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>
67 #include <linux/userfaultfd.h>
69 #ifndef __NR_userfaultfd
71 #define __NR_userfaultfd 323
72 #elif defined(__i386__)
73 #define __NR_userfaultfd 374
74 #elif defined(__powewrpc__)
75 #define __NR_userfaultfd 364
76 #elif defined(__s390__)
77 #define __NR_userfaultfd 355
79 #error "missing __NR_userfaultfd definition"
83 static unsigned long nr_cpus
, nr_pages
, nr_pages_per_cpu
, page_size
;
85 #define BOUNCE_RANDOM (1<<0)
86 #define BOUNCE_RACINGFAULTS (1<<1)
87 #define BOUNCE_VERIFY (1<<2)
88 #define BOUNCE_POLL (1<<3)
91 static unsigned long long *count_verify
;
92 static int uffd
, finished
, *pipefd
;
93 static char *area_src
, *area_dst
;
94 static char *zeropage
;
97 /* pthread_mutex_t starts at page offset 0 */
98 #define area_mutex(___area, ___nr) \
99 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
101 * count is placed in the page after pthread_mutex_t naturally aligned
102 * to avoid non alignment faults on non-x86 archs.
104 #define area_count(___area, ___nr) \
105 ((volatile unsigned long long *) ((unsigned long) \
106 ((___area) + (___nr)*page_size + \
107 sizeof(pthread_mutex_t) + \
108 sizeof(unsigned long long) - 1) & \
109 ~(unsigned long)(sizeof(unsigned long long) \
112 static int my_bcmp(char *str1
, char *str2
, size_t n
)
115 for (i
= 0; i
< n
; i
++)
116 if (str1
[i
] != str2
[i
])
121 static void *locking_thread(void *arg
)
123 unsigned long cpu
= (unsigned long) arg
;
124 struct random_data rand
;
125 unsigned long page_nr
= *(&(page_nr
)); /* uninitialized warning */
127 unsigned long long count
;
132 if (bounces
& BOUNCE_RANDOM
) {
133 seed
= (unsigned int) time(NULL
) - bounces
;
134 if (!(bounces
& BOUNCE_RACINGFAULTS
))
136 bzero(&rand
, sizeof(rand
));
137 bzero(&randstate
, sizeof(randstate
));
138 if (initstate_r(seed
, randstate
, sizeof(randstate
), &rand
))
139 fprintf(stderr
, "srandom_r error\n"), exit(1);
142 if (!(bounces
& BOUNCE_RACINGFAULTS
))
143 page_nr
+= cpu
* nr_pages_per_cpu
;
147 if (bounces
& BOUNCE_RANDOM
) {
148 if (random_r(&rand
, &rand_nr
))
149 fprintf(stderr
, "random_r 1 error\n"), exit(1);
151 if (sizeof(page_nr
) > sizeof(rand_nr
)) {
152 if (random_r(&rand
, &rand_nr
))
153 fprintf(stderr
, "random_r 2 error\n"), exit(1);
154 page_nr
|= (((unsigned long) rand_nr
) << 16) <<
162 if (bounces
& BOUNCE_VERIFY
) {
163 count
= *area_count(area_dst
, page_nr
);
166 "page_nr %lu wrong count %Lu %Lu\n",
168 count_verify
[page_nr
]), exit(1);
172 * We can't use bcmp (or memcmp) because that
173 * returns 0 erroneously if the memory is
174 * changing under it (even if the end of the
175 * page is never changing and always
179 if (!my_bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
182 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
184 count_verify
[page_nr
]), exit(1);
189 /* uncomment the below line to test with mutex */
190 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
191 while (!bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
197 /* uncomment below line to test with mutex */
198 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
201 "page_nr %lu all zero thread %lu %p %lu\n",
202 page_nr
, cpu
, area_dst
+ page_nr
* page_size
,
210 pthread_mutex_lock(area_mutex(area_dst
, page_nr
));
211 count
= *area_count(area_dst
, page_nr
);
212 if (count
!= count_verify
[page_nr
]) {
214 "page_nr %lu memory corruption %Lu %Lu\n",
216 count_verify
[page_nr
]), exit(1);
219 *area_count(area_dst
, page_nr
) = count_verify
[page_nr
] = count
;
220 pthread_mutex_unlock(area_mutex(area_dst
, page_nr
));
222 if (time(NULL
) - start
> 1)
224 "userfault too slow %ld "
225 "possible false positive with overcommit\n",
232 static int copy_page(unsigned long offset
)
234 struct uffdio_copy uffdio_copy
;
236 if (offset
>= nr_pages
* page_size
)
237 fprintf(stderr
, "unexpected offset %lu\n",
239 uffdio_copy
.dst
= (unsigned long) area_dst
+ offset
;
240 uffdio_copy
.src
= (unsigned long) area_src
+ offset
;
241 uffdio_copy
.len
= page_size
;
242 uffdio_copy
.mode
= 0;
243 uffdio_copy
.copy
= 0;
244 if (ioctl(uffd
, UFFDIO_COPY
, &uffdio_copy
)) {
245 /* real retval in ufdio_copy.copy */
246 if (uffdio_copy
.copy
!= -EEXIST
)
247 fprintf(stderr
, "UFFDIO_COPY error %Ld\n",
248 uffdio_copy
.copy
), exit(1);
249 } else if (uffdio_copy
.copy
!= page_size
) {
250 fprintf(stderr
, "UFFDIO_COPY unexpected copy %Ld\n",
251 uffdio_copy
.copy
), exit(1);
257 static void *uffd_poll_thread(void *arg
)
259 unsigned long cpu
= (unsigned long) arg
;
260 struct pollfd pollfd
[2];
263 unsigned long offset
;
265 unsigned long userfaults
= 0;
268 pollfd
[0].events
= POLLIN
;
269 pollfd
[1].fd
= pipefd
[cpu
*2];
270 pollfd
[1].events
= POLLIN
;
273 ret
= poll(pollfd
, 2, -1);
275 fprintf(stderr
, "poll error %d\n", ret
), exit(1);
277 perror("poll"), exit(1);
278 if (pollfd
[1].revents
& POLLIN
) {
279 if (read(pollfd
[1].fd
, &tmp_chr
, 1) != 1)
280 fprintf(stderr
, "read pipefd error\n"),
284 if (!(pollfd
[0].revents
& POLLIN
))
285 fprintf(stderr
, "pollfd[0].revents %d\n",
286 pollfd
[0].revents
), exit(1);
287 ret
= read(uffd
, &msg
, sizeof(msg
));
291 perror("nonblocking read error"), exit(1);
293 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
294 fprintf(stderr
, "unexpected msg event %u\n",
296 if (msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
297 fprintf(stderr
, "unexpected write fault\n"), exit(1);
298 offset
= (char *)(unsigned long)msg
.arg
.pagefault
.address
-
300 offset
&= ~(page_size
-1);
301 if (copy_page(offset
))
304 return (void *)userfaults
;
307 pthread_mutex_t uffd_read_mutex
= PTHREAD_MUTEX_INITIALIZER
;
309 static void *uffd_read_thread(void *arg
)
311 unsigned long *this_cpu_userfaults
;
313 unsigned long offset
;
316 this_cpu_userfaults
= (unsigned long *) arg
;
317 *this_cpu_userfaults
= 0;
319 pthread_mutex_unlock(&uffd_read_mutex
);
320 /* from here cancellation is ok */
323 ret
= read(uffd
, &msg
, sizeof(msg
));
324 if (ret
!= sizeof(msg
)) {
326 perror("blocking read error"), exit(1);
328 fprintf(stderr
, "short read\n"), exit(1);
330 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
331 fprintf(stderr
, "unexpected msg event %u\n",
333 if (bounces
& BOUNCE_VERIFY
&&
334 msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
335 fprintf(stderr
, "unexpected write fault\n"), exit(1);
336 offset
= (char *)(unsigned long)msg
.arg
.pagefault
.address
-
338 offset
&= ~(page_size
-1);
339 if (copy_page(offset
))
340 (*this_cpu_userfaults
)++;
345 static void *background_thread(void *arg
)
347 unsigned long cpu
= (unsigned long) arg
;
348 unsigned long page_nr
;
350 for (page_nr
= cpu
* nr_pages_per_cpu
;
351 page_nr
< (cpu
+1) * nr_pages_per_cpu
;
353 copy_page(page_nr
* page_size
);
358 static int stress(unsigned long *userfaults
)
361 pthread_t locking_threads
[nr_cpus
];
362 pthread_t uffd_threads
[nr_cpus
];
363 pthread_t background_threads
[nr_cpus
];
364 void **_userfaults
= (void **) userfaults
;
367 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
368 if (pthread_create(&locking_threads
[cpu
], &attr
,
369 locking_thread
, (void *)cpu
))
371 if (bounces
& BOUNCE_POLL
) {
372 if (pthread_create(&uffd_threads
[cpu
], &attr
,
373 uffd_poll_thread
, (void *)cpu
))
376 if (pthread_create(&uffd_threads
[cpu
], &attr
,
380 pthread_mutex_lock(&uffd_read_mutex
);
382 if (pthread_create(&background_threads
[cpu
], &attr
,
383 background_thread
, (void *)cpu
))
386 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
387 if (pthread_join(background_threads
[cpu
], NULL
))
391 * Be strict and immediately zap area_src, the whole area has
392 * been transferred already by the background treads. The
393 * area_src could then be faulted in in a racy way by still
394 * running uffdio_threads reading zeropages after we zapped
395 * area_src (but they're guaranteed to get -EEXIST from
396 * UFFDIO_COPY without writing zero pages into area_dst
397 * because the background threads already completed).
399 if (madvise(area_src
, nr_pages
* page_size
, MADV_DONTNEED
)) {
404 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
406 if (bounces
& BOUNCE_POLL
) {
407 if (write(pipefd
[cpu
*2+1], &c
, 1) != 1) {
408 fprintf(stderr
, "pipefd write error\n");
411 if (pthread_join(uffd_threads
[cpu
], &_userfaults
[cpu
]))
414 if (pthread_cancel(uffd_threads
[cpu
]))
416 if (pthread_join(uffd_threads
[cpu
], NULL
))
422 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
423 if (pthread_join(locking_threads
[cpu
], NULL
))
429 static int userfaultfd_stress(void)
434 struct uffdio_register uffdio_register
;
435 struct uffdio_api uffdio_api
;
438 unsigned long userfaults
[nr_cpus
];
440 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
441 fprintf(stderr
, "out of memory\n");
445 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
446 fprintf(stderr
, "out of memory\n");
451 uffd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
454 "userfaultfd syscall not available in this kernel\n");
457 uffd_flags
= fcntl(uffd
, F_GETFD
, NULL
);
459 uffdio_api
.api
= UFFD_API
;
460 uffdio_api
.features
= 0;
461 if (ioctl(uffd
, UFFDIO_API
, &uffdio_api
)) {
462 fprintf(stderr
, "UFFDIO_API\n");
465 if (uffdio_api
.api
!= UFFD_API
) {
466 fprintf(stderr
, "UFFDIO_API error %Lu\n", uffdio_api
.api
);
470 count_verify
= malloc(nr_pages
* sizeof(unsigned long long));
472 perror("count_verify");
476 for (nr
= 0; nr
< nr_pages
; nr
++) {
477 *area_mutex(area_src
, nr
) = (pthread_mutex_t
)
478 PTHREAD_MUTEX_INITIALIZER
;
479 count_verify
[nr
] = *area_count(area_src
, nr
) = 1;
482 pipefd
= malloc(sizeof(int) * nr_cpus
* 2);
487 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
488 if (pipe2(&pipefd
[cpu
*2], O_CLOEXEC
| O_NONBLOCK
)) {
494 if (posix_memalign(&area
, page_size
, page_size
)) {
495 fprintf(stderr
, "out of memory\n");
499 bzero(zeropage
, page_size
);
501 pthread_mutex_lock(&uffd_read_mutex
);
503 pthread_attr_init(&attr
);
504 pthread_attr_setstacksize(&attr
, 16*1024*1024);
507 unsigned long expected_ioctls
;
509 printf("bounces: %d, mode:", bounces
);
510 if (bounces
& BOUNCE_RANDOM
)
512 if (bounces
& BOUNCE_RACINGFAULTS
)
514 if (bounces
& BOUNCE_VERIFY
)
516 if (bounces
& BOUNCE_POLL
)
521 if (bounces
& BOUNCE_POLL
)
522 fcntl(uffd
, F_SETFL
, uffd_flags
| O_NONBLOCK
);
524 fcntl(uffd
, F_SETFL
, uffd_flags
& ~O_NONBLOCK
);
527 uffdio_register
.range
.start
= (unsigned long) area_dst
;
528 uffdio_register
.range
.len
= nr_pages
* page_size
;
529 uffdio_register
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
530 if (ioctl(uffd
, UFFDIO_REGISTER
, &uffdio_register
)) {
531 fprintf(stderr
, "register failure\n");
534 expected_ioctls
= (1 << _UFFDIO_WAKE
) |
535 (1 << _UFFDIO_COPY
) |
536 (1 << _UFFDIO_ZEROPAGE
);
537 if ((uffdio_register
.ioctls
& expected_ioctls
) !=
540 "unexpected missing ioctl for anon memory\n");
545 * The madvise done previously isn't enough: some
546 * uffd_thread could have read userfaults (one of
547 * those already resolved by the background thread)
548 * and it may be in the process of calling
549 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
550 * area_src and it would map a zero page in it (of
551 * course such a UFFDIO_COPY is perfectly safe as it'd
552 * return -EEXIST). The problem comes at the next
553 * bounce though: that racing UFFDIO_COPY would
554 * generate zeropages in the area_src, so invalidating
555 * the previous MADV_DONTNEED. Without this additional
556 * MADV_DONTNEED those zeropages leftovers in the
557 * area_src would lead to -EEXIST failure during the
558 * next bounce, effectively leaving a zeropage in the
561 * Try to comment this out madvise to see the memory
562 * corruption being caught pretty quick.
564 * khugepaged is also inhibited to collapse THP after
565 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
566 * required to MADV_DONTNEED here.
568 if (madvise(area_dst
, nr_pages
* page_size
, MADV_DONTNEED
)) {
574 if (stress(userfaults
))
578 if (ioctl(uffd
, UFFDIO_UNREGISTER
, &uffdio_register
.range
)) {
579 fprintf(stderr
, "register failure\n");
584 if (bounces
& BOUNCE_VERIFY
) {
585 for (nr
= 0; nr
< nr_pages
; nr
++) {
586 if (my_bcmp(area_dst
,
587 area_dst
+ nr
* page_size
,
588 sizeof(pthread_mutex_t
))) {
590 "error mutex 2 %lu\n",
594 if (*area_count(area_dst
, nr
) != count_verify
[nr
]) {
596 "error area_count %Lu %Lu %lu\n",
597 *area_count(area_src
, nr
),
605 /* prepare next bounce */
610 printf("userfaults:");
611 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
612 printf(" %lu", userfaults
[cpu
]);
619 int main(int argc
, char **argv
)
622 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
623 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
624 page_size
= sysconf(_SC_PAGE_SIZE
);
625 if ((unsigned long) area_count(NULL
, 0) + sizeof(unsigned long long) >
627 fprintf(stderr
, "Impossible to run this test\n"), exit(2);
628 nr_pages_per_cpu
= atol(argv
[1]) * 1024*1024 / page_size
/
630 if (!nr_pages_per_cpu
) {
631 fprintf(stderr
, "invalid MiB\n");
632 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
634 bounces
= atoi(argv
[2]);
636 fprintf(stderr
, "invalid bounces\n");
637 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
639 nr_pages
= nr_pages_per_cpu
* nr_cpus
;
640 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
641 nr_pages
, nr_pages_per_cpu
);
642 return userfaultfd_stress();