tools/testing/selftests/vm/userfaultfd.c

   1 /*
   2  * Stress userfaultfd syscall.
   3  *
   4  *  Copyright (C) 2015  Red Hat, Inc.
   5  *
   6  *  This work is licensed under the terms of the GNU GPL, version 2. See
   7  *  the COPYING file in the top-level directory.
   8  *
   9  * This test allocates two virtual areas and bounces the physical
  10  * memory across the two virtual areas (from area_src to area_dst)
  11  * using userfaultfd.
  12  *
  13  * There are three threads running per CPU:
  14  *
  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.
  19  *
  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.
  23  *
  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
  27  *    area.
  28  *
  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
  32  * opposite direction.
  33  *
  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).
  37  *
  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.
  40  *
  41  * # 100MiB 99999 bounces
  42  * ./userfaultfd 100 99999
  43  *
  44  * # 1GiB 99 bounces
  45  * ./userfaultfd 1000 99
  46  *
  47  * # 10MiB-~6GiB 999 bounces, continue forever unless an error triggers
  48  * while ./userfaultfd $[RANDOM % 6000 + 10] 999; do true; done
  49  */
  50
  51 #define _GNU_SOURCE
  52 #include <stdio.h>
  53 #include <errno.h>
  54 #include <unistd.h>
  55 #include <stdlib.h>
  56 #include <sys/types.h>
  57 #include <sys/stat.h>
  58 #include <fcntl.h>
  59 #include <time.h>
  60 #include <signal.h>
  61 #include <poll.h>
  62 #include <string.h>
  63 #include <sys/mman.h>
  64 #include <sys/syscall.h>
  65 #include <sys/ioctl.h>
  66 #include <sys/wait.h>
  67 #include <pthread.h>
  68 #include <linux/userfaultfd.h>
  69 #include <setjmp.h>
  70 #include <stdbool.h>
  71
  72 #ifdef __NR_userfaultfd
  73
  74 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
  75
  76 #define BOUNCE_RANDOM           (1<<0)
  77 #define BOUNCE_RACINGFAULTS     (1<<1)
  78 #define BOUNCE_VERIFY           (1<<2)
  79 #define BOUNCE_POLL             (1<<3)
  80 static int bounces;
  81
  82 #define TEST_ANON       1
  83 #define TEST_HUGETLB    2
  84 #define TEST_SHMEM      3
  85 static int test_type;
  86
  87 /* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
  88 #define ALARM_INTERVAL_SECS 10
  89 static volatile bool test_uffdio_copy_eexist = true;
  90 static volatile bool test_uffdio_zeropage_eexist = true;
  91
  92 static bool map_shared;
  93 static int huge_fd;
  94 static char *huge_fd_off0;
  95 static unsigned long long *count_verify;
  96 static int uffd, uffd_flags, finished, *pipefd;
  97 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
  98 static char *zeropage;
  99 pthread_attr_t attr;
 100
 101 /* pthread_mutex_t starts at page offset 0 */
 102 #define area_mutex(___area, ___nr)                                      \
 103         ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
 104 /*
 105  * count is placed in the page after pthread_mutex_t naturally aligned
 106  * to avoid non alignment faults on non-x86 archs.
 107  */
 108 #define area_count(___area, ___nr)                                      \
 109         ((volatile unsigned long long *) ((unsigned long)               \
 110                                  ((___area) + (___nr)*page_size +       \
 111                                   sizeof(pthread_mutex_t) +             \
 112                                   sizeof(unsigned long long) - 1) &     \
 113                                  ~(unsigned long)(sizeof(unsigned long long) \
 114                                                   -  1)))
 115
 116 static int anon_release_pages(char *rel_area)
 117 {
 118         int ret = 0;
 119
 120         if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
 121                 perror("madvise");
 122                 ret = 1;
 123         }
 124
 125         return ret;
 126 }
 127
 128 static void anon_allocate_area(void **alloc_area)
 129 {
 130         if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
 131                 fprintf(stderr, "out of memory\n");
 132                 *alloc_area = NULL;
 133         }
 134 }
 135
 136 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 137 {
 138 }
 139
 140 /* HugeTLB memory */
 141 static int hugetlb_release_pages(char *rel_area)
 142 {
 143         int ret = 0;
 144
 145         if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 146                                 rel_area == huge_fd_off0 ? 0 :
 147                                 nr_pages * page_size,
 148                                 nr_pages * page_size)) {
 149                 perror("fallocate");
 150                 ret = 1;
 151         }
 152
 153         return ret;
 154 }
 155
 156
 157 static void hugetlb_allocate_area(void **alloc_area)
 158 {
 159         void *area_alias = NULL;
 160         char **alloc_area_alias;
 161         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 162                            (map_shared ? MAP_SHARED : MAP_PRIVATE) |
 163                            MAP_HUGETLB,
 164                            huge_fd, *alloc_area == area_src ? 0 :
 165                            nr_pages * page_size);
 166         if (*alloc_area == MAP_FAILED) {
 167                 fprintf(stderr, "mmap of hugetlbfs file failed\n");
 168                 *alloc_area = NULL;
 169         }
 170
 171         if (map_shared) {
 172                 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 173                                   MAP_SHARED | MAP_HUGETLB,
 174                                   huge_fd, *alloc_area == area_src ? 0 :
 175                                   nr_pages * page_size);
 176                 if (area_alias == MAP_FAILED) {
 177                         if (munmap(*alloc_area, nr_pages * page_size) < 0)
 178                                 perror("hugetlb munmap"), exit(1);
 179                         *alloc_area = NULL;
 180                         return;
 181                 }
 182         }
 183         if (*alloc_area == area_src) {
 184                 huge_fd_off0 = *alloc_area;
 185                 alloc_area_alias = &area_src_alias;
 186         } else {
 187                 alloc_area_alias = &area_dst_alias;
 188         }
 189         if (area_alias)
 190                 *alloc_area_alias = area_alias;
 191 }
 192
 193 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 194 {
 195         if (!map_shared)
 196                 return;
 197         /*
 198          * We can't zap just the pagetable with hugetlbfs because
 199          * MADV_DONTEED won't work. So exercise -EEXIST on a alias
 200          * mapping where the pagetables are not established initially,
 201          * this way we'll exercise the -EEXEC at the fs level.
 202          */
 203         *start = (unsigned long) area_dst_alias + offset;
 204 }
 205
 206 /* Shared memory */
 207 static int shmem_release_pages(char *rel_area)
 208 {
 209         int ret = 0;
 210
 211         if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
 212                 perror("madvise");
 213                 ret = 1;
 214         }
 215
 216         return ret;
 217 }
 218
 219 static void shmem_allocate_area(void **alloc_area)
 220 {
 221         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 222                            MAP_ANONYMOUS | MAP_SHARED, -1, 0);
 223         if (*alloc_area == MAP_FAILED) {
 224                 fprintf(stderr, "shared memory mmap failed\n");
 225                 *alloc_area = NULL;
 226         }
 227 }
 228
 229 struct uffd_test_ops {
 230         unsigned long expected_ioctls;
 231         void (*allocate_area)(void **alloc_area);
 232         int (*release_pages)(char *rel_area);
 233         void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
 234 };
 235
 236 #define ANON_EXPECTED_IOCTLS            ((1 << _UFFDIO_WAKE) | \
 237                                          (1 << _UFFDIO_COPY) | \
 238                                          (1 << _UFFDIO_ZEROPAGE))
 239
 240 static struct uffd_test_ops anon_uffd_test_ops = {
 241         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 242         .allocate_area  = anon_allocate_area,
 243         .release_pages  = anon_release_pages,
 244         .alias_mapping = noop_alias_mapping,
 245 };
 246
 247 static struct uffd_test_ops shmem_uffd_test_ops = {
 248         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 249         .allocate_area  = shmem_allocate_area,
 250         .release_pages  = shmem_release_pages,
 251         .alias_mapping = noop_alias_mapping,
 252 };
 253
 254 static struct uffd_test_ops hugetlb_uffd_test_ops = {
 255         .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
 256         .allocate_area  = hugetlb_allocate_area,
 257         .release_pages  = hugetlb_release_pages,
 258         .alias_mapping = hugetlb_alias_mapping,
 259 };
 260
 261 static struct uffd_test_ops *uffd_test_ops;
 262
 263 static int my_bcmp(char *str1, char *str2, size_t n)
 264 {
 265         unsigned long i;
 266         for (i = 0; i < n; i++)
 267                 if (str1[i] != str2[i])
 268                         return 1;
 269         return 0;
 270 }
 271
 272 static void *locking_thread(void *arg)
 273 {
 274         unsigned long cpu = (unsigned long) arg;
 275         struct random_data rand;
 276         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 277         int32_t rand_nr;
 278         unsigned long long count;
 279         char randstate[64];
 280         unsigned int seed;
 281         time_t start;
 282
 283         if (bounces & BOUNCE_RANDOM) {
 284                 seed = (unsigned int) time(NULL) - bounces;
 285                 if (!(bounces & BOUNCE_RACINGFAULTS))
 286                         seed += cpu;
 287                 bzero(&rand, sizeof(rand));
 288                 bzero(&randstate, sizeof(randstate));
 289                 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
 290                         fprintf(stderr, "srandom_r error\n"), exit(1);
 291         } else {
 292                 page_nr = -bounces;
 293                 if (!(bounces & BOUNCE_RACINGFAULTS))
 294                         page_nr += cpu * nr_pages_per_cpu;
 295         }
 296
 297         while (!finished) {
 298                 if (bounces & BOUNCE_RANDOM) {
 299                         if (random_r(&rand, &rand_nr))
 300                                 fprintf(stderr, "random_r 1 error\n"), exit(1);
 301                         page_nr = rand_nr;
 302                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 303                                 if (random_r(&rand, &rand_nr))
 304                                         fprintf(stderr, "random_r 2 error\n"), exit(1);
 305                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 306                                            16;
 307                         }
 308                 } else
 309                         page_nr += 1;
 310                 page_nr %= nr_pages;
 311
 312                 start = time(NULL);
 313                 if (bounces & BOUNCE_VERIFY) {
 314                         count = *area_count(area_dst, page_nr);
 315                         if (!count)
 316                                 fprintf(stderr,
 317                                         "page_nr %lu wrong count %Lu %Lu\n",
 318                                         page_nr, count,
 319                                         count_verify[page_nr]), exit(1);
 320
 321
 322                         /*
 323                          * We can't use bcmp (or memcmp) because that
 324                          * returns 0 erroneously if the memory is
 325                          * changing under it (even if the end of the
 326                          * page is never changing and always
 327                          * different).
 328                          */
 329 #if 1
 330                         if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
 331                                      page_size))
 332                                 fprintf(stderr,
 333                                         "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
 334                                         page_nr, count,
 335                                         count_verify[page_nr]), exit(1);
 336 #else
 337                         unsigned long loops;
 338
 339                         loops = 0;
 340                         /* uncomment the below line to test with mutex */
 341                         /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
 342                         while (!bcmp(area_dst + page_nr * page_size, zeropage,
 343                                      page_size)) {
 344                                 loops += 1;
 345                                 if (loops > 10)
 346                                         break;
 347                         }
 348                         /* uncomment below line to test with mutex */
 349                         /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
 350                         if (loops) {
 351                                 fprintf(stderr,
 352                                         "page_nr %lu all zero thread %lu %p %lu\n",
 353                                         page_nr, cpu, area_dst + page_nr * page_size,
 354                                         loops);
 355                                 if (loops > 10)
 356                                         exit(1);
 357                         }
 358 #endif
 359                 }
 360
 361                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 362                 count = *area_count(area_dst, page_nr);
 363                 if (count != count_verify[page_nr]) {
 364                         fprintf(stderr,
 365                                 "page_nr %lu memory corruption %Lu %Lu\n",
 366                                 page_nr, count,
 367                                 count_verify[page_nr]), exit(1);
 368                 }
 369                 count++;
 370                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 371                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 372
 373                 if (time(NULL) - start > 1)
 374                         fprintf(stderr,
 375                                 "userfault too slow %ld "
 376                                 "possible false positive with overcommit\n",
 377                                 time(NULL) - start);
 378         }
 379
 380         return NULL;
 381 }
 382
 383 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 384                             unsigned long offset)
 385 {
 386         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 387                                      uffdio_copy->len,
 388                                      offset);
 389         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 390                 /* real retval in ufdio_copy.copy */
 391                 if (uffdio_copy->copy != -EEXIST)
 392                         fprintf(stderr, "UFFDIO_COPY retry error %Ld\n",
 393                                 uffdio_copy->copy), exit(1);
 394         } else {
 395                 fprintf(stderr, "UFFDIO_COPY retry unexpected %Ld\n",
 396                         uffdio_copy->copy), exit(1);
 397         }
 398 }
 399
 400 static int __copy_page(int ufd, unsigned long offset, bool retry)
 401 {
 402         struct uffdio_copy uffdio_copy;
 403
 404         if (offset >= nr_pages * page_size)
 405                 fprintf(stderr, "unexpected offset %lu\n",
 406                         offset), exit(1);
 407         uffdio_copy.dst = (unsigned long) area_dst + offset;
 408         uffdio_copy.src = (unsigned long) area_src + offset;
 409         uffdio_copy.len = page_size;
 410         uffdio_copy.mode = 0;
 411         uffdio_copy.copy = 0;
 412         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 413                 /* real retval in ufdio_copy.copy */
 414                 if (uffdio_copy.copy != -EEXIST)
 415                         fprintf(stderr, "UFFDIO_COPY error %Ld\n",
 416                                 uffdio_copy.copy), exit(1);
 417         } else if (uffdio_copy.copy != page_size) {
 418                 fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
 419                         uffdio_copy.copy), exit(1);
 420         } else {
 421                 if (test_uffdio_copy_eexist && retry) {
 422                         test_uffdio_copy_eexist = false;
 423                         retry_copy_page(ufd, &uffdio_copy, offset);
 424                 }
 425                 return 1;
 426         }
 427         return 0;
 428 }
 429
 430 static int copy_page_retry(int ufd, unsigned long offset)
 431 {
 432         return __copy_page(ufd, offset, true);
 433 }
 434
 435 static int copy_page(int ufd, unsigned long offset)
 436 {
 437         return __copy_page(ufd, offset, false);
 438 }
 439
 440 static void *uffd_poll_thread(void *arg)
 441 {
 442         unsigned long cpu = (unsigned long) arg;
 443         struct pollfd pollfd[2];
 444         struct uffd_msg msg;
 445         struct uffdio_register uffd_reg;
 446         int ret;
 447         unsigned long offset;
 448         char tmp_chr;
 449         unsigned long userfaults = 0;
 450
 451         pollfd[0].fd = uffd;
 452         pollfd[0].events = POLLIN;
 453         pollfd[1].fd = pipefd[cpu*2];
 454         pollfd[1].events = POLLIN;
 455
 456         for (;;) {
 457                 ret = poll(pollfd, 2, -1);
 458                 if (!ret)
 459                         fprintf(stderr, "poll error %d\n", ret), exit(1);
 460                 if (ret < 0)
 461                         perror("poll"), exit(1);
 462                 if (pollfd[1].revents & POLLIN) {
 463                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
 464                                 fprintf(stderr, "read pipefd error\n"),
 465                                         exit(1);
 466                         break;
 467                 }
 468                 if (!(pollfd[0].revents & POLLIN))
 469                         fprintf(stderr, "pollfd[0].revents %d\n",
 470                                 pollfd[0].revents), exit(1);
 471                 ret = read(uffd, &msg, sizeof(msg));
 472                 if (ret < 0) {
 473                         if (errno == EAGAIN)
 474                                 continue;
 475                         perror("nonblocking read error"), exit(1);
 476                 }
 477                 switch (msg.event) {
 478                 default:
 479                         fprintf(stderr, "unexpected msg event %u\n",
 480                                 msg.event), exit(1);
 481                         break;
 482                 case UFFD_EVENT_PAGEFAULT:
 483                         if (msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 484                                 fprintf(stderr, "unexpected write fault\n"), exit(1);
 485                         offset = (char *)(unsigned long)msg.arg.pagefault.address -
 486                                 area_dst;
 487                         offset &= ~(page_size-1);
 488                         if (copy_page(uffd, offset))
 489                                 userfaults++;
 490                         break;
 491                 case UFFD_EVENT_FORK:
 492                         close(uffd);
 493                         uffd = msg.arg.fork.ufd;
 494                         pollfd[0].fd = uffd;
 495                         break;
 496                 case UFFD_EVENT_REMOVE:
 497                         uffd_reg.range.start = msg.arg.remove.start;
 498                         uffd_reg.range.len = msg.arg.remove.end -
 499                                 msg.arg.remove.start;
 500                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
 501                                 fprintf(stderr, "remove failure\n"), exit(1);
 502                         break;
 503                 case UFFD_EVENT_REMAP:
 504                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 505                         break;
 506                 }
 507         }
 508         return (void *)userfaults;
 509 }
 510
 511 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 512
 513 static void *uffd_read_thread(void *arg)
 514 {
 515         unsigned long *this_cpu_userfaults;
 516         struct uffd_msg msg;
 517         unsigned long offset;
 518         int ret;
 519
 520         this_cpu_userfaults = (unsigned long *) arg;
 521         *this_cpu_userfaults = 0;
 522
 523         pthread_mutex_unlock(&uffd_read_mutex);
 524         /* from here cancellation is ok */
 525
 526         for (;;) {
 527                 ret = read(uffd, &msg, sizeof(msg));
 528                 if (ret != sizeof(msg)) {
 529                         if (ret < 0)
 530                                 perror("blocking read error"), exit(1);
 531                         else
 532                                 fprintf(stderr, "short read\n"), exit(1);
 533                 }
 534                 if (msg.event != UFFD_EVENT_PAGEFAULT)
 535                         fprintf(stderr, "unexpected msg event %u\n",
 536                                 msg.event), exit(1);
 537                 if (bounces & BOUNCE_VERIFY &&
 538                     msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 539                         fprintf(stderr, "unexpected write fault\n"), exit(1);
 540                 offset = (char *)(unsigned long)msg.arg.pagefault.address -
 541                          area_dst;
 542                 offset &= ~(page_size-1);
 543                 if (copy_page(uffd, offset))
 544                         (*this_cpu_userfaults)++;
 545         }
 546         return (void *)NULL;
 547 }
 548
 549 static void *background_thread(void *arg)
 550 {
 551         unsigned long cpu = (unsigned long) arg;
 552         unsigned long page_nr;
 553
 554         for (page_nr = cpu * nr_pages_per_cpu;
 555              page_nr < (cpu+1) * nr_pages_per_cpu;
 556              page_nr++)
 557                 copy_page_retry(uffd, page_nr * page_size);
 558
 559         return NULL;
 560 }
 561
 562 static int stress(unsigned long *userfaults)
 563 {
 564         unsigned long cpu;
 565         pthread_t locking_threads[nr_cpus];
 566         pthread_t uffd_threads[nr_cpus];
 567         pthread_t background_threads[nr_cpus];
 568         void **_userfaults = (void **) userfaults;
 569
 570         finished = 0;
 571         for (cpu = 0; cpu < nr_cpus; cpu++) {
 572                 if (pthread_create(&locking_threads[cpu], &attr,
 573                                    locking_thread, (void *)cpu))
 574                         return 1;
 575                 if (bounces & BOUNCE_POLL) {
 576                         if (pthread_create(&uffd_threads[cpu], &attr,
 577                                            uffd_poll_thread, (void *)cpu))
 578                                 return 1;
 579                 } else {
 580                         if (pthread_create(&uffd_threads[cpu], &attr,
 581                                            uffd_read_thread,
 582                                            &_userfaults[cpu]))
 583                                 return 1;
 584                         pthread_mutex_lock(&uffd_read_mutex);
 585                 }
 586                 if (pthread_create(&background_threads[cpu], &attr,
 587                                    background_thread, (void *)cpu))
 588                         return 1;
 589         }
 590         for (cpu = 0; cpu < nr_cpus; cpu++)
 591                 if (pthread_join(background_threads[cpu], NULL))
 592                         return 1;
 593
 594         /*
 595          * Be strict and immediately zap area_src, the whole area has
 596          * been transferred already by the background treads. The
 597          * area_src could then be faulted in in a racy way by still
 598          * running uffdio_threads reading zeropages after we zapped
 599          * area_src (but they're guaranteed to get -EEXIST from
 600          * UFFDIO_COPY without writing zero pages into area_dst
 601          * because the background threads already completed).
 602          */
 603         if (uffd_test_ops->release_pages(area_src))
 604                 return 1;
 605
 606         for (cpu = 0; cpu < nr_cpus; cpu++) {
 607                 char c;
 608                 if (bounces & BOUNCE_POLL) {
 609                         if (write(pipefd[cpu*2+1], &c, 1) != 1) {
 610                                 fprintf(stderr, "pipefd write error\n");
 611                                 return 1;
 612                         }
 613                         if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
 614                                 return 1;
 615                 } else {
 616                         if (pthread_cancel(uffd_threads[cpu]))
 617                                 return 1;
 618                         if (pthread_join(uffd_threads[cpu], NULL))
 619                                 return 1;
 620                 }
 621         }
 622
 623         finished = 1;
 624         for (cpu = 0; cpu < nr_cpus; cpu++)
 625                 if (pthread_join(locking_threads[cpu], NULL))
 626                         return 1;
 627
 628         return 0;
 629 }
 630
 631 static int userfaultfd_open(int features)
 632 {
 633         struct uffdio_api uffdio_api;
 634
 635         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 636         if (uffd < 0) {
 637                 fprintf(stderr,
 638                         "userfaultfd syscall not available in this kernel\n");
 639                 return 1;
 640         }
 641         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 642
 643         uffdio_api.api = UFFD_API;
 644         uffdio_api.features = features;
 645         if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
 646                 fprintf(stderr, "UFFDIO_API\n");
 647                 return 1;
 648         }
 649         if (uffdio_api.api != UFFD_API) {
 650                 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
 651                 return 1;
 652         }
 653
 654         return 0;
 655 }
 656
 657 sigjmp_buf jbuf, *sigbuf;
 658
 659 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 660 {
 661         if (sig == SIGBUS) {
 662                 if (sigbuf)
 663                         siglongjmp(*sigbuf, 1);
 664                 abort();
 665         }
 666 }
 667
 668 /*
 669  * For non-cooperative userfaultfd test we fork() a process that will
 670  * generate pagefaults, will mremap the area monitored by the
 671  * userfaultfd and at last this process will release the monitored
 672  * area.
 673  * For the anonymous and shared memory the area is divided into two
 674  * parts, the first part is accessed before mremap, and the second
 675  * part is accessed after mremap. Since hugetlbfs does not support
 676  * mremap, the entire monitored area is accessed in a single pass for
 677  * HUGETLB_TEST.
 678  * The release of the pages currently generates event for shmem and
 679  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 680  * for hugetlb.
 681  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 682  * monitored area, generate pagefaults and test that signal is delivered.
 683  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 684  * test robustness use case - we release monitored area, fork a process
 685  * that will generate pagefaults and verify signal is generated.
 686  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 687  * feature. Using monitor thread, verify no userfault events are generated.
 688  */
 689 static int faulting_process(int signal_test)
 690 {
 691         unsigned long nr;
 692         unsigned long long count;
 693         unsigned long split_nr_pages;
 694         unsigned long lastnr;
 695         struct sigaction act;
 696         unsigned long signalled = 0;
 697
 698         if (test_type != TEST_HUGETLB)
 699                 split_nr_pages = (nr_pages + 1) / 2;
 700         else
 701                 split_nr_pages = nr_pages;
 702
 703         if (signal_test) {
 704                 sigbuf = &jbuf;
 705                 memset(&act, 0, sizeof(act));
 706                 act.sa_sigaction = sighndl;
 707                 act.sa_flags = SA_SIGINFO;
 708                 if (sigaction(SIGBUS, &act, 0)) {
 709                         perror("sigaction");
 710                         return 1;
 711                 }
 712                 lastnr = (unsigned long)-1;
 713         }
 714
 715         for (nr = 0; nr < split_nr_pages; nr++) {
 716                 if (signal_test) {
 717                         if (sigsetjmp(*sigbuf, 1) != 0) {
 718                                 if (nr == lastnr) {
 719                                         fprintf(stderr, "Signal repeated\n");
 720                                         return 1;
 721                                 }
 722
 723                                 lastnr = nr;
 724                                 if (signal_test == 1) {
 725                                         if (copy_page(uffd, nr * page_size))
 726                                                 signalled++;
 727                                 } else {
 728                                         signalled++;
 729                                         continue;
 730                                 }
 731                         }
 732                 }
 733
 734                 count = *area_count(area_dst, nr);
 735                 if (count != count_verify[nr]) {
 736                         fprintf(stderr,
 737                                 "nr %lu memory corruption %Lu %Lu\n",
 738                                 nr, count,
 739                                 count_verify[nr]), exit(1);
 740                 }
 741         }
 742
 743         if (signal_test)
 744                 return signalled != split_nr_pages;
 745
 746         if (test_type == TEST_HUGETLB)
 747                 return 0;
 748
 749         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 750                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 751         if (area_dst == MAP_FAILED)
 752                 perror("mremap"), exit(1);
 753
 754         for (; nr < nr_pages; nr++) {
 755                 count = *area_count(area_dst, nr);
 756                 if (count != count_verify[nr]) {
 757                         fprintf(stderr,
 758                                 "nr %lu memory corruption %Lu %Lu\n",
 759                                 nr, count,
 760                                 count_verify[nr]), exit(1);
 761                 }
 762         }
 763
 764         if (uffd_test_ops->release_pages(area_dst))
 765                 return 1;
 766
 767         for (nr = 0; nr < nr_pages; nr++) {
 768                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
 769                         fprintf(stderr, "nr %lu is not zero\n", nr), exit(1);
 770         }
 771
 772         return 0;
 773 }
 774
 775 static void retry_uffdio_zeropage(int ufd,
 776                                   struct uffdio_zeropage *uffdio_zeropage,
 777                                   unsigned long offset)
 778 {
 779         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
 780                                      uffdio_zeropage->range.len,
 781                                      offset);
 782         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
 783                 if (uffdio_zeropage->zeropage != -EEXIST)
 784                         fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
 785                                 uffdio_zeropage->zeropage), exit(1);
 786         } else {
 787                 fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
 788                         uffdio_zeropage->zeropage), exit(1);
 789         }
 790 }
 791
 792 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
 793 {
 794         struct uffdio_zeropage uffdio_zeropage;
 795         int ret;
 796         unsigned long has_zeropage;
 797
 798         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
 799
 800         if (offset >= nr_pages * page_size)
 801                 fprintf(stderr, "unexpected offset %lu\n",
 802                         offset), exit(1);
 803         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
 804         uffdio_zeropage.range.len = page_size;
 805         uffdio_zeropage.mode = 0;
 806         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
 807         if (ret) {
 808                 /* real retval in ufdio_zeropage.zeropage */
 809                 if (has_zeropage) {
 810                         if (uffdio_zeropage.zeropage == -EEXIST)
 811                                 fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n"),
 812                                         exit(1);
 813                         else
 814                                 fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
 815                                         uffdio_zeropage.zeropage), exit(1);
 816                 } else {
 817                         if (uffdio_zeropage.zeropage != -EINVAL)
 818                                 fprintf(stderr,
 819                                         "UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
 820                                         uffdio_zeropage.zeropage), exit(1);
 821                 }
 822         } else if (has_zeropage) {
 823                 if (uffdio_zeropage.zeropage != page_size) {
 824                         fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
 825                                 uffdio_zeropage.zeropage), exit(1);
 826                 } else {
 827                         if (test_uffdio_zeropage_eexist && retry) {
 828                                 test_uffdio_zeropage_eexist = false;
 829                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
 830                                                       offset);
 831                         }
 832                         return 1;
 833                 }
 834         } else {
 835                 fprintf(stderr,
 836                         "UFFDIO_ZEROPAGE succeeded %Ld\n",
 837                         uffdio_zeropage.zeropage), exit(1);
 838         }
 839
 840         return 0;
 841 }
 842
 843 static int uffdio_zeropage(int ufd, unsigned long offset)
 844 {
 845         return __uffdio_zeropage(ufd, offset, false);
 846 }
 847
 848 /* exercise UFFDIO_ZEROPAGE */
 849 static int userfaultfd_zeropage_test(void)
 850 {
 851         struct uffdio_register uffdio_register;
 852         unsigned long expected_ioctls;
 853
 854         printf("testing UFFDIO_ZEROPAGE: ");
 855         fflush(stdout);
 856
 857         if (uffd_test_ops->release_pages(area_dst))
 858                 return 1;
 859
 860         if (userfaultfd_open(0) < 0)
 861                 return 1;
 862         uffdio_register.range.start = (unsigned long) area_dst;
 863         uffdio_register.range.len = nr_pages * page_size;
 864         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 865         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 866                 fprintf(stderr, "register failure\n"), exit(1);
 867
 868         expected_ioctls = uffd_test_ops->expected_ioctls;
 869         if ((uffdio_register.ioctls & expected_ioctls) !=
 870             expected_ioctls)
 871                 fprintf(stderr,
 872                         "unexpected missing ioctl for anon memory\n"),
 873                         exit(1);
 874
 875         if (uffdio_zeropage(uffd, 0)) {
 876                 if (my_bcmp(area_dst, zeropage, page_size))
 877                         fprintf(stderr, "zeropage is not zero\n"), exit(1);
 878         }
 879
 880         close(uffd);
 881         printf("done.\n");
 882         return 0;
 883 }
 884
 885 static int userfaultfd_events_test(void)
 886 {
 887         struct uffdio_register uffdio_register;
 888         unsigned long expected_ioctls;
 889         unsigned long userfaults;
 890         pthread_t uffd_mon;
 891         int err, features;
 892         pid_t pid;
 893         char c;
 894
 895         printf("testing events (fork, remap, remove): ");
 896         fflush(stdout);
 897
 898         if (uffd_test_ops->release_pages(area_dst))
 899                 return 1;
 900
 901         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
 902                 UFFD_FEATURE_EVENT_REMOVE;
 903         if (userfaultfd_open(features) < 0)
 904                 return 1;
 905         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 906
 907         uffdio_register.range.start = (unsigned long) area_dst;
 908         uffdio_register.range.len = nr_pages * page_size;
 909         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 910         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 911                 fprintf(stderr, "register failure\n"), exit(1);
 912
 913         expected_ioctls = uffd_test_ops->expected_ioctls;
 914         if ((uffdio_register.ioctls & expected_ioctls) !=
 915             expected_ioctls)
 916                 fprintf(stderr,
 917                         "unexpected missing ioctl for anon memory\n"),
 918                         exit(1);
 919
 920         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
 921                 perror("uffd_poll_thread create"), exit(1);
 922
 923         pid = fork();
 924         if (pid < 0)
 925                 perror("fork"), exit(1);
 926
 927         if (!pid)
 928                 return faulting_process(0);
 929
 930         waitpid(pid, &err, 0);
 931         if (err)
 932                 fprintf(stderr, "faulting process failed\n"), exit(1);
 933
 934         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
 935                 perror("pipe write"), exit(1);
 936         if (pthread_join(uffd_mon, (void **)&userfaults))
 937                 return 1;
 938
 939         close(uffd);
 940         printf("userfaults: %ld\n", userfaults);
 941
 942         return userfaults != nr_pages;
 943 }
 944
 945 static int userfaultfd_sig_test(void)
 946 {
 947         struct uffdio_register uffdio_register;
 948         unsigned long expected_ioctls;
 949         unsigned long userfaults;
 950         pthread_t uffd_mon;
 951         int err, features;
 952         pid_t pid;
 953         char c;
 954
 955         printf("testing signal delivery: ");
 956         fflush(stdout);
 957
 958         if (uffd_test_ops->release_pages(area_dst))
 959                 return 1;
 960
 961         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
 962         if (userfaultfd_open(features) < 0)
 963                 return 1;
 964         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 965
 966         uffdio_register.range.start = (unsigned long) area_dst;
 967         uffdio_register.range.len = nr_pages * page_size;
 968         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 969         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 970                 fprintf(stderr, "register failure\n"), exit(1);
 971
 972         expected_ioctls = uffd_test_ops->expected_ioctls;
 973         if ((uffdio_register.ioctls & expected_ioctls) !=
 974             expected_ioctls)
 975                 fprintf(stderr,
 976                         "unexpected missing ioctl for anon memory\n"),
 977                         exit(1);
 978
 979         if (faulting_process(1))
 980                 fprintf(stderr, "faulting process failed\n"), exit(1);
 981
 982         if (uffd_test_ops->release_pages(area_dst))
 983                 return 1;
 984
 985         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
 986                 perror("uffd_poll_thread create"), exit(1);
 987
 988         pid = fork();
 989         if (pid < 0)
 990                 perror("fork"), exit(1);
 991
 992         if (!pid)
 993                 exit(faulting_process(2));
 994
 995         waitpid(pid, &err, 0);
 996         if (err)
 997                 fprintf(stderr, "faulting process failed\n"), exit(1);
 998
 999         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1000                 perror("pipe write"), exit(1);
1001         if (pthread_join(uffd_mon, (void **)&userfaults))
1002                 return 1;
1003
1004         printf("done.\n");
1005         if (userfaults)
1006                 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1007                         userfaults);
1008         close(uffd);
1009         return userfaults != 0;
1010 }
1011 static int userfaultfd_stress(void)
1012 {
1013         void *area;
1014         char *tmp_area;
1015         unsigned long nr;
1016         struct uffdio_register uffdio_register;
1017         unsigned long cpu;
1018         int err;
1019         unsigned long userfaults[nr_cpus];
1020
1021         uffd_test_ops->allocate_area((void **)&area_src);
1022         if (!area_src)
1023                 return 1;
1024         uffd_test_ops->allocate_area((void **)&area_dst);
1025         if (!area_dst)
1026                 return 1;
1027
1028         if (userfaultfd_open(0) < 0)
1029                 return 1;
1030
1031         count_verify = malloc(nr_pages * sizeof(unsigned long long));
1032         if (!count_verify) {
1033                 perror("count_verify");
1034                 return 1;
1035         }
1036
1037         for (nr = 0; nr < nr_pages; nr++) {
1038                 *area_mutex(area_src, nr) = (pthread_mutex_t)
1039                         PTHREAD_MUTEX_INITIALIZER;
1040                 count_verify[nr] = *area_count(area_src, nr) = 1;
1041                 /*
1042                  * In the transition between 255 to 256, powerpc will
1043                  * read out of order in my_bcmp and see both bytes as
1044                  * zero, so leave a placeholder below always non-zero
1045                  * after the count, to avoid my_bcmp to trigger false
1046                  * positives.
1047                  */
1048                 *(area_count(area_src, nr) + 1) = 1;
1049         }
1050
1051         pipefd = malloc(sizeof(int) * nr_cpus * 2);
1052         if (!pipefd) {
1053                 perror("pipefd");
1054                 return 1;
1055         }
1056         for (cpu = 0; cpu < nr_cpus; cpu++) {
1057                 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1058                         perror("pipe");
1059                         return 1;
1060                 }
1061         }
1062
1063         if (posix_memalign(&area, page_size, page_size)) {
1064                 fprintf(stderr, "out of memory\n");
1065                 return 1;
1066         }
1067         zeropage = area;
1068         bzero(zeropage, page_size);
1069
1070         pthread_mutex_lock(&uffd_read_mutex);
1071
1072         pthread_attr_init(&attr);
1073         pthread_attr_setstacksize(&attr, 16*1024*1024);
1074
1075         err = 0;
1076         while (bounces--) {
1077                 unsigned long expected_ioctls;
1078
1079                 printf("bounces: %d, mode:", bounces);
1080                 if (bounces & BOUNCE_RANDOM)
1081                         printf(" rnd");
1082                 if (bounces & BOUNCE_RACINGFAULTS)
1083                         printf(" racing");
1084                 if (bounces & BOUNCE_VERIFY)
1085                         printf(" ver");
1086                 if (bounces & BOUNCE_POLL)
1087                         printf(" poll");
1088                 printf(", ");
1089                 fflush(stdout);
1090
1091                 if (bounces & BOUNCE_POLL)
1092                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1093                 else
1094                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1095
1096                 /* register */
1097                 uffdio_register.range.start = (unsigned long) area_dst;
1098                 uffdio_register.range.len = nr_pages * page_size;
1099                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1100                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1101                         fprintf(stderr, "register failure\n");
1102                         return 1;
1103                 }
1104                 expected_ioctls = uffd_test_ops->expected_ioctls;
1105                 if ((uffdio_register.ioctls & expected_ioctls) !=
1106                     expected_ioctls) {
1107                         fprintf(stderr,
1108                                 "unexpected missing ioctl for anon memory\n");
1109                         return 1;
1110                 }
1111
1112                 if (area_dst_alias) {
1113                         uffdio_register.range.start = (unsigned long)
1114                                 area_dst_alias;
1115                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1116                                 fprintf(stderr, "register failure alias\n");
1117                                 return 1;
1118                         }
1119                 }
1120
1121                 /*
1122                  * The madvise done previously isn't enough: some
1123                  * uffd_thread could have read userfaults (one of
1124                  * those already resolved by the background thread)
1125                  * and it may be in the process of calling
1126                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1127                  * area_src and it would map a zero page in it (of
1128                  * course such a UFFDIO_COPY is perfectly safe as it'd
1129                  * return -EEXIST). The problem comes at the next
1130                  * bounce though: that racing UFFDIO_COPY would
1131                  * generate zeropages in the area_src, so invalidating
1132                  * the previous MADV_DONTNEED. Without this additional
1133                  * MADV_DONTNEED those zeropages leftovers in the
1134                  * area_src would lead to -EEXIST failure during the
1135                  * next bounce, effectively leaving a zeropage in the
1136                  * area_dst.
1137                  *
1138                  * Try to comment this out madvise to see the memory
1139                  * corruption being caught pretty quick.
1140                  *
1141                  * khugepaged is also inhibited to collapse THP after
1142                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1143                  * required to MADV_DONTNEED here.
1144                  */
1145                 if (uffd_test_ops->release_pages(area_dst))
1146                         return 1;
1147
1148                 /* bounce pass */
1149                 if (stress(userfaults))
1150                         return 1;
1151
1152                 /* unregister */
1153                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1154                         fprintf(stderr, "unregister failure\n");
1155                         return 1;
1156                 }
1157                 if (area_dst_alias) {
1158                         uffdio_register.range.start = (unsigned long) area_dst;
1159                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1160                                   &uffdio_register.range)) {
1161                                 fprintf(stderr, "unregister failure alias\n");
1162                                 return 1;
1163                         }
1164                 }
1165
1166                 /* verification */
1167                 if (bounces & BOUNCE_VERIFY) {
1168                         for (nr = 0; nr < nr_pages; nr++) {
1169                                 if (*area_count(area_dst, nr) != count_verify[nr]) {
1170                                         fprintf(stderr,
1171                                                 "error area_count %Lu %Lu %lu\n",
1172                                                 *area_count(area_src, nr),
1173                                                 count_verify[nr],
1174                                                 nr);
1175                                         err = 1;
1176                                         bounces = 0;
1177                                 }
1178                         }
1179                 }
1180
1181                 /* prepare next bounce */
1182                 tmp_area = area_src;
1183                 area_src = area_dst;
1184                 area_dst = tmp_area;
1185
1186                 tmp_area = area_src_alias;
1187                 area_src_alias = area_dst_alias;
1188                 area_dst_alias = tmp_area;
1189
1190                 printf("userfaults:");
1191                 for (cpu = 0; cpu < nr_cpus; cpu++)
1192                         printf(" %lu", userfaults[cpu]);
1193                 printf("\n");
1194         }
1195
1196         if (err)
1197                 return err;
1198
1199         close(uffd);
1200         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1201                 || userfaultfd_events_test();
1202 }
1203
1204 /*
1205  * Copied from mlock2-tests.c
1206  */
1207 unsigned long default_huge_page_size(void)
1208 {
1209         unsigned long hps = 0;
1210         char *line = NULL;
1211         size_t linelen = 0;
1212         FILE *f = fopen("/proc/meminfo", "r");
1213
1214         if (!f)
1215                 return 0;
1216         while (getline(&line, &linelen, f) > 0) {
1217                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1218                         hps <<= 10;
1219                         break;
1220                 }
1221         }
1222
1223         free(line);
1224         fclose(f);
1225         return hps;
1226 }
1227
1228 static void set_test_type(const char *type)
1229 {
1230         if (!strcmp(type, "anon")) {
1231                 test_type = TEST_ANON;
1232                 uffd_test_ops = &anon_uffd_test_ops;
1233         } else if (!strcmp(type, "hugetlb")) {
1234                 test_type = TEST_HUGETLB;
1235                 uffd_test_ops = &hugetlb_uffd_test_ops;
1236         } else if (!strcmp(type, "hugetlb_shared")) {
1237                 map_shared = true;
1238                 test_type = TEST_HUGETLB;
1239                 uffd_test_ops = &hugetlb_uffd_test_ops;
1240         } else if (!strcmp(type, "shmem")) {
1241                 map_shared = true;
1242                 test_type = TEST_SHMEM;
1243                 uffd_test_ops = &shmem_uffd_test_ops;
1244         } else {
1245                 fprintf(stderr, "Unknown test type: %s\n", type), exit(1);
1246         }
1247
1248         if (test_type == TEST_HUGETLB)
1249                 page_size = default_huge_page_size();
1250         else
1251                 page_size = sysconf(_SC_PAGE_SIZE);
1252
1253         if (!page_size)
1254                 fprintf(stderr, "Unable to determine page size\n"),
1255                                 exit(2);
1256         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1257             > page_size)
1258                 fprintf(stderr, "Impossible to run this test\n"), exit(2);
1259 }
1260
1261 static void sigalrm(int sig)
1262 {
1263         if (sig != SIGALRM)
1264                 abort();
1265         test_uffdio_copy_eexist = true;
1266         test_uffdio_zeropage_eexist = true;
1267         alarm(ALARM_INTERVAL_SECS);
1268 }
1269
1270 int main(int argc, char **argv)
1271 {
1272         if (argc < 4)
1273                 fprintf(stderr, "Usage: <test type> <MiB> <bounces> [hugetlbfs_file]\n"),
1274                                 exit(1);
1275
1276         if (signal(SIGALRM, sigalrm) == SIG_ERR)
1277                 fprintf(stderr, "failed to arm SIGALRM"), exit(1);
1278         alarm(ALARM_INTERVAL_SECS);
1279
1280         set_test_type(argv[1]);
1281
1282         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1283         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1284                 nr_cpus;
1285         if (!nr_pages_per_cpu) {
1286                 fprintf(stderr, "invalid MiB\n");
1287                 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
1288         }
1289
1290         bounces = atoi(argv[3]);
1291         if (bounces <= 0) {
1292                 fprintf(stderr, "invalid bounces\n");
1293                 fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
1294         }
1295         nr_pages = nr_pages_per_cpu * nr_cpus;
1296
1297         if (test_type == TEST_HUGETLB) {
1298                 if (argc < 5)
1299                         fprintf(stderr, "Usage: hugetlb <MiB> <bounces> <hugetlbfs_file>\n"),
1300                                 exit(1);
1301                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1302                 if (huge_fd < 0) {
1303                         fprintf(stderr, "Open of %s failed", argv[3]);
1304                         perror("open");
1305                         exit(1);
1306                 }
1307                 if (ftruncate(huge_fd, 0)) {
1308                         fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1309                         perror("ftruncate");
1310                         exit(1);
1311                 }
1312         }
1313         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1314                nr_pages, nr_pages_per_cpu);
1315         return userfaultfd_stress();
1316 }
1317
1318 #else /* __NR_userfaultfd */
1319
1320 #warning "missing __NR_userfaultfd definition"
1321
1322 int main(void)
1323 {
1324         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1325         return 0;
1326 }
1327
1328 #endif /* __NR_userfaultfd */