]>
Commit | Line | Data |
---|---|---|
fa6c0231 ZY |
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 <stdio.h> | |
9 | #include <stdlib.h> | |
fbe37501 | 10 | #include <stdarg.h> |
fa6c0231 ZY |
11 | #include <unistd.h> |
12 | #include <inttypes.h> | |
13 | #include <string.h> | |
14 | #include <fcntl.h> | |
15 | #include <sys/mman.h> | |
fbe37501 | 16 | #include <sys/mount.h> |
fa6c0231 ZY |
17 | #include <malloc.h> |
18 | #include <stdbool.h> | |
19 | ||
20 | uint64_t pagesize; | |
21 | unsigned int pageshift; | |
22 | uint64_t pmd_pagesize; | |
23 | ||
24 | #define PMD_SIZE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size" | |
25 | #define SPLIT_DEBUGFS "/sys/kernel/debug/split_huge_pages" | |
26 | #define SMAP_PATH "/proc/self/smaps" | |
27 | #define INPUT_MAX 80 | |
28 | ||
fbe37501 ZY |
29 | #define PID_FMT "%d,0x%lx,0x%lx" |
30 | #define PATH_FMT "%s,0x%lx,0x%lx" | |
31 | ||
fa6c0231 ZY |
32 | #define PFN_MASK ((1UL<<55)-1) |
33 | #define KPF_THP (1UL<<22) | |
34 | ||
35 | int is_backed_by_thp(char *vaddr, int pagemap_file, int kpageflags_file) | |
36 | { | |
37 | uint64_t paddr; | |
38 | uint64_t page_flags; | |
39 | ||
40 | if (pagemap_file) { | |
41 | pread(pagemap_file, &paddr, sizeof(paddr), | |
42 | ((long)vaddr >> pageshift) * sizeof(paddr)); | |
43 | ||
44 | if (kpageflags_file) { | |
45 | pread(kpageflags_file, &page_flags, sizeof(page_flags), | |
46 | (paddr & PFN_MASK) * sizeof(page_flags)); | |
47 | ||
48 | return !!(page_flags & KPF_THP); | |
49 | } | |
50 | } | |
51 | return 0; | |
52 | } | |
53 | ||
54 | ||
55 | static uint64_t read_pmd_pagesize(void) | |
56 | { | |
57 | int fd; | |
58 | char buf[20]; | |
59 | ssize_t num_read; | |
60 | ||
61 | fd = open(PMD_SIZE_PATH, O_RDONLY); | |
62 | if (fd == -1) { | |
63 | perror("Open hpage_pmd_size failed"); | |
64 | exit(EXIT_FAILURE); | |
65 | } | |
66 | num_read = read(fd, buf, 19); | |
67 | if (num_read < 1) { | |
68 | close(fd); | |
69 | perror("Read hpage_pmd_size failed"); | |
70 | exit(EXIT_FAILURE); | |
71 | } | |
72 | buf[num_read] = '\0'; | |
73 | close(fd); | |
74 | ||
75 | return strtoul(buf, NULL, 10); | |
76 | } | |
77 | ||
78 | static int write_file(const char *path, const char *buf, size_t buflen) | |
79 | { | |
80 | int fd; | |
81 | ssize_t numwritten; | |
82 | ||
83 | fd = open(path, O_WRONLY); | |
84 | if (fd == -1) | |
85 | return 0; | |
86 | ||
87 | numwritten = write(fd, buf, buflen - 1); | |
88 | close(fd); | |
89 | if (numwritten < 1) | |
90 | return 0; | |
91 | ||
92 | return (unsigned int) numwritten; | |
93 | } | |
94 | ||
fbe37501 | 95 | static void write_debugfs(const char *fmt, ...) |
fa6c0231 ZY |
96 | { |
97 | char input[INPUT_MAX]; | |
98 | int ret; | |
fbe37501 ZY |
99 | va_list argp; |
100 | ||
101 | va_start(argp, fmt); | |
102 | ret = vsnprintf(input, INPUT_MAX, fmt, argp); | |
103 | va_end(argp); | |
fa6c0231 | 104 | |
fa6c0231 ZY |
105 | if (ret >= INPUT_MAX) { |
106 | printf("%s: Debugfs input is too long\n", __func__); | |
107 | exit(EXIT_FAILURE); | |
108 | } | |
109 | ||
110 | if (!write_file(SPLIT_DEBUGFS, input, ret + 1)) { | |
111 | perror(SPLIT_DEBUGFS); | |
112 | exit(EXIT_FAILURE); | |
113 | } | |
114 | } | |
115 | ||
116 | #define MAX_LINE_LENGTH 500 | |
117 | ||
118 | static bool check_for_pattern(FILE *fp, const char *pattern, char *buf) | |
119 | { | |
120 | while (fgets(buf, MAX_LINE_LENGTH, fp) != NULL) { | |
121 | if (!strncmp(buf, pattern, strlen(pattern))) | |
122 | return true; | |
123 | } | |
124 | return false; | |
125 | } | |
126 | ||
127 | static uint64_t check_huge(void *addr) | |
128 | { | |
129 | uint64_t thp = 0; | |
130 | int ret; | |
131 | FILE *fp; | |
132 | char buffer[MAX_LINE_LENGTH]; | |
133 | char addr_pattern[MAX_LINE_LENGTH]; | |
134 | ||
135 | ret = snprintf(addr_pattern, MAX_LINE_LENGTH, "%08lx-", | |
136 | (unsigned long) addr); | |
137 | if (ret >= MAX_LINE_LENGTH) { | |
138 | printf("%s: Pattern is too long\n", __func__); | |
139 | exit(EXIT_FAILURE); | |
140 | } | |
141 | ||
142 | ||
143 | fp = fopen(SMAP_PATH, "r"); | |
144 | if (!fp) { | |
145 | printf("%s: Failed to open file %s\n", __func__, SMAP_PATH); | |
146 | exit(EXIT_FAILURE); | |
147 | } | |
148 | if (!check_for_pattern(fp, addr_pattern, buffer)) | |
149 | goto err_out; | |
150 | ||
151 | /* | |
152 | * Fetch the AnonHugePages: in the same block and check the number of | |
153 | * hugepages. | |
154 | */ | |
155 | if (!check_for_pattern(fp, "AnonHugePages:", buffer)) | |
156 | goto err_out; | |
157 | ||
158 | if (sscanf(buffer, "AnonHugePages:%10ld kB", &thp) != 1) { | |
159 | printf("Reading smap error\n"); | |
160 | exit(EXIT_FAILURE); | |
161 | } | |
162 | ||
163 | err_out: | |
164 | fclose(fp); | |
165 | return thp; | |
166 | } | |
167 | ||
168 | void split_pmd_thp(void) | |
169 | { | |
170 | char *one_page; | |
171 | size_t len = 4 * pmd_pagesize; | |
172 | uint64_t thp_size; | |
173 | size_t i; | |
174 | ||
175 | one_page = memalign(pmd_pagesize, len); | |
176 | ||
177 | if (!one_page) { | |
178 | printf("Fail to allocate memory\n"); | |
179 | exit(EXIT_FAILURE); | |
180 | } | |
181 | ||
182 | madvise(one_page, len, MADV_HUGEPAGE); | |
183 | ||
184 | for (i = 0; i < len; i++) | |
185 | one_page[i] = (char)i; | |
186 | ||
187 | thp_size = check_huge(one_page); | |
188 | if (!thp_size) { | |
189 | printf("No THP is allocated\n"); | |
190 | exit(EXIT_FAILURE); | |
191 | } | |
192 | ||
193 | /* split all THPs */ | |
fbe37501 ZY |
194 | write_debugfs(PID_FMT, getpid(), (uint64_t)one_page, |
195 | (uint64_t)one_page + len); | |
fa6c0231 ZY |
196 | |
197 | for (i = 0; i < len; i++) | |
198 | if (one_page[i] != (char)i) { | |
199 | printf("%ld byte corrupted\n", i); | |
200 | exit(EXIT_FAILURE); | |
201 | } | |
202 | ||
203 | ||
204 | thp_size = check_huge(one_page); | |
205 | if (thp_size) { | |
206 | printf("Still %ld kB AnonHugePages not split\n", thp_size); | |
207 | exit(EXIT_FAILURE); | |
208 | } | |
209 | ||
210 | printf("Split huge pages successful\n"); | |
211 | free(one_page); | |
212 | } | |
213 | ||
214 | void split_pte_mapped_thp(void) | |
215 | { | |
216 | char *one_page, *pte_mapped, *pte_mapped2; | |
217 | size_t len = 4 * pmd_pagesize; | |
218 | uint64_t thp_size; | |
219 | size_t i; | |
220 | const char *pagemap_template = "/proc/%d/pagemap"; | |
221 | const char *kpageflags_proc = "/proc/kpageflags"; | |
222 | char pagemap_proc[255]; | |
223 | int pagemap_fd; | |
224 | int kpageflags_fd; | |
225 | ||
226 | if (snprintf(pagemap_proc, 255, pagemap_template, getpid()) < 0) { | |
227 | perror("get pagemap proc error"); | |
228 | exit(EXIT_FAILURE); | |
229 | } | |
230 | pagemap_fd = open(pagemap_proc, O_RDONLY); | |
231 | ||
232 | if (pagemap_fd == -1) { | |
233 | perror("read pagemap:"); | |
234 | exit(EXIT_FAILURE); | |
235 | } | |
236 | ||
237 | kpageflags_fd = open(kpageflags_proc, O_RDONLY); | |
238 | ||
239 | if (kpageflags_fd == -1) { | |
240 | perror("read kpageflags:"); | |
241 | exit(EXIT_FAILURE); | |
242 | } | |
243 | ||
244 | one_page = mmap((void *)(1UL << 30), len, PROT_READ | PROT_WRITE, | |
245 | MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); | |
246 | ||
247 | madvise(one_page, len, MADV_HUGEPAGE); | |
248 | ||
249 | for (i = 0; i < len; i++) | |
250 | one_page[i] = (char)i; | |
251 | ||
252 | thp_size = check_huge(one_page); | |
253 | if (!thp_size) { | |
254 | printf("No THP is allocated\n"); | |
255 | exit(EXIT_FAILURE); | |
256 | } | |
257 | ||
258 | /* remap the first pagesize of first THP */ | |
259 | pte_mapped = mremap(one_page, pagesize, pagesize, MREMAP_MAYMOVE); | |
260 | ||
261 | /* remap the Nth pagesize of Nth THP */ | |
262 | for (i = 1; i < 4; i++) { | |
263 | pte_mapped2 = mremap(one_page + pmd_pagesize * i + pagesize * i, | |
264 | pagesize, pagesize, | |
265 | MREMAP_MAYMOVE|MREMAP_FIXED, | |
266 | pte_mapped + pagesize * i); | |
267 | if (pte_mapped2 == (char *)-1) { | |
268 | perror("mremap failed"); | |
269 | exit(EXIT_FAILURE); | |
270 | } | |
271 | } | |
272 | ||
273 | /* smap does not show THPs after mremap, use kpageflags instead */ | |
274 | thp_size = 0; | |
275 | for (i = 0; i < pagesize * 4; i++) | |
276 | if (i % pagesize == 0 && | |
277 | is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd)) | |
278 | thp_size++; | |
279 | ||
280 | if (thp_size != 4) { | |
281 | printf("Some THPs are missing during mremap\n"); | |
282 | exit(EXIT_FAILURE); | |
283 | } | |
284 | ||
285 | /* split all remapped THPs */ | |
fbe37501 | 286 | write_debugfs(PID_FMT, getpid(), (uint64_t)pte_mapped, |
fa6c0231 ZY |
287 | (uint64_t)pte_mapped + pagesize * 4); |
288 | ||
289 | /* smap does not show THPs after mremap, use kpageflags instead */ | |
290 | thp_size = 0; | |
291 | for (i = 0; i < pagesize * 4; i++) { | |
292 | if (pte_mapped[i] != (char)i) { | |
293 | printf("%ld byte corrupted\n", i); | |
294 | exit(EXIT_FAILURE); | |
295 | } | |
296 | if (i % pagesize == 0 && | |
297 | is_backed_by_thp(&pte_mapped[i], pagemap_fd, kpageflags_fd)) | |
298 | thp_size++; | |
299 | } | |
300 | ||
301 | if (thp_size) { | |
302 | printf("Still %ld THPs not split\n", thp_size); | |
303 | exit(EXIT_FAILURE); | |
304 | } | |
305 | ||
306 | printf("Split PTE-mapped huge pages successful\n"); | |
307 | munmap(one_page, len); | |
308 | close(pagemap_fd); | |
309 | close(kpageflags_fd); | |
310 | } | |
311 | ||
fbe37501 ZY |
312 | void split_file_backed_thp(void) |
313 | { | |
314 | int status; | |
315 | int fd; | |
316 | ssize_t num_written; | |
317 | char tmpfs_template[] = "/tmp/thp_split_XXXXXX"; | |
318 | const char *tmpfs_loc = mkdtemp(tmpfs_template); | |
319 | char testfile[INPUT_MAX]; | |
320 | uint64_t pgoff_start = 0, pgoff_end = 1024; | |
321 | ||
322 | printf("Please enable pr_debug in split_huge_pages_in_file() if you need more info.\n"); | |
323 | ||
324 | status = mount("tmpfs", tmpfs_loc, "tmpfs", 0, "huge=always,size=4m"); | |
325 | ||
326 | if (status) { | |
327 | printf("Unable to create a tmpfs for testing\n"); | |
328 | exit(EXIT_FAILURE); | |
329 | } | |
330 | ||
331 | status = snprintf(testfile, INPUT_MAX, "%s/thp_file", tmpfs_loc); | |
332 | if (status >= INPUT_MAX) { | |
333 | printf("Fail to create file-backed THP split testing file\n"); | |
334 | goto cleanup; | |
335 | } | |
336 | ||
337 | fd = open(testfile, O_CREAT|O_WRONLY); | |
338 | if (fd == -1) { | |
339 | perror("Cannot open testing file\n"); | |
340 | goto cleanup; | |
341 | } | |
342 | ||
343 | /* write something to the file, so a file-backed THP can be allocated */ | |
9c7516d6 | 344 | num_written = write(fd, tmpfs_loc, strlen(tmpfs_loc) + 1); |
fbe37501 ZY |
345 | close(fd); |
346 | ||
347 | if (num_written < 1) { | |
348 | printf("Fail to write data to testing file\n"); | |
349 | goto cleanup; | |
350 | } | |
351 | ||
352 | /* split the file-backed THP */ | |
353 | write_debugfs(PATH_FMT, testfile, pgoff_start, pgoff_end); | |
354 | ||
355 | status = unlink(testfile); | |
356 | if (status) | |
357 | perror("Cannot remove testing file\n"); | |
358 | ||
359 | cleanup: | |
360 | status = umount(tmpfs_loc); | |
361 | if (status) { | |
362 | printf("Unable to umount %s\n", tmpfs_loc); | |
363 | exit(EXIT_FAILURE); | |
364 | } | |
365 | status = rmdir(tmpfs_loc); | |
366 | if (status) { | |
367 | perror("cannot remove tmp dir"); | |
368 | exit(EXIT_FAILURE); | |
369 | } | |
370 | ||
371 | printf("file-backed THP split test done, please check dmesg for more information\n"); | |
372 | } | |
373 | ||
fa6c0231 ZY |
374 | int main(int argc, char **argv) |
375 | { | |
376 | if (geteuid() != 0) { | |
377 | printf("Please run the benchmark as root\n"); | |
378 | exit(EXIT_FAILURE); | |
379 | } | |
380 | ||
381 | pagesize = getpagesize(); | |
382 | pageshift = ffs(pagesize) - 1; | |
383 | pmd_pagesize = read_pmd_pagesize(); | |
384 | ||
385 | split_pmd_thp(); | |
386 | split_pte_mapped_thp(); | |
fbe37501 | 387 | split_file_backed_thp(); |
fa6c0231 ZY |
388 | |
389 | return 0; | |
390 | } |