv)
1455 {
1456 	printf("Usage : %s <OPTIONS>\n",
1457 		argv[0]);
1458 	printf("OPTIONS:\n");
1459 	printf("	[-1 loops] Number of loops for delay injection 1\n");
1460 	printf("	[-2 loops] Number of loops for delay injection 2\n");
1461 	printf("	[-3 loops] Number of loops for delay injection 3\n");
1462 	printf("	[-4 loops] Number of loops for delay injection 4\n");
1463 	printf("	[-5 loops] Number of loops for delay injection 5\n");
1464 	printf("	[-6 loops] Number of loops for delay injection 6\n");
1465 	printf("	[-7 loops] Number of loops for delay injection 7 (-1 to enable -m)\n");
1466 	printf("	[-8 loops] Number of loops for delay injection 8 (-1 to enable -m)\n");
1467 	printf("	[-9 loops] Number of loops for delay injection 9 (-1 to enable -m)\n");
1468 	printf("	[-m N] Yield/sleep/kill every modulo N (default 0: disabled) (>= 0)\n");
1469 	printf("	[-y] Yield\n");
1470 	printf("	[-k] Kill thread with signal\n");
1471 	printf("	[-s S] S: =0: disabled (default), >0: sleep time (ms)\n");
1472 	printf("	[-t N] Number of threads (default 200)\n");
1473 	printf("	[-r N] Number of repetitions per thread (default 5000)\n");
1474 	printf("	[-d] Disable rseq system call (no initialization)\n");
1475 	printf("	[-D M] Disable rseq for each M threads\n");
1476 	printf("	[-T test] Choose test: (s)pinlock, (l)ist, (b)uffer, (m)emcpy, (i)ncrement, membarrie(r)\n");
1477 	printf("	[-M] Push into buffer and memcpy buffer with memory barriers.\n");
1478 	printf("	[-v] Verbose output.\n");
1479 	printf("	[-h] Show this help.\n");
1480 	printf("\n");
1481 }
1482 
1483 int main(int argc, char **argv)
1484 {
1485 	int i;
1486 
1487 	for (i = 1; i < argc; i++) {
1488 		if (argv[i][0] != '-')
1489 			continue;
1490 		switch (argv[i][1]) {
1491 		case '1':
1492 		case '2':
1493 		case '3':
1494 		case '4':
1495 		case '5':
1496 		case '6':
1497 		case '7':
1498 		case '8':
1499 		case '9':
1500 			if (argc < i + 2) {
1501 				show_usage(argc, argv);
1502 				goto error;
1503 			}
1504 			loop_cnt[argv[i][1] - '0'] = atol(argv[i + 1]);
1505 			i++;
1506 			break;
1507 		case 'm':
1508 			if (argc < i + 2) {
1509 				show_usage(argc, argv);
1510 				goto error;
1511 			}
1512 			opt_modulo = atol(argv[i + 1]);
1513 			if (opt_modulo < 0) {
1514 				show_usage(argc, argv);
1515 				goto error;
1516 			}
1517 			i++;
1518 			break;
1519 		case 's':
1520 			if (argc < i + 2) {
1521 				show_usage(argc, argv);
1522 				goto error;
1523 			}
1524 			opt_sleep = atol(argv[i + 1]);
1525 			if (opt_sleep < 0) {
1526 				show_usage(argc, argv);
1527 				goto error;
1528 			}
1529 			i++;
1530 			break;
1531 		case 'y':
1532 			opt_yield = 1;
1533 			break;
1534 		case 'k':
1535 			opt_signal = 1;
1536 			break;
1537 		case 'd':
1538 			opt_disable_rseq = 1;
1539 			break;
1540 		case 'D':
1541 			if (argc < i + 2) {
1542 				show_usage(argc, argv);
1543 				goto error;
1544 			}
1545 			opt_disable_mod = atol(argv[i + 1]);
1546 			if (opt_disable_mod < 0) {
1547 				show_usage(argc, argv);
1548 				goto error;
1549 			}
1550 			i++;
1551 			break;
1552 		case 't':
1553 			if (argc < i + 2) {
1554 				show_usage(argc, argv);
1555 				goto error;
1556 			}
1557 			opt_threads = atol(argv[i + 1]);
1558 			if (opt_threads < 0) {
1559 				show_usage(argc, argv);
1560 				goto error;
1561 			}
1562 			i++;
1563 			break;
1564 		case 'r':
1565 			if (argc < i + 2) {
1566 				show_usage(argc, argv);
1567 				goto error;
1568 			}
1569 			opt_reps = atoll(argv[i + 1]);
1570 			if (opt_reps < 0) {
1571 				show_usage(argc, argv);
1572 				goto error;
1573 			}
1574 			i++;
1575 			break;
1576 		case 'h':
1577 			show_usage(argc, argv);
1578 			goto end;
1579 		case 'T':
1580 			if (argc < i + 2) {
1581 				show_usage(argc, argv);
1582 				goto error;
1583 			}
1584 			opt_test = *argv[i + 1];
1585 			switch (opt_test) {
1586 			case 's':
1587 			case 'l':
1588 			case 'i':
1589 			case 'b':
1590 			case 'm':
1591 			case 'r':
1592 				break;
1593 			default:
1594 				show_usage(argc, argv);
1595 				goto error;
1596 			}
1597 			i++;
1598 			break;
1599 		case 'v':
1600 			verbose = 1;
1601 			break;
1602 		case 'M':
1603 			opt_mo = RSEQ_MO_RELEASE;
1604 			break;
1605 		default:
1606 			show_usage(argc, argv);
1607 			goto error;
1608 		}
1609 	}
1610 
1611 	loop_cnt_1 = loop_cnt[1];
1612 	loop_cnt_2 = loop_cnt[2];
1613 	loop_cnt_3 = loop_cnt[3];
1614 	loop_cnt_4 = loop_cnt[4];
1615 	loop_cnt_5 = loop_cnt[5];
1616 	loop_cnt_6 = loop_cnt[6];
1617 
1618 	if (set_signal_handler())
1619 		goto error;
1620 
1621 	if (!opt_disable_rseq && rseq_register_current_thread())
1622 		goto error;
1623 	if (!opt_disable_rseq && !rseq_validate_cpu_id()) {
1624 		fprintf(stderr, "Error: cpu id getter unavailable\n");
1625 		goto error;
1626 	}
1627 	switch (opt_test) {
1628 	case 's':
1629 		printf_verbose("spinlock\n");
1630 		test_percpu_spinlock();
1631 		break;
1632 	case 'l':
1633 		printf_verbose("linked list\n");
1634 		test_percpu_list();
1635 		break;
1636 	case 'b':
1637 		printf_verbose("buffer\n");
1638 		test_percpu_buffer();
1639 		break;
1640 	case 'm':
1641 		printf_verbose("memcpy buffer\n");
1642 		test_percpu_memcpy_buffer();
1643 		break;
1644 	case 'i':
1645 		printf_verbose("counter increment\n");
1646 		test_percpu_inc();
1647 		break;
1648 	case 'r':
1649 		printf_verbose("membarrier\n");
1650 		test_membarrier();
1651 		break;
1652 	}
1653 	if (!opt_disable_rseq && rseq_unregister_current_thread())
1654 		abort();
1655 end:
1656 	return 0;
1657 
1658 error:
1659 	return -1;
1660 }
1661