• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Stress userfaultfd syscall.
4  *
5  *  Copyright (C) 2015  Red Hat, Inc.
6  *
7  * This test allocates two virtual areas and bounces the physical
8  * memory across the two virtual areas (from area_src to area_dst)
9  * using userfaultfd.
10  *
11  * There are three threads running per CPU:
12  *
13  * 1) one per-CPU thread takes a per-page pthread_mutex in a random
14  *    page of the area_dst (while the physical page may still be in
15  *    area_src), and increments a per-page counter in the same page,
16  *    and checks its value against a verification region.
17  *
18  * 2) another per-CPU thread handles the userfaults generated by
19  *    thread 1 above. userfaultfd blocking reads or poll() modes are
20  *    exercised interleaved.
21  *
22  * 3) one last per-CPU thread transfers the memory in the background
23  *    at maximum bandwidth (if not already transferred by thread
24  *    2). Each cpu thread takes cares of transferring a portion of the
25  *    area.
26  *
27  * When all threads of type 3 completed the transfer, one bounce is
28  * complete. area_src and area_dst are then swapped. All threads are
29  * respawned and so the bounce is immediately restarted in the
30  * opposite direction.
31  *
32  * per-CPU threads 1 by triggering userfaults inside
33  * pthread_mutex_lock will also verify the atomicity of the memory
34  * transfer (UFFDIO_COPY).
35  */
36 
37 #define _GNU_SOURCE
38 #include <stdio.h>
39 #include <errno.h>
40 #include <unistd.h>
41 #include <stdlib.h>
42 #include <sys/types.h>
43 #include <sys/stat.h>
44 #include <fcntl.h>
45 #include <time.h>
46 #include <signal.h>
47 #include <poll.h>
48 #include <string.h>
49 #include <linux/mman.h>
50 #include <sys/mman.h>
51 #include <sys/syscall.h>
52 #include <sys/ioctl.h>
53 #include <sys/wait.h>
54 #include <pthread.h>
55 #include <linux/userfaultfd.h>
56 #include <setjmp.h>
57 #include <stdbool.h>
58 #include <assert.h>
59 
60 #include "../kselftest.h"
61 
62 #ifdef __NR_userfaultfd
63 
64 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
65 
66 #define BOUNCE_RANDOM		(1<<0)
67 #define BOUNCE_RACINGFAULTS	(1<<1)
68 #define BOUNCE_VERIFY		(1<<2)
69 #define BOUNCE_POLL		(1<<3)
70 static int bounces;
71 
72 #define TEST_ANON	1
73 #define TEST_HUGETLB	2
74 #define TEST_SHMEM	3
75 static int test_type;
76 
77 /* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
78 #define ALARM_INTERVAL_SECS 10
79 static volatile bool test_uffdio_copy_eexist = true;
80 static volatile bool test_uffdio_zeropage_eexist = true;
81 /* Whether to test uffd write-protection */
82 static bool test_uffdio_wp = false;
83 
84 static bool map_shared;
85 static int huge_fd;
86 static char *huge_fd_off0;
87 static unsigned long long *count_verify;
88 static int uffd, uffd_flags, finished, *pipefd;
89 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
90 static char *zeropage;
91 pthread_attr_t attr;
92 
93 /* Userfaultfd test statistics */
94 struct uffd_stats {
95 	int cpu;
96 	unsigned long missing_faults;
97 	unsigned long wp_faults;
98 };
99 
100 /* pthread_mutex_t starts at page offset 0 */
101 #define area_mutex(___area, ___nr)					\
102 	((pthread_mutex_t *) ((___area) + (___nr)*page_size))
103 /*
104  * count is placed in the page after pthread_mutex_t naturally aligned
105  * to avoid non alignment faults on non-x86 archs.
106  */
107 #define area_count(___area, ___nr)					\
108 	((volatile unsigned long long *) ((unsigned long)		\
109 				 ((___area) + (___nr)*page_size +	\
110 				  sizeof(pthread_mutex_t) +		\
111 				  sizeof(unsigned long long) - 1) &	\
112 				 ~(unsigned long)(sizeof(unsigned long long) \
113 						  -  1)))
114 
115 const char *examples =
116     "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
117     "./userfaultfd anon 100 99999\n\n"
118     "# Run share memory test on 1GiB region with 99 bounces:\n"
119     "./userfaultfd shmem 1000 99\n\n"
120     "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
121     "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
122     "# Run the same hugetlb test but using shmem:\n"
123     "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
124     "# 10MiB-~6GiB 999 bounces anonymous test, "
125     "continue forever unless an error triggers\n"
126     "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
127 
usage(void)128 static void usage(void)
129 {
130 	fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
131 		"[hugetlbfs_file]\n\n");
132 	fprintf(stderr, "Supported <test type>: anon, hugetlb, "
133 		"hugetlb_shared, shmem\n\n");
134 	fprintf(stderr, "Examples:\n\n");
135 	fprintf(stderr, "%s", examples);
136 	exit(1);
137 }
138 
uffd_stats_reset(struct uffd_stats * uffd_stats,unsigned long n_cpus)139 static void uffd_stats_reset(struct uffd_stats *uffd_stats,
140 			     unsigned long n_cpus)
141 {
142 	int i;
143 
144 	for (i = 0; i < n_cpus; i++) {
145 		uffd_stats[i].cpu = i;
146 		uffd_stats[i].missing_faults = 0;
147 		uffd_stats[i].wp_faults = 0;
148 	}
149 }
150 
uffd_stats_report(struct uffd_stats * stats,int n_cpus)151 static void uffd_stats_report(struct uffd_stats *stats, int n_cpus)
152 {
153 	int i;
154 	unsigned long long miss_total = 0, wp_total = 0;
155 
156 	for (i = 0; i < n_cpus; i++) {
157 		miss_total += stats[i].missing_faults;
158 		wp_total += stats[i].wp_faults;
159 	}
160 
161 	printf("userfaults: %llu missing (", miss_total);
162 	for (i = 0; i < n_cpus; i++)
163 		printf("%lu+", stats[i].missing_faults);
164 	printf("\b), %llu wp (", wp_total);
165 	for (i = 0; i < n_cpus; i++)
166 		printf("%lu+", stats[i].wp_faults);
167 	printf("\b)\n");
168 }
169 
anon_release_pages(char * rel_area)170 static int anon_release_pages(char *rel_area)
171 {
172 	int ret = 0;
173 
174 	if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
175 		perror("madvise");
176 		ret = 1;
177 	}
178 
179 	return ret;
180 }
181 
anon_allocate_area(void ** alloc_area)182 static void anon_allocate_area(void **alloc_area)
183 {
184 	*alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
185 			   MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
186 	if (*alloc_area == MAP_FAILED) {
187 		fprintf(stderr, "mmap of anonymous memory failed");
188 		*alloc_area = NULL;
189 	}
190 }
191 
noop_alias_mapping(__u64 * start,size_t len,unsigned long offset)192 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
193 {
194 }
195 
196 /* HugeTLB memory */
hugetlb_release_pages(char * rel_area)197 static int hugetlb_release_pages(char *rel_area)
198 {
199 	int ret = 0;
200 
201 	if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
202 				rel_area == huge_fd_off0 ? 0 :
203 				nr_pages * page_size,
204 				nr_pages * page_size)) {
205 		perror("fallocate");
206 		ret = 1;
207 	}
208 
209 	return ret;
210 }
211 
hugetlb_allocate_area(void ** alloc_area)212 static void hugetlb_allocate_area(void **alloc_area)
213 {
214 	void *area_alias = NULL;
215 	char **alloc_area_alias;
216 
217 	*alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
218 			   (map_shared ? MAP_SHARED : MAP_PRIVATE) |
219 			   MAP_HUGETLB,
220 			   huge_fd, *alloc_area == area_src ? 0 :
221 			   nr_pages * page_size);
222 	if (*alloc_area == MAP_FAILED) {
223 		perror("mmap of hugetlbfs file failed");
224 		goto fail;
225 	}
226 
227 	if (map_shared) {
228 		area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
229 				  MAP_SHARED | MAP_HUGETLB,
230 				  huge_fd, *alloc_area == area_src ? 0 :
231 				  nr_pages * page_size);
232 		if (area_alias == MAP_FAILED) {
233 			perror("mmap of hugetlb file alias failed");
234 			goto fail_munmap;
235 		}
236 	}
237 
238 	if (*alloc_area == area_src) {
239 		huge_fd_off0 = *alloc_area;
240 		alloc_area_alias = &area_src_alias;
241 	} else {
242 		alloc_area_alias = &area_dst_alias;
243 	}
244 	if (area_alias)
245 		*alloc_area_alias = area_alias;
246 
247 	return;
248 
249 fail_munmap:
250 	if (munmap(*alloc_area, nr_pages * page_size) < 0) {
251 		perror("hugetlb munmap");
252 		exit(1);
253 	}
254 fail:
255 	*alloc_area = NULL;
256 }
257 
hugetlb_alias_mapping(__u64 * start,size_t len,unsigned long offset)258 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
259 {
260 	if (!map_shared)
261 		return;
262 	/*
263 	 * We can't zap just the pagetable with hugetlbfs because
264 	 * MADV_DONTEED won't work. So exercise -EEXIST on a alias
265 	 * mapping where the pagetables are not established initially,
266 	 * this way we'll exercise the -EEXEC at the fs level.
267 	 */
268 	*start = (unsigned long) area_dst_alias + offset;
269 }
270 
271 /* Shared memory */
shmem_release_pages(char * rel_area)272 static int shmem_release_pages(char *rel_area)
273 {
274 	int ret = 0;
275 
276 	if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
277 		perror("madvise");
278 		ret = 1;
279 	}
280 
281 	return ret;
282 }
283 
shmem_allocate_area(void ** alloc_area)284 static void shmem_allocate_area(void **alloc_area)
285 {
286 	*alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
287 			   MAP_ANONYMOUS | MAP_SHARED, -1, 0);
288 	if (*alloc_area == MAP_FAILED) {
289 		fprintf(stderr, "shared memory mmap failed\n");
290 		*alloc_area = NULL;
291 	}
292 }
293 
294 struct uffd_test_ops {
295 	unsigned long expected_ioctls;
296 	void (*allocate_area)(void **alloc_area);
297 	int (*release_pages)(char *rel_area);
298 	void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
299 };
300 
301 #define SHMEM_EXPECTED_IOCTLS		((1 << _UFFDIO_WAKE) | \
302 					 (1 << _UFFDIO_COPY) | \
303 					 (1 << _UFFDIO_ZEROPAGE))
304 
305 #define ANON_EXPECTED_IOCTLS		((1 << _UFFDIO_WAKE) | \
306 					 (1 << _UFFDIO_COPY) | \
307 					 (1 << _UFFDIO_ZEROPAGE) | \
308 					 (1 << _UFFDIO_WRITEPROTECT))
309 
310 static struct uffd_test_ops anon_uffd_test_ops = {
311 	.expected_ioctls = ANON_EXPECTED_IOCTLS,
312 	.allocate_area	= anon_allocate_area,
313 	.release_pages	= anon_release_pages,
314 	.alias_mapping = noop_alias_mapping,
315 };
316 
317 static struct uffd_test_ops shmem_uffd_test_ops = {
318 	.expected_ioctls = SHMEM_EXPECTED_IOCTLS,
319 	.allocate_area	= shmem_allocate_area,
320 	.release_pages	= shmem_release_pages,
321 	.alias_mapping = noop_alias_mapping,
322 };
323 
324 static struct uffd_test_ops hugetlb_uffd_test_ops = {
325 	.expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
326 	.allocate_area	= hugetlb_allocate_area,
327 	.release_pages	= hugetlb_release_pages,
328 	.alias_mapping = hugetlb_alias_mapping,
329 };
330 
331 static struct uffd_test_ops *uffd_test_ops;
332 
my_bcmp(char * str1,char * str2,size_t n)333 static int my_bcmp(char *str1, char *str2, size_t n)
334 {
335 	unsigned long i;
336 	for (i = 0; i < n; i++)
337 		if (str1[i] != str2[i])
338 			return 1;
339 	return 0;
340 }
341 
wp_range(int ufd,__u64 start,__u64 len,bool wp)342 static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
343 {
344 	struct uffdio_writeprotect prms = { 0 };
345 
346 	/* Write protection page faults */
347 	prms.range.start = start;
348 	prms.range.len = len;
349 	/* Undo write-protect, do wakeup after that */
350 	prms.mode = wp ? UFFDIO_WRITEPROTECT_MODE_WP : 0;
351 
352 	if (ioctl(ufd, UFFDIO_WRITEPROTECT, &prms)) {
353 		fprintf(stderr, "clear WP failed for address 0x%Lx\n", start);
354 		exit(1);
355 	}
356 }
357 
locking_thread(void * arg)358 static void *locking_thread(void *arg)
359 {
360 	unsigned long cpu = (unsigned long) arg;
361 	struct random_data rand;
362 	unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
363 	int32_t rand_nr;
364 	unsigned long long count;
365 	char randstate[64];
366 	unsigned int seed;
367 	time_t start;
368 
369 	if (bounces & BOUNCE_RANDOM) {
370 		seed = (unsigned int) time(NULL) - bounces;
371 		if (!(bounces & BOUNCE_RACINGFAULTS))
372 			seed += cpu;
373 		bzero(&rand, sizeof(rand));
374 		bzero(&randstate, sizeof(randstate));
375 		if (initstate_r(seed, randstate, sizeof(randstate), &rand)) {
376 			fprintf(stderr, "srandom_r error\n");
377 			exit(1);
378 		}
379 	} else {
380 		page_nr = -bounces;
381 		if (!(bounces & BOUNCE_RACINGFAULTS))
382 			page_nr += cpu * nr_pages_per_cpu;
383 	}
384 
385 	while (!finished) {
386 		if (bounces & BOUNCE_RANDOM) {
387 			if (random_r(&rand, &rand_nr)) {
388 				fprintf(stderr, "random_r 1 error\n");
389 				exit(1);
390 			}
391 			page_nr = rand_nr;
392 			if (sizeof(page_nr) > sizeof(rand_nr)) {
393 				if (random_r(&rand, &rand_nr)) {
394 					fprintf(stderr, "random_r 2 error\n");
395 					exit(1);
396 				}
397 				page_nr |= (((unsigned long) rand_nr) << 16) <<
398 					   16;
399 			}
400 		} else
401 			page_nr += 1;
402 		page_nr %= nr_pages;
403 
404 		start = time(NULL);
405 		if (bounces & BOUNCE_VERIFY) {
406 			count = *area_count(area_dst, page_nr);
407 			if (!count) {
408 				fprintf(stderr,
409 					"page_nr %lu wrong count %Lu %Lu\n",
410 					page_nr, count,
411 					count_verify[page_nr]);
412 				exit(1);
413 			}
414 
415 
416 			/*
417 			 * We can't use bcmp (or memcmp) because that
418 			 * returns 0 erroneously if the memory is
419 			 * changing under it (even if the end of the
420 			 * page is never changing and always
421 			 * different).
422 			 */
423 #if 1
424 			if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
425 				     page_size)) {
426 				fprintf(stderr,
427 					"my_bcmp page_nr %lu wrong count %Lu %Lu\n",
428 					page_nr, count, count_verify[page_nr]);
429 				exit(1);
430 			}
431 #else
432 			unsigned long loops;
433 
434 			loops = 0;
435 			/* uncomment the below line to test with mutex */
436 			/* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
437 			while (!bcmp(area_dst + page_nr * page_size, zeropage,
438 				     page_size)) {
439 				loops += 1;
440 				if (loops > 10)
441 					break;
442 			}
443 			/* uncomment below line to test with mutex */
444 			/* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
445 			if (loops) {
446 				fprintf(stderr,
447 					"page_nr %lu all zero thread %lu %p %lu\n",
448 					page_nr, cpu, area_dst + page_nr * page_size,
449 					loops);
450 				if (loops > 10)
451 					exit(1);
452 			}
453 #endif
454 		}
455 
456 		pthread_mutex_lock(area_mutex(area_dst, page_nr));
457 		count = *area_count(area_dst, page_nr);
458 		if (count != count_verify[page_nr]) {
459 			fprintf(stderr,
460 				"page_nr %lu memory corruption %Lu %Lu\n",
461 				page_nr, count,
462 				count_verify[page_nr]); exit(1);
463 		}
464 		count++;
465 		*area_count(area_dst, page_nr) = count_verify[page_nr] = count;
466 		pthread_mutex_unlock(area_mutex(area_dst, page_nr));
467 
468 		if (time(NULL) - start > 1)
469 			fprintf(stderr,
470 				"userfault too slow %ld "
471 				"possible false positive with overcommit\n",
472 				time(NULL) - start);
473 	}
474 
475 	return NULL;
476 }
477 
retry_copy_page(int ufd,struct uffdio_copy * uffdio_copy,unsigned long offset)478 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
479 			    unsigned long offset)
480 {
481 	uffd_test_ops->alias_mapping(&uffdio_copy->dst,
482 				     uffdio_copy->len,
483 				     offset);
484 	if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
485 		/* real retval in ufdio_copy.copy */
486 		if (uffdio_copy->copy != -EEXIST) {
487 			fprintf(stderr, "UFFDIO_COPY retry error %Ld\n",
488 				uffdio_copy->copy);
489 			exit(1);
490 		}
491 	} else {
492 		fprintf(stderr,	"UFFDIO_COPY retry unexpected %Ld\n",
493 			uffdio_copy->copy); exit(1);
494 	}
495 }
496 
__copy_page(int ufd,unsigned long offset,bool retry)497 static int __copy_page(int ufd, unsigned long offset, bool retry)
498 {
499 	struct uffdio_copy uffdio_copy;
500 
501 	if (offset >= nr_pages * page_size) {
502 		fprintf(stderr, "unexpected offset %lu\n", offset);
503 		exit(1);
504 	}
505 	uffdio_copy.dst = (unsigned long) area_dst + offset;
506 	uffdio_copy.src = (unsigned long) area_src + offset;
507 	uffdio_copy.len = page_size;
508 	if (test_uffdio_wp)
509 		uffdio_copy.mode = UFFDIO_COPY_MODE_WP;
510 	else
511 		uffdio_copy.mode = 0;
512 	uffdio_copy.copy = 0;
513 	if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
514 		/* real retval in ufdio_copy.copy */
515 		if (uffdio_copy.copy != -EEXIST) {
516 			fprintf(stderr, "UFFDIO_COPY error %Ld\n",
517 				uffdio_copy.copy);
518 			exit(1);
519 		}
520 	} else if (uffdio_copy.copy != page_size) {
521 		fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
522 			uffdio_copy.copy); exit(1);
523 	} else {
524 		if (test_uffdio_copy_eexist && retry) {
525 			test_uffdio_copy_eexist = false;
526 			retry_copy_page(ufd, &uffdio_copy, offset);
527 		}
528 		return 1;
529 	}
530 	return 0;
531 }
532 
copy_page_retry(int ufd,unsigned long offset)533 static int copy_page_retry(int ufd, unsigned long offset)
534 {
535 	return __copy_page(ufd, offset, true);
536 }
537 
copy_page(int ufd,unsigned long offset)538 static int copy_page(int ufd, unsigned long offset)
539 {
540 	return __copy_page(ufd, offset, false);
541 }
542 
uffd_read_msg(int ufd,struct uffd_msg * msg)543 static int uffd_read_msg(int ufd, struct uffd_msg *msg)
544 {
545 	int ret = read(uffd, msg, sizeof(*msg));
546 
547 	if (ret != sizeof(*msg)) {
548 		if (ret < 0) {
549 			if (errno == EAGAIN)
550 				return 1;
551 			perror("blocking read error");
552 		} else {
553 			fprintf(stderr, "short read\n");
554 		}
555 		exit(1);
556 	}
557 
558 	return 0;
559 }
560 
uffd_handle_page_fault(struct uffd_msg * msg,struct uffd_stats * stats)561 static void uffd_handle_page_fault(struct uffd_msg *msg,
562 				   struct uffd_stats *stats)
563 {
564 	unsigned long offset;
565 
566 	if (msg->event != UFFD_EVENT_PAGEFAULT) {
567 		fprintf(stderr, "unexpected msg event %u\n", msg->event);
568 		exit(1);
569 	}
570 
571 	if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP) {
572 		wp_range(uffd, msg->arg.pagefault.address, page_size, false);
573 		stats->wp_faults++;
574 	} else {
575 		/* Missing page faults */
576 		if (bounces & BOUNCE_VERIFY &&
577 		    msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE) {
578 			fprintf(stderr, "unexpected write fault\n");
579 			exit(1);
580 		}
581 
582 		offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
583 		offset &= ~(page_size-1);
584 
585 		if (copy_page(uffd, offset))
586 			stats->missing_faults++;
587 	}
588 }
589 
uffd_poll_thread(void * arg)590 static void *uffd_poll_thread(void *arg)
591 {
592 	struct uffd_stats *stats = (struct uffd_stats *)arg;
593 	unsigned long cpu = stats->cpu;
594 	struct pollfd pollfd[2];
595 	struct uffd_msg msg;
596 	struct uffdio_register uffd_reg;
597 	int ret;
598 	char tmp_chr;
599 
600 	pollfd[0].fd = uffd;
601 	pollfd[0].events = POLLIN;
602 	pollfd[1].fd = pipefd[cpu*2];
603 	pollfd[1].events = POLLIN;
604 
605 	for (;;) {
606 		ret = poll(pollfd, 2, -1);
607 		if (!ret) {
608 			fprintf(stderr, "poll error %d\n", ret);
609 			exit(1);
610 		}
611 		if (ret < 0) {
612 			perror("poll");
613 			exit(1);
614 		}
615 		if (pollfd[1].revents & POLLIN) {
616 			if (read(pollfd[1].fd, &tmp_chr, 1) != 1) {
617 				fprintf(stderr, "read pipefd error\n");
618 				exit(1);
619 			}
620 			break;
621 		}
622 		if (!(pollfd[0].revents & POLLIN)) {
623 			fprintf(stderr, "pollfd[0].revents %d\n",
624 				pollfd[0].revents);
625 			exit(1);
626 		}
627 		if (uffd_read_msg(uffd, &msg))
628 			continue;
629 		switch (msg.event) {
630 		default:
631 			fprintf(stderr, "unexpected msg event %u\n",
632 				msg.event); exit(1);
633 			break;
634 		case UFFD_EVENT_PAGEFAULT:
635 			uffd_handle_page_fault(&msg, stats);
636 			break;
637 		case UFFD_EVENT_FORK:
638 			close(uffd);
639 			uffd = msg.arg.fork.ufd;
640 			pollfd[0].fd = uffd;
641 			break;
642 		case UFFD_EVENT_REMOVE:
643 			uffd_reg.range.start = msg.arg.remove.start;
644 			uffd_reg.range.len = msg.arg.remove.end -
645 				msg.arg.remove.start;
646 			if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range)) {
647 				fprintf(stderr, "remove failure\n");
648 				exit(1);
649 			}
650 			break;
651 		case UFFD_EVENT_REMAP:
652 			area_dst = (char *)(unsigned long)msg.arg.remap.to;
653 			break;
654 		}
655 	}
656 
657 	return NULL;
658 }
659 
660 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
661 
uffd_read_thread(void * arg)662 static void *uffd_read_thread(void *arg)
663 {
664 	struct uffd_stats *stats = (struct uffd_stats *)arg;
665 	struct uffd_msg msg;
666 
667 	pthread_mutex_unlock(&uffd_read_mutex);
668 	/* from here cancellation is ok */
669 
670 	for (;;) {
671 		if (uffd_read_msg(uffd, &msg))
672 			continue;
673 		uffd_handle_page_fault(&msg, stats);
674 	}
675 
676 	return NULL;
677 }
678 
background_thread(void * arg)679 static void *background_thread(void *arg)
680 {
681 	unsigned long cpu = (unsigned long) arg;
682 	unsigned long page_nr, start_nr, mid_nr, end_nr;
683 
684 	start_nr = cpu * nr_pages_per_cpu;
685 	end_nr = (cpu+1) * nr_pages_per_cpu;
686 	mid_nr = (start_nr + end_nr) / 2;
687 
688 	/* Copy the first half of the pages */
689 	for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
690 		copy_page_retry(uffd, page_nr * page_size);
691 
692 	/*
693 	 * If we need to test uffd-wp, set it up now.  Then we'll have
694 	 * at least the first half of the pages mapped already which
695 	 * can be write-protected for testing
696 	 */
697 	if (test_uffdio_wp)
698 		wp_range(uffd, (unsigned long)area_dst + start_nr * page_size,
699 			nr_pages_per_cpu * page_size, true);
700 
701 	/*
702 	 * Continue the 2nd half of the page copying, handling write
703 	 * protection faults if any
704 	 */
705 	for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
706 		copy_page_retry(uffd, page_nr * page_size);
707 
708 	return NULL;
709 }
710 
stress(struct uffd_stats * uffd_stats)711 static int stress(struct uffd_stats *uffd_stats)
712 {
713 	unsigned long cpu;
714 	pthread_t locking_threads[nr_cpus];
715 	pthread_t uffd_threads[nr_cpus];
716 	pthread_t background_threads[nr_cpus];
717 
718 	finished = 0;
719 	for (cpu = 0; cpu < nr_cpus; cpu++) {
720 		if (pthread_create(&locking_threads[cpu], &attr,
721 				   locking_thread, (void *)cpu))
722 			return 1;
723 		if (bounces & BOUNCE_POLL) {
724 			if (pthread_create(&uffd_threads[cpu], &attr,
725 					   uffd_poll_thread,
726 					   (void *)&uffd_stats[cpu]))
727 				return 1;
728 		} else {
729 			if (pthread_create(&uffd_threads[cpu], &attr,
730 					   uffd_read_thread,
731 					   (void *)&uffd_stats[cpu]))
732 				return 1;
733 			pthread_mutex_lock(&uffd_read_mutex);
734 		}
735 		if (pthread_create(&background_threads[cpu], &attr,
736 				   background_thread, (void *)cpu))
737 			return 1;
738 	}
739 	for (cpu = 0; cpu < nr_cpus; cpu++)
740 		if (pthread_join(background_threads[cpu], NULL))
741 			return 1;
742 
743 	/*
744 	 * Be strict and immediately zap area_src, the whole area has
745 	 * been transferred already by the background treads. The
746 	 * area_src could then be faulted in in a racy way by still
747 	 * running uffdio_threads reading zeropages after we zapped
748 	 * area_src (but they're guaranteed to get -EEXIST from
749 	 * UFFDIO_COPY without writing zero pages into area_dst
750 	 * because the background threads already completed).
751 	 */
752 	if (uffd_test_ops->release_pages(area_src))
753 		return 1;
754 
755 
756 	finished = 1;
757 	for (cpu = 0; cpu < nr_cpus; cpu++)
758 		if (pthread_join(locking_threads[cpu], NULL))
759 			return 1;
760 
761 	for (cpu = 0; cpu < nr_cpus; cpu++) {
762 		char c;
763 		if (bounces & BOUNCE_POLL) {
764 			if (write(pipefd[cpu*2+1], &c, 1) != 1) {
765 				fprintf(stderr, "pipefd write error\n");
766 				return 1;
767 			}
768 			if (pthread_join(uffd_threads[cpu],
769 					 (void *)&uffd_stats[cpu]))
770 				return 1;
771 		} else {
772 			if (pthread_cancel(uffd_threads[cpu]))
773 				return 1;
774 			if (pthread_join(uffd_threads[cpu], NULL))
775 				return 1;
776 		}
777 	}
778 
779 	return 0;
780 }
781 
userfaultfd_open(int features)782 static int userfaultfd_open(int features)
783 {
784 	struct uffdio_api uffdio_api;
785 
786 	uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
787 	if (uffd < 0) {
788 		fprintf(stderr,
789 			"userfaultfd syscall not available in this kernel\n");
790 		return 1;
791 	}
792 	uffd_flags = fcntl(uffd, F_GETFD, NULL);
793 
794 	uffdio_api.api = UFFD_API;
795 	uffdio_api.features = features;
796 	if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
797 		fprintf(stderr, "UFFDIO_API\n");
798 		return 1;
799 	}
800 	if (uffdio_api.api != UFFD_API) {
801 		fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
802 		return 1;
803 	}
804 
805 	return 0;
806 }
807 
808 sigjmp_buf jbuf, *sigbuf;
809 
sighndl(int sig,siginfo_t * siginfo,void * ptr)810 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
811 {
812 	if (sig == SIGBUS) {
813 		if (sigbuf)
814 			siglongjmp(*sigbuf, 1);
815 		abort();
816 	}
817 }
818 
819 /*
820  * For non-cooperative userfaultfd test we fork() a process that will
821  * generate pagefaults, will mremap the area monitored by the
822  * userfaultfd and at last this process will release the monitored
823  * area.
824  * For the anonymous and shared memory the area is divided into two
825  * parts, the first part is accessed before mremap, and the second
826  * part is accessed after mremap. Since hugetlbfs does not support
827  * mremap, the entire monitored area is accessed in a single pass for
828  * HUGETLB_TEST.
829  * The release of the pages currently generates event for shmem and
830  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
831  * for hugetlb.
832  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
833  * monitored area, generate pagefaults and test that signal is delivered.
834  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
835  * test robustness use case - we release monitored area, fork a process
836  * that will generate pagefaults and verify signal is generated.
837  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
838  * feature. Using monitor thread, verify no userfault events are generated.
839  */
faulting_process(int signal_test)840 static int faulting_process(int signal_test)
841 {
842 	unsigned long nr;
843 	unsigned long long count;
844 	unsigned long split_nr_pages;
845 	unsigned long lastnr;
846 	struct sigaction act;
847 	unsigned long signalled = 0;
848 
849 	if (test_type != TEST_HUGETLB)
850 		split_nr_pages = (nr_pages + 1) / 2;
851 	else
852 		split_nr_pages = nr_pages;
853 
854 	if (signal_test) {
855 		sigbuf = &jbuf;
856 		memset(&act, 0, sizeof(act));
857 		act.sa_sigaction = sighndl;
858 		act.sa_flags = SA_SIGINFO;
859 		if (sigaction(SIGBUS, &act, 0)) {
860 			perror("sigaction");
861 			return 1;
862 		}
863 		lastnr = (unsigned long)-1;
864 	}
865 
866 	for (nr = 0; nr < split_nr_pages; nr++) {
867 		int steps = 1;
868 		unsigned long offset = nr * page_size;
869 
870 		if (signal_test) {
871 			if (sigsetjmp(*sigbuf, 1) != 0) {
872 				if (steps == 1 && nr == lastnr) {
873 					fprintf(stderr, "Signal repeated\n");
874 					return 1;
875 				}
876 
877 				lastnr = nr;
878 				if (signal_test == 1) {
879 					if (steps == 1) {
880 						/* This is a MISSING request */
881 						steps++;
882 						if (copy_page(uffd, offset))
883 							signalled++;
884 					} else {
885 						/* This is a WP request */
886 						assert(steps == 2);
887 						wp_range(uffd,
888 							 (__u64)area_dst +
889 							 offset,
890 							 page_size, false);
891 					}
892 				} else {
893 					signalled++;
894 					continue;
895 				}
896 			}
897 		}
898 
899 		count = *area_count(area_dst, nr);
900 		if (count != count_verify[nr]) {
901 			fprintf(stderr,
902 				"nr %lu memory corruption %Lu %Lu\n",
903 				nr, count,
904 				count_verify[nr]);
905 	        }
906 		/*
907 		 * Trigger write protection if there is by writting
908 		 * the same value back.
909 		 */
910 		*area_count(area_dst, nr) = count;
911 	}
912 
913 	if (signal_test)
914 		return signalled != split_nr_pages;
915 
916 	if (test_type == TEST_HUGETLB)
917 		return 0;
918 
919 	area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
920 			  MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
921 	if (area_dst == MAP_FAILED) {
922 		perror("mremap");
923 		exit(1);
924 	}
925 
926 	for (; nr < nr_pages; nr++) {
927 		count = *area_count(area_dst, nr);
928 		if (count != count_verify[nr]) {
929 			fprintf(stderr,
930 				"nr %lu memory corruption %Lu %Lu\n",
931 				nr, count,
932 				count_verify[nr]); exit(1);
933 		}
934 		/*
935 		 * Trigger write protection if there is by writting
936 		 * the same value back.
937 		 */
938 		*area_count(area_dst, nr) = count;
939 	}
940 
941 	if (uffd_test_ops->release_pages(area_dst))
942 		return 1;
943 
944 	for (nr = 0; nr < nr_pages; nr++) {
945 		if (my_bcmp(area_dst + nr * page_size, zeropage, page_size)) {
946 			fprintf(stderr, "nr %lu is not zero\n", nr);
947 			exit(1);
948 		}
949 	}
950 
951 	return 0;
952 }
953 
retry_uffdio_zeropage(int ufd,struct uffdio_zeropage * uffdio_zeropage,unsigned long offset)954 static void retry_uffdio_zeropage(int ufd,
955 				  struct uffdio_zeropage *uffdio_zeropage,
956 				  unsigned long offset)
957 {
958 	uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
959 				     uffdio_zeropage->range.len,
960 				     offset);
961 	if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
962 		if (uffdio_zeropage->zeropage != -EEXIST) {
963 			fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
964 				uffdio_zeropage->zeropage);
965 			exit(1);
966 		}
967 	} else {
968 		fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
969 			uffdio_zeropage->zeropage); exit(1);
970 	}
971 }
972 
__uffdio_zeropage(int ufd,unsigned long offset,bool retry)973 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
974 {
975 	struct uffdio_zeropage uffdio_zeropage;
976 	int ret;
977 	unsigned long has_zeropage;
978 
979 	has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
980 
981 	if (offset >= nr_pages * page_size) {
982 		fprintf(stderr, "unexpected offset %lu\n", offset);
983 		exit(1);
984 	}
985 	uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
986 	uffdio_zeropage.range.len = page_size;
987 	uffdio_zeropage.mode = 0;
988 	ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
989 	if (ret) {
990 		/* real retval in ufdio_zeropage.zeropage */
991 		if (has_zeropage) {
992 			if (uffdio_zeropage.zeropage == -EEXIST) {
993 				fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n");
994 				exit(1);
995 			} else {
996 				fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
997 					uffdio_zeropage.zeropage);
998 				exit(1);
999 			}
1000 		} else {
1001 			if (uffdio_zeropage.zeropage != -EINVAL) {
1002 				fprintf(stderr,
1003 					"UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
1004 					uffdio_zeropage.zeropage);
1005 				exit(1);
1006 			}
1007 		}
1008 	} else if (has_zeropage) {
1009 		if (uffdio_zeropage.zeropage != page_size) {
1010 			fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
1011 				uffdio_zeropage.zeropage); exit(1);
1012 		} else {
1013 			if (test_uffdio_zeropage_eexist && retry) {
1014 				test_uffdio_zeropage_eexist = false;
1015 				retry_uffdio_zeropage(ufd, &uffdio_zeropage,
1016 						      offset);
1017 			}
1018 			return 1;
1019 		}
1020 	} else {
1021 		fprintf(stderr,
1022 			"UFFDIO_ZEROPAGE succeeded %Ld\n",
1023 			uffdio_zeropage.zeropage); exit(1);
1024 	}
1025 
1026 	return 0;
1027 }
1028 
uffdio_zeropage(int ufd,unsigned long offset)1029 static int uffdio_zeropage(int ufd, unsigned long offset)
1030 {
1031 	return __uffdio_zeropage(ufd, offset, false);
1032 }
1033 
1034 /* exercise UFFDIO_ZEROPAGE */
userfaultfd_zeropage_test(void)1035 static int userfaultfd_zeropage_test(void)
1036 {
1037 	struct uffdio_register uffdio_register;
1038 	unsigned long expected_ioctls;
1039 
1040 	printf("testing UFFDIO_ZEROPAGE: ");
1041 	fflush(stdout);
1042 
1043 	if (uffd_test_ops->release_pages(area_dst))
1044 		return 1;
1045 
1046 	if (userfaultfd_open(0) < 0)
1047 		return 1;
1048 	uffdio_register.range.start = (unsigned long) area_dst;
1049 	uffdio_register.range.len = nr_pages * page_size;
1050 	uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1051 	if (test_uffdio_wp)
1052 		uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1053 	if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1054 		fprintf(stderr, "register failure\n");
1055 		exit(1);
1056 	}
1057 
1058 	expected_ioctls = uffd_test_ops->expected_ioctls;
1059 	if ((uffdio_register.ioctls & expected_ioctls) !=
1060 	    expected_ioctls) {
1061 		fprintf(stderr,
1062 			"unexpected missing ioctl for anon memory\n");
1063 		exit(1);
1064 	}
1065 
1066 	if (uffdio_zeropage(uffd, 0)) {
1067 		if (my_bcmp(area_dst, zeropage, page_size)) {
1068 			fprintf(stderr, "zeropage is not zero\n");
1069 			exit(1);
1070 		}
1071 	}
1072 
1073 	close(uffd);
1074 	printf("done.\n");
1075 	return 0;
1076 }
1077 
userfaultfd_events_test(void)1078 static int userfaultfd_events_test(void)
1079 {
1080 	struct uffdio_register uffdio_register;
1081 	unsigned long expected_ioctls;
1082 	pthread_t uffd_mon;
1083 	int err, features;
1084 	pid_t pid;
1085 	char c;
1086 	struct uffd_stats stats = { 0 };
1087 
1088 	printf("testing events (fork, remap, remove): ");
1089 	fflush(stdout);
1090 
1091 	if (uffd_test_ops->release_pages(area_dst))
1092 		return 1;
1093 
1094 	features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
1095 		UFFD_FEATURE_EVENT_REMOVE;
1096 	if (userfaultfd_open(features) < 0)
1097 		return 1;
1098 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1099 
1100 	uffdio_register.range.start = (unsigned long) area_dst;
1101 	uffdio_register.range.len = nr_pages * page_size;
1102 	uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1103 	if (test_uffdio_wp)
1104 		uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1105 	if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1106 		fprintf(stderr, "register failure\n");
1107 		exit(1);
1108 	}
1109 
1110 	expected_ioctls = uffd_test_ops->expected_ioctls;
1111 	if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1112 		fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1113 		exit(1);
1114 	}
1115 
1116 	if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1117 		perror("uffd_poll_thread create");
1118 		exit(1);
1119 	}
1120 
1121 	pid = fork();
1122 	if (pid < 0) {
1123 		perror("fork");
1124 		exit(1);
1125 	}
1126 
1127 	if (!pid)
1128 		return faulting_process(0);
1129 
1130 	waitpid(pid, &err, 0);
1131 	if (err) {
1132 		fprintf(stderr, "faulting process failed\n");
1133 		exit(1);
1134 	}
1135 
1136 	if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1137 		perror("pipe write");
1138 		exit(1);
1139 	}
1140 	if (pthread_join(uffd_mon, NULL))
1141 		return 1;
1142 
1143 	close(uffd);
1144 
1145 	uffd_stats_report(&stats, 1);
1146 
1147 	return stats.missing_faults != nr_pages;
1148 }
1149 
userfaultfd_sig_test(void)1150 static int userfaultfd_sig_test(void)
1151 {
1152 	struct uffdio_register uffdio_register;
1153 	unsigned long expected_ioctls;
1154 	unsigned long userfaults;
1155 	pthread_t uffd_mon;
1156 	int err, features;
1157 	pid_t pid;
1158 	char c;
1159 	struct uffd_stats stats = { 0 };
1160 
1161 	printf("testing signal delivery: ");
1162 	fflush(stdout);
1163 
1164 	if (uffd_test_ops->release_pages(area_dst))
1165 		return 1;
1166 
1167 	features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
1168 	if (userfaultfd_open(features) < 0)
1169 		return 1;
1170 	fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1171 
1172 	uffdio_register.range.start = (unsigned long) area_dst;
1173 	uffdio_register.range.len = nr_pages * page_size;
1174 	uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1175 	if (test_uffdio_wp)
1176 		uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1177 	if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1178 		fprintf(stderr, "register failure\n");
1179 		exit(1);
1180 	}
1181 
1182 	expected_ioctls = uffd_test_ops->expected_ioctls;
1183 	if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1184 		fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1185 		exit(1);
1186 	}
1187 
1188 	if (faulting_process(1)) {
1189 		fprintf(stderr, "faulting process failed\n");
1190 		exit(1);
1191 	}
1192 
1193 	if (uffd_test_ops->release_pages(area_dst))
1194 		return 1;
1195 
1196 	if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1197 		perror("uffd_poll_thread create");
1198 		exit(1);
1199 	}
1200 
1201 	pid = fork();
1202 	if (pid < 0) {
1203 		perror("fork");
1204 		exit(1);
1205 	}
1206 
1207 	if (!pid)
1208 		exit(faulting_process(2));
1209 
1210 	waitpid(pid, &err, 0);
1211 	if (err) {
1212 		fprintf(stderr, "faulting process failed\n");
1213 		exit(1);
1214 	}
1215 
1216 	if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1217 		perror("pipe write");
1218 		exit(1);
1219 	}
1220 	if (pthread_join(uffd_mon, (void **)&userfaults))
1221 		return 1;
1222 
1223 	printf("done.\n");
1224 	if (userfaults)
1225 		fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1226 			userfaults);
1227 	close(uffd);
1228 	return userfaults != 0;
1229 }
1230 
userfaultfd_stress(void)1231 static int userfaultfd_stress(void)
1232 {
1233 	void *area;
1234 	char *tmp_area;
1235 	unsigned long nr;
1236 	struct uffdio_register uffdio_register;
1237 	unsigned long cpu;
1238 	int err;
1239 	struct uffd_stats uffd_stats[nr_cpus];
1240 
1241 	uffd_test_ops->allocate_area((void **)&area_src);
1242 	if (!area_src)
1243 		return 1;
1244 	uffd_test_ops->allocate_area((void **)&area_dst);
1245 	if (!area_dst)
1246 		return 1;
1247 
1248 	if (userfaultfd_open(0) < 0)
1249 		return 1;
1250 
1251 	count_verify = malloc(nr_pages * sizeof(unsigned long long));
1252 	if (!count_verify) {
1253 		perror("count_verify");
1254 		return 1;
1255 	}
1256 
1257 	for (nr = 0; nr < nr_pages; nr++) {
1258 		*area_mutex(area_src, nr) = (pthread_mutex_t)
1259 			PTHREAD_MUTEX_INITIALIZER;
1260 		count_verify[nr] = *area_count(area_src, nr) = 1;
1261 		/*
1262 		 * In the transition between 255 to 256, powerpc will
1263 		 * read out of order in my_bcmp and see both bytes as
1264 		 * zero, so leave a placeholder below always non-zero
1265 		 * after the count, to avoid my_bcmp to trigger false
1266 		 * positives.
1267 		 */
1268 		*(area_count(area_src, nr) + 1) = 1;
1269 	}
1270 
1271 	pipefd = malloc(sizeof(int) * nr_cpus * 2);
1272 	if (!pipefd) {
1273 		perror("pipefd");
1274 		return 1;
1275 	}
1276 	for (cpu = 0; cpu < nr_cpus; cpu++) {
1277 		if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1278 			perror("pipe");
1279 			return 1;
1280 		}
1281 	}
1282 
1283 	if (posix_memalign(&area, page_size, page_size)) {
1284 		fprintf(stderr, "out of memory\n");
1285 		return 1;
1286 	}
1287 	zeropage = area;
1288 	bzero(zeropage, page_size);
1289 
1290 	pthread_mutex_lock(&uffd_read_mutex);
1291 
1292 	pthread_attr_init(&attr);
1293 	pthread_attr_setstacksize(&attr, 16*1024*1024);
1294 
1295 	err = 0;
1296 	while (bounces--) {
1297 		unsigned long expected_ioctls;
1298 
1299 		printf("bounces: %d, mode:", bounces);
1300 		if (bounces & BOUNCE_RANDOM)
1301 			printf(" rnd");
1302 		if (bounces & BOUNCE_RACINGFAULTS)
1303 			printf(" racing");
1304 		if (bounces & BOUNCE_VERIFY)
1305 			printf(" ver");
1306 		if (bounces & BOUNCE_POLL)
1307 			printf(" poll");
1308 		printf(", ");
1309 		fflush(stdout);
1310 
1311 		if (bounces & BOUNCE_POLL)
1312 			fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1313 		else
1314 			fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1315 
1316 		/* register */
1317 		uffdio_register.range.start = (unsigned long) area_dst;
1318 		uffdio_register.range.len = nr_pages * page_size;
1319 		uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1320 		if (test_uffdio_wp)
1321 			uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1322 		if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1323 			fprintf(stderr, "register failure\n");
1324 			return 1;
1325 		}
1326 		expected_ioctls = uffd_test_ops->expected_ioctls;
1327 		if ((uffdio_register.ioctls & expected_ioctls) !=
1328 		    expected_ioctls) {
1329 			fprintf(stderr,
1330 				"unexpected missing ioctl for anon memory\n");
1331 			return 1;
1332 		}
1333 
1334 		if (area_dst_alias) {
1335 			uffdio_register.range.start = (unsigned long)
1336 				area_dst_alias;
1337 			if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1338 				fprintf(stderr, "register failure alias\n");
1339 				return 1;
1340 			}
1341 		}
1342 
1343 		/*
1344 		 * The madvise done previously isn't enough: some
1345 		 * uffd_thread could have read userfaults (one of
1346 		 * those already resolved by the background thread)
1347 		 * and it may be in the process of calling
1348 		 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1349 		 * area_src and it would map a zero page in it (of
1350 		 * course such a UFFDIO_COPY is perfectly safe as it'd
1351 		 * return -EEXIST). The problem comes at the next
1352 		 * bounce though: that racing UFFDIO_COPY would
1353 		 * generate zeropages in the area_src, so invalidating
1354 		 * the previous MADV_DONTNEED. Without this additional
1355 		 * MADV_DONTNEED those zeropages leftovers in the
1356 		 * area_src would lead to -EEXIST failure during the
1357 		 * next bounce, effectively leaving a zeropage in the
1358 		 * area_dst.
1359 		 *
1360 		 * Try to comment this out madvise to see the memory
1361 		 * corruption being caught pretty quick.
1362 		 *
1363 		 * khugepaged is also inhibited to collapse THP after
1364 		 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1365 		 * required to MADV_DONTNEED here.
1366 		 */
1367 		if (uffd_test_ops->release_pages(area_dst))
1368 			return 1;
1369 
1370 		uffd_stats_reset(uffd_stats, nr_cpus);
1371 
1372 		/* bounce pass */
1373 		if (stress(uffd_stats))
1374 			return 1;
1375 
1376 		/* Clear all the write protections if there is any */
1377 		if (test_uffdio_wp)
1378 			wp_range(uffd, (unsigned long)area_dst,
1379 				 nr_pages * page_size, false);
1380 
1381 		/* unregister */
1382 		if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1383 			fprintf(stderr, "unregister failure\n");
1384 			return 1;
1385 		}
1386 		if (area_dst_alias) {
1387 			uffdio_register.range.start = (unsigned long) area_dst;
1388 			if (ioctl(uffd, UFFDIO_UNREGISTER,
1389 				  &uffdio_register.range)) {
1390 				fprintf(stderr, "unregister failure alias\n");
1391 				return 1;
1392 			}
1393 		}
1394 
1395 		/* verification */
1396 		if (bounces & BOUNCE_VERIFY) {
1397 			for (nr = 0; nr < nr_pages; nr++) {
1398 				if (*area_count(area_dst, nr) != count_verify[nr]) {
1399 					fprintf(stderr,
1400 						"error area_count %Lu %Lu %lu\n",
1401 						*area_count(area_src, nr),
1402 						count_verify[nr],
1403 						nr);
1404 					err = 1;
1405 					bounces = 0;
1406 				}
1407 			}
1408 		}
1409 
1410 		/* prepare next bounce */
1411 		tmp_area = area_src;
1412 		area_src = area_dst;
1413 		area_dst = tmp_area;
1414 
1415 		tmp_area = area_src_alias;
1416 		area_src_alias = area_dst_alias;
1417 		area_dst_alias = tmp_area;
1418 
1419 		uffd_stats_report(uffd_stats, nr_cpus);
1420 	}
1421 
1422 	if (err)
1423 		return err;
1424 
1425 	close(uffd);
1426 	return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1427 		|| userfaultfd_events_test();
1428 }
1429 
1430 /*
1431  * Copied from mlock2-tests.c
1432  */
default_huge_page_size(void)1433 unsigned long default_huge_page_size(void)
1434 {
1435 	unsigned long hps = 0;
1436 	char *line = NULL;
1437 	size_t linelen = 0;
1438 	FILE *f = fopen("/proc/meminfo", "r");
1439 
1440 	if (!f)
1441 		return 0;
1442 	while (getline(&line, &linelen, f) > 0) {
1443 		if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1444 			hps <<= 10;
1445 			break;
1446 		}
1447 	}
1448 
1449 	free(line);
1450 	fclose(f);
1451 	return hps;
1452 }
1453 
set_test_type(const char * type)1454 static void set_test_type(const char *type)
1455 {
1456 	if (!strcmp(type, "anon")) {
1457 		test_type = TEST_ANON;
1458 		uffd_test_ops = &anon_uffd_test_ops;
1459 		/* Only enable write-protect test for anonymous test */
1460 		test_uffdio_wp = true;
1461 	} else if (!strcmp(type, "hugetlb")) {
1462 		test_type = TEST_HUGETLB;
1463 		uffd_test_ops = &hugetlb_uffd_test_ops;
1464 	} else if (!strcmp(type, "hugetlb_shared")) {
1465 		map_shared = true;
1466 		test_type = TEST_HUGETLB;
1467 		uffd_test_ops = &hugetlb_uffd_test_ops;
1468 	} else if (!strcmp(type, "shmem")) {
1469 		map_shared = true;
1470 		test_type = TEST_SHMEM;
1471 		uffd_test_ops = &shmem_uffd_test_ops;
1472 	} else {
1473 		fprintf(stderr, "Unknown test type: %s\n", type); exit(1);
1474 	}
1475 
1476 	if (test_type == TEST_HUGETLB)
1477 		page_size = default_huge_page_size();
1478 	else
1479 		page_size = sysconf(_SC_PAGE_SIZE);
1480 
1481 	if (!page_size) {
1482 		fprintf(stderr, "Unable to determine page size\n");
1483 		exit(2);
1484 	}
1485 	if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1486 	    > page_size) {
1487 		fprintf(stderr, "Impossible to run this test\n");
1488 		exit(2);
1489 	}
1490 }
1491 
sigalrm(int sig)1492 static void sigalrm(int sig)
1493 {
1494 	if (sig != SIGALRM)
1495 		abort();
1496 	test_uffdio_copy_eexist = true;
1497 	test_uffdio_zeropage_eexist = true;
1498 	alarm(ALARM_INTERVAL_SECS);
1499 }
1500 
main(int argc,char ** argv)1501 int main(int argc, char **argv)
1502 {
1503 	if (argc < 4)
1504 		usage();
1505 
1506 	if (signal(SIGALRM, sigalrm) == SIG_ERR) {
1507 		fprintf(stderr, "failed to arm SIGALRM");
1508 		exit(1);
1509 	}
1510 	alarm(ALARM_INTERVAL_SECS);
1511 
1512 	set_test_type(argv[1]);
1513 
1514 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1515 	nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1516 		nr_cpus;
1517 	if (!nr_pages_per_cpu) {
1518 		fprintf(stderr, "invalid MiB\n");
1519 		usage();
1520 	}
1521 
1522 	bounces = atoi(argv[3]);
1523 	if (bounces <= 0) {
1524 		fprintf(stderr, "invalid bounces\n");
1525 		usage();
1526 	}
1527 	nr_pages = nr_pages_per_cpu * nr_cpus;
1528 
1529 	if (test_type == TEST_HUGETLB) {
1530 		if (argc < 5)
1531 			usage();
1532 		huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1533 		if (huge_fd < 0) {
1534 			fprintf(stderr, "Open of %s failed", argv[3]);
1535 			perror("open");
1536 			exit(1);
1537 		}
1538 		if (ftruncate(huge_fd, 0)) {
1539 			fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1540 			perror("ftruncate");
1541 			exit(1);
1542 		}
1543 	}
1544 	printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1545 	       nr_pages, nr_pages_per_cpu);
1546 	return userfaultfd_stress();
1547 }
1548 
1549 #else /* __NR_userfaultfd */
1550 
1551 #warning "missing __NR_userfaultfd definition"
1552 
main(void)1553 int main(void)
1554 {
1555 	printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1556 	return KSFT_SKIP;
1557 }
1558 
1559 #endif /* __NR_userfaultfd */
1560