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         uffd_test_ops->release_pages(area_src);
 418         uffd_test_ops->release_pages(area_dst);
 419
 420         userfaultfd_open(features);
 421
 422         count_verify = malloc(nr_pages * sizeof(unsigned long long));
 423         if (!count_verify)
 424                 err("count_verify");
 425
 426         for (nr = 0; nr < nr_pages; nr++) {
 427                 *area_mutex(area_src, nr) =
 428                         (pthread_mutex_t)PTHREAD_MUTEX_INITIALIZER;
 429                 count_verify[nr] = *area_count(area_src, nr) = 1;
 430                 /*
 431                  * In the transition between 255 to 256, powerpc will
 432                  * read out of order in my_bcmp and see both bytes as
 433                  * zero, so leave a placeholder below always non-zero
 434                  * after the count, to avoid my_bcmp to trigger false
 435                  * positives.
 436                  */
 437                 *(area_count(area_src, nr) + 1) = 1;
 438         }
 439
 440         pipefd = malloc(sizeof(int) * nr_cpus * 2);
 441         if (!pipefd)
 442                 err("pipefd");
 443         for (cpu = 0; cpu < nr_cpus; cpu++)
 444                 if (pipe2(&pipefd[cpu * 2], O_CLOEXEC | O_NONBLOCK))
 445                         err("pipe");
 446 }
 447
 448 static inline void uffd_test_ctx_init(uint64_t features)
 449 {
 450         uffd_test_ctx_init_ext(&features);
 451 }
 452
 453 static int my_bcmp(char *str1, char *str2, size_t n)
 454 {
 455         unsigned long i;
 456         for (i = 0; i < n; i++)
 457                 if (str1[i] != str2[i])
 458                         return 1;
 459         return 0;
 460 }
 461
 462 static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
 463 {
 464         struct uffdio_writeprotect prms;
 465
 466         /* Write protection page faults */
 467         prms.range.start = start;
 468         prms.range.len = len;
 469         /* Undo write-protect, do wakeup after that */
 470         prms.mode = wp ? UFFDIO_WRITEPROTECT_MODE_WP : 0;
 471
 472         if (ioctl(ufd, UFFDIO_WRITEPROTECT, &prms))
 473                 err("clear WP failed: address=0x%"PRIx64, (uint64_t)start);
 474 }
 475
 476 static void continue_range(int ufd, __u64 start, __u64 len)
 477 {
 478         struct uffdio_continue req;
 479         int ret;
 480
 481         req.range.start = start;
 482         req.range.len = len;
 483         req.mode = 0;
 484
 485         if (ioctl(ufd, UFFDIO_CONTINUE, &req))
 486                 err("UFFDIO_CONTINUE failed for address 0x%" PRIx64,
 487                     (uint64_t)start);
 488
 489         /*
 490          * Error handling within the kernel for continue is subtly different
 491          * from copy or zeropage, so it may be a source of bugs. Trigger an
 492          * error (-EEXIST) on purpose, to verify doing so doesn't cause a BUG.
 493          */
 494         req.mapped = 0;
 495         ret = ioctl(ufd, UFFDIO_CONTINUE, &req);
 496         if (ret >= 0 || req.mapped != -EEXIST)
 497                 err("failed to exercise UFFDIO_CONTINUE error handling, ret=%d, mapped=%" PRId64,
 498                     ret, (int64_t) req.mapped);
 499 }
 500
 501 static void *locking_thread(void *arg)
 502 {
 503         unsigned long cpu = (unsigned long) arg;
 504         struct random_data rand;
 505         unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
 506         int32_t rand_nr;
 507         unsigned long long count;
 508         char randstate[64];
 509         unsigned int seed;
 510
 511         if (bounces & BOUNCE_RANDOM) {
 512                 seed = (unsigned int) time(NULL) - bounces;
 513                 if (!(bounces & BOUNCE_RACINGFAULTS))
 514                         seed += cpu;
 515                 bzero(&rand, sizeof(rand));
 516                 bzero(&randstate, sizeof(randstate));
 517                 if (initstate_r(seed, randstate, sizeof(randstate), &rand))
 518                         err("initstate_r failed");
 519         } else {
 520                 page_nr = -bounces;
 521                 if (!(bounces & BOUNCE_RACINGFAULTS))
 522                         page_nr += cpu * nr_pages_per_cpu;
 523         }
 524
 525         while (!finished) {
 526                 if (bounces & BOUNCE_RANDOM) {
 527                         if (random_r(&rand, &rand_nr))
 528                                 err("random_r failed");
 529                         page_nr = rand_nr;
 530                         if (sizeof(page_nr) > sizeof(rand_nr)) {
 531                                 if (random_r(&rand, &rand_nr))
 532                                         err("random_r failed");
 533                                 page_nr |= (((unsigned long) rand_nr) << 16) <<
 534                                            16;
 535                         }
 536                 } else
 537                         page_nr += 1;
 538                 page_nr %= nr_pages;
 539                 pthread_mutex_lock(area_mutex(area_dst, page_nr));
 540                 count = *area_count(area_dst, page_nr);
 541                 if (count != count_verify[page_nr])
 542                         err("page_nr %lu memory corruption %llu %llu",
 543                             page_nr, count, count_verify[page_nr]);
 544                 count++;
 545                 *area_count(area_dst, page_nr) = count_verify[page_nr] = count;
 546                 pthread_mutex_unlock(area_mutex(area_dst, page_nr));
 547         }
 548
 549         return NULL;
 550 }
 551
 552 static void retry_copy_page(int ufd, struct uffdio_copy *uffdio_copy,
 553                             unsigned long offset)
 554 {
 555         uffd_test_ops->alias_mapping(&uffdio_copy->dst,
 556                                      uffdio_copy->len,
 557                                      offset);
 558         if (ioctl(ufd, UFFDIO_COPY, uffdio_copy)) {
 559                 /* real retval in ufdio_copy.copy */
 560                 if (uffdio_copy->copy != -EEXIST)
 561                         err("UFFDIO_COPY retry error: %"PRId64,
 562                             (int64_t)uffdio_copy->copy);
 563         } else {
 564                 err("UFFDIO_COPY retry unexpected: %"PRId64,
 565                     (int64_t)uffdio_copy->copy);
 566         }
 567 }
 568
 569 static void wake_range(int ufd, unsigned long addr, unsigned long len)
 570 {
 571         struct uffdio_range uffdio_wake;
 572
 573         uffdio_wake.start = addr;
 574         uffdio_wake.len = len;
 575
 576         if (ioctl(ufd, UFFDIO_WAKE, &uffdio_wake))
 577                 fprintf(stderr, "error waking %lu\n",
 578                         addr), exit(1);
 579 }
 580
 581 static int __copy_page(int ufd, unsigned long offset, bool retry)
 582 {
 583         struct uffdio_copy uffdio_copy;
 584
 585         if (offset >= nr_pages * page_size)
 586                 err("unexpected offset %lu\n", offset);
 587         uffdio_copy.dst = (unsigned long) area_dst + offset;
 588         uffdio_copy.src = (unsigned long) area_src + offset;
 589         uffdio_copy.len = page_size;
 590         if (test_uffdio_wp)
 591                 uffdio_copy.mode = UFFDIO_COPY_MODE_WP;
 592         else
 593                 uffdio_copy.mode = 0;
 594         uffdio_copy.copy = 0;
 595         if (ioctl(ufd, UFFDIO_COPY, &uffdio_copy)) {
 596                 /* real retval in ufdio_copy.copy */
 597                 if (uffdio_copy.copy != -EEXIST)
 598                         err("UFFDIO_COPY error: %"PRId64,
 599                             (int64_t)uffdio_copy.copy);
 600                 wake_range(ufd, uffdio_copy.dst, page_size);
 601         } else if (uffdio_copy.copy != page_size) {
 602                 err("UFFDIO_COPY error: %"PRId64, (int64_t)uffdio_copy.copy);
 603         } else {
 604                 if (test_uffdio_copy_eexist && retry) {
 605                         test_uffdio_copy_eexist = false;
 606                         retry_copy_page(ufd, &uffdio_copy, offset);
 607                 }
 608                 return 1;
 609         }
 610         return 0;
 611 }
 612
 613 static int copy_page_retry(int ufd, unsigned long offset)
 614 {
 615         return __copy_page(ufd, offset, true);
 616 }
 617
 618 static int copy_page(int ufd, unsigned long offset)
 619 {
 620         return __copy_page(ufd, offset, false);
 621 }
 622
 623 static int uffd_read_msg(int ufd, struct uffd_msg *msg)
 624 {
 625         int ret = read(uffd, msg, sizeof(*msg));
 626
 627         if (ret != sizeof(*msg)) {
 628                 if (ret < 0) {
 629                         if (errno == EAGAIN)
 630                                 return 1;
 631                         err("blocking read error");
 632                 } else {
 633                         err("short read");
 634                 }
 635         }
 636
 637         return 0;
 638 }
 639
 640 static void uffd_handle_page_fault(struct uffd_msg *msg,
 641                                    struct uffd_stats *stats)
 642 {
 643         unsigned long offset;
 644
 645         if (msg->event != UFFD_EVENT_PAGEFAULT)
 646                 err("unexpected msg event %u", msg->event);
 647
 648         if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP) {
 649                 /* Write protect page faults */
 650                 wp_range(uffd, msg->arg.pagefault.address, page_size, false);
 651                 stats->wp_faults++;
 652         } else if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_MINOR) {
 653                 uint8_t *area;
 654                 int b;
 655
 656                 /*
 657                  * Minor page faults
 658                  *
 659                  * To prove we can modify the original range for testing
 660                  * purposes, we're going to bit flip this range before
 661                  * continuing.
 662                  *
 663                  * Note that this requires all minor page fault tests operate on
 664                  * area_dst (non-UFFD-registered) and area_dst_alias
 665                  * (UFFD-registered).
 666                  */
 667
 668                 area = (uint8_t *)(area_dst +
 669                                    ((char *)msg->arg.pagefault.address -
 670                                     area_dst_alias));
 671                 for (b = 0; b < page_size; ++b)
 672                         area[b] = ~area[b];
 673                 continue_range(uffd, msg->arg.pagefault.address, page_size);
 674                 stats->minor_faults++;
 675         } else {
 676                 /* Missing page faults */
 677                 if (msg->arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
 678                         err("unexpected write fault");
 679
 680                 offset = (char *)(unsigned long)msg->arg.pagefault.address - area_dst;
 681                 offset &= ~(page_size-1);
 682
 683                 if (copy_page(uffd, offset))
 684                         stats->missing_faults++;
 685         }
 686 }
 687
 688 static void *uffd_poll_thread(void *arg)
 689 {
 690         struct uffd_stats *stats = (struct uffd_stats *)arg;
 691         unsigned long cpu = stats->cpu;
 692         struct pollfd pollfd[2];
 693         struct uffd_msg msg;
 694         struct uffdio_register uffd_reg;
 695         int ret;
 696         char tmp_chr;
 697
 698         pollfd[0].fd = uffd;
 699         pollfd[0].events = POLLIN;
 700         pollfd[1].fd = pipefd[cpu*2];
 701         pollfd[1].events = POLLIN;
 702
 703         for (;;) {
 704                 ret = poll(pollfd, 2, -1);
 705                 if (ret <= 0)
 706                         err("poll error: %d", ret);
 707                 if (pollfd[1].revents & POLLIN) {
 708                         if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
 709                                 err("read pipefd error");
 710                         break;
 711                 }
 712                 if (!(pollfd[0].revents & POLLIN))
 713                         err("pollfd[0].revents %d", pollfd[0].revents);
 714                 if (uffd_read_msg(uffd, &msg))
 715                         continue;
 716                 switch (msg.event) {
 717                 default:
 718                         err("unexpected msg event %u\n", msg.event);
 719                         break;
 720                 case UFFD_EVENT_PAGEFAULT:
 721                         uffd_handle_page_fault(&msg, stats);
 722                         break;
 723                 case UFFD_EVENT_FORK:
 724                         close(uffd);
 725                         uffd = msg.arg.fork.ufd;
 726                         pollfd[0].fd = uffd;
 727                         break;
 728                 case UFFD_EVENT_REMOVE:
 729                         uffd_reg.range.start = msg.arg.remove.start;
 730                         uffd_reg.range.len = msg.arg.remove.end -
 731                                 msg.arg.remove.start;
 732                         if (ioctl(uffd, UFFDIO_UNREGISTER, &uffd_reg.range))
 733                                 err("remove failure");
 734                         break;
 735                 case UFFD_EVENT_REMAP:
 736                         area_dst = (char *)(unsigned long)msg.arg.remap.to;
 737                         break;
 738                 }
 739         }
 740
 741         return NULL;
 742 }
 743
 744 pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
 745
 746 static void *uffd_read_thread(void *arg)
 747 {
 748         struct uffd_stats *stats = (struct uffd_stats *)arg;
 749         struct uffd_msg msg;
 750
 751         pthread_mutex_unlock(&uffd_read_mutex);
 752         /* from here cancellation is ok */
 753
 754         for (;;) {
 755                 if (uffd_read_msg(uffd, &msg))
 756                         continue;
 757                 uffd_handle_page_fault(&msg, stats);
 758         }
 759
 760         return NULL;
 761 }
 762
 763 static void *background_thread(void *arg)
 764 {
 765         unsigned long cpu = (unsigned long) arg;
 766         unsigned long page_nr, start_nr, mid_nr, end_nr;
 767
 768         start_nr = cpu * nr_pages_per_cpu;
 769         end_nr = (cpu+1) * nr_pages_per_cpu;
 770         mid_nr = (start_nr + end_nr) / 2;
 771
 772         /* Copy the first half of the pages */
 773         for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
 774                 copy_page_retry(uffd, page_nr * page_size);
 775
 776         /*
 777          * If we need to test uffd-wp, set it up now.  Then we'll have
 778          * at least the first half of the pages mapped already which
 779          * can be write-protected for testing
 780          */
 781         if (test_uffdio_wp)
 782                 wp_range(uffd, (unsigned long)area_dst + start_nr * page_size,
 783                         nr_pages_per_cpu * page_size, true);
 784
 785         /*
 786          * Continue the 2nd half of the page copying, handling write
 787          * protection faults if any
 788          */
 789         for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
 790                 copy_page_retry(uffd, page_nr * page_size);
 791
 792         return NULL;
 793 }
 794
 795 static int stress(struct uffd_stats *uffd_stats)
 796 {
 797         unsigned long cpu;
 798         pthread_t locking_threads[nr_cpus];
 799         pthread_t uffd_threads[nr_cpus];
 800         pthread_t background_threads[nr_cpus];
 801
 802         finished = 0;
 803         for (cpu = 0; cpu < nr_cpus; cpu++) {
 804                 if (pthread_create(&locking_threads[cpu], &attr,
 805                                    locking_thread, (void *)cpu))
 806                         return 1;
 807                 if (bounces & BOUNCE_POLL) {
 808                         if (pthread_create(&uffd_threads[cpu], &attr,
 809                                            uffd_poll_thread,
 810                                            (void *)&uffd_stats[cpu]))
 811                                 return 1;
 812                 } else {
 813                         if (pthread_create(&uffd_threads[cpu], &attr,
 814                                            uffd_read_thread,
 815                                            (void *)&uffd_stats[cpu]))
 816                                 return 1;
 817                         pthread_mutex_lock(&uffd_read_mutex);
 818                 }
 819                 if (pthread_create(&background_threads[cpu], &attr,
 820                                    background_thread, (void *)cpu))
 821                         return 1;
 822         }
 823         for (cpu = 0; cpu < nr_cpus; cpu++)
 824                 if (pthread_join(background_threads[cpu], NULL))
 825                         return 1;
 826
 827         /*
 828          * Be strict and immediately zap area_src, the whole area has
 829          * been transferred already by the background treads. The
 830          * area_src could then be faulted in in a racy way by still
 831          * running uffdio_threads reading zeropages after we zapped
 832          * area_src (but they're guaranteed to get -EEXIST from
 833          * UFFDIO_COPY without writing zero pages into area_dst
 834          * because the background threads already completed).
 835          */
 836         uffd_test_ops->release_pages(area_src);
 837
 838         finished = 1;
 839         for (cpu = 0; cpu < nr_cpus; cpu++)
 840                 if (pthread_join(locking_threads[cpu], NULL))
 841                         return 1;
 842
 843         for (cpu = 0; cpu < nr_cpus; cpu++) {
 844                 char c;
 845                 if (bounces & BOUNCE_POLL) {
 846                         if (write(pipefd[cpu*2+1], &c, 1) != 1)
 847                                 err("pipefd write error");
 848                         if (pthread_join(uffd_threads[cpu],
 849                                          (void *)&uffd_stats[cpu]))
 850                                 return 1;
 851                 } else {
 852                         if (pthread_cancel(uffd_threads[cpu]))
 853                                 return 1;
 854                         if (pthread_join(uffd_threads[cpu], NULL))
 855                                 return 1;
 856                 }
 857         }
 858
 859         return 0;
 860 }
 861
 862 sigjmp_buf jbuf, *sigbuf;
 863
 864 static void sighndl(int sig, siginfo_t *siginfo, void *ptr)
 865 {
 866         if (sig == SIGBUS) {
 867                 if (sigbuf)
 868                         siglongjmp(*sigbuf, 1);
 869                 abort();
 870         }
 871 }
 872
 873 /*
 874  * For non-cooperative userfaultfd test we fork() a process that will
 875  * generate pagefaults, will mremap the area monitored by the
 876  * userfaultfd and at last this process will release the monitored
 877  * area.
 878  * For the anonymous and shared memory the area is divided into two
 879  * parts, the first part is accessed before mremap, and the second
 880  * part is accessed after mremap. Since hugetlbfs does not support
 881  * mremap, the entire monitored area is accessed in a single pass for
 882  * HUGETLB_TEST.
 883  * The release of the pages currently generates event for shmem and
 884  * anonymous memory (UFFD_EVENT_REMOVE), hence it is not checked
 885  * for hugetlb.
 886  * For signal test(UFFD_FEATURE_SIGBUS), signal_test = 1, we register
 887  * monitored area, generate pagefaults and test that signal is delivered.
 888  * Use UFFDIO_COPY to allocate missing page and retry. For signal_test = 2
 889  * test robustness use case - we release monitored area, fork a process
 890  * that will generate pagefaults and verify signal is generated.
 891  * This also tests UFFD_FEATURE_EVENT_FORK event along with the signal
 892  * feature. Using monitor thread, verify no userfault events are generated.
 893  */
 894 static int faulting_process(int signal_test)
 895 {
 896         unsigned long nr;
 897         unsigned long long count;
 898         unsigned long split_nr_pages;
 899         unsigned long lastnr;
 900         struct sigaction act;
 901         unsigned long signalled = 0;
 902
 903         if (test_type != TEST_HUGETLB)
 904                 split_nr_pages = (nr_pages + 1) / 2;
 905         else
 906                 split_nr_pages = nr_pages;
 907
 908         if (signal_test) {
 909                 sigbuf = &jbuf;
 910                 memset(&act, 0, sizeof(act));
 911                 act.sa_sigaction = sighndl;
 912                 act.sa_flags = SA_SIGINFO;
 913                 if (sigaction(SIGBUS, &act, 0))
 914                         err("sigaction");
 915                 lastnr = (unsigned long)-1;
 916         }
 917
 918         for (nr = 0; nr < split_nr_pages; nr++) {
 919                 int steps = 1;
 920                 unsigned long offset = nr * page_size;
 921
 922                 if (signal_test) {
 923                         if (sigsetjmp(*sigbuf, 1) != 0) {
 924                                 if (steps == 1 && nr == lastnr)
 925                                         err("Signal repeated");
 926
 927                                 lastnr = nr;
 928                                 if (signal_test == 1) {
 929                                         if (steps == 1) {
 930                                                 /* This is a MISSING request */
 931                                                 steps++;
 932                                                 if (copy_page(uffd, offset))
 933                                                         signalled++;
 934                                         } else {
 935                                                 /* This is a WP request */
 936                                                 assert(steps == 2);
 937                                                 wp_range(uffd,
 938                                                          (__u64)area_dst +
 939                                                          offset,
 940                                                          page_size, false);
 941                                         }
 942                                 } else {
 943                                         signalled++;
 944                                         continue;
 945                                 }
 946                         }
 947                 }
 948
 949                 count = *area_count(area_dst, nr);
 950                 if (count != count_verify[nr])
 951                         err("nr %lu memory corruption %llu %llu\n",
 952                             nr, count, count_verify[nr]);
 953                 /*
 954                  * Trigger write protection if there is by writing
 955                  * the same value back.
 956                  */
 957                 *area_count(area_dst, nr) = count;
 958         }
 959
 960         if (signal_test)
 961                 return signalled != split_nr_pages;
 962
 963         if (test_type == TEST_HUGETLB)
 964                 return 0;
 965
 966         area_dst = mremap(area_dst, nr_pages * page_size,  nr_pages * page_size,
 967                           MREMAP_MAYMOVE | MREMAP_FIXED, area_src);
 968         if (area_dst == MAP_FAILED)
 969                 err("mremap");
 970         /* Reset area_src since we just clobbered it */
 971         area_src = NULL;
 972
 973         for (; nr < nr_pages; nr++) {
 974                 count = *area_count(area_dst, nr);
 975                 if (count != count_verify[nr]) {
 976                         err("nr %lu memory corruption %llu %llu\n",
 977                             nr, count, count_verify[nr]);
 978                 }
 979                 /*
 980                  * Trigger write protection if there is by writing
 981                  * the same value back.
 982                  */
 983                 *area_count(area_dst, nr) = count;
 984         }
 985
 986         uffd_test_ops->release_pages(area_dst);
 987
 988         for (nr = 0; nr < nr_pages; nr++)
 989                 if (my_bcmp(area_dst + nr * page_size, zeropage, page_size))
 990                         err("nr %lu is not zero", nr);
 991
 992         return 0;
 993 }
 994
 995 static void retry_uffdio_zeropage(int ufd,
 996                                   struct uffdio_zeropage *uffdio_zeropage,
 997                                   unsigned long offset)
 998 {
 999         uffd_test_ops->alias_mapping(&uffdio_zeropage->range.start,
1000                                      uffdio_zeropage->range.len,
1001                                      offset);
1002         if (ioctl(ufd, UFFDIO_ZEROPAGE, uffdio_zeropage)) {
1003                 if (uffdio_zeropage->zeropage != -EEXIST)
1004                         err("UFFDIO_ZEROPAGE error: %"PRId64,
1005                             (int64_t)uffdio_zeropage->zeropage);
1006         } else {
1007                 err("UFFDIO_ZEROPAGE error: %"PRId64,
1008                     (int64_t)uffdio_zeropage->zeropage);
1009         }
1010 }
1011
1012 static int __uffdio_zeropage(int ufd, unsigned long offset, bool retry)
1013 {
1014         struct uffdio_zeropage uffdio_zeropage;
1015         int ret;
1016         unsigned long has_zeropage;
1017         __s64 res;
1018
1019         has_zeropage = uffd_test_ops->expected_ioctls & (1 << _UFFDIO_ZEROPAGE);
1020
1021         if (offset >= nr_pages * page_size)
1022                 err("unexpected offset %lu", offset);
1023         uffdio_zeropage.range.start = (unsigned long) area_dst + offset;
1024         uffdio_zeropage.range.len = page_size;
1025         uffdio_zeropage.mode = 0;
1026         ret = ioctl(ufd, UFFDIO_ZEROPAGE, &uffdio_zeropage);
1027         res = uffdio_zeropage.zeropage;
1028         if (ret) {
1029                 /* real retval in ufdio_zeropage.zeropage */
1030                 if (has_zeropage)
1031                         err("UFFDIO_ZEROPAGE error: %"PRId64, (int64_t)res);
1032                 else if (res != -EINVAL)
1033                         err("UFFDIO_ZEROPAGE not -EINVAL");
1034         } else if (has_zeropage) {
1035                 if (res != page_size) {
1036                         err("UFFDIO_ZEROPAGE unexpected size");
1037                 } else {
1038                         if (test_uffdio_zeropage_eexist && retry) {
1039                                 test_uffdio_zeropage_eexist = false;
1040                                 retry_uffdio_zeropage(ufd, &uffdio_zeropage,
1041                                                       offset);
1042                         }
1043                         return 1;
1044                 }
1045         } else
1046                 err("UFFDIO_ZEROPAGE succeeded");
1047
1048         return 0;
1049 }
1050
1051 static int uffdio_zeropage(int ufd, unsigned long offset)
1052 {
1053         return __uffdio_zeropage(ufd, offset, false);
1054 }
1055
1056 /* exercise UFFDIO_ZEROPAGE */
1057 static int userfaultfd_zeropage_test(void)
1058 {
1059         struct uffdio_register uffdio_register;
1060         unsigned long expected_ioctls;
1061
1062         printf("testing UFFDIO_ZEROPAGE: ");
1063         fflush(stdout);
1064
1065         uffd_test_ctx_init(0);
1066
1067         uffdio_register.range.start = (unsigned long) area_dst;
1068         uffdio_register.range.len = nr_pages * page_size;
1069         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1070         if (test_uffdio_wp)
1071                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1072         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1073                 err("register failure");
1074
1075         expected_ioctls = uffd_test_ops->expected_ioctls;
1076         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1077                 err("unexpected missing ioctl for anon memory");
1078
1079         if (uffdio_zeropage(uffd, 0))
1080                 if (my_bcmp(area_dst, zeropage, page_size))
1081                         err("zeropage is not zero");
1082
1083         printf("done.\n");
1084         return 0;
1085 }
1086
1087 static int userfaultfd_events_test(void)
1088 {
1089         struct uffdio_register uffdio_register;
1090         unsigned long expected_ioctls;
1091         pthread_t uffd_mon;
1092         int err, features;
1093         pid_t pid;
1094         char c;
1095         struct uffd_stats stats = { 0 };
1096
1097         printf("testing events (fork, remap, remove): ");
1098         fflush(stdout);
1099
1100         features = UFFD_FEATURE_EVENT_FORK | UFFD_FEATURE_EVENT_REMAP |
1101                 UFFD_FEATURE_EVENT_REMOVE;
1102         uffd_test_ctx_init(features);
1103
1104         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1105
1106         uffdio_register.range.start = (unsigned long) area_dst;
1107         uffdio_register.range.len = nr_pages * page_size;
1108         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1109         if (test_uffdio_wp)
1110                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1111         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1112                 err("register failure");
1113
1114         expected_ioctls = uffd_test_ops->expected_ioctls;
1115         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1116                 err("unexpected missing ioctl for anon memory");
1117
1118         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1119                 err("uffd_poll_thread create");
1120
1121         pid = fork();
1122         if (pid < 0)
1123                 err("fork");
1124
1125         if (!pid)
1126                 exit(faulting_process(0));
1127
1128         waitpid(pid, &err, 0);
1129         if (err)
1130                 err("faulting process failed");
1131         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1132                 err("pipe write");
1133         if (pthread_join(uffd_mon, NULL))
1134                 return 1;
1135
1136         uffd_stats_report(&stats, 1);
1137
1138         return stats.missing_faults != nr_pages;
1139 }
1140
1141 static int userfaultfd_sig_test(void)
1142 {
1143         struct uffdio_register uffdio_register;
1144         unsigned long expected_ioctls;
1145         unsigned long userfaults;
1146         pthread_t uffd_mon;
1147         int err, features;
1148         pid_t pid;
1149         char c;
1150         struct uffd_stats stats = { 0 };
1151
1152         printf("testing signal delivery: ");
1153         fflush(stdout);
1154
1155         features = UFFD_FEATURE_EVENT_FORK|UFFD_FEATURE_SIGBUS;
1156         uffd_test_ctx_init(features);
1157
1158         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1159
1160         uffdio_register.range.start = (unsigned long) area_dst;
1161         uffdio_register.range.len = nr_pages * page_size;
1162         uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1163         if (test_uffdio_wp)
1164                 uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1165         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1166                 err("register failure");
1167
1168         expected_ioctls = uffd_test_ops->expected_ioctls;
1169         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1170                 err("unexpected missing ioctl for anon memory");
1171
1172         if (faulting_process(1))
1173                 err("faulting process failed");
1174
1175         uffd_test_ops->release_pages(area_dst);
1176
1177         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1178                 err("uffd_poll_thread create");
1179
1180         pid = fork();
1181         if (pid < 0)
1182                 err("fork");
1183
1184         if (!pid)
1185                 exit(faulting_process(2));
1186
1187         waitpid(pid, &err, 0);
1188         if (err)
1189                 err("faulting process failed");
1190         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1191                 err("pipe write");
1192         if (pthread_join(uffd_mon, (void **)&userfaults))
1193                 return 1;
1194
1195         printf("done.\n");
1196         if (userfaults)
1197                 err("Signal test failed, userfaults: %ld", userfaults);
1198
1199         return userfaults != 0;
1200 }
1201
1202 static int userfaultfd_minor_test(void)
1203 {
1204         struct uffdio_register uffdio_register;
1205         unsigned long expected_ioctls;
1206         unsigned long p;
1207         pthread_t uffd_mon;
1208         uint8_t expected_byte;
1209         void *expected_page;
1210         char c;
1211         struct uffd_stats stats = { 0 };
1212         uint64_t req_features, features_out;
1213
1214         if (!test_uffdio_minor)
1215                 return 0;
1216
1217         printf("testing minor faults: ");
1218         fflush(stdout);
1219
1220         if (test_type == TEST_HUGETLB)
1221                 req_features = UFFD_FEATURE_MINOR_HUGETLBFS;
1222         else if (test_type == TEST_SHMEM)
1223                 req_features = UFFD_FEATURE_MINOR_SHMEM;
1224         else
1225                 return 1;
1226
1227         features_out = req_features;
1228         uffd_test_ctx_init_ext(&features_out);
1229         /* If kernel reports required features aren't supported, skip test. */
1230         if ((features_out & req_features) != req_features) {
1231                 printf("skipping test due to lack of feature support\n");
1232                 fflush(stdout);
1233                 return 0;
1234         }
1235
1236         uffdio_register.range.start = (unsigned long)area_dst_alias;
1237         uffdio_register.range.len = nr_pages * page_size;
1238         uffdio_register.mode = UFFDIO_REGISTER_MODE_MINOR;
1239         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1240                 err("register failure");
1241
1242         expected_ioctls = uffd_test_ops->expected_ioctls;
1243         expected_ioctls |= 1 << _UFFDIO_CONTINUE;
1244         if ((uffdio_register.ioctls & expected_ioctls) != expected_ioctls)
1245                 err("unexpected missing ioctl(s)");
1246
1247         /*
1248          * After registering with UFFD, populate the non-UFFD-registered side of
1249          * the shared mapping. This should *not* trigger any UFFD minor faults.
1250          */
1251         for (p = 0; p < nr_pages; ++p) {
1252                 memset(area_dst + (p * page_size), p % ((uint8_t)-1),
1253                        page_size);
1254         }
1255
1256         if (pthread_create(&uffd_mon, &attr, uffd_poll_thread, &stats))
1257                 err("uffd_poll_thread create");
1258
1259         /*
1260          * Read each of the pages back using the UFFD-registered mapping. We
1261          * expect that the first time we touch a page, it will result in a minor
1262          * fault. uffd_poll_thread will resolve the fault by bit-flipping the
1263          * page's contents, and then issuing a CONTINUE ioctl.
1264          */
1265
1266         if (posix_memalign(&expected_page, page_size, page_size))
1267                 err("out of memory");
1268
1269         for (p = 0; p < nr_pages; ++p) {
1270                 expected_byte = ~((uint8_t)(p % ((uint8_t)-1)));
1271                 memset(expected_page, expected_byte, page_size);
1272                 if (my_bcmp(expected_page, area_dst_alias + (p * page_size),
1273                             page_size))
1274                         err("unexpected page contents after minor fault");
1275         }
1276
1277         if (write(pipefd[1], &c, sizeof(c)) != sizeof(c))
1278                 err("pipe write");
1279         if (pthread_join(uffd_mon, NULL))
1280                 return 1;
1281
1282         uffd_stats_report(&stats, 1);
1283
1284         return stats.missing_faults != 0 || stats.minor_faults != nr_pages;
1285 }
1286
1287 #define BIT_ULL(nr)                   (1ULL << (nr))
1288 #define PM_SOFT_DIRTY                 BIT_ULL(55)
1289 #define PM_MMAP_EXCLUSIVE             BIT_ULL(56)
1290 #define PM_UFFD_WP                    BIT_ULL(57)
1291 #define PM_FILE                       BIT_ULL(61)
1292 #define PM_SWAP                       BIT_ULL(62)
1293 #define PM_PRESENT                    BIT_ULL(63)
1294
1295 static int pagemap_open(void)
1296 {
1297         int fd = open("/proc/self/pagemap", O_RDONLY);
1298
1299         if (fd < 0)
1300                 err("open pagemap");
1301
1302         return fd;
1303 }
1304
1305 static uint64_t pagemap_read_vaddr(int fd, void *vaddr)
1306 {
1307         uint64_t value;
1308         int ret;
1309
1310         ret = pread(fd, &value, sizeof(uint64_t),
1311                     ((uint64_t)vaddr >> 12) * sizeof(uint64_t));
1312         if (ret != sizeof(uint64_t))
1313                 err("pread() on pagemap failed");
1314
1315         return value;
1316 }
1317
1318 /* This macro let __LINE__ works in err() */
1319 #define  pagemap_check_wp(value, wp) do {                               \
1320                 if (!!(value & PM_UFFD_WP) != wp)                       \
1321                         err("pagemap uffd-wp bit error: 0x%"PRIx64, value); \
1322         } while (0)
1323
1324 static int pagemap_test_fork(bool present)
1325 {
1326         pid_t child = fork();
1327         uint64_t value;
1328         int fd, result;
1329
1330         if (!child) {
1331                 /* Open the pagemap fd of the child itself */
1332                 fd = pagemap_open();
1333                 value = pagemap_read_vaddr(fd, area_dst);
1334                 /*
1335                  * After fork() uffd-wp bit should be gone as long as we're
1336                  * without UFFD_FEATURE_EVENT_FORK
1337                  */
1338                 pagemap_check_wp(value, false);
1339                 /* Succeed */
1340                 exit(0);
1341         }
1342         waitpid(child, &result, 0);
1343         return result;
1344 }
1345
1346 static void userfaultfd_pagemap_test(unsigned int test_pgsize)
1347 {
1348         struct uffdio_register uffdio_register;
1349         int pagemap_fd;
1350         uint64_t value;
1351
1352         /* Pagemap tests uffd-wp only */
1353         if (!test_uffdio_wp)
1354                 return;
1355
1356         /* Not enough memory to test this page size */
1357         if (test_pgsize > nr_pages * page_size)
1358                 return;
1359
1360         printf("testing uffd-wp with pagemap (pgsize=%u): ", test_pgsize);
1361         /* Flush so it doesn't flush twice in parent/child later */
1362         fflush(stdout);
1363
1364         uffd_test_ctx_init(0);
1365
1366         if (test_pgsize > page_size) {
1367                 /* This is a thp test */
1368                 if (madvise(area_dst, nr_pages * page_size, MADV_HUGEPAGE))
1369                         err("madvise(MADV_HUGEPAGE) failed");
1370         } else if (test_pgsize == page_size) {
1371                 /* This is normal page test; force no thp */
1372                 if (madvise(area_dst, nr_pages * page_size, MADV_NOHUGEPAGE))
1373                         err("madvise(MADV_NOHUGEPAGE) failed");
1374         }
1375
1376         uffdio_register.range.start = (unsigned long) area_dst;
1377         uffdio_register.range.len = nr_pages * page_size;
1378         uffdio_register.mode = UFFDIO_REGISTER_MODE_WP;
1379         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1380                 err("register failed");
1381
1382         pagemap_fd = pagemap_open();
1383
1384         /* Touch the page */
1385         *area_dst = 1;
1386         wp_range(uffd, (uint64_t)area_dst, test_pgsize, true);
1387         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1388         pagemap_check_wp(value, true);
1389         /* Make sure uffd-wp bit dropped when fork */
1390         if (pagemap_test_fork(true))
1391                 err("Detected stall uffd-wp bit in child");
1392
1393         /* Exclusive required or PAGEOUT won't work */
1394         if (!(value & PM_MMAP_EXCLUSIVE))
1395                 err("multiple mapping detected: 0x%"PRIx64, value);
1396
1397         if (madvise(area_dst, test_pgsize, MADV_PAGEOUT))
1398                 err("madvise(MADV_PAGEOUT) failed");
1399
1400         /* Uffd-wp should persist even swapped out */
1401         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1402         pagemap_check_wp(value, true);
1403         /* Make sure uffd-wp bit dropped when fork */
1404         if (pagemap_test_fork(false))
1405                 err("Detected stall uffd-wp bit in child");
1406
1407         /* Unprotect; this tests swap pte modifications */
1408         wp_range(uffd, (uint64_t)area_dst, page_size, false);
1409         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1410         pagemap_check_wp(value, false);
1411
1412         /* Fault in the page from disk */
1413         *area_dst = 2;
1414         value = pagemap_read_vaddr(pagemap_fd, area_dst);
1415         pagemap_check_wp(value, false);
1416
1417         close(pagemap_fd);
1418         printf("done\n");
1419 }
1420
1421 static int userfaultfd_stress(void)
1422 {
1423         void *area;
1424         char *tmp_area;
1425         unsigned long nr;
1426         struct uffdio_register uffdio_register;
1427         struct uffd_stats uffd_stats[nr_cpus];
1428
1429         uffd_test_ctx_init(0);
1430
1431         if (posix_memalign(&area, page_size, page_size))
1432                 err("out of memory");
1433         zeropage = area;
1434         bzero(zeropage, page_size);
1435
1436         pthread_mutex_lock(&uffd_read_mutex);
1437
1438         pthread_attr_init(&attr);
1439         pthread_attr_setstacksize(&attr, 16*1024*1024);
1440
1441         while (bounces--) {
1442                 unsigned long expected_ioctls;
1443
1444                 printf("bounces: %d, mode:", bounces);
1445                 if (bounces & BOUNCE_RANDOM)
1446                         printf(" rnd");
1447                 if (bounces & BOUNCE_RACINGFAULTS)
1448                         printf(" racing");
1449                 if (bounces & BOUNCE_VERIFY)
1450                         printf(" ver");
1451                 if (bounces & BOUNCE_POLL)
1452                         printf(" poll");
1453                 else
1454                         printf(" read");
1455                 printf(", ");
1456                 fflush(stdout);
1457
1458                 if (bounces & BOUNCE_POLL)
1459                         fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
1460                 else
1461                         fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
1462
1463                 /* register */
1464                 uffdio_register.range.start = (unsigned long) area_dst;
1465                 uffdio_register.range.len = nr_pages * page_size;
1466                 uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
1467                 if (test_uffdio_wp)
1468                         uffdio_register.mode |= UFFDIO_REGISTER_MODE_WP;
1469                 if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1470                         err("register failure");
1471                 expected_ioctls = uffd_test_ops->expected_ioctls;
1472                 if ((uffdio_register.ioctls & expected_ioctls) !=
1473                     expected_ioctls)
1474                         err("unexpected missing ioctl for anon memory");
1475
1476                 if (area_dst_alias) {
1477                         uffdio_register.range.start = (unsigned long)
1478                                 area_dst_alias;
1479                         if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register))
1480                                 err("register failure alias");
1481                 }
1482
1483                 /*
1484                  * The madvise done previously isn't enough: some
1485                  * uffd_thread could have read userfaults (one of
1486                  * those already resolved by the background thread)
1487                  * and it may be in the process of calling
1488                  * UFFDIO_COPY. UFFDIO_COPY will read the zapped
1489                  * area_src and it would map a zero page in it (of
1490                  * course such a UFFDIO_COPY is perfectly safe as it'd
1491                  * return -EEXIST). The problem comes at the next
1492                  * bounce though: that racing UFFDIO_COPY would
1493                  * generate zeropages in the area_src, so invalidating
1494                  * the previous MADV_DONTNEED. Without this additional
1495                  * MADV_DONTNEED those zeropages leftovers in the
1496                  * area_src would lead to -EEXIST failure during the
1497                  * next bounce, effectively leaving a zeropage in the
1498                  * area_dst.
1499                  *
1500                  * Try to comment this out madvise to see the memory
1501                  * corruption being caught pretty quick.
1502                  *
1503                  * khugepaged is also inhibited to collapse THP after
1504                  * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
1505                  * required to MADV_DONTNEED here.
1506                  */
1507                 uffd_test_ops->release_pages(area_dst);
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                         err("unregister failure");
1523                 if (area_dst_alias) {
1524                         uffdio_register.range.start = (unsigned long) area_dst;
1525                         if (ioctl(uffd, UFFDIO_UNREGISTER,
1526                                   &uffdio_register.range))
1527                                 err("unregister failure alias");
1528                 }
1529
1530                 /* verification */
1531                 if (bounces & BOUNCE_VERIFY)
1532                         for (nr = 0; nr < nr_pages; nr++)
1533                                 if (*area_count(area_dst, nr) != count_verify[nr])
1534                                         err("error area_count %llu %llu %lu\n",
1535                                             *area_count(area_src, nr),
1536                                             count_verify[nr], nr);
1537
1538                 /* prepare next bounce */
1539                 tmp_area = area_src;
1540                 area_src = area_dst;
1541                 area_dst = tmp_area;
1542
1543                 tmp_area = area_src_alias;
1544                 area_src_alias = area_dst_alias;
1545                 area_dst_alias = tmp_area;
1546
1547                 uffd_stats_report(uffd_stats, nr_cpus);
1548         }
1549
1550         if (test_type == TEST_ANON) {
1551                 /*
1552                  * shmem/hugetlb won't be able to run since they have different
1553                  * behavior on fork() (file-backed memory normally drops ptes
1554                  * directly when fork), meanwhile the pagemap test will verify
1555                  * pgtable entry of fork()ed child.
1556                  */
1557                 userfaultfd_pagemap_test(page_size);
1558                 /*
1559                  * Hard-code for x86_64 for now for 2M THP, as x86_64 is
1560                  * currently the only one that supports uffd-wp
1561                  */
1562                 userfaultfd_pagemap_test(page_size * 512);
1563         }
1564
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                 test_uffdio_minor = true;
1614         } else {
1615                 err("Unknown test type: %s", type);
1616         }
1617
1618         if (test_type == TEST_HUGETLB)
1619                 page_size = default_huge_page_size();
1620         else
1621                 page_size = sysconf(_SC_PAGE_SIZE);
1622
1623         if (!page_size)
1624                 err("Unable to determine page size");
1625         if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
1626             > page_size)
1627                 err("Impossible to run this test");
1628 }
1629
1630 static void sigalrm(int sig)
1631 {
1632         if (sig != SIGALRM)
1633                 abort();
1634         test_uffdio_copy_eexist = true;
1635         test_uffdio_zeropage_eexist = true;
1636         alarm(ALARM_INTERVAL_SECS);
1637 }
1638
1639 int main(int argc, char **argv)
1640 {
1641         if (argc < 4)
1642                 usage();
1643
1644         if (signal(SIGALRM, sigalrm) == SIG_ERR)
1645                 err("failed to arm SIGALRM");
1646         alarm(ALARM_INTERVAL_SECS);
1647
1648         set_test_type(argv[1]);
1649
1650         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1651         nr_pages_per_cpu = atol(argv[2]) * 1024*1024 / page_size /
1652                 nr_cpus;
1653         if (!nr_pages_per_cpu) {
1654                 _err("invalid MiB");
1655                 usage();
1656         }
1657
1658         bounces = atoi(argv[3]);
1659         if (bounces <= 0) {
1660                 _err("invalid bounces");
1661                 usage();
1662         }
1663         nr_pages = nr_pages_per_cpu * nr_cpus;
1664
1665         if (test_type == TEST_HUGETLB) {
1666                 if (argc < 5)
1667                         usage();
1668                 huge_fd = open(argv[4], O_CREAT | O_RDWR, 0755);
1669                 if (huge_fd < 0)
1670                         err("Open of %s failed", argv[4]);
1671                 if (ftruncate(huge_fd, 0))
1672                         err("ftruncate %s to size 0 failed", argv[4]);
1673         } else if (test_type == TEST_SHMEM) {
1674                 shm_fd = memfd_create(argv[0], 0);
1675                 if (shm_fd < 0)
1676                         err("memfd_create");
1677                 if (ftruncate(shm_fd, nr_pages * page_size * 2))
1678                         err("ftruncate");
1679                 if (fallocate(shm_fd,
1680                               FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 0,
1681                               nr_pages * page_size * 2))
1682                         err("fallocate");
1683         }
1684         printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
1685                nr_pages, nr_pages_per_cpu);
1686         return userfaultfd_stress();
1687 }
1688
1689 #else /* __NR_userfaultfd */
1690
1691 #warning "missing __NR_userfaultfd definition"
1692
1693 int main(void)
1694 {
1695         printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
1696         return KSFT_SKIP;
1697 }
1698
1699 #endif /* __NR_userfaultfd */