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
  39 #define _GNU_SOURCE
  40 #include <stdio.h>
  41 #include <errno.h>
  42 #include <unistd.h>
  43 #include <stdlib.h>
  44 #include <sys/types.h>
  45 #include <sys/stat.h>
  46 #include <fcntl.h>
  47 #include <time.h>
  48 #include <signal.h>
  49 #include <poll.h>
  50 #include <string.h>
  51 #include <sys/mman.h>
  52 #include <sys/syscall.h>
  53 #include <sys/ioctl.h>
  54 #include <sys/wait.h>
  55 #include <pthread.h>
  56 #include <linux/userfaultfd.h>
  57 #include <setjmp.h>
  58 #include <stdbool.h>
  59
  60 #include "../kselftest.h"
  61
  62 #ifdef __NR_userfaultfd
  63
  64 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
  65
  66 #define BOUNCE_RANDOM           (1<<0)
  67 #define BOUNCE_RACINGFAULTS     (1<<1)
  68 #define BOUNCE_VERIFY           (1<<2)
  69 #define BOUNCE_POLL             (1<<3)
  70 static int bounces;
  71
  72 #define TEST_ANON       1
  73 #define TEST_HUGETLB    2
  74 #define TEST_SHMEM      3
  75 static int test_type;
  76
  77 /* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
  78 #define ALARM_INTERVAL_SECS 10
  79 static volatile bool test_uffdio_copy_eexist = true;
  80 static volatile bool test_uffdio_zeropage_eexist = true;
  81
  82 static bool map_shared;
  83 static int huge_fd;
  84 static char *huge_fd_off0;
  85 static unsigned long long *count_verify;
  86 static int uffd, uffd_flags, finished, *pipefd;
  87 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
  88 static char *zeropage;
  89 pthread_attr_t attr;
  90
  91 /* pthread_mutex_t starts at page offset 0 */
  92 #define area_mutex(___area, ___nr)                                      \
  93         ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
  94 /*
  95  * count is placed in the page after pthread_mutex_t naturally aligned
  96  * to avoid non alignment faults on non-x86 archs.
  97  */
  98 #define area_count(___area, ___nr)                                      \
  99         ((volatile unsigned long long *) ((unsigned long)               \
 100                                  ((___area) + (___nr)*page_size +       \
 101                                   sizeof(pthread_mutex_t) +             \
 102                                   sizeof(unsigned long long) - 1) &     \
 103                                  ~(unsigned long)(sizeof(unsigned long long) \
 104                                                   -  1)))
 105
 106 const char *examples =
 107     "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
 108     "./userfaultfd anon 100 99999\n\n"
 109     "# Run share memory test on 1GiB region with 99 bounces:\n"
 110     "./userfaultfd shmem 1000 99\n\n"
 111     "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
 112     "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
 113     "# Run the same hugetlb test but using shmem:\n"
 114     "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
 115     "# 10MiB-~6GiB 999 bounces anonymous test, "
 116     "continue forever unless an error triggers\n"
 117     "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
 118
 119 static void usage(void)
 120 {
 121         fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
 122                 "[hugetlbfs_file]\n\n");
 123         fprintf(stderr, "Supported <test type>: anon, hugetlb, "
 124                 "hugetlb_shared, shmem\n\n");
 125         fprintf(stderr, "Examples:\n\n");
 126         fprintf(stderr, examples);
 127         exit(1);
 128 }
 129
 130 static int anon_release_pages(char *rel_area)
 131 {
 132         int ret = 0;
 133
 134         if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
 135                 perror("madvise");
 136                 ret = 1;
 137         }
 138
 139         return ret;
 140 }
 141
 142 static void anon_allocate_area(void **alloc_area)
 143 {
 144         if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
 145                 fprintf(stderr, "out of memory\n");
 146                 *alloc_area = NULL;
 147         }
 148 }
 149
 150 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 151 {
 152 }
 153
 154 /* HugeTLB memory */
 155 static int hugetlb_release_pages(char *rel_area)
 156 {
 157         int ret = 0;
 158
 159         if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 160                                 rel_area == huge_fd_off0 ? 0 :
 161                                 nr_pages * page_size,
 162                                 nr_pages * page_size)) {
 163                 perror("fallocate");
 164                 ret = 1;
 165         }
 166
 167         return ret;
 168 }
 169
 170
 171 static void hugetlb_allocate_area(void **alloc_area)
 172 {
 173         void *area_alias = NULL;
 174         char **alloc_area_alias;
 175         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 176                            (map_shared ? MAP_SHARED : MAP_PRIVATE) |
 177                            MAP_HUGETLB,
 178                            huge_fd, *alloc_area == area_src ? 0 :
 179                            nr_pages * page_size);
 180         if (*alloc_area == MAP_FAILED) {
 181                 fprintf(stderr, "mmap of hugetlbfs file failed\n");
 182                 *alloc_area = NULL;
 183         }
 184
 185         if (map_shared) {
 186                 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 187                                   MAP_SHARED | MAP_HUGETLB,
 188                                   huge_fd, *alloc_area == area_src ? 0 :
 189                                   nr_pages * page_size);
 190                 if (area_alias == MAP_FAILED) {
 191                         if (munmap(*alloc_area, nr_pages * page_size) < 0)
 192                                 perror("hugetlb munmap"), exit(1);
 193                         *alloc_area = NULL;
 194                         return;
 195                 }
 196         }
 197         if (*alloc_area == area_src) {
 198                 huge_fd_off0 = *alloc_area;
 199                 alloc_area_alias = &area_src_alias;
 200         } else {
 201                 alloc_area_alias = &area_dst_alias;
 202         }
 203         if (area_alias)
 204                 *alloc_area_alias = area_alias;
 205 }
 206
 207 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 208 {
 209         if (!map_shared)
 210                 return;
 211         /*
 212          * We can't zap just the pagetable with hugetlbfs because
 213          * MADV_DONTEED won't work. So exercise -EEXIST on a alias
 214          * mapping where the pagetables are not established initially,
 215          * this way we'll exercise the -EEXEC at the fs level.
 216          */
 217         *start = (unsigned long) area_dst_alias + offset;
 218 }
 219
 220 /* Shared memory */
 221 static int shmem_release_pages(char *rel_area)
 222 {
 223         int ret = 0;
 224
 225         if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
 226                 perror("madvise");
 227                 ret = 1;
 228         }
 229
 230         return ret;
 231 }
 232
 233 static void shmem_allocate_area(void **alloc_area)
 234 {
 235         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 236                            MAP_ANONYMOUS | MAP_SHARED, -1, 0);
 237         if (*alloc_area == MAP_FAILED) {
 238                 fprintf(stderr, "shared memory mmap failed\n");
 239                 *alloc_area = NULL;
 240         }
 241 }
 242
 243 struct uffd_test_ops {
 244         unsigned long expected_ioctls;
 245         void (*allocate_area)(void **alloc_area);
 246         int (*release_pages)(char *rel_area);
 247         void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
 248 };
 249
 250 #define ANON_EXPECTED_IOCTLS            ((1 << _UFFDIO_WAKE) | \
 251                                          (1 << _UFFDIO_COPY) | \
 252                                          (1 << _UFFDIO_ZEROPAGE))
 253
 254 static struct uffd_test_ops anon_uffd_test_ops = {
 255         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 256         .allocate_area  = anon_allocate_area,
 257         .release_pages  = anon_release_pages,
 258         .alias_mapping = noop_alias_mapping,
 259 };
 260
 261 static struct uffd_test_ops shmem_uffd_test_ops = {
 262         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 263         .allocate_area  = shmem_allocate_area,
 264         .release_pages  = shmem_release_pages,
 265         .alias_mapping = noop_alias_mapping,
 266 };
 267
 268 static struct uffd_test_ops hugetlb_uffd_test_ops = {
 269         .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
 270         .allocate_area  = hugetlb_allocate_area,
 271         .release_pages  = hugetlb_release_pages,
 272         .alias_mapping = hugetlb_alias_mapping,
 273 };
 274
 275 static struct uffd_test_ops *uffd_test_ops;
 276
 277 static int my_bcmp(char *str1, char *str2, size_t n)
 278 {
 279         unsigned long i;
 280         for (i = 0; i < n; i++)
 281                 if (str1[i] != str2[i])
 282                         return 1;
 283         return 0;
 284 }
 285
 286 static void *locking_thread(void *arg)
 287 {
 288         unsigned long cpu = (unsigned long) arg;
 289         struct random_data rand;
 290         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 291         int32_t rand_nr;
 292         unsigned long long count;
 293         char randstate[64];
 294         unsigned int seed;
 295         time_t start;
 296
 297         if (bounces & BOUNCE_RANDOM) {
 298                 seed = (unsigned int) time(NULL) - bounces;
 299                 if (!(bounces & BOUNCE_RACINGFAULTS))
 300                         seed += cpu;
 301                 bzero(&rand, sizeof(rand));
 302                 bzero(&randstate, sizeof(randstate));
 303                 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
 304                         fprintf(stderr, "srandom_r error\n"), exit(1);
 305         } else {
 306                 page_nr = -bounces;
 307                 if (!(bounces & BOUNCE_RACINGFAULTS))
 308                         page_nr += cpu * nr_pages_per_cpu;
 309         }
 310
 311         while (!finished) {
 312                 if (bounces & BOUNCE_RANDOM) {
 313                         if (random_r(&rand, &rand_nr))
 314                                 fprintf(stderr, "random_r 1 error\n"), exit(1);
 315                         page_nr = rand_nr;
 316                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 317                                 if (random_r(&rand, &rand_nr))
 318                                         fprintf(stderr, "random_r 2 error\n"), exit(1);
 319                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 320                                            16;
 321                         }
 322                 } else
 323                         page_nr += 1;
 324                 page_nr %= nr_pages;
 325
 326                 start = time(NULL);
 327                 if (bounces & BOUNCE_VERIFY) {
 328                         count = *area_count(area_dst, page_nr);
 329                         if (!count)
 330                                 fprintf(stderr,
 331                                         "page_nr %lu wrong count %Lu %Lu\n",
 332                                         page_nr, count,
 333                                         count_verify[page_nr]), exit(1);
 334
 335
 336                         /*
 337                          * We can't use bcmp (or memcmp) because that
 338                          * returns 0 erroneously if the memory is
 339                          * changing under it (even if the end of the
 340                          * page is never changing and always
 341                          * different).
 342                          */
 343 #if 1
 344                         if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
 345                                      page_size))
 346                                 fprintf(stderr,
 347                                         "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
 348                                         page_nr, count,
 349                                         count_verify[page_nr]), exit(1);
 350 #else
 351                         unsigned long loops;
 352
 353                         loops = 0;
 354                         /* uncomment the below line to test with mutex */
 355                         /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
 356                         while (!bcmp(area_dst + page_nr * page_size, zeropage,
 357                                      page_size)) {
 358                                 loops += 1;
 359                                 if (loops > 10)
 360                                         break;
 361                         }
 362                         /* uncomment below line to test with mutex */
 363                         /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
 364                         if (loops) {
 365                                 fprintf(stderr,
 366                                         "page_nr %lu all zero thread %lu %p %lu\n",
 367                                         page_nr, cpu, area_dst + page_nr * page_size,
 368                                         loops);
 369                                 if (loops > 10)
 370                                         exit(1);
 371                         }
 372 #endif
 373                 }
 374
 375                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 376                 count = *area_count(area_dst, page_nr);
 377                 if (count != count_verify[page_nr]) {
 378                         fprintf(stderr,
 379                                 "page_nr %lu memory corruption %Lu %Lu\n",
 380                                 page_nr, count,
 381                                 count_verify[page_nr]), exit(1);
 382                 }
 383                 count++;
 384                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 385                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 386
 387                 if (time(NULL) - start > 1)
 388                         fprintf(stderr,
 389                                 "userfault too slow %ld "
 390                                 "possible false positive with overcommit\n",
 391                                 time(NULL) - start);
 392         }
 393
 394         return NULL;
 395 }
 396
 397 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 398                             unsigned long offset)
 399 {
 400         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 401                                      uffdio_copy->len,
 402                                      offset);
 403         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 404                 /* real retval in ufdio_copy.copy */
 405                 if (uffdio_copy->copy != -EEXIST)
 406                         fprintf(stderr, "UFFDIO_COPY retry error %Ld\n",
 407                                 uffdio_copy->copy), exit(1);
 408         } else {
 409                 fprintf(stderr, "UFFDIO_COPY retry unexpected %Ld\n",
 410                         uffdio_copy->copy), exit(1);
 411         }
 412 }
 413
 414 static int __copy_page(int ufd, unsigned long offset, bool retry)
 415 {
 416         struct uffdio_copy uffdio_copy;
 417
 418         if (offset >= nr_pages * page_size)
 419                 fprintf(stderr, "unexpected offset %lu\n",
 420                         offset), exit(1);
 421         uffdio_copy.dst = (unsigned long) area_dst + offset;
 422         uffdio_copy.src = (unsigned long) area_src + offset;
 423         uffdio_copy.len = page_size;
 424         uffdio_copy.mode = 0;
 425         uffdio_copy.copy = 0;
 426         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 427                 /* real retval in ufdio_copy.copy */
 428                 if (uffdio_copy.copy != -EEXIST)
 429                         fprintf(stderr, "UFFDIO_COPY error %Ld\n",
 430                                 uffdio_copy.copy), exit(1);
 431         } else if (uffdio_copy.copy != page_size) {
 432                 fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
 433                         uffdio_copy.copy), exit(1);
 434         } else {
 435                 if (test_uffdio_copy_eexist && retry) {
 436                         test_uffdio_copy_eexist = false;
 437                         retry_copy_page(ufd, &uffdio_copy, offset);
 438                 }
 439                 return 1;
 440         }
 441         return 0;
 442 }
 443
 444 static int copy_page_retry(int ufd, unsigned long offset)
 445 {
 446         return __copy_page(ufd, offset, true);
 447 }
 448
 449 static int copy_page(int ufd, unsigned long offset)
 450 {
 451         return __copy_page(ufd, offset, false);
 452 }
 453
 454 static void *uffd_poll_thread(void *arg)
 455 {
 456         unsigned long cpu = (unsigned long) arg;
 457         struct pollfd pollfd[2];
 458         struct uffd_msg msg;
 459         struct uffdio_register uffd_reg;
 460         int ret;
 461         unsigned long offset;
 462         char tmp_chr;
 463         unsigned long userfaults = 0;
 464
 465         pollfd[0].fd = uffd;
 466         pollfd[0].events = POLLIN;
 467         pollfd[1].fd = pipefd[cpu*2];
 468         pollfd[1].events = POLLIN;
 469
 470         for (;;) {
 471                 ret = poll(pollfd, 2, -1);
 472                 if (!ret)
 473                         fprintf(stderr, "poll error %d\n", ret), exit(1);
 474                 if (ret < 0)
 475                         perror("poll"), exit(1);
 476                 if (pollfd[1].revents & POLLIN) {
 477                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
 478                                 fprintf(stderr, "read pipefd error\n"),
 479                                         exit(1);
 480                         break;
 481                 }
 482                 if (!(pollfd[0].revents & POLLIN))
 483                         fprintf(stderr, "pollfd[0].revents %d\n",
 484                                 pollfd[0].revents), exit(1);
 485                 ret = read(uffd, &msg, sizeof(msg));
 486                 if (ret < 0) {
 487                         if (errno == EAGAIN)
 488                                 continue;
 489                         perror("nonblocking read error"), exit(1);
 490                 }
 491                 switch (msg.event) {
 492                 default:
 493                         fprintf(stderr, "unexpected msg event %u\n",
 494                                 msg.event), exit(1);
 495                         break;
 496                 case UFFD_EVENT_PAGEFAULT:
 497                         if (msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 498                                 fprintf(stderr, "unexpected write fault\n"), exit(1);
 499                         offset = (char *)(unsigned long)msg.arg.pagefault.address -
 500                                 area_dst;
 501                         offset &= ~(page_size-1);
 502                         if (copy_page(uffd, offset))
 503                                 userfaults++;
 504                         break;
 505                 case UFFD_EVENT_FORK:
 506                         close(uffd);
 507                         uffd = msg.arg.fork.ufd;
 508                         pollfd[0].fd = uffd;
 509                         break;
 510                 case UFFD_EVENT_REMOVE:
 511                         uffd_reg.range.start = msg.arg.remove.start;
 512                         uffd_reg.range.len = msg.arg.remove.end -
 513                                 msg.arg.remove.start;
 514                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
 515                                 fprintf(stderr, "remove failure\n"), exit(1);
 516                         break;
 517                 case UFFD_EVENT_REMAP:
 518                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 519                         break;
 520                 }
 521         }
 522         return (void *)userfaults;
 523 }
 524
 525 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 526
 527 static void *uffd_read_thread(void *arg)
 528 {
 529         unsigned long *this_cpu_userfaults;
 530         struct uffd_msg msg;
 531         unsigned long offset;
 532         int ret;
 533
 534         this_cpu_userfaults = (unsigned long *) arg;
 535         *this_cpu_userfaults = 0;
 536
 537         pthread_mutex_unlock(&uffd_read_mutex);
 538         /* from here cancellation is ok */
 539
 540         for (;;) {
 541                 ret = read(uffd, &msg, sizeof(msg));
 542                 if (ret != sizeof(msg)) {
 543                         if (ret < 0)
 544                                 perror("blocking read error"), exit(1);
 545                         else
 546                                 fprintf(stderr, "short read\n"), exit(1);
 547                 }
 548                 if (msg.event != UFFD_EVENT_PAGEFAULT)
 549                         fprintf(stderr, "unexpected msg event %u\n",
 550                                 msg.event), exit(1);
 551                 if (bounces & BOUNCE_VERIFY &&
 552                     msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 553                         fprintf(stderr, "unexpected write fault\n"), exit(1);
 554                 offset = (char *)(unsigned long)msg.arg.pagefault.address -
 555                          area_dst;
 556                 offset &= ~(page_size-1);
 557                 if (copy_page(uffd, offset))
 558                         (*this_cpu_userfaults)++;
 559         }
 560         return (void *)NULL;
 561 }
 562
 563 static void *background_thread(void *arg)
 564 {
 565         unsigned long cpu = (unsigned long) arg;
 566         unsigned long page_nr;
 567
 568         for (page_nr = cpu * nr_pages_per_cpu;
 569              page_nr < (cpu+1) * nr_pages_per_cpu;
 570              page_nr++)
 571                 copy_page_retry(uffd, page_nr * page_size);
 572
 573         return NULL;
 574 }
 575
 576 static int stress(unsigned long *userfaults)
 577 {
 578         unsigned long cpu;
 579         pthread_t locking_threads[nr_cpus];
 580         pthread_t uffd_threads[nr_cpus];
 581         pthread_t background_threads[nr_cpus];
 582         void **_userfaults = (void **) userfaults;
 583
 584         finished = 0;
 585         for (cpu = 0; cpu < nr_cpus; cpu++) {
 586                 if (pthread_create(&locking_threads[cpu], &attr,
 587                                    locking_thread, (void *)cpu))
 588                         return 1;
 589                 if (bounces & BOUNCE_POLL) {
 590                         if (pthread_create(&uffd_threads[cpu], &attr,
 591                                            uffd_poll_thread, (void *)cpu))
 592                                 return 1;
 593                 } else {
 594                         if (pthread_create(&uffd_threads[cpu], &attr,
 595                                            uffd_read_thread,
 596                                            &_userfaults[cpu]))
 597                                 return 1;
 598                         pthread_mutex_lock(&uffd_read_mutex);
 599                 }
 600                 if (pthread_create(&background_threads[cpu], &attr,
 601                                    background_thread, (void *)cpu))
 602                         return 1;
 603         }
 604         for (cpu = 0; cpu < nr_cpus; cpu++)
 605                 if (pthread_join(background_threads[cpu], NULL))
 606                         return 1;
 607
 608         /*
 609          * Be strict and immediately zap area_src, the whole area has
 610          * been transferred already by the background treads. The
 611          * area_src could then be faulted in in a racy way by still
 612          * running uffdio_threads reading zeropages after we zapped
 613          * area_src (but they're guaranteed to get -EEXIST from
 614          * UFFDIO_COPY without writing zero pages into area_dst
 615          * because the background threads already completed).
 616          */
 617         if (uffd_test_ops->release_pages(area_src))
 618                 return 1;
 619
 620         for (cpu = 0; cpu < nr_cpus; cpu++) {
 621                 char c;
 622                 if (bounces & BOUNCE_POLL) {
 623                         if (write(pipefd[cpu*2+1], &c, 1) != 1) {
 624                                 fprintf(stderr, "pipefd write error\n");
 625                                 return 1;
 626                         }
 627                         if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
 628                                 return 1;
 629                 } else {
 630                         if (pthread_cancel(uffd_threads[cpu]))
 631                                 return 1;
 632                         if (pthread_join(uffd_threads[cpu], NULL))
 633                                 return 1;
 634                 }
 635         }
 636
 637         finished = 1;
 638         for (cpu = 0; cpu < nr_cpus; cpu++)
 639                 if (pthread_join(locking_threads[cpu], NULL))
 640                         return 1;
 641
 642         return 0;
 643 }
 644
 645 static int userfaultfd_open(int features)
 646 {
 647         struct uffdio_api uffdio_api;
 648
 649         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 650         if (uffd < 0) {
 651                 fprintf(stderr,
 652                         "userfaultfd syscall not available in this kernel\n");
 653                 return 1;
 654         }
 655         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 656
 657         uffdio_api.api = UFFD_API;
 658         uffdio_api.features = features;
 659         if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
 660                 fprintf(stderr, "UFFDIO_API\n");
 661                 return 1;
 662         }
 663         if (uffdio_api.api != UFFD_API) {
 664                 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
 665                 return 1;
 666         }
 667
 668         return 0;
 669 }
 670
 671 sigjmp_buf jbuf, *sigbuf;
 672
 673 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 674 {
 675         if (sig == SIGBUS) {
 676                 if (sigbuf)
 677                         siglongjmp(*sigbuf, 1);
 678                 abort();
 679         }
 680 }
 681
 682 /*
 683  * For non-cooperative userfaultfd test we fork() a process that will
 684  * generate pagefaults, will mremap the area monitored by the
 685  * userfaultfd and at last this process will release the monitored
 686  * area.
 687  * For the anonymous and shared memory the area is divided into two
 688  * parts, the first part is accessed before mremap, and the second
 689  * part is accessed after mremap. Since hugetlbfs does not support
 690  * mremap, the entire monitored area is accessed in a single pass for
 691  * HUGETLB_TEST.
 692  * The release of the pages currently generates event for shmem and
 693  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 694  * for hugetlb.
 695  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 696  * monitored area, generate pagefaults and test that signal is delivered.
 697  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 698  * test robustness use case - we release monitored area, fork a process
 699  * that will generate pagefaults and verify signal is generated.
 700  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 701  * feature. Using monitor thread, verify no userfault events are generated.
 702  */
 703 static int faulting_process(int signal_test)
 704 {
 705         unsigned long nr;
 706         unsigned long long count;
 707         unsigned long split_nr_pages;
 708         unsigned long lastnr;
 709         struct sigaction act;
 710         unsigned long signalled = 0;
 711
 712         if (test_type != TEST_HUGETLB)
 713                 split_nr_pages = (nr_pages + 1) / 2;
 714         else
 715                 split_nr_pages = nr_pages;
 716
 717         if (signal_test) {
 718                 sigbuf = &jbuf;
 719                 memset(&act, 0, sizeof(act));
 720                 act.sa_sigaction = sighndl;
 721                 act.sa_flags = SA_SIGINFO;
 722                 if (sigaction(SIGBUS, &act, 0)) {
 723                         perror("sigaction");
 724                         return 1;
 725                 }
 726                 lastnr = (unsigned long)-1;
 727         }
 728
 729         for (nr = 0; nr < split_nr_pages; nr++) {
 730                 if (signal_test) {
 731                         if (sigsetjmp(*sigbuf, 1) != 0) {
 732                                 if (nr == lastnr) {
 733                                         fprintf(stderr, "Signal repeated\n");
 734                                         return 1;
 735                                 }
 736
 737                                 lastnr = nr;
 738                                 if (signal_test == 1) {
 739                                         if (copy_page(uffd, nr * page_size))
 740                                                 signalled++;
 741                                 } else {
 742                                         signalled++;
 743                                         continue;
 744                                 }
 745                         }
 746                 }
 747
 748                 count = *area_count(area_dst, nr);
 749                 if (count != count_verify[nr]) {
 750                         fprintf(stderr,
 751                                 "nr %lu memory corruption %Lu %Lu\n",
 752                                 nr, count,
 753                                 count_verify[nr]), exit(1);
 754                 }
 755         }
 756
 757         if (signal_test)
 758                 return signalled != split_nr_pages;
 759
 760         if (test_type == TEST_HUGETLB)
 761                 return 0;
 762
 763         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 764                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 765         if (area_dst == MAP_FAILED)
 766                 perror("mremap"), exit(1);
 767
 768         for (; nr < nr_pages; nr++) {
 769                 count = *area_count(area_dst, nr);
 770                 if (count != count_verify[nr]) {
 771                         fprintf(stderr,
 772                                 "nr %lu memory corruption %Lu %Lu\n",
 773                                 nr, count,
 774                                 count_verify[nr]), exit(1);
 775                 }
 776         }
 777
 778         if (uffd_test_ops->release_pages(area_dst))
 779                 return 1;
 780
 781         for (nr = 0; nr < nr_pages; nr++) {
 782                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
 783                         fprintf(stderr, "nr %lu is not zero\n", nr), exit(1);
 784         }
 785
 786         return 0;
 787 }
 788
 789 static void retry_uffdio_zeropage(int ufd,
 790                                   struct uffdio_zeropage *uffdio_zeropage,
 791                                   unsigned long offset)
 792 {
 793         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
 794                                      uffdio_zeropage->range.len,
 795                                      offset);
 796         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
 797                 if (uffdio_zeropage->zeropage != -EEXIST)
 798                         fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
 799                                 uffdio_zeropage->zeropage), exit(1);
 800         } else {
 801                 fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
 802                         uffdio_zeropage->zeropage), exit(1);
 803         }
 804 }
 805
 806 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
 807 {
 808         struct uffdio_zeropage uffdio_zeropage;
 809         int ret;
 810         unsigned long has_zeropage;
 811
 812         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
 813
 814         if (offset >= nr_pages * page_size)
 815                 fprintf(stderr, "unexpected offset %lu\n",
 816                         offset), exit(1);
 817         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
 818         uffdio_zeropage.range.len = page_size;
 819         uffdio_zeropage.mode = 0;
 820         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
 821         if (ret) {
 822                 /* real retval in ufdio_zeropage.zeropage */
 823                 if (has_zeropage) {
 824                         if (uffdio_zeropage.zeropage == -EEXIST)
 825                                 fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n"),
 826                                         exit(1);
 827                         else
 828                                 fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
 829                                         uffdio_zeropage.zeropage), exit(1);
 830                 } else {
 831                         if (uffdio_zeropage.zeropage != -EINVAL)
 832                                 fprintf(stderr,
 833                                         "UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
 834                                         uffdio_zeropage.zeropage), exit(1);
 835                 }
 836         } else if (has_zeropage) {
 837                 if (uffdio_zeropage.zeropage != page_size) {
 838                         fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
 839                                 uffdio_zeropage.zeropage), exit(1);
 840                 } else {
 841                         if (test_uffdio_zeropage_eexist && retry) {
 842                                 test_uffdio_zeropage_eexist = false;
 843                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
 844                                                       offset);
 845                         }
 846                         return 1;
 847                 }
 848         } else {
 849                 fprintf(stderr,
 850                         "UFFDIO_ZEROPAGE succeeded %Ld\n",
 851                         uffdio_zeropage.zeropage), exit(1);
 852         }
 853
 854         return 0;
 855 }
 856
 857 static int uffdio_zeropage(int ufd, unsigned long offset)
 858 {
 859         return __uffdio_zeropage(ufd, offset, false);
 860 }
 861
 862 /* exercise UFFDIO_ZEROPAGE */
 863 static int userfaultfd_zeropage_test(void)
 864 {
 865         struct uffdio_register uffdio_register;
 866         unsigned long expected_ioctls;
 867
 868         printf("testing UFFDIO_ZEROPAGE: ");
 869         fflush(stdout);
 870
 871         if (uffd_test_ops->release_pages(area_dst))
 872                 return 1;
 873
 874         if (userfaultfd_open(0) < 0)
 875                 return 1;
 876         uffdio_register.range.start = (unsigned long) area_dst;
 877         uffdio_register.range.len = nr_pages * page_size;
 878         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 879         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 880                 fprintf(stderr, "register failure\n"), exit(1);
 881
 882         expected_ioctls = uffd_test_ops->expected_ioctls;
 883         if ((uffdio_register.ioctls & expected_ioctls) !=
 884             expected_ioctls)
 885                 fprintf(stderr,
 886                         "unexpected missing ioctl for anon memory\n"),
 887                         exit(1);
 888
 889         if (uffdio_zeropage(uffd, 0)) {
 890                 if (my_bcmp(area_dst, zeropage, page_size))
 891                         fprintf(stderr, "zeropage is not zero\n"), exit(1);
 892         }
 893
 894         close(uffd);
 895         printf("done.\n");
 896         return 0;
 897 }
 898
 899 static int userfaultfd_events_test(void)
 900 {
 901         struct uffdio_register uffdio_register;
 902         unsigned long expected_ioctls;
 903         unsigned long userfaults;
 904         pthread_t uffd_mon;
 905         int err, features;
 906         pid_t pid;
 907         char c;
 908
 909         printf("testing events (fork, remap, remove): ");
 910         fflush(stdout);
 911
 912         if (uffd_test_ops->release_pages(area_dst))
 913                 return 1;
 914
 915         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
 916                 UFFD_FEATURE_EVENT_REMOVE;
 917         if (userfaultfd_open(features) < 0)
 918                 return 1;
 919         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 920
 921         uffdio_register.range.start = (unsigned long) area_dst;
 922         uffdio_register.range.len = nr_pages * page_size;
 923         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 924         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 925                 fprintf(stderr, "register failure\n"), exit(1);
 926
 927         expected_ioctls = uffd_test_ops->expected_ioctls;
 928         if ((uffdio_register.ioctls & expected_ioctls) !=
 929             expected_ioctls)
 930                 fprintf(stderr,
 931                         "unexpected missing ioctl for anon memory\n"),
 932                         exit(1);
 933
 934         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
 935                 perror("uffd_poll_thread create"), exit(1);
 936
 937         pid = fork();
 938         if (pid < 0)
 939                 perror("fork"), exit(1);
 940
 941         if (!pid)
 942                 return faulting_process(0);
 943
 944         waitpid(pid, &err, 0);
 945         if (err)
 946                 fprintf(stderr, "faulting process failed\n"), exit(1);
 947
 948         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
 949                 perror("pipe write"), exit(1);
 950         if (pthread_join(uffd_mon, (void **)&userfaults))
 951                 return 1;
 952
 953         close(uffd);
 954         printf("userfaults: %ld\n", userfaults);
 955
 956         return userfaults != nr_pages;
 957 }
 958
 959 static int userfaultfd_sig_test(void)
 960 {
 961         struct uffdio_register uffdio_register;
 962         unsigned long expected_ioctls;
 963         unsigned long userfaults;
 964         pthread_t uffd_mon;
 965         int err, features;
 966         pid_t pid;
 967         char c;
 968
 969         printf("testing signal delivery: ");
 970         fflush(stdout);
 971
 972         if (uffd_test_ops->release_pages(area_dst))
 973                 return 1;
 974
 975         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
 976         if (userfaultfd_open(features) < 0)
 977                 return 1;
 978         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 979
 980         uffdio_register.range.start = (unsigned long) area_dst;
 981         uffdio_register.range.len = nr_pages * page_size;
 982         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 983         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 984                 fprintf(stderr, "register failure\n"), exit(1);
 985
 986         expected_ioctls = uffd_test_ops->expected_ioctls;
 987         if ((uffdio_register.ioctls & expected_ioctls) !=
 988             expected_ioctls)
 989                 fprintf(stderr,
 990                         "unexpected missing ioctl for anon memory\n"),
 991                         exit(1);
 992
 993         if (faulting_process(1))
 994                 fprintf(stderr, "faulting process failed\n"), exit(1);
 995
 996         if (uffd_test_ops->release_pages(area_dst))
 997                 return 1;
 998
 999         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
1000                 perror("uffd_poll_thread create"), exit(1);
1001
1002         pid = fork();
1003         if (pid < 0)
1004                 perror("fork"), exit(1);
1005
1006         if (!pid)
1007                 exit(faulting_process(2));
1008
1009         waitpid(pid, &err, 0);
1010         if (err)
1011                 fprintf(stderr, "faulting process failed\n"), exit(1);
1012
1013         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1014                 perror("pipe write"), exit(1);
1015         if (pthread_join(uffd_mon, (void **)&userfaults))
1016                 return 1;
1017
1018         printf("done.\n");
1019         if (userfaults)
1020                 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1021                         userfaults);
1022         close(uffd);
1023         return userfaults != 0;
1024 }
1025 static int userfaultfd_stress(void)
1026 {
1027         void *area;
1028         char *tmp_area;
1029         unsigned long nr;
1030         struct uffdio_register uffdio_register;
1031         unsigned long cpu;
1032         int err;
1033         unsigned long userfaults[nr_cpus];
1034
1035         uffd_test_ops->allocate_area((void **)&area_src);
1036         if (!area_src)
1037                 return 1;
1038         uffd_test_ops->allocate_area((void **)&area_dst);
1039         if (!area_dst)
1040                 return 1;
1041
1042         if (userfaultfd_open(0) < 0)
1043                 return 1;
1044
1045         count_verify = malloc(nr_pages * sizeof(unsigned long long));
1046         if (!count_verify) {
1047                 perror("count_verify");
1048                 return 1;
1049         }
1050
1051         for (nr = 0; nr < nr_pages; nr++) {
1052                 *area_mutex(area_src, nr) = (pthread_mutex_t)
1053                         PTHREAD_MUTEX_INITIALIZER;
1054                 count_verify[nr] = *area_count(area_src, nr) = 1;
1055                 /*
1056                  * In the transition between 255 to 256, powerpc will
1057                  * read out of order in my_bcmp and see both bytes as
1058                  * zero, so leave a placeholder below always non-zero
1059                  * after the count, to avoid my_bcmp to trigger false
1060                  * positives.
1061                  */
1062                 *(area_count(area_src, nr) + 1) = 1;
1063         }
1064
1065         pipefd = malloc(sizeof(int) * nr_cpus * 2);
1066         if (!pipefd) {
1067                 perror("pipefd");
1068                 return 1;
1069         }
1070         for (cpu = 0; cpu < nr_cpus; cpu++) {
1071                 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1072                         perror("pipe");
1073                         return 1;
1074                 }
1075         }
1076
1077         if (posix_memalign(&area, page_size, page_size)) {
1078                 fprintf(stderr, "out of memory\n");
1079                 return 1;
1080         }
1081         zeropage = area;
1082         bzero(zeropage, page_size);
1083
1084         pthread_mutex_lock(&uffd_read_mutex);
1085
1086         pthread_attr_init(&attr);
1087         pthread_attr_setstacksize(&attr, 16*1024*1024);
1088
1089         err = 0;
1090         while (bounces--) {
1091                 unsigned long expected_ioctls;
1092
1093                 printf("bounces: %d, mode:", bounces);
1094                 if (bounces & BOUNCE_RANDOM)
1095                         printf(" rnd");
1096                 if (bounces & BOUNCE_RACINGFAULTS)
1097                         printf(" racing");
1098                 if (bounces & BOUNCE_VERIFY)
1099                         printf(" ver");
1100                 if (bounces & BOUNCE_POLL)
1101                         printf(" poll");
1102                 printf(", ");
1103                 fflush(stdout);
1104
1105                 if (bounces & BOUNCE_POLL)
1106                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1107                 else
1108                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1109
1110                 /* register */
1111                 uffdio_register.range.start = (unsigned long) area_dst;
1112                 uffdio_register.range.len = nr_pages * page_size;
1113                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1114                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1115                         fprintf(stderr, "register failure\n");
1116                         return 1;
1117                 }
1118                 expected_ioctls = uffd_test_ops->expected_ioctls;
1119                 if ((uffdio_register.ioctls & expected_ioctls) !=
1120                     expected_ioctls) {
1121                         fprintf(stderr,
1122                                 "unexpected missing ioctl for anon memory\n");
1123                         return 1;
1124                 }
1125
1126                 if (area_dst_alias) {
1127                         uffdio_register.range.start = (unsigned long)
1128                                 area_dst_alias;
1129                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1130                                 fprintf(stderr, "register failure alias\n");
1131                                 return 1;
1132                         }
1133                 }
1134
1135                 /*
1136                  * The madvise done previously isn't enough: some
1137                  * uffd_thread could have read userfaults (one of
1138                  * those already resolved by the background thread)
1139                  * and it may be in the process of calling
1140                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1141                  * area_src and it would map a zero page in it (of
1142                  * course such a UFFDIO_COPY is perfectly safe as it'd
1143                  * return -EEXIST). The problem comes at the next
1144                  * bounce though: that racing UFFDIO_COPY would
1145                  * generate zeropages in the area_src, so invalidating
1146                  * the previous MADV_DONTNEED. Without this additional
1147                  * MADV_DONTNEED those zeropages leftovers in the
1148                  * area_src would lead to -EEXIST failure during the
1149                  * next bounce, effectively leaving a zeropage in the
1150                  * area_dst.
1151                  *
1152                  * Try to comment this out madvise to see the memory
1153                  * corruption being caught pretty quick.
1154                  *
1155                  * khugepaged is also inhibited to collapse THP after
1156                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1157                  * required to MADV_DONTNEED here.
1158                  */
1159                 if (uffd_test_ops->release_pages(area_dst))
1160                         return 1;
1161
1162                 /* bounce pass */
1163                 if (stress(userfaults))
1164                         return 1;
1165
1166                 /* unregister */
1167                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1168                         fprintf(stderr, "unregister failure\n");
1169                         return 1;
1170                 }
1171                 if (area_dst_alias) {
1172                         uffdio_register.range.start = (unsigned long) area_dst;
1173                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1174                                   &uffdio_register.range)) {
1175                                 fprintf(stderr, "unregister failure alias\n");
1176                                 return 1;
1177                         }
1178                 }
1179
1180                 /* verification */
1181                 if (bounces & BOUNCE_VERIFY) {
1182                         for (nr = 0; nr < nr_pages; nr++) {
1183                                 if (*area_count(area_dst, nr) != count_verify[nr]) {
1184                                         fprintf(stderr,
1185                                                 "error area_count %Lu %Lu %lu\n",
1186                                                 *area_count(area_src, nr),
1187                                                 count_verify[nr],
1188                                                 nr);
1189                                         err = 1;
1190                                         bounces = 0;
1191                                 }
1192                         }
1193                 }
1194
1195                 /* prepare next bounce */
1196                 tmp_area = area_src;
1197                 area_src = area_dst;
1198                 area_dst = tmp_area;
1199
1200                 tmp_area = area_src_alias;
1201                 area_src_alias = area_dst_alias;
1202                 area_dst_alias = tmp_area;
1203
1204                 printf("userfaults:");
1205                 for (cpu = 0; cpu < nr_cpus; cpu++)
1206                         printf(" %lu", userfaults[cpu]);
1207                 printf("\n");
1208         }
1209
1210         if (err)
1211                 return err;
1212
1213         close(uffd);
1214         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1215                 || userfaultfd_events_test();
1216 }
1217
1218 /*
1219  * Copied from mlock2-tests.c
1220  */
1221 unsigned long default_huge_page_size(void)
1222 {
1223         unsigned long hps = 0;
1224         char *line = NULL;
1225         size_t linelen = 0;
1226         FILE *f = fopen("/proc/meminfo", "r");
1227
1228         if (!f)
1229                 return 0;
1230         while (getline(&line, &linelen, f) > 0) {
1231                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1232                         hps <<= 10;
1233                         break;
1234                 }
1235         }
1236
1237         free(line);
1238         fclose(f);
1239         return hps;
1240 }
1241
1242 static void set_test_type(const char *type)
1243 {
1244         if (!strcmp(type, "anon")) {
1245                 test_type = TEST_ANON;
1246                 uffd_test_ops = &anon_uffd_test_ops;
1247         } else if (!strcmp(type, "hugetlb")) {
1248                 test_type = TEST_HUGETLB;
1249                 uffd_test_ops = &hugetlb_uffd_test_ops;
1250         } else if (!strcmp(type, "hugetlb_shared")) {
1251                 map_shared = true;
1252                 test_type = TEST_HUGETLB;
1253                 uffd_test_ops = &hugetlb_uffd_test_ops;
1254         } else if (!strcmp(type, "shmem")) {
1255                 map_shared = true;
1256                 test_type = TEST_SHMEM;
1257                 uffd_test_ops = &shmem_uffd_test_ops;
1258         } else {
1259                 fprintf(stderr, "Unknown test type: %s\n", type), exit(1);
1260         }
1261
1262         if (test_type == TEST_HUGETLB)
1263                 page_size = default_huge_page_size();
1264         else
1265                 page_size = sysconf(_SC_PAGE_SIZE);
1266
1267         if (!page_size)
1268                 fprintf(stderr, "Unable to determine page size\n"),
1269                                 exit(2);
1270         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1271             > page_size)
1272                 fprintf(stderr, "Impossible to run this test\n"), exit(2);
1273 }
1274
1275 static void sigalrm(int sig)
1276 {
1277         if (sig != SIGALRM)
1278                 abort();
1279         test_uffdio_copy_eexist = true;
1280         test_uffdio_zeropage_eexist = true;
1281         alarm(ALARM_INTERVAL_SECS);
1282 }
1283
1284 int main(int argc, char **argv)
1285 {
1286         if (argc < 4)
1287                 usage();
1288
1289         if (signal(SIGALRM, sigalrm) == SIG_ERR)
1290                 fprintf(stderr, "failed to arm SIGALRM"), exit(1);
1291         alarm(ALARM_INTERVAL_SECS);
1292
1293         set_test_type(argv[1]);
1294
1295         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1296         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1297                 nr_cpus;
1298         if (!nr_pages_per_cpu) {
1299                 fprintf(stderr, "invalid MiB\n");
1300                 usage();
1301         }
1302
1303         bounces = atoi(argv[3]);
1304         if (bounces <= 0) {
1305                 fprintf(stderr, "invalid bounces\n");
1306                 usage();
1307         }
1308         nr_pages = nr_pages_per_cpu * nr_cpus;
1309
1310         if (test_type == TEST_HUGETLB) {
1311                 if (argc < 5)
1312                         usage();
1313                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1314                 if (huge_fd < 0) {
1315                         fprintf(stderr, "Open of %s failed", argv[3]);
1316                         perror("open");
1317                         exit(1);
1318                 }
1319                 if (ftruncate(huge_fd, 0)) {
1320                         fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1321                         perror("ftruncate");
1322                         exit(1);
1323                 }
1324         }
1325         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1326                nr_pages, nr_pages_per_cpu);
1327         return userfaultfd_stress();
1328 }
1329
1330 #else /* __NR_userfaultfd */
1331
1332 #warning "missing __NR_userfaultfd definition"
1333
1334 int main(void)
1335 {
1336         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1337         return KSFT_SKIP;
1338 }
1339
1340 #endif /* __NR_userfaultfd */