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, "%s", 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 int uffd_read_msg(int ufd, struct uffd_msg *msg)
 455 {
 456         int ret = read(uffd, msg, sizeof(*msg));
 457
 458         if (ret != sizeof(*msg)) {
 459                 if (ret < 0) {
 460                         if (errno == EAGAIN)
 461                                 return 1;
 462                         else
 463                                 perror("blocking read error"), exit(1);
 464                 } else {
 465                         fprintf(stderr, "short read\n"), exit(1);
 466                 }
 467         }
 468
 469         return 0;
 470 }
 471
 472 /* Return 1 if page fault handled by us; otherwise 0 */
 473 static int uffd_handle_page_fault(struct uffd_msg *msg)
 474 {
 475         unsigned long offset;
 476
 477         if (msg->event != UFFD_EVENT_PAGEFAULT)
 478                 fprintf(stderr, "unexpected msg event %u\n",
 479                         msg->event), exit(1);
 480
 481         if (bounces & BOUNCE_VERIFY &&
 482             msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 483                 fprintf(stderr, "unexpected write fault\n"), exit(1);
 484
 485         offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
 486         offset &= ~(page_size-1);
 487
 488         return copy_page(uffd, offset);
 489 }
 490
 491 static void *uffd_poll_thread(void *arg)
 492 {
 493         unsigned long cpu = (unsigned long) arg;
 494         struct pollfd pollfd[2];
 495         struct uffd_msg msg;
 496         struct uffdio_register uffd_reg;
 497         int ret;
 498         char tmp_chr;
 499         unsigned long userfaults = 0;
 500
 501         pollfd[0].fd = uffd;
 502         pollfd[0].events = POLLIN;
 503         pollfd[1].fd = pipefd[cpu*2];
 504         pollfd[1].events = POLLIN;
 505
 506         for (;;) {
 507                 ret = poll(pollfd, 2, -1);
 508                 if (!ret)
 509                         fprintf(stderr, "poll error %d\n", ret), exit(1);
 510                 if (ret < 0)
 511                         perror("poll"), exit(1);
 512                 if (pollfd[1].revents & POLLIN) {
 513                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
 514                                 fprintf(stderr, "read pipefd error\n"),
 515                                         exit(1);
 516                         break;
 517                 }
 518                 if (!(pollfd[0].revents & POLLIN))
 519                         fprintf(stderr, "pollfd[0].revents %d\n",
 520                                 pollfd[0].revents), exit(1);
 521                 if (uffd_read_msg(uffd, &msg))
 522                         continue;
 523                 switch (msg.event) {
 524                 default:
 525                         fprintf(stderr, "unexpected msg event %u\n",
 526                                 msg.event), exit(1);
 527                         break;
 528                 case UFFD_EVENT_PAGEFAULT:
 529                         userfaults += uffd_handle_page_fault(&msg);
 530                         break;
 531                 case UFFD_EVENT_FORK:
 532                         close(uffd);
 533                         uffd = msg.arg.fork.ufd;
 534                         pollfd[0].fd = uffd;
 535                         break;
 536                 case UFFD_EVENT_REMOVE:
 537                         uffd_reg.range.start = msg.arg.remove.start;
 538                         uffd_reg.range.len = msg.arg.remove.end -
 539                                 msg.arg.remove.start;
 540                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
 541                                 fprintf(stderr, "remove failure\n"), exit(1);
 542                         break;
 543                 case UFFD_EVENT_REMAP:
 544                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 545                         break;
 546                 }
 547         }
 548         return (void *)userfaults;
 549 }
 550
 551 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 552
 553 static void *uffd_read_thread(void *arg)
 554 {
 555         unsigned long *this_cpu_userfaults;
 556         struct uffd_msg msg;
 557
 558         this_cpu_userfaults = (unsigned long *) arg;
 559         *this_cpu_userfaults = 0;
 560
 561         pthread_mutex_unlock(&uffd_read_mutex);
 562         /* from here cancellation is ok */
 563
 564         for (;;) {
 565                 if (uffd_read_msg(uffd, &msg))
 566                         continue;
 567                 (*this_cpu_userfaults) += uffd_handle_page_fault(&msg);
 568         }
 569         return (void *)NULL;
 570 }
 571
 572 static void *background_thread(void *arg)
 573 {
 574         unsigned long cpu = (unsigned long) arg;
 575         unsigned long page_nr;
 576
 577         for (page_nr = cpu * nr_pages_per_cpu;
 578              page_nr < (cpu+1) * nr_pages_per_cpu;
 579              page_nr++)
 580                 copy_page_retry(uffd, page_nr * page_size);
 581
 582         return NULL;
 583 }
 584
 585 static int stress(unsigned long *userfaults)
 586 {
 587         unsigned long cpu;
 588         pthread_t locking_threads[nr_cpus];
 589         pthread_t uffd_threads[nr_cpus];
 590         pthread_t background_threads[nr_cpus];
 591         void **_userfaults = (void **) userfaults;
 592
 593         finished = 0;
 594         for (cpu = 0; cpu < nr_cpus; cpu++) {
 595                 if (pthread_create(&locking_threads[cpu], &attr,
 596                                    locking_thread, (void *)cpu))
 597                         return 1;
 598                 if (bounces & BOUNCE_POLL) {
 599                         if (pthread_create(&uffd_threads[cpu], &attr,
 600                                            uffd_poll_thread, (void *)cpu))
 601                                 return 1;
 602                 } else {
 603                         if (pthread_create(&uffd_threads[cpu], &attr,
 604                                            uffd_read_thread,
 605                                            &_userfaults[cpu]))
 606                                 return 1;
 607                         pthread_mutex_lock(&uffd_read_mutex);
 608                 }
 609                 if (pthread_create(&background_threads[cpu], &attr,
 610                                    background_thread, (void *)cpu))
 611                         return 1;
 612         }
 613         for (cpu = 0; cpu < nr_cpus; cpu++)
 614                 if (pthread_join(background_threads[cpu], NULL))
 615                         return 1;
 616
 617         /*
 618          * Be strict and immediately zap area_src, the whole area has
 619          * been transferred already by the background treads. The
 620          * area_src could then be faulted in in a racy way by still
 621          * running uffdio_threads reading zeropages after we zapped
 622          * area_src (but they're guaranteed to get -EEXIST from
 623          * UFFDIO_COPY without writing zero pages into area_dst
 624          * because the background threads already completed).
 625          */
 626         if (uffd_test_ops->release_pages(area_src))
 627                 return 1;
 628
 629
 630         finished = 1;
 631         for (cpu = 0; cpu < nr_cpus; cpu++)
 632                 if (pthread_join(locking_threads[cpu], NULL))
 633                         return 1;
 634
 635         for (cpu = 0; cpu < nr_cpus; cpu++) {
 636                 char c;
 637                 if (bounces & BOUNCE_POLL) {
 638                         if (write(pipefd[cpu*2+1], &c, 1) != 1) {
 639                                 fprintf(stderr, "pipefd write error\n");
 640                                 return 1;
 641                         }
 642                         if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
 643                                 return 1;
 644                 } else {
 645                         if (pthread_cancel(uffd_threads[cpu]))
 646                                 return 1;
 647                         if (pthread_join(uffd_threads[cpu], NULL))
 648                                 return 1;
 649                 }
 650         }
 651
 652         return 0;
 653 }
 654
 655 static int userfaultfd_open(int features)
 656 {
 657         struct uffdio_api uffdio_api;
 658
 659         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 660         if (uffd < 0) {
 661                 fprintf(stderr,
 662                         "userfaultfd syscall not available in this kernel\n");
 663                 return 1;
 664         }
 665         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 666
 667         uffdio_api.api = UFFD_API;
 668         uffdio_api.features = features;
 669         if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
 670                 fprintf(stderr, "UFFDIO_API\n");
 671                 return 1;
 672         }
 673         if (uffdio_api.api != UFFD_API) {
 674                 fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
 675                 return 1;
 676         }
 677
 678         return 0;
 679 }
 680
 681 sigjmp_buf jbuf, *sigbuf;
 682
 683 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 684 {
 685         if (sig == SIGBUS) {
 686                 if (sigbuf)
 687                         siglongjmp(*sigbuf, 1);
 688                 abort();
 689         }
 690 }
 691
 692 /*
 693  * For non-cooperative userfaultfd test we fork() a process that will
 694  * generate pagefaults, will mremap the area monitored by the
 695  * userfaultfd and at last this process will release the monitored
 696  * area.
 697  * For the anonymous and shared memory the area is divided into two
 698  * parts, the first part is accessed before mremap, and the second
 699  * part is accessed after mremap. Since hugetlbfs does not support
 700  * mremap, the entire monitored area is accessed in a single pass for
 701  * HUGETLB_TEST.
 702  * The release of the pages currently generates event for shmem and
 703  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 704  * for hugetlb.
 705  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 706  * monitored area, generate pagefaults and test that signal is delivered.
 707  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 708  * test robustness use case - we release monitored area, fork a process
 709  * that will generate pagefaults and verify signal is generated.
 710  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 711  * feature. Using monitor thread, verify no userfault events are generated.
 712  */
 713 static int faulting_process(int signal_test)
 714 {
 715         unsigned long nr;
 716         unsigned long long count;
 717         unsigned long split_nr_pages;
 718         unsigned long lastnr;
 719         struct sigaction act;
 720         unsigned long signalled = 0;
 721
 722         if (test_type != TEST_HUGETLB)
 723                 split_nr_pages = (nr_pages + 1) / 2;
 724         else
 725                 split_nr_pages = nr_pages;
 726
 727         if (signal_test) {
 728                 sigbuf = &jbuf;
 729                 memset(&act, 0, sizeof(act));
 730                 act.sa_sigaction = sighndl;
 731                 act.sa_flags = SA_SIGINFO;
 732                 if (sigaction(SIGBUS, &act, 0)) {
 733                         perror("sigaction");
 734                         return 1;
 735                 }
 736                 lastnr = (unsigned long)-1;
 737         }
 738
 739         for (nr = 0; nr < split_nr_pages; nr++) {
 740                 if (signal_test) {
 741                         if (sigsetjmp(*sigbuf, 1) != 0) {
 742                                 if (nr == lastnr) {
 743                                         fprintf(stderr, "Signal repeated\n");
 744                                         return 1;
 745                                 }
 746
 747                                 lastnr = nr;
 748                                 if (signal_test == 1) {
 749                                         if (copy_page(uffd, nr * page_size))
 750                                                 signalled++;
 751                                 } else {
 752                                         signalled++;
 753                                         continue;
 754                                 }
 755                         }
 756                 }
 757
 758                 count = *area_count(area_dst, nr);
 759                 if (count != count_verify[nr]) {
 760                         fprintf(stderr,
 761                                 "nr %lu memory corruption %Lu %Lu\n",
 762                                 nr, count,
 763                                 count_verify[nr]), exit(1);
 764                 }
 765         }
 766
 767         if (signal_test)
 768                 return signalled != split_nr_pages;
 769
 770         if (test_type == TEST_HUGETLB)
 771                 return 0;
 772
 773         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 774                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 775         if (area_dst == MAP_FAILED)
 776                 perror("mremap"), exit(1);
 777
 778         for (; nr < nr_pages; nr++) {
 779                 count = *area_count(area_dst, nr);
 780                 if (count != count_verify[nr]) {
 781                         fprintf(stderr,
 782                                 "nr %lu memory corruption %Lu %Lu\n",
 783                                 nr, count,
 784                                 count_verify[nr]), exit(1);
 785                 }
 786         }
 787
 788         if (uffd_test_ops->release_pages(area_dst))
 789                 return 1;
 790
 791         for (nr = 0; nr < nr_pages; nr++) {
 792                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
 793                         fprintf(stderr, "nr %lu is not zero\n", nr), exit(1);
 794         }
 795
 796         return 0;
 797 }
 798
 799 static void retry_uffdio_zeropage(int ufd,
 800                                   struct uffdio_zeropage *uffdio_zeropage,
 801                                   unsigned long offset)
 802 {
 803         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
 804                                      uffdio_zeropage->range.len,
 805                                      offset);
 806         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
 807                 if (uffdio_zeropage->zeropage != -EEXIST)
 808                         fprintf(stderr, "UFFDIO_ZEROPAGE retry error %Ld\n",
 809                                 uffdio_zeropage->zeropage), exit(1);
 810         } else {
 811                 fprintf(stderr, "UFFDIO_ZEROPAGE retry unexpected %Ld\n",
 812                         uffdio_zeropage->zeropage), exit(1);
 813         }
 814 }
 815
 816 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
 817 {
 818         struct uffdio_zeropage uffdio_zeropage;
 819         int ret;
 820         unsigned long has_zeropage;
 821
 822         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
 823
 824         if (offset >= nr_pages * page_size)
 825                 fprintf(stderr, "unexpected offset %lu\n",
 826                         offset), exit(1);
 827         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
 828         uffdio_zeropage.range.len = page_size;
 829         uffdio_zeropage.mode = 0;
 830         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
 831         if (ret) {
 832                 /* real retval in ufdio_zeropage.zeropage */
 833                 if (has_zeropage) {
 834                         if (uffdio_zeropage.zeropage == -EEXIST)
 835                                 fprintf(stderr, "UFFDIO_ZEROPAGE -EEXIST\n"),
 836                                         exit(1);
 837                         else
 838                                 fprintf(stderr, "UFFDIO_ZEROPAGE error %Ld\n",
 839                                         uffdio_zeropage.zeropage), exit(1);
 840                 } else {
 841                         if (uffdio_zeropage.zeropage != -EINVAL)
 842                                 fprintf(stderr,
 843                                         "UFFDIO_ZEROPAGE not -EINVAL %Ld\n",
 844                                         uffdio_zeropage.zeropage), exit(1);
 845                 }
 846         } else if (has_zeropage) {
 847                 if (uffdio_zeropage.zeropage != page_size) {
 848                         fprintf(stderr, "UFFDIO_ZEROPAGE unexpected %Ld\n",
 849                                 uffdio_zeropage.zeropage), exit(1);
 850                 } else {
 851                         if (test_uffdio_zeropage_eexist && retry) {
 852                                 test_uffdio_zeropage_eexist = false;
 853                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
 854                                                       offset);
 855                         }
 856                         return 1;
 857                 }
 858         } else {
 859                 fprintf(stderr,
 860                         "UFFDIO_ZEROPAGE succeeded %Ld\n",
 861                         uffdio_zeropage.zeropage), exit(1);
 862         }
 863
 864         return 0;
 865 }
 866
 867 static int uffdio_zeropage(int ufd, unsigned long offset)
 868 {
 869         return __uffdio_zeropage(ufd, offset, false);
 870 }
 871
 872 /* exercise UFFDIO_ZEROPAGE */
 873 static int userfaultfd_zeropage_test(void)
 874 {
 875         struct uffdio_register uffdio_register;
 876         unsigned long expected_ioctls;
 877
 878         printf("testing UFFDIO_ZEROPAGE: ");
 879         fflush(stdout);
 880
 881         if (uffd_test_ops->release_pages(area_dst))
 882                 return 1;
 883
 884         if (userfaultfd_open(0) < 0)
 885                 return 1;
 886         uffdio_register.range.start = (unsigned long) area_dst;
 887         uffdio_register.range.len = nr_pages * page_size;
 888         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 889         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 890                 fprintf(stderr, "register failure\n"), exit(1);
 891
 892         expected_ioctls = uffd_test_ops->expected_ioctls;
 893         if ((uffdio_register.ioctls & expected_ioctls) !=
 894             expected_ioctls)
 895                 fprintf(stderr,
 896                         "unexpected missing ioctl for anon memory\n"),
 897                         exit(1);
 898
 899         if (uffdio_zeropage(uffd, 0)) {
 900                 if (my_bcmp(area_dst, zeropage, page_size))
 901                         fprintf(stderr, "zeropage is not zero\n"), exit(1);
 902         }
 903
 904         close(uffd);
 905         printf("done.\n");
 906         return 0;
 907 }
 908
 909 static int userfaultfd_events_test(void)
 910 {
 911         struct uffdio_register uffdio_register;
 912         unsigned long expected_ioctls;
 913         unsigned long userfaults;
 914         pthread_t uffd_mon;
 915         int err, features;
 916         pid_t pid;
 917         char c;
 918
 919         printf("testing events (fork, remap, remove): ");
 920         fflush(stdout);
 921
 922         if (uffd_test_ops->release_pages(area_dst))
 923                 return 1;
 924
 925         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
 926                 UFFD_FEATURE_EVENT_REMOVE;
 927         if (userfaultfd_open(features) < 0)
 928                 return 1;
 929         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 930
 931         uffdio_register.range.start = (unsigned long) area_dst;
 932         uffdio_register.range.len = nr_pages * page_size;
 933         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 934         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 935                 fprintf(stderr, "register failure\n"), exit(1);
 936
 937         expected_ioctls = uffd_test_ops->expected_ioctls;
 938         if ((uffdio_register.ioctls & expected_ioctls) !=
 939             expected_ioctls)
 940                 fprintf(stderr,
 941                         "unexpected missing ioctl for anon memory\n"),
 942                         exit(1);
 943
 944         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
 945                 perror("uffd_poll_thread create"), exit(1);
 946
 947         pid = fork();
 948         if (pid < 0)
 949                 perror("fork"), exit(1);
 950
 951         if (!pid)
 952                 return faulting_process(0);
 953
 954         waitpid(pid, &err, 0);
 955         if (err)
 956                 fprintf(stderr, "faulting process failed\n"), exit(1);
 957
 958         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
 959                 perror("pipe write"), exit(1);
 960         if (pthread_join(uffd_mon, (void **)&userfaults))
 961                 return 1;
 962
 963         close(uffd);
 964         printf("userfaults: %ld\n", userfaults);
 965
 966         return userfaults != nr_pages;
 967 }
 968
 969 static int userfaultfd_sig_test(void)
 970 {
 971         struct uffdio_register uffdio_register;
 972         unsigned long expected_ioctls;
 973         unsigned long userfaults;
 974         pthread_t uffd_mon;
 975         int err, features;
 976         pid_t pid;
 977         char c;
 978
 979         printf("testing signal delivery: ");
 980         fflush(stdout);
 981
 982         if (uffd_test_ops->release_pages(area_dst))
 983                 return 1;
 984
 985         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
 986         if (userfaultfd_open(features) < 0)
 987                 return 1;
 988         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
 989
 990         uffdio_register.range.start = (unsigned long) area_dst;
 991         uffdio_register.range.len = nr_pages * page_size;
 992         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
 993         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
 994                 fprintf(stderr, "register failure\n"), exit(1);
 995
 996         expected_ioctls = uffd_test_ops->expected_ioctls;
 997         if ((uffdio_register.ioctls & expected_ioctls) !=
 998             expected_ioctls)
 999                 fprintf(stderr,
1000                         "unexpected missing ioctl for anon memory\n"),
1001                         exit(1);
1002
1003         if (faulting_process(1))
1004                 fprintf(stderr, "faulting process failed\n"), exit(1);
1005
1006         if (uffd_test_ops->release_pages(area_dst))
1007                 return 1;
1008
1009         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, NULL))
1010                 perror("uffd_poll_thread create"), exit(1);
1011
1012         pid = fork();
1013         if (pid < 0)
1014                 perror("fork"), exit(1);
1015
1016         if (!pid)
1017                 exit(faulting_process(2));
1018
1019         waitpid(pid, &err, 0);
1020         if (err)
1021                 fprintf(stderr, "faulting process failed\n"), exit(1);
1022
1023         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1024                 perror("pipe write"), exit(1);
1025         if (pthread_join(uffd_mon, (void **)&userfaults))
1026                 return 1;
1027
1028         printf("done.\n");
1029         if (userfaults)
1030                 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1031                         userfaults);
1032         close(uffd);
1033         return userfaults != 0;
1034 }
1035 static int userfaultfd_stress(void)
1036 {
1037         void *area;
1038         char *tmp_area;
1039         unsigned long nr;
1040         struct uffdio_register uffdio_register;
1041         unsigned long cpu;
1042         int err;
1043         unsigned long userfaults[nr_cpus];
1044
1045         uffd_test_ops->allocate_area((void **)&area_src);
1046         if (!area_src)
1047                 return 1;
1048         uffd_test_ops->allocate_area((void **)&area_dst);
1049         if (!area_dst)
1050                 return 1;
1051
1052         if (userfaultfd_open(0) < 0)
1053                 return 1;
1054
1055         count_verify = malloc(nr_pages * sizeof(unsigned long long));
1056         if (!count_verify) {
1057                 perror("count_verify");
1058                 return 1;
1059         }
1060
1061         for (nr = 0; nr < nr_pages; nr++) {
1062                 *area_mutex(area_src, nr) = (pthread_mutex_t)
1063                         PTHREAD_MUTEX_INITIALIZER;
1064                 count_verify[nr] = *area_count(area_src, nr) = 1;
1065                 /*
1066                  * In the transition between 255 to 256, powerpc will
1067                  * read out of order in my_bcmp and see both bytes as
1068                  * zero, so leave a placeholder below always non-zero
1069                  * after the count, to avoid my_bcmp to trigger false
1070                  * positives.
1071                  */
1072                 *(area_count(area_src, nr) + 1) = 1;
1073         }
1074
1075         pipefd = malloc(sizeof(int) * nr_cpus * 2);
1076         if (!pipefd) {
1077                 perror("pipefd");
1078                 return 1;
1079         }
1080         for (cpu = 0; cpu < nr_cpus; cpu++) {
1081                 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1082                         perror("pipe");
1083                         return 1;
1084                 }
1085         }
1086
1087         if (posix_memalign(&area, page_size, page_size)) {
1088                 fprintf(stderr, "out of memory\n");
1089                 return 1;
1090         }
1091         zeropage = area;
1092         bzero(zeropage, page_size);
1093
1094         pthread_mutex_lock(&uffd_read_mutex);
1095
1096         pthread_attr_init(&attr);
1097         pthread_attr_setstacksize(&attr, 16*1024*1024);
1098
1099         err = 0;
1100         while (bounces--) {
1101                 unsigned long expected_ioctls;
1102
1103                 printf("bounces: %d, mode:", bounces);
1104                 if (bounces & BOUNCE_RANDOM)
1105                         printf(" rnd");
1106                 if (bounces & BOUNCE_RACINGFAULTS)
1107                         printf(" racing");
1108                 if (bounces & BOUNCE_VERIFY)
1109                         printf(" ver");
1110                 if (bounces & BOUNCE_POLL)
1111                         printf(" poll");
1112                 printf(", ");
1113                 fflush(stdout);
1114
1115                 if (bounces & BOUNCE_POLL)
1116                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1117                 else
1118                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1119
1120                 /* register */
1121                 uffdio_register.range.start = (unsigned long) area_dst;
1122                 uffdio_register.range.len = nr_pages * page_size;
1123                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1124                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1125                         fprintf(stderr, "register failure\n");
1126                         return 1;
1127                 }
1128                 expected_ioctls = uffd_test_ops->expected_ioctls;
1129                 if ((uffdio_register.ioctls & expected_ioctls) !=
1130                     expected_ioctls) {
1131                         fprintf(stderr,
1132                                 "unexpected missing ioctl for anon memory\n");
1133                         return 1;
1134                 }
1135
1136                 if (area_dst_alias) {
1137                         uffdio_register.range.start = (unsigned long)
1138                                 area_dst_alias;
1139                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1140                                 fprintf(stderr, "register failure alias\n");
1141                                 return 1;
1142                         }
1143                 }
1144
1145                 /*
1146                  * The madvise done previously isn't enough: some
1147                  * uffd_thread could have read userfaults (one of
1148                  * those already resolved by the background thread)
1149                  * and it may be in the process of calling
1150                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1151                  * area_src and it would map a zero page in it (of
1152                  * course such a UFFDIO_COPY is perfectly safe as it'd
1153                  * return -EEXIST). The problem comes at the next
1154                  * bounce though: that racing UFFDIO_COPY would
1155                  * generate zeropages in the area_src, so invalidating
1156                  * the previous MADV_DONTNEED. Without this additional
1157                  * MADV_DONTNEED those zeropages leftovers in the
1158                  * area_src would lead to -EEXIST failure during the
1159                  * next bounce, effectively leaving a zeropage in the
1160                  * area_dst.
1161                  *
1162                  * Try to comment this out madvise to see the memory
1163                  * corruption being caught pretty quick.
1164                  *
1165                  * khugepaged is also inhibited to collapse THP after
1166                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1167                  * required to MADV_DONTNEED here.
1168                  */
1169                 if (uffd_test_ops->release_pages(area_dst))
1170                         return 1;
1171
1172                 /* bounce pass */
1173                 if (stress(userfaults))
1174                         return 1;
1175
1176                 /* unregister */
1177                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1178                         fprintf(stderr, "unregister failure\n");
1179                         return 1;
1180                 }
1181                 if (area_dst_alias) {
1182                         uffdio_register.range.start = (unsigned long) area_dst;
1183                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1184                                   &uffdio_register.range)) {
1185                                 fprintf(stderr, "unregister failure alias\n");
1186                                 return 1;
1187                         }
1188                 }
1189
1190                 /* verification */
1191                 if (bounces & BOUNCE_VERIFY) {
1192                         for (nr = 0; nr < nr_pages; nr++) {
1193                                 if (*area_count(area_dst, nr) != count_verify[nr]) {
1194                                         fprintf(stderr,
1195                                                 "error area_count %Lu %Lu %lu\n",
1196                                                 *area_count(area_src, nr),
1197                                                 count_verify[nr],
1198                                                 nr);
1199                                         err = 1;
1200                                         bounces = 0;
1201                                 }
1202                         }
1203                 }
1204
1205                 /* prepare next bounce */
1206                 tmp_area = area_src;
1207                 area_src = area_dst;
1208                 area_dst = tmp_area;
1209
1210                 tmp_area = area_src_alias;
1211                 area_src_alias = area_dst_alias;
1212                 area_dst_alias = tmp_area;
1213
1214                 printf("userfaults:");
1215                 for (cpu = 0; cpu < nr_cpus; cpu++)
1216                         printf(" %lu", userfaults[cpu]);
1217                 printf("\n");
1218         }
1219
1220         if (err)
1221                 return err;
1222
1223         close(uffd);
1224         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1225                 || userfaultfd_events_test();
1226 }
1227
1228 /*
1229  * Copied from mlock2-tests.c
1230  */
1231 unsigned long default_huge_page_size(void)
1232 {
1233         unsigned long hps = 0;
1234         char *line = NULL;
1235         size_t linelen = 0;
1236         FILE *f = fopen("/proc/meminfo", "r");
1237
1238         if (!f)
1239                 return 0;
1240         while (getline(&line, &linelen, f) > 0) {
1241                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1242                         hps <<= 10;
1243                         break;
1244                 }
1245         }
1246
1247         free(line);
1248         fclose(f);
1249         return hps;
1250 }
1251
1252 static void set_test_type(const char *type)
1253 {
1254         if (!strcmp(type, "anon")) {
1255                 test_type = TEST_ANON;
1256                 uffd_test_ops = &anon_uffd_test_ops;
1257         } else if (!strcmp(type, "hugetlb")) {
1258                 test_type = TEST_HUGETLB;
1259                 uffd_test_ops = &hugetlb_uffd_test_ops;
1260         } else if (!strcmp(type, "hugetlb_shared")) {
1261                 map_shared = true;
1262                 test_type = TEST_HUGETLB;
1263                 uffd_test_ops = &hugetlb_uffd_test_ops;
1264         } else if (!strcmp(type, "shmem")) {
1265                 map_shared = true;
1266                 test_type = TEST_SHMEM;
1267                 uffd_test_ops = &shmem_uffd_test_ops;
1268         } else {
1269                 fprintf(stderr, "Unknown test type: %s\n", type), exit(1);
1270         }
1271
1272         if (test_type == TEST_HUGETLB)
1273                 page_size = default_huge_page_size();
1274         else
1275                 page_size = sysconf(_SC_PAGE_SIZE);
1276
1277         if (!page_size)
1278                 fprintf(stderr, "Unable to determine page size\n"),
1279                                 exit(2);
1280         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1281             > page_size)
1282                 fprintf(stderr, "Impossible to run this test\n"), exit(2);
1283 }
1284
1285 static void sigalrm(int sig)
1286 {
1287         if (sig != SIGALRM)
1288                 abort();
1289         test_uffdio_copy_eexist = true;
1290         test_uffdio_zeropage_eexist = true;
1291         alarm(ALARM_INTERVAL_SECS);
1292 }
1293
1294 int main(int argc, char **argv)
1295 {
1296         if (argc < 4)
1297                 usage();
1298
1299         if (signal(SIGALRM, sigalrm) == SIG_ERR)
1300                 fprintf(stderr, "failed to arm SIGALRM"), exit(1);
1301         alarm(ALARM_INTERVAL_SECS);
1302
1303         set_test_type(argv[1]);
1304
1305         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1306         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1307                 nr_cpus;
1308         if (!nr_pages_per_cpu) {
1309                 fprintf(stderr, "invalid MiB\n");
1310                 usage();
1311         }
1312
1313         bounces = atoi(argv[3]);
1314         if (bounces <= 0) {
1315                 fprintf(stderr, "invalid bounces\n");
1316                 usage();
1317         }
1318         nr_pages = nr_pages_per_cpu * nr_cpus;
1319
1320         if (test_type == TEST_HUGETLB) {
1321                 if (argc < 5)
1322                         usage();
1323                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1324                 if (huge_fd < 0) {
1325                         fprintf(stderr, "Open of %s failed", argv[3]);
1326                         perror("open");
1327                         exit(1);
1328                 }
1329                 if (ftruncate(huge_fd, 0)) {
1330                         fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1331                         perror("ftruncate");
1332                         exit(1);
1333                 }
1334         }
1335         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1336                nr_pages, nr_pages_per_cpu);
1337         return userfaultfd_stress();
1338 }
1339
1340 #else /* __NR_userfaultfd */
1341
1342 #warning "missing __NR_userfaultfd definition"
1343
1344 int main(void)
1345 {
1346         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1347         return KSFT_SKIP;
1348 }
1349
1350 #endif /* __NR_userfaultfd */