tools/testing/selftests/vm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Stress userfaultfd syscall.
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  *
   7  * This test allocates two virtual areas and bounces the physical
   8  * memory across the two virtual areas (from area_src to area_dst)
   9  * using userfaultfd.
  10  *
  11  * There are three threads running per CPU:
  12  *
  13  * 1) one per-CPU thread takes a per-page pthread_mutex in a random
  14  *    page of the area_dst (while the physical page may still be in
  15  *    area_src), and increments a per-page counter in the same page,
  16  *    and checks its value against a verification region.
  17  *
  18  * 2) another per-CPU thread handles the userfaults generated by
  19  *    thread 1 above. userfaultfd blocking reads or poll() modes are
  20  *    exercised interleaved.
  21  *
  22  * 3) one last per-CPU thread transfers the memory in the background
  23  *    at maximum bandwidth (if not already transferred by thread
  24  *    2). Each cpu thread takes cares of transferring a portion of the
  25  *    area.
  26  *
  27  * When all threads of type 3 completed the transfer, one bounce is
  28  * complete. area_src and area_dst are then swapped. All threads are
  29  * respawned and so the bounce is immediately restarted in the
  30  * opposite direction.
  31  *
  32  * per-CPU threads 1 by triggering userfaults inside
  33  * pthread_mutex_lock will also verify the atomicity of the memory
  34  * transfer (UFFDIO_COPY).
  35  */
  36
  37 #define _GNU_SOURCE
  38 #include <stdio.h>
  39 #include <errno.h>
  40 #include <unistd.h>
  41 #include <stdlib.h>
  42 #include <sys/types.h>
  43 #include <sys/stat.h>
  44 #include <fcntl.h>
  45 #include <time.h>
  46 #include <signal.h>
  47 #include <poll.h>
  48 #include <string.h>
  49 #include <sys/mman.h>
  50 #include <sys/syscall.h>
  51 #include <sys/ioctl.h>
  52 #include <sys/wait.h>
  53 #include <pthread.h>
  54 #include <linux/userfaultfd.h>
  55 #include <setjmp.h>
  56 #include <stdbool.h>
  57 #include <assert.h>
  58 #include <inttypes.h>
  59 #include <stdint.h>
  60
  61 #include "../kselftest.h"
  62
  63 #ifdef __NR_userfaultfd
  64
  65 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
  66
  67 #define BOUNCE_RANDOM           (1<<0)
  68 #define BOUNCE_RACINGFAULTS     (1<<1)
  69 #define BOUNCE_VERIFY           (1<<2)
  70 #define BOUNCE_POLL             (1<<3)
  71 static int bounces;
  72
  73 #define TEST_ANON       1
  74 #define TEST_HUGETLB    2
  75 #define TEST_SHMEM      3
  76 static int test_type;
  77
  78 /* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
  79 #define ALARM_INTERVAL_SECS 10
  80 static volatile bool test_uffdio_copy_eexist = true;
  81 static volatile bool test_uffdio_zeropage_eexist = true;
  82 /* Whether to test uffd write-protection */
  83 static bool test_uffdio_wp = false;
  84
  85 static bool map_shared;
  86 static int huge_fd;
  87 static char *huge_fd_off0;
  88 static unsigned long long *count_verify;
  89 static int uffd, uffd_flags, finished, *pipefd;
  90 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
  91 static char *zeropage;
  92 pthread_attr_t attr;
  93
  94 /* Userfaultfd test statistics */
  95 struct uffd_stats {
  96         int cpu;
  97         unsigned long missing_faults;
  98         unsigned long wp_faults;
  99 };
 100
 101 /* pthread_mutex_t starts at page offset 0 */
 102 #define area_mutex(___area, ___nr)                                      \
 103         ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
 104 /*
 105  * count is placed in the page after pthread_mutex_t naturally aligned
 106  * to avoid non alignment faults on non-x86 archs.
 107  */
 108 #define area_count(___area, ___nr)                                      \
 109         ((volatile unsigned long long *) ((unsigned long)               \
 110                                  ((___area) + (___nr)*page_size +       \
 111                                   sizeof(pthread_mutex_t) +             \
 112                                   sizeof(unsigned long long) - 1) &     \
 113                                  ~(unsigned long)(sizeof(unsigned long long) \
 114                                                   -  1)))
 115
 116 const char *examples =
 117     "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
 118     "./userfaultfd anon 100 99999\n\n"
 119     "# Run share memory test on 1GiB region with 99 bounces:\n"
 120     "./userfaultfd shmem 1000 99\n\n"
 121     "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
 122     "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
 123     "# Run the same hugetlb test but using shmem:\n"
 124     "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
 125     "# 10MiB-~6GiB 999 bounces anonymous test, "
 126     "continue forever unless an error triggers\n"
 127     "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
 128
 129 static void usage(void)
 130 {
 131         fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
 132                 "[hugetlbfs_file]\n\n");
 133         fprintf(stderr, "Supported <test type>: anon, hugetlb, "
 134                 "hugetlb_shared, shmem\n\n");
 135         fprintf(stderr, "Examples:\n\n");
 136         fprintf(stderr, "%s", examples);
 137         exit(1);
 138 }
 139
 140 #define uffd_error(code, fmt, ...)                                             \
 141         do {                                                                   \
 142                 fprintf(stderr, fmt, ##__VA_ARGS__);                           \
 143                 fprintf(stderr, ": %" PRId64 "\n", (int64_t)(code));           \
 144                 exit(1);                                                       \
 145         } while (0)
 146
 147 static void uffd_stats_reset(struct uffd_stats *uffd_stats,
 148                              unsigned long n_cpus)
 149 {
 150         int i;
 151
 152         for (i = 0; i < n_cpus; i++) {
 153                 uffd_stats[i].cpu = i;
 154                 uffd_stats[i].missing_faults = 0;
 155                 uffd_stats[i].wp_faults = 0;
 156         }
 157 }
 158
 159 static void uffd_stats_report(struct uffd_stats *stats, int n_cpus)
 160 {
 161         int i;
 162         unsigned long long miss_total = 0, wp_total = 0;
 163
 164         for (i = 0; i < n_cpus; i++) {
 165                 miss_total += stats[i].missing_faults;
 166                 wp_total += stats[i].wp_faults;
 167         }
 168
 169         printf("userfaults: %llu missing (", miss_total);
 170         for (i = 0; i < n_cpus; i++)
 171                 printf("%lu+", stats[i].missing_faults);
 172         printf("\b), %llu wp (", wp_total);
 173         for (i = 0; i < n_cpus; i++)
 174                 printf("%lu+", stats[i].wp_faults);
 175         printf("\b)\n");
 176 }
 177
 178 static int anon_release_pages(char *rel_area)
 179 {
 180         int ret = 0;
 181
 182         if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
 183                 perror("madvise");
 184                 ret = 1;
 185         }
 186
 187         return ret;
 188 }
 189
 190 static void anon_allocate_area(void **alloc_area)
 191 {
 192         if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
 193                 fprintf(stderr, "out of memory\n");
 194                 *alloc_area = NULL;
 195         }
 196 }
 197
 198 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 199 {
 200 }
 201
 202 /* HugeTLB memory */
 203 static int hugetlb_release_pages(char *rel_area)
 204 {
 205         int ret = 0;
 206
 207         if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 208                                 rel_area == huge_fd_off0 ? 0 :
 209                                 nr_pages * page_size,
 210                                 nr_pages * page_size)) {
 211                 perror("fallocate");
 212                 ret = 1;
 213         }
 214
 215         return ret;
 216 }
 217
 218 static void hugetlb_allocate_area(void **alloc_area)
 219 {
 220         void *area_alias = NULL;
 221         char **alloc_area_alias;
 222
 223         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 224                            (map_shared ? MAP_SHARED : MAP_PRIVATE) |
 225                            MAP_HUGETLB,
 226                            huge_fd, *alloc_area == area_src ? 0 :
 227                            nr_pages * page_size);
 228         if (*alloc_area == MAP_FAILED) {
 229                 perror("mmap of hugetlbfs file failed");
 230                 goto fail;
 231         }
 232
 233         if (map_shared) {
 234                 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 235                                   MAP_SHARED | MAP_HUGETLB,
 236                                   huge_fd, *alloc_area == area_src ? 0 :
 237                                   nr_pages * page_size);
 238                 if (area_alias == MAP_FAILED) {
 239                         perror("mmap of hugetlb file alias failed");
 240                         goto fail_munmap;
 241                 }
 242         }
 243
 244         if (*alloc_area == area_src) {
 245                 huge_fd_off0 = *alloc_area;
 246                 alloc_area_alias = &area_src_alias;
 247         } else {
 248                 alloc_area_alias = &area_dst_alias;
 249         }
 250         if (area_alias)
 251                 *alloc_area_alias = area_alias;
 252
 253         return;
 254
 255 fail_munmap:
 256         if (munmap(*alloc_area, nr_pages * page_size) < 0) {
 257                 perror("hugetlb munmap");
 258                 exit(1);
 259         }
 260 fail:
 261         *alloc_area = NULL;
 262 }
 263
 264 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 265 {
 266         if (!map_shared)
 267                 return;
 268         /*
 269          * We can't zap just the pagetable with hugetlbfs because
 270          * MADV_DONTEED won't work. So exercise -EEXIST on a alias
 271          * mapping where the pagetables are not established initially,
 272          * this way we'll exercise the -EEXEC at the fs level.
 273          */
 274         *start = (unsigned long) area_dst_alias + offset;
 275 }
 276
 277 /* Shared memory */
 278 static int shmem_release_pages(char *rel_area)
 279 {
 280         int ret = 0;
 281
 282         if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
 283                 perror("madvise");
 284                 ret = 1;
 285         }
 286
 287         return ret;
 288 }
 289
 290 static void shmem_allocate_area(void **alloc_area)
 291 {
 292         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 293                            MAP_ANONYMOUS | MAP_SHARED, -1, 0);
 294         if (*alloc_area == MAP_FAILED) {
 295                 fprintf(stderr, "shared memory mmap failed\n");
 296                 *alloc_area = NULL;
 297         }
 298 }
 299
 300 struct uffd_test_ops {
 301         unsigned long expected_ioctls;
 302         void (*allocate_area)(void **alloc_area);
 303         int (*release_pages)(char *rel_area);
 304         void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
 305 };
 306
 307 #define SHMEM_EXPECTED_IOCTLS           ((1 << _UFFDIO_WAKE) | \
 308                                          (1 << _UFFDIO_COPY) | \
 309                                          (1 << _UFFDIO_ZEROPAGE))
 310
 311 #define ANON_EXPECTED_IOCTLS            ((1 << _UFFDIO_WAKE) | \
 312                                          (1 << _UFFDIO_COPY) | \
 313                                          (1 << _UFFDIO_ZEROPAGE) | \
 314                                          (1 << _UFFDIO_WRITEPROTECT))
 315
 316 static struct uffd_test_ops anon_uffd_test_ops = {
 317         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 318         .allocate_area  = anon_allocate_area,
 319         .release_pages  = anon_release_pages,
 320         .alias_mapping = noop_alias_mapping,
 321 };
 322
 323 static struct uffd_test_ops shmem_uffd_test_ops = {
 324         .expected_ioctls = SHMEM_EXPECTED_IOCTLS,
 325         .allocate_area  = shmem_allocate_area,
 326         .release_pages  = shmem_release_pages,
 327         .alias_mapping = noop_alias_mapping,
 328 };
 329
 330 static struct uffd_test_ops hugetlb_uffd_test_ops = {
 331         .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC,
 332         .allocate_area  = hugetlb_allocate_area,
 333         .release_pages  = hugetlb_release_pages,
 334         .alias_mapping = hugetlb_alias_mapping,
 335 };
 336
 337 static struct uffd_test_ops *uffd_test_ops;
 338
 339 static int my_bcmp(char *str1, char *str2, size_t n)
 340 {
 341         unsigned long i;
 342         for (i = 0; i < n; i++)
 343                 if (str1[i] != str2[i])
 344                         return 1;
 345         return 0;
 346 }
 347
 348 static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
 349 {
 350         struct uffdio_writeprotect prms;
 351
 352         /* Write protection page faults */
 353         prms.range.start = start;
 354         prms.range.len = len;
 355         /* Undo write-protect, do wakeup after that */
 356         prms.mode = wp ? UFFDIO_WRITEPROTECT_MODE_WP : 0;
 357
 358         if (ioctl(ufd, UFFDIO_WRITEPROTECT, &prms)) {
 359                 fprintf(stderr, "clear WP failed for address 0x%" PRIx64 "\n",
 360                         (uint64_t)start);
 361                 exit(1);
 362         }
 363 }
 364
 365 static void *locking_thread(void *arg)
 366 {
 367         unsigned long cpu = (unsigned long) arg;
 368         struct random_data rand;
 369         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 370         int32_t rand_nr;
 371         unsigned long long count;
 372         char randstate[64];
 373         unsigned int seed;
 374         time_t start;
 375
 376         if (bounces & BOUNCE_RANDOM) {
 377                 seed = (unsigned int) time(NULL) - bounces;
 378                 if (!(bounces & BOUNCE_RACINGFAULTS))
 379                         seed += cpu;
 380                 bzero(&rand, sizeof(rand));
 381                 bzero(&randstate, sizeof(randstate));
 382                 if (initstate_r(seed, randstate, sizeof(randstate), &rand)) {
 383                         fprintf(stderr, "srandom_r error\n");
 384                         exit(1);
 385                 }
 386         } else {
 387                 page_nr = -bounces;
 388                 if (!(bounces & BOUNCE_RACINGFAULTS))
 389                         page_nr += cpu * nr_pages_per_cpu;
 390         }
 391
 392         while (!finished) {
 393                 if (bounces & BOUNCE_RANDOM) {
 394                         if (random_r(&rand, &rand_nr)) {
 395                                 fprintf(stderr, "random_r 1 error\n");
 396                                 exit(1);
 397                         }
 398                         page_nr = rand_nr;
 399                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 400                                 if (random_r(&rand, &rand_nr)) {
 401                                         fprintf(stderr, "random_r 2 error\n");
 402                                         exit(1);
 403                                 }
 404                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 405                                            16;
 406                         }
 407                 } else
 408                         page_nr += 1;
 409                 page_nr %= nr_pages;
 410
 411                 start = time(NULL);
 412                 if (bounces & BOUNCE_VERIFY) {
 413                         count = *area_count(area_dst, page_nr);
 414                         if (!count) {
 415                                 fprintf(stderr,
 416                                         "page_nr %lu wrong count %Lu %Lu\n",
 417                                         page_nr, count,
 418                                         count_verify[page_nr]);
 419                                 exit(1);
 420                         }
 421
 422
 423                         /*
 424                          * We can't use bcmp (or memcmp) because that
 425                          * returns 0 erroneously if the memory is
 426                          * changing under it (even if the end of the
 427                          * page is never changing and always
 428                          * different).
 429                          */
 430 #if 1
 431                         if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
 432                                      page_size)) {
 433                                 fprintf(stderr,
 434                                         "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
 435                                         page_nr, count, count_verify[page_nr]);
 436                                 exit(1);
 437                         }
 438 #else
 439                         unsigned long loops;
 440
 441                         loops = 0;
 442                         /* uncomment the below line to test with mutex */
 443                         /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
 444                         while (!bcmp(area_dst + page_nr * page_size, zeropage,
 445                                      page_size)) {
 446                                 loops += 1;
 447                                 if (loops > 10)
 448                                         break;
 449                         }
 450                         /* uncomment below line to test with mutex */
 451                         /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
 452                         if (loops) {
 453                                 fprintf(stderr,
 454                                         "page_nr %lu all zero thread %lu %p %lu\n",
 455                                         page_nr, cpu, area_dst + page_nr * page_size,
 456                                         loops);
 457                                 if (loops > 10)
 458                                         exit(1);
 459                         }
 460 #endif
 461                 }
 462
 463                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 464                 count = *area_count(area_dst, page_nr);
 465                 if (count != count_verify[page_nr]) {
 466                         fprintf(stderr,
 467                                 "page_nr %lu memory corruption %Lu %Lu\n",
 468                                 page_nr, count,
 469                                 count_verify[page_nr]); exit(1);
 470                 }
 471                 count++;
 472                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 473                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 474
 475                 if (time(NULL) - start > 1)
 476                         fprintf(stderr,
 477                                 "userfault too slow %ld "
 478                                 "possible false positive with overcommit\n",
 479                                 time(NULL) - start);
 480         }
 481
 482         return NULL;
 483 }
 484
 485 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 486                             unsigned long offset)
 487 {
 488         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 489                                      uffdio_copy->len,
 490                                      offset);
 491         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 492                 /* real retval in ufdio_copy.copy */
 493                 if (uffdio_copy->copy != -EEXIST) {
 494                         uffd_error(uffdio_copy->copy,
 495                                    "UFFDIO_COPY retry error");
 496                 }
 497         } else
 498                 uffd_error(uffdio_copy->copy, "UFFDIO_COPY retry unexpected");
 499 }
 500
 501 static int __copy_page(int ufd, unsigned long offset, bool retry)
 502 {
 503         struct uffdio_copy uffdio_copy;
 504
 505         if (offset >= nr_pages * page_size) {
 506                 fprintf(stderr, "unexpected offset %lu\n", offset);
 507                 exit(1);
 508         }
 509         uffdio_copy.dst = (unsigned long) area_dst + offset;
 510         uffdio_copy.src = (unsigned long) area_src + offset;
 511         uffdio_copy.len = page_size;
 512         if (test_uffdio_wp)
 513                 uffdio_copy.mode = UFFDIO_COPY_MODE_WP;
 514         else
 515                 uffdio_copy.mode = 0;
 516         uffdio_copy.copy = 0;
 517         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 518                 /* real retval in ufdio_copy.copy */
 519                 if (uffdio_copy.copy != -EEXIST)
 520                         uffd_error(uffdio_copy.copy, "UFFDIO_COPY error");
 521         } else if (uffdio_copy.copy != page_size) {
 522                 uffd_error(uffdio_copy.copy, "UFFDIO_COPY unexpected copy");
 523         } else {
 524                 if (test_uffdio_copy_eexist && retry) {
 525                         test_uffdio_copy_eexist = false;
 526                         retry_copy_page(ufd, &uffdio_copy, offset);
 527                 }
 528                 return 1;
 529         }
 530         return 0;
 531 }
 532
 533 static int copy_page_retry(int ufd, unsigned long offset)
 534 {
 535         return __copy_page(ufd, offset, true);
 536 }
 537
 538 static int copy_page(int ufd, unsigned long offset)
 539 {
 540         return __copy_page(ufd, offset, false);
 541 }
 542
 543 static int uffd_read_msg(int ufd, struct uffd_msg *msg)
 544 {
 545         int ret = read(uffd, msg, sizeof(*msg));
 546
 547         if (ret != sizeof(*msg)) {
 548                 if (ret < 0) {
 549                         if (errno == EAGAIN)
 550                                 return 1;
 551                         perror("blocking read error");
 552                 } else {
 553                         fprintf(stderr, "short read\n");
 554                 }
 555                 exit(1);
 556         }
 557
 558         return 0;
 559 }
 560
 561 static void uffd_handle_page_fault(struct uffd_msg *msg,
 562                                    struct uffd_stats *stats)
 563 {
 564         unsigned long offset;
 565
 566         if (msg->event != UFFD_EVENT_PAGEFAULT) {
 567                 fprintf(stderr, "unexpected msg event %u\n", msg->event);
 568                 exit(1);
 569         }
 570
 571         if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP) {
 572                 wp_range(uffd, msg->arg.pagefault.address, page_size, false);
 573                 stats->wp_faults++;
 574         } else {
 575                 /* Missing page faults */
 576                 if (bounces & BOUNCE_VERIFY &&
 577                     msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE) {
 578                         fprintf(stderr, "unexpected write fault\n");
 579                         exit(1);
 580                 }
 581
 582                 offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
 583                 offset &= ~(page_size-1);
 584
 585                 if (copy_page(uffd, offset))
 586                         stats->missing_faults++;
 587         }
 588 }
 589
 590 static void *uffd_poll_thread(void *arg)
 591 {
 592         struct uffd_stats *stats = (struct uffd_stats *)arg;
 593         unsigned long cpu = stats->cpu;
 594         struct pollfd pollfd[2];
 595         struct uffd_msg msg;
 596         struct uffdio_register uffd_reg;
 597         int ret;
 598         char tmp_chr;
 599
 600         pollfd[0].fd = uffd;
 601         pollfd[0].events = POLLIN;
 602         pollfd[1].fd = pipefd[cpu*2];
 603         pollfd[1].events = POLLIN;
 604
 605         for (;;) {
 606                 ret = poll(pollfd, 2, -1);
 607                 if (!ret) {
 608                         fprintf(stderr, "poll error %d\n", ret);
 609                         exit(1);
 610                 }
 611                 if (ret < 0) {
 612                         perror("poll");
 613                         exit(1);
 614                 }
 615                 if (pollfd[1].revents & POLLIN) {
 616                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1) {
 617                                 fprintf(stderr, "read pipefd error\n");
 618                                 exit(1);
 619                         }
 620                         break;
 621                 }
 622                 if (!(pollfd[0].revents & POLLIN)) {
 623                         fprintf(stderr, "pollfd[0].revents %d\n",
 624                                 pollfd[0].revents);
 625                         exit(1);
 626                 }
 627                 if (uffd_read_msg(uffd, &msg))
 628                         continue;
 629                 switch (msg.event) {
 630                 default:
 631                         fprintf(stderr, "unexpected msg event %u\n",
 632                                 msg.event); exit(1);
 633                         break;
 634                 case UFFD_EVENT_PAGEFAULT:
 635                         uffd_handle_page_fault(&msg, stats);
 636                         break;
 637                 case UFFD_EVENT_FORK:
 638                         close(uffd);
 639                         uffd = msg.arg.fork.ufd;
 640                         pollfd[0].fd = uffd;
 641                         break;
 642                 case UFFD_EVENT_REMOVE:
 643                         uffd_reg.range.start = msg.arg.remove.start;
 644                         uffd_reg.range.len = msg.arg.remove.end -
 645                                 msg.arg.remove.start;
 646                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range)) {
 647                                 fprintf(stderr, "remove failure\n");
 648                                 exit(1);
 649                         }
 650                         break;
 651                 case UFFD_EVENT_REMAP:
 652                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 653                         break;
 654                 }
 655         }
 656
 657         return NULL;
 658 }
 659
 660 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 661
 662 static void *uffd_read_thread(void *arg)
 663 {
 664         struct uffd_stats *stats = (struct uffd_stats *)arg;
 665         struct uffd_msg msg;
 666
 667         pthread_mutex_unlock(&uffd_read_mutex);
 668         /* from here cancellation is ok */
 669
 670         for (;;) {
 671                 if (uffd_read_msg(uffd, &msg))
 672                         continue;
 673                 uffd_handle_page_fault(&msg, stats);
 674         }
 675
 676         return NULL;
 677 }
 678
 679 static void *background_thread(void *arg)
 680 {
 681         unsigned long cpu = (unsigned long) arg;
 682         unsigned long page_nr, start_nr, mid_nr, end_nr;
 683
 684         start_nr = cpu * nr_pages_per_cpu;
 685         end_nr = (cpu+1) * nr_pages_per_cpu;
 686         mid_nr = (start_nr + end_nr) / 2;
 687
 688         /* Copy the first half of the pages */
 689         for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
 690                 copy_page_retry(uffd, page_nr * page_size);
 691
 692         /*
 693          * If we need to test uffd-wp, set it up now.  Then we'll have
 694          * at least the first half of the pages mapped already which
 695          * can be write-protected for testing
 696          */
 697         if (test_uffdio_wp)
 698                 wp_range(uffd, (unsigned long)area_dst + start_nr * page_size,
 699                         nr_pages_per_cpu * page_size, true);
 700
 701         /*
 702          * Continue the 2nd half of the page copying, handling write
 703          * protection faults if any
 704          */
 705         for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
 706                 copy_page_retry(uffd, page_nr * page_size);
 707
 708         return NULL;
 709 }
 710
 711 static int stress(struct uffd_stats *uffd_stats)
 712 {
 713         unsigned long cpu;
 714         pthread_t locking_threads[nr_cpus];
 715         pthread_t uffd_threads[nr_cpus];
 716         pthread_t background_threads[nr_cpus];
 717
 718         finished = 0;
 719         for (cpu = 0; cpu < nr_cpus; cpu++) {
 720                 if (pthread_create(&locking_threads[cpu], &attr,
 721                                    locking_thread, (void *)cpu))
 722                         return 1;
 723                 if (bounces & BOUNCE_POLL) {
 724                         if (pthread_create(&uffd_threads[cpu], &attr,
 725                                            uffd_poll_thread,
 726                                            (void *)&uffd_stats[cpu]))
 727                                 return 1;
 728                 } else {
 729                         if (pthread_create(&uffd_threads[cpu], &attr,
 730                                            uffd_read_thread,
 731                                            (void *)&uffd_stats[cpu]))
 732                                 return 1;
 733                         pthread_mutex_lock(&uffd_read_mutex);
 734                 }
 735                 if (pthread_create(&background_threads[cpu], &attr,
 736                                    background_thread, (void *)cpu))
 737                         return 1;
 738         }
 739         for (cpu = 0; cpu < nr_cpus; cpu++)
 740                 if (pthread_join(background_threads[cpu], NULL))
 741                         return 1;
 742
 743         /*
 744          * Be strict and immediately zap area_src, the whole area has
 745          * been transferred already by the background treads. The
 746          * area_src could then be faulted in in a racy way by still
 747          * running uffdio_threads reading zeropages after we zapped
 748          * area_src (but they're guaranteed to get -EEXIST from
 749          * UFFDIO_COPY without writing zero pages into area_dst
 750          * because the background threads already completed).
 751          */
 752         if (uffd_test_ops->release_pages(area_src))
 753                 return 1;
 754
 755
 756         finished = 1;
 757         for (cpu = 0; cpu < nr_cpus; cpu++)
 758                 if (pthread_join(locking_threads[cpu], NULL))
 759                         return 1;
 760
 761         for (cpu = 0; cpu < nr_cpus; cpu++) {
 762                 char c;
 763                 if (bounces & BOUNCE_POLL) {
 764                         if (write(pipefd[cpu*2+1], &c, 1) != 1) {
 765                                 fprintf(stderr, "pipefd write error\n");
 766                                 return 1;
 767                         }
 768                         if (pthread_join(uffd_threads[cpu],
 769                                          (void *)&uffd_stats[cpu]))
 770                                 return 1;
 771                 } else {
 772                         if (pthread_cancel(uffd_threads[cpu]))
 773                                 return 1;
 774                         if (pthread_join(uffd_threads[cpu], NULL))
 775                                 return 1;
 776                 }
 777         }
 778
 779         return 0;
 780 }
 781
 782 static int userfaultfd_open(int features)
 783 {
 784         struct uffdio_api uffdio_api;
 785
 786         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 787         if (uffd < 0) {
 788                 fprintf(stderr,
 789                         "userfaultfd syscall not available in this kernel\n");
 790                 return 1;
 791         }
 792         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 793
 794         uffdio_api.api = UFFD_API;
 795         uffdio_api.features = features;
 796         if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
 797                 fprintf(stderr, "UFFDIO_API failed.\nPlease make sure to "
 798                         "run with either root or ptrace capability.\n");
 799                 return 1;
 800         }
 801         if (uffdio_api.api != UFFD_API) {
 802                 fprintf(stderr, "UFFDIO_API error: %" PRIu64 "\n",
 803                         (uint64_t)uffdio_api.api);
 804                 return 1;
 805         }
 806
 807         return 0;
 808 }
 809
 810 sigjmp_buf jbuf, *sigbuf;
 811
 812 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 813 {
 814         if (sig == SIGBUS) {
 815                 if (sigbuf)
 816                         siglongjmp(*sigbuf, 1);
 817                 abort();
 818         }
 819 }
 820
 821 /*
 822  * For non-cooperative userfaultfd test we fork() a process that will
 823  * generate pagefaults, will mremap the area monitored by the
 824  * userfaultfd and at last this process will release the monitored
 825  * area.
 826  * For the anonymous and shared memory the area is divided into two
 827  * parts, the first part is accessed before mremap, and the second
 828  * part is accessed after mremap. Since hugetlbfs does not support
 829  * mremap, the entire monitored area is accessed in a single pass for
 830  * HUGETLB_TEST.
 831  * The release of the pages currently generates event for shmem and
 832  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 833  * for hugetlb.
 834  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 835  * monitored area, generate pagefaults and test that signal is delivered.
 836  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 837  * test robustness use case - we release monitored area, fork a process
 838  * that will generate pagefaults and verify signal is generated.
 839  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 840  * feature. Using monitor thread, verify no userfault events are generated.
 841  */
 842 static int faulting_process(int signal_test)
 843 {
 844         unsigned long nr;
 845         unsigned long long count;
 846         unsigned long split_nr_pages;
 847         unsigned long lastnr;
 848         struct sigaction act;
 849         unsigned long signalled = 0;
 850
 851         if (test_type != TEST_HUGETLB)
 852                 split_nr_pages = (nr_pages + 1) / 2;
 853         else
 854                 split_nr_pages = nr_pages;
 855
 856         if (signal_test) {
 857                 sigbuf = &jbuf;
 858                 memset(&act, 0, sizeof(act));
 859                 act.sa_sigaction = sighndl;
 860                 act.sa_flags = SA_SIGINFO;
 861                 if (sigaction(SIGBUS, &act, 0)) {
 862                         perror("sigaction");
 863                         return 1;
 864                 }
 865                 lastnr = (unsigned long)-1;
 866         }
 867
 868         for (nr = 0; nr < split_nr_pages; nr++) {
 869                 int steps = 1;
 870                 unsigned long offset = nr * page_size;
 871
 872                 if (signal_test) {
 873                         if (sigsetjmp(*sigbuf, 1) != 0) {
 874                                 if (steps == 1 && nr == lastnr) {
 875                                         fprintf(stderr, "Signal repeated\n");
 876                                         return 1;
 877                                 }
 878
 879                                 lastnr = nr;
 880                                 if (signal_test == 1) {
 881                                         if (steps == 1) {
 882                                                 /* This is a MISSING request */
 883                                                 steps++;
 884                                                 if (copy_page(uffd, offset))
 885                                                         signalled++;
 886                                         } else {
 887                                                 /* This is a WP request */
 888                                                 assert(steps == 2);
 889                                                 wp_range(uffd,
 890                                                          (__u64)area_dst +
 891                                                          offset,
 892                                                          page_size, false);
 893                                         }
 894                                 } else {
 895                                         signalled++;
 896                                         continue;
 897                                 }
 898                         }
 899                 }
 900
 901                 count = *area_count(area_dst, nr);
 902                 if (count != count_verify[nr]) {
 903                         fprintf(stderr,
 904                                 "nr %lu memory corruption %Lu %Lu\n",
 905                                 nr, count,
 906                                 count_verify[nr]);
 907                 }
 908                 /*
 909                  * Trigger write protection if there is by writing
 910                  * the same value back.
 911                  */
 912                 *area_count(area_dst, nr) = count;
 913         }
 914
 915         if (signal_test)
 916                 return signalled != split_nr_pages;
 917
 918         if (test_type == TEST_HUGETLB)
 919                 return 0;
 920
 921         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 922                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 923         if (area_dst == MAP_FAILED) {
 924                 perror("mremap");
 925                 exit(1);
 926         }
 927
 928         for (; nr < nr_pages; nr++) {
 929                 count = *area_count(area_dst, nr);
 930                 if (count != count_verify[nr]) {
 931                         fprintf(stderr,
 932                                 "nr %lu memory corruption %Lu %Lu\n",
 933                                 nr, count,
 934                                 count_verify[nr]); exit(1);
 935                 }
 936                 /*
 937                  * Trigger write protection if there is by writing
 938                  * the same value back.
 939                  */
 940                 *area_count(area_dst, nr) = count;
 941         }
 942
 943         if (uffd_test_ops->release_pages(area_dst))
 944                 return 1;
 945
 946         for (nr = 0; nr < nr_pages; nr++) {
 947                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size)) {
 948                         fprintf(stderr, "nr %lu is not zero\n", nr);
 949                         exit(1);
 950                 }
 951         }
 952
 953         return 0;
 954 }
 955
 956 static void retry_uffdio_zeropage(int ufd,
 957                                   struct uffdio_zeropage *uffdio_zeropage,
 958                                   unsigned long offset)
 959 {
 960         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
 961                                      uffdio_zeropage->range.len,
 962                                      offset);
 963         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
 964                 if (uffdio_zeropage->zeropage != -EEXIST) {
 965                         uffd_error(uffdio_zeropage->zeropage,
 966                                    "UFFDIO_ZEROPAGE retry error");
 967                 }
 968         } else {
 969                 uffd_error(uffdio_zeropage->zeropage,
 970                            "UFFDIO_ZEROPAGE retry unexpected");
 971         }
 972 }
 973
 974 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
 975 {
 976         struct uffdio_zeropage uffdio_zeropage;
 977         int ret;
 978         unsigned long has_zeropage;
 979         __s64 res;
 980
 981         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
 982
 983         if (offset >= nr_pages * page_size) {
 984                 fprintf(stderr, "unexpected offset %lu\n", offset);
 985                 exit(1);
 986         }
 987         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
 988         uffdio_zeropage.range.len = page_size;
 989         uffdio_zeropage.mode = 0;
 990         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
 991         res = uffdio_zeropage.zeropage;
 992         if (ret) {
 993                 /* real retval in ufdio_zeropage.zeropage */
 994                 if (has_zeropage) {
 995                         uffd_error(res, "UFFDIO_ZEROPAGE %s",
 996                                    res == -EEXIST ? "-EEXIST" : "error");
 997                 } else if (res != -EINVAL)
 998                         uffd_error(res, "UFFDIO_ZEROPAGE not -EINVAL");
 999         } else if (has_zeropage) {
1000                 if (res != page_size) {
1001                         uffd_error(res, "UFFDIO_ZEROPAGE unexpected");
1002                 } else {
1003                         if (test_uffdio_zeropage_eexist && retry) {
1004                                 test_uffdio_zeropage_eexist = false;
1005                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
1006                                                       offset);
1007                         }
1008                         return 1;
1009                 }
1010         } else
1011                 uffd_error(res, "UFFDIO_ZEROPAGE succeeded");
1012
1013         return 0;
1014 }
1015
1016 static int uffdio_zeropage(int ufd, unsigned long offset)
1017 {
1018         return __uffdio_zeropage(ufd, offset, false);
1019 }
1020
1021 /* exercise UFFDIO_ZEROPAGE */
1022 static int userfaultfd_zeropage_test(void)
1023 {
1024         struct uffdio_register uffdio_register;
1025         unsigned long expected_ioctls;
1026
1027         printf("testing UFFDIO_ZEROPAGE: ");
1028         fflush(stdout);
1029
1030         if (uffd_test_ops->release_pages(area_dst))
1031                 return 1;
1032
1033         if (userfaultfd_open(0))
1034                 return 1;
1035         uffdio_register.range.start = (unsigned long) area_dst;
1036         uffdio_register.range.len = nr_pages * page_size;
1037         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1038         if (test_uffdio_wp)
1039                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1040         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1041                 fprintf(stderr, "register failure\n");
1042                 exit(1);
1043         }
1044
1045         expected_ioctls = uffd_test_ops->expected_ioctls;
1046         if ((uffdio_register.ioctls & expected_ioctls) !=
1047             expected_ioctls) {
1048                 fprintf(stderr,
1049                         "unexpected missing ioctl for anon memory\n");
1050                 exit(1);
1051         }
1052
1053         if (uffdio_zeropage(uffd, 0)) {
1054                 if (my_bcmp(area_dst, zeropage, page_size)) {
1055                         fprintf(stderr, "zeropage is not zero\n");
1056                         exit(1);
1057                 }
1058         }
1059
1060         close(uffd);
1061         printf("done.\n");
1062         return 0;
1063 }
1064
1065 static int userfaultfd_events_test(void)
1066 {
1067         struct uffdio_register uffdio_register;
1068         unsigned long expected_ioctls;
1069         pthread_t uffd_mon;
1070         int err, features;
1071         pid_t pid;
1072         char c;
1073         struct uffd_stats stats = { 0 };
1074
1075         printf("testing events (fork, remap, remove): ");
1076         fflush(stdout);
1077
1078         if (uffd_test_ops->release_pages(area_dst))
1079                 return 1;
1080
1081         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
1082                 UFFD_FEATURE_EVENT_REMOVE;
1083         if (userfaultfd_open(features))
1084                 return 1;
1085         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1086
1087         uffdio_register.range.start = (unsigned long) area_dst;
1088         uffdio_register.range.len = nr_pages * page_size;
1089         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1090         if (test_uffdio_wp)
1091                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1092         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1093                 fprintf(stderr, "register failure\n");
1094                 exit(1);
1095         }
1096
1097         expected_ioctls = uffd_test_ops->expected_ioctls;
1098         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1099                 fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1100                 exit(1);
1101         }
1102
1103         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1104                 perror("uffd_poll_thread create");
1105                 exit(1);
1106         }
1107
1108         pid = fork();
1109         if (pid < 0) {
1110                 perror("fork");
1111                 exit(1);
1112         }
1113
1114         if (!pid)
1115                 return faulting_process(0);
1116
1117         waitpid(pid, &err, 0);
1118         if (err) {
1119                 fprintf(stderr, "faulting process failed\n");
1120                 exit(1);
1121         }
1122
1123         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1124                 perror("pipe write");
1125                 exit(1);
1126         }
1127         if (pthread_join(uffd_mon, NULL))
1128                 return 1;
1129
1130         close(uffd);
1131
1132         uffd_stats_report(&stats, 1);
1133
1134         return stats.missing_faults != nr_pages;
1135 }
1136
1137 static int userfaultfd_sig_test(void)
1138 {
1139         struct uffdio_register uffdio_register;
1140         unsigned long expected_ioctls;
1141         unsigned long userfaults;
1142         pthread_t uffd_mon;
1143         int err, features;
1144         pid_t pid;
1145         char c;
1146         struct uffd_stats stats = { 0 };
1147
1148         printf("testing signal delivery: ");
1149         fflush(stdout);
1150
1151         if (uffd_test_ops->release_pages(area_dst))
1152                 return 1;
1153
1154         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
1155         if (userfaultfd_open(features))
1156                 return 1;
1157         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1158
1159         uffdio_register.range.start = (unsigned long) area_dst;
1160         uffdio_register.range.len = nr_pages * page_size;
1161         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1162         if (test_uffdio_wp)
1163                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1164         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1165                 fprintf(stderr, "register failure\n");
1166                 exit(1);
1167         }
1168
1169         expected_ioctls = uffd_test_ops->expected_ioctls;
1170         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1171                 fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1172                 exit(1);
1173         }
1174
1175         if (faulting_process(1)) {
1176                 fprintf(stderr, "faulting process failed\n");
1177                 exit(1);
1178         }
1179
1180         if (uffd_test_ops->release_pages(area_dst))
1181                 return 1;
1182
1183         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1184                 perror("uffd_poll_thread create");
1185                 exit(1);
1186         }
1187
1188         pid = fork();
1189         if (pid < 0) {
1190                 perror("fork");
1191                 exit(1);
1192         }
1193
1194         if (!pid)
1195                 exit(faulting_process(2));
1196
1197         waitpid(pid, &err, 0);
1198         if (err) {
1199                 fprintf(stderr, "faulting process failed\n");
1200                 exit(1);
1201         }
1202
1203         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1204                 perror("pipe write");
1205                 exit(1);
1206         }
1207         if (pthread_join(uffd_mon, (void **)&userfaults))
1208                 return 1;
1209
1210         printf("done.\n");
1211         if (userfaults)
1212                 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1213                         userfaults);
1214         close(uffd);
1215         return userfaults != 0;
1216 }
1217
1218 static int userfaultfd_stress(void)
1219 {
1220         void *area;
1221         char *tmp_area;
1222         unsigned long nr;
1223         struct uffdio_register uffdio_register;
1224         unsigned long cpu;
1225         int err;
1226         struct uffd_stats uffd_stats[nr_cpus];
1227
1228         uffd_test_ops->allocate_area((void **)&area_src);
1229         if (!area_src)
1230                 return 1;
1231         uffd_test_ops->allocate_area((void **)&area_dst);
1232         if (!area_dst)
1233                 return 1;
1234
1235         if (userfaultfd_open(0))
1236                 return 1;
1237
1238         count_verify = malloc(nr_pages * sizeof(unsigned long long));
1239         if (!count_verify) {
1240                 perror("count_verify");
1241                 return 1;
1242         }
1243
1244         for (nr = 0; nr < nr_pages; nr++) {
1245                 *area_mutex(area_src, nr) = (pthread_mutex_t)
1246                         PTHREAD_MUTEX_INITIALIZER;
1247                 count_verify[nr] = *area_count(area_src, nr) = 1;
1248                 /*
1249                  * In the transition between 255 to 256, powerpc will
1250                  * read out of order in my_bcmp and see both bytes as
1251                  * zero, so leave a placeholder below always non-zero
1252                  * after the count, to avoid my_bcmp to trigger false
1253                  * positives.
1254                  */
1255                 *(area_count(area_src, nr) + 1) = 1;
1256         }
1257
1258         pipefd = malloc(sizeof(int) * nr_cpus * 2);
1259         if (!pipefd) {
1260                 perror("pipefd");
1261                 return 1;
1262         }
1263         for (cpu = 0; cpu < nr_cpus; cpu++) {
1264                 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1265                         perror("pipe");
1266                         return 1;
1267                 }
1268         }
1269
1270         if (posix_memalign(&area, page_size, page_size)) {
1271                 fprintf(stderr, "out of memory\n");
1272                 return 1;
1273         }
1274         zeropage = area;
1275         bzero(zeropage, page_size);
1276
1277         pthread_mutex_lock(&uffd_read_mutex);
1278
1279         pthread_attr_init(&attr);
1280         pthread_attr_setstacksize(&attr, 16*1024*1024);
1281
1282         err = 0;
1283         while (bounces--) {
1284                 unsigned long expected_ioctls;
1285
1286                 printf("bounces: %d, mode:", bounces);
1287                 if (bounces & BOUNCE_RANDOM)
1288                         printf(" rnd");
1289                 if (bounces & BOUNCE_RACINGFAULTS)
1290                         printf(" racing");
1291                 if (bounces & BOUNCE_VERIFY)
1292                         printf(" ver");
1293                 if (bounces & BOUNCE_POLL)
1294                         printf(" poll");
1295                 else
1296                         printf(" read");
1297                 printf(", ");
1298                 fflush(stdout);
1299
1300                 if (bounces & BOUNCE_POLL)
1301                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1302                 else
1303                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1304
1305                 /* register */
1306                 uffdio_register.range.start = (unsigned long) area_dst;
1307                 uffdio_register.range.len = nr_pages * page_size;
1308                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1309                 if (test_uffdio_wp)
1310                         uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1311                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1312                         fprintf(stderr, "register failure\n");
1313                         return 1;
1314                 }
1315                 expected_ioctls = uffd_test_ops->expected_ioctls;
1316                 if ((uffdio_register.ioctls & expected_ioctls) !=
1317                     expected_ioctls) {
1318                         fprintf(stderr,
1319                                 "unexpected missing ioctl for anon memory\n");
1320                         return 1;
1321                 }
1322
1323                 if (area_dst_alias) {
1324                         uffdio_register.range.start = (unsigned long)
1325                                 area_dst_alias;
1326                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1327                                 fprintf(stderr, "register failure alias\n");
1328                                 return 1;
1329                         }
1330                 }
1331
1332                 /*
1333                  * The madvise done previously isn't enough: some
1334                  * uffd_thread could have read userfaults (one of
1335                  * those already resolved by the background thread)
1336                  * and it may be in the process of calling
1337                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1338                  * area_src and it would map a zero page in it (of
1339                  * course such a UFFDIO_COPY is perfectly safe as it'd
1340                  * return -EEXIST). The problem comes at the next
1341                  * bounce though: that racing UFFDIO_COPY would
1342                  * generate zeropages in the area_src, so invalidating
1343                  * the previous MADV_DONTNEED. Without this additional
1344                  * MADV_DONTNEED those zeropages leftovers in the
1345                  * area_src would lead to -EEXIST failure during the
1346                  * next bounce, effectively leaving a zeropage in the
1347                  * area_dst.
1348                  *
1349                  * Try to comment this out madvise to see the memory
1350                  * corruption being caught pretty quick.
1351                  *
1352                  * khugepaged is also inhibited to collapse THP after
1353                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1354                  * required to MADV_DONTNEED here.
1355                  */
1356                 if (uffd_test_ops->release_pages(area_dst))
1357                         return 1;
1358
1359                 uffd_stats_reset(uffd_stats, nr_cpus);
1360
1361                 /* bounce pass */
1362                 if (stress(uffd_stats))
1363                         return 1;
1364
1365                 /* Clear all the write protections if there is any */
1366                 if (test_uffdio_wp)
1367                         wp_range(uffd, (unsigned long)area_dst,
1368                                  nr_pages * page_size, false);
1369
1370                 /* unregister */
1371                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1372                         fprintf(stderr, "unregister failure\n");
1373                         return 1;
1374                 }
1375                 if (area_dst_alias) {
1376                         uffdio_register.range.start = (unsigned long) area_dst;
1377                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1378                                   &uffdio_register.range)) {
1379                                 fprintf(stderr, "unregister failure alias\n");
1380                                 return 1;
1381                         }
1382                 }
1383
1384                 /* verification */
1385                 if (bounces & BOUNCE_VERIFY) {
1386                         for (nr = 0; nr < nr_pages; nr++) {
1387                                 if (*area_count(area_dst, nr) != count_verify[nr]) {
1388                                         fprintf(stderr,
1389                                                 "error area_count %Lu %Lu %lu\n",
1390                                                 *area_count(area_src, nr),
1391                                                 count_verify[nr],
1392                                                 nr);
1393                                         err = 1;
1394                                         bounces = 0;
1395                                 }
1396                         }
1397                 }
1398
1399                 /* prepare next bounce */
1400                 tmp_area = area_src;
1401                 area_src = area_dst;
1402                 area_dst = tmp_area;
1403
1404                 tmp_area = area_src_alias;
1405                 area_src_alias = area_dst_alias;
1406                 area_dst_alias = tmp_area;
1407
1408                 uffd_stats_report(uffd_stats, nr_cpus);
1409         }
1410
1411         if (err)
1412                 return err;
1413
1414         close(uffd);
1415         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1416                 || userfaultfd_events_test();
1417 }
1418
1419 /*
1420  * Copied from mlock2-tests.c
1421  */
1422 unsigned long default_huge_page_size(void)
1423 {
1424         unsigned long hps = 0;
1425         char *line = NULL;
1426         size_t linelen = 0;
1427         FILE *f = fopen("/proc/meminfo", "r");
1428
1429         if (!f)
1430                 return 0;
1431         while (getline(&line, &linelen, f) > 0) {
1432                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1433                         hps <<= 10;
1434                         break;
1435                 }
1436         }
1437
1438         free(line);
1439         fclose(f);
1440         return hps;
1441 }
1442
1443 static void set_test_type(const char *type)
1444 {
1445         if (!strcmp(type, "anon")) {
1446                 test_type = TEST_ANON;
1447                 uffd_test_ops = &anon_uffd_test_ops;
1448                 /* Only enable write-protect test for anonymous test */
1449                 test_uffdio_wp = true;
1450         } else if (!strcmp(type, "hugetlb")) {
1451                 test_type = TEST_HUGETLB;
1452                 uffd_test_ops = &hugetlb_uffd_test_ops;
1453         } else if (!strcmp(type, "hugetlb_shared")) {
1454                 map_shared = true;
1455                 test_type = TEST_HUGETLB;
1456                 uffd_test_ops = &hugetlb_uffd_test_ops;
1457         } else if (!strcmp(type, "shmem")) {
1458                 map_shared = true;
1459                 test_type = TEST_SHMEM;
1460                 uffd_test_ops = &shmem_uffd_test_ops;
1461         } else {
1462                 fprintf(stderr, "Unknown test type: %s\n", type); exit(1);
1463         }
1464
1465         if (test_type == TEST_HUGETLB)
1466                 page_size = default_huge_page_size();
1467         else
1468                 page_size = sysconf(_SC_PAGE_SIZE);
1469
1470         if (!page_size) {
1471                 fprintf(stderr, "Unable to determine page size\n");
1472                 exit(2);
1473         }
1474         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1475             > page_size) {
1476                 fprintf(stderr, "Impossible to run this test\n");
1477                 exit(2);
1478         }
1479 }
1480
1481 static void sigalrm(int sig)
1482 {
1483         if (sig != SIGALRM)
1484                 abort();
1485         test_uffdio_copy_eexist = true;
1486         test_uffdio_zeropage_eexist = true;
1487         alarm(ALARM_INTERVAL_SECS);
1488 }
1489
1490 int main(int argc, char **argv)
1491 {
1492         if (argc < 4)
1493                 usage();
1494
1495         if (signal(SIGALRM, sigalrm) == SIG_ERR) {
1496                 fprintf(stderr, "failed to arm SIGALRM");
1497                 exit(1);
1498         }
1499         alarm(ALARM_INTERVAL_SECS);
1500
1501         set_test_type(argv[1]);
1502
1503         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1504         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1505                 nr_cpus;
1506         if (!nr_pages_per_cpu) {
1507                 fprintf(stderr, "invalid MiB\n");
1508                 usage();
1509         }
1510
1511         bounces = atoi(argv[3]);
1512         if (bounces <= 0) {
1513                 fprintf(stderr, "invalid bounces\n");
1514                 usage();
1515         }
1516         nr_pages = nr_pages_per_cpu * nr_cpus;
1517
1518         if (test_type == TEST_HUGETLB) {
1519                 if (argc < 5)
1520                         usage();
1521                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1522                 if (huge_fd < 0) {
1523                         fprintf(stderr, "Open of %s failed", argv[3]);
1524                         perror("open");
1525                         exit(1);
1526                 }
1527                 if (ftruncate(huge_fd, 0)) {
1528                         fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1529                         perror("ftruncate");
1530                         exit(1);
1531                 }
1532         }
1533         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1534                nr_pages, nr_pages_per_cpu);
1535         return userfaultfd_stress();
1536 }
1537
1538 #else /* __NR_userfaultfd */
1539
1540 #warning "missing __NR_userfaultfd definition"
1541
1542 int main(void)
1543 {
1544         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1545         return KSFT_SKIP;
1546 }
1547
1548 #endif /* __NR_userfaultfd */