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
70 #error "missing __NR_userfaultfd definition"
73 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
75 #define BOUNCE_RANDOM (1<<0)
76 #define BOUNCE_RACINGFAULTS (1<<1)
77 #define BOUNCE_VERIFY (1<<2)
78 #define BOUNCE_POLL (1<<3)
81 static unsigned long long *count_verify;
82 static int uffd, finished, *pipefd;
83 static char *area_src, *area_dst;
84 static char *zeropage;
87 /* pthread_mutex_t starts at page offset 0 */
88 #define area_mutex(___area, ___nr) \
89 ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
91 * count is placed in the page after pthread_mutex_t naturally aligned
92 * to avoid non alignment faults on non-x86 archs.
94 #define area_count(___area, ___nr) \
95 ((volatile unsigned long long *) ((unsigned long) \
96 ((___area) + (___nr)*page_size + \
97 sizeof(pthread_mutex_t) + \
98 sizeof(unsigned long long) - 1) & \
99 ~(unsigned long)(sizeof(unsigned long long) \
102 static int my_bcmp(char *str1, char *str2, size_t n)
105 for (i = 0; i < n; i++)
106 if (str1[i] != str2[i])
111 static void *locking_thread(void *arg)
113 unsigned long cpu = (unsigned long) arg;
114 struct random_data rand;
115 unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
117 unsigned long long count;
122 if (bounces & BOUNCE_RANDOM) {
123 seed = (unsigned int) time(NULL) - bounces;
124 if (!(bounces & BOUNCE_RACINGFAULTS))
126 bzero(&rand, sizeof(rand));
127 bzero(&randstate, sizeof(randstate));
128 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
129 fprintf(stderr, "srandom_r error\n"), exit(1);
132 if (!(bounces & BOUNCE_RACINGFAULTS))
133 page_nr += cpu * nr_pages_per_cpu;
137 if (bounces & BOUNCE_RANDOM) {
138 if (random_r(&rand, &rand_nr))
139 fprintf(stderr, "random_r 1 error\n"), exit(1);
141 if (sizeof(page_nr) > sizeof(rand_nr)) {
142 if (random_r(&rand, &rand_nr))
143 fprintf(stderr, "random_r 2 error\n"), exit(1);
144 page_nr |= (((unsigned long) rand_nr) << 16) <<
152 if (bounces & BOUNCE_VERIFY) {
153 count = *area_count(area_dst, page_nr);
156 "page_nr %lu wrong count %Lu %Lu\n",
158 count_verify[page_nr]), exit(1);
162 * We can't use bcmp (or memcmp) because that
163 * returns 0 erroneously if the memory is
164 * changing under it (even if the end of the
165 * page is never changing and always
169 if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
172 "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
174 count_verify[page_nr]), exit(1);
179 /* uncomment the below line to test with mutex */
180 /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
181 while (!bcmp(area_dst + page_nr * page_size, zeropage,
187 /* uncomment below line to test with mutex */
188 /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
191 "page_nr %lu all zero thread %lu %p %lu\n",
192 page_nr, cpu, area_dst + page_nr * page_size,
200 pthread_mutex_lock(area_mutex(area_dst, page_nr));
201 count = *area_count(area_dst, page_nr);
202 if (count != count_verify[page_nr]) {
204 "page_nr %lu memory corruption %Lu %Lu\n",
206 count_verify[page_nr]), exit(1);
209 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
210 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
212 if (time(NULL) - start > 1)
214 "userfault too slow %ld "
215 "possible false positive with overcommit\n",
222 static int copy_page(unsigned long offset)
224 struct uffdio_copy uffdio_copy;
226 if (offset >= nr_pages * page_size)
227 fprintf(stderr, "unexpected offset %lu\n",
229 uffdio_copy.dst = (unsigned long) area_dst + offset;
230 uffdio_copy.src = (unsigned long) area_src + offset;
231 uffdio_copy.len = page_size;
232 uffdio_copy.mode = 0;
233 uffdio_copy.copy = 0;
234 if (ioctl(uffd, UFFDIO_COPY, &uffdio_copy)) {
235 /* real retval in ufdio_copy.copy */
236 if (uffdio_copy.copy != -EEXIST)
237 fprintf(stderr, "UFFDIO_COPY error %Ld\n",
238 uffdio_copy.copy), exit(1);
239 } else if (uffdio_copy.copy != page_size) {
240 fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
241 uffdio_copy.copy), exit(1);
247 static void *uffd_poll_thread(void *arg)
249 unsigned long cpu = (unsigned long) arg;
250 struct pollfd pollfd[2];
253 unsigned long offset;
255 unsigned long userfaults = 0;
258 pollfd[0].events = POLLIN;
259 pollfd[1].fd = pipefd[cpu*2];
260 pollfd[1].events = POLLIN;
263 ret = poll(pollfd, 2, -1);
265 fprintf(stderr, "poll error %d\n", ret), exit(1);
267 perror("poll"), exit(1);
268 if (pollfd[1].revents & POLLIN) {
269 if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
270 fprintf(stderr, "read pipefd error\n"),
274 if (!(pollfd[0].revents & POLLIN))
275 fprintf(stderr, "pollfd[0].revents %d\n",
276 pollfd[0].revents), exit(1);
277 ret = read(uffd, &msg, sizeof(msg));
281 perror("nonblocking read error"), exit(1);
283 if (msg.event != UFFD_EVENT_PAGEFAULT)
284 fprintf(stderr, "unexpected msg event %u\n",
286 if (msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
287 fprintf(stderr, "unexpected write fault\n"), exit(1);
288 offset = (char *)(unsigned long)msg.arg.pagefault.address -
290 offset &= ~(page_size-1);
291 if (copy_page(offset))
294 return (void *)userfaults;
297 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
299 static void *uffd_read_thread(void *arg)
301 unsigned long *this_cpu_userfaults;
303 unsigned long offset;
306 this_cpu_userfaults = (unsigned long *) arg;
307 *this_cpu_userfaults = 0;
309 pthread_mutex_unlock(&uffd_read_mutex);
310 /* from here cancellation is ok */
313 ret = read(uffd, &msg, sizeof(msg));
314 if (ret != sizeof(msg)) {
316 perror("blocking read error"), exit(1);
318 fprintf(stderr, "short read\n"), exit(1);
320 if (msg.event != UFFD_EVENT_PAGEFAULT)
321 fprintf(stderr, "unexpected msg event %u\n",
323 if (bounces & BOUNCE_VERIFY &&
324 msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
325 fprintf(stderr, "unexpected write fault\n"), exit(1);
326 offset = (char *)(unsigned long)msg.arg.pagefault.address -
328 offset &= ~(page_size-1);
329 if (copy_page(offset))
330 (*this_cpu_userfaults)++;
335 static void *background_thread(void *arg)
337 unsigned long cpu = (unsigned long) arg;
338 unsigned long page_nr;
340 for (page_nr = cpu * nr_pages_per_cpu;
341 page_nr < (cpu+1) * nr_pages_per_cpu;
343 copy_page(page_nr * page_size);
348 static int stress(unsigned long *userfaults)
351 pthread_t locking_threads[nr_cpus];
352 pthread_t uffd_threads[nr_cpus];
353 pthread_t background_threads[nr_cpus];
354 void **_userfaults = (void **) userfaults;
357 for (cpu = 0; cpu < nr_cpus; cpu++) {
358 if (pthread_create(&locking_threads[cpu], &attr,
359 locking_thread, (void *)cpu))
361 if (bounces & BOUNCE_POLL) {
362 if (pthread_create(&uffd_threads[cpu], &attr,
363 uffd_poll_thread, (void *)cpu))
366 if (pthread_create(&uffd_threads[cpu], &attr,
370 pthread_mutex_lock(&uffd_read_mutex);
372 if (pthread_create(&background_threads[cpu], &attr,
373 background_thread, (void *)cpu))
376 for (cpu = 0; cpu < nr_cpus; cpu++)
377 if (pthread_join(background_threads[cpu], NULL))
381 * Be strict and immediately zap area_src, the whole area has
382 * been transferred already by the background treads. The
383 * area_src could then be faulted in in a racy way by still
384 * running uffdio_threads reading zeropages after we zapped
385 * area_src (but they're guaranteed to get -EEXIST from
386 * UFFDIO_COPY without writing zero pages into area_dst
387 * because the background threads already completed).
389 if (madvise(area_src, nr_pages * page_size, MADV_DONTNEED)) {
394 for (cpu = 0; cpu < nr_cpus; cpu++) {
396 if (bounces & BOUNCE_POLL) {
397 if (write(pipefd[cpu*2+1], &c, 1) != 1) {
398 fprintf(stderr, "pipefd write error\n");
401 if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
404 if (pthread_cancel(uffd_threads[cpu]))
406 if (pthread_join(uffd_threads[cpu], NULL))
412 for (cpu = 0; cpu < nr_cpus; cpu++)
413 if (pthread_join(locking_threads[cpu], NULL))
419 static int userfaultfd_stress(void)
424 struct uffdio_register uffdio_register;
425 struct uffdio_api uffdio_api;
428 unsigned long userfaults[nr_cpus];
430 if (posix_memalign(&area, page_size, nr_pages * page_size)) {
431 fprintf(stderr, "out of memory\n");
435 if (posix_memalign(&area, page_size, nr_pages * page_size)) {
436 fprintf(stderr, "out of memory\n");
441 uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
444 "userfaultfd syscall not available in this kernel\n");
447 uffd_flags = fcntl(uffd, F_GETFD, NULL);
449 uffdio_api.api = UFFD_API;
450 uffdio_api.features = 0;
451 if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
452 fprintf(stderr, "UFFDIO_API\n");
455 if (uffdio_api.api != UFFD_API) {
456 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
460 count_verify = malloc(nr_pages * sizeof(unsigned long long));
462 perror("count_verify");
466 for (nr = 0; nr < nr_pages; nr++) {
467 *area_mutex(area_src, nr) = (pthread_mutex_t)
468 PTHREAD_MUTEX_INITIALIZER;
469 count_verify[nr] = *area_count(area_src, nr) = 1;
472 pipefd = malloc(sizeof(int) * nr_cpus * 2);
477 for (cpu = 0; cpu < nr_cpus; cpu++) {
478 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
484 if (posix_memalign(&area, page_size, page_size)) {
485 fprintf(stderr, "out of memory\n");
489 bzero(zeropage, page_size);
491 pthread_mutex_lock(&uffd_read_mutex);
493 pthread_attr_init(&attr);
494 pthread_attr_setstacksize(&attr, 16*1024*1024);
497 unsigned long expected_ioctls;
499 printf("bounces: %d, mode:", bounces);
500 if (bounces & BOUNCE_RANDOM)
502 if (bounces & BOUNCE_RACINGFAULTS)
504 if (bounces & BOUNCE_VERIFY)
506 if (bounces & BOUNCE_POLL)
511 if (bounces & BOUNCE_POLL)
512 fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
514 fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
517 uffdio_register.range.start = (unsigned long) area_dst;
518 uffdio_register.range.len = nr_pages * page_size;
519 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
520 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
521 fprintf(stderr, "register failure\n");
524 expected_ioctls = (1 << _UFFDIO_WAKE) |
525 (1 << _UFFDIO_COPY) |
526 (1 << _UFFDIO_ZEROPAGE);
527 if ((uffdio_register.ioctls & expected_ioctls) !=
530 "unexpected missing ioctl for anon memory\n");
535 * The madvise done previously isn't enough: some
536 * uffd_thread could have read userfaults (one of
537 * those already resolved by the background thread)
538 * and it may be in the process of calling
539 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
540 * area_src and it would map a zero page in it (of
541 * course such a UFFDIO_COPY is perfectly safe as it'd
542 * return -EEXIST). The problem comes at the next
543 * bounce though: that racing UFFDIO_COPY would
544 * generate zeropages in the area_src, so invalidating
545 * the previous MADV_DONTNEED. Without this additional
546 * MADV_DONTNEED those zeropages leftovers in the
547 * area_src would lead to -EEXIST failure during the
548 * next bounce, effectively leaving a zeropage in the
551 * Try to comment this out madvise to see the memory
552 * corruption being caught pretty quick.
554 * khugepaged is also inhibited to collapse THP after
555 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
556 * required to MADV_DONTNEED here.
558 if (madvise(area_dst, nr_pages * page_size, MADV_DONTNEED)) {
564 if (stress(userfaults))
568 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
569 fprintf(stderr, "register failure\n");
574 if (bounces & BOUNCE_VERIFY) {
575 for (nr = 0; nr < nr_pages; nr++) {
576 if (my_bcmp(area_dst,
577 area_dst + nr * page_size,
578 sizeof(pthread_mutex_t))) {
580 "error mutex 2 %lu\n",
584 if (*area_count(area_dst, nr) != count_verify[nr]) {
586 "error area_count %Lu %Lu %lu\n",
587 *area_count(area_src, nr),
595 /* prepare next bounce */
600 printf("userfaults:");
601 for (cpu = 0; cpu < nr_cpus; cpu++)
602 printf(" %lu", userfaults[cpu]);
609 int main(int argc, char **argv)
612 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
613 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
614 page_size = sysconf(_SC_PAGE_SIZE);
615 if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) >
617 fprintf(stderr, "Impossible to run this test\n"), exit(2);
618 nr_pages_per_cpu = atol(argv[1]) * 1024*1024 / page_size /
620 if (!nr_pages_per_cpu) {
621 fprintf(stderr, "invalid MiB\n");
622 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
624 bounces = atoi(argv[2]);
626 fprintf(stderr, "invalid bounces\n");
627 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
629 nr_pages = nr_pages_per_cpu * nr_cpus;
630 printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
631 nr_pages, nr_pages_per_cpu);
632 return userfaultfd_stress();