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 shm_fd;
  89 static int huge_fd;
  90 static char *huge_fd_off0;
  91 static unsigned long long *count_verify;
  92 static int uffd = -1;
  93 static int uffd_flags, finished, *pipefd;
  94 static char *area_src, *area_src_alias, *area_dst, *area_dst_alias;
  95 static char *zeropage;
  96 pthread_attr_t attr;
  97
  98 /* Userfaultfd test statistics */
  99 struct uffd_stats {
 100         int cpu;
 101         unsigned long missing_faults;
 102         unsigned long wp_faults;
 103         unsigned long minor_faults;
 104 };
 105
 106 /* pthread_mutex_t starts at page offset 0 */
 107 #define area_mutex(___area, ___nr)                                      \
 108         ((pthread_mutex_t *) ((___area) + (___nr)*page_size))
 109 /*
 110  * count is placed in the page after pthread_mutex_t naturally aligned
 111  * to avoid non alignment faults on non-x86 archs.
 112  */
 113 #define area_count(___area, ___nr)                                      \
 114         ((volatile unsigned long long *) ((unsigned long)               \
 115                                  ((___area) + (___nr)*page_size +       \
 116                                   sizeof(pthread_mutex_t) +             \
 117                                   sizeof(unsigned long long) - 1) &     \
 118                                  ~(unsigned long)(sizeof(unsigned long long) \
 119                                                   -  1)))
 120
 121 const char *examples =
 122     "# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
 123     "./userfaultfd anon 100 99999\n\n"
 124     "# Run share memory test on 1GiB region with 99 bounces:\n"
 125     "./userfaultfd shmem 1000 99\n\n"
 126     "# Run hugetlb memory test on 256MiB region with 50 bounces (using /dev/hugepages/hugefile):\n"
 127     "./userfaultfd hugetlb 256 50 /dev/hugepages/hugefile\n\n"
 128     "# Run the same hugetlb test but using shmem:\n"
 129     "./userfaultfd hugetlb_shared 256 50 /dev/hugepages/hugefile\n\n"
 130     "# 10MiB-~6GiB 999 bounces anonymous test, "
 131     "continue forever unless an error triggers\n"
 132     "while ./userfaultfd anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";
 133
 134 static void usage(void)
 135 {
 136         fprintf(stderr, "\nUsage: ./userfaultfd <test type> <MiB> <bounces> "
 137                 "[hugetlbfs_file]\n\n");
 138         fprintf(stderr, "Supported <test type>: anon, hugetlb, "
 139                 "hugetlb_shared, shmem\n\n");
 140         fprintf(stderr, "Examples:\n\n");
 141         fprintf(stderr, "%s", examples);
 142         exit(1);
 143 }
 144
 145 #define _err(fmt, ...)                                          \
 146         do {                                                    \
 147                 int ret = errno;                                \
 148                 fprintf(stderr, "ERROR: " fmt, ##__VA_ARGS__);  \
 149                 fprintf(stderr, " (errno=%d, line=%d)\n",       \
 150                         ret, __LINE__);                         \
 151         } while (0)
 152
 153 #define err(fmt, ...)                           \
 154         do {                                    \
 155                 _err(fmt, ##__VA_ARGS__);       \
 156                 exit(1);                        \
 157         } while (0)
 158
 159 static void uffd_stats_reset(struct uffd_stats *uffd_stats,
 160                              unsigned long n_cpus)
 161 {
 162         int i;
 163
 164         for (i = 0; i < n_cpus; i++) {
 165                 uffd_stats[i].cpu = i;
 166                 uffd_stats[i].missing_faults = 0;
 167                 uffd_stats[i].wp_faults = 0;
 168                 uffd_stats[i].minor_faults = 0;
 169         }
 170 }
 171
 172 static void uffd_stats_report(struct uffd_stats *stats, int n_cpus)
 173 {
 174         int i;
 175         unsigned long long miss_total = 0, wp_total = 0, minor_total = 0;
 176
 177         for (i = 0; i < n_cpus; i++) {
 178                 miss_total += stats[i].missing_faults;
 179                 wp_total += stats[i].wp_faults;
 180                 minor_total += stats[i].minor_faults;
 181         }
 182
 183         printf("userfaults: ");
 184         if (miss_total) {
 185                 printf("%llu missing (", miss_total);
 186                 for (i = 0; i < n_cpus; i++)
 187                         printf("%lu+", stats[i].missing_faults);
 188                 printf("\b) ");
 189         }
 190         if (wp_total) {
 191                 printf("%llu wp (", wp_total);
 192                 for (i = 0; i < n_cpus; i++)
 193                         printf("%lu+", stats[i].wp_faults);
 194                 printf("\b) ");
 195         }
 196         if (minor_total) {
 197                 printf("%llu minor (", minor_total);
 198                 for (i = 0; i < n_cpus; i++)
 199                         printf("%lu+", stats[i].minor_faults);
 200                 printf("\b)");
 201         }
 202         printf("\n");
 203 }
 204
 205 static void anon_release_pages(char *rel_area)
 206 {
 207         if (madvise(rel_area, nr_pages * page_size, MADV_DONTNEED))
 208                 err("madvise(MADV_DONTNEED) failed");
 209 }
 210
 211 static void anon_allocate_area(void **alloc_area)
 212 {
 213         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 214                            MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
 215         if (*alloc_area == MAP_FAILED)
 216                 err("mmap of anonymous memory failed");
 217 }
 218
 219 static void noop_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 220 {
 221 }
 222
 223 static void hugetlb_release_pages(char *rel_area)
 224 {
 225         if (fallocate(huge_fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 226                       rel_area == huge_fd_off0 ? 0 : nr_pages * page_size,
 227                       nr_pages * page_size))
 228                 err("fallocate() failed");
 229 }
 230
 231 static void hugetlb_allocate_area(void **alloc_area)
 232 {
 233         void *area_alias = NULL;
 234         char **alloc_area_alias;
 235
 236         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 237                            (map_shared ? MAP_SHARED : MAP_PRIVATE) |
 238                            MAP_HUGETLB,
 239                            huge_fd, *alloc_area == area_src ? 0 :
 240                            nr_pages * page_size);
 241         if (*alloc_area == MAP_FAILED)
 242                 err("mmap of hugetlbfs file failed");
 243
 244         if (map_shared) {
 245                 area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 246                                   MAP_SHARED | MAP_HUGETLB,
 247                                   huge_fd, *alloc_area == area_src ? 0 :
 248                                   nr_pages * page_size);
 249                 if (area_alias == MAP_FAILED)
 250                         err("mmap of hugetlb file alias failed");
 251         }
 252
 253         if (*alloc_area == area_src) {
 254                 huge_fd_off0 = *alloc_area;
 255                 alloc_area_alias = &area_src_alias;
 256         } else {
 257                 alloc_area_alias = &area_dst_alias;
 258         }
 259         if (area_alias)
 260                 *alloc_area_alias = area_alias;
 261 }
 262
 263 static void hugetlb_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 264 {
 265         if (!map_shared)
 266                 return;
 267         /*
 268          * We can't zap just the pagetable with hugetlbfs because
 269          * MADV_DONTEED won't work. So exercise -EEXIST on a alias
 270          * mapping where the pagetables are not established initially,
 271          * this way we'll exercise the -EEXEC at the fs level.
 272          */
 273         *start = (unsigned long) area_dst_alias + offset;
 274 }
 275
 276 static void shmem_release_pages(char *rel_area)
 277 {
 278         if (madvise(rel_area, nr_pages * page_size, MADV_REMOVE))
 279                 err("madvise(MADV_REMOVE) failed");
 280 }
 281
 282 static void shmem_allocate_area(void **alloc_area)
 283 {
 284         void *area_alias = NULL;
 285         bool is_src = alloc_area == (void **)&area_src;
 286         unsigned long offset = is_src ? 0 : nr_pages * page_size;
 287
 288         *alloc_area = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 289                            MAP_SHARED, shm_fd, offset);
 290         if (*alloc_area == MAP_FAILED)
 291                 err("mmap of memfd failed");
 292
 293         area_alias = mmap(NULL, nr_pages * page_size, PROT_READ | PROT_WRITE,
 294                           MAP_SHARED, shm_fd, offset);
 295         if (area_alias == MAP_FAILED)
 296                 err("mmap of memfd alias failed");
 297
 298         if (is_src)
 299                 area_src_alias = area_alias;
 300         else
 301                 area_dst_alias = area_alias;
 302 }
 303
 304 static void shmem_alias_mapping(__u64 *start, size_t len, unsigned long offset)
 305 {
 306         *start = (unsigned long)area_dst_alias + offset;
 307 }
 308
 309 struct uffd_test_ops {
 310         unsigned long expected_ioctls;
 311         void (*allocate_area)(void **alloc_area);
 312         void (*release_pages)(char *rel_area);
 313         void (*alias_mapping)(__u64 *start, size_t len, unsigned long offset);
 314 };
 315
 316 #define SHMEM_EXPECTED_IOCTLS           ((1 << _UFFDIO_WAKE) | \
 317                                          (1 << _UFFDIO_COPY) | \
 318                                          (1 << _UFFDIO_ZEROPAGE))
 319
 320 #define ANON_EXPECTED_IOCTLS            ((1 << _UFFDIO_WAKE) | \
 321                                          (1 << _UFFDIO_COPY) | \
 322                                          (1 << _UFFDIO_ZEROPAGE) | \
 323                                          (1 << _UFFDIO_WRITEPROTECT))
 324
 325 static struct uffd_test_ops anon_uffd_test_ops = {
 326         .expected_ioctls = ANON_EXPECTED_IOCTLS,
 327         .allocate_area  = anon_allocate_area,
 328         .release_pages  = anon_release_pages,
 329         .alias_mapping = noop_alias_mapping,
 330 };
 331
 332 static struct uffd_test_ops shmem_uffd_test_ops = {
 333         .expected_ioctls = SHMEM_EXPECTED_IOCTLS,
 334         .allocate_area  = shmem_allocate_area,
 335         .release_pages  = shmem_release_pages,
 336         .alias_mapping = shmem_alias_mapping,
 337 };
 338
 339 static struct uffd_test_ops hugetlb_uffd_test_ops = {
 340         .expected_ioctls = UFFD_API_RANGE_IOCTLS_BASIC & ~(1 << _UFFDIO_CONTINUE),
 341         .allocate_area  = hugetlb_allocate_area,
 342         .release_pages  = hugetlb_release_pages,
 343         .alias_mapping = hugetlb_alias_mapping,
 344 };
 345
 346 static struct uffd_test_ops *uffd_test_ops;
 347
 348 static void userfaultfd_open(uint64_t *features)
 349 {
 350         struct uffdio_api uffdio_api;
 351
 352         uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK | UFFD_USER_MODE_ONLY);
 353         if (uffd < 0)
 354                 err("userfaultfd syscall not available in this kernel");
 355         uffd_flags = fcntl(uffd, F_GETFD, NULL);
 356
 357         uffdio_api.api = UFFD_API;
 358         uffdio_api.features = *features;
 359         if (ioctl(uffd, UFFDIO_API, &uffdio_api))
 360                 err("UFFDIO_API failed.\nPlease make sure to "
 361                     "run with either root or ptrace capability.");
 362         if (uffdio_api.api != UFFD_API)
 363                 err("UFFDIO_API error: %" PRIu64, (uint64_t)uffdio_api.api);
 364
 365         *features = uffdio_api.features;
 366 }
 367
 368 static inline void munmap_area(void **area)
 369 {
 370         if (*area)
 371                 if (munmap(*area, nr_pages * page_size))
 372                         err("munmap");
 373
 374         *area = NULL;
 375 }
 376
 377 static void uffd_test_ctx_clear(void)
 378 {
 379         size_t i;
 380
 381         if (pipefd) {
 382                 for (i = 0; i < nr_cpus * 2; ++i) {
 383                         if (close(pipefd[i]))
 384                                 err("close pipefd");
 385                 }
 386                 free(pipefd);
 387                 pipefd = NULL;
 388         }
 389
 390         if (count_verify) {
 391                 free(count_verify);
 392                 count_verify = NULL;
 393         }
 394
 395         if (uffd != -1) {
 396                 if (close(uffd))
 397                         err("close uffd");
 398                 uffd = -1;
 399         }
 400
 401         huge_fd_off0 = NULL;
 402         munmap_area((void **)&area_src);
 403         munmap_area((void **)&area_src_alias);
 404         munmap_area((void **)&area_dst);
 405         munmap_area((void **)&area_dst_alias);
 406 }
 407
 408 static void uffd_test_ctx_init_ext(uint64_t *features)
 409 {
 410         unsigned long nr, cpu;
 411
 412         uffd_test_ctx_clear();
 413
 414         uffd_test_ops->allocate_area((void **)&area_src);
 415         uffd_test_ops->allocate_area((void **)&area_dst);
 416
 417         userfaultfd_open(features);
 418
 419         count_verify = malloc(nr_pages * sizeof(unsigned long long));
 420         if (!count_verify)
 421                 err("count_verify");
 422
 423         for (nr = 0; nr < nr_pages; nr++) {
 424                 *area_mutex(area_src, nr) =
 425                         (pthread_mutex_t)PTHREAD_MUTEX_INITIALIZER;
 426                 count_verify[nr] = *area_count(area_src, nr) = 1;
 427                 /*
 428                  * In the transition between 255 to 256, powerpc will
 429                  * read out of order in my_bcmp and see both bytes as
 430                  * zero, so leave a placeholder below always non-zero
 431                  * after the count, to avoid my_bcmp to trigger false
 432                  * positives.
 433                  */
 434                 *(area_count(area_src, nr) + 1) = 1;
 435         }
 436
 437         /*
 438          * After initialization of area_src, we must explicitly release pages
 439          * for area_dst to make sure it's fully empty.  Otherwise we could have
 440          * some area_dst pages be errornously initialized with zero pages,
 441          * hence we could hit memory corruption later in the test.
 442          *
 443          * One example is when THP is globally enabled, above allocate_area()
 444          * calls could have the two areas merged into a single VMA (as they
 445          * will have the same VMA flags so they're mergeable).  When we
 446          * initialize the area_src above, it's possible that some part of
 447          * area_dst could have been faulted in via one huge THP that will be
 448          * shared between area_src and area_dst.  It could cause some of the
 449          * area_dst won't be trapped by missing userfaults.
 450          *
 451          * This release_pages() will guarantee even if that happened, we'll
 452          * proactively split the thp and drop any accidentally initialized
 453          * pages within area_dst.
 454          */
 455         uffd_test_ops->release_pages(area_dst);
 456
 457         pipefd = malloc(sizeof(int) * nr_cpus * 2);
 458         if (!pipefd)
 459                 err("pipefd");
 460         for (cpu = 0; cpu < nr_cpus; cpu++)
 461                 if (pipe2(&pipefd[cpu * 2], O_CLOEXEC | O_NONBLOCK))
 462                         err("pipe");
 463 }
 464
 465 static inline void uffd_test_ctx_init(uint64_t features)
 466 {
 467         uffd_test_ctx_init_ext(&features);
 468 }
 469
 470 static int my_bcmp(char *str1, char *str2, size_t n)
 471 {
 472         unsigned long i;
 473         for (i = 0; i < n; i++)
 474                 if (str1[i] != str2[i])
 475                         return 1;
 476         return 0;
 477 }
 478
 479 static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
 480 {
 481         struct uffdio_writeprotect prms;
 482
 483         /* Write protection page faults */
 484         prms.range.start = start;
 485         prms.range.len = len;
 486         /* Undo write-protect, do wakeup after that */
 487         prms.mode = wp ? UFFDIO_WRITEPROTECT_MODE_WP : 0;
 488
 489         if (ioctl(ufd, UFFDIO_WRITEPROTECT, &prms))
 490                 err("clear WP failed: address=0x%"PRIx64, (uint64_t)start);
 491 }
 492
 493 static void continue_range(int ufd, __u64 start, __u64 len)
 494 {
 495         struct uffdio_continue req;
 496         int ret;
 497
 498         req.range.start = start;
 499         req.range.len = len;
 500         req.mode = 0;
 501
 502         if (ioctl(ufd, UFFDIO_CONTINUE, &req))
 503                 err("UFFDIO_CONTINUE failed for address 0x%" PRIx64,
 504                     (uint64_t)start);
 505
 506         /*
 507          * Error handling within the kernel for continue is subtly different
 508          * from copy or zeropage, so it may be a source of bugs. Trigger an
 509          * error (-EEXIST) on purpose, to verify doing so doesn't cause a BUG.
 510          */
 511         req.mapped = 0;
 512         ret = ioctl(ufd, UFFDIO_CONTINUE, &req);
 513         if (ret >= 0 || req.mapped != -EEXIST)
 514                 err("failed to exercise UFFDIO_CONTINUE error handling, ret=%d, mapped=%" PRId64,
 515                     ret, (int64_t) req.mapped);
 516 }
 517
 518 static void *locking_thread(void *arg)
 519 {
 520         unsigned long cpu = (unsigned long) arg;
 521         struct random_data rand;
 522         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 523         int32_t rand_nr;
 524         unsigned long long count;
 525         char randstate[64];
 526         unsigned int seed;
 527
 528         if (bounces & BOUNCE_RANDOM) {
 529                 seed = (unsigned int) time(NULL) - bounces;
 530                 if (!(bounces & BOUNCE_RACINGFAULTS))
 531                         seed += cpu;
 532                 bzero(&rand, sizeof(rand));
 533                 bzero(&randstate, sizeof(randstate));
 534                 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
 535                         err("initstate_r failed");
 536         } else {
 537                 page_nr = -bounces;
 538                 if (!(bounces & BOUNCE_RACINGFAULTS))
 539                         page_nr += cpu * nr_pages_per_cpu;
 540         }
 541
 542         while (!finished) {
 543                 if (bounces & BOUNCE_RANDOM) {
 544                         if (random_r(&rand, &rand_nr))
 545                                 err("random_r failed");
 546                         page_nr = rand_nr;
 547                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 548                                 if (random_r(&rand, &rand_nr))
 549                                         err("random_r failed");
 550                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 551                                            16;
 552                         }
 553                 } else
 554                         page_nr += 1;
 555                 page_nr %= nr_pages;
 556                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 557                 count = *area_count(area_dst, page_nr);
 558                 if (count != count_verify[page_nr])
 559                         err("page_nr %lu memory corruption %llu %llu",
 560                             page_nr, count, count_verify[page_nr]);
 561                 count++;
 562                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 563                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 564         }
 565
 566         return NULL;
 567 }
 568
 569 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 570                             unsigned long offset)
 571 {
 572         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 573                                      uffdio_copy->len,
 574                                      offset);
 575         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 576                 /* real retval in ufdio_copy.copy */
 577                 if (uffdio_copy->copy != -EEXIST)
 578                         err("UFFDIO_COPY retry error: %"PRId64,
 579                             (int64_t)uffdio_copy->copy);
 580         } else {
 581                 err("UFFDIO_COPY retry unexpected: %"PRId64,
 582                     (int64_t)uffdio_copy->copy);
 583         }
 584 }
 585
 586 static void wake_range(int ufd, unsigned long addr, unsigned long len)
 587 {
 588         struct uffdio_range uffdio_wake;
 589
 590         uffdio_wake.start = addr;
 591         uffdio_wake.len = len;
 592
 593         if (ioctl(ufd, UFFDIO_WAKE, &uffdio_wake))
 594                 fprintf(stderr, "error waking %lu\n",
 595                         addr), exit(1);
 596 }
 597
 598 static int __copy_page(int ufd, unsigned long offset, bool retry)
 599 {
 600         struct uffdio_copy uffdio_copy;
 601
 602         if (offset >= nr_pages * page_size)
 603                 err("unexpected offset %lu\n", offset);
 604         uffdio_copy.dst = (unsigned long) area_dst + offset;
 605         uffdio_copy.src = (unsigned long) area_src + offset;
 606         uffdio_copy.len = page_size;
 607         if (test_uffdio_wp)
 608                 uffdio_copy.mode = UFFDIO_COPY_MODE_WP;
 609         else
 610                 uffdio_copy.mode = 0;
 611         uffdio_copy.copy = 0;
 612         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 613                 /* real retval in ufdio_copy.copy */
 614                 if (uffdio_copy.copy != -EEXIST)
 615                         err("UFFDIO_COPY error: %"PRId64,
 616                             (int64_t)uffdio_copy.copy);
 617                 wake_range(ufd, uffdio_copy.dst, page_size);
 618         } else if (uffdio_copy.copy != page_size) {
 619                 err("UFFDIO_COPY error: %"PRId64, (int64_t)uffdio_copy.copy);
 620         } else {
 621                 if (test_uffdio_copy_eexist && retry) {
 622                         test_uffdio_copy_eexist = false;
 623                         retry_copy_page(ufd, &uffdio_copy, offset);
 624                 }
 625                 return 1;
 626         }
 627         return 0;
 628 }
 629
 630 static int copy_page_retry(int ufd, unsigned long offset)
 631 {
 632         return __copy_page(ufd, offset, true);
 633 }
 634
 635 static int copy_page(int ufd, unsigned long offset)
 636 {
 637         return __copy_page(ufd, offset, false);
 638 }
 639
 640 static int uffd_read_msg(int ufd, struct uffd_msg *msg)
 641 {
 642         int ret = read(uffd, msg, sizeof(*msg));
 643
 644         if (ret != sizeof(*msg)) {
 645                 if (ret < 0) {
 646                         if (errno == EAGAIN)
 647                                 return 1;
 648                         err("blocking read error");
 649                 } else {
 650                         err("short read");
 651                 }
 652         }
 653
 654         return 0;
 655 }
 656
 657 static void uffd_handle_page_fault(struct uffd_msg *msg,
 658                                    struct uffd_stats *stats)
 659 {
 660         unsigned long offset;
 661
 662         if (msg->event != UFFD_EVENT_PAGEFAULT)
 663                 err("unexpected msg event %u", msg->event);
 664
 665         if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP) {
 666                 /* Write protect page faults */
 667                 wp_range(uffd, msg->arg.pagefault.address, page_size, false);
 668                 stats->wp_faults++;
 669         } else if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_MINOR) {
 670                 uint8_t *area;
 671                 int b;
 672
 673                 /*
 674                  * Minor page faults
 675                  *
 676                  * To prove we can modify the original range for testing
 677                  * purposes, we're going to bit flip this range before
 678                  * continuing.
 679                  *
 680                  * Note that this requires all minor page fault tests operate on
 681                  * area_dst (non-UFFD-registered) and area_dst_alias
 682                  * (UFFD-registered).
 683                  */
 684
 685                 area = (uint8_t *)(area_dst +
 686                                    ((char *)msg->arg.pagefault.address -
 687                                     area_dst_alias));
 688                 for (b = 0; b < page_size; ++b)
 689                         area[b] = ~area[b];
 690                 continue_range(uffd, msg->arg.pagefault.address, page_size);
 691                 stats->minor_faults++;
 692         } else {
 693                 /* Missing page faults */
 694                 if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 695                         err("unexpected write fault");
 696
 697                 offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
 698                 offset &= ~(page_size-1);
 699
 700                 if (copy_page(uffd, offset))
 701                         stats->missing_faults++;
 702         }
 703 }
 704
 705 static void *uffd_poll_thread(void *arg)
 706 {
 707         struct uffd_stats *stats = (struct uffd_stats *)arg;
 708         unsigned long cpu = stats->cpu;
 709         struct pollfd pollfd[2];
 710         struct uffd_msg msg;
 711         struct uffdio_register uffd_reg;
 712         int ret;
 713         char tmp_chr;
 714
 715         pollfd[0].fd = uffd;
 716         pollfd[0].events = POLLIN;
 717         pollfd[1].fd = pipefd[cpu*2];
 718         pollfd[1].events = POLLIN;
 719
 720         for (;;) {
 721                 ret = poll(pollfd, 2, -1);
 722                 if (ret <= 0)
 723                         err("poll error: %d", ret);
 724                 if (pollfd[1].revents & POLLIN) {
 725                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
 726                                 err("read pipefd error");
 727                         break;
 728                 }
 729                 if (!(pollfd[0].revents & POLLIN))
 730                         err("pollfd[0].revents %d", pollfd[0].revents);
 731                 if (uffd_read_msg(uffd, &msg))
 732                         continue;
 733                 switch (msg.event) {
 734                 default:
 735                         err("unexpected msg event %u\n", msg.event);
 736                         break;
 737                 case UFFD_EVENT_PAGEFAULT:
 738                         uffd_handle_page_fault(&msg, stats);
 739                         break;
 740                 case UFFD_EVENT_FORK:
 741                         close(uffd);
 742                         uffd = msg.arg.fork.ufd;
 743                         pollfd[0].fd = uffd;
 744                         break;
 745                 case UFFD_EVENT_REMOVE:
 746                         uffd_reg.range.start = msg.arg.remove.start;
 747                         uffd_reg.range.len = msg.arg.remove.end -
 748                                 msg.arg.remove.start;
 749                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
 750                                 err("remove failure");
 751                         break;
 752                 case UFFD_EVENT_REMAP:
 753                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 754                         break;
 755                 }
 756         }
 757
 758         return NULL;
 759 }
 760
 761 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 762
 763 static void *uffd_read_thread(void *arg)
 764 {
 765         struct uffd_stats *stats = (struct uffd_stats *)arg;
 766         struct uffd_msg msg;
 767
 768         pthread_mutex_unlock(&uffd_read_mutex);
 769         /* from here cancellation is ok */
 770
 771         for (;;) {
 772                 if (uffd_read_msg(uffd, &msg))
 773                         continue;
 774                 uffd_handle_page_fault(&msg, stats);
 775         }
 776
 777         return NULL;
 778 }
 779
 780 static void *background_thread(void *arg)
 781 {
 782         unsigned long cpu = (unsigned long) arg;
 783         unsigned long page_nr, start_nr, mid_nr, end_nr;
 784
 785         start_nr = cpu * nr_pages_per_cpu;
 786         end_nr = (cpu+1) * nr_pages_per_cpu;
 787         mid_nr = (start_nr + end_nr) / 2;
 788
 789         /* Copy the first half of the pages */
 790         for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
 791                 copy_page_retry(uffd, page_nr * page_size);
 792
 793         /*
 794          * If we need to test uffd-wp, set it up now.  Then we'll have
 795          * at least the first half of the pages mapped already which
 796          * can be write-protected for testing
 797          */
 798         if (test_uffdio_wp)
 799                 wp_range(uffd, (unsigned long)area_dst + start_nr * page_size,
 800                         nr_pages_per_cpu * page_size, true);
 801
 802         /*
 803          * Continue the 2nd half of the page copying, handling write
 804          * protection faults if any
 805          */
 806         for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
 807                 copy_page_retry(uffd, page_nr * page_size);
 808
 809         return NULL;
 810 }
 811
 812 static int stress(struct uffd_stats *uffd_stats)
 813 {
 814         unsigned long cpu;
 815         pthread_t locking_threads[nr_cpus];
 816         pthread_t uffd_threads[nr_cpus];
 817         pthread_t background_threads[nr_cpus];
 818
 819         finished = 0;
 820         for (cpu = 0; cpu < nr_cpus; cpu++) {
 821                 if (pthread_create(&locking_threads[cpu], &attr,
 822                                    locking_thread, (void *)cpu))
 823                         return 1;
 824                 if (bounces & BOUNCE_POLL) {
 825                         if (pthread_create(&uffd_threads[cpu], &attr,
 826                                            uffd_poll_thread,
 827                                            (void *)&uffd_stats[cpu]))
 828                                 return 1;
 829                 } else {
 830                         if (pthread_create(&uffd_threads[cpu], &attr,
 831                                            uffd_read_thread,
 832                                            (void *)&uffd_stats[cpu]))
 833                                 return 1;
 834                         pthread_mutex_lock(&uffd_read_mutex);
 835                 }
 836                 if (pthread_create(&background_threads[cpu], &attr,
 837                                    background_thread, (void *)cpu))
 838                         return 1;
 839         }
 840         for (cpu = 0; cpu < nr_cpus; cpu++)
 841                 if (pthread_join(background_threads[cpu], NULL))
 842                         return 1;
 843
 844         /*
 845          * Be strict and immediately zap area_src, the whole area has
 846          * been transferred already by the background treads. The
 847          * area_src could then be faulted in in a racy way by still
 848          * running uffdio_threads reading zeropages after we zapped
 849          * area_src (but they're guaranteed to get -EEXIST from
 850          * UFFDIO_COPY without writing zero pages into area_dst
 851          * because the background threads already completed).
 852          */
 853         uffd_test_ops->release_pages(area_src);
 854
 855         finished = 1;
 856         for (cpu = 0; cpu < nr_cpus; cpu++)
 857                 if (pthread_join(locking_threads[cpu], NULL))
 858                         return 1;
 859
 860         for (cpu = 0; cpu < nr_cpus; cpu++) {
 861                 char c;
 862                 if (bounces & BOUNCE_POLL) {
 863                         if (write(pipefd[cpu*2+1], &c, 1) != 1)
 864                                 err("pipefd write error");
 865                         if (pthread_join(uffd_threads[cpu],
 866                                          (void *)&uffd_stats[cpu]))
 867                                 return 1;
 868                 } else {
 869                         if (pthread_cancel(uffd_threads[cpu]))
 870                                 return 1;
 871                         if (pthread_join(uffd_threads[cpu], NULL))
 872                                 return 1;
 873                 }
 874         }
 875
 876         return 0;
 877 }
 878
 879 sigjmp_buf jbuf, *sigbuf;
 880
 881 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 882 {
 883         if (sig == SIGBUS) {
 884                 if (sigbuf)
 885                         siglongjmp(*sigbuf, 1);
 886                 abort();
 887         }
 888 }
 889
 890 /*
 891  * For non-cooperative userfaultfd test we fork() a process that will
 892  * generate pagefaults, will mremap the area monitored by the
 893  * userfaultfd and at last this process will release the monitored
 894  * area.
 895  * For the anonymous and shared memory the area is divided into two
 896  * parts, the first part is accessed before mremap, and the second
 897  * part is accessed after mremap. Since hugetlbfs does not support
 898  * mremap, the entire monitored area is accessed in a single pass for
 899  * HUGETLB_TEST.
 900  * The release of the pages currently generates event for shmem and
 901  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 902  * for hugetlb.
 903  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 904  * monitored area, generate pagefaults and test that signal is delivered.
 905  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 906  * test robustness use case - we release monitored area, fork a process
 907  * that will generate pagefaults and verify signal is generated.
 908  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 909  * feature. Using monitor thread, verify no userfault events are generated.
 910  */
 911 static int faulting_process(int signal_test)
 912 {
 913         unsigned long nr;
 914         unsigned long long count;
 915         unsigned long split_nr_pages;
 916         unsigned long lastnr;
 917         struct sigaction act;
 918         unsigned long signalled = 0;
 919
 920         if (test_type != TEST_HUGETLB)
 921                 split_nr_pages = (nr_pages + 1) / 2;
 922         else
 923                 split_nr_pages = nr_pages;
 924
 925         if (signal_test) {
 926                 sigbuf = &jbuf;
 927                 memset(&act, 0, sizeof(act));
 928                 act.sa_sigaction = sighndl;
 929                 act.sa_flags = SA_SIGINFO;
 930                 if (sigaction(SIGBUS, &act, 0))
 931                         err("sigaction");
 932                 lastnr = (unsigned long)-1;
 933         }
 934
 935         for (nr = 0; nr < split_nr_pages; nr++) {
 936                 int steps = 1;
 937                 unsigned long offset = nr * page_size;
 938
 939                 if (signal_test) {
 940                         if (sigsetjmp(*sigbuf, 1) != 0) {
 941                                 if (steps == 1 && nr == lastnr)
 942                                         err("Signal repeated");
 943
 944                                 lastnr = nr;
 945                                 if (signal_test == 1) {
 946                                         if (steps == 1) {
 947                                                 /* This is a MISSING request */
 948                                                 steps++;
 949                                                 if (copy_page(uffd, offset))
 950                                                         signalled++;
 951                                         } else {
 952                                                 /* This is a WP request */
 953                                                 assert(steps == 2);
 954                                                 wp_range(uffd,
 955                                                          (__u64)area_dst +
 956                                                          offset,
 957                                                          page_size, false);
 958                                         }
 959                                 } else {
 960                                         signalled++;
 961                                         continue;
 962                                 }
 963                         }
 964                 }
 965
 966                 count = *area_count(area_dst, nr);
 967                 if (count != count_verify[nr])
 968                         err("nr %lu memory corruption %llu %llu\n",
 969                             nr, count, count_verify[nr]);
 970                 /*
 971                  * Trigger write protection if there is by writing
 972                  * the same value back.
 973                  */
 974                 *area_count(area_dst, nr) = count;
 975         }
 976
 977         if (signal_test)
 978                 return signalled != split_nr_pages;
 979
 980         if (test_type == TEST_HUGETLB)
 981                 return 0;
 982
 983         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 984                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 985         if (area_dst == MAP_FAILED)
 986                 err("mremap");
 987         /* Reset area_src since we just clobbered it */
 988         area_src = NULL;
 989
 990         for (; nr < nr_pages; nr++) {
 991                 count = *area_count(area_dst, nr);
 992                 if (count != count_verify[nr]) {
 993                         err("nr %lu memory corruption %llu %llu\n",
 994                             nr, count, count_verify[nr]);
 995                 }
 996                 /*
 997                  * Trigger write protection if there is by writing
 998                  * the same value back.
 999                  */
1000                 *area_count(area_dst, nr) = count;
1001         }
1002
1003         uffd_test_ops->release_pages(area_dst);
1004
1005         for (nr = 0; nr < nr_pages; nr++)
1006                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
1007                         err("nr %lu is not zero", nr);
1008
1009         return 0;
1010 }
1011
1012 static void retry_uffdio_zeropage(int ufd,
1013                                   struct uffdio_zeropage *uffdio_zeropage,
1014                                   unsigned long offset)
1015 {
1016         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
1017                                      uffdio_zeropage->range.len,
1018                                      offset);
1019         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
1020                 if (uffdio_zeropage->zeropage != -EEXIST)
1021                         err("UFFDIO_ZEROPAGE error: %"PRId64,
1022                             (int64_t)uffdio_zeropage->zeropage);
1023         } else {
1024                 err("UFFDIO_ZEROPAGE error: %"PRId64,
1025                     (int64_t)uffdio_zeropage->zeropage);
1026         }
1027 }
1028
1029 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
1030 {
1031         struct uffdio_zeropage uffdio_zeropage;
1032         int ret;
1033         unsigned long has_zeropage;
1034         __s64 res;
1035
1036         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
1037
1038         if (offset >= nr_pages * page_size)
1039                 err("unexpected offset %lu", offset);
1040         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
1041         uffdio_zeropage.range.len = page_size;
1042         uffdio_zeropage.mode = 0;
1043         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
1044         res = uffdio_zeropage.zeropage;
1045         if (ret) {
1046                 /* real retval in ufdio_zeropage.zeropage */
1047                 if (has_zeropage)
1048                         err("UFFDIO_ZEROPAGE error: %"PRId64, (int64_t)res);
1049                 else if (res != -EINVAL)
1050                         err("UFFDIO_ZEROPAGE not -EINVAL");
1051         } else if (has_zeropage) {
1052                 if (res != page_size) {
1053                         err("UFFDIO_ZEROPAGE unexpected size");
1054                 } else {
1055                         if (test_uffdio_zeropage_eexist && retry) {
1056                                 test_uffdio_zeropage_eexist = false;
1057                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
1058                                                       offset);
1059                         }
1060                         return 1;
1061                 }
1062         } else
1063                 err("UFFDIO_ZEROPAGE succeeded");
1064
1065         return 0;
1066 }
1067
1068 static int uffdio_zeropage(int ufd, unsigned long offset)
1069 {
1070         return __uffdio_zeropage(ufd, offset, false);
1071 }
1072
1073 /* exercise UFFDIO_ZEROPAGE */
1074 static int userfaultfd_zeropage_test(void)
1075 {
1076         struct uffdio_register uffdio_register;
1077         unsigned long expected_ioctls;
1078
1079         printf("testing UFFDIO_ZEROPAGE: ");
1080         fflush(stdout);
1081
1082         uffd_test_ctx_init(0);
1083
1084         uffdio_register.range.start = (unsigned long) area_dst;
1085         uffdio_register.range.len = nr_pages * page_size;
1086         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1087         if (test_uffdio_wp)
1088                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1089         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1090                 err("register failure");
1091
1092         expected_ioctls = uffd_test_ops->expected_ioctls;
1093         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1094                 err("unexpected missing ioctl for anon memory");
1095
1096         if (uffdio_zeropage(uffd, 0))
1097                 if (my_bcmp(area_dst, zeropage, page_size))
1098                         err("zeropage is not zero");
1099
1100         printf("done.\n");
1101         return 0;
1102 }
1103
1104 static int userfaultfd_events_test(void)
1105 {
1106         struct uffdio_register uffdio_register;
1107         unsigned long expected_ioctls;
1108         pthread_t uffd_mon;
1109         int err, features;
1110         pid_t pid;
1111         char c;
1112         struct uffd_stats stats = { 0 };
1113
1114         printf("testing events (fork, remap, remove): ");
1115         fflush(stdout);
1116
1117         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
1118                 UFFD_FEATURE_EVENT_REMOVE;
1119         uffd_test_ctx_init(features);
1120
1121         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1122
1123         uffdio_register.range.start = (unsigned long) area_dst;
1124         uffdio_register.range.len = nr_pages * page_size;
1125         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1126         if (test_uffdio_wp)
1127                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1128         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1129                 err("register failure");
1130
1131         expected_ioctls = uffd_test_ops->expected_ioctls;
1132         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1133                 err("unexpected missing ioctl for anon memory");
1134
1135         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1136                 err("uffd_poll_thread create");
1137
1138         pid = fork();
1139         if (pid < 0)
1140                 err("fork");
1141
1142         if (!pid)
1143                 exit(faulting_process(0));
1144
1145         waitpid(pid, &err, 0);
1146         if (err)
1147                 err("faulting process failed");
1148         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1149                 err("pipe write");
1150         if (pthread_join(uffd_mon, NULL))
1151                 return 1;
1152
1153         uffd_stats_report(&stats, 1);
1154
1155         return stats.missing_faults != nr_pages;
1156 }
1157
1158 static int userfaultfd_sig_test(void)
1159 {
1160         struct uffdio_register uffdio_register;
1161         unsigned long expected_ioctls;
1162         unsigned long userfaults;
1163         pthread_t uffd_mon;
1164         int err, features;
1165         pid_t pid;
1166         char c;
1167         struct uffd_stats stats = { 0 };
1168
1169         printf("testing signal delivery: ");
1170         fflush(stdout);
1171
1172         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
1173         uffd_test_ctx_init(features);
1174
1175         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1176
1177         uffdio_register.range.start = (unsigned long) area_dst;
1178         uffdio_register.range.len = nr_pages * page_size;
1179         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1180         if (test_uffdio_wp)
1181                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1182         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1183                 err("register failure");
1184
1185         expected_ioctls = uffd_test_ops->expected_ioctls;
1186         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1187                 err("unexpected missing ioctl for anon memory");
1188
1189         if (faulting_process(1))
1190                 err("faulting process failed");
1191
1192         uffd_test_ops->release_pages(area_dst);
1193
1194         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1195                 err("uffd_poll_thread create");
1196
1197         pid = fork();
1198         if (pid < 0)
1199                 err("fork");
1200
1201         if (!pid)
1202                 exit(faulting_process(2));
1203
1204         waitpid(pid, &err, 0);
1205         if (err)
1206                 err("faulting process failed");
1207         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1208                 err("pipe write");
1209         if (pthread_join(uffd_mon, (void **)&userfaults))
1210                 return 1;
1211
1212         printf("done.\n");
1213         if (userfaults)
1214                 err("Signal test failed, userfaults: %ld", userfaults);
1215
1216         return userfaults != 0;
1217 }
1218
1219 static int userfaultfd_minor_test(void)
1220 {
1221         struct uffdio_register uffdio_register;
1222         unsigned long expected_ioctls;
1223         unsigned long p;
1224         pthread_t uffd_mon;
1225         uint8_t expected_byte;
1226         void *expected_page;
1227         char c;
1228         struct uffd_stats stats = { 0 };
1229         uint64_t req_features, features_out;
1230
1231         if (!test_uffdio_minor)
1232                 return 0;
1233
1234         printf("testing minor faults: ");
1235         fflush(stdout);
1236
1237         if (test_type == TEST_HUGETLB)
1238                 req_features = UFFD_FEATURE_MINOR_HUGETLBFS;
1239         else if (test_type == TEST_SHMEM)
1240                 req_features = UFFD_FEATURE_MINOR_SHMEM;
1241         else
1242                 return 1;
1243
1244         features_out = req_features;
1245         uffd_test_ctx_init_ext(&features_out);
1246         /* If kernel reports required features aren't supported, skip test. */
1247         if ((features_out & req_features) != req_features) {
1248                 printf("skipping test due to lack of feature support\n");
1249                 fflush(stdout);
1250                 return 0;
1251         }
1252
1253         uffdio_register.range.start = (unsigned long)area_dst_alias;
1254         uffdio_register.range.len = nr_pages * page_size;
1255         uffdio_register.mode = UFFDIO_REGISTER_MODE_MINOR;
1256         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1257                 err("register failure");
1258
1259         expected_ioctls = uffd_test_ops->expected_ioctls;
1260         expected_ioctls |= 1 << _UFFDIO_CONTINUE;
1261         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1262                 err("unexpected missing ioctl(s)");
1263
1264         /*
1265          * After registering with UFFD, populate the non-UFFD-registered side of
1266          * the shared mapping. This should *not* trigger any UFFD minor faults.
1267          */
1268         for (p = 0; p < nr_pages; ++p) {
1269                 memset(area_dst + (p * page_size), p % ((uint8_t)-1),
1270                        page_size);
1271         }
1272
1273         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1274                 err("uffd_poll_thread create");
1275
1276         /*
1277          * Read each of the pages back using the UFFD-registered mapping. We
1278          * expect that the first time we touch a page, it will result in a minor
1279          * fault. uffd_poll_thread will resolve the fault by bit-flipping the
1280          * page's contents, and then issuing a CONTINUE ioctl.
1281          */
1282
1283         if (posix_memalign(&expected_page, page_size, page_size))
1284                 err("out of memory");
1285
1286         for (p = 0; p < nr_pages; ++p) {
1287                 expected_byte = ~((uint8_t)(p % ((uint8_t)-1)));
1288                 memset(expected_page, expected_byte, page_size);
1289                 if (my_bcmp(expected_page, area_dst_alias + (p * page_size),
1290                             page_size))
1291                         err("unexpected page contents after minor fault");
1292         }
1293
1294         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1295                 err("pipe write");
1296         if (pthread_join(uffd_mon, NULL))
1297                 return 1;
1298
1299         uffd_stats_report(&stats, 1);
1300
1301         return stats.missing_faults != 0 || stats.minor_faults != nr_pages;
1302 }
1303
1304 #define BIT_ULL(nr)                   (1ULL << (nr))
1305 #define PM_SOFT_DIRTY                 BIT_ULL(55)
1306 #define PM_MMAP_EXCLUSIVE             BIT_ULL(56)
1307 #define PM_UFFD_WP                    BIT_ULL(57)
1308 #define PM_FILE                       BIT_ULL(61)
1309 #define PM_SWAP                       BIT_ULL(62)
1310 #define PM_PRESENT                    BIT_ULL(63)
1311
1312 static int pagemap_open(void)
1313 {
1314         int fd = open("/proc/self/pagemap", O_RDONLY);
1315
1316         if (fd < 0)
1317                 err("open pagemap");
1318
1319         return fd;
1320 }
1321
1322 static uint64_t pagemap_read_vaddr(int fd, void *vaddr)
1323 {
1324         uint64_t value;
1325         int ret;
1326
1327         ret = pread(fd, &value, sizeof(uint64_t),
1328                     ((uint64_t)vaddr >> 12) * sizeof(uint64_t));
1329         if (ret != sizeof(uint64_t))
1330                 err("pread() on pagemap failed");
1331
1332         return value;
1333 }
1334
1335 /* This macro let __LINE__ works in err() */
1336 #define  pagemap_check_wp(value, wp) do {                               \
1337                 if (!!(value & PM_UFFD_WP) != wp)                       \
1338                         err("pagemap uffd-wp bit error: 0x%"PRIx64, value); \
1339         } while (0)
1340
1341 static int pagemap_test_fork(bool present)
1342 {
1343         pid_t child = fork();
1344         uint64_t value;
1345         int fd, result;
1346
1347         if (!child) {
1348                 /* Open the pagemap fd of the child itself */
1349                 fd = pagemap_open();
1350                 value = pagemap_read_vaddr(fd, area_dst);
1351                 /*
1352                  * After fork() uffd-wp bit should be gone as long as we're
1353                  * without UFFD_FEATURE_EVENT_FORK
1354                  */
1355                 pagemap_check_wp(value, false);
1356                 /* Succeed */
1357                 exit(0);
1358         }
1359         waitpid(child, &result, 0);
1360         return result;
1361 }
1362
1363 static void userfaultfd_pagemap_test(unsigned int test_pgsize)
1364 {
1365         struct uffdio_register uffdio_register;
1366         int pagemap_fd;
1367         uint64_t value;
1368
1369         /* Pagemap tests uffd-wp only */
1370         if (!test_uffdio_wp)
1371                 return;
1372
1373         /* Not enough memory to test this page size */
1374         if (test_pgsize > nr_pages * page_size)
1375                 return;
1376
1377         printf("testing uffd-wp with pagemap (pgsize=%u): ", test_pgsize);
1378         /* Flush so it doesn't flush twice in parent/child later */
1379         fflush(stdout);
1380
1381         uffd_test_ctx_init(0);
1382
1383         if (test_pgsize > page_size) {
1384                 /* This is a thp test */
1385                 if (madvise(area_dst, nr_pages * page_size, MADV_HUGEPAGE))
1386                         err("madvise(MADV_HUGEPAGE) failed");
1387         } else if (test_pgsize == page_size) {
1388                 /* This is normal page test; force no thp */
1389                 if (madvise(area_dst, nr_pages * page_size, MADV_NOHUGEPAGE))
1390                         err("madvise(MADV_NOHUGEPAGE) failed");
1391         }
1392
1393         uffdio_register.range.start = (unsigned long) area_dst;
1394         uffdio_register.range.len = nr_pages * page_size;
1395         uffdio_register.mode = UFFDIO_REGISTER_MODE_WP;
1396         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1397                 err("register failed");
1398
1399         pagemap_fd = pagemap_open();
1400
1401         /* Touch the page */
1402         *area_dst = 1;
1403         wp_range(uffd, (uint64_t)area_dst, test_pgsize, true);
1404         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1405         pagemap_check_wp(value, true);
1406         /* Make sure uffd-wp bit dropped when fork */
1407         if (pagemap_test_fork(true))
1408                 err("Detected stall uffd-wp bit in child");
1409
1410         /* Exclusive required or PAGEOUT won't work */
1411         if (!(value & PM_MMAP_EXCLUSIVE))
1412                 err("multiple mapping detected: 0x%"PRIx64, value);
1413
1414         if (madvise(area_dst, test_pgsize, MADV_PAGEOUT))
1415                 err("madvise(MADV_PAGEOUT) failed");
1416
1417         /* Uffd-wp should persist even swapped out */
1418         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1419         pagemap_check_wp(value, true);
1420         /* Make sure uffd-wp bit dropped when fork */
1421         if (pagemap_test_fork(false))
1422                 err("Detected stall uffd-wp bit in child");
1423
1424         /* Unprotect; this tests swap pte modifications */
1425         wp_range(uffd, (uint64_t)area_dst, page_size, false);
1426         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1427         pagemap_check_wp(value, false);
1428
1429         /* Fault in the page from disk */
1430         *area_dst = 2;
1431         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1432         pagemap_check_wp(value, false);
1433
1434         close(pagemap_fd);
1435         printf("done\n");
1436 }
1437
1438 static int userfaultfd_stress(void)
1439 {
1440         void *area;
1441         char *tmp_area;
1442         unsigned long nr;
1443         struct uffdio_register uffdio_register;
1444         struct uffd_stats uffd_stats[nr_cpus];
1445
1446         uffd_test_ctx_init(0);
1447
1448         if (posix_memalign(&area, page_size, page_size))
1449                 err("out of memory");
1450         zeropage = area;
1451         bzero(zeropage, page_size);
1452
1453         pthread_mutex_lock(&uffd_read_mutex);
1454
1455         pthread_attr_init(&attr);
1456         pthread_attr_setstacksize(&attr, 16*1024*1024);
1457
1458         while (bounces--) {
1459                 unsigned long expected_ioctls;
1460
1461                 printf("bounces: %d, mode:", bounces);
1462                 if (bounces & BOUNCE_RANDOM)
1463                         printf(" rnd");
1464                 if (bounces & BOUNCE_RACINGFAULTS)
1465                         printf(" racing");
1466                 if (bounces & BOUNCE_VERIFY)
1467                         printf(" ver");
1468                 if (bounces & BOUNCE_POLL)
1469                         printf(" poll");
1470                 else
1471                         printf(" read");
1472                 printf(", ");
1473                 fflush(stdout);
1474
1475                 if (bounces & BOUNCE_POLL)
1476                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1477                 else
1478                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1479
1480                 /* register */
1481                 uffdio_register.range.start = (unsigned long) area_dst;
1482                 uffdio_register.range.len = nr_pages * page_size;
1483                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1484                 if (test_uffdio_wp)
1485                         uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1486                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1487                         err("register failure");
1488                 expected_ioctls = uffd_test_ops->expected_ioctls;
1489                 if ((uffdio_register.ioctls & expected_ioctls) !=
1490                     expected_ioctls)
1491                         err("unexpected missing ioctl for anon memory");
1492
1493                 if (area_dst_alias) {
1494                         uffdio_register.range.start = (unsigned long)
1495                                 area_dst_alias;
1496                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1497                                 err("register failure alias");
1498                 }
1499
1500                 /*
1501                  * The madvise done previously isn't enough: some
1502                  * uffd_thread could have read userfaults (one of
1503                  * those already resolved by the background thread)
1504                  * and it may be in the process of calling
1505                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1506                  * area_src and it would map a zero page in it (of
1507                  * course such a UFFDIO_COPY is perfectly safe as it'd
1508                  * return -EEXIST). The problem comes at the next
1509                  * bounce though: that racing UFFDIO_COPY would
1510                  * generate zeropages in the area_src, so invalidating
1511                  * the previous MADV_DONTNEED. Without this additional
1512                  * MADV_DONTNEED those zeropages leftovers in the
1513                  * area_src would lead to -EEXIST failure during the
1514                  * next bounce, effectively leaving a zeropage in the
1515                  * area_dst.
1516                  *
1517                  * Try to comment this out madvise to see the memory
1518                  * corruption being caught pretty quick.
1519                  *
1520                  * khugepaged is also inhibited to collapse THP after
1521                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1522                  * required to MADV_DONTNEED here.
1523                  */
1524                 uffd_test_ops->release_pages(area_dst);
1525
1526                 uffd_stats_reset(uffd_stats, nr_cpus);
1527
1528                 /* bounce pass */
1529                 if (stress(uffd_stats))
1530                         return 1;
1531
1532                 /* Clear all the write protections if there is any */
1533                 if (test_uffdio_wp)
1534                         wp_range(uffd, (unsigned long)area_dst,
1535                                  nr_pages * page_size, false);
1536
1537                 /* unregister */
1538                 if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range))
1539                         err("unregister failure");
1540                 if (area_dst_alias) {
1541                         uffdio_register.range.start = (unsigned long) area_dst;
1542                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1543                                   &uffdio_register.range))
1544                                 err("unregister failure alias");
1545                 }
1546
1547                 /* verification */
1548                 if (bounces & BOUNCE_VERIFY)
1549                         for (nr = 0; nr < nr_pages; nr++)
1550                                 if (*area_count(area_dst, nr) != count_verify[nr])
1551                                         err("error area_count %llu %llu %lu\n",
1552                                             *area_count(area_src, nr),
1553                                             count_verify[nr], nr);
1554
1555                 /* prepare next bounce */
1556                 tmp_area = area_src;
1557                 area_src = area_dst;
1558                 area_dst = tmp_area;
1559
1560                 tmp_area = area_src_alias;
1561                 area_src_alias = area_dst_alias;
1562                 area_dst_alias = tmp_area;
1563
1564                 uffd_stats_report(uffd_stats, nr_cpus);
1565         }
1566
1567         if (test_type == TEST_ANON) {
1568                 /*
1569                  * shmem/hugetlb won't be able to run since they have different
1570                  * behavior on fork() (file-backed memory normally drops ptes
1571                  * directly when fork), meanwhile the pagemap test will verify
1572                  * pgtable entry of fork()ed child.
1573                  */
1574                 userfaultfd_pagemap_test(page_size);
1575                 /*
1576                  * Hard-code for x86_64 for now for 2M THP, as x86_64 is
1577                  * currently the only one that supports uffd-wp
1578                  */
1579                 userfaultfd_pagemap_test(page_size * 512);
1580         }
1581
1582         return userfaultfd_zeropage_test() || userfaultfd_sig_test()
1583                 || userfaultfd_events_test() || userfaultfd_minor_test();
1584 }
1585
1586 /*
1587  * Copied from mlock2-tests.c
1588  */
1589 unsigned long default_huge_page_size(void)
1590 {
1591         unsigned long hps = 0;
1592         char *line = NULL;
1593         size_t linelen = 0;
1594         FILE *f = fopen("/proc/meminfo", "r");
1595
1596         if (!f)
1597                 return 0;
1598         while (getline(&line, &linelen, f) > 0) {
1599                 if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
1600                         hps <<= 10;
1601                         break;
1602                 }
1603         }
1604
1605         free(line);
1606         fclose(f);
1607         return hps;
1608 }
1609
1610 static void set_test_type(const char *type)
1611 {
1612         if (!strcmp(type, "anon")) {
1613                 test_type = TEST_ANON;
1614                 uffd_test_ops = &anon_uffd_test_ops;
1615                 /* Only enable write-protect test for anonymous test */
1616                 test_uffdio_wp = true;
1617         } else if (!strcmp(type, "hugetlb")) {
1618                 test_type = TEST_HUGETLB;
1619                 uffd_test_ops = &hugetlb_uffd_test_ops;
1620         } else if (!strcmp(type, "hugetlb_shared")) {
1621                 map_shared = true;
1622                 test_type = TEST_HUGETLB;
1623                 uffd_test_ops = &hugetlb_uffd_test_ops;
1624                 /* Minor faults require shared hugetlb; only enable here. */
1625                 test_uffdio_minor = true;
1626         } else if (!strcmp(type, "shmem")) {
1627                 map_shared = true;
1628                 test_type = TEST_SHMEM;
1629                 uffd_test_ops = &shmem_uffd_test_ops;
1630                 test_uffdio_minor = true;
1631         } else {
1632                 err("Unknown test type: %s", type);
1633         }
1634
1635         if (test_type == TEST_HUGETLB)
1636                 page_size = default_huge_page_size();
1637         else
1638                 page_size = sysconf(_SC_PAGE_SIZE);
1639
1640         if (!page_size)
1641                 err("Unable to determine page size");
1642         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1643             > page_size)
1644                 err("Impossible to run this test");
1645 }
1646
1647 static void sigalrm(int sig)
1648 {
1649         if (sig != SIGALRM)
1650                 abort();
1651         test_uffdio_copy_eexist = true;
1652         test_uffdio_zeropage_eexist = true;
1653         alarm(ALARM_INTERVAL_SECS);
1654 }
1655
1656 int main(int argc, char **argv)
1657 {
1658         if (argc < 4)
1659                 usage();
1660
1661         if (signal(SIGALRM, sigalrm) == SIG_ERR)
1662                 err("failed to arm SIGALRM");
1663         alarm(ALARM_INTERVAL_SECS);
1664
1665         set_test_type(argv[1]);
1666
1667         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1668         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1669                 nr_cpus;
1670         if (!nr_pages_per_cpu) {
1671                 _err("invalid MiB");
1672                 usage();
1673         }
1674
1675         bounces = atoi(argv[3]);
1676         if (bounces <= 0) {
1677                 _err("invalid bounces");
1678                 usage();
1679         }
1680         nr_pages = nr_pages_per_cpu * nr_cpus;
1681
1682         if (test_type == TEST_HUGETLB) {
1683                 if (argc < 5)
1684                         usage();
1685                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1686                 if (huge_fd < 0)
1687                         err("Open of %s failed", argv[4]);
1688                 if (ftruncate(huge_fd, 0))
1689                         err("ftruncate %s to size 0 failed", argv[4]);
1690         } else if (test_type == TEST_SHMEM) {
1691                 shm_fd = memfd_create(argv[0], 0);
1692                 if (shm_fd < 0)
1693                         err("memfd_create");
1694                 if (ftruncate(shm_fd, nr_pages * page_size * 2))
1695                         err("ftruncate");
1696                 if (fallocate(shm_fd,
1697                               FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 0,
1698                               nr_pages * page_size * 2))
1699                         err("fallocate");
1700         }
1701         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1702                nr_pages, nr_pages_per_cpu);
1703         return userfaultfd_stress();
1704 }
1705
1706 #else /* __NR_userfaultfd */
1707
1708 #warning "missing __NR_userfaultfd definition"
1709
1710 int main(void)
1711 {
1712         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1713         return KSFT_SKIP;
1714 }
1715
1716 #endif /* __NR_userfaultfd */