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 /* Whether to test uffd minor faults */
  85 static bool test_uffdio_minor = false;
  86
  87 static bool map_shared;
  88 static int huge_fd;
  89 static char *huge_fd_off0;
  90 static unsigned long long *count_verify;
  91 static int uffd, uffd_flags, finished, *pipefd;
  92 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
  93 static char *zeropage;
  94 pthread_attr_t attr;
  95
  96 /* Userfaultfd test statistics */
  97 struct uffd_stats {
  98         int cpu;
  99         unsigned long missing_faults;
 100         unsigned long wp_faults;
 101         unsigned long minor_faults;
 102 };
 103
 104 /* pthread_mutex_t starts at page offset 0 */
 105 #define area_mutex(___area, ___nr)                                      \
 106         ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
 107 /*
 108  * count is placed in the page after pthread_mutex_t naturally aligned
 109  * to avoid non alignment faults on non-x86 archs.
 110  */
 111 #define area_count(___area, ___nr)                                      \
 112         ((volatile unsigned long long *) ((unsigned long)               \
 113                                  ((___area) + (___nr)*page_size +       \
 114                                   sizeof(pthread_mutex_t) +             \
 115                                   sizeof(unsigned long long) - 1) &     \
 116                                  ~(unsigned long)(sizeof(unsigned long long) \
 117                                                   -  1)))
 118
 119 const char *examples =
 120     "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
 121     "./userfaultfd anon 100 99999\n\n"
 122     "# Run share memory test on 1GiB region with 99 bounces:\n"
 123     "./userfaultfd shmem 1000 99\n\n"
 124     "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
 125     "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
 126     "# Run the same hugetlb test but using shmem:\n"
 127     "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
 128     "# 10MiB-~6GiB 999 bounces anonymous test, "
 129     "continue forever unless an error triggers\n"
 130     "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
 131
 132 static void usage(void)
 133 {
 134         fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
 135                 "[hugetlbfs_file]\n\n");
 136         fprintf(stderr, "Supported <test type>: anon, hugetlb, "
 137                 "hugetlb_shared, shmem\n\n");
 138         fprintf(stderr, "Examples:\n\n");
 139         fprintf(stderr, "%s", examples);
 140         exit(1);
 141 }
 142
 143 #define uffd_error(code, fmt, ...)                                             \
 144         do {                                                                   \
 145                 fprintf(stderr, fmt, ##__VA_ARGS__);                           \
 146                 fprintf(stderr, ": %" PRId64 "\n", (int64_t)(code));           \
 147                 exit(1);                                                       \
 148         } while (0)
 149
 150 static void uffd_stats_reset(struct uffd_stats *uffd_stats,
 151                              unsigned long n_cpus)
 152 {
 153         int i;
 154
 155         for (i = 0; i < n_cpus; i++) {
 156                 uffd_stats[i].cpu = i;
 157                 uffd_stats[i].missing_faults = 0;
 158                 uffd_stats[i].wp_faults = 0;
 159                 uffd_stats[i].minor_faults = 0;
 160         }
 161 }
 162
 163 static void uffd_stats_report(struct uffd_stats *stats, int n_cpus)
 164 {
 165         int i;
 166         unsigned long long miss_total = 0, wp_total = 0, minor_total = 0;
 167
 168         for (i = 0; i < n_cpus; i++) {
 169                 miss_total += stats[i].missing_faults;
 170                 wp_total += stats[i].wp_faults;
 171                 minor_total += stats[i].minor_faults;
 172         }
 173
 174         printf("userfaults: %llu missing (", miss_total);
 175         for (i = 0; i < n_cpus; i++)
 176                 printf("%lu+", stats[i].missing_faults);
 177         printf("\b), %llu wp (", wp_total);
 178         for (i = 0; i < n_cpus; i++)
 179                 printf("%lu+", stats[i].wp_faults);
 180         printf("\b), %llu minor (", minor_total);
 181         for (i = 0; i < n_cpus; i++)
 182                 printf("%lu+", stats[i].minor_faults);
 183         printf("\b)\n");
 184 }
 185
 186 static int anon_release_pages(char *rel_area)
 187 {
 188         int ret = 0;
 189
 190         if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED)) {
 191                 perror("madvise");
 192                 ret = 1;
 193         }
 194
 195         return ret;
 196 }
 197
 198 static void anon_allocate_area(void **alloc_area)
 199 {
 200         if (posix_memalign(alloc_area, page_size, nr_pages * page_size)) {
 201                 fprintf(stderr, "out of memory\n");
 202                 *alloc_area = NULL;
 203         }
 204 }
 205
 206 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 207 {
 208 }
 209
 210 /* HugeTLB memory */
 211 static int hugetlb_release_pages(char *rel_area)
 212 {
 213         int ret = 0;
 214
 215         if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 216                                 rel_area == huge_fd_off0 ? 0 :
 217                                 nr_pages * page_size,
 218                                 nr_pages * page_size)) {
 219                 perror("fallocate");
 220                 ret = 1;
 221         }
 222
 223         return ret;
 224 }
 225
 226 static void hugetlb_allocate_area(void **alloc_area)
 227 {
 228         void *area_alias = NULL;
 229         char **alloc_area_alias;
 230
 231         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 232                            (map_shared ? MAP_SHARED : MAP_PRIVATE) |
 233                            MAP_HUGETLB,
 234                            huge_fd, *alloc_area == area_src ? 0 :
 235                            nr_pages * page_size);
 236         if (*alloc_area == MAP_FAILED) {
 237                 perror("mmap of hugetlbfs file failed");
 238                 goto fail;
 239         }
 240
 241         if (map_shared) {
 242                 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 243                                   MAP_SHARED | MAP_HUGETLB,
 244                                   huge_fd, *alloc_area == area_src ? 0 :
 245                                   nr_pages * page_size);
 246                 if (area_alias == MAP_FAILED) {
 247                         perror("mmap of hugetlb file alias failed");
 248                         goto fail_munmap;
 249                 }
 250         }
 251
 252         if (*alloc_area == area_src) {
 253                 huge_fd_off0 = *alloc_area;
 254                 alloc_area_alias = &area_src_alias;
 255         } else {
 256                 alloc_area_alias = &area_dst_alias;
 257         }
 258         if (area_alias)
 259                 *alloc_area_alias = area_alias;
 260
 261         return;
 262
 263 fail_munmap:
 264         if (munmap(*alloc_area, nr_pages * page_size) < 0) {
 265                 perror("hugetlb munmap");
 266                 exit(1);
 267         }
 268 fail:
 269         *alloc_area = NULL;
 270 }
 271
 272 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 273 {
 274         if (!map_shared)
 275                 return;
 276         /*
 277          * We can't zap just the pagetable with hugetlbfs because
 278          * MADV_DONTEED won't work. So exercise -EEXIST on a alias
 279          * mapping where the pagetables are not established initially,
 280          * this way we'll exercise the -EEXEC at the fs level.
 281          */
 282         *start = (unsigned long) area_dst_alias + offset;
 283 }
 284
 285 /* Shared memory */
 286 static int shmem_release_pages(char *rel_area)
 287 {
 288         int ret = 0;
 289
 290         if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE)) {
 291                 perror("madvise");
 292                 ret = 1;
 293         }
 294
 295         return ret;
 296 }
 297
 298 static void shmem_allocate_area(void **alloc_area)
 299 {
 300         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 301                            MAP_ANONYMOUS | MAP_SHARED, -1, 0);
 302         if (*alloc_area == MAP_FAILED) {
 303                 fprintf(stderr, "shared memory mmap failed\n");
 304                 *alloc_area = NULL;
 305         }
 306 }
 307
 308 struct uffd_test_ops {
 309         unsigned long expected_ioctls;
 310         void (*allocate_area)(void **alloc_area);
 311         int (*release_pages)(char *rel_area);
 312         void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
 313 };
 314
 315 #define SHMEM_EXPECTED_IOCTLS           ((1 << _UFFDIO_WAKE) | \
 316                                          (1 << _UFFDIO_COPY) | \
 317                                          (1 << _UFFDIO_ZEROPAGE))
 318
 319 #define ANON_EXPECTED_IOCTLS            ((1 << _UFFDIO_WAKE) | \
 320                                          (1 << _UFFDIO_COPY) | \
 321                                          (1 << _UFFDIO_ZEROPAGE) | \
 322                                          (1 << _UFFDIO_WRITEPROTECT))
 323
 324 static struct uffd_test_ops anon_uffd_test_ops = {
 325         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 326         .allocate_area  = anon_allocate_area,
 327         .release_pages  = anon_release_pages,
 328         .alias_mapping = noop_alias_mapping,
 329 };
 330
 331 static struct uffd_test_ops shmem_uffd_test_ops = {
 332         .expected_ioctls = SHMEM_EXPECTED_IOCTLS,
 333         .allocate_area  = shmem_allocate_area,
 334         .release_pages  = shmem_release_pages,
 335         .alias_mapping = noop_alias_mapping,
 336 };
 337
 338 static struct uffd_test_ops hugetlb_uffd_test_ops = {
 339         .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC & ~(1 << _UFFDIO_CONTINUE),
 340         .allocate_area  = hugetlb_allocate_area,
 341         .release_pages  = hugetlb_release_pages,
 342         .alias_mapping = hugetlb_alias_mapping,
 343 };
 344
 345 static struct uffd_test_ops *uffd_test_ops;
 346
 347 static int my_bcmp(char *str1, char *str2, size_t n)
 348 {
 349         unsigned long i;
 350         for (i = 0; i < n; i++)
 351                 if (str1[i] != str2[i])
 352                         return 1;
 353         return 0;
 354 }
 355
 356 static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
 357 {
 358         struct uffdio_writeprotect prms;
 359
 360         /* Write protection page faults */
 361         prms.range.start = start;
 362         prms.range.len = len;
 363         /* Undo write-protect, do wakeup after that */
 364         prms.mode = wp ? UFFDIO_WRITEPROTECT_MODE_WP : 0;
 365
 366         if (ioctl(ufd, UFFDIO_WRITEPROTECT, &prms)) {
 367                 fprintf(stderr, "clear WP failed for address 0x%" PRIx64 "\n",
 368                         (uint64_t)start);
 369                 exit(1);
 370         }
 371 }
 372
 373 static void continue_range(int ufd, __u64 start, __u64 len)
 374 {
 375         struct uffdio_continue req;
 376
 377         req.range.start = start;
 378         req.range.len = len;
 379         req.mode = 0;
 380
 381         if (ioctl(ufd, UFFDIO_CONTINUE, &req)) {
 382                 fprintf(stderr,
 383                         "UFFDIO_CONTINUE failed for address 0x%" PRIx64 "\n",
 384                         (uint64_t)start);
 385                 exit(1);
 386         }
 387 }
 388
 389 static void *locking_thread(void *arg)
 390 {
 391         unsigned long cpu = (unsigned long) arg;
 392         struct random_data rand;
 393         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 394         int32_t rand_nr;
 395         unsigned long long count;
 396         char randstate[64];
 397         unsigned int seed;
 398         time_t start;
 399
 400         if (bounces & BOUNCE_RANDOM) {
 401                 seed = (unsigned int) time(NULL) - bounces;
 402                 if (!(bounces & BOUNCE_RACINGFAULTS))
 403                         seed += cpu;
 404                 bzero(&rand, sizeof(rand));
 405                 bzero(&randstate, sizeof(randstate));
 406                 if (initstate_r(seed, randstate, sizeof(randstate), &rand)) {
 407                         fprintf(stderr, "srandom_r error\n");
 408                         exit(1);
 409                 }
 410         } else {
 411                 page_nr = -bounces;
 412                 if (!(bounces & BOUNCE_RACINGFAULTS))
 413                         page_nr += cpu * nr_pages_per_cpu;
 414         }
 415
 416         while (!finished) {
 417                 if (bounces & BOUNCE_RANDOM) {
 418                         if (random_r(&rand, &rand_nr)) {
 419                                 fprintf(stderr, "random_r 1 error\n");
 420                                 exit(1);
 421                         }
 422                         page_nr = rand_nr;
 423                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 424                                 if (random_r(&rand, &rand_nr)) {
 425                                         fprintf(stderr, "random_r 2 error\n");
 426                                         exit(1);
 427                                 }
 428                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 429                                            16;
 430                         }
 431                 } else
 432                         page_nr += 1;
 433                 page_nr %= nr_pages;
 434
 435                 start = time(NULL);
 436                 if (bounces & BOUNCE_VERIFY) {
 437                         count = *area_count(area_dst, page_nr);
 438                         if (!count) {
 439                                 fprintf(stderr,
 440                                         "page_nr %lu wrong count %Lu %Lu\n",
 441                                         page_nr, count,
 442                                         count_verify[page_nr]);
 443                                 exit(1);
 444                         }
 445
 446
 447                         /*
 448                          * We can't use bcmp (or memcmp) because that
 449                          * returns 0 erroneously if the memory is
 450                          * changing under it (even if the end of the
 451                          * page is never changing and always
 452                          * different).
 453                          */
 454 #if 1
 455                         if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
 456                                      page_size)) {
 457                                 fprintf(stderr,
 458                                         "my_bcmp page_nr %lu wrong count %Lu %Lu\n",
 459                                         page_nr, count, count_verify[page_nr]);
 460                                 exit(1);
 461                         }
 462 #else
 463                         unsigned long loops;
 464
 465                         loops = 0;
 466                         /* uncomment the below line to test with mutex */
 467                         /* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
 468                         while (!bcmp(area_dst + page_nr * page_size, zeropage,
 469                                      page_size)) {
 470                                 loops += 1;
 471                                 if (loops > 10)
 472                                         break;
 473                         }
 474                         /* uncomment below line to test with mutex */
 475                         /* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
 476                         if (loops) {
 477                                 fprintf(stderr,
 478                                         "page_nr %lu all zero thread %lu %p %lu\n",
 479                                         page_nr, cpu, area_dst + page_nr * page_size,
 480                                         loops);
 481                                 if (loops > 10)
 482                                         exit(1);
 483                         }
 484 #endif
 485                 }
 486
 487                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 488                 count = *area_count(area_dst, page_nr);
 489                 if (count != count_verify[page_nr]) {
 490                         fprintf(stderr,
 491                                 "page_nr %lu memory corruption %Lu %Lu\n",
 492                                 page_nr, count,
 493                                 count_verify[page_nr]); exit(1);
 494                 }
 495                 count++;
 496                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 497                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 498
 499                 if (time(NULL) - start > 1)
 500                         fprintf(stderr,
 501                                 "userfault too slow %ld "
 502                                 "possible false positive with overcommit\n",
 503                                 time(NULL) - start);
 504         }
 505
 506         return NULL;
 507 }
 508
 509 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 510                             unsigned long offset)
 511 {
 512         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 513                                      uffdio_copy->len,
 514                                      offset);
 515         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 516                 /* real retval in ufdio_copy.copy */
 517                 if (uffdio_copy->copy != -EEXIST) {
 518                         uffd_error(uffdio_copy->copy,
 519                                    "UFFDIO_COPY retry error");
 520                 }
 521         } else
 522                 uffd_error(uffdio_copy->copy, "UFFDIO_COPY retry unexpected");
 523 }
 524
 525 static int __copy_page(int ufd, unsigned long offset, bool retry)
 526 {
 527         struct uffdio_copy uffdio_copy;
 528
 529         if (offset >= nr_pages * page_size) {
 530                 fprintf(stderr, "unexpected offset %lu\n", offset);
 531                 exit(1);
 532         }
 533         uffdio_copy.dst = (unsigned long) area_dst + offset;
 534         uffdio_copy.src = (unsigned long) area_src + offset;
 535         uffdio_copy.len = page_size;
 536         if (test_uffdio_wp)
 537                 uffdio_copy.mode = UFFDIO_COPY_MODE_WP;
 538         else
 539                 uffdio_copy.mode = 0;
 540         uffdio_copy.copy = 0;
 541         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 542                 /* real retval in ufdio_copy.copy */
 543                 if (uffdio_copy.copy != -EEXIST)
 544                         uffd_error(uffdio_copy.copy, "UFFDIO_COPY error");
 545         } else if (uffdio_copy.copy != page_size) {
 546                 uffd_error(uffdio_copy.copy, "UFFDIO_COPY unexpected copy");
 547         } else {
 548                 if (test_uffdio_copy_eexist && retry) {
 549                         test_uffdio_copy_eexist = false;
 550                         retry_copy_page(ufd, &uffdio_copy, offset);
 551                 }
 552                 return 1;
 553         }
 554         return 0;
 555 }
 556
 557 static int copy_page_retry(int ufd, unsigned long offset)
 558 {
 559         return __copy_page(ufd, offset, true);
 560 }
 561
 562 static int copy_page(int ufd, unsigned long offset)
 563 {
 564         return __copy_page(ufd, offset, false);
 565 }
 566
 567 static int uffd_read_msg(int ufd, struct uffd_msg *msg)
 568 {
 569         int ret = read(uffd, msg, sizeof(*msg));
 570
 571         if (ret != sizeof(*msg)) {
 572                 if (ret < 0) {
 573                         if (errno == EAGAIN)
 574                                 return 1;
 575                         perror("blocking read error");
 576                 } else {
 577                         fprintf(stderr, "short read\n");
 578                 }
 579                 exit(1);
 580         }
 581
 582         return 0;
 583 }
 584
 585 static void uffd_handle_page_fault(struct uffd_msg *msg,
 586                                    struct uffd_stats *stats)
 587 {
 588         unsigned long offset;
 589
 590         if (msg->event != UFFD_EVENT_PAGEFAULT) {
 591                 fprintf(stderr, "unexpected msg event %u\n", msg->event);
 592                 exit(1);
 593         }
 594
 595         if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP) {
 596                 /* Write protect page faults */
 597                 wp_range(uffd, msg->arg.pagefault.address, page_size, false);
 598                 stats->wp_faults++;
 599         } else if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_MINOR) {
 600                 uint8_t *area;
 601                 int b;
 602
 603                 /*
 604                  * Minor page faults
 605                  *
 606                  * To prove we can modify the original range for testing
 607                  * purposes, we're going to bit flip this range before
 608                  * continuing.
 609                  *
 610                  * Note that this requires all minor page fault tests operate on
 611                  * area_dst (non-UFFD-registered) and area_dst_alias
 612                  * (UFFD-registered).
 613                  */
 614
 615                 area = (uint8_t *)(area_dst +
 616                                    ((char *)msg->arg.pagefault.address -
 617                                     area_dst_alias));
 618                 for (b = 0; b < page_size; ++b)
 619                         area[b] = ~area[b];
 620                 continue_range(uffd, msg->arg.pagefault.address, page_size);
 621                 stats->minor_faults++;
 622         } else {
 623                 /* Missing page faults */
 624                 if (bounces & BOUNCE_VERIFY &&
 625                     msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE) {
 626                         fprintf(stderr, "unexpected write fault\n");
 627                         exit(1);
 628                 }
 629
 630                 offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
 631                 offset &= ~(page_size-1);
 632
 633                 if (copy_page(uffd, offset))
 634                         stats->missing_faults++;
 635         }
 636 }
 637
 638 static void *uffd_poll_thread(void *arg)
 639 {
 640         struct uffd_stats *stats = (struct uffd_stats *)arg;
 641         unsigned long cpu = stats->cpu;
 642         struct pollfd pollfd[2];
 643         struct uffd_msg msg;
 644         struct uffdio_register uffd_reg;
 645         int ret;
 646         char tmp_chr;
 647
 648         pollfd[0].fd = uffd;
 649         pollfd[0].events = POLLIN;
 650         pollfd[1].fd = pipefd[cpu*2];
 651         pollfd[1].events = POLLIN;
 652
 653         for (;;) {
 654                 ret = poll(pollfd, 2, -1);
 655                 if (!ret) {
 656                         fprintf(stderr, "poll error %d\n", ret);
 657                         exit(1);
 658                 }
 659                 if (ret < 0) {
 660                         perror("poll");
 661                         exit(1);
 662                 }
 663                 if (pollfd[1].revents & POLLIN) {
 664                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1) {
 665                                 fprintf(stderr, "read pipefd error\n");
 666                                 exit(1);
 667                         }
 668                         break;
 669                 }
 670                 if (!(pollfd[0].revents & POLLIN)) {
 671                         fprintf(stderr, "pollfd[0].revents %d\n",
 672                                 pollfd[0].revents);
 673                         exit(1);
 674                 }
 675                 if (uffd_read_msg(uffd, &msg))
 676                         continue;
 677                 switch (msg.event) {
 678                 default:
 679                         fprintf(stderr, "unexpected msg event %u\n",
 680                                 msg.event); exit(1);
 681                         break;
 682                 case UFFD_EVENT_PAGEFAULT:
 683                         uffd_handle_page_fault(&msg, stats);
 684                         break;
 685                 case UFFD_EVENT_FORK:
 686                         close(uffd);
 687                         uffd = msg.arg.fork.ufd;
 688                         pollfd[0].fd = uffd;
 689                         break;
 690                 case UFFD_EVENT_REMOVE:
 691                         uffd_reg.range.start = msg.arg.remove.start;
 692                         uffd_reg.range.len = msg.arg.remove.end -
 693                                 msg.arg.remove.start;
 694                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range)) {
 695                                 fprintf(stderr, "remove failure\n");
 696                                 exit(1);
 697                         }
 698                         break;
 699                 case UFFD_EVENT_REMAP:
 700                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 701                         break;
 702                 }
 703         }
 704
 705         return NULL;
 706 }
 707
 708 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 709
 710 static void *uffd_read_thread(void *arg)
 711 {
 712         struct uffd_stats *stats = (struct uffd_stats *)arg;
 713         struct uffd_msg msg;
 714
 715         pthread_mutex_unlock(&uffd_read_mutex);
 716         /* from here cancellation is ok */
 717
 718         for (;;) {
 719                 if (uffd_read_msg(uffd, &msg))
 720                         continue;
 721                 uffd_handle_page_fault(&msg, stats);
 722         }
 723
 724         return NULL;
 725 }
 726
 727 static void *background_thread(void *arg)
 728 {
 729         unsigned long cpu = (unsigned long) arg;
 730         unsigned long page_nr, start_nr, mid_nr, end_nr;
 731
 732         start_nr = cpu * nr_pages_per_cpu;
 733         end_nr = (cpu+1) * nr_pages_per_cpu;
 734         mid_nr = (start_nr + end_nr) / 2;
 735
 736         /* Copy the first half of the pages */
 737         for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
 738                 copy_page_retry(uffd, page_nr * page_size);
 739
 740         /*
 741          * If we need to test uffd-wp, set it up now.  Then we'll have
 742          * at least the first half of the pages mapped already which
 743          * can be write-protected for testing
 744          */
 745         if (test_uffdio_wp)
 746                 wp_range(uffd, (unsigned long)area_dst + start_nr * page_size,
 747                         nr_pages_per_cpu * page_size, true);
 748
 749         /*
 750          * Continue the 2nd half of the page copying, handling write
 751          * protection faults if any
 752          */
 753         for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
 754                 copy_page_retry(uffd, page_nr * page_size);
 755
 756         return NULL;
 757 }
 758
 759 static int stress(struct uffd_stats *uffd_stats)
 760 {
 761         unsigned long cpu;
 762         pthread_t locking_threads[nr_cpus];
 763         pthread_t uffd_threads[nr_cpus];
 764         pthread_t background_threads[nr_cpus];
 765
 766         finished = 0;
 767         for (cpu = 0; cpu < nr_cpus; cpu++) {
 768                 if (pthread_create(&locking_threads[cpu], &attr,
 769                                    locking_thread, (void *)cpu))
 770                         return 1;
 771                 if (bounces & BOUNCE_POLL) {
 772                         if (pthread_create(&uffd_threads[cpu], &attr,
 773                                            uffd_poll_thread,
 774                                            (void *)&uffd_stats[cpu]))
 775                                 return 1;
 776                 } else {
 777                         if (pthread_create(&uffd_threads[cpu], &attr,
 778                                            uffd_read_thread,
 779                                            (void *)&uffd_stats[cpu]))
 780                                 return 1;
 781                         pthread_mutex_lock(&uffd_read_mutex);
 782                 }
 783                 if (pthread_create(&background_threads[cpu], &attr,
 784                                    background_thread, (void *)cpu))
 785                         return 1;
 786         }
 787         for (cpu = 0; cpu < nr_cpus; cpu++)
 788                 if (pthread_join(background_threads[cpu], NULL))
 789                         return 1;
 790
 791         /*
 792          * Be strict and immediately zap area_src, the whole area has
 793          * been transferred already by the background treads. The
 794          * area_src could then be faulted in in a racy way by still
 795          * running uffdio_threads reading zeropages after we zapped
 796          * area_src (but they're guaranteed to get -EEXIST from
 797          * UFFDIO_COPY without writing zero pages into area_dst
 798          * because the background threads already completed).
 799          */
 800         if (uffd_test_ops->release_pages(area_src))
 801                 return 1;
 802
 803
 804         finished = 1;
 805         for (cpu = 0; cpu < nr_cpus; cpu++)
 806                 if (pthread_join(locking_threads[cpu], NULL))
 807                         return 1;
 808
 809         for (cpu = 0; cpu < nr_cpus; cpu++) {
 810                 char c;
 811                 if (bounces & BOUNCE_POLL) {
 812                         if (write(pipefd[cpu*2+1], &c, 1) != 1) {
 813                                 fprintf(stderr, "pipefd write error\n");
 814                                 return 1;
 815                         }
 816                         if (pthread_join(uffd_threads[cpu],
 817                                          (void *)&uffd_stats[cpu]))
 818                                 return 1;
 819                 } else {
 820                         if (pthread_cancel(uffd_threads[cpu]))
 821                                 return 1;
 822                         if (pthread_join(uffd_threads[cpu], NULL))
 823                                 return 1;
 824                 }
 825         }
 826
 827         return 0;
 828 }
 829
 830 static int userfaultfd_open_ext(uint64_t *features)
 831 {
 832         struct uffdio_api uffdio_api;
 833
 834         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 835         if (uffd < 0) {
 836                 fprintf(stderr,
 837                         "userfaultfd syscall not available in this kernel\n");
 838                 return 1;
 839         }
 840         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 841
 842         uffdio_api.api = UFFD_API;
 843         uffdio_api.features = *features;
 844         if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
 845                 fprintf(stderr, "UFFDIO_API failed.\nPlease make sure to "
 846                         "run with either root or ptrace capability.\n");
 847                 return 1;
 848         }
 849         if (uffdio_api.api != UFFD_API) {
 850                 fprintf(stderr, "UFFDIO_API error: %" PRIu64 "\n",
 851                         (uint64_t)uffdio_api.api);
 852                 return 1;
 853         }
 854
 855         *features = uffdio_api.features;
 856         return 0;
 857 }
 858
 859 static int userfaultfd_open(uint64_t features)
 860 {
 861         return userfaultfd_open_ext(&features);
 862 }
 863
 864 sigjmp_buf jbuf, *sigbuf;
 865
 866 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 867 {
 868         if (sig == SIGBUS) {
 869                 if (sigbuf)
 870                         siglongjmp(*sigbuf, 1);
 871                 abort();
 872         }
 873 }
 874
 875 /*
 876  * For non-cooperative userfaultfd test we fork() a process that will
 877  * generate pagefaults, will mremap the area monitored by the
 878  * userfaultfd and at last this process will release the monitored
 879  * area.
 880  * For the anonymous and shared memory the area is divided into two
 881  * parts, the first part is accessed before mremap, and the second
 882  * part is accessed after mremap. Since hugetlbfs does not support
 883  * mremap, the entire monitored area is accessed in a single pass for
 884  * HUGETLB_TEST.
 885  * The release of the pages currently generates event for shmem and
 886  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 887  * for hugetlb.
 888  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 889  * monitored area, generate pagefaults and test that signal is delivered.
 890  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 891  * test robustness use case - we release monitored area, fork a process
 892  * that will generate pagefaults and verify signal is generated.
 893  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 894  * feature. Using monitor thread, verify no userfault events are generated.
 895  */
 896 static int faulting_process(int signal_test)
 897 {
 898         unsigned long nr;
 899         unsigned long long count;
 900         unsigned long split_nr_pages;
 901         unsigned long lastnr;
 902         struct sigaction act;
 903         unsigned long signalled = 0;
 904
 905         if (test_type != TEST_HUGETLB)
 906                 split_nr_pages = (nr_pages + 1) / 2;
 907         else
 908                 split_nr_pages = nr_pages;
 909
 910         if (signal_test) {
 911                 sigbuf = &jbuf;
 912                 memset(&act, 0, sizeof(act));
 913                 act.sa_sigaction = sighndl;
 914                 act.sa_flags = SA_SIGINFO;
 915                 if (sigaction(SIGBUS, &act, 0)) {
 916                         perror("sigaction");
 917                         return 1;
 918                 }
 919                 lastnr = (unsigned long)-1;
 920         }
 921
 922         for (nr = 0; nr < split_nr_pages; nr++) {
 923                 int steps = 1;
 924                 unsigned long offset = nr * page_size;
 925
 926                 if (signal_test) {
 927                         if (sigsetjmp(*sigbuf, 1) != 0) {
 928                                 if (steps == 1 && nr == lastnr) {
 929                                         fprintf(stderr, "Signal repeated\n");
 930                                         return 1;
 931                                 }
 932
 933                                 lastnr = nr;
 934                                 if (signal_test == 1) {
 935                                         if (steps == 1) {
 936                                                 /* This is a MISSING request */
 937                                                 steps++;
 938                                                 if (copy_page(uffd, offset))
 939                                                         signalled++;
 940                                         } else {
 941                                                 /* This is a WP request */
 942                                                 assert(steps == 2);
 943                                                 wp_range(uffd,
 944                                                          (__u64)area_dst +
 945                                                          offset,
 946                                                          page_size, false);
 947                                         }
 948                                 } else {
 949                                         signalled++;
 950                                         continue;
 951                                 }
 952                         }
 953                 }
 954
 955                 count = *area_count(area_dst, nr);
 956                 if (count != count_verify[nr]) {
 957                         fprintf(stderr,
 958                                 "nr %lu memory corruption %Lu %Lu\n",
 959                                 nr, count,
 960                                 count_verify[nr]);
 961                 }
 962                 /*
 963                  * Trigger write protection if there is by writing
 964                  * the same value back.
 965                  */
 966                 *area_count(area_dst, nr) = count;
 967         }
 968
 969         if (signal_test)
 970                 return signalled != split_nr_pages;
 971
 972         if (test_type == TEST_HUGETLB)
 973                 return 0;
 974
 975         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 976                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 977         if (area_dst == MAP_FAILED) {
 978                 perror("mremap");
 979                 exit(1);
 980         }
 981
 982         for (; nr < nr_pages; nr++) {
 983                 count = *area_count(area_dst, nr);
 984                 if (count != count_verify[nr]) {
 985                         fprintf(stderr,
 986                                 "nr %lu memory corruption %Lu %Lu\n",
 987                                 nr, count,
 988                                 count_verify[nr]); exit(1);
 989                 }
 990                 /*
 991                  * Trigger write protection if there is by writing
 992                  * the same value back.
 993                  */
 994                 *area_count(area_dst, nr) = count;
 995         }
 996
 997         if (uffd_test_ops->release_pages(area_dst))
 998                 return 1;
 999
1000         for (nr = 0; nr < nr_pages; nr++) {
1001                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size)) {
1002                         fprintf(stderr, "nr %lu is not zero\n", nr);
1003                         exit(1);
1004                 }
1005         }
1006
1007         return 0;
1008 }
1009
1010 static void retry_uffdio_zeropage(int ufd,
1011                                   struct uffdio_zeropage *uffdio_zeropage,
1012                                   unsigned long offset)
1013 {
1014         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
1015                                      uffdio_zeropage->range.len,
1016                                      offset);
1017         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
1018                 if (uffdio_zeropage->zeropage != -EEXIST) {
1019                         uffd_error(uffdio_zeropage->zeropage,
1020                                    "UFFDIO_ZEROPAGE retry error");
1021                 }
1022         } else {
1023                 uffd_error(uffdio_zeropage->zeropage,
1024                            "UFFDIO_ZEROPAGE retry unexpected");
1025         }
1026 }
1027
1028 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
1029 {
1030         struct uffdio_zeropage uffdio_zeropage;
1031         int ret;
1032         unsigned long has_zeropage;
1033         __s64 res;
1034
1035         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
1036
1037         if (offset >= nr_pages * page_size) {
1038                 fprintf(stderr, "unexpected offset %lu\n", offset);
1039                 exit(1);
1040         }
1041         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
1042         uffdio_zeropage.range.len = page_size;
1043         uffdio_zeropage.mode = 0;
1044         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
1045         res = uffdio_zeropage.zeropage;
1046         if (ret) {
1047                 /* real retval in ufdio_zeropage.zeropage */
1048                 if (has_zeropage) {
1049                         uffd_error(res, "UFFDIO_ZEROPAGE %s",
1050                                    res == -EEXIST ? "-EEXIST" : "error");
1051                 } else if (res != -EINVAL)
1052                         uffd_error(res, "UFFDIO_ZEROPAGE not -EINVAL");
1053         } else if (has_zeropage) {
1054                 if (res != page_size) {
1055                         uffd_error(res, "UFFDIO_ZEROPAGE unexpected");
1056                 } else {
1057                         if (test_uffdio_zeropage_eexist && retry) {
1058                                 test_uffdio_zeropage_eexist = false;
1059                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
1060                                                       offset);
1061                         }
1062                         return 1;
1063                 }
1064         } else
1065                 uffd_error(res, "UFFDIO_ZEROPAGE succeeded");
1066
1067         return 0;
1068 }
1069
1070 static int uffdio_zeropage(int ufd, unsigned long offset)
1071 {
1072         return __uffdio_zeropage(ufd, offset, false);
1073 }
1074
1075 /* exercise UFFDIO_ZEROPAGE */
1076 static int userfaultfd_zeropage_test(void)
1077 {
1078         struct uffdio_register uffdio_register;
1079         unsigned long expected_ioctls;
1080
1081         printf("testing UFFDIO_ZEROPAGE: ");
1082         fflush(stdout);
1083
1084         if (uffd_test_ops->release_pages(area_dst))
1085                 return 1;
1086
1087         if (userfaultfd_open(0))
1088                 return 1;
1089         uffdio_register.range.start = (unsigned long) area_dst;
1090         uffdio_register.range.len = nr_pages * page_size;
1091         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1092         if (test_uffdio_wp)
1093                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1094         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1095                 fprintf(stderr, "register failure\n");
1096                 exit(1);
1097         }
1098
1099         expected_ioctls = uffd_test_ops->expected_ioctls;
1100         if ((uffdio_register.ioctls & expected_ioctls) !=
1101             expected_ioctls) {
1102                 fprintf(stderr,
1103                         "unexpected missing ioctl for anon memory\n");
1104                 exit(1);
1105         }
1106
1107         if (uffdio_zeropage(uffd, 0)) {
1108                 if (my_bcmp(area_dst, zeropage, page_size)) {
1109                         fprintf(stderr, "zeropage is not zero\n");
1110                         exit(1);
1111                 }
1112         }
1113
1114         close(uffd);
1115         printf("done.\n");
1116         return 0;
1117 }
1118
1119 static int userfaultfd_events_test(void)
1120 {
1121         struct uffdio_register uffdio_register;
1122         unsigned long expected_ioctls;
1123         pthread_t uffd_mon;
1124         int err, features;
1125         pid_t pid;
1126         char c;
1127         struct uffd_stats stats = { 0 };
1128
1129         printf("testing events (fork, remap, remove): ");
1130         fflush(stdout);
1131
1132         if (uffd_test_ops->release_pages(area_dst))
1133                 return 1;
1134
1135         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
1136                 UFFD_FEATURE_EVENT_REMOVE;
1137         if (userfaultfd_open(features))
1138                 return 1;
1139         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1140
1141         uffdio_register.range.start = (unsigned long) area_dst;
1142         uffdio_register.range.len = nr_pages * page_size;
1143         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1144         if (test_uffdio_wp)
1145                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1146         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1147                 fprintf(stderr, "register failure\n");
1148                 exit(1);
1149         }
1150
1151         expected_ioctls = uffd_test_ops->expected_ioctls;
1152         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1153                 fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1154                 exit(1);
1155         }
1156
1157         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1158                 perror("uffd_poll_thread create");
1159                 exit(1);
1160         }
1161
1162         pid = fork();
1163         if (pid < 0) {
1164                 perror("fork");
1165                 exit(1);
1166         }
1167
1168         if (!pid)
1169                 exit(faulting_process(0));
1170
1171         waitpid(pid, &err, 0);
1172         if (err) {
1173                 fprintf(stderr, "faulting process failed\n");
1174                 exit(1);
1175         }
1176
1177         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1178                 perror("pipe write");
1179                 exit(1);
1180         }
1181         if (pthread_join(uffd_mon, NULL))
1182                 return 1;
1183
1184         close(uffd);
1185
1186         uffd_stats_report(&stats, 1);
1187
1188         return stats.missing_faults != nr_pages;
1189 }
1190
1191 static int userfaultfd_sig_test(void)
1192 {
1193         struct uffdio_register uffdio_register;
1194         unsigned long expected_ioctls;
1195         unsigned long userfaults;
1196         pthread_t uffd_mon;
1197         int err, features;
1198         pid_t pid;
1199         char c;
1200         struct uffd_stats stats = { 0 };
1201
1202         printf("testing signal delivery: ");
1203         fflush(stdout);
1204
1205         if (uffd_test_ops->release_pages(area_dst))
1206                 return 1;
1207
1208         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
1209         if (userfaultfd_open(features))
1210                 return 1;
1211         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1212
1213         uffdio_register.range.start = (unsigned long) area_dst;
1214         uffdio_register.range.len = nr_pages * page_size;
1215         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1216         if (test_uffdio_wp)
1217                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1218         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1219                 fprintf(stderr, "register failure\n");
1220                 exit(1);
1221         }
1222
1223         expected_ioctls = uffd_test_ops->expected_ioctls;
1224         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1225                 fprintf(stderr, "unexpected missing ioctl for anon memory\n");
1226                 exit(1);
1227         }
1228
1229         if (faulting_process(1)) {
1230                 fprintf(stderr, "faulting process failed\n");
1231                 exit(1);
1232         }
1233
1234         if (uffd_test_ops->release_pages(area_dst))
1235                 return 1;
1236
1237         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1238                 perror("uffd_poll_thread create");
1239                 exit(1);
1240         }
1241
1242         pid = fork();
1243         if (pid < 0) {
1244                 perror("fork");
1245                 exit(1);
1246         }
1247
1248         if (!pid)
1249                 exit(faulting_process(2));
1250
1251         waitpid(pid, &err, 0);
1252         if (err) {
1253                 fprintf(stderr, "faulting process failed\n");
1254                 exit(1);
1255         }
1256
1257         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1258                 perror("pipe write");
1259                 exit(1);
1260         }
1261         if (pthread_join(uffd_mon, (void **)&userfaults))
1262                 return 1;
1263
1264         printf("done.\n");
1265         if (userfaults)
1266                 fprintf(stderr, "Signal test failed, userfaults: %ld\n",
1267                         userfaults);
1268         close(uffd);
1269         return userfaults != 0;
1270 }
1271
1272 static int userfaultfd_minor_test(void)
1273 {
1274         struct uffdio_register uffdio_register;
1275         unsigned long expected_ioctls;
1276         unsigned long p;
1277         pthread_t uffd_mon;
1278         uint8_t expected_byte;
1279         void *expected_page;
1280         char c;
1281         struct uffd_stats stats = { 0 };
1282         uint64_t features = UFFD_FEATURE_MINOR_HUGETLBFS;
1283
1284         if (!test_uffdio_minor)
1285                 return 0;
1286
1287         printf("testing minor faults: ");
1288         fflush(stdout);
1289
1290         if (uffd_test_ops->release_pages(area_dst))
1291                 return 1;
1292
1293         if (userfaultfd_open_ext(&features))
1294                 return 1;
1295         /* If kernel reports the feature isn't supported, skip the test. */
1296         if (!(features & UFFD_FEATURE_MINOR_HUGETLBFS)) {
1297                 printf("skipping test due to lack of feature support\n");
1298                 fflush(stdout);
1299                 return 0;
1300         }
1301
1302         uffdio_register.range.start = (unsigned long)area_dst_alias;
1303         uffdio_register.range.len = nr_pages * page_size;
1304         uffdio_register.mode = UFFDIO_REGISTER_MODE_MINOR;
1305         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1306                 fprintf(stderr, "register failure\n");
1307                 exit(1);
1308         }
1309
1310         expected_ioctls = uffd_test_ops->expected_ioctls;
1311         expected_ioctls |= 1 << _UFFDIO_CONTINUE;
1312         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls) {
1313                 fprintf(stderr, "unexpected missing ioctl(s)\n");
1314                 exit(1);
1315         }
1316
1317         /*
1318          * After registering with UFFD, populate the non-UFFD-registered side of
1319          * the shared mapping. This should *not* trigger any UFFD minor faults.
1320          */
1321         for (p = 0; p < nr_pages; ++p) {
1322                 memset(area_dst + (p * page_size), p % ((uint8_t)-1),
1323                        page_size);
1324         }
1325
1326         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats)) {
1327                 perror("uffd_poll_thread create");
1328                 exit(1);
1329         }
1330
1331         /*
1332          * Read each of the pages back using the UFFD-registered mapping. We
1333          * expect that the first time we touch a page, it will result in a minor
1334          * fault. uffd_poll_thread will resolve the fault by bit-flipping the
1335          * page's contents, and then issuing a CONTINUE ioctl.
1336          */
1337
1338         if (posix_memalign(&expected_page, page_size, page_size)) {
1339                 fprintf(stderr, "out of memory\n");
1340                 return 1;
1341         }
1342
1343         for (p = 0; p < nr_pages; ++p) {
1344                 expected_byte = ~((uint8_t)(p % ((uint8_t)-1)));
1345                 memset(expected_page, expected_byte, page_size);
1346                 if (my_bcmp(expected_page, area_dst_alias + (p * page_size),
1347                             page_size)) {
1348                         fprintf(stderr,
1349                                 "unexpected page contents after minor fault\n");
1350                         exit(1);
1351                 }
1352         }
1353
1354         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c)) {
1355                 perror("pipe write");
1356                 exit(1);
1357         }
1358         if (pthread_join(uffd_mon, NULL))
1359                 return 1;
1360
1361         close(uffd);
1362
1363         uffd_stats_report(&stats, 1);
1364
1365         return stats.missing_faults != 0 || stats.minor_faults != nr_pages;
1366 }
1367
1368 static int userfaultfd_stress(void)
1369 {
1370         void *area;
1371         char *tmp_area;
1372         unsigned long nr;
1373         struct uffdio_register uffdio_register;
1374         unsigned long cpu;
1375         int err;
1376         struct uffd_stats uffd_stats[nr_cpus];
1377
1378         uffd_test_ops->allocate_area((void **)&area_src);
1379         if (!area_src)
1380                 return 1;
1381         uffd_test_ops->allocate_area((void **)&area_dst);
1382         if (!area_dst)
1383                 return 1;
1384
1385         if (userfaultfd_open(0))
1386                 return 1;
1387
1388         count_verify = malloc(nr_pages * sizeof(unsigned long long));
1389         if (!count_verify) {
1390                 perror("count_verify");
1391                 return 1;
1392         }
1393
1394         for (nr = 0; nr < nr_pages; nr++) {
1395                 *area_mutex(area_src, nr) = (pthread_mutex_t)
1396                         PTHREAD_MUTEX_INITIALIZER;
1397                 count_verify[nr] = *area_count(area_src, nr) = 1;
1398                 /*
1399                  * In the transition between 255 to 256, powerpc will
1400                  * read out of order in my_bcmp and see both bytes as
1401                  * zero, so leave a placeholder below always non-zero
1402                  * after the count, to avoid my_bcmp to trigger false
1403                  * positives.
1404                  */
1405                 *(area_count(area_src, nr) + 1) = 1;
1406         }
1407
1408         pipefd = malloc(sizeof(int) * nr_cpus * 2);
1409         if (!pipefd) {
1410                 perror("pipefd");
1411                 return 1;
1412         }
1413         for (cpu = 0; cpu < nr_cpus; cpu++) {
1414                 if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
1415                         perror("pipe");
1416                         return 1;
1417                 }
1418         }
1419
1420         if (posix_memalign(&area, page_size, page_size)) {
1421                 fprintf(stderr, "out of memory\n");
1422                 return 1;
1423         }
1424         zeropage = area;
1425         bzero(zeropage, page_size);
1426
1427         pthread_mutex_lock(&uffd_read_mutex);
1428
1429         pthread_attr_init(&attr);
1430         pthread_attr_setstacksize(&attr, 16*1024*1024);
1431
1432         err = 0;
1433         while (bounces--) {
1434                 unsigned long expected_ioctls;
1435
1436                 printf("bounces: %d, mode:", bounces);
1437                 if (bounces & BOUNCE_RANDOM)
1438                         printf(" rnd");
1439                 if (bounces & BOUNCE_RACINGFAULTS)
1440                         printf(" racing");
1441                 if (bounces & BOUNCE_VERIFY)
1442                         printf(" ver");
1443                 if (bounces & BOUNCE_POLL)
1444                         printf(" poll");
1445                 else
1446                         printf(" read");
1447                 printf(", ");
1448                 fflush(stdout);
1449
1450                 if (bounces & BOUNCE_POLL)
1451                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1452                 else
1453                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1454
1455                 /* register */
1456                 uffdio_register.range.start = (unsigned long) area_dst;
1457                 uffdio_register.range.len = nr_pages * page_size;
1458                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1459                 if (test_uffdio_wp)
1460                         uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1461                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1462                         fprintf(stderr, "register failure\n");
1463                         return 1;
1464                 }
1465                 expected_ioctls = uffd_test_ops->expected_ioctls;
1466                 if ((uffdio_register.ioctls & expected_ioctls) !=
1467                     expected_ioctls) {
1468                         fprintf(stderr,
1469                                 "unexpected missing ioctl for anon memory\n");
1470                         return 1;
1471                 }
1472
1473                 if (area_dst_alias) {
1474                         uffdio_register.range.start = (unsigned long)
1475                                 area_dst_alias;
1476                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
1477                                 fprintf(stderr, "register failure alias\n");
1478                                 return 1;
1479                         }
1480                 }
1481
1482                 /*
1483                  * The madvise done previously isn't enough: some
1484                  * uffd_thread could have read userfaults (one of
1485                  * those already resolved by the background thread)
1486                  * and it may be in the process of calling
1487                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1488                  * area_src and it would map a zero page in it (of
1489                  * course such a UFFDIO_COPY is perfectly safe as it'd
1490                  * return -EEXIST). The problem comes at the next
1491                  * bounce though: that racing UFFDIO_COPY would
1492                  * generate zeropages in the area_src, so invalidating
1493                  * the previous MADV_DONTNEED. Without this additional
1494                  * MADV_DONTNEED those zeropages leftovers in the
1495                  * area_src would lead to -EEXIST failure during the
1496                  * next bounce, effectively leaving a zeropage in the
1497                  * area_dst.
1498                  *
1499                  * Try to comment this out madvise to see the memory
1500                  * corruption being caught pretty quick.
1501                  *
1502                  * khugepaged is also inhibited to collapse THP after
1503                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1504                  * required to MADV_DONTNEED here.
1505                  */
1506                 if (uffd_test_ops->release_pages(area_dst))
1507                         return 1;
1508
1509                 uffd_stats_reset(uffd_stats, nr_cpus);
1510
1511                 /* bounce pass */
1512                 if (stress(uffd_stats))
1513                         return 1;
1514
1515                 /* Clear all the write protections if there is any */
1516                 if (test_uffdio_wp)
1517                         wp_range(uffd, (unsigned long)area_dst,
1518                                  nr_pages * page_size, false);
1519
1520                 /* unregister */
1521                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
1522                         fprintf(stderr, "unregister failure\n");
1523                         return 1;
1524                 }
1525                 if (area_dst_alias) {
1526                         uffdio_register.range.start = (unsigned long) area_dst;
1527                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1528                                   &uffdio_register.range)) {
1529                                 fprintf(stderr, "unregister failure alias\n");
1530                                 return 1;
1531                         }
1532                 }
1533
1534                 /* verification */
1535                 if (bounces & BOUNCE_VERIFY) {
1536                         for (nr = 0; nr < nr_pages; nr++) {
1537                                 if (*area_count(area_dst, nr) != count_verify[nr]) {
1538                                         fprintf(stderr,
1539                                                 "error area_count %Lu %Lu %lu\n",
1540                                                 *area_count(area_src, nr),
1541                                                 count_verify[nr],
1542                                                 nr);
1543                                         err = 1;
1544                                         bounces = 0;
1545                                 }
1546                         }
1547                 }
1548
1549                 /* prepare next bounce */
1550                 tmp_area = area_src;
1551                 area_src = area_dst;
1552                 area_dst = tmp_area;
1553
1554                 tmp_area = area_src_alias;
1555                 area_src_alias = area_dst_alias;
1556                 area_dst_alias = tmp_area;
1557
1558                 uffd_stats_report(uffd_stats, nr_cpus);
1559         }
1560
1561         if (err)
1562                 return err;
1563
1564         close(uffd);
1565         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1566                 || userfaultfd_events_test() || userfaultfd_minor_test();
1567 }
1568
1569 /*
1570  * Copied from mlock2-tests.c
1571  */
1572 unsigned long default_huge_page_size(void)
1573 {
1574         unsigned long hps = 0;
1575         char *line = NULL;
1576         size_t linelen = 0;
1577         FILE *f = fopen("/proc/meminfo", "r");
1578
1579         if (!f)
1580                 return 0;
1581         while (getline(&line, &linelen, f) > 0) {
1582                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1583                         hps <<= 10;
1584                         break;
1585                 }
1586         }
1587
1588         free(line);
1589         fclose(f);
1590         return hps;
1591 }
1592
1593 static void set_test_type(const char *type)
1594 {
1595         if (!strcmp(type, "anon")) {
1596                 test_type = TEST_ANON;
1597                 uffd_test_ops = &anon_uffd_test_ops;
1598                 /* Only enable write-protect test for anonymous test */
1599                 test_uffdio_wp = true;
1600         } else if (!strcmp(type, "hugetlb")) {
1601                 test_type = TEST_HUGETLB;
1602                 uffd_test_ops = &hugetlb_uffd_test_ops;
1603         } else if (!strcmp(type, "hugetlb_shared")) {
1604                 map_shared = true;
1605                 test_type = TEST_HUGETLB;
1606                 uffd_test_ops = &hugetlb_uffd_test_ops;
1607                 /* Minor faults require shared hugetlb; only enable here. */
1608                 test_uffdio_minor = true;
1609         } else if (!strcmp(type, "shmem")) {
1610                 map_shared = true;
1611                 test_type = TEST_SHMEM;
1612                 uffd_test_ops = &shmem_uffd_test_ops;
1613         } else {
1614                 fprintf(stderr, "Unknown test type: %s\n", type); exit(1);
1615         }
1616
1617         if (test_type == TEST_HUGETLB)
1618                 page_size = default_huge_page_size();
1619         else
1620                 page_size = sysconf(_SC_PAGE_SIZE);
1621
1622         if (!page_size) {
1623                 fprintf(stderr, "Unable to determine page size\n");
1624                 exit(2);
1625         }
1626         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1627             > page_size) {
1628                 fprintf(stderr, "Impossible to run this test\n");
1629                 exit(2);
1630         }
1631 }
1632
1633 static void sigalrm(int sig)
1634 {
1635         if (sig != SIGALRM)
1636                 abort();
1637         test_uffdio_copy_eexist = true;
1638         test_uffdio_zeropage_eexist = true;
1639         alarm(ALARM_INTERVAL_SECS);
1640 }
1641
1642 int main(int argc, char **argv)
1643 {
1644         if (argc < 4)
1645                 usage();
1646
1647         if (signal(SIGALRM, sigalrm) == SIG_ERR) {
1648                 fprintf(stderr, "failed to arm SIGALRM");
1649                 exit(1);
1650         }
1651         alarm(ALARM_INTERVAL_SECS);
1652
1653         set_test_type(argv[1]);
1654
1655         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1656         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1657                 nr_cpus;
1658         if (!nr_pages_per_cpu) {
1659                 fprintf(stderr, "invalid MiB\n");
1660                 usage();
1661         }
1662
1663         bounces = atoi(argv[3]);
1664         if (bounces <= 0) {
1665                 fprintf(stderr, "invalid bounces\n");
1666                 usage();
1667         }
1668         nr_pages = nr_pages_per_cpu * nr_cpus;
1669
1670         if (test_type == TEST_HUGETLB) {
1671                 if (argc < 5)
1672                         usage();
1673                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1674                 if (huge_fd < 0) {
1675                         fprintf(stderr, "Open of %s failed", argv[3]);
1676                         perror("open");
1677                         exit(1);
1678                 }
1679                 if (ftruncate(huge_fd, 0)) {
1680                         fprintf(stderr, "ftruncate %s to size 0 failed", argv[3]);
1681                         perror("ftruncate");
1682                         exit(1);
1683                 }
1684         }
1685         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1686                nr_pages, nr_pages_per_cpu);
1687         return userfaultfd_stress();
1688 }
1689
1690 #else /* __NR_userfaultfd */
1691
1692 #warning "missing __NR_userfaultfd definition"
1693
1694 int main(void)
1695 {
1696         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1697         return KSFT_SKIP;
1698 }
1699
1700 #endif /* __NR_userfaultfd */