1baa489faSSeongJae Park // SPDX-License-Identifier: GPL-2.0
2baa489faSSeongJae Park #include <string.h>
3baa489faSSeongJae Park #include <fcntl.h>
4c4277cb6SPeter Xu #include <dirent.h>
5c4277cb6SPeter Xu #include <sys/ioctl.h>
678391f64SPeter Xu #include <linux/userfaultfd.h>
778391f64SPeter Xu #include <sys/syscall.h>
8baa489faSSeongJae Park #include <unistd.h>
9baa489faSSeongJae Park #include "../kselftest.h"
10baa489faSSeongJae Park #include "vm_util.h"
11baa489faSSeongJae Park
12baa489faSSeongJae Park #define PMD_SIZE_FILE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
13baa489faSSeongJae Park #define SMAP_FILE_PATH "/proc/self/smaps"
14baa489faSSeongJae Park #define MAX_LINE_LENGTH 500
15af605d26SPeter Xu
16af605d26SPeter Xu unsigned int __page_size;
17af605d26SPeter Xu unsigned int __page_shift;
18baa489faSSeongJae Park
pagemap_get_entry(int fd,char * start)19baa489faSSeongJae Park uint64_t pagemap_get_entry(int fd, char *start)
20baa489faSSeongJae Park {
21baa489faSSeongJae Park const unsigned long pfn = (unsigned long)start / getpagesize();
22baa489faSSeongJae Park uint64_t entry;
23baa489faSSeongJae Park int ret;
24baa489faSSeongJae Park
25baa489faSSeongJae Park ret = pread(fd, &entry, sizeof(entry), pfn * sizeof(entry));
26baa489faSSeongJae Park if (ret != sizeof(entry))
27baa489faSSeongJae Park ksft_exit_fail_msg("reading pagemap failed\n");
28baa489faSSeongJae Park return entry;
29baa489faSSeongJae Park }
30baa489faSSeongJae Park
pagemap_is_softdirty(int fd,char * start)31baa489faSSeongJae Park bool pagemap_is_softdirty(int fd, char *start)
329f74696bSPeter Xu {
33baa489faSSeongJae Park return pagemap_get_entry(fd, start) & PM_SOFT_DIRTY;
34baa489faSSeongJae Park }
35baa489faSSeongJae Park
pagemap_is_swapped(int fd,char * start)36baa489faSSeongJae Park bool pagemap_is_swapped(int fd, char *start)
379f74696bSPeter Xu {
38baa489faSSeongJae Park return pagemap_get_entry(fd, start) & PM_SWAP;
39baa489faSSeongJae Park }
40baa489faSSeongJae Park
pagemap_is_populated(int fd,char * start)41baa489faSSeongJae Park bool pagemap_is_populated(int fd, char *start)
429f74696bSPeter Xu {
43baa489faSSeongJae Park return pagemap_get_entry(fd, start) & (PM_PRESENT | PM_SWAP);
44baa489faSSeongJae Park }
45baa489faSSeongJae Park
pagemap_get_pfn(int fd,char * start)46baa489faSSeongJae Park unsigned long pagemap_get_pfn(int fd, char *start)
47baa489faSSeongJae Park {
48baa489faSSeongJae Park uint64_t entry = pagemap_get_entry(fd, start);
49baa489faSSeongJae Park
509f74696bSPeter Xu /* If present (63th bit), PFN is at bit 0 -- 54. */
51baa489faSSeongJae Park if (entry & PM_PRESENT)
52baa489faSSeongJae Park return entry & 0x007fffffffffffffull;
53baa489faSSeongJae Park return -1ul;
54baa489faSSeongJae Park }
55baa489faSSeongJae Park
clear_softdirty(void)56baa489faSSeongJae Park void clear_softdirty(void)
57baa489faSSeongJae Park {
58baa489faSSeongJae Park int ret;
59baa489faSSeongJae Park const char *ctrl = "4";
60baa489faSSeongJae Park int fd = open("/proc/self/clear_refs", O_WRONLY);
61baa489faSSeongJae Park
62baa489faSSeongJae Park if (fd < 0)
63baa489faSSeongJae Park ksft_exit_fail_msg("opening clear_refs failed\n");
64baa489faSSeongJae Park ret = write(fd, ctrl, strlen(ctrl));
65baa489faSSeongJae Park close(fd);
66baa489faSSeongJae Park if (ret != strlen(ctrl))
67baa489faSSeongJae Park ksft_exit_fail_msg("writing clear_refs failed\n");
68baa489faSSeongJae Park }
69baa489faSSeongJae Park
check_for_pattern(FILE * fp,const char * pattern,char * buf,size_t len)70baa489faSSeongJae Park bool check_for_pattern(FILE *fp, const char *pattern, char *buf, size_t len)
71baa489faSSeongJae Park {
72baa489faSSeongJae Park while (fgets(buf, len, fp)) {
73baa489faSSeongJae Park if (!strncmp(buf, pattern, strlen(pattern)))
74baa489faSSeongJae Park return true;
75baa489faSSeongJae Park }
76baa489faSSeongJae Park return false;
77baa489faSSeongJae Park }
78baa489faSSeongJae Park
read_pmd_pagesize(void)79baa489faSSeongJae Park uint64_t read_pmd_pagesize(void)
80baa489faSSeongJae Park {
81baa489faSSeongJae Park int fd;
82baa489faSSeongJae Park char buf[20];
83baa489faSSeongJae Park ssize_t num_read;
84baa489faSSeongJae Park
85baa489faSSeongJae Park fd = open(PMD_SIZE_FILE_PATH, O_RDONLY);
86d6e61afbSDavid Hildenbrand if (fd == -1)
87baa489faSSeongJae Park return 0;
88baa489faSSeongJae Park
89baa489faSSeongJae Park num_read = read(fd, buf, 19);
90baa489faSSeongJae Park if (num_read < 1) {
91d6e61afbSDavid Hildenbrand close(fd);
92baa489faSSeongJae Park return 0;
93baa489faSSeongJae Park }
94baa489faSSeongJae Park buf[num_read] = '\0';
95baa489faSSeongJae Park close(fd);
96baa489faSSeongJae Park
97baa489faSSeongJae Park return strtoul(buf, NULL, 10);
98baa489faSSeongJae Park }
99baa489faSSeongJae Park
__check_huge(void * addr,char * pattern,int nr_hpages,uint64_t hpage_size)100baa489faSSeongJae Park bool __check_huge(void *addr, char *pattern, int nr_hpages,
101baa489faSSeongJae Park uint64_t hpage_size)
102baa489faSSeongJae Park {
103baa489faSSeongJae Park uint64_t thp = -1;
104baa489faSSeongJae Park int ret;
105baa489faSSeongJae Park FILE *fp;
106baa489faSSeongJae Park char buffer[MAX_LINE_LENGTH];
107baa489faSSeongJae Park char addr_pattern[MAX_LINE_LENGTH];
108baa489faSSeongJae Park
109baa489faSSeongJae Park ret = snprintf(addr_pattern, MAX_LINE_LENGTH, "%08lx-",
110baa489faSSeongJae Park (unsigned long) addr);
111baa489faSSeongJae Park if (ret >= MAX_LINE_LENGTH)
112baa489faSSeongJae Park ksft_exit_fail_msg("%s: Pattern is too long\n", __func__);
113baa489faSSeongJae Park
114baa489faSSeongJae Park fp = fopen(SMAP_FILE_PATH, "r");
115baa489faSSeongJae Park if (!fp)
116baa489faSSeongJae Park ksft_exit_fail_msg("%s: Failed to open file %s\n", __func__, SMAP_FILE_PATH);
117baa489faSSeongJae Park
118baa489faSSeongJae Park if (!check_for_pattern(fp, addr_pattern, buffer, sizeof(buffer)))
119baa489faSSeongJae Park goto err_out;
120baa489faSSeongJae Park
121baa489faSSeongJae Park /*
122baa489faSSeongJae Park * Fetch the pattern in the same block and check the number of
123baa489faSSeongJae Park * hugepages.
124baa489faSSeongJae Park */
125baa489faSSeongJae Park if (!check_for_pattern(fp, pattern, buffer, sizeof(buffer)))
126baa489faSSeongJae Park goto err_out;
127baa489faSSeongJae Park
128baa489faSSeongJae Park snprintf(addr_pattern, MAX_LINE_LENGTH, "%s%%9ld kB", pattern);
129baa489faSSeongJae Park
130baa489faSSeongJae Park if (sscanf(buffer, addr_pattern, &thp) != 1)
131baa489faSSeongJae Park ksft_exit_fail_msg("Reading smap error\n");
132baa489faSSeongJae Park
133baa489faSSeongJae Park err_out:
134baa489faSSeongJae Park fclose(fp);
135baa489faSSeongJae Park return thp == (nr_hpages * (hpage_size >> 10));
136baa489faSSeongJae Park }
137baa489faSSeongJae Park
check_huge_anon(void * addr,int nr_hpages,uint64_t hpage_size)138baa489faSSeongJae Park bool check_huge_anon(void *addr, int nr_hpages, uint64_t hpage_size)
139baa489faSSeongJae Park {
140baa489faSSeongJae Park return __check_huge(addr, "AnonHugePages: ", nr_hpages, hpage_size);
141baa489faSSeongJae Park }
142baa489faSSeongJae Park
check_huge_file(void * addr,int nr_hpages,uint64_t hpage_size)143baa489faSSeongJae Park bool check_huge_file(void *addr, int nr_hpages, uint64_t hpage_size)
144baa489faSSeongJae Park {
145baa489faSSeongJae Park return __check_huge(addr, "FilePmdMapped:", nr_hpages, hpage_size);
146baa489faSSeongJae Park }
147baa489faSSeongJae Park
check_huge_shmem(void * addr,int nr_hpages,uint64_t hpage_size)148baa489faSSeongJae Park bool check_huge_shmem(void *addr, int nr_hpages, uint64_t hpage_size)
149baa489faSSeongJae Park {
150baa489faSSeongJae Park return __check_huge(addr, "ShmemPmdMapped:", nr_hpages, hpage_size);
151af605d26SPeter Xu }
152af605d26SPeter Xu
allocate_transhuge(void * ptr,int pagemap_fd)153af605d26SPeter Xu int64_t allocate_transhuge(void *ptr, int pagemap_fd)
154af605d26SPeter Xu {
155af605d26SPeter Xu uint64_t ent[2];
156af605d26SPeter Xu
157af605d26SPeter Xu /* drop pmd */
158af605d26SPeter Xu if (mmap(ptr, HPAGE_SIZE, PROT_READ | PROT_WRITE,
159af605d26SPeter Xu MAP_FIXED | MAP_ANONYMOUS |
160af605d26SPeter Xu MAP_NORESERVE | MAP_PRIVATE, -1, 0) != ptr)
161af605d26SPeter Xu errx(2, "mmap transhuge");
162af605d26SPeter Xu
163af605d26SPeter Xu if (madvise(ptr, HPAGE_SIZE, MADV_HUGEPAGE))
164af605d26SPeter Xu err(2, "MADV_HUGEPAGE");
165af605d26SPeter Xu
166af605d26SPeter Xu /* allocate transparent huge page */
167af605d26SPeter Xu *(volatile void **)ptr = ptr;
168af605d26SPeter Xu
169af605d26SPeter Xu if (pread(pagemap_fd, ent, sizeof(ent),
170af605d26SPeter Xu (uintptr_t)ptr >> (pshift() - 3)) != sizeof(ent))
171af605d26SPeter Xu err(2, "read pagemap");
172af605d26SPeter Xu
173af605d26SPeter Xu if (PAGEMAP_PRESENT(ent[0]) && PAGEMAP_PRESENT(ent[1]) &&
174af605d26SPeter Xu PAGEMAP_PFN(ent[0]) + 1 == PAGEMAP_PFN(ent[1]) &&
175af605d26SPeter Xu !(PAGEMAP_PFN(ent[0]) & ((1 << (HPAGE_SHIFT - pshift())) - 1)))
176af605d26SPeter Xu return PAGEMAP_PFN(ent[0]);
177af605d26SPeter Xu
178af605d26SPeter Xu return -1;
179bd4d67e7SPeter Xu }
180bd4d67e7SPeter Xu
default_huge_page_size(void)181bd4d67e7SPeter Xu unsigned long default_huge_page_size(void)
182bd4d67e7SPeter Xu {
183bd4d67e7SPeter Xu unsigned long hps = 0;
184bd4d67e7SPeter Xu char *line = NULL;
185bd4d67e7SPeter Xu size_t linelen = 0;
186bd4d67e7SPeter Xu FILE *f = fopen("/proc/meminfo", "r");
187bd4d67e7SPeter Xu
188bd4d67e7SPeter Xu if (!f)
189bd4d67e7SPeter Xu return 0;
190bd4d67e7SPeter Xu while (getline(&line, &linelen, f) > 0) {
191bd4d67e7SPeter Xu if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) {
192bd4d67e7SPeter Xu hps <<= 10;
193bd4d67e7SPeter Xu break;
194bd4d67e7SPeter Xu }
195bd4d67e7SPeter Xu }
196bd4d67e7SPeter Xu
197bd4d67e7SPeter Xu free(line);
198bd4d67e7SPeter Xu fclose(f);
199bd4d67e7SPeter Xu return hps;
200c4277cb6SPeter Xu }
201*c3315502SPeter Xu
detect_hugetlb_page_sizes(size_t sizes[],int max)202*c3315502SPeter Xu int detect_hugetlb_page_sizes(size_t sizes[], int max)
203*c3315502SPeter Xu {
204c4277cb6SPeter Xu DIR *dir = opendir("/sys/kernel/mm/hugepages/");
205c4277cb6SPeter Xu int count = 0;
206c4277cb6SPeter Xu
207c4277cb6SPeter Xu if (!dir)
208c4277cb6SPeter Xu return 0;
209c4277cb6SPeter Xu
210c4277cb6SPeter Xu while (count < max) {
211c4277cb6SPeter Xu struct dirent *entry = readdir(dir);
212c4277cb6SPeter Xu size_t kb;
213c4277cb6SPeter Xu
214c4277cb6SPeter Xu if (!entry)
215c4277cb6SPeter Xu break;
216c4277cb6SPeter Xu if (entry->d_type != DT_DIR)
217c4277cb6SPeter Xu continue;
218c4277cb6SPeter Xu if (sscanf(entry->d_name, "hugepages-%zukB", &kb) != 1)
219c4277cb6SPeter Xu continue;
220c4277cb6SPeter Xu sizes[count++] = kb * 1024;
221c4277cb6SPeter Xu ksft_print_msg("[INFO] detected hugetlb page size: %zu KiB\n",
222*c3315502SPeter Xu kb);
223*c3315502SPeter Xu }
224c4277cb6SPeter Xu closedir(dir);
225c4277cb6SPeter Xu return count;
226c4277cb6SPeter Xu }
227c4277cb6SPeter Xu
228*c3315502SPeter Xu /* If `ioctls' non-NULL, the allowed ioctls will be returned into the var */
uffd_register_with_ioctls(int uffd,void * addr,uint64_t len,bool miss,bool wp,bool minor,uint64_t * ioctls)229*c3315502SPeter Xu int uffd_register_with_ioctls(int uffd, void *addr, uint64_t len,
230*c3315502SPeter Xu bool miss, bool wp, bool minor, uint64_t *ioctls)
231*c3315502SPeter Xu {
232*c3315502SPeter Xu struct uffdio_register uffdio_register = { 0 };
233*c3315502SPeter Xu uint64_t mode = 0;
234*c3315502SPeter Xu int ret = 0;
235c4277cb6SPeter Xu
236c4277cb6SPeter Xu if (miss)
237c4277cb6SPeter Xu mode |= UFFDIO_REGISTER_MODE_MISSING;
238c4277cb6SPeter Xu if (wp)
239c4277cb6SPeter Xu mode |= UFFDIO_REGISTER_MODE_WP;
240c4277cb6SPeter Xu if (minor)
241c4277cb6SPeter Xu mode |= UFFDIO_REGISTER_MODE_MINOR;
242c4277cb6SPeter Xu
243c4277cb6SPeter Xu uffdio_register.range.start = (unsigned long)addr;
244c4277cb6SPeter Xu uffdio_register.range.len = len;
24578391f64SPeter Xu uffdio_register.mode = mode;
24678391f64SPeter Xu
24778391f64SPeter Xu if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register) == -1)
24878391f64SPeter Xu ret = -errno;
24978391f64SPeter Xu else if (ioctls)
25078391f64SPeter Xu *ioctls = uffdio_register.ioctls;
25178391f64SPeter Xu
25278391f64SPeter Xu return ret;
25378391f64SPeter Xu }
25478391f64SPeter Xu
uffd_register(int uffd,void * addr,uint64_t len,bool miss,bool wp,bool minor)25578391f64SPeter Xu int uffd_register(int uffd, void *addr, uint64_t len,
25678391f64SPeter Xu bool miss, bool wp, bool minor)
25778391f64SPeter Xu {
25878391f64SPeter Xu return uffd_register_with_ioctls(uffd, addr, len,
25978391f64SPeter Xu miss, wp, minor, NULL);
26078391f64SPeter Xu }
26178391f64SPeter Xu
uffd_unregister(int uffd,void * addr,uint64_t len)26278391f64SPeter Xu int uffd_unregister(int uffd, void *addr, uint64_t len)
26378391f64SPeter Xu {
26478391f64SPeter Xu struct uffdio_range range = { .start = (uintptr_t)addr, .len = len };
26578391f64SPeter Xu int ret = 0;
26678391f64SPeter Xu
26716a45b57SPeter Xu if (ioctl(uffd, UFFDIO_UNREGISTER, &range) == -1)
26816a45b57SPeter Xu ret = -errno;
26916a45b57SPeter Xu
27016a45b57SPeter Xu return ret;
27116a45b57SPeter Xu }
27216a45b57SPeter Xu