Loading...
Note: File does not exist in v3.15.
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 <sys/mman.h>
50#include <sys/syscall.h>
51#include <sys/ioctl.h>
52#include <sys/wait.h>
53#include <pthread.h>
54#include <linux/userfaultfd.h>
55#include <setjmp.h>
56#include <stdbool.h>
57
58#include "../kselftest.h"
59
60#ifdef __NR_userfaultfd
61
62static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
63
64#define BOUNCE_RANDOM (1<<0)
65#define BOUNCE_RACINGFAULTS (1<<1)
66#define BOUNCE_VERIFY (1<<2)
67#define BOUNCE_POLL (1<<3)
68static int bounces;
69
70#define TEST_ANON 1
71#define TEST_HUGETLB 2
72#define TEST_SHMEM 3
73static int test_type;
74
75/* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
76#define ALARM_INTERVAL_SECS 10
77static volatile bool test_uffdio_copy_eexist = true;
78static volatile bool test_uffdio_zeropage_eexist = true;
79
80static bool map_shared;
81static int huge_fd;
82static char *huge_fd_off0;
83static unsigned long long *count_verify;
84static int uffd, uffd_flags, finished, *pipefd;
85static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
86static char *zeropage;
87pthread_attr_t attr;
88
89/* pthread_mutex_t starts at page offset 0 */
90#define area_mutex(___area, ___nr) \
91 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
92/*
93 * count is placed in the page after pthread_mutex_t naturally aligned
94 * to avoid non alignment faults on non-x86 archs.
95 */
96#define area_count(___area, ___nr) \
97 ((volatile unsigned long long *) ((unsigned long) \
98 ((___area) + (___nr)*page_size + \
99 sizeof(pthread_mutex_t) + \
100 sizeof(unsigned long long) - 1) & \
101 ~(unsigned long)(sizeof(unsigned long long) \
102 - 1)))
103
104const char *examples =
105 "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
106 "./userfaultfd anon 100 99999\n\n"
107 "# Run share memory test on 1GiB region with 99 bounces:\n"
108 "./userfaultfd shmem 1000 99\n\n"
109 "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
110 "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
111 "# Run the same hugetlb test but using shmem:\n"
112 "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
113 "# 10MiB-~6GiB 999 bounces anonymous test, "
114 "continue forever unless an error triggers\n"
115 "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
116
117static void usage(void)
118{
119 fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
120 "[hugetlbfs_file]\n\n");
121 fprintf(stderr, "Supported <test type>: anon, hugetlb, "
122 "hugetlb_shared, shmem\n\n");
123 fprintf(stderr, "Examples:\n\n");
124 fprintf(stderr, "%s", examples);
125 exit(1);
126}
127
128static int anon_release_pages(char *rel_area)
129{
130 int ret = 0;
131
132 if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
133 perror("madvise");
134 ret = 1;
135 }
136
137 return ret;
138}
139
140static void anon_allocate_area(void **alloc_area)
141{
142 if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
143 fprintf(stderr, "out of memory\n");
144 *alloc_area = NULL;
145 }
146}
147
148static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
149{
150}
151
152/* HugeTLB memory */
153static int hugetlb_release_pages(char *rel_area)
154{
155 int ret = 0;
156
157 if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
158 rel_area == huge_fd_off0 ? 0 :
159 nr_pages * page_size,
160 nr_pages * page_size)) {
161 perror("fallocate");
162 ret = 1;
163 }
164
165 return ret;
166}
167
168
169static void hugetlb_allocate_area(void **alloc_area)
170{
171 void *area_alias = NULL;
172 char **alloc_area_alias;
173 *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
174 (map_shared ? MAP_SHARED : MAP_PRIVATE) |
175 MAP_HUGETLB,
176 huge_fd, *alloc_area == area_src ? 0 :
177 nr_pages * page_size);
178 if (*alloc_area == MAP_FAILED) {
179 fprintf(stderr, "mmap of hugetlbfs file failed\n");
180 *alloc_area = NULL;
181 }
182
183 if (map_shared) {
184 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
185 MAP_SHARED | MAP_HUGETLB,
186 huge_fd, *alloc_area == area_src ? 0 :
187 nr_pages * page_size);
188 if (area_alias == MAP_FAILED) {
189 if (munmap(*alloc_area, nr_pages * page_size) < 0)
190 perror("hugetlb munmap"), exit(1);
191 *alloc_area = NULL;
192 return;
193 }
194 }
195 if (*alloc_area == area_src) {
196 huge_fd_off0 = *alloc_area;
197 alloc_area_alias = &area_src_alias;
198 } else {
199 alloc_area_alias = &area_dst_alias;
200 }
201 if (area_alias)
202 *alloc_area_alias = area_alias;
203}
204
205static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
206{
207 if (!map_shared)
208 return;
209 /*
210 * We can't zap just the pagetable with hugetlbfs because
211 * MADV_DONTEED won't work. So exercise -EEXIST on a alias
212 * mapping where the pagetables are not established initially,
213 * this way we'll exercise the -EEXEC at the fs level.
214 */
215 *start = (unsigned long) area_dst_alias + offset;
216}
217
218/* Shared memory */
219static int shmem_release_pages(char *rel_area)
220{
221 int ret = 0;
222
223 if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
224 perror("madvise");
225 ret = 1;
226 }
227
228 return ret;
229}
230
231static void shmem_allocate_area(void **alloc_area)
232{
233 *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
234 MAP_ANONYMOUS | MAP_SHARED, -1, 0);
235 if (*alloc_area == MAP_FAILED) {
236 fprintf(stderr, "shared memory mmap failed\n");
237 *alloc_area = NULL;
238 }
239}
240
241struct uffd_test_ops {
242 unsigned long expected_ioctls;
243 void (*allocate_area)(void **alloc_area);
244 int (*release_pages)(char *rel_area);
245 void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
246};
247
248#define ANON_EXPECTED_IOCTLS ((1 << _UFFDIO_WAKE) | \
249 (1 << _UFFDIO_COPY) | \
250 (1 << _UFFDIO_ZEROPAGE))
251
252static struct uffd_test_ops anon_uffd_test_ops = {
253 .expected_ioctls = ANON_EXPECTED_IOCTLS,
254 .allocate_area = anon_allocate_area,
255 .release_pages = anon_release_pages,
256 .alias_mapping = noop_alias_mapping,
257};
258
259static struct uffd_test_ops shmem_uffd_test_ops = {
260 .expected_ioctls = ANON_EXPECTED_IOCTLS,
261 .allocate_area = shmem_allocate_area,
262 .release_pages = shmem_release_pages,
263 .alias_mapping = noop_alias_mapping,
264};
265
266static struct uffd_test_ops hugetlb_uffd_test_ops = {
267 .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
268 .allocate_area = hugetlb_allocate_area,
269 .release_pages = hugetlb_release_pages,
270 .alias_mapping = hugetlb_alias_mapping,
271};
272
273static struct uffd_test_ops *uffd_test_ops;
274
275static int my_bcmp(char *str1, char *str2, size_t n)
276{
277 unsigned long i;
278 for (i = 0; i < n; i++)
279 if (str1[i] != str2[i])
280 return 1;
281 return 0;
282}
283
284static void *locking_thread(void *arg)
285{
286 unsigned long cpu = (unsigned long) arg;
287 struct random_data rand;
288 unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
289 int32_t rand_nr;
290 unsigned long long count;
291 char randstate[64];
292 unsigned int seed;
293 time_t start;
294
295 if (bounces & BOUNCE_RANDOM) {
296 seed = (unsigned int) time(NULL) - bounces;
297 if (!(bounces & BOUNCE_RACINGFAULTS))
298 seed += cpu;
299 bzero(&rand, sizeof(rand));
300 bzero(&randstate, sizeof(randstate));
301 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
302 fprintf(stderr, "srandom_r error\n"), exit(1);
303 } else {
304 page_nr = -bounces;
305 if (!(bounces & BOUNCE_RACINGFAULTS))
306 page_nr += cpu * nr_pages_per_cpu;
307 }
308
309 while (!finished) {
310 if (bounces & BOUNCE_RANDOM) {
311 if (random_r(&rand, &rand_nr))
312 fprintf(stderr, "random_r 1 error\n"), exit(1);
313 page_nr = rand_nr;
314 if (sizeof(page_nr) > sizeof(rand_nr)) {
315 if (random_r(&rand, &rand_nr))
316 fprintf(stderr, "random_r 2 error\n"), exit(1);
317 page_nr |= (((unsigned long) rand_nr) << 16) <<
318 16;
319 }
320 } else
321 page_nr += 1;
322 page_nr %= nr_pages;
323
324 start = time(NULL);
325 if (bounces & BOUNCE_VERIFY) {
326 count = *area_count(area_dst, page_nr);
327 if (!count)
328 fprintf(stderr,
329 "page_nr %lu wrong count %Lu %Lu\n",
330 page_nr, count,
331 count_verify[page_nr]), exit(1);
332
333
334 /*
335 * We can't use bcmp (or memcmp) because that
336 * returns 0 erroneously if the memory is
337 * changing under it (even if the end of the
338 * page is never changing and always
339 * different).
340 */
341#if 1
342 if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
343 page_size))
344 fprintf(stderr,
345 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
346 page_nr, count,
347 count_verify[page_nr]), exit(1);
348#else
349 unsigned long loops;
350
351 loops = 0;
352 /* uncomment the below line to test with mutex */
353 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
354 while (!bcmp(area_dst + page_nr * page_size, zeropage,
355 page_size)) {
356 loops += 1;
357 if (loops > 10)
358 break;
359 }
360 /* uncomment below line to test with mutex */
361 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
362 if (loops) {
363 fprintf(stderr,
364 "page_nr %lu all zero thread %lu %p %lu\n",
365 page_nr, cpu, area_dst + page_nr * page_size,
366 loops);
367 if (loops > 10)
368 exit(1);
369 }
370#endif
371 }
372
373 pthread_mutex_lock(area_mutex(area_dst, page_nr));
374 count = *area_count(area_dst, page_nr);
375 if (count != count_verify[page_nr]) {
376 fprintf(stderr,
377 "page_nr %lu memory corruption %Lu %Lu\n",
378 page_nr, count,
379 count_verify[page_nr]), exit(1);
380 }
381 count++;
382 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
383 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
384
385 if (time(NULL) - start > 1)
386 fprintf(stderr,
387 "userfault too slow %ld "
388 "possible false positive with overcommit\n",
389 time(NULL) - start);
390 }
391
392 return NULL;
393}
394
395static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
396 unsigned long offset)
397{
398 uffd_test_ops->alias_mapping(&uffdio_copy->dst,
399 uffdio_copy->len,
400 offset);
401 if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
402 /* real retval in ufdio_copy.copy */
403 if (uffdio_copy->copy != -EEXIST)
404 fprintf(stderr, "UFFDIO_COPY retry error %Ld\n",
405 uffdio_copy->copy), exit(1);
406 } else {
407 fprintf(stderr, "UFFDIO_COPY retry unexpected %Ld\n",
408 uffdio_copy->copy), exit(1);
409 }
410}
411
412static int __copy_page(int ufd, unsigned long offset, bool retry)
413{
414 struct uffdio_copy uffdio_copy;
415
416 if (offset >= nr_pages * page_size)
417 fprintf(stderr, "unexpected offset %lu\n",
418 offset), exit(1);
419 uffdio_copy.dst = (unsigned long) area_dst + offset;
420 uffdio_copy.src = (unsigned long) area_src + offset;
421 uffdio_copy.len = page_size;
422 uffdio_copy.mode = 0;
423 uffdio_copy.copy = 0;
424 if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
425 /* real retval in ufdio_copy.copy */
426 if (uffdio_copy.copy != -EEXIST)
427 fprintf(stderr, "UFFDIO_COPY error %Ld\n",
428 uffdio_copy.copy), exit(1);
429 } else if (uffdio_copy.copy != page_size) {
430 fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
431 uffdio_copy.copy), exit(1);
432 } else {
433 if (test_uffdio_copy_eexist && retry) {
434 test_uffdio_copy_eexist = false;
435 retry_copy_page(ufd, &uffdio_copy, offset);
436 }
437 return 1;
438 }
439 return 0;
440}
441
442static int copy_page_retry(int ufd, unsigned long offset)
443{
444 return __copy_page(ufd, offset, true);
445}
446
447static int copy_page(int ufd, unsigned long offset)
448{
449 return __copy_page(ufd, offset, false);
450}
451
452static int uffd_read_msg(int ufd, struct uffd_msg *msg)
453{
454 int ret = read(uffd, msg, sizeof(*msg));
455
456 if (ret != sizeof(*msg)) {
457 if (ret < 0) {
458 if (errno == EAGAIN)
459 return 1;
460 else
461 perror("blocking read error"), exit(1);
462 } else {
463 fprintf(stderr, "short read\n"), exit(1);
464 }
465 }
466
467 return 0;
468}
469
470/* Return 1 if page fault handled by us; otherwise 0 */
471static int uffd_handle_page_fault(struct uffd_msg *msg)
472{
473 unsigned long offset;
474
475 if (msg->event != UFFD_EVENT_PAGEFAULT)
476 fprintf(stderr, "unexpected msg event %u\n",
477 msg->event), exit(1);
478
479 if (bounces & BOUNCE_VERIFY &&
480 msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
481 fprintf(stderr, "unexpected write fault\n"), exit(1);
482
483 offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
484 offset &= ~(page_size-1);
485
486 return copy_page(uffd, offset);
487}
488
489static void *uffd_poll_thread(void *arg)
490{
491 unsigned long cpu = (unsigned long) arg;
492 struct pollfd pollfd[2];
493 struct uffd_msg msg;
494 struct uffdio_register uffd_reg;
495 int ret;
496 char tmp_chr;
497 unsigned long userfaults = 0;
498
499 pollfd[0].fd = uffd;
500 pollfd[0].events = POLLIN;
501 pollfd[1].fd = pipefd[cpu*2];
502 pollfd[1].events = POLLIN;
503
504 for (;;) {
505 ret = poll(pollfd, 2, -1);
506 if (!ret)
507 fprintf(stderr, "poll error %d\n", ret), exit(1);
508 if (ret < 0)
509 perror("poll"), exit(1);
510 if (pollfd[1].revents & POLLIN) {
511 if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
512 fprintf(stderr, "read pipefd error\n"),
513 exit(1);
514 break;
515 }
516 if (!(pollfd[0].revents & POLLIN))
517 fprintf(stderr, "pollfd[0].revents %d\n",
518 pollfd[0].revents), exit(1);
519 if (uffd_read_msg(uffd, &msg))
520 continue;
521 switch (msg.event) {
522 default:
523 fprintf(stderr, "unexpected msg event %u\n",
524 msg.event), exit(1);
525 break;
526 case UFFD_EVENT_PAGEFAULT:
527 userfaults += uffd_handle_page_fault(&msg);
528 break;
529 case UFFD_EVENT_FORK:
530 close(uffd);
531 uffd = msg.arg.fork.ufd;
532 pollfd[0].fd = uffd;
533 break;
534 case UFFD_EVENT_REMOVE:
535 uffd_reg.range.start = msg.arg.remove.start;
536 uffd_reg.range.len = msg.arg.remove.end -
537 msg.arg.remove.start;
538 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
539 fprintf(stderr, "remove failure\n"), exit(1);
540 break;
541 case UFFD_EVENT_REMAP:
542 area_dst = (char *)(unsigned long)msg.arg.remap.to;
543 break;
544 }
545 }
546 return (void *)userfaults;
547}
548
549pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
550
551static void *uffd_read_thread(void *arg)
552{
553 unsigned long *this_cpu_userfaults;
554 struct uffd_msg msg;
555
556 this_cpu_userfaults = (unsigned long *) arg;
557 *this_cpu_userfaults = 0;
558
559 pthread_mutex_unlock(&uffd_read_mutex);
560 /* from here cancellation is ok */
561
562 for (;;) {
563 if (uffd_read_msg(uffd, &msg))
564 continue;
565 (*this_cpu_userfaults) += uffd_handle_page_fault(&msg);
566 }
567 return (void *)NULL;
568}
569
570static void *background_thread(void *arg)
571{
572 unsigned long cpu = (unsigned long) arg;
573 unsigned long page_nr;
574
575 for (page_nr = cpu * nr_pages_per_cpu;
576 page_nr < (cpu+1) * nr_pages_per_cpu;
577 page_nr++)
578 copy_page_retry(uffd, page_nr * page_size);
579
580 return NULL;
581}
582
583static int stress(unsigned long *userfaults)
584{
585 unsigned long cpu;
586 pthread_t locking_threads[nr_cpus];
587 pthread_t uffd_threads[nr_cpus];
588 pthread_t background_threads[nr_cpus];
589 void **_userfaults = (void **) userfaults;
590
591 finished = 0;
592 for (cpu = 0; cpu < nr_cpus; cpu++) {
593 if (pthread_create(&locking_threads[cpu], &attr,
594 locking_thread, (void *)cpu))
595 return 1;
596 if (bounces & BOUNCE_POLL) {
597 if (pthread_create(&uffd_threads[cpu], &attr,
598 uffd_poll_thread, (void *)cpu))
599 return 1;
600 } else {
601 if (pthread_create(&uffd_threads[cpu], &attr,
602 uffd_read_thread,
603 &_userfaults[cpu]))
604 return 1;
605 pthread_mutex_lock(&uffd_read_mutex);
606 }
607 if (pthread_create(&background_threads[cpu], &attr,
608 background_thread, (void *)cpu))
609 return 1;
610 }
611 for (cpu = 0; cpu < nr_cpus; cpu++)
612 if (pthread_join(background_threads[cpu], NULL))
613 return 1;
614
615 /*
616 * Be strict and immediately zap area_src, the whole area has
617 * been transferred already by the background treads. The
618 * area_src could then be faulted in in a racy way by still
619 * running uffdio_threads reading zeropages after we zapped
620 * area_src (but they're guaranteed to get -EEXIST from
621 * UFFDIO_COPY without writing zero pages into area_dst
622 * because the background threads already completed).
623 */
624 if (uffd_test_ops->release_pages(area_src))
625 return 1;
626
627
628 finished = 1;
629 for (cpu = 0; cpu < nr_cpus; cpu++)
630 if (pthread_join(locking_threads[cpu], NULL))
631 return 1;
632
633 for (cpu = 0; cpu < nr_cpus; cpu++) {
634 char c;
635 if (bounces & BOUNCE_POLL) {
636 if (write(pipefd[cpu*2+1], &c, 1) != 1) {
637 fprintf(stderr, "pipefd write error\n");
638 return 1;
639 }
640 if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
641 return 1;
642 } else {
643 if (pthread_cancel(uffd_threads[cpu]))
644 return 1;
645 if (pthread_join(uffd_threads[cpu], NULL))
646 return 1;
647 }
648 }
649
650 return 0;
651}
652
653static int userfaultfd_open(int features)
654{
655 struct uffdio_api uffdio_api;
656
657 uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
658 if (uffd < 0) {
659 fprintf(stderr,
660 "userfaultfd syscall not available in this kernel\n");
661 return 1;
662 }
663 uffd_flags = fcntl(uffd, F_GETFD, NULL);
664
665 uffdio_api.api = UFFD_API;
666 uffdio_api.features = features;
667 if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
668 fprintf(stderr, "UFFDIO_API\n");
669 return 1;
670 }
671 if (uffdio_api.api != UFFD_API) {
672 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
673 return 1;
674 }
675
676 return 0;
677}
678
679sigjmp_buf jbuf, *sigbuf;
680
681static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
682{
683 if (sig == SIGBUS) {
684 if (sigbuf)
685 siglongjmp(*sigbuf, 1);
686 abort();
687 }
688}
689
690/*
691 * For non-cooperative userfaultfd test we fork() a process that will
692 * generate pagefaults, will mremap the area monitored by the
693 * userfaultfd and at last this process will release the monitored
694 * area.
695 * For the anonymous and shared memory the area is divided into two
696 * parts, the first part is accessed before mremap, and the second
697 * part is accessed after mremap. Since hugetlbfs does not support
698 * mremap, the entire monitored area is accessed in a single pass for
699 * HUGETLB_TEST.
700 * The release of the pages currently generates event for shmem and
701 * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
702 * for hugetlb.
703 * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
704 * monitored area, generate pagefaults and test that signal is delivered.
705 * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
706 * test robustness use case - we release monitored area, fork a process
707 * that will generate pagefaults and verify signal is generated.
708 * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
709 * feature. Using monitor thread, verify no userfault events are generated.
710 */
711static int faulting_process(int signal_test)
712{
713 unsigned long nr;
714 unsigned long long count;
715 unsigned long split_nr_pages;
716 unsigned long lastnr;
717 struct sigaction act;
718 unsigned long signalled = 0;
719
720 if (test_type != TEST_HUGETLB)
721 split_nr_pages = (nr_pages + 1) / 2;
722 else
723 split_nr_pages = nr_pages;
724
725 if (signal_test) {
726 sigbuf = &jbuf;
727 memset(&act, 0, sizeof(act));
728 act.sa_sigaction = sighndl;
729 act.sa_flags = SA_SIGINFO;
730 if (sigaction(SIGBUS, &act, 0)) {
731 perror("sigaction");
732 return 1;
733 }
734 lastnr = (unsigned long)-1;
735 }
736
737 for (nr = 0; nr < split_nr_pages; nr++) {
738 if (signal_test) {
739 if (sigsetjmp(*sigbuf, 1) != 0) {
740 if (nr == lastnr) {
741 fprintf(stderr, "Signal repeated\n");
742 return 1;
743 }
744
745 lastnr = nr;
746 if (signal_test == 1) {
747 if (copy_page(uffd, nr * page_size))
748 signalled++;
749 } else {
750 signalled++;
751 continue;
752 }
753 }
754 }
755
756 count = *area_count(area_dst, nr);
757 if (count != count_verify[nr]) {
758 fprintf(stderr,
759 "nr %lu memory corruption %Lu %Lu\n",
760 nr, count,
761 count_verify[nr]), exit(1);
762 }
763 }
764
765 if (signal_test)
766 return signalled != split_nr_pages;
767
768 if (test_type == TEST_HUGETLB)
769 return 0;
770
771 area_dst = mremap(area_dst, nr_pages * page_size, nr_pages * page_size,
772 MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
773 if (area_dst == MAP_FAILED)
774 perror("mremap"), exit(1);
775
776 for (; nr < nr_pages; nr++) {
777 count = *area_count(area_dst, nr);
778 if (count != count_verify[nr]) {
779 fprintf(stderr,
780 "nr %lu memory corruption %Lu %Lu\n",
781 nr, count,
782 count_verify[nr]), exit(1);
783 }
784 }
785
786 if (uffd_test_ops->release_pages(area_dst))
787 return 1;
788
789 for (nr = 0; nr < nr_pages; nr++) {
790 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
791 fprintf(stderr, "nr %lu is not zero\n", nr), exit(1);
792 }
793
794 return 0;
795}
796
797static void retry_uffdio_zeropage(int ufd,
798 struct uffdio_zeropage *uffdio_zeropage,
799 unsigned long offset)
800{
801 uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
802 uffdio_zeropage->range.len,
803 offset);
804 if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
805 if (uffdio_zeropage->zeropage != -EEXIST)
806 fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
807 uffdio_zeropage->zeropage), exit(1);
808 } else {
809 fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
810 uffdio_zeropage->zeropage), exit(1);
811 }
812}
813
814static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
815{
816 struct uffdio_zeropage uffdio_zeropage;
817 int ret;
818 unsigned long has_zeropage;
819
820 has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
821
822 if (offset >= nr_pages * page_size)
823 fprintf(stderr, "unexpected offset %lu\n",
824 offset), exit(1);
825 uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
826 uffdio_zeropage.range.len = page_size;
827 uffdio_zeropage.mode = 0;
828 ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
829 if (ret) {
830 /* real retval in ufdio_zeropage.zeropage */
831 if (has_zeropage) {
832 if (uffdio_zeropage.zeropage == -EEXIST)
833 fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n"),
834 exit(1);
835 else
836 fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
837 uffdio_zeropage.zeropage), exit(1);
838 } else {
839 if (uffdio_zeropage.zeropage != -EINVAL)
840 fprintf(stderr,
841 "UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
842 uffdio_zeropage.zeropage), exit(1);
843 }
844 } else if (has_zeropage) {
845 if (uffdio_zeropage.zeropage != page_size) {
846 fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
847 uffdio_zeropage.zeropage), exit(1);
848 } else {
849 if (test_uffdio_zeropage_eexist && retry) {
850 test_uffdio_zeropage_eexist = false;
851 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
852 offset);
853 }
854 return 1;
855 }
856 } else {
857 fprintf(stderr,
858 "UFFDIO_ZEROPAGE succeeded %Ld\n",
859 uffdio_zeropage.zeropage), exit(1);
860 }
861
862 return 0;
863}
864
865static int uffdio_zeropage(int ufd, unsigned long offset)
866{
867 return __uffdio_zeropage(ufd, offset, false);
868}
869
870/* exercise UFFDIO_ZEROPAGE */
871static int userfaultfd_zeropage_test(void)
872{
873 struct uffdio_register uffdio_register;
874 unsigned long expected_ioctls;
875
876 printf("testing UFFDIO_ZEROPAGE: ");
877 fflush(stdout);
878
879 if (uffd_test_ops->release_pages(area_dst))
880 return 1;
881
882 if (userfaultfd_open(0) < 0)
883 return 1;
884 uffdio_register.range.start = (unsigned long) area_dst;
885 uffdio_register.range.len = nr_pages * page_size;
886 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
887 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
888 fprintf(stderr, "register failure\n"), exit(1);
889
890 expected_ioctls = uffd_test_ops->expected_ioctls;
891 if ((uffdio_register.ioctls & expected_ioctls) !=
892 expected_ioctls)
893 fprintf(stderr,
894 "unexpected missing ioctl for anon memory\n"),
895 exit(1);
896
897 if (uffdio_zeropage(uffd, 0)) {
898 if (my_bcmp(area_dst, zeropage, page_size))
899 fprintf(stderr, "zeropage is not zero\n"), exit(1);
900 }
901
902 close(uffd);
903 printf("done.\n");
904 return 0;
905}
906
907static int userfaultfd_events_test(void)
908{
909 struct uffdio_register uffdio_register;
910 unsigned long expected_ioctls;
911 unsigned long userfaults;
912 pthread_t uffd_mon;
913 int err, features;
914 pid_t pid;
915 char c;
916
917 printf("testing events (fork, remap, remove): ");
918 fflush(stdout);
919
920 if (uffd_test_ops->release_pages(area_dst))
921 return 1;
922
923 features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
924 UFFD_FEATURE_EVENT_REMOVE;
925 if (userfaultfd_open(features) < 0)
926 return 1;
927 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
928
929 uffdio_register.range.start = (unsigned long) area_dst;
930 uffdio_register.range.len = nr_pages * page_size;
931 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
932 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
933 fprintf(stderr, "register failure\n"), exit(1);
934
935 expected_ioctls = uffd_test_ops->expected_ioctls;
936 if ((uffdio_register.ioctls & expected_ioctls) !=
937 expected_ioctls)
938 fprintf(stderr,
939 "unexpected missing ioctl for anon memory\n"),
940 exit(1);
941
942 if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
943 perror("uffd_poll_thread create"), exit(1);
944
945 pid = fork();
946 if (pid < 0)
947 perror("fork"), exit(1);
948
949 if (!pid)
950 return faulting_process(0);
951
952 waitpid(pid, &err, 0);
953 if (err)
954 fprintf(stderr, "faulting process failed\n"), exit(1);
955
956 if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
957 perror("pipe write"), exit(1);
958 if (pthread_join(uffd_mon, (void **)&userfaults))
959 return 1;
960
961 close(uffd);
962 printf("userfaults: %ld\n", userfaults);
963
964 return userfaults != nr_pages;
965}
966
967static int userfaultfd_sig_test(void)
968{
969 struct uffdio_register uffdio_register;
970 unsigned long expected_ioctls;
971 unsigned long userfaults;
972 pthread_t uffd_mon;
973 int err, features;
974 pid_t pid;
975 char c;
976
977 printf("testing signal delivery: ");
978 fflush(stdout);
979
980 if (uffd_test_ops->release_pages(area_dst))
981 return 1;
982
983 features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
984 if (userfaultfd_open(features) < 0)
985 return 1;
986 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
987
988 uffdio_register.range.start = (unsigned long) area_dst;
989 uffdio_register.range.len = nr_pages * page_size;
990 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
991 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
992 fprintf(stderr, "register failure\n"), exit(1);
993
994 expected_ioctls = uffd_test_ops->expected_ioctls;
995 if ((uffdio_register.ioctls & expected_ioctls) !=
996 expected_ioctls)
997 fprintf(stderr,
998 "unexpected missing ioctl for anon memory\n"),
999 exit(1);
1000
1001 if (faulting_process(1))
1002 fprintf(stderr, "faulting process failed\n"), exit(1);
1003
1004 if (uffd_test_ops->release_pages(area_dst))
1005 return 1;
1006
1007 if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
1008 perror("uffd_poll_thread create"), exit(1);
1009
1010 pid = fork();
1011 if (pid < 0)
1012 perror("fork"), exit(1);
1013
1014 if (!pid)
1015 exit(faulting_process(2));
1016
1017 waitpid(pid, &err, 0);
1018 if (err)
1019 fprintf(stderr, "faulting process failed\n"), exit(1);
1020
1021 if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1022 perror("pipe write"), exit(1);
1023 if (pthread_join(uffd_mon, (void **)&userfaults))
1024 return 1;
1025
1026 printf("done.\n");
1027 if (userfaults)
1028 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1029 userfaults);
1030 close(uffd);
1031 return userfaults != 0;
1032}
1033static int userfaultfd_stress(void)
1034{
1035 void *area;
1036 char *tmp_area;
1037 unsigned long nr;
1038 struct uffdio_register uffdio_register;
1039 unsigned long cpu;
1040 int err;
1041 unsigned long userfaults[nr_cpus];
1042
1043 uffd_test_ops->allocate_area((void **)&area_src);
1044 if (!area_src)
1045 return 1;
1046 uffd_test_ops->allocate_area((void **)&area_dst);
1047 if (!area_dst)
1048 return 1;
1049
1050 if (userfaultfd_open(0) < 0)
1051 return 1;
1052
1053 count_verify = malloc(nr_pages * sizeof(unsigned long long));
1054 if (!count_verify) {
1055 perror("count_verify");
1056 return 1;
1057 }
1058
1059 for (nr = 0; nr < nr_pages; nr++) {
1060 *area_mutex(area_src, nr) = (pthread_mutex_t)
1061 PTHREAD_MUTEX_INITIALIZER;
1062 count_verify[nr] = *area_count(area_src, nr) = 1;
1063 /*
1064 * In the transition between 255 to 256, powerpc will
1065 * read out of order in my_bcmp and see both bytes as
1066 * zero, so leave a placeholder below always non-zero
1067 * after the count, to avoid my_bcmp to trigger false
1068 * positives.
1069 */
1070 *(area_count(area_src, nr) + 1) = 1;
1071 }
1072
1073 pipefd = malloc(sizeof(int) * nr_cpus * 2);
1074 if (!pipefd) {
1075 perror("pipefd");
1076 return 1;
1077 }
1078 for (cpu = 0; cpu < nr_cpus; cpu++) {
1079 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1080 perror("pipe");
1081 return 1;
1082 }
1083 }
1084
1085 if (posix_memalign(&area, page_size, page_size)) {
1086 fprintf(stderr, "out of memory\n");
1087 return 1;
1088 }
1089 zeropage = area;
1090 bzero(zeropage, page_size);
1091
1092 pthread_mutex_lock(&uffd_read_mutex);
1093
1094 pthread_attr_init(&attr);
1095 pthread_attr_setstacksize(&attr, 16*1024*1024);
1096
1097 err = 0;
1098 while (bounces--) {
1099 unsigned long expected_ioctls;
1100
1101 printf("bounces: %d, mode:", bounces);
1102 if (bounces & BOUNCE_RANDOM)
1103 printf(" rnd");
1104 if (bounces & BOUNCE_RACINGFAULTS)
1105 printf(" racing");
1106 if (bounces & BOUNCE_VERIFY)
1107 printf(" ver");
1108 if (bounces & BOUNCE_POLL)
1109 printf(" poll");
1110 printf(", ");
1111 fflush(stdout);
1112
1113 if (bounces & BOUNCE_POLL)
1114 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1115 else
1116 fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1117
1118 /* register */
1119 uffdio_register.range.start = (unsigned long) area_dst;
1120 uffdio_register.range.len = nr_pages * page_size;
1121 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1122 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1123 fprintf(stderr, "register failure\n");
1124 return 1;
1125 }
1126 expected_ioctls = uffd_test_ops->expected_ioctls;
1127 if ((uffdio_register.ioctls & expected_ioctls) !=
1128 expected_ioctls) {
1129 fprintf(stderr,
1130 "unexpected missing ioctl for anon memory\n");
1131 return 1;
1132 }
1133
1134 if (area_dst_alias) {
1135 uffdio_register.range.start = (unsigned long)
1136 area_dst_alias;
1137 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1138 fprintf(stderr, "register failure alias\n");
1139 return 1;
1140 }
1141 }
1142
1143 /*
1144 * The madvise done previously isn't enough: some
1145 * uffd_thread could have read userfaults (one of
1146 * those already resolved by the background thread)
1147 * and it may be in the process of calling
1148 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1149 * area_src and it would map a zero page in it (of
1150 * course such a UFFDIO_COPY is perfectly safe as it'd
1151 * return -EEXIST). The problem comes at the next
1152 * bounce though: that racing UFFDIO_COPY would
1153 * generate zeropages in the area_src, so invalidating
1154 * the previous MADV_DONTNEED. Without this additional
1155 * MADV_DONTNEED those zeropages leftovers in the
1156 * area_src would lead to -EEXIST failure during the
1157 * next bounce, effectively leaving a zeropage in the
1158 * area_dst.
1159 *
1160 * Try to comment this out madvise to see the memory
1161 * corruption being caught pretty quick.
1162 *
1163 * khugepaged is also inhibited to collapse THP after
1164 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1165 * required to MADV_DONTNEED here.
1166 */
1167 if (uffd_test_ops->release_pages(area_dst))
1168 return 1;
1169
1170 /* bounce pass */
1171 if (stress(userfaults))
1172 return 1;
1173
1174 /* unregister */
1175 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1176 fprintf(stderr, "unregister failure\n");
1177 return 1;
1178 }
1179 if (area_dst_alias) {
1180 uffdio_register.range.start = (unsigned long) area_dst;
1181 if (ioctl(uffd, UFFDIO_UNREGISTER,
1182 &uffdio_register.range)) {
1183 fprintf(stderr, "unregister failure alias\n");
1184 return 1;
1185 }
1186 }
1187
1188 /* verification */
1189 if (bounces & BOUNCE_VERIFY) {
1190 for (nr = 0; nr < nr_pages; nr++) {
1191 if (*area_count(area_dst, nr) != count_verify[nr]) {
1192 fprintf(stderr,
1193 "error area_count %Lu %Lu %lu\n",
1194 *area_count(area_src, nr),
1195 count_verify[nr],
1196 nr);
1197 err = 1;
1198 bounces = 0;
1199 }
1200 }
1201 }
1202
1203 /* prepare next bounce */
1204 tmp_area = area_src;
1205 area_src = area_dst;
1206 area_dst = tmp_area;
1207
1208 tmp_area = area_src_alias;
1209 area_src_alias = area_dst_alias;
1210 area_dst_alias = tmp_area;
1211
1212 printf("userfaults:");
1213 for (cpu = 0; cpu < nr_cpus; cpu++)
1214 printf(" %lu", userfaults[cpu]);
1215 printf("\n");
1216 }
1217
1218 if (err)
1219 return err;
1220
1221 close(uffd);
1222 return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1223 || userfaultfd_events_test();
1224}
1225
1226/*
1227 * Copied from mlock2-tests.c
1228 */
1229unsigned long default_huge_page_size(void)
1230{
1231 unsigned long hps = 0;
1232 char *line = NULL;
1233 size_t linelen = 0;
1234 FILE *f = fopen("/proc/meminfo", "r");
1235
1236 if (!f)
1237 return 0;
1238 while (getline(&line, &linelen, f) > 0) {
1239 if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) {
1240 hps <<= 10;
1241 break;
1242 }
1243 }
1244
1245 free(line);
1246 fclose(f);
1247 return hps;
1248}
1249
1250static void set_test_type(const char *type)
1251{
1252 if (!strcmp(type, "anon")) {
1253 test_type = TEST_ANON;
1254 uffd_test_ops = &anon_uffd_test_ops;
1255 } else if (!strcmp(type, "hugetlb")) {
1256 test_type = TEST_HUGETLB;
1257 uffd_test_ops = &hugetlb_uffd_test_ops;
1258 } else if (!strcmp(type, "hugetlb_shared")) {
1259 map_shared = true;
1260 test_type = TEST_HUGETLB;
1261 uffd_test_ops = &hugetlb_uffd_test_ops;
1262 } else if (!strcmp(type, "shmem")) {
1263 map_shared = true;
1264 test_type = TEST_SHMEM;
1265 uffd_test_ops = &shmem_uffd_test_ops;
1266 } else {
1267 fprintf(stderr, "Unknown test type: %s\n", type), exit(1);
1268 }
1269
1270 if (test_type == TEST_HUGETLB)
1271 page_size = default_huge_page_size();
1272 else
1273 page_size = sysconf(_SC_PAGE_SIZE);
1274
1275 if (!page_size)
1276 fprintf(stderr, "Unable to determine page size\n"),
1277 exit(2);
1278 if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1279 > page_size)
1280 fprintf(stderr, "Impossible to run this test\n"), exit(2);
1281}
1282
1283static void sigalrm(int sig)
1284{
1285 if (sig != SIGALRM)
1286 abort();
1287 test_uffdio_copy_eexist = true;
1288 test_uffdio_zeropage_eexist = true;
1289 alarm(ALARM_INTERVAL_SECS);
1290}
1291
1292int main(int argc, char **argv)
1293{
1294 if (argc < 4)
1295 usage();
1296
1297 if (signal(SIGALRM, sigalrm) == SIG_ERR)
1298 fprintf(stderr, "failed to arm SIGALRM"), exit(1);
1299 alarm(ALARM_INTERVAL_SECS);
1300
1301 set_test_type(argv[1]);
1302
1303 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1304 nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1305 nr_cpus;
1306 if (!nr_pages_per_cpu) {
1307 fprintf(stderr, "invalid MiB\n");
1308 usage();
1309 }
1310
1311 bounces = atoi(argv[3]);
1312 if (bounces <= 0) {
1313 fprintf(stderr, "invalid bounces\n");
1314 usage();
1315 }
1316 nr_pages = nr_pages_per_cpu * nr_cpus;
1317
1318 if (test_type == TEST_HUGETLB) {
1319 if (argc < 5)
1320 usage();
1321 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1322 if (huge_fd < 0) {
1323 fprintf(stderr, "Open of %s failed", argv[3]);
1324 perror("open");
1325 exit(1);
1326 }
1327 if (ftruncate(huge_fd, 0)) {
1328 fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1329 perror("ftruncate");
1330 exit(1);
1331 }
1332 }
1333 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1334 nr_pages, nr_pages_per_cpu);
1335 return userfaultfd_stress();
1336}
1337
1338#else /* __NR_userfaultfd */
1339
1340#warning "missing __NR_userfaultfd definition"
1341
1342int main(void)
1343{
1344 printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1345 return KSFT_SKIP;
1346}
1347
1348#endif /* __NR_userfaultfd */