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 PATH_FMT "%s,0x%lx,0x%lx,%d"
35
36 #define PFN_MASK ((1UL<<55)-1)
37 #define KPF_THP (1UL<<22)
38
is_backed_by_thp(char * vaddr,int pagemap_file,int kpageflags_file)39 int is_backed_by_thp(char *vaddr, int pagemap_file, int kpageflags_file)
40 {
41 uint64_t paddr;
42 uint64_t page_flags;
43
44 if (pagemap_file) {
45 pread(pagemap_file, &paddr, sizeof(paddr),
46 ((long)vaddr >> pageshift) * sizeof(paddr));
47
48 if (kpageflags_file) {
49 pread(kpageflags_file, &page_flags, sizeof(page_flags),
50 (paddr & PFN_MASK) * sizeof(page_flags));
51
52 return !!(page_flags & KPF_THP);
53 }
54 }
55 return 0;
56 }
57
write_file(const char * path,const char * buf,size_t buflen)58 static void write_file(const char *path, const char *buf, size_t buflen)
59 {
60 int fd;
61 ssize_t numwritten;
62
63 fd = open(path, O_WRONLY);
64 if (fd == -1)
65 ksft_exit_fail_msg("%s open failed: %s\n", path, strerror(errno));
66
67 numwritten = write(fd, buf, buflen - 1);
68 close(fd);
69 if (numwritten < 1)
70 ksft_exit_fail_msg("Write failed\n");
71 }
72
write_debugfs(const char * fmt,...)73 static void write_debugfs(const char *fmt, ...)
74 {
75 char input[INPUT_MAX];
76 int ret;
77 va_list argp;
78
79 va_start(argp, fmt);
80 ret = vsnprintf(input, INPUT_MAX, fmt, argp);
81 va_end(argp);
82
83 if (ret >= INPUT_MAX)
84 ksft_exit_fail_msg("%s: Debugfs input is too long\n", __func__);
85
86 write_file(SPLIT_DEBUGFS, input, ret + 1);
87 }
88
allocate_zero_filled_hugepage(size_t len)89 static char *allocate_zero_filled_hugepage(size_t len)
90 {
91 char *result;
92 size_t i;
93
94 result = memalign(pmd_pagesize, len);
95 if (!result) {
96 printf("Fail to allocate memory\n");
97 exit(EXIT_FAILURE);
98 }
99
100 madvise(result, len, MADV_HUGEPAGE);
101
102 for (i = 0; i < len; i++)
103 result[i] = (char)0;
104
105 return result;
106 }
107
verify_rss_anon_split_huge_page_all_zeroes(char * one_page,int nr_hpages,size_t len)108 static void verify_rss_anon_split_huge_page_all_zeroes(char *one_page, int nr_hpages, size_t len)
109 {
110 unsigned long rss_anon_before, rss_anon_after;
111 size_t i;
112
113 if (!check_huge_anon(one_page, 4, pmd_pagesize))
114 ksft_exit_fail_msg("No THP is allocated\n");
115
116 rss_anon_before = rss_anon();
117 if (!rss_anon_before)
118 ksft_exit_fail_msg("No RssAnon is allocated before split\n");
119
120 /* split all THPs */
121 write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
122 (uint64_t)one_page + len, 0);
123
124 for (i = 0; i < len; i++)
125 if (one_page[i] != (char)0)
126 ksft_exit_fail_msg("%ld byte corrupted\n", i);
127
128 if (!check_huge_anon(one_page, 0, pmd_pagesize))
129 ksft_exit_fail_msg("Still AnonHugePages not split\n");
130
131 rss_anon_after = rss_anon();
132 if (rss_anon_after >= rss_anon_before)
133 ksft_exit_fail_msg("Incorrect RssAnon value. Before: %ld After: %ld\n",
134 rss_anon_before, rss_anon_after);
135 }
136
split_pmd_zero_pages(void)137 void split_pmd_zero_pages(void)
138 {
139 char *one_page;
140 int nr_hpages = 4;
141 size_t len = nr_hpages * pmd_pagesize;
142
143 one_page = allocate_zero_filled_hugepage(len);
144 verify_rss_anon_split_huge_page_all_zeroes(one_page, nr_hpages, len);
145 ksft_test_result_pass("Split zero filled huge pages successful\n");
146 free(one_page);
147 }
148
split_pmd_thp_to_order(int order)149 void split_pmd_thp_to_order(int order)
150 {
151 char *one_page;
152 size_t len = 4 * pmd_pagesize;
153 size_t i;
154
155 one_page = memalign(pmd_pagesize, len);
156 if (!one_page)
157 ksft_exit_fail_msg("Fail to allocate memory: %s\n", strerror(errno));
158
159 madvise(one_page, len, MADV_HUGEPAGE);
160
161 for (i = 0; i < len; i++)
162 one_page[i] = (char)i;
163
164 if (!check_huge_anon(one_page, 4, pmd_pagesize))
165 ksft_exit_fail_msg("No THP is allocated\n");
166
167 /* split all THPs */
168 write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
169 (uint64_t)one_page + len, order);
170
171 for (i = 0; i < len; i++)
172 if (one_page[i] != (char)i)
173 ksft_exit_fail_msg("%ld byte corrupted\n", i);
174
175
176 if (!check_huge_anon(one_page, 0, pmd_pagesize))
177 ksft_exit_fail_msg("Still AnonHugePages not split\n");
178
179 ksft_test_result_pass("Split huge pages to order %d successful\n", order);
180 free(one_page);
181 }
182
split_pte_mapped_thp(void)183 void split_pte_mapped_thp(void)
184 {
185 char *one_page, *pte_mapped, *pte_mapped2;
186 size_t len = 4 * pmd_pagesize;
187 uint64_t thp_size;
188 size_t i;
189 const char *pagemap_template = "/proc/%d/pagemap";
190 const char *kpageflags_proc = "/proc/kpageflags";
191 char pagemap_proc[255];
192 int pagemap_fd;
193 int kpageflags_fd;
194
195 if (snprintf(pagemap_proc, 255, pagemap_template, getpid()) < 0)
196 ksft_exit_fail_msg("get pagemap proc error: %s\n", strerror(errno));
197
198 pagemap_fd = open(pagemap_proc, O_RDONLY);
199 if (pagemap_fd == -1)
200 ksft_exit_fail_msg("read pagemap: %s\n", strerror(errno));
201
202 kpageflags_fd = open(kpageflags_proc, O_RDONLY);
203 if (kpageflags_fd == -1)
204 ksft_exit_fail_msg("read kpageflags: %s\n", strerror(errno));
205
206 one_page = mmap((void *)(1UL << 30), len, PROT_READ | PROT_WRITE,
207 MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
208 if (one_page == MAP_FAILED)
209 ksft_exit_fail_msg("Fail to allocate memory: %s\n", strerror(errno));
210
211 madvise(one_page, len, MADV_HUGEPAGE);
212
213 for (i = 0; i < len; i++)
214 one_page[i] = (char)i;
215
216 if (!check_huge_anon(one_page, 4, pmd_pagesize))
217 ksft_exit_fail_msg("No THP is allocated\n");
218
219 /* remap the first pagesize of first THP */
220 pte_mapped = mremap(one_page, pagesize, pagesize, MREMAP_MAYMOVE);
221
222 /* remap the Nth pagesize of Nth THP */
223 for (i = 1; i < 4; i++) {
224 pte_mapped2 = mremap(one_page + pmd_pagesize * i + pagesize * i,
225 pagesize, pagesize,
226 MREMAP_MAYMOVE|MREMAP_FIXED,
227 pte_mapped + pagesize * i);
228 if (pte_mapped2 == MAP_FAILED)
229 ksft_exit_fail_msg("mremap failed: %s\n", strerror(errno));
230 }
231
232 /* smap does not show THPs after mremap, use kpageflags instead */
233 thp_size = 0;
234 for (i = 0; i < pagesize * 4; i++)
235 if (i % pagesize == 0 &&
236 is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd))
237 thp_size++;
238
239 if (thp_size != 4)
240 ksft_exit_fail_msg("Some THPs are missing during mremap\n");
241
242 /* split all remapped THPs */
243 write_debugfs(PID_FMT, getpid(), (uint64_t)pte_mapped,
244 (uint64_t)pte_mapped + pagesize * 4, 0);
245
246 /* smap does not show THPs after mremap, use kpageflags instead */
247 thp_size = 0;
248 for (i = 0; i < pagesize * 4; i++) {
249 if (pte_mapped[i] != (char)i)
250 ksft_exit_fail_msg("%ld byte corrupted\n", i);
251
252 if (i % pagesize == 0 &&
253 is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd))
254 thp_size++;
255 }
256
257 if (thp_size)
258 ksft_exit_fail_msg("Still %ld THPs not split\n", thp_size);
259
260 ksft_test_result_pass("Split PTE-mapped huge pages successful\n");
261 munmap(one_page, len);
262 close(pagemap_fd);
263 close(kpageflags_fd);
264 }
265
split_file_backed_thp(int order)266 void split_file_backed_thp(int order)
267 {
268 int status;
269 int fd;
270 char tmpfs_template[] = "/tmp/thp_split_XXXXXX";
271 const char *tmpfs_loc = mkdtemp(tmpfs_template);
272 char testfile[INPUT_MAX];
273 ssize_t num_written, num_read;
274 char *file_buf1, *file_buf2;
275 uint64_t pgoff_start = 0, pgoff_end = 1024;
276 int i;
277
278 ksft_print_msg("Please enable pr_debug in split_huge_pages_in_file() for more info.\n");
279
280 file_buf1 = (char *)malloc(pmd_pagesize);
281 file_buf2 = (char *)malloc(pmd_pagesize);
282
283 if (!file_buf1 || !file_buf2) {
284 ksft_print_msg("cannot allocate file buffers\n");
285 goto out;
286 }
287
288 for (i = 0; i < pmd_pagesize; i++)
289 file_buf1[i] = (char)i;
290 memset(file_buf2, 0, pmd_pagesize);
291
292 status = mount("tmpfs", tmpfs_loc, "tmpfs", 0, "huge=always,size=4m");
293
294 if (status)
295 ksft_exit_fail_msg("Unable to create a tmpfs for testing\n");
296
297 status = snprintf(testfile, INPUT_MAX, "%s/thp_file", tmpfs_loc);
298 if (status >= INPUT_MAX) {
299 ksft_exit_fail_msg("Fail to create file-backed THP split testing file\n");
300 goto cleanup;
301 }
302
303 fd = open(testfile, O_CREAT|O_RDWR, 0664);
304 if (fd == -1) {
305 ksft_perror("Cannot open testing file");
306 goto cleanup;
307 }
308
309 /* write pmd size data to the file, so a file-backed THP can be allocated */
310 num_written = write(fd, file_buf1, pmd_pagesize);
311
312 if (num_written == -1 || num_written != pmd_pagesize) {
313 ksft_perror("Failed to write data to testing file");
314 goto close_file;
315 }
316
317 /* split the file-backed THP */
318 write_debugfs(PATH_FMT, testfile, pgoff_start, pgoff_end, order);
319
320 /* check file content after split */
321 status = lseek(fd, 0, SEEK_SET);
322 if (status == -1) {
323 ksft_perror("Cannot lseek file");
324 goto close_file;
325 }
326
327 num_read = read(fd, file_buf2, num_written);
328 if (num_read == -1 || num_read != num_written) {
329 ksft_perror("Cannot read file content back");
330 goto close_file;
331 }
332
333 if (strncmp(file_buf1, file_buf2, pmd_pagesize) != 0) {
334 ksft_print_msg("File content changed\n");
335 goto close_file;
336 }
337
338 close(fd);
339 status = unlink(testfile);
340 if (status) {
341 ksft_perror("Cannot remove testing file");
342 goto cleanup;
343 }
344
345 status = umount(tmpfs_loc);
346 if (status) {
347 rmdir(tmpfs_loc);
348 ksft_exit_fail_msg("Unable to umount %s\n", tmpfs_loc);
349 }
350
351 status = rmdir(tmpfs_loc);
352 if (status)
353 ksft_exit_fail_msg("cannot remove tmp dir: %s\n", strerror(errno));
354
355 ksft_print_msg("Please check dmesg for more information\n");
356 ksft_test_result_pass("File-backed THP split to order %d test done\n", order);
357 return;
358
359 close_file:
360 close(fd);
361 cleanup:
362 umount(tmpfs_loc);
363 rmdir(tmpfs_loc);
364 out:
365 ksft_exit_fail_msg("Error occurred\n");
366 }
367
prepare_thp_fs(const char * xfs_path,char * thp_fs_template,const char ** thp_fs_loc)368 bool prepare_thp_fs(const char *xfs_path, char *thp_fs_template,
369 const char **thp_fs_loc)
370 {
371 if (xfs_path) {
372 *thp_fs_loc = xfs_path;
373 return false;
374 }
375
376 *thp_fs_loc = mkdtemp(thp_fs_template);
377
378 if (!*thp_fs_loc)
379 ksft_exit_fail_msg("cannot create temp folder\n");
380
381 return true;
382 }
383
cleanup_thp_fs(const char * thp_fs_loc,bool created_tmp)384 void cleanup_thp_fs(const char *thp_fs_loc, bool created_tmp)
385 {
386 int status;
387
388 if (!created_tmp)
389 return;
390
391 status = rmdir(thp_fs_loc);
392 if (status)
393 ksft_exit_fail_msg("cannot remove tmp dir: %s\n",
394 strerror(errno));
395 }
396
create_pagecache_thp_and_fd(const char * testfile,size_t fd_size,int * fd,char ** addr)397 int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd,
398 char **addr)
399 {
400 size_t i;
401 int dummy = 0;
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 dummy += *(*addr + i);
444 asm volatile("" : "+r" (dummy));
445
446 if (!check_huge_file(*addr, fd_size / pmd_pagesize, pmd_pagesize)) {
447 ksft_print_msg("No large pagecache folio generated, please provide a filesystem supporting large folio\n");
448 munmap(*addr, fd_size);
449 close(*fd);
450 unlink(testfile);
451 ksft_test_result_skip("Pagecache folio split skipped\n");
452 return -2;
453 }
454 return 0;
455 err_out_close:
456 close(*fd);
457 err_out_unlink:
458 unlink(testfile);
459 ksft_exit_fail_msg("Failed to create large pagecache folios\n");
460 return -1;
461 }
462
split_thp_in_pagecache_to_order_at(size_t fd_size,const char * fs_loc,int order,int offset)463 void split_thp_in_pagecache_to_order_at(size_t fd_size, const char *fs_loc,
464 int order, int offset)
465 {
466 int fd;
467 char *addr;
468 size_t i;
469 char testfile[INPUT_MAX];
470 int err = 0;
471
472 err = snprintf(testfile, INPUT_MAX, "%s/test", fs_loc);
473
474 if (err < 0)
475 ksft_exit_fail_msg("cannot generate right test file name\n");
476
477 err = create_pagecache_thp_and_fd(testfile, fd_size, &fd, &addr);
478 if (err)
479 return;
480 err = 0;
481
482 if (offset == -1)
483 write_debugfs(PID_FMT, getpid(), (uint64_t)addr,
484 (uint64_t)addr + fd_size, order);
485 else
486 write_debugfs(PID_FMT, getpid(), (uint64_t)addr,
487 (uint64_t)addr + fd_size, order, offset);
488
489 for (i = 0; i < fd_size; i++)
490 if (*(addr + i) != (char)i) {
491 ksft_print_msg("%lu byte corrupted in the file\n", i);
492 err = EXIT_FAILURE;
493 goto out;
494 }
495
496 if (!check_huge_file(addr, 0, pmd_pagesize)) {
497 ksft_print_msg("Still FilePmdMapped not split\n");
498 err = EXIT_FAILURE;
499 goto out;
500 }
501
502 out:
503 munmap(addr, fd_size);
504 close(fd);
505 unlink(testfile);
506 if (offset == -1) {
507 if (err)
508 ksft_exit_fail_msg("Split PMD-mapped pagecache folio to order %d failed\n", order);
509 ksft_test_result_pass("Split PMD-mapped pagecache folio to order %d passed\n", order);
510 } else {
511 if (err)
512 ksft_exit_fail_msg("Split PMD-mapped pagecache folio to order %d at in-folio offset %d failed\n", order, offset);
513 ksft_test_result_pass("Split PMD-mapped pagecache folio to order %d at in-folio offset %d passed\n", order, offset);
514 }
515 }
516
main(int argc,char ** argv)517 int main(int argc, char **argv)
518 {
519 int i;
520 size_t fd_size;
521 char *optional_xfs_path = NULL;
522 char fs_loc_template[] = "/tmp/thp_fs_XXXXXX";
523 const char *fs_loc;
524 bool created_tmp;
525 int offset;
526
527 ksft_print_header();
528
529 if (geteuid() != 0) {
530 ksft_print_msg("Please run the benchmark as root\n");
531 ksft_finished();
532 }
533
534 if (argc > 1)
535 optional_xfs_path = argv[1];
536
537 ksft_set_plan(1+8+1+9+9+8*4+2);
538
539 pagesize = getpagesize();
540 pageshift = ffs(pagesize) - 1;
541 pmd_pagesize = read_pmd_pagesize();
542 if (!pmd_pagesize)
543 ksft_exit_fail_msg("Reading PMD pagesize failed\n");
544
545 fd_size = 2 * pmd_pagesize;
546
547 split_pmd_zero_pages();
548
549 for (i = 0; i < 9; i++)
550 if (i != 1)
551 split_pmd_thp_to_order(i);
552
553 split_pte_mapped_thp();
554 for (i = 0; i < 9; i++)
555 split_file_backed_thp(i);
556
557 created_tmp = prepare_thp_fs(optional_xfs_path, fs_loc_template,
558 &fs_loc);
559 for (i = 8; i >= 0; i--)
560 split_thp_in_pagecache_to_order_at(fd_size, fs_loc, i, -1);
561
562 for (i = 0; i < 9; i++)
563 for (offset = 0;
564 offset < pmd_pagesize / pagesize;
565 offset += MAX(pmd_pagesize / pagesize / 4, 1 << i))
566 split_thp_in_pagecache_to_order_at(fd_size, fs_loc, i, offset);
567 cleanup_thp_fs(fs_loc, created_tmp);
568
569 ksft_finished();
570
571 return 0;
572 }
573