1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * KSM functional tests
4 *
5 * Copyright 2022, Red Hat, Inc.
6 *
7 * Author(s): David Hildenbrand <david@redhat.com>
8 */
9 #define _GNU_SOURCE
10 #include <stdlib.h>
11 #include <string.h>
12 #include <stdbool.h>
13 #include <stdint.h>
14 #include <unistd.h>
15 #include <errno.h>
16 #include <fcntl.h>
17 #include <sys/mman.h>
18 #include <sys/prctl.h>
19 #include <sys/syscall.h>
20 #include <sys/ioctl.h>
21 #include <sys/wait.h>
22 #include <linux/userfaultfd.h>
23
24 #include "../kselftest.h"
25 #include "vm_util.h"
26
27 #define KiB 1024u
28 #define MiB (1024 * KiB)
29
30 static int mem_fd;
31 static int ksm_fd;
32 static int ksm_full_scans_fd;
33 static int proc_self_ksm_stat_fd;
34 static int proc_self_ksm_merging_pages_fd;
35 static int ksm_use_zero_pages_fd;
36 static int pagemap_fd;
37 static size_t pagesize;
38
range_maps_duplicates(char * addr,unsigned long size)39 static bool range_maps_duplicates(char *addr, unsigned long size)
40 {
41 unsigned long offs_a, offs_b, pfn_a, pfn_b;
42
43 /*
44 * There is no easy way to check if there are KSM pages mapped into
45 * this range. We only check that the range does not map the same PFN
46 * twice by comparing each pair of mapped pages.
47 */
48 for (offs_a = 0; offs_a < size; offs_a += pagesize) {
49 pfn_a = pagemap_get_pfn(pagemap_fd, addr + offs_a);
50 /* Page not present or PFN not exposed by the kernel. */
51 if (pfn_a == -1ul || !pfn_a)
52 continue;
53
54 for (offs_b = offs_a + pagesize; offs_b < size;
55 offs_b += pagesize) {
56 pfn_b = pagemap_get_pfn(pagemap_fd, addr + offs_b);
57 if (pfn_b == -1ul || !pfn_b)
58 continue;
59 if (pfn_a == pfn_b)
60 return true;
61 }
62 }
63 return false;
64 }
65
get_my_ksm_zero_pages(void)66 static long get_my_ksm_zero_pages(void)
67 {
68 char buf[200];
69 char *substr_ksm_zero;
70 size_t value_pos;
71 ssize_t read_size;
72 unsigned long my_ksm_zero_pages;
73
74 if (!proc_self_ksm_stat_fd)
75 return 0;
76
77 read_size = pread(proc_self_ksm_stat_fd, buf, sizeof(buf) - 1, 0);
78 if (read_size < 0)
79 return -errno;
80
81 buf[read_size] = 0;
82
83 substr_ksm_zero = strstr(buf, "ksm_zero_pages");
84 if (!substr_ksm_zero)
85 return 0;
86
87 value_pos = strcspn(substr_ksm_zero, "0123456789");
88 my_ksm_zero_pages = strtol(substr_ksm_zero + value_pos, NULL, 10);
89
90 return my_ksm_zero_pages;
91 }
92
get_my_merging_pages(void)93 static long get_my_merging_pages(void)
94 {
95 char buf[10];
96 ssize_t ret;
97
98 if (proc_self_ksm_merging_pages_fd < 0)
99 return proc_self_ksm_merging_pages_fd;
100
101 ret = pread(proc_self_ksm_merging_pages_fd, buf, sizeof(buf) - 1, 0);
102 if (ret <= 0)
103 return -errno;
104 buf[ret] = 0;
105
106 return strtol(buf, NULL, 10);
107 }
108
ksm_get_full_scans(void)109 static long ksm_get_full_scans(void)
110 {
111 char buf[10];
112 ssize_t ret;
113
114 ret = pread(ksm_full_scans_fd, buf, sizeof(buf) - 1, 0);
115 if (ret <= 0)
116 return -errno;
117 buf[ret] = 0;
118
119 return strtol(buf, NULL, 10);
120 }
121
ksm_merge(void)122 static int ksm_merge(void)
123 {
124 long start_scans, end_scans;
125
126 /* Wait for two full scans such that any possible merging happened. */
127 start_scans = ksm_get_full_scans();
128 if (start_scans < 0)
129 return start_scans;
130 if (write(ksm_fd, "1", 1) != 1)
131 return -errno;
132 do {
133 end_scans = ksm_get_full_scans();
134 if (end_scans < 0)
135 return end_scans;
136 } while (end_scans < start_scans + 2);
137
138 return 0;
139 }
140
ksm_unmerge(void)141 static int ksm_unmerge(void)
142 {
143 if (write(ksm_fd, "2", 1) != 1)
144 return -errno;
145 return 0;
146 }
147
mmap_and_merge_range(char val,unsigned long size,int prot,bool use_prctl)148 static char *mmap_and_merge_range(char val, unsigned long size, int prot,
149 bool use_prctl)
150 {
151 char *map;
152 int ret;
153
154 /* Stabilize accounting by disabling KSM completely. */
155 if (ksm_unmerge()) {
156 ksft_test_result_fail("Disabling (unmerging) KSM failed\n");
157 goto unmap;
158 }
159
160 if (get_my_merging_pages() > 0) {
161 ksft_test_result_fail("Still pages merged\n");
162 goto unmap;
163 }
164
165 map = mmap(NULL, size, PROT_READ|PROT_WRITE,
166 MAP_PRIVATE|MAP_ANON, -1, 0);
167 if (map == MAP_FAILED) {
168 ksft_test_result_fail("mmap() failed\n");
169 return MAP_FAILED;
170 }
171
172 /* Don't use THP. Ignore if THP are not around on a kernel. */
173 if (madvise(map, size, MADV_NOHUGEPAGE) && errno != EINVAL) {
174 ksft_test_result_fail("MADV_NOHUGEPAGE failed\n");
175 goto unmap;
176 }
177
178 /* Make sure each page contains the same values to merge them. */
179 memset(map, val, size);
180
181 if (mprotect(map, size, prot)) {
182 ksft_test_result_skip("mprotect() failed\n");
183 goto unmap;
184 }
185
186 if (use_prctl) {
187 ret = prctl(PR_SET_MEMORY_MERGE, 1, 0, 0, 0);
188 if (ret < 0 && errno == EINVAL) {
189 ksft_test_result_skip("PR_SET_MEMORY_MERGE not supported\n");
190 goto unmap;
191 } else if (ret) {
192 ksft_test_result_fail("PR_SET_MEMORY_MERGE=1 failed\n");
193 goto unmap;
194 }
195 } else if (madvise(map, size, MADV_MERGEABLE)) {
196 ksft_test_result_fail("MADV_MERGEABLE failed\n");
197 goto unmap;
198 }
199
200 /* Run KSM to trigger merging and wait. */
201 if (ksm_merge()) {
202 ksft_test_result_fail("Running KSM failed\n");
203 goto unmap;
204 }
205
206 /*
207 * Check if anything was merged at all. Ignore the zero page that is
208 * accounted differently (depending on kernel support).
209 */
210 if (val && !get_my_merging_pages()) {
211 ksft_test_result_fail("No pages got merged\n");
212 goto unmap;
213 }
214
215 return map;
216 unmap:
217 munmap(map, size);
218 return MAP_FAILED;
219 }
220
test_unmerge(void)221 static void test_unmerge(void)
222 {
223 const unsigned int size = 2 * MiB;
224 char *map;
225
226 ksft_print_msg("[RUN] %s\n", __func__);
227
228 map = mmap_and_merge_range(0xcf, size, PROT_READ | PROT_WRITE, false);
229 if (map == MAP_FAILED)
230 return;
231
232 if (madvise(map, size, MADV_UNMERGEABLE)) {
233 ksft_test_result_fail("MADV_UNMERGEABLE failed\n");
234 goto unmap;
235 }
236
237 ksft_test_result(!range_maps_duplicates(map, size),
238 "Pages were unmerged\n");
239 unmap:
240 munmap(map, size);
241 }
242
test_unmerge_zero_pages(void)243 static void test_unmerge_zero_pages(void)
244 {
245 const unsigned int size = 2 * MiB;
246 char *map;
247 unsigned int offs;
248 unsigned long pages_expected;
249
250 ksft_print_msg("[RUN] %s\n", __func__);
251
252 if (proc_self_ksm_stat_fd < 0) {
253 ksft_test_result_skip("open(\"/proc/self/ksm_stat\") failed\n");
254 return;
255 }
256 if (ksm_use_zero_pages_fd < 0) {
257 ksft_test_result_skip("open \"/sys/kernel/mm/ksm/use_zero_pages\" failed\n");
258 return;
259 }
260 if (write(ksm_use_zero_pages_fd, "1", 1) != 1) {
261 ksft_test_result_skip("write \"/sys/kernel/mm/ksm/use_zero_pages\" failed\n");
262 return;
263 }
264
265 /* Let KSM deduplicate zero pages. */
266 map = mmap_and_merge_range(0x00, size, PROT_READ | PROT_WRITE, false);
267 if (map == MAP_FAILED)
268 return;
269
270 /* Check if ksm_zero_pages is updated correctly after KSM merging */
271 pages_expected = size / pagesize;
272 if (pages_expected != get_my_ksm_zero_pages()) {
273 ksft_test_result_fail("'ksm_zero_pages' updated after merging\n");
274 goto unmap;
275 }
276
277 /* Try to unmerge half of the region */
278 if (madvise(map, size / 2, MADV_UNMERGEABLE)) {
279 ksft_test_result_fail("MADV_UNMERGEABLE failed\n");
280 goto unmap;
281 }
282
283 /* Check if ksm_zero_pages is updated correctly after unmerging */
284 pages_expected /= 2;
285 if (pages_expected != get_my_ksm_zero_pages()) {
286 ksft_test_result_fail("'ksm_zero_pages' updated after unmerging\n");
287 goto unmap;
288 }
289
290 /* Trigger unmerging of the other half by writing to the pages. */
291 for (offs = size / 2; offs < size; offs += pagesize)
292 *((unsigned int *)&map[offs]) = offs;
293
294 /* Now we should have no zeropages remaining. */
295 if (get_my_ksm_zero_pages()) {
296 ksft_test_result_fail("'ksm_zero_pages' updated after write fault\n");
297 goto unmap;
298 }
299
300 /* Check if ksm zero pages are really unmerged */
301 ksft_test_result(!range_maps_duplicates(map, size),
302 "KSM zero pages were unmerged\n");
303 unmap:
304 munmap(map, size);
305 }
306
test_unmerge_discarded(void)307 static void test_unmerge_discarded(void)
308 {
309 const unsigned int size = 2 * MiB;
310 char *map;
311
312 ksft_print_msg("[RUN] %s\n", __func__);
313
314 map = mmap_and_merge_range(0xcf, size, PROT_READ | PROT_WRITE, false);
315 if (map == MAP_FAILED)
316 return;
317
318 /* Discard half of all mapped pages so we have pte_none() entries. */
319 if (madvise(map, size / 2, MADV_DONTNEED)) {
320 ksft_test_result_fail("MADV_DONTNEED failed\n");
321 goto unmap;
322 }
323
324 if (madvise(map, size, MADV_UNMERGEABLE)) {
325 ksft_test_result_fail("MADV_UNMERGEABLE failed\n");
326 goto unmap;
327 }
328
329 ksft_test_result(!range_maps_duplicates(map, size),
330 "Pages were unmerged\n");
331 unmap:
332 munmap(map, size);
333 }
334
335 #ifdef __NR_userfaultfd
test_unmerge_uffd_wp(void)336 static void test_unmerge_uffd_wp(void)
337 {
338 struct uffdio_writeprotect uffd_writeprotect;
339 const unsigned int size = 2 * MiB;
340 struct uffdio_api uffdio_api;
341 char *map;
342 int uffd;
343
344 ksft_print_msg("[RUN] %s\n", __func__);
345
346 map = mmap_and_merge_range(0xcf, size, PROT_READ | PROT_WRITE, false);
347 if (map == MAP_FAILED)
348 return;
349
350 /* See if UFFD is around. */
351 uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
352 if (uffd < 0) {
353 ksft_test_result_skip("__NR_userfaultfd failed\n");
354 goto unmap;
355 }
356
357 /* See if UFFD-WP is around. */
358 uffdio_api.api = UFFD_API;
359 uffdio_api.features = UFFD_FEATURE_PAGEFAULT_FLAG_WP;
360 if (ioctl(uffd, UFFDIO_API, &uffdio_api) < 0) {
361 ksft_test_result_fail("UFFDIO_API failed\n");
362 goto close_uffd;
363 }
364 if (!(uffdio_api.features & UFFD_FEATURE_PAGEFAULT_FLAG_WP)) {
365 ksft_test_result_skip("UFFD_FEATURE_PAGEFAULT_FLAG_WP not available\n");
366 goto close_uffd;
367 }
368
369 /* Register UFFD-WP, no need for an actual handler. */
370 if (uffd_register(uffd, map, size, false, true, false)) {
371 ksft_test_result_fail("UFFDIO_REGISTER_MODE_WP failed\n");
372 goto close_uffd;
373 }
374
375 /* Write-protect the range using UFFD-WP. */
376 uffd_writeprotect.range.start = (unsigned long) map;
377 uffd_writeprotect.range.len = size;
378 uffd_writeprotect.mode = UFFDIO_WRITEPROTECT_MODE_WP;
379 if (ioctl(uffd, UFFDIO_WRITEPROTECT, &uffd_writeprotect)) {
380 ksft_test_result_fail("UFFDIO_WRITEPROTECT failed\n");
381 goto close_uffd;
382 }
383
384 if (madvise(map, size, MADV_UNMERGEABLE)) {
385 ksft_test_result_fail("MADV_UNMERGEABLE failed\n");
386 goto close_uffd;
387 }
388
389 ksft_test_result(!range_maps_duplicates(map, size),
390 "Pages were unmerged\n");
391 close_uffd:
392 close(uffd);
393 unmap:
394 munmap(map, size);
395 }
396 #endif
397
398 /* Verify that KSM can be enabled / queried with prctl. */
test_prctl(void)399 static void test_prctl(void)
400 {
401 int ret;
402
403 ksft_print_msg("[RUN] %s\n", __func__);
404
405 ret = prctl(PR_SET_MEMORY_MERGE, 1, 0, 0, 0);
406 if (ret < 0 && errno == EINVAL) {
407 ksft_test_result_skip("PR_SET_MEMORY_MERGE not supported\n");
408 return;
409 } else if (ret) {
410 ksft_test_result_fail("PR_SET_MEMORY_MERGE=1 failed\n");
411 return;
412 }
413
414 ret = prctl(PR_GET_MEMORY_MERGE, 0, 0, 0, 0);
415 if (ret < 0) {
416 ksft_test_result_fail("PR_GET_MEMORY_MERGE failed\n");
417 return;
418 } else if (ret != 1) {
419 ksft_test_result_fail("PR_SET_MEMORY_MERGE=1 not effective\n");
420 return;
421 }
422
423 ret = prctl(PR_SET_MEMORY_MERGE, 0, 0, 0, 0);
424 if (ret) {
425 ksft_test_result_fail("PR_SET_MEMORY_MERGE=0 failed\n");
426 return;
427 }
428
429 ret = prctl(PR_GET_MEMORY_MERGE, 0, 0, 0, 0);
430 if (ret < 0) {
431 ksft_test_result_fail("PR_GET_MEMORY_MERGE failed\n");
432 return;
433 } else if (ret != 0) {
434 ksft_test_result_fail("PR_SET_MEMORY_MERGE=0 not effective\n");
435 return;
436 }
437
438 ksft_test_result_pass("Setting/clearing PR_SET_MEMORY_MERGE works\n");
439 }
440
441 /* Verify that prctl ksm flag is inherited. */
test_prctl_fork(void)442 static void test_prctl_fork(void)
443 {
444 int ret, status;
445 pid_t child_pid;
446
447 ksft_print_msg("[RUN] %s\n", __func__);
448
449 ret = prctl(PR_SET_MEMORY_MERGE, 1, 0, 0, 0);
450 if (ret < 0 && errno == EINVAL) {
451 ksft_test_result_skip("PR_SET_MEMORY_MERGE not supported\n");
452 return;
453 } else if (ret) {
454 ksft_test_result_fail("PR_SET_MEMORY_MERGE=1 failed\n");
455 return;
456 }
457
458 child_pid = fork();
459 if (!child_pid) {
460 exit(prctl(PR_GET_MEMORY_MERGE, 0, 0, 0, 0));
461 } else if (child_pid < 0) {
462 ksft_test_result_fail("fork() failed\n");
463 return;
464 }
465
466 if (waitpid(child_pid, &status, 0) < 0) {
467 ksft_test_result_fail("waitpid() failed\n");
468 return;
469 } else if (WEXITSTATUS(status) != 1) {
470 ksft_test_result_fail("unexpected PR_GET_MEMORY_MERGE result in child\n");
471 return;
472 }
473
474 if (prctl(PR_SET_MEMORY_MERGE, 0, 0, 0, 0)) {
475 ksft_test_result_fail("PR_SET_MEMORY_MERGE=0 failed\n");
476 return;
477 }
478
479 ksft_test_result_pass("PR_SET_MEMORY_MERGE value is inherited\n");
480 }
481
test_prctl_unmerge(void)482 static void test_prctl_unmerge(void)
483 {
484 const unsigned int size = 2 * MiB;
485 char *map;
486
487 ksft_print_msg("[RUN] %s\n", __func__);
488
489 map = mmap_and_merge_range(0xcf, size, PROT_READ | PROT_WRITE, true);
490 if (map == MAP_FAILED)
491 return;
492
493 if (prctl(PR_SET_MEMORY_MERGE, 0, 0, 0, 0)) {
494 ksft_test_result_fail("PR_SET_MEMORY_MERGE=0 failed\n");
495 goto unmap;
496 }
497
498 ksft_test_result(!range_maps_duplicates(map, size),
499 "Pages were unmerged\n");
500 unmap:
501 munmap(map, size);
502 }
503
test_prot_none(void)504 static void test_prot_none(void)
505 {
506 const unsigned int size = 2 * MiB;
507 char *map;
508 int i;
509
510 ksft_print_msg("[RUN] %s\n", __func__);
511
512 map = mmap_and_merge_range(0x11, size, PROT_NONE, false);
513 if (map == MAP_FAILED)
514 goto unmap;
515
516 /* Store a unique value in each page on one half using ptrace */
517 for (i = 0; i < size / 2; i += pagesize) {
518 lseek(mem_fd, (uintptr_t) map + i, SEEK_SET);
519 if (write(mem_fd, &i, sizeof(i)) != sizeof(i)) {
520 ksft_test_result_fail("ptrace write failed\n");
521 goto unmap;
522 }
523 }
524
525 /* Trigger unsharing on the other half. */
526 if (madvise(map + size / 2, size / 2, MADV_UNMERGEABLE)) {
527 ksft_test_result_fail("MADV_UNMERGEABLE failed\n");
528 goto unmap;
529 }
530
531 ksft_test_result(!range_maps_duplicates(map, size),
532 "Pages were unmerged\n");
533 unmap:
534 munmap(map, size);
535 }
536
main(int argc,char ** argv)537 int main(int argc, char **argv)
538 {
539 unsigned int tests = 7;
540 int err;
541
542 #ifdef __NR_userfaultfd
543 tests++;
544 #endif
545
546 ksft_print_header();
547 ksft_set_plan(tests);
548
549 pagesize = getpagesize();
550
551 mem_fd = open("/proc/self/mem", O_RDWR);
552 if (mem_fd < 0)
553 ksft_exit_fail_msg("opening /proc/self/mem failed\n");
554 ksm_fd = open("/sys/kernel/mm/ksm/run", O_RDWR);
555 if (ksm_fd < 0)
556 ksft_exit_skip("open(\"/sys/kernel/mm/ksm/run\") failed\n");
557 ksm_full_scans_fd = open("/sys/kernel/mm/ksm/full_scans", O_RDONLY);
558 if (ksm_full_scans_fd < 0)
559 ksft_exit_skip("open(\"/sys/kernel/mm/ksm/full_scans\") failed\n");
560 pagemap_fd = open("/proc/self/pagemap", O_RDONLY);
561 if (pagemap_fd < 0)
562 ksft_exit_skip("open(\"/proc/self/pagemap\") failed\n");
563 proc_self_ksm_stat_fd = open("/proc/self/ksm_stat", O_RDONLY);
564 proc_self_ksm_merging_pages_fd = open("/proc/self/ksm_merging_pages",
565 O_RDONLY);
566 ksm_use_zero_pages_fd = open("/sys/kernel/mm/ksm/use_zero_pages", O_RDWR);
567
568 test_unmerge();
569 test_unmerge_zero_pages();
570 test_unmerge_discarded();
571 #ifdef __NR_userfaultfd
572 test_unmerge_uffd_wp();
573 #endif
574
575 test_prot_none();
576
577 test_prctl();
578 test_prctl_fork();
579 test_prctl_unmerge();
580
581 err = ksft_get_fail_cnt();
582 if (err)
583 ksft_exit_fail_msg("%d out of %d tests failed\n",
584 err, ksft_test_num());
585 return ksft_exit_pass();
586 }
587