2 * Stress userfaultfd syscall.
4 * Copyright (C) 2015 Red Hat, Inc.
6 * This work is licensed under the terms of the GNU GPL, version 2. See
7 * the COPYING file in the top-level directory.
9 * This test allocates two virtual areas and bounces the physical
10 * memory across the two virtual areas (from area_src to area_dst)
13 * There are three threads running per CPU:
15 * 1) one per-CPU thread takes a per-page pthread_mutex in a random
16 * page of the area_dst (while the physical page may still be in
17 * area_src), and increments a per-page counter in the same page,
18 * and checks its value against a verification region.
20 * 2) another per-CPU thread handles the userfaults generated by
21 * thread 1 above. userfaultfd blocking reads or poll() modes are
22 * exercised interleaved.
24 * 3) one last per-CPU thread transfers the memory in the background
25 * at maximum bandwidth (if not already transferred by thread
26 * 2). Each cpu thread takes cares of transferring a portion of the
29 * When all threads of type 3 completed the transfer, one bounce is
30 * complete. area_src and area_dst are then swapped. All threads are
31 * respawned and so the bounce is immediately restarted in the
34 * per-CPU threads 1 by triggering userfaults inside
35 * pthread_mutex_lock will also verify the atomicity of the memory
36 * transfer (UFFDIO_COPY).
38 * The program takes two parameters: the amounts of physical memory in
39 * megabytes (MiB) of the area and the number of bounces to execute.
41 * # 100MiB 99999 bounces
42 * ./userfaultfd 100 99999
45 * ./userfaultfd 1000 99
47 * # 10MiB-~6GiB 999 bounces, continue forever unless an error triggers
48 * while ./userfaultfd $[RANDOM % 6000 + 10] 999; do true; done
56 #include <sys/types.h>
64 #include <sys/syscall.h>
65 #include <sys/ioctl.h>
68 #include <linux/userfaultfd.h>
72 #ifdef __NR_userfaultfd
74 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
76 #define BOUNCE_RANDOM (1<<0)
77 #define BOUNCE_RACINGFAULTS (1<<1)
78 #define BOUNCE_VERIFY (1<<2)
79 #define BOUNCE_POLL (1<<3)
83 #define TEST_HUGETLB 2
87 /* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
88 #define ALARM_INTERVAL_SECS 10
89 static volatile bool test_uffdio_copy_eexist = true;
90 static volatile bool test_uffdio_zeropage_eexist = true;
92 static bool map_shared;
94 static char *huge_fd_off0;
95 static unsigned long long *count_verify;
96 static int uffd, uffd_flags, finished, *pipefd;
97 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
98 static char *zeropage;
101 /* pthread_mutex_t starts at page offset 0 */
102 #define area_mutex(___area, ___nr) \
103 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
105 * count is placed in the page after pthread_mutex_t naturally aligned
106 * to avoid non alignment faults on non-x86 archs.
108 #define area_count(___area, ___nr) \
109 ((volatile unsigned long long *) ((unsigned long) \
110 ((___area) + (___nr)*page_size + \
111 sizeof(pthread_mutex_t) + \
112 sizeof(unsigned long long) - 1) & \
113 ~(unsigned long)(sizeof(unsigned long long) \
116 static int anon_release_pages(char *rel_area)
120 if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
128 static void anon_allocate_area(void **alloc_area)
130 if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
131 fprintf(stderr, "out of memory\n");
136 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
141 static int hugetlb_release_pages(char *rel_area)
145 if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
146 rel_area == huge_fd_off0 ? 0 :
147 nr_pages * page_size,
148 nr_pages * page_size)) {
157 static void hugetlb_allocate_area(void **alloc_area)
159 void *area_alias = NULL;
160 char **alloc_area_alias;
161 *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
162 (map_shared ? MAP_SHARED : MAP_PRIVATE) |
164 huge_fd, *alloc_area == area_src ? 0 :
165 nr_pages * page_size);
166 if (*alloc_area == MAP_FAILED) {
167 fprintf(stderr, "mmap of hugetlbfs file failed\n");
172 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
173 MAP_SHARED | MAP_HUGETLB,
174 huge_fd, *alloc_area == area_src ? 0 :
175 nr_pages * page_size);
176 if (area_alias == MAP_FAILED) {
177 if (munmap(*alloc_area, nr_pages * page_size) < 0)
178 perror("hugetlb munmap"), exit(1);
183 if (*alloc_area == area_src) {
184 huge_fd_off0 = *alloc_area;
185 alloc_area_alias = &area_src_alias;
187 alloc_area_alias = &area_dst_alias;
190 *alloc_area_alias = area_alias;
193 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
198 * We can't zap just the pagetable with hugetlbfs because
199 * MADV_DONTEED won't work. So exercise -EEXIST on a alias
200 * mapping where the pagetables are not established initially,
201 * this way we'll exercise the -EEXEC at the fs level.
203 *start = (unsigned long) area_dst_alias + offset;
207 static int shmem_release_pages(char *rel_area)
211 if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
219 static void shmem_allocate_area(void **alloc_area)
221 *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
222 MAP_ANONYMOUS | MAP_SHARED, -1, 0);
223 if (*alloc_area == MAP_FAILED) {
224 fprintf(stderr, "shared memory mmap failed\n");
229 struct uffd_test_ops {
230 unsigned long expected_ioctls;
231 void (*allocate_area)(void **alloc_area);
232 int (*release_pages)(char *rel_area);
233 void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
236 #define ANON_EXPECTED_IOCTLS ((1 << _UFFDIO_WAKE) | \
237 (1 << _UFFDIO_COPY) | \
238 (1 << _UFFDIO_ZEROPAGE))
240 static struct uffd_test_ops anon_uffd_test_ops = {
241 .expected_ioctls = ANON_EXPECTED_IOCTLS,
242 .allocate_area = anon_allocate_area,
243 .release_pages = anon_release_pages,
244 .alias_mapping = noop_alias_mapping,
247 static struct uffd_test_ops shmem_uffd_test_ops = {
248 .expected_ioctls = ANON_EXPECTED_IOCTLS,
249 .allocate_area = shmem_allocate_area,
250 .release_pages = shmem_release_pages,
251 .alias_mapping = noop_alias_mapping,
254 static struct uffd_test_ops hugetlb_uffd_test_ops = {
255 .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
256 .allocate_area = hugetlb_allocate_area,
257 .release_pages = hugetlb_release_pages,
258 .alias_mapping = hugetlb_alias_mapping,
261 static struct uffd_test_ops *uffd_test_ops;
263 static int my_bcmp(char *str1, char *str2, size_t n)
266 for (i = 0; i < n; i++)
267 if (str1[i] != str2[i])
272 static void *locking_thread(void *arg)
274 unsigned long cpu = (unsigned long) arg;
275 struct random_data rand;
276 unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
278 unsigned long long count;
283 if (bounces & BOUNCE_RANDOM) {
284 seed = (unsigned int) time(NULL) - bounces;
285 if (!(bounces & BOUNCE_RACINGFAULTS))
287 bzero(&rand, sizeof(rand));
288 bzero(&randstate, sizeof(randstate));
289 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
290 fprintf(stderr, "srandom_r error\n"), exit(1);
293 if (!(bounces & BOUNCE_RACINGFAULTS))
294 page_nr += cpu * nr_pages_per_cpu;
298 if (bounces & BOUNCE_RANDOM) {
299 if (random_r(&rand, &rand_nr))
300 fprintf(stderr, "random_r 1 error\n"), exit(1);
302 if (sizeof(page_nr) > sizeof(rand_nr)) {
303 if (random_r(&rand, &rand_nr))
304 fprintf(stderr, "random_r 2 error\n"), exit(1);
305 page_nr |= (((unsigned long) rand_nr) << 16) <<
313 if (bounces & BOUNCE_VERIFY) {
314 count = *area_count(area_dst, page_nr);
317 "page_nr %lu wrong count %Lu %Lu\n",
319 count_verify[page_nr]), exit(1);
323 * We can't use bcmp (or memcmp) because that
324 * returns 0 erroneously if the memory is
325 * changing under it (even if the end of the
326 * page is never changing and always
330 if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
333 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
335 count_verify[page_nr]), exit(1);
340 /* uncomment the below line to test with mutex */
341 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
342 while (!bcmp(area_dst + page_nr * page_size, zeropage,
348 /* uncomment below line to test with mutex */
349 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
352 "page_nr %lu all zero thread %lu %p %lu\n",
353 page_nr, cpu, area_dst + page_nr * page_size,
361 pthread_mutex_lock(area_mutex(area_dst, page_nr));
362 count = *area_count(area_dst, page_nr);
363 if (count != count_verify[page_nr]) {
365 "page_nr %lu memory corruption %Lu %Lu\n",
367 count_verify[page_nr]), exit(1);
370 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
371 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
373 if (time(NULL) - start > 1)
375 "userfault too slow %ld "
376 "possible false positive with overcommit\n",
383 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
384 unsigned long offset)
386 uffd_test_ops->alias_mapping(&uffdio_copy->dst,
389 if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
390 /* real retval in ufdio_copy.copy */
391 if (uffdio_copy->copy != -EEXIST)
392 fprintf(stderr, "UFFDIO_COPY retry error %Ld\n",
393 uffdio_copy->copy), exit(1);
395 fprintf(stderr, "UFFDIO_COPY retry unexpected %Ld\n",
396 uffdio_copy->copy), exit(1);
400 static int __copy_page(int ufd, unsigned long offset, bool retry)
402 struct uffdio_copy uffdio_copy;
404 if (offset >= nr_pages * page_size)
405 fprintf(stderr, "unexpected offset %lu\n",
407 uffdio_copy.dst = (unsigned long) area_dst + offset;
408 uffdio_copy.src = (unsigned long) area_src + offset;
409 uffdio_copy.len = page_size;
410 uffdio_copy.mode = 0;
411 uffdio_copy.copy = 0;
412 if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
413 /* real retval in ufdio_copy.copy */
414 if (uffdio_copy.copy != -EEXIST)
415 fprintf(stderr, "UFFDIO_COPY error %Ld\n",
416 uffdio_copy.copy), exit(1);
417 } else if (uffdio_copy.copy != page_size) {
418 fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
419 uffdio_copy.copy), exit(1);
421 if (test_uffdio_copy_eexist && retry) {
422 test_uffdio_copy_eexist = false;
423 retry_copy_page(ufd, &uffdio_copy, offset);
430 static int copy_page_retry(int ufd, unsigned long offset)
432 return __copy_page(ufd, offset, true);
435 static int copy_page(int ufd, unsigned long offset)
437 return __copy_page(ufd, offset, false);
440 static void *uffd_poll_thread(void *arg)
442 unsigned long cpu = (unsigned long) arg;
443 struct pollfd pollfd[2];
445 struct uffdio_register uffd_reg;
447 unsigned long offset;
449 unsigned long userfaults = 0;
452 pollfd[0].events = POLLIN;
453 pollfd[1].fd = pipefd[cpu*2];
454 pollfd[1].events = POLLIN;
457 ret = poll(pollfd, 2, -1);
459 fprintf(stderr, "poll error %d\n", ret), exit(1);
461 perror("poll"), exit(1);
462 if (pollfd[1].revents & POLLIN) {
463 if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
464 fprintf(stderr, "read pipefd error\n"),
468 if (!(pollfd[0].revents & POLLIN))
469 fprintf(stderr, "pollfd[0].revents %d\n",
470 pollfd[0].revents), exit(1);
471 ret = read(uffd, &msg, sizeof(msg));
475 perror("nonblocking read error"), exit(1);
479 fprintf(stderr, "unexpected msg event %u\n",
482 case UFFD_EVENT_PAGEFAULT:
483 if (msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
484 fprintf(stderr, "unexpected write fault\n"), exit(1);
485 offset = (char *)(unsigned long)msg.arg.pagefault.address -
487 offset &= ~(page_size-1);
488 if (copy_page(uffd, offset))
491 case UFFD_EVENT_FORK:
493 uffd = msg.arg.fork.ufd;
496 case UFFD_EVENT_REMOVE:
497 uffd_reg.range.start = msg.arg.remove.start;
498 uffd_reg.range.len = msg.arg.remove.end -
499 msg.arg.remove.start;
500 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
501 fprintf(stderr, "remove failure\n"), exit(1);
503 case UFFD_EVENT_REMAP:
504 area_dst = (char *)(unsigned long)msg.arg.remap.to;
508 return (void *)userfaults;
511 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
513 static void *uffd_read_thread(void *arg)
515 unsigned long *this_cpu_userfaults;
517 unsigned long offset;
520 this_cpu_userfaults = (unsigned long *) arg;
521 *this_cpu_userfaults = 0;
523 pthread_mutex_unlock(&uffd_read_mutex);
524 /* from here cancellation is ok */
527 ret = read(uffd, &msg, sizeof(msg));
528 if (ret != sizeof(msg)) {
530 perror("blocking read error"), exit(1);
532 fprintf(stderr, "short read\n"), exit(1);
534 if (msg.event != UFFD_EVENT_PAGEFAULT)
535 fprintf(stderr, "unexpected msg event %u\n",
537 if (bounces & BOUNCE_VERIFY &&
538 msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
539 fprintf(stderr, "unexpected write fault\n"), exit(1);
540 offset = (char *)(unsigned long)msg.arg.pagefault.address -
542 offset &= ~(page_size-1);
543 if (copy_page(uffd, offset))
544 (*this_cpu_userfaults)++;
549 static void *background_thread(void *arg)
551 unsigned long cpu = (unsigned long) arg;
552 unsigned long page_nr;
554 for (page_nr = cpu * nr_pages_per_cpu;
555 page_nr < (cpu+1) * nr_pages_per_cpu;
557 copy_page_retry(uffd, page_nr * page_size);
562 static int stress(unsigned long *userfaults)
565 pthread_t locking_threads[nr_cpus];
566 pthread_t uffd_threads[nr_cpus];
567 pthread_t background_threads[nr_cpus];
568 void **_userfaults = (void **) userfaults;
571 for (cpu = 0; cpu < nr_cpus; cpu++) {
572 if (pthread_create(&locking_threads[cpu], &attr,
573 locking_thread, (void *)cpu))
575 if (bounces & BOUNCE_POLL) {
576 if (pthread_create(&uffd_threads[cpu], &attr,
577 uffd_poll_thread, (void *)cpu))
580 if (pthread_create(&uffd_threads[cpu], &attr,
584 pthread_mutex_lock(&uffd_read_mutex);
586 if (pthread_create(&background_threads[cpu], &attr,
587 background_thread, (void *)cpu))
590 for (cpu = 0; cpu < nr_cpus; cpu++)
591 if (pthread_join(background_threads[cpu], NULL))
595 * Be strict and immediately zap area_src, the whole area has
596 * been transferred already by the background treads. The
597 * area_src could then be faulted in in a racy way by still
598 * running uffdio_threads reading zeropages after we zapped
599 * area_src (but they're guaranteed to get -EEXIST from
600 * UFFDIO_COPY without writing zero pages into area_dst
601 * because the background threads already completed).
603 if (uffd_test_ops->release_pages(area_src))
606 for (cpu = 0; cpu < nr_cpus; cpu++) {
608 if (bounces & BOUNCE_POLL) {
609 if (write(pipefd[cpu*2+1], &c, 1) != 1) {
610 fprintf(stderr, "pipefd write error\n");
613 if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
616 if (pthread_cancel(uffd_threads[cpu]))
618 if (pthread_join(uffd_threads[cpu], NULL))
624 for (cpu = 0; cpu < nr_cpus; cpu++)
625 if (pthread_join(locking_threads[cpu], NULL))
631 static int userfaultfd_open(int features)
633 struct uffdio_api uffdio_api;
635 uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
638 "userfaultfd syscall not available in this kernel\n");
641 uffd_flags = fcntl(uffd, F_GETFD, NULL);
643 uffdio_api.api = UFFD_API;
644 uffdio_api.features = features;
645 if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
646 fprintf(stderr, "UFFDIO_API\n");
649 if (uffdio_api.api != UFFD_API) {
650 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
657 sigjmp_buf jbuf, *sigbuf;
659 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
663 siglongjmp(*sigbuf, 1);
669 * For non-cooperative userfaultfd test we fork() a process that will
670 * generate pagefaults, will mremap the area monitored by the
671 * userfaultfd and at last this process will release the monitored
673 * For the anonymous and shared memory the area is divided into two
674 * parts, the first part is accessed before mremap, and the second
675 * part is accessed after mremap. Since hugetlbfs does not support
676 * mremap, the entire monitored area is accessed in a single pass for
678 * The release of the pages currently generates event for shmem and
679 * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
681 * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
682 * monitored area, generate pagefaults and test that signal is delivered.
683 * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
684 * test robustness use case - we release monitored area, fork a process
685 * that will generate pagefaults and verify signal is generated.
686 * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
687 * feature. Using monitor thread, verify no userfault events are generated.
689 static int faulting_process(int signal_test)
692 unsigned long long count;
693 unsigned long split_nr_pages;
694 unsigned long lastnr;
695 struct sigaction act;
696 unsigned long signalled = 0;
698 if (test_type != TEST_HUGETLB)
699 split_nr_pages = (nr_pages + 1) / 2;
701 split_nr_pages = nr_pages;
705 memset(&act, 0, sizeof(act));
706 act.sa_sigaction = sighndl;
707 act.sa_flags = SA_SIGINFO;
708 if (sigaction(SIGBUS, &act, 0)) {
712 lastnr = (unsigned long)-1;
715 for (nr = 0; nr < split_nr_pages; nr++) {
717 if (sigsetjmp(*sigbuf, 1) != 0) {
719 fprintf(stderr, "Signal repeated\n");
724 if (signal_test == 1) {
725 if (copy_page(uffd, nr * page_size))
734 count = *area_count(area_dst, nr);
735 if (count != count_verify[nr]) {
737 "nr %lu memory corruption %Lu %Lu\n",
739 count_verify[nr]), exit(1);
744 return signalled != split_nr_pages;
746 if (test_type == TEST_HUGETLB)
749 area_dst = mremap(area_dst, nr_pages * page_size, nr_pages * page_size,
750 MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
751 if (area_dst == MAP_FAILED)
752 perror("mremap"), exit(1);
754 for (; nr < nr_pages; nr++) {
755 count = *area_count(area_dst, nr);
756 if (count != count_verify[nr]) {
758 "nr %lu memory corruption %Lu %Lu\n",
760 count_verify[nr]), exit(1);
764 if (uffd_test_ops->release_pages(area_dst))
767 for (nr = 0; nr < nr_pages; nr++) {
768 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
769 fprintf(stderr, "nr %lu is not zero\n", nr), exit(1);
775 static void retry_uffdio_zeropage(int ufd,
776 struct uffdio_zeropage *uffdio_zeropage,
777 unsigned long offset)
779 uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
780 uffdio_zeropage->range.len,
782 if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
783 if (uffdio_zeropage->zeropage != -EEXIST)
784 fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
785 uffdio_zeropage->zeropage), exit(1);
787 fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
788 uffdio_zeropage->zeropage), exit(1);
792 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
794 struct uffdio_zeropage uffdio_zeropage;
796 unsigned long has_zeropage;
798 has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
800 if (offset >= nr_pages * page_size)
801 fprintf(stderr, "unexpected offset %lu\n",
803 uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
804 uffdio_zeropage.range.len = page_size;
805 uffdio_zeropage.mode = 0;
806 ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
808 /* real retval in ufdio_zeropage.zeropage */
810 if (uffdio_zeropage.zeropage == -EEXIST)
811 fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n"),
814 fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
815 uffdio_zeropage.zeropage), exit(1);
817 if (uffdio_zeropage.zeropage != -EINVAL)
819 "UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
820 uffdio_zeropage.zeropage), exit(1);
822 } else if (has_zeropage) {
823 if (uffdio_zeropage.zeropage != page_size) {
824 fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
825 uffdio_zeropage.zeropage), exit(1);
827 if (test_uffdio_zeropage_eexist && retry) {
828 test_uffdio_zeropage_eexist = false;
829 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
836 "UFFDIO_ZEROPAGE succeeded %Ld\n",
837 uffdio_zeropage.zeropage), exit(1);
843 static int uffdio_zeropage(int ufd, unsigned long offset)
845 return __uffdio_zeropage(ufd, offset, false);
848 /* exercise UFFDIO_ZEROPAGE */
849 static int userfaultfd_zeropage_test(void)
851 struct uffdio_register uffdio_register;
852 unsigned long expected_ioctls;
854 printf("testing UFFDIO_ZEROPAGE: ");
857 if (uffd_test_ops->release_pages(area_dst))
860 if (userfaultfd_open(0) < 0)
862 uffdio_register.range.start = (unsigned long) area_dst;
863 uffdio_register.range.len = nr_pages * page_size;
864 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
865 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
866 fprintf(stderr, "register failure\n"), exit(1);
868 expected_ioctls = uffd_test_ops->expected_ioctls;
869 if ((uffdio_register.ioctls & expected_ioctls) !=
872 "unexpected missing ioctl for anon memory\n"),
875 if (uffdio_zeropage(uffd, 0)) {
876 if (my_bcmp(area_dst, zeropage, page_size))
877 fprintf(stderr, "zeropage is not zero\n"), exit(1);
885 static int userfaultfd_events_test(void)
887 struct uffdio_register uffdio_register;
888 unsigned long expected_ioctls;
889 unsigned long userfaults;
895 printf("testing events (fork, remap, remove): ");
898 if (uffd_test_ops->release_pages(area_dst))
901 features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
902 UFFD_FEATURE_EVENT_REMOVE;
903 if (userfaultfd_open(features) < 0)
905 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
907 uffdio_register.range.start = (unsigned long) area_dst;
908 uffdio_register.range.len = nr_pages * page_size;
909 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
910 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
911 fprintf(stderr, "register failure\n"), exit(1);
913 expected_ioctls = uffd_test_ops->expected_ioctls;
914 if ((uffdio_register.ioctls & expected_ioctls) !=
917 "unexpected missing ioctl for anon memory\n"),
920 if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
921 perror("uffd_poll_thread create"), exit(1);
925 perror("fork"), exit(1);
928 return faulting_process(0);
930 waitpid(pid, &err, 0);
932 fprintf(stderr, "faulting process failed\n"), exit(1);
934 if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
935 perror("pipe write"), exit(1);
936 if (pthread_join(uffd_mon, (void **)&userfaults))
940 printf("userfaults: %ld\n", userfaults);
942 return userfaults != nr_pages;
945 static int userfaultfd_sig_test(void)
947 struct uffdio_register uffdio_register;
948 unsigned long expected_ioctls;
949 unsigned long userfaults;
955 printf("testing signal delivery: ");
958 if (uffd_test_ops->release_pages(area_dst))
961 features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
962 if (userfaultfd_open(features) < 0)
964 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
966 uffdio_register.range.start = (unsigned long) area_dst;
967 uffdio_register.range.len = nr_pages * page_size;
968 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
969 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
970 fprintf(stderr, "register failure\n"), exit(1);
972 expected_ioctls = uffd_test_ops->expected_ioctls;
973 if ((uffdio_register.ioctls & expected_ioctls) !=
976 "unexpected missing ioctl for anon memory\n"),
979 if (faulting_process(1))
980 fprintf(stderr, "faulting process failed\n"), exit(1);
982 if (uffd_test_ops->release_pages(area_dst))
985 if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
986 perror("uffd_poll_thread create"), exit(1);
990 perror("fork"), exit(1);
993 exit(faulting_process(2));
995 waitpid(pid, &err, 0);
997 fprintf(stderr, "faulting process failed\n"), exit(1);
999 if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1000 perror("pipe write"), exit(1);
1001 if (pthread_join(uffd_mon, (void **)&userfaults))
1006 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1009 return userfaults != 0;
1011 static int userfaultfd_stress(void)
1016 struct uffdio_register uffdio_register;
1019 unsigned long userfaults[nr_cpus];
1021 uffd_test_ops->allocate_area((void **)&area_src);
1024 uffd_test_ops->allocate_area((void **)&area_dst);
1028 if (userfaultfd_open(0) < 0)
1031 count_verify = malloc(nr_pages * sizeof(unsigned long long));
1032 if (!count_verify) {
1033 perror("count_verify");
1037 for (nr = 0; nr < nr_pages; nr++) {
1038 *area_mutex(area_src, nr) = (pthread_mutex_t)
1039 PTHREAD_MUTEX_INITIALIZER;
1040 count_verify[nr] = *area_count(area_src, nr) = 1;
1042 * In the transition between 255 to 256, powerpc will
1043 * read out of order in my_bcmp and see both bytes as
1044 * zero, so leave a placeholder below always non-zero
1045 * after the count, to avoid my_bcmp to trigger false
1048 *(area_count(area_src, nr) + 1) = 1;
1051 pipefd = malloc(sizeof(int) * nr_cpus * 2);
1056 for (cpu = 0; cpu < nr_cpus; cpu++) {
1057 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1063 if (posix_memalign(&area, page_size, page_size)) {
1064 fprintf(stderr, "out of memory\n");
1068 bzero(zeropage, page_size);
1070 pthread_mutex_lock(&uffd_read_mutex);
1072 pthread_attr_init(&attr);
1073 pthread_attr_setstacksize(&attr, 16*1024*1024);
1077 unsigned long expected_ioctls;
1079 printf("bounces: %d, mode:", bounces);
1080 if (bounces & BOUNCE_RANDOM)
1082 if (bounces & BOUNCE_RACINGFAULTS)
1084 if (bounces & BOUNCE_VERIFY)
1086 if (bounces & BOUNCE_POLL)
1091 if (bounces & BOUNCE_POLL)
1092 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1094 fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1097 uffdio_register.range.start = (unsigned long) area_dst;
1098 uffdio_register.range.len = nr_pages * page_size;
1099 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1100 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1101 fprintf(stderr, "register failure\n");
1104 expected_ioctls = uffd_test_ops->expected_ioctls;
1105 if ((uffdio_register.ioctls & expected_ioctls) !=
1108 "unexpected missing ioctl for anon memory\n");
1112 if (area_dst_alias) {
1113 uffdio_register.range.start = (unsigned long)
1115 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1116 fprintf(stderr, "register failure alias\n");
1122 * The madvise done previously isn't enough: some
1123 * uffd_thread could have read userfaults (one of
1124 * those already resolved by the background thread)
1125 * and it may be in the process of calling
1126 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1127 * area_src and it would map a zero page in it (of
1128 * course such a UFFDIO_COPY is perfectly safe as it'd
1129 * return -EEXIST). The problem comes at the next
1130 * bounce though: that racing UFFDIO_COPY would
1131 * generate zeropages in the area_src, so invalidating
1132 * the previous MADV_DONTNEED. Without this additional
1133 * MADV_DONTNEED those zeropages leftovers in the
1134 * area_src would lead to -EEXIST failure during the
1135 * next bounce, effectively leaving a zeropage in the
1138 * Try to comment this out madvise to see the memory
1139 * corruption being caught pretty quick.
1141 * khugepaged is also inhibited to collapse THP after
1142 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1143 * required to MADV_DONTNEED here.
1145 if (uffd_test_ops->release_pages(area_dst))
1149 if (stress(userfaults))
1153 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1154 fprintf(stderr, "unregister failure\n");
1157 if (area_dst_alias) {
1158 uffdio_register.range.start = (unsigned long) area_dst;
1159 if (ioctl(uffd, UFFDIO_UNREGISTER,
1160 &uffdio_register.range)) {
1161 fprintf(stderr, "unregister failure alias\n");
1167 if (bounces & BOUNCE_VERIFY) {
1168 for (nr = 0; nr < nr_pages; nr++) {
1169 if (*area_count(area_dst, nr) != count_verify[nr]) {
1171 "error area_count %Lu %Lu %lu\n",
1172 *area_count(area_src, nr),
1181 /* prepare next bounce */
1182 tmp_area = area_src;
1183 area_src = area_dst;
1184 area_dst = tmp_area;
1186 tmp_area = area_src_alias;
1187 area_src_alias = area_dst_alias;
1188 area_dst_alias = tmp_area;
1190 printf("userfaults:");
1191 for (cpu = 0; cpu < nr_cpus; cpu++)
1192 printf(" %lu", userfaults[cpu]);
1200 return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1201 || userfaultfd_events_test();
1205 * Copied from mlock2-tests.c
1207 unsigned long default_huge_page_size(void)
1209 unsigned long hps = 0;
1212 FILE *f = fopen("/proc/meminfo", "r");
1216 while (getline(&line, &linelen, f) > 0) {
1217 if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) {
1228 static void set_test_type(const char *type)
1230 if (!strcmp(type, "anon")) {
1231 test_type = TEST_ANON;
1232 uffd_test_ops = &anon_uffd_test_ops;
1233 } else if (!strcmp(type, "hugetlb")) {
1234 test_type = TEST_HUGETLB;
1235 uffd_test_ops = &hugetlb_uffd_test_ops;
1236 } else if (!strcmp(type, "hugetlb_shared")) {
1238 test_type = TEST_HUGETLB;
1239 uffd_test_ops = &hugetlb_uffd_test_ops;
1240 } else if (!strcmp(type, "shmem")) {
1242 test_type = TEST_SHMEM;
1243 uffd_test_ops = &shmem_uffd_test_ops;
1245 fprintf(stderr, "Unknown test type: %s\n", type), exit(1);
1248 if (test_type == TEST_HUGETLB)
1249 page_size = default_huge_page_size();
1251 page_size = sysconf(_SC_PAGE_SIZE);
1254 fprintf(stderr, "Unable to determine page size\n"),
1256 if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1258 fprintf(stderr, "Impossible to run this test\n"), exit(2);
1261 static void sigalrm(int sig)
1265 test_uffdio_copy_eexist = true;
1266 test_uffdio_zeropage_eexist = true;
1267 alarm(ALARM_INTERVAL_SECS);
1270 int main(int argc, char **argv)
1273 fprintf(stderr, "Usage: <test type> <MiB> <bounces> [hugetlbfs_file]\n"),
1276 if (signal(SIGALRM, sigalrm) == SIG_ERR)
1277 fprintf(stderr, "failed to arm SIGALRM"), exit(1);
1278 alarm(ALARM_INTERVAL_SECS);
1280 set_test_type(argv[1]);
1282 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1283 nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1285 if (!nr_pages_per_cpu) {
1286 fprintf(stderr, "invalid MiB\n");
1287 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
1290 bounces = atoi(argv[3]);
1292 fprintf(stderr, "invalid bounces\n");
1293 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
1295 nr_pages = nr_pages_per_cpu * nr_cpus;
1297 if (test_type == TEST_HUGETLB) {
1299 fprintf(stderr, "Usage: hugetlb <MiB> <bounces> <hugetlbfs_file>\n"),
1301 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1303 fprintf(stderr, "Open of %s failed", argv[3]);
1307 if (ftruncate(huge_fd, 0)) {
1308 fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1309 perror("ftruncate");
1313 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1314 nr_pages, nr_pages_per_cpu);
1315 return userfaultfd_stress();
1318 #else /* __NR_userfaultfd */
1320 #warning "missing __NR_userfaultfd definition"
1324 printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1328 #endif /* __NR_userfaultfd */