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
75 #elif defined(__s390__)
76 #define __NR_userfaultfd 355
78 #error "missing __NR_userfaultfd definition"
81 static unsigned long nr_cpus
, nr_pages
, nr_pages_per_cpu
, page_size
;
83 #define BOUNCE_RANDOM (1<<0)
84 #define BOUNCE_RACINGFAULTS (1<<1)
85 #define BOUNCE_VERIFY (1<<2)
86 #define BOUNCE_POLL (1<<3)
89 static unsigned long long *count_verify
;
90 static int uffd
, finished
, *pipefd
;
91 static char *area_src
, *area_dst
;
92 static char *zeropage
;
95 /* pthread_mutex_t starts at page offset 0 */
96 #define area_mutex(___area, ___nr) \
97 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
99 * count is placed in the page after pthread_mutex_t naturally aligned
100 * to avoid non alignment faults on non-x86 archs.
102 #define area_count(___area, ___nr) \
103 ((volatile unsigned long long *) ((unsigned long) \
104 ((___area) + (___nr)*page_size + \
105 sizeof(pthread_mutex_t) + \
106 sizeof(unsigned long long) - 1) & \
107 ~(unsigned long)(sizeof(unsigned long long) \
110 static int my_bcmp(char *str1
, char *str2
, size_t n
)
113 for (i
= 0; i
< n
; i
++)
114 if (str1
[i
] != str2
[i
])
119 static void *locking_thread(void *arg
)
121 unsigned long cpu
= (unsigned long) arg
;
122 struct random_data rand
;
123 unsigned long page_nr
= *(&(page_nr
)); /* uninitialized warning */
125 unsigned long long count
;
130 if (bounces
& BOUNCE_RANDOM
) {
131 seed
= (unsigned int) time(NULL
) - bounces
;
132 if (!(bounces
& BOUNCE_RACINGFAULTS
))
134 bzero(&rand
, sizeof(rand
));
135 bzero(&randstate
, sizeof(randstate
));
136 if (initstate_r(seed
, randstate
, sizeof(randstate
), &rand
))
137 fprintf(stderr
, "srandom_r error\n"), exit(1);
140 if (!(bounces
& BOUNCE_RACINGFAULTS
))
141 page_nr
+= cpu
* nr_pages_per_cpu
;
145 if (bounces
& BOUNCE_RANDOM
) {
146 if (random_r(&rand
, &rand_nr
))
147 fprintf(stderr
, "random_r 1 error\n"), exit(1);
149 if (sizeof(page_nr
) > sizeof(rand_nr
)) {
150 if (random_r(&rand
, &rand_nr
))
151 fprintf(stderr
, "random_r 2 error\n"), exit(1);
152 page_nr
|= (((unsigned long) rand_nr
) << 16) <<
160 if (bounces
& BOUNCE_VERIFY
) {
161 count
= *area_count(area_dst
, page_nr
);
164 "page_nr %lu wrong count %Lu %Lu\n",
166 count_verify
[page_nr
]), exit(1);
170 * We can't use bcmp (or memcmp) because that
171 * returns 0 erroneously if the memory is
172 * changing under it (even if the end of the
173 * page is never changing and always
177 if (!my_bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
180 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
182 count_verify
[page_nr
]), exit(1);
187 /* uncomment the below line to test with mutex */
188 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
189 while (!bcmp(area_dst
+ page_nr
* page_size
, zeropage
,
195 /* uncomment below line to test with mutex */
196 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
199 "page_nr %lu all zero thread %lu %p %lu\n",
200 page_nr
, cpu
, area_dst
+ page_nr
* page_size
,
208 pthread_mutex_lock(area_mutex(area_dst
, page_nr
));
209 count
= *area_count(area_dst
, page_nr
);
210 if (count
!= count_verify
[page_nr
]) {
212 "page_nr %lu memory corruption %Lu %Lu\n",
214 count_verify
[page_nr
]), exit(1);
217 *area_count(area_dst
, page_nr
) = count_verify
[page_nr
] = count
;
218 pthread_mutex_unlock(area_mutex(area_dst
, page_nr
));
220 if (time(NULL
) - start
> 1)
222 "userfault too slow %ld "
223 "possible false positive with overcommit\n",
230 static int copy_page(unsigned long offset
)
232 struct uffdio_copy uffdio_copy
;
234 if (offset
>= nr_pages
* page_size
)
235 fprintf(stderr
, "unexpected offset %lu\n",
237 uffdio_copy
.dst
= (unsigned long) area_dst
+ offset
;
238 uffdio_copy
.src
= (unsigned long) area_src
+ offset
;
239 uffdio_copy
.len
= page_size
;
240 uffdio_copy
.mode
= 0;
241 uffdio_copy
.copy
= 0;
242 if (ioctl(uffd
, UFFDIO_COPY
, &uffdio_copy
)) {
243 /* real retval in ufdio_copy.copy */
244 if (uffdio_copy
.copy
!= -EEXIST
)
245 fprintf(stderr
, "UFFDIO_COPY error %Ld\n",
246 uffdio_copy
.copy
), exit(1);
247 } else if (uffdio_copy
.copy
!= page_size
) {
248 fprintf(stderr
, "UFFDIO_COPY unexpected copy %Ld\n",
249 uffdio_copy
.copy
), exit(1);
255 static void *uffd_poll_thread(void *arg
)
257 unsigned long cpu
= (unsigned long) arg
;
258 struct pollfd pollfd
[2];
261 unsigned long offset
;
263 unsigned long userfaults
= 0;
266 pollfd
[0].events
= POLLIN
;
267 pollfd
[1].fd
= pipefd
[cpu
*2];
268 pollfd
[1].events
= POLLIN
;
271 ret
= poll(pollfd
, 2, -1);
273 fprintf(stderr
, "poll error %d\n", ret
), exit(1);
275 perror("poll"), exit(1);
276 if (pollfd
[1].revents
& POLLIN
) {
277 if (read(pollfd
[1].fd
, &tmp_chr
, 1) != 1)
278 fprintf(stderr
, "read pipefd error\n"),
282 if (!(pollfd
[0].revents
& POLLIN
))
283 fprintf(stderr
, "pollfd[0].revents %d\n",
284 pollfd
[0].revents
), exit(1);
285 ret
= read(uffd
, &msg
, sizeof(msg
));
289 perror("nonblocking read error"), exit(1);
291 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
292 fprintf(stderr
, "unexpected msg event %u\n",
294 if (msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
295 fprintf(stderr
, "unexpected write fault\n"), exit(1);
296 offset
= (char *)(unsigned long)msg
.arg
.pagefault
.address
-
298 offset
&= ~(page_size
-1);
299 if (copy_page(offset
))
302 return (void *)userfaults
;
305 pthread_mutex_t uffd_read_mutex
= PTHREAD_MUTEX_INITIALIZER
;
307 static void *uffd_read_thread(void *arg
)
309 unsigned long *this_cpu_userfaults
;
311 unsigned long offset
;
314 this_cpu_userfaults
= (unsigned long *) arg
;
315 *this_cpu_userfaults
= 0;
317 pthread_mutex_unlock(&uffd_read_mutex
);
318 /* from here cancellation is ok */
321 ret
= read(uffd
, &msg
, sizeof(msg
));
322 if (ret
!= sizeof(msg
)) {
324 perror("blocking read error"), exit(1);
326 fprintf(stderr
, "short read\n"), exit(1);
328 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
)
329 fprintf(stderr
, "unexpected msg event %u\n",
331 if (bounces
& BOUNCE_VERIFY
&&
332 msg
.arg
.pagefault
.flags
& UFFD_PAGEFAULT_FLAG_WRITE
)
333 fprintf(stderr
, "unexpected write fault\n"), exit(1);
334 offset
= (char *)(unsigned long)msg
.arg
.pagefault
.address
-
336 offset
&= ~(page_size
-1);
337 if (copy_page(offset
))
338 (*this_cpu_userfaults
)++;
343 static void *background_thread(void *arg
)
345 unsigned long cpu
= (unsigned long) arg
;
346 unsigned long page_nr
;
348 for (page_nr
= cpu
* nr_pages_per_cpu
;
349 page_nr
< (cpu
+1) * nr_pages_per_cpu
;
351 copy_page(page_nr
* page_size
);
356 static int stress(unsigned long *userfaults
)
359 pthread_t locking_threads
[nr_cpus
];
360 pthread_t uffd_threads
[nr_cpus
];
361 pthread_t background_threads
[nr_cpus
];
362 void **_userfaults
= (void **) userfaults
;
365 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
366 if (pthread_create(&locking_threads
[cpu
], &attr
,
367 locking_thread
, (void *)cpu
))
369 if (bounces
& BOUNCE_POLL
) {
370 if (pthread_create(&uffd_threads
[cpu
], &attr
,
371 uffd_poll_thread
, (void *)cpu
))
374 if (pthread_create(&uffd_threads
[cpu
], &attr
,
378 pthread_mutex_lock(&uffd_read_mutex
);
380 if (pthread_create(&background_threads
[cpu
], &attr
,
381 background_thread
, (void *)cpu
))
384 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
385 if (pthread_join(background_threads
[cpu
], NULL
))
389 * Be strict and immediately zap area_src, the whole area has
390 * been transferred already by the background treads. The
391 * area_src could then be faulted in in a racy way by still
392 * running uffdio_threads reading zeropages after we zapped
393 * area_src (but they're guaranteed to get -EEXIST from
394 * UFFDIO_COPY without writing zero pages into area_dst
395 * because the background threads already completed).
397 if (madvise(area_src
, nr_pages
* page_size
, MADV_DONTNEED
)) {
402 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
404 if (bounces
& BOUNCE_POLL
) {
405 if (write(pipefd
[cpu
*2+1], &c
, 1) != 1) {
406 fprintf(stderr
, "pipefd write error\n");
409 if (pthread_join(uffd_threads
[cpu
], &_userfaults
[cpu
]))
412 if (pthread_cancel(uffd_threads
[cpu
]))
414 if (pthread_join(uffd_threads
[cpu
], NULL
))
420 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
421 if (pthread_join(locking_threads
[cpu
], NULL
))
427 static int userfaultfd_stress(void)
432 struct uffdio_register uffdio_register
;
433 struct uffdio_api uffdio_api
;
436 unsigned long userfaults
[nr_cpus
];
438 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
439 fprintf(stderr
, "out of memory\n");
443 if (posix_memalign(&area
, page_size
, nr_pages
* page_size
)) {
444 fprintf(stderr
, "out of memory\n");
449 uffd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
452 "userfaultfd syscall not available in this kernel\n");
455 uffd_flags
= fcntl(uffd
, F_GETFD
, NULL
);
457 uffdio_api
.api
= UFFD_API
;
458 uffdio_api
.features
= 0;
459 if (ioctl(uffd
, UFFDIO_API
, &uffdio_api
)) {
460 fprintf(stderr
, "UFFDIO_API\n");
463 if (uffdio_api
.api
!= UFFD_API
) {
464 fprintf(stderr
, "UFFDIO_API error %Lu\n", uffdio_api
.api
);
468 count_verify
= malloc(nr_pages
* sizeof(unsigned long long));
470 perror("count_verify");
474 for (nr
= 0; nr
< nr_pages
; nr
++) {
475 *area_mutex(area_src
, nr
) = (pthread_mutex_t
)
476 PTHREAD_MUTEX_INITIALIZER
;
477 count_verify
[nr
] = *area_count(area_src
, nr
) = 1;
480 pipefd
= malloc(sizeof(int) * nr_cpus
* 2);
485 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
486 if (pipe2(&pipefd
[cpu
*2], O_CLOEXEC
| O_NONBLOCK
)) {
492 if (posix_memalign(&area
, page_size
, page_size
)) {
493 fprintf(stderr
, "out of memory\n");
497 bzero(zeropage
, page_size
);
499 pthread_mutex_lock(&uffd_read_mutex
);
501 pthread_attr_init(&attr
);
502 pthread_attr_setstacksize(&attr
, 16*1024*1024);
505 unsigned long expected_ioctls
;
507 printf("bounces: %d, mode:", bounces
);
508 if (bounces
& BOUNCE_RANDOM
)
510 if (bounces
& BOUNCE_RACINGFAULTS
)
512 if (bounces
& BOUNCE_VERIFY
)
514 if (bounces
& BOUNCE_POLL
)
519 if (bounces
& BOUNCE_POLL
)
520 fcntl(uffd
, F_SETFL
, uffd_flags
| O_NONBLOCK
);
522 fcntl(uffd
, F_SETFL
, uffd_flags
& ~O_NONBLOCK
);
525 uffdio_register
.range
.start
= (unsigned long) area_dst
;
526 uffdio_register
.range
.len
= nr_pages
* page_size
;
527 uffdio_register
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
528 if (ioctl(uffd
, UFFDIO_REGISTER
, &uffdio_register
)) {
529 fprintf(stderr
, "register failure\n");
532 expected_ioctls
= (1 << _UFFDIO_WAKE
) |
533 (1 << _UFFDIO_COPY
) |
534 (1 << _UFFDIO_ZEROPAGE
);
535 if ((uffdio_register
.ioctls
& expected_ioctls
) !=
538 "unexpected missing ioctl for anon memory\n");
543 * The madvise done previously isn't enough: some
544 * uffd_thread could have read userfaults (one of
545 * those already resolved by the background thread)
546 * and it may be in the process of calling
547 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
548 * area_src and it would map a zero page in it (of
549 * course such a UFFDIO_COPY is perfectly safe as it'd
550 * return -EEXIST). The problem comes at the next
551 * bounce though: that racing UFFDIO_COPY would
552 * generate zeropages in the area_src, so invalidating
553 * the previous MADV_DONTNEED. Without this additional
554 * MADV_DONTNEED those zeropages leftovers in the
555 * area_src would lead to -EEXIST failure during the
556 * next bounce, effectively leaving a zeropage in the
559 * Try to comment this out madvise to see the memory
560 * corruption being caught pretty quick.
562 * khugepaged is also inhibited to collapse THP after
563 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
564 * required to MADV_DONTNEED here.
566 if (madvise(area_dst
, nr_pages
* page_size
, MADV_DONTNEED
)) {
572 if (stress(userfaults
))
576 if (ioctl(uffd
, UFFDIO_UNREGISTER
, &uffdio_register
.range
)) {
577 fprintf(stderr
, "register failure\n");
582 if (bounces
& BOUNCE_VERIFY
) {
583 for (nr
= 0; nr
< nr_pages
; nr
++) {
584 if (my_bcmp(area_dst
,
585 area_dst
+ nr
* page_size
,
586 sizeof(pthread_mutex_t
))) {
588 "error mutex 2 %lu\n",
592 if (*area_count(area_dst
, nr
) != count_verify
[nr
]) {
594 "error area_count %Lu %Lu %lu\n",
595 *area_count(area_src
, nr
),
603 /* prepare next bounce */
608 printf("userfaults:");
609 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
610 printf(" %lu", userfaults
[cpu
]);
617 int main(int argc
, char **argv
)
620 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
621 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
622 page_size
= sysconf(_SC_PAGE_SIZE
);
623 if ((unsigned long) area_count(NULL
, 0) + sizeof(unsigned long long) >
625 fprintf(stderr
, "Impossible to run this test\n"), exit(2);
626 nr_pages_per_cpu
= atol(argv
[1]) * 1024*1024 / page_size
/
628 if (!nr_pages_per_cpu
) {
629 fprintf(stderr
, "invalid MiB\n");
630 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
632 bounces
= atoi(argv
[2]);
634 fprintf(stderr
, "invalid bounces\n");
635 fprintf(stderr
, "Usage: <MiB> <bounces>\n"), exit(1);
637 nr_pages
= nr_pages_per_cpu
* nr_cpus
;
638 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
639 nr_pages
, nr_pages_per_cpu
);
640 return userfaultfd_stress();