tools/testing/selftests/mm/mremap_test.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2020 Google LLC
   4  */
   5 #define _GNU_SOURCE
   6
   7 #include <errno.h>
   8 #include <stdlib.h>
   9 #include <stdio.h>
  10 #include <string.h>
  11 #include <sys/mman.h>
  12 #include <time.h>
  13 #include <stdbool.h>
  14
  15 #include "../kselftest.h"
  16
  17 #define EXPECT_SUCCESS 0
  18 #define EXPECT_FAILURE 1
  19 #define NON_OVERLAPPING 0
  20 #define OVERLAPPING 1
  21 #define NS_PER_SEC 1000000000ULL
  22 #define VALIDATION_DEFAULT_THRESHOLD 4  /* 4MB */
  23 #define VALIDATION_NO_THRESHOLD 0       /* Verify the entire region */
  24
  25 #define MIN(X, Y) ((X) < (Y) ? (X) : (Y))
  26
  27 struct config {
  28         unsigned long long src_alignment;
  29         unsigned long long dest_alignment;
  30         unsigned long long region_size;
  31         int overlapping;
  32 };
  33
  34 struct test {
  35         const char *name;
  36         struct config config;
  37         int expect_failure;
  38 };
  39
  40 enum {
  41         _1KB = 1ULL << 10,      /* 1KB -> not page aligned */
  42         _4KB = 4ULL << 10,
  43         _8KB = 8ULL << 10,
  44         _1MB = 1ULL << 20,
  45         _2MB = 2ULL << 20,
  46         _4MB = 4ULL << 20,
  47         _1GB = 1ULL << 30,
  48         _2GB = 2ULL << 30,
  49         PMD = _2MB,
  50         PUD = _1GB,
  51 };
  52
  53 #define PTE page_size
  54
  55 #define MAKE_TEST(source_align, destination_align, size,        \
  56                   overlaps, should_fail, test_name)             \
  57 (struct test){                                                  \
  58         .name = test_name,                                      \
  59         .config = {                                             \
  60                 .src_alignment = source_align,                  \
  61                 .dest_alignment = destination_align,            \
  62                 .region_size = size,                            \
  63                 .overlapping = overlaps,                        \
  64         },                                                      \
  65         .expect_failure = should_fail                           \
  66 }
  67
  68 /*
  69  * Returns false if the requested remap region overlaps with an
  70  * existing mapping (e.g text, stack) else returns true.
  71  */
  72 static bool is_remap_region_valid(void *addr, unsigned long long size)
  73 {
  74         void *remap_addr = NULL;
  75         bool ret = true;
  76
  77         /* Use MAP_FIXED_NOREPLACE flag to ensure region is not mapped */
  78         remap_addr = mmap(addr, size, PROT_READ | PROT_WRITE,
  79                                          MAP_FIXED_NOREPLACE | MAP_ANONYMOUS | MAP_SHARED,
  80                                          -1, 0);
  81
  82         if (remap_addr == MAP_FAILED) {
  83                 if (errno == EEXIST)
  84                         ret = false;
  85         } else {
  86                 munmap(remap_addr, size);
  87         }
  88
  89         return ret;
  90 }
  91
  92 /* Returns mmap_min_addr sysctl tunable from procfs */
  93 static unsigned long long get_mmap_min_addr(void)
  94 {
  95         FILE *fp;
  96         int n_matched;
  97         static unsigned long long addr;
  98
  99         if (addr)
 100                 return addr;
 101
 102         fp = fopen("/proc/sys/vm/mmap_min_addr", "r");
 103         if (fp == NULL) {
 104                 ksft_print_msg("Failed to open /proc/sys/vm/mmap_min_addr: %s\n",
 105                         strerror(errno));
 106                 exit(KSFT_SKIP);
 107         }
 108
 109         n_matched = fscanf(fp, "%llu", &addr);
 110         if (n_matched != 1) {
 111                 ksft_print_msg("Failed to read /proc/sys/vm/mmap_min_addr: %s\n",
 112                         strerror(errno));
 113                 fclose(fp);
 114                 exit(KSFT_SKIP);
 115         }
 116
 117         fclose(fp);
 118         return addr;
 119 }
 120
 121 /*
 122  * This test validates that merge is called when expanding a mapping.
 123  * Mapping containing three pages is created, middle page is unmapped
 124  * and then the mapping containing the first page is expanded so that
 125  * it fills the created hole. The two parts should merge creating
 126  * single mapping with three pages.
 127  */
 128 static void mremap_expand_merge(unsigned long page_size)
 129 {
 130         char *test_name = "mremap expand merge";
 131         FILE *fp;
 132         char *line = NULL;
 133         size_t len = 0;
 134         bool success = false;
 135         char *start = mmap(NULL, 3 * page_size, PROT_READ | PROT_WRITE,
 136                            MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 137
 138         munmap(start + page_size, page_size);
 139         mremap(start, page_size, 2 * page_size, 0);
 140
 141         fp = fopen("/proc/self/maps", "r");
 142         if (fp == NULL) {
 143                 ksft_test_result_fail("%s\n", test_name);
 144                 return;
 145         }
 146
 147         while (getline(&line, &len, fp) != -1) {
 148                 char *first = strtok(line, "- ");
 149                 void *first_val = (void *)strtol(first, NULL, 16);
 150                 char *second = strtok(NULL, "- ");
 151                 void *second_val = (void *) strtol(second, NULL, 16);
 152
 153                 if (first_val == start && second_val == start + 3 * page_size) {
 154                         success = true;
 155                         break;
 156                 }
 157         }
 158         if (success)
 159                 ksft_test_result_pass("%s\n", test_name);
 160         else
 161                 ksft_test_result_fail("%s\n", test_name);
 162         fclose(fp);
 163 }
 164
 165 /*
 166  * Returns the start address of the mapping on success, else returns
 167  * NULL on failure.
 168  */
 169 static void *get_source_mapping(struct config c)
 170 {
 171         unsigned long long addr = 0ULL;
 172         void *src_addr = NULL;
 173         unsigned long long mmap_min_addr;
 174
 175         mmap_min_addr = get_mmap_min_addr();
 176
 177 retry:
 178         addr += c.src_alignment;
 179         if (addr < mmap_min_addr)
 180                 goto retry;
 181
 182         src_addr = mmap((void *) addr, c.region_size, PROT_READ | PROT_WRITE,
 183                                         MAP_FIXED_NOREPLACE | MAP_ANONYMOUS | MAP_SHARED,
 184                                         -1, 0);
 185         if (src_addr == MAP_FAILED) {
 186                 if (errno == EPERM || errno == EEXIST)
 187                         goto retry;
 188                 goto error;
 189         }
 190         /*
 191          * Check that the address is aligned to the specified alignment.
 192          * Addresses which have alignments that are multiples of that
 193          * specified are not considered valid. For instance, 1GB address is
 194          * 2MB-aligned, however it will not be considered valid for a
 195          * requested alignment of 2MB. This is done to reduce coincidental
 196          * alignment in the tests.
 197          */
 198         if (((unsigned long long) src_addr & (c.src_alignment - 1)) ||
 199                         !((unsigned long long) src_addr & c.src_alignment)) {
 200                 munmap(src_addr, c.region_size);
 201                 goto retry;
 202         }
 203
 204         if (!src_addr)
 205                 goto error;
 206
 207         return src_addr;
 208 error:
 209         ksft_print_msg("Failed to map source region: %s\n",
 210                         strerror(errno));
 211         return NULL;
 212 }
 213
 214 /* Returns the time taken for the remap on success else returns -1. */
 215 static long long remap_region(struct config c, unsigned int threshold_mb,
 216                               char pattern_seed)
 217 {
 218         void *addr, *src_addr, *dest_addr;
 219         unsigned long long i;
 220         struct timespec t_start = {0, 0}, t_end = {0, 0};
 221         long long  start_ns, end_ns, align_mask, ret, offset;
 222         unsigned long long threshold;
 223
 224         if (threshold_mb == VALIDATION_NO_THRESHOLD)
 225                 threshold = c.region_size;
 226         else
 227                 threshold = MIN(threshold_mb * _1MB, c.region_size);
 228
 229         src_addr = get_source_mapping(c);
 230         if (!src_addr) {
 231                 ret = -1;
 232                 goto out;
 233         }
 234
 235         /* Set byte pattern */
 236         srand(pattern_seed);
 237         for (i = 0; i < threshold; i++)
 238                 memset((char *) src_addr + i, (char) rand(), 1);
 239
 240         /* Mask to zero out lower bits of address for alignment */
 241         align_mask = ~(c.dest_alignment - 1);
 242         /* Offset of destination address from the end of the source region */
 243         offset = (c.overlapping) ? -c.dest_alignment : c.dest_alignment;
 244         addr = (void *) (((unsigned long long) src_addr + c.region_size
 245                           + offset) & align_mask);
 246
 247         /* See comment in get_source_mapping() */
 248         if (!((unsigned long long) addr & c.dest_alignment))
 249                 addr = (void *) ((unsigned long long) addr | c.dest_alignment);
 250
 251         /* Don't destroy existing mappings unless expected to overlap */
 252         while (!is_remap_region_valid(addr, c.region_size) && !c.overlapping) {
 253                 /* Check for unsigned overflow */
 254                 if (addr + c.dest_alignment < addr) {
 255                         ksft_print_msg("Couldn't find a valid region to remap to\n");
 256                         ret = -1;
 257                         goto out;
 258                 }
 259                 addr += c.dest_alignment;
 260         }
 261
 262         clock_gettime(CLOCK_MONOTONIC, &t_start);
 263         dest_addr = mremap(src_addr, c.region_size, c.region_size,
 264                                           MREMAP_MAYMOVE|MREMAP_FIXED, (char *) addr);
 265         clock_gettime(CLOCK_MONOTONIC, &t_end);
 266
 267         if (dest_addr == MAP_FAILED) {
 268                 ksft_print_msg("mremap failed: %s\n", strerror(errno));
 269                 ret = -1;
 270                 goto clean_up_src;
 271         }
 272
 273         /* Verify byte pattern after remapping */
 274         srand(pattern_seed);
 275         for (i = 0; i < threshold; i++) {
 276                 char c = (char) rand();
 277
 278                 if (((char *) dest_addr)[i] != c) {
 279                         ksft_print_msg("Data after remap doesn't match at offset %d\n",
 280                                        i);
 281                         ksft_print_msg("Expected: %#x\t Got: %#x\n", c & 0xff,
 282                                         ((char *) dest_addr)[i] & 0xff);
 283                         ret = -1;
 284                         goto clean_up_dest;
 285                 }
 286         }
 287
 288         start_ns = t_start.tv_sec * NS_PER_SEC + t_start.tv_nsec;
 289         end_ns = t_end.tv_sec * NS_PER_SEC + t_end.tv_nsec;
 290         ret = end_ns - start_ns;
 291
 292 /*
 293  * Since the destination address is specified using MREMAP_FIXED, subsequent
 294  * mremap will unmap any previous mapping at the address range specified by
 295  * dest_addr and region_size. This significantly affects the remap time of
 296  * subsequent tests. So we clean up mappings after each test.
 297  */
 298 clean_up_dest:
 299         munmap(dest_addr, c.region_size);
 300 clean_up_src:
 301         munmap(src_addr, c.region_size);
 302 out:
 303         return ret;
 304 }
 305
 306 static void run_mremap_test_case(struct test test_case, int *failures,
 307                                  unsigned int threshold_mb,
 308                                  unsigned int pattern_seed)
 309 {
 310         long long remap_time = remap_region(test_case.config, threshold_mb,
 311                                             pattern_seed);
 312
 313         if (remap_time < 0) {
 314                 if (test_case.expect_failure)
 315                         ksft_test_result_xfail("%s\n\tExpected mremap failure\n",
 316                                               test_case.name);
 317                 else {
 318                         ksft_test_result_fail("%s\n", test_case.name);
 319                         *failures += 1;
 320                 }
 321         } else {
 322                 /*
 323                  * Comparing mremap time is only applicable if entire region
 324                  * was faulted in.
 325                  */
 326                 if (threshold_mb == VALIDATION_NO_THRESHOLD ||
 327                     test_case.config.region_size <= threshold_mb * _1MB)
 328                         ksft_test_result_pass("%s\n\tmremap time: %12lldns\n",
 329                                               test_case.name, remap_time);
 330                 else
 331                         ksft_test_result_pass("%s\n", test_case.name);
 332         }
 333 }
 334
 335 static void usage(const char *cmd)
 336 {
 337         fprintf(stderr,
 338                 "Usage: %s [[-t <threshold_mb>] [-p <pattern_seed>]]\n"
 339                 "-t\t only validate threshold_mb of the remapped region\n"
 340                 "  \t if 0 is supplied no threshold is used; all tests\n"
 341                 "  \t are run and remapped regions validated fully.\n"
 342                 "  \t The default threshold used is 4MB.\n"
 343                 "-p\t provide a seed to generate the random pattern for\n"
 344                 "  \t validating the remapped region.\n", cmd);
 345 }
 346
 347 static int parse_args(int argc, char **argv, unsigned int *threshold_mb,
 348                       unsigned int *pattern_seed)
 349 {
 350         const char *optstr = "t:p:";
 351         int opt;
 352
 353         while ((opt = getopt(argc, argv, optstr)) != -1) {
 354                 switch (opt) {
 355                 case 't':
 356                         *threshold_mb = atoi(optarg);
 357                         break;
 358                 case 'p':
 359                         *pattern_seed = atoi(optarg);
 360                         break;
 361                 default:
 362                         usage(argv[0]);
 363                         return -1;
 364                 }
 365         }
 366
 367         if (optind < argc) {
 368                 usage(argv[0]);
 369                 return -1;
 370         }
 371
 372         return 0;
 373 }
 374
 375 #define MAX_TEST 13
 376 #define MAX_PERF_TEST 3
 377 int main(int argc, char **argv)
 378 {
 379         int failures = 0;
 380         int i, run_perf_tests;
 381         unsigned int threshold_mb = VALIDATION_DEFAULT_THRESHOLD;
 382         unsigned int pattern_seed;
 383         int num_expand_tests = 1;
 384         struct test test_cases[MAX_TEST];
 385         struct test perf_test_cases[MAX_PERF_TEST];
 386         int page_size;
 387         time_t t;
 388
 389         pattern_seed = (unsigned int) time(&t);
 390
 391         if (parse_args(argc, argv, &threshold_mb, &pattern_seed) < 0)
 392                 exit(EXIT_FAILURE);
 393
 394         ksft_print_msg("Test configs:\n\tthreshold_mb=%u\n\tpattern_seed=%u\n\n",
 395                        threshold_mb, pattern_seed);
 396
 397         page_size = sysconf(_SC_PAGESIZE);
 398
 399         /* Expected mremap failures */
 400         test_cases[0] = MAKE_TEST(page_size, page_size, page_size,
 401                                   OVERLAPPING, EXPECT_FAILURE,
 402                                   "mremap - Source and Destination Regions Overlapping");
 403
 404         test_cases[1] = MAKE_TEST(page_size, page_size/4, page_size,
 405                                   NON_OVERLAPPING, EXPECT_FAILURE,
 406                                   "mremap - Destination Address Misaligned (1KB-aligned)");
 407         test_cases[2] = MAKE_TEST(page_size/4, page_size, page_size,
 408                                   NON_OVERLAPPING, EXPECT_FAILURE,
 409                                   "mremap - Source Address Misaligned (1KB-aligned)");
 410
 411         /* Src addr PTE aligned */
 412         test_cases[3] = MAKE_TEST(PTE, PTE, PTE * 2,
 413                                   NON_OVERLAPPING, EXPECT_SUCCESS,
 414                                   "8KB mremap - Source PTE-aligned, Destination PTE-aligned");
 415
 416         /* Src addr 1MB aligned */
 417         test_cases[4] = MAKE_TEST(_1MB, PTE, _2MB, NON_OVERLAPPING, EXPECT_SUCCESS,
 418                                   "2MB mremap - Source 1MB-aligned, Destination PTE-aligned");
 419         test_cases[5] = MAKE_TEST(_1MB, _1MB, _2MB, NON_OVERLAPPING, EXPECT_SUCCESS,
 420                                   "2MB mremap - Source 1MB-aligned, Destination 1MB-aligned");
 421
 422         /* Src addr PMD aligned */
 423         test_cases[6] = MAKE_TEST(PMD, PTE, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
 424                                   "4MB mremap - Source PMD-aligned, Destination PTE-aligned");
 425         test_cases[7] = MAKE_TEST(PMD, _1MB, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
 426                                   "4MB mremap - Source PMD-aligned, Destination 1MB-aligned");
 427         test_cases[8] = MAKE_TEST(PMD, PMD, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
 428                                   "4MB mremap - Source PMD-aligned, Destination PMD-aligned");
 429
 430         /* Src addr PUD aligned */
 431         test_cases[9] = MAKE_TEST(PUD, PTE, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 432                                   "2GB mremap - Source PUD-aligned, Destination PTE-aligned");
 433         test_cases[10] = MAKE_TEST(PUD, _1MB, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 434                                    "2GB mremap - Source PUD-aligned, Destination 1MB-aligned");
 435         test_cases[11] = MAKE_TEST(PUD, PMD, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 436                                    "2GB mremap - Source PUD-aligned, Destination PMD-aligned");
 437         test_cases[12] = MAKE_TEST(PUD, PUD, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 438                                    "2GB mremap - Source PUD-aligned, Destination PUD-aligned");
 439
 440         perf_test_cases[0] =  MAKE_TEST(page_size, page_size, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 441                                         "1GB mremap - Source PTE-aligned, Destination PTE-aligned");
 442         /*
 443          * mremap 1GB region - Page table level aligned time
 444          * comparison.
 445          */
 446         perf_test_cases[1] = MAKE_TEST(PMD, PMD, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 447                                        "1GB mremap - Source PMD-aligned, Destination PMD-aligned");
 448         perf_test_cases[2] = MAKE_TEST(PUD, PUD, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
 449                                        "1GB mremap - Source PUD-aligned, Destination PUD-aligned");
 450
 451         run_perf_tests =  (threshold_mb == VALIDATION_NO_THRESHOLD) ||
 452                                 (threshold_mb * _1MB >= _1GB);
 453
 454         ksft_set_plan(ARRAY_SIZE(test_cases) + (run_perf_tests ?
 455                       ARRAY_SIZE(perf_test_cases) : 0) + num_expand_tests);
 456
 457         for (i = 0; i < ARRAY_SIZE(test_cases); i++)
 458                 run_mremap_test_case(test_cases[i], &failures, threshold_mb,
 459                                      pattern_seed);
 460
 461         mremap_expand_merge(page_size);
 462
 463         if (run_perf_tests) {
 464                 ksft_print_msg("\n%s\n",
 465                  "mremap HAVE_MOVE_PMD/PUD optimization time comparison for 1GB region:");
 466                 for (i = 0; i < ARRAY_SIZE(perf_test_cases); i++)
 467                         run_mremap_test_case(perf_test_cases[i], &failures,
 468                                              threshold_mb, pattern_seed);
 469         }
 470
 471         if (failures > 0)
 472                 ksft_exit_fail();
 473         else
 474                 ksft_exit_pass();
 475 }