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 "../../../../include/uapi/linux/userfaultfd.h"
70 #define __NR_userfaultfd 323
71 #elif defined(__i386__)
72 #define __NR_userfaultfd 374
73 #elif defined(__powewrpc__)
74 #define __NR_userfaultfd 364
76 #error "missing __NR_userfaultfd definition"
79 static unsigned long nr_cpus
, nr_pages
, nr_pages_per_cpu
, page_size
;
81 #define BOUNCE_RANDOM (1<<0)
82 #define BOUNCE_RACINGFAULTS (1<<1)
83 #define BOUNCE_VERIFY (1<<2)
84 #define BOUNCE_POLL (1<<3)
87 static unsigned long long *count_verify
;
88 static int uffd
, finished
, *pipefd
;
89 static char *area_src
, *area_dst
;
90 static char *zeropage
;
93 /* pthread_mutex_t starts at page offset 0 */
94 #define area_mutex(___area, ___nr) \
95 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
97 * count is placed in the page after pthread_mutex_t naturally aligned
98 * to avoid non alignment faults on non-x86 archs.
100 #define area_count(___area, ___nr) \
101 ((volatile unsigned long long *) ((unsigned long) \
102 ((___area) + (___nr)*page_size + \
103 sizeof(pthread_mutex_t) + \
104 sizeof(unsigned long long) - 1) & \
105 ~(unsigned long)(sizeof(unsigned long long) \
108 static int my_bcmp(char *str1
, char *str2
, size_t n
)
111 for (i
= 0; i
< n
; i
++)
112 if (str1
[i
] != str2
[i
])
117 static void *locking_thread(void *arg
)
119 unsigned long cpu
= (unsigned long) arg
;
120 struct random_data rand
;
121 unsigned long page_nr
= *(&(page_nr
)); /* uninitialized warning */
123 unsigned long long count
;
128 if (bounces
& BOUNCE_RANDOM
) {
129 seed
= (unsigned int) time(NULL
) - bounces
;
130 if (!(bounces
& BOUNCE_RACINGFAULTS
))
132 bzero(&rand
, sizeof(rand
));
133 bzero(&randstate
, sizeof(randstate
));
134 if (initstate_r(seed
, randstate
, sizeof(randstate
), &rand
))
135 fprintf(stderr
, "srandom_r error\n"), exit(1);
138 if (!(bounces
& BOUNCE_RACINGFAULTS
))
139 page_nr
+= cpu
* nr_pages_per_cpu
;
143 if (bounces
& BOUNCE_RANDOM
) {
144 if (random_r(&rand
, &rand_nr
))
145 fprintf(stderr
, "random_r 1 error\n"), exit(1);
147 if (sizeof(page_nr
) > sizeof(rand_nr
)) {
148 if (random_r(&rand
, &rand_nr
))
149 fprintf(stderr
, "random_r 2 error\n"), exit(1);
150 page_nr
|= ((unsigned long) rand_nr
) << 32;
157 if (bounces
& BOUNCE_VERIFY
) {
158 count
= *area_count(area_dst
, page_nr
);
161 "page_nr %lu wrong count %Lu %Lu\n",
163 count_verify
[page_nr
]), exit(1);
167 * We can't use bcmp (or memcmp) because that
168 * returns 0 erroneously if the memory is
169 * changing under it (even if the end of the
170 * page is never changing and always
174 if (!my_bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
177 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
179 count_verify
[page_nr
]), exit(1);
184 /* uncomment the below line to test with mutex */
185 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
186 while (!bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
192 /* uncomment below line to test with mutex */
193 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
196 "page_nr %lu all zero thread %lu %p %lu\n",
197 page_nr
, cpu
, area_dst
+ page_nr
* page_size
,
205 pthread_mutex_lock(area_mutex(area_dst
, page_nr
));
206 count
= *area_count(area_dst
, page_nr
);
207 if (count
!= count_verify
[page_nr
]) {
209 "page_nr %lu memory corruption %Lu %Lu\n",
211 count_verify
[page_nr
]), exit(1);
214 *area_count(area_dst
, page_nr
) = count_verify
[page_nr
] = count
;
215 pthread_mutex_unlock(area_mutex(area_dst
, page_nr
));
217 if (time(NULL
) - start
> 1)
219 "userfault too slow %ld "
220 "possible false positive with overcommit\n",
227 static int copy_page(unsigned long offset
)
229 struct uffdio_copy uffdio_copy
;
231 if (offset
>= nr_pages
* page_size
)
232 fprintf(stderr
, "unexpected offset %lu\n",
234 uffdio_copy
.dst
= (unsigned long) area_dst
+ offset
;
235 uffdio_copy
.src
= (unsigned long) area_src
+ offset
;
236 uffdio_copy
.len
= page_size
;
237 uffdio_copy
.mode
= 0;
238 uffdio_copy
.copy
= 0;
239 if (ioctl(uffd
, UFFDIO_COPY
, &uffdio_copy
)) {
240 /* real retval in ufdio_copy.copy */
241 if (uffdio_copy
.copy
!= -EEXIST
)
242 fprintf(stderr
, "UFFDIO_COPY error %Ld\n",
243 uffdio_copy
.copy
), exit(1);
244 } else if (uffdio_copy
.copy
!= page_size
) {
245 fprintf(stderr
, "UFFDIO_COPY unexpected copy %Ld\n",
246 uffdio_copy
.copy
), exit(1);
252 static void *uffd_poll_thread(void *arg
)
254 unsigned long cpu
= (unsigned long) arg
;
255 struct pollfd pollfd
[2];
258 unsigned long offset
;
260 unsigned long userfaults
= 0;
263 pollfd
[0].events
= POLLIN
;
264 pollfd
[1].fd
= pipefd
[cpu
*2];
265 pollfd
[1].events
= POLLIN
;
268 ret
= poll(pollfd
, 2, -1);
270 fprintf(stderr
, "poll error %d\n", ret
), exit(1);
272 perror("poll"), exit(1);
273 if (pollfd
[1].revents
& POLLIN
) {
274 if (read(pollfd
[1].fd
, &tmp_chr
, 1) != 1)
275 fprintf(stderr
, "read pipefd error\n"),
279 if (!(pollfd
[0].revents
& POLLIN
))
280 fprintf(stderr
, "pollfd[0].revents %d\n",
281 pollfd
[0].revents
), exit(1);
282 ret
= read(uffd
, &msg
, sizeof(msg
));
286 perror("nonblocking read error"), exit(1);
288 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
289 fprintf(stderr
, "unexpected msg event %u\n",
291 if (msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
292 fprintf(stderr
, "unexpected write fault\n"), exit(1);
293 offset
= (char *)msg
.arg
.pagefault
.address
- area_dst
;
294 offset
&= ~(page_size
-1);
295 if (copy_page(offset
))
298 return (void *)userfaults
;
301 pthread_mutex_t uffd_read_mutex
= PTHREAD_MUTEX_INITIALIZER
;
303 static void *uffd_read_thread(void *arg
)
305 unsigned long *this_cpu_userfaults
;
307 unsigned long offset
;
310 this_cpu_userfaults
= (unsigned long *) arg
;
311 *this_cpu_userfaults
= 0;
313 pthread_mutex_unlock(&uffd_read_mutex
);
314 /* from here cancellation is ok */
317 ret
= read(uffd
, &msg
, sizeof(msg
));
318 if (ret
!= sizeof(msg
)) {
320 perror("blocking read error"), exit(1);
322 fprintf(stderr
, "short read\n"), exit(1);
324 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
325 fprintf(stderr
, "unexpected msg event %u\n",
327 if (bounces
& BOUNCE_VERIFY
&&
328 msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
329 fprintf(stderr
, "unexpected write fault\n"), exit(1);
330 offset
= (char *)msg
.arg
.pagefault
.address
- area_dst
;
331 offset
&= ~(page_size
-1);
332 if (copy_page(offset
))
333 (*this_cpu_userfaults
)++;
338 static void *background_thread(void *arg
)
340 unsigned long cpu
= (unsigned long) arg
;
341 unsigned long page_nr
;
343 for (page_nr
= cpu
* nr_pages_per_cpu
;
344 page_nr
< (cpu
+1) * nr_pages_per_cpu
;
346 copy_page(page_nr
* page_size
);
351 static int stress(unsigned long *userfaults
)
354 pthread_t locking_threads
[nr_cpus
];
355 pthread_t uffd_threads
[nr_cpus
];
356 pthread_t background_threads
[nr_cpus
];
357 void **_userfaults
= (void **) userfaults
;
360 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
361 if (pthread_create(&locking_threads
[cpu
], &attr
,
362 locking_thread
, (void *)cpu
))
364 if (bounces
& BOUNCE_POLL
) {
365 if (pthread_create(&uffd_threads
[cpu
], &attr
,
366 uffd_poll_thread
, (void *)cpu
))
369 if (pthread_create(&uffd_threads
[cpu
], &attr
,
373 pthread_mutex_lock(&uffd_read_mutex
);
375 if (pthread_create(&background_threads
[cpu
], &attr
,
376 background_thread
, (void *)cpu
))
379 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
380 if (pthread_join(background_threads
[cpu
], NULL
))
384 * Be strict and immediately zap area_src, the whole area has
385 * been transferred already by the background treads. The
386 * area_src could then be faulted in in a racy way by still
387 * running uffdio_threads reading zeropages after we zapped
388 * area_src (but they're guaranteed to get -EEXIST from
389 * UFFDIO_COPY without writing zero pages into area_dst
390 * because the background threads already completed).
392 if (madvise(area_src
, nr_pages
* page_size
, MADV_DONTNEED
)) {
397 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
399 if (bounces
& BOUNCE_POLL
) {
400 if (write(pipefd
[cpu
*2+1], &c
, 1) != 1) {
401 fprintf(stderr
, "pipefd write error\n");
404 if (pthread_join(uffd_threads
[cpu
], &_userfaults
[cpu
]))
407 if (pthread_cancel(uffd_threads
[cpu
]))
409 if (pthread_join(uffd_threads
[cpu
], NULL
))
415 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
416 if (pthread_join(locking_threads
[cpu
], NULL
))
422 static int userfaultfd_stress(void)
427 struct uffdio_register uffdio_register
;
428 struct uffdio_api uffdio_api
;
431 unsigned long userfaults
[nr_cpus
];
433 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
434 fprintf(stderr
, "out of memory\n");
438 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
439 fprintf(stderr
, "out of memory\n");
444 uffd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
447 "userfaultfd syscall not available in this kernel\n");
450 uffd_flags
= fcntl(uffd
, F_GETFD
, NULL
);
452 uffdio_api
.api
= UFFD_API
;
453 uffdio_api
.features
= 0;
454 if (ioctl(uffd
, UFFDIO_API
, &uffdio_api
)) {
455 fprintf(stderr
, "UFFDIO_API\n");
458 if (uffdio_api
.api
!= UFFD_API
) {
459 fprintf(stderr
, "UFFDIO_API error %Lu\n", uffdio_api
.api
);
463 count_verify
= malloc(nr_pages
* sizeof(unsigned long long));
465 perror("count_verify");
469 for (nr
= 0; nr
< nr_pages
; nr
++) {
470 *area_mutex(area_src
, nr
) = (pthread_mutex_t
)
471 PTHREAD_MUTEX_INITIALIZER
;
472 count_verify
[nr
] = *area_count(area_src
, nr
) = 1;
475 pipefd
= malloc(sizeof(int) * nr_cpus
* 2);
480 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
481 if (pipe2(&pipefd
[cpu
*2], O_CLOEXEC
| O_NONBLOCK
)) {
487 if (posix_memalign(&area
, page_size
, page_size
)) {
488 fprintf(stderr
, "out of memory\n");
492 bzero(zeropage
, page_size
);
494 pthread_mutex_lock(&uffd_read_mutex
);
496 pthread_attr_init(&attr
);
497 pthread_attr_setstacksize(&attr
, 16*1024*1024);
500 unsigned long expected_ioctls
;
502 printf("bounces: %d, mode:", bounces
);
503 if (bounces
& BOUNCE_RANDOM
)
505 if (bounces
& BOUNCE_RACINGFAULTS
)
507 if (bounces
& BOUNCE_VERIFY
)
509 if (bounces
& BOUNCE_POLL
)
514 if (bounces
& BOUNCE_POLL
)
515 fcntl(uffd
, F_SETFL
, uffd_flags
| O_NONBLOCK
);
517 fcntl(uffd
, F_SETFL
, uffd_flags
& ~O_NONBLOCK
);
520 uffdio_register
.range
.start
= (unsigned long) area_dst
;
521 uffdio_register
.range
.len
= nr_pages
* page_size
;
522 uffdio_register
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
523 if (ioctl(uffd
, UFFDIO_REGISTER
, &uffdio_register
)) {
524 fprintf(stderr
, "register failure\n");
527 expected_ioctls
= (1 << _UFFDIO_WAKE
) |
528 (1 << _UFFDIO_COPY
) |
529 (1 << _UFFDIO_ZEROPAGE
);
530 if ((uffdio_register
.ioctls
& expected_ioctls
) !=
533 "unexpected missing ioctl for anon memory\n");
538 * The madvise done previously isn't enough: some
539 * uffd_thread could have read userfaults (one of
540 * those already resolved by the background thread)
541 * and it may be in the process of calling
542 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
543 * area_src and it would map a zero page in it (of
544 * course such a UFFDIO_COPY is perfectly safe as it'd
545 * return -EEXIST). The problem comes at the next
546 * bounce though: that racing UFFDIO_COPY would
547 * generate zeropages in the area_src, so invalidating
548 * the previous MADV_DONTNEED. Without this additional
549 * MADV_DONTNEED those zeropages leftovers in the
550 * area_src would lead to -EEXIST failure during the
551 * next bounce, effectively leaving a zeropage in the
554 * Try to comment this out madvise to see the memory
555 * corruption being caught pretty quick.
557 * khugepaged is also inhibited to collapse THP after
558 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
559 * required to MADV_DONTNEED here.
561 if (madvise(area_dst
, nr_pages
* page_size
, MADV_DONTNEED
)) {
567 if (stress(userfaults
))
571 if (ioctl(uffd
, UFFDIO_UNREGISTER
, &uffdio_register
.range
)) {
572 fprintf(stderr
, "register failure\n");
577 if (bounces
& BOUNCE_VERIFY
) {
578 for (nr
= 0; nr
< nr_pages
; nr
++) {
579 if (my_bcmp(area_dst
,
580 area_dst
+ nr
* page_size
,
581 sizeof(pthread_mutex_t
))) {
583 "error mutex 2 %lu\n",
587 if (*area_count(area_dst
, nr
) != count_verify
[nr
]) {
589 "error area_count %Lu %Lu %lu\n",
590 *area_count(area_src
, nr
),
598 /* prepare next bounce */
603 printf("userfaults:");
604 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
605 printf(" %lu", userfaults
[cpu
]);
612 int main(int argc
, char **argv
)
615 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
616 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
617 page_size
= sysconf(_SC_PAGE_SIZE
);
618 if ((unsigned long) area_count(NULL
, 0) + sizeof(unsigned long long) >
620 fprintf(stderr
, "Impossible to run this test\n"), exit(2);
621 nr_pages_per_cpu
= atol(argv
[1]) * 1024*1024 / page_size
/
623 if (!nr_pages_per_cpu
) {
624 fprintf(stderr
, "invalid MiB\n");
625 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
627 bounces
= atoi(argv
[2]);
629 fprintf(stderr
, "invalid bounces\n");
630 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
632 nr_pages
= nr_pages_per_cpu
* nr_cpus
;
633 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
634 nr_pages
, nr_pages_per_cpu
);
635 return userfaultfd_stress();