1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * A test of splitting PMD THPs and PTE-mapped THPs from a specified virtual
4 * address range in a process via <debugfs>/split_huge_pages interface.
5 */
6
7 #define _GNU_SOURCE
8 #include <assert.h>
9 #include <stdio.h>
10 #include <stdlib.h>
11 #include <stdarg.h>
12 #include <unistd.h>
13 #include <inttypes.h>
14 #include <string.h>
15 #include <fcntl.h>
16 #include <sys/mman.h>
17 #include <sys/mount.h>
18 #include <sys/param.h>
19 #include <malloc.h>
20 #include <stdbool.h>
21 #include <time.h>
22 #include "vm_util.h"
23 #include "../kselftest.h"
24
25 uint64_t pagesize;
26 unsigned int pageshift;
27 uint64_t pmd_pagesize;
28
29 #define SPLIT_DEBUGFS "/sys/kernel/debug/split_huge_pages"
30 #define SMAP_PATH "/proc/self/smaps"
31 #define INPUT_MAX 80
32
33 #define PID_FMT "%d,0x%lx,0x%lx,%d"
34 #define PID_FMT_OFFSET "%d,0x%lx,0x%lx,%d,%d"
35 #define PATH_FMT "%s,0x%lx,0x%lx,%d"
36
37 #define PFN_MASK ((1UL<<55)-1)
38 #define KPF_THP (1UL<<22)
39
is_backed_by_thp(char * vaddr,int pagemap_file,int kpageflags_file)40 int is_backed_by_thp(char *vaddr, int pagemap_file, int kpageflags_file)
41 {
42 uint64_t paddr;
43 uint64_t page_flags;
44
45 if (pagemap_file) {
46 pread(pagemap_file, &paddr, sizeof(paddr),
47 ((long)vaddr >> pageshift) * sizeof(paddr));
48
49 if (kpageflags_file) {
50 pread(kpageflags_file, &page_flags, sizeof(page_flags),
51 (paddr & PFN_MASK) * sizeof(page_flags));
52
53 return !!(page_flags & KPF_THP);
54 }
55 }
56 return 0;
57 }
58
write_file(const char * path,const char * buf,size_t buflen)59 static void write_file(const char *path, const char *buf, size_t buflen)
60 {
61 int fd;
62 ssize_t numwritten;
63
64 fd = open(path, O_WRONLY);
65 if (fd == -1)
66 ksft_exit_fail_msg("%s open failed: %s\n", path, strerror(errno));
67
68 numwritten = write(fd, buf, buflen - 1);
69 close(fd);
70 if (numwritten < 1)
71 ksft_exit_fail_msg("Write failed\n");
72 }
73
write_debugfs(const char * fmt,...)74 static void write_debugfs(const char *fmt, ...)
75 {
76 char input[INPUT_MAX];
77 int ret;
78 va_list argp;
79
80 va_start(argp, fmt);
81 ret = vsnprintf(input, INPUT_MAX, fmt, argp);
82 va_end(argp);
83
84 if (ret >= INPUT_MAX)
85 ksft_exit_fail_msg("%s: Debugfs input is too long\n", __func__);
86
87 write_file(SPLIT_DEBUGFS, input, ret + 1);
88 }
89
allocate_zero_filled_hugepage(size_t len)90 static char *allocate_zero_filled_hugepage(size_t len)
91 {
92 char *result;
93 size_t i;
94
95 result = memalign(pmd_pagesize, len);
96 if (!result) {
97 printf("Fail to allocate memory\n");
98 exit(EXIT_FAILURE);
99 }
100
101 madvise(result, len, MADV_HUGEPAGE);
102
103 for (i = 0; i < len; i++)
104 result[i] = (char)0;
105
106 return result;
107 }
108
verify_rss_anon_split_huge_page_all_zeroes(char * one_page,int nr_hpages,size_t len)109 static void verify_rss_anon_split_huge_page_all_zeroes(char *one_page, int nr_hpages, size_t len)
110 {
111 unsigned long rss_anon_before, rss_anon_after;
112 size_t i;
113
114 if (!check_huge_anon(one_page, 4, pmd_pagesize))
115 ksft_exit_fail_msg("No THP is allocated\n");
116
117 rss_anon_before = rss_anon();
118 if (!rss_anon_before)
119 ksft_exit_fail_msg("No RssAnon is allocated before split\n");
120
121 /* split all THPs */
122 write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
123 (uint64_t)one_page + len, 0);
124
125 for (i = 0; i < len; i++)
126 if (one_page[i] != (char)0)
127 ksft_exit_fail_msg("%ld byte corrupted\n", i);
128
129 if (!check_huge_anon(one_page, 0, pmd_pagesize))
130 ksft_exit_fail_msg("Still AnonHugePages not split\n");
131
132 rss_anon_after = rss_anon();
133 if (rss_anon_after >= rss_anon_before)
134 ksft_exit_fail_msg("Incorrect RssAnon value. Before: %ld After: %ld\n",
135 rss_anon_before, rss_anon_after);
136 }
137
split_pmd_zero_pages(void)138 void split_pmd_zero_pages(void)
139 {
140 char *one_page;
141 int nr_hpages = 4;
142 size_t len = nr_hpages * pmd_pagesize;
143
144 one_page = allocate_zero_filled_hugepage(len);
145 verify_rss_anon_split_huge_page_all_zeroes(one_page, nr_hpages, len);
146 ksft_test_result_pass("Split zero filled huge pages successful\n");
147 free(one_page);
148 }
149
split_pmd_thp_to_order(int order)150 void split_pmd_thp_to_order(int order)
151 {
152 char *one_page;
153 size_t len = 4 * pmd_pagesize;
154 size_t i;
155
156 one_page = memalign(pmd_pagesize, len);
157 if (!one_page)
158 ksft_exit_fail_msg("Fail to allocate memory: %s\n", strerror(errno));
159
160 madvise(one_page, len, MADV_HUGEPAGE);
161
162 for (i = 0; i < len; i++)
163 one_page[i] = (char)i;
164
165 if (!check_huge_anon(one_page, 4, pmd_pagesize))
166 ksft_exit_fail_msg("No THP is allocated\n");
167
168 /* split all THPs */
169 write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
170 (uint64_t)one_page + len, order);
171
172 for (i = 0; i < len; i++)
173 if (one_page[i] != (char)i)
174 ksft_exit_fail_msg("%ld byte corrupted\n", i);
175
176
177 if (!check_huge_anon(one_page, 0, pmd_pagesize))
178 ksft_exit_fail_msg("Still AnonHugePages not split\n");
179
180 ksft_test_result_pass("Split huge pages to order %d successful\n", order);
181 free(one_page);
182 }
183
split_pte_mapped_thp(void)184 void split_pte_mapped_thp(void)
185 {
186 char *one_page, *pte_mapped, *pte_mapped2;
187 size_t len = 4 * pmd_pagesize;
188 uint64_t thp_size;
189 size_t i;
190 const char *pagemap_template = "/proc/%d/pagemap";
191 const char *kpageflags_proc = "/proc/kpageflags";
192 char pagemap_proc[255];
193 int pagemap_fd;
194 int kpageflags_fd;
195
196 if (snprintf(pagemap_proc, 255, pagemap_template, getpid()) < 0)
197 ksft_exit_fail_msg("get pagemap proc error: %s\n", strerror(errno));
198
199 pagemap_fd = open(pagemap_proc, O_RDONLY);
200 if (pagemap_fd == -1)
201 ksft_exit_fail_msg("read pagemap: %s\n", strerror(errno));
202
203 kpageflags_fd = open(kpageflags_proc, O_RDONLY);
204 if (kpageflags_fd == -1)
205 ksft_exit_fail_msg("read kpageflags: %s\n", strerror(errno));
206
207 one_page = mmap((void *)(1UL << 30), len, PROT_READ | PROT_WRITE,
208 MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
209 if (one_page == MAP_FAILED)
210 ksft_exit_fail_msg("Fail to allocate memory: %s\n", strerror(errno));
211
212 madvise(one_page, len, MADV_HUGEPAGE);
213
214 for (i = 0; i < len; i++)
215 one_page[i] = (char)i;
216
217 if (!check_huge_anon(one_page, 4, pmd_pagesize))
218 ksft_exit_fail_msg("No THP is allocated\n");
219
220 /* remap the first pagesize of first THP */
221 pte_mapped = mremap(one_page, pagesize, pagesize, MREMAP_MAYMOVE);
222
223 /* remap the Nth pagesize of Nth THP */
224 for (i = 1; i < 4; i++) {
225 pte_mapped2 = mremap(one_page + pmd_pagesize * i + pagesize * i,
226 pagesize, pagesize,
227 MREMAP_MAYMOVE|MREMAP_FIXED,
228 pte_mapped + pagesize * i);
229 if (pte_mapped2 == MAP_FAILED)
230 ksft_exit_fail_msg("mremap failed: %s\n", strerror(errno));
231 }
232
233 /* smap does not show THPs after mremap, use kpageflags instead */
234 thp_size = 0;
235 for (i = 0; i < pagesize * 4; i++)
236 if (i % pagesize == 0 &&
237 is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd))
238 thp_size++;
239
240 if (thp_size != 4)
241 ksft_exit_fail_msg("Some THPs are missing during mremap\n");
242
243 /* split all remapped THPs */
244 write_debugfs(PID_FMT, getpid(), (uint64_t)pte_mapped,
245 (uint64_t)pte_mapped + pagesize * 4, 0);
246
247 /* smap does not show THPs after mremap, use kpageflags instead */
248 thp_size = 0;
249 for (i = 0; i < pagesize * 4; i++) {
250 if (pte_mapped[i] != (char)i)
251 ksft_exit_fail_msg("%ld byte corrupted\n", i);
252
253 if (i % pagesize == 0 &&
254 is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd))
255 thp_size++;
256 }
257
258 if (thp_size)
259 ksft_exit_fail_msg("Still %ld THPs not split\n", thp_size);
260
261 ksft_test_result_pass("Split PTE-mapped huge pages successful\n");
262 munmap(one_page, len);
263 close(pagemap_fd);
264 close(kpageflags_fd);
265 }
266
split_file_backed_thp(int order)267 void split_file_backed_thp(int order)
268 {
269 int status;
270 int fd;
271 char tmpfs_template[] = "/tmp/thp_split_XXXXXX";
272 const char *tmpfs_loc = mkdtemp(tmpfs_template);
273 char testfile[INPUT_MAX];
274 ssize_t num_written, num_read;
275 char *file_buf1, *file_buf2;
276 uint64_t pgoff_start = 0, pgoff_end = 1024;
277 int i;
278
279 ksft_print_msg("Please enable pr_debug in split_huge_pages_in_file() for more info.\n");
280
281 file_buf1 = (char *)malloc(pmd_pagesize);
282 file_buf2 = (char *)malloc(pmd_pagesize);
283
284 if (!file_buf1 || !file_buf2) {
285 ksft_print_msg("cannot allocate file buffers\n");
286 goto out;
287 }
288
289 for (i = 0; i < pmd_pagesize; i++)
290 file_buf1[i] = (char)i;
291 memset(file_buf2, 0, pmd_pagesize);
292
293 status = mount("tmpfs", tmpfs_loc, "tmpfs", 0, "huge=always,size=4m");
294
295 if (status)
296 ksft_exit_fail_msg("Unable to create a tmpfs for testing\n");
297
298 status = snprintf(testfile, INPUT_MAX, "%s/thp_file", tmpfs_loc);
299 if (status >= INPUT_MAX) {
300 ksft_exit_fail_msg("Fail to create file-backed THP split testing file\n");
301 goto cleanup;
302 }
303
304 fd = open(testfile, O_CREAT|O_RDWR, 0664);
305 if (fd == -1) {
306 ksft_perror("Cannot open testing file");
307 goto cleanup;
308 }
309
310 /* write pmd size data to the file, so a file-backed THP can be allocated */
311 num_written = write(fd, file_buf1, pmd_pagesize);
312
313 if (num_written == -1 || num_written != pmd_pagesize) {
314 ksft_perror("Failed to write data to testing file");
315 goto close_file;
316 }
317
318 /* split the file-backed THP */
319 write_debugfs(PATH_FMT, testfile, pgoff_start, pgoff_end, order);
320
321 /* check file content after split */
322 status = lseek(fd, 0, SEEK_SET);
323 if (status == -1) {
324 ksft_perror("Cannot lseek file");
325 goto close_file;
326 }
327
328 num_read = read(fd, file_buf2, num_written);
329 if (num_read == -1 || num_read != num_written) {
330 ksft_perror("Cannot read file content back");
331 goto close_file;
332 }
333
334 if (strncmp(file_buf1, file_buf2, pmd_pagesize) != 0) {
335 ksft_print_msg("File content changed\n");
336 goto close_file;
337 }
338
339 close(fd);
340 status = unlink(testfile);
341 if (status) {
342 ksft_perror("Cannot remove testing file");
343 goto cleanup;
344 }
345
346 status = umount(tmpfs_loc);
347 if (status) {
348 rmdir(tmpfs_loc);
349 ksft_exit_fail_msg("Unable to umount %s\n", tmpfs_loc);
350 }
351
352 status = rmdir(tmpfs_loc);
353 if (status)
354 ksft_exit_fail_msg("cannot remove tmp dir: %s\n", strerror(errno));
355
356 ksft_print_msg("Please check dmesg for more information\n");
357 ksft_test_result_pass("File-backed THP split to order %d test done\n", order);
358 return;
359
360 close_file:
361 close(fd);
362 cleanup:
363 umount(tmpfs_loc);
364 rmdir(tmpfs_loc);
365 out:
366 ksft_exit_fail_msg("Error occurred\n");
367 }
368
prepare_thp_fs(const char * xfs_path,char * thp_fs_template,const char ** thp_fs_loc)369 bool prepare_thp_fs(const char *xfs_path, char *thp_fs_template,
370 const char **thp_fs_loc)
371 {
372 if (xfs_path) {
373 *thp_fs_loc = xfs_path;
374 return false;
375 }
376
377 *thp_fs_loc = mkdtemp(thp_fs_template);
378
379 if (!*thp_fs_loc)
380 ksft_exit_fail_msg("cannot create temp folder\n");
381
382 return true;
383 }
384
cleanup_thp_fs(const char * thp_fs_loc,bool created_tmp)385 void cleanup_thp_fs(const char *thp_fs_loc, bool created_tmp)
386 {
387 int status;
388
389 if (!created_tmp)
390 return;
391
392 status = rmdir(thp_fs_loc);
393 if (status)
394 ksft_exit_fail_msg("cannot remove tmp dir: %s\n",
395 strerror(errno));
396 }
397
create_pagecache_thp_and_fd(const char * testfile,size_t fd_size,int * fd,char ** addr)398 int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd,
399 char **addr)
400 {
401 size_t i;
402 unsigned char buf[1024];
403
404 srand(time(NULL));
405
406 *fd = open(testfile, O_CREAT | O_RDWR, 0664);
407 if (*fd == -1)
408 ksft_exit_fail_msg("Failed to create a file at %s\n", testfile);
409
410 assert(fd_size % sizeof(buf) == 0);
411 for (i = 0; i < sizeof(buf); i++)
412 buf[i] = (unsigned char)i;
413 for (i = 0; i < fd_size; i += sizeof(buf))
414 write(*fd, buf, sizeof(buf));
415
416 close(*fd);
417 sync();
418 *fd = open("/proc/sys/vm/drop_caches", O_WRONLY);
419 if (*fd == -1) {
420 ksft_perror("open drop_caches");
421 goto err_out_unlink;
422 }
423 if (write(*fd, "3", 1) != 1) {
424 ksft_perror("write to drop_caches");
425 goto err_out_unlink;
426 }
427 close(*fd);
428
429 *fd = open(testfile, O_RDWR);
430 if (*fd == -1) {
431 ksft_perror("Failed to open testfile\n");
432 goto err_out_unlink;
433 }
434
435 *addr = mmap(NULL, fd_size, PROT_READ|PROT_WRITE, MAP_SHARED, *fd, 0);
436 if (*addr == (char *)-1) {
437 ksft_perror("cannot mmap");
438 goto err_out_close;
439 }
440 madvise(*addr, fd_size, MADV_HUGEPAGE);
441
442 for (size_t i = 0; i < fd_size; i++)
443 FORCE_READ((*addr + i));
444
445 if (!check_huge_file(*addr, fd_size / pmd_pagesize, pmd_pagesize)) {
446 ksft_print_msg("No large pagecache folio generated, please provide a filesystem supporting large folio\n");
447 munmap(*addr, fd_size);
448 close(*fd);
449 unlink(testfile);
450 ksft_test_result_skip("Pagecache folio split skipped\n");
451 return -2;
452 }
453 return 0;
454 err_out_close:
455 close(*fd);
456 err_out_unlink:
457 unlink(testfile);
458 ksft_exit_fail_msg("Failed to create large pagecache folios\n");
459 return -1;
460 }
461
split_thp_in_pagecache_to_order_at(size_t fd_size,const char * fs_loc,int order,int offset)462 void split_thp_in_pagecache_to_order_at(size_t fd_size, const char *fs_loc,
463 int order, int offset)
464 {
465 int fd;
466 char *addr;
467 size_t i;
468 char testfile[INPUT_MAX];
469 int err = 0;
470
471 err = snprintf(testfile, INPUT_MAX, "%s/test", fs_loc);
472
473 if (err < 0)
474 ksft_exit_fail_msg("cannot generate right test file name\n");
475
476 err = create_pagecache_thp_and_fd(testfile, fd_size, &fd, &addr);
477 if (err)
478 return;
479 err = 0;
480
481 if (offset == -1)
482 write_debugfs(PID_FMT, getpid(), (uint64_t)addr,
483 (uint64_t)addr + fd_size, order);
484 else
485 write_debugfs(PID_FMT_OFFSET, getpid(), (uint64_t)addr,
486 (uint64_t)addr + fd_size, order, offset);
487
488 for (i = 0; i < fd_size; i++)
489 if (*(addr + i) != (char)i) {
490 ksft_print_msg("%lu byte corrupted in the file\n", i);
491 err = EXIT_FAILURE;
492 goto out;
493 }
494
495 if (!check_huge_file(addr, 0, pmd_pagesize)) {
496 ksft_print_msg("Still FilePmdMapped not split\n");
497 err = EXIT_FAILURE;
498 goto out;
499 }
500
501 out:
502 munmap(addr, fd_size);
503 close(fd);
504 unlink(testfile);
505 if (offset == -1) {
506 if (err)
507 ksft_exit_fail_msg("Split PMD-mapped pagecache folio to order %d failed\n", order);
508 ksft_test_result_pass("Split PMD-mapped pagecache folio to order %d passed\n", order);
509 } else {
510 if (err)
511 ksft_exit_fail_msg("Split PMD-mapped pagecache folio to order %d at in-folio offset %d failed\n", order, offset);
512 ksft_test_result_pass("Split PMD-mapped pagecache folio to order %d at in-folio offset %d passed\n", order, offset);
513 }
514 }
515
main(int argc,char ** argv)516 int main(int argc, char **argv)
517 {
518 int i;
519 size_t fd_size;
520 char *optional_xfs_path = NULL;
521 char fs_loc_template[] = "/tmp/thp_fs_XXXXXX";
522 const char *fs_loc;
523 bool created_tmp;
524 int offset;
525
526 ksft_print_header();
527
528 if (geteuid() != 0) {
529 ksft_print_msg("Please run the benchmark as root\n");
530 ksft_finished();
531 }
532
533 if (argc > 1)
534 optional_xfs_path = argv[1];
535
536 ksft_set_plan(1+8+1+9+9+8*4+2);
537
538 pagesize = getpagesize();
539 pageshift = ffs(pagesize) - 1;
540 pmd_pagesize = read_pmd_pagesize();
541 if (!pmd_pagesize)
542 ksft_exit_fail_msg("Reading PMD pagesize failed\n");
543
544 fd_size = 2 * pmd_pagesize;
545
546 split_pmd_zero_pages();
547
548 for (i = 0; i < 9; i++)
549 if (i != 1)
550 split_pmd_thp_to_order(i);
551
552 split_pte_mapped_thp();
553 for (i = 0; i < 9; i++)
554 split_file_backed_thp(i);
555
556 created_tmp = prepare_thp_fs(optional_xfs_path, fs_loc_template,
557 &fs_loc);
558 for (i = 8; i >= 0; i--)
559 split_thp_in_pagecache_to_order_at(fd_size, fs_loc, i, -1);
560
561 for (i = 0; i < 9; i++)
562 for (offset = 0;
563 offset < pmd_pagesize / pagesize;
564 offset += MAX(pmd_pagesize / pagesize / 4, 1 << i))
565 split_thp_in_pagecache_to_order_at(fd_size, fs_loc, i, offset);
566 cleanup_thp_fs(fs_loc, created_tmp);
567
568 ksft_finished();
569
570 return 0;
571 }
572