162306a36Sopenharmony_ci// SPDX-License-Identifier: GPL-2.0
262306a36Sopenharmony_ci#include <string.h>
362306a36Sopenharmony_ci#include <fcntl.h>
462306a36Sopenharmony_ci#include <dirent.h>
562306a36Sopenharmony_ci#include <sys/ioctl.h>
662306a36Sopenharmony_ci#include <linux/userfaultfd.h>
762306a36Sopenharmony_ci#include <sys/syscall.h>
862306a36Sopenharmony_ci#include <unistd.h>
962306a36Sopenharmony_ci#include "../kselftest.h"
1062306a36Sopenharmony_ci#include "vm_util.h"
1162306a36Sopenharmony_ci
1262306a36Sopenharmony_ci#define PMD_SIZE_FILE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
1362306a36Sopenharmony_ci#define SMAP_FILE_PATH "/proc/self/smaps"
1462306a36Sopenharmony_ci#define MAX_LINE_LENGTH 500
1562306a36Sopenharmony_ci
1662306a36Sopenharmony_ciunsigned int __page_size;
1762306a36Sopenharmony_ciunsigned int __page_shift;
1862306a36Sopenharmony_ci
1962306a36Sopenharmony_ciuint64_t pagemap_get_entry(int fd, char *start)
2062306a36Sopenharmony_ci{
2162306a36Sopenharmony_ci	const unsigned long pfn = (unsigned long)start / getpagesize();
2262306a36Sopenharmony_ci	uint64_t entry;
2362306a36Sopenharmony_ci	int ret;
2462306a36Sopenharmony_ci
2562306a36Sopenharmony_ci	ret = pread(fd, &entry, sizeof(entry), pfn * sizeof(entry));
2662306a36Sopenharmony_ci	if (ret != sizeof(entry))
2762306a36Sopenharmony_ci		ksft_exit_fail_msg("reading pagemap failed\n");
2862306a36Sopenharmony_ci	return entry;
2962306a36Sopenharmony_ci}
3062306a36Sopenharmony_ci
3162306a36Sopenharmony_cibool pagemap_is_softdirty(int fd, char *start)
3262306a36Sopenharmony_ci{
3362306a36Sopenharmony_ci	return pagemap_get_entry(fd, start) & PM_SOFT_DIRTY;
3462306a36Sopenharmony_ci}
3562306a36Sopenharmony_ci
3662306a36Sopenharmony_cibool pagemap_is_swapped(int fd, char *start)
3762306a36Sopenharmony_ci{
3862306a36Sopenharmony_ci	return pagemap_get_entry(fd, start) & PM_SWAP;
3962306a36Sopenharmony_ci}
4062306a36Sopenharmony_ci
4162306a36Sopenharmony_cibool pagemap_is_populated(int fd, char *start)
4262306a36Sopenharmony_ci{
4362306a36Sopenharmony_ci	return pagemap_get_entry(fd, start) & (PM_PRESENT | PM_SWAP);
4462306a36Sopenharmony_ci}
4562306a36Sopenharmony_ci
4662306a36Sopenharmony_ciunsigned long pagemap_get_pfn(int fd, char *start)
4762306a36Sopenharmony_ci{
4862306a36Sopenharmony_ci	uint64_t entry = pagemap_get_entry(fd, start);
4962306a36Sopenharmony_ci
5062306a36Sopenharmony_ci	/* If present (63th bit), PFN is at bit 0 -- 54. */
5162306a36Sopenharmony_ci	if (entry & PM_PRESENT)
5262306a36Sopenharmony_ci		return entry & 0x007fffffffffffffull;
5362306a36Sopenharmony_ci	return -1ul;
5462306a36Sopenharmony_ci}
5562306a36Sopenharmony_ci
5662306a36Sopenharmony_civoid clear_softdirty(void)
5762306a36Sopenharmony_ci{
5862306a36Sopenharmony_ci	int ret;
5962306a36Sopenharmony_ci	const char *ctrl = "4";
6062306a36Sopenharmony_ci	int fd = open("/proc/self/clear_refs", O_WRONLY);
6162306a36Sopenharmony_ci
6262306a36Sopenharmony_ci	if (fd < 0)
6362306a36Sopenharmony_ci		ksft_exit_fail_msg("opening clear_refs failed\n");
6462306a36Sopenharmony_ci	ret = write(fd, ctrl, strlen(ctrl));
6562306a36Sopenharmony_ci	close(fd);
6662306a36Sopenharmony_ci	if (ret != strlen(ctrl))
6762306a36Sopenharmony_ci		ksft_exit_fail_msg("writing clear_refs failed\n");
6862306a36Sopenharmony_ci}
6962306a36Sopenharmony_ci
7062306a36Sopenharmony_cibool check_for_pattern(FILE *fp, const char *pattern, char *buf, size_t len)
7162306a36Sopenharmony_ci{
7262306a36Sopenharmony_ci	while (fgets(buf, len, fp)) {
7362306a36Sopenharmony_ci		if (!strncmp(buf, pattern, strlen(pattern)))
7462306a36Sopenharmony_ci			return true;
7562306a36Sopenharmony_ci	}
7662306a36Sopenharmony_ci	return false;
7762306a36Sopenharmony_ci}
7862306a36Sopenharmony_ci
7962306a36Sopenharmony_ciuint64_t read_pmd_pagesize(void)
8062306a36Sopenharmony_ci{
8162306a36Sopenharmony_ci	int fd;
8262306a36Sopenharmony_ci	char buf[20];
8362306a36Sopenharmony_ci	ssize_t num_read;
8462306a36Sopenharmony_ci
8562306a36Sopenharmony_ci	fd = open(PMD_SIZE_FILE_PATH, O_RDONLY);
8662306a36Sopenharmony_ci	if (fd == -1)
8762306a36Sopenharmony_ci		return 0;
8862306a36Sopenharmony_ci
8962306a36Sopenharmony_ci	num_read = read(fd, buf, 19);
9062306a36Sopenharmony_ci	if (num_read < 1) {
9162306a36Sopenharmony_ci		close(fd);
9262306a36Sopenharmony_ci		return 0;
9362306a36Sopenharmony_ci	}
9462306a36Sopenharmony_ci	buf[num_read] = '\0';
9562306a36Sopenharmony_ci	close(fd);
9662306a36Sopenharmony_ci
9762306a36Sopenharmony_ci	return strtoul(buf, NULL, 10);
9862306a36Sopenharmony_ci}
9962306a36Sopenharmony_ci
10062306a36Sopenharmony_cibool __check_huge(void *addr, char *pattern, int nr_hpages,
10162306a36Sopenharmony_ci		  uint64_t hpage_size)
10262306a36Sopenharmony_ci{
10362306a36Sopenharmony_ci	uint64_t thp = -1;
10462306a36Sopenharmony_ci	int ret;
10562306a36Sopenharmony_ci	FILE *fp;
10662306a36Sopenharmony_ci	char buffer[MAX_LINE_LENGTH];
10762306a36Sopenharmony_ci	char addr_pattern[MAX_LINE_LENGTH];
10862306a36Sopenharmony_ci
10962306a36Sopenharmony_ci	ret = snprintf(addr_pattern, MAX_LINE_LENGTH, "%08lx-",
11062306a36Sopenharmony_ci		       (unsigned long) addr);
11162306a36Sopenharmony_ci	if (ret >= MAX_LINE_LENGTH)
11262306a36Sopenharmony_ci		ksft_exit_fail_msg("%s: Pattern is too long\n", __func__);
11362306a36Sopenharmony_ci
11462306a36Sopenharmony_ci	fp = fopen(SMAP_FILE_PATH, "r");
11562306a36Sopenharmony_ci	if (!fp)
11662306a36Sopenharmony_ci		ksft_exit_fail_msg("%s: Failed to open file %s\n", __func__, SMAP_FILE_PATH);
11762306a36Sopenharmony_ci
11862306a36Sopenharmony_ci	if (!check_for_pattern(fp, addr_pattern, buffer, sizeof(buffer)))
11962306a36Sopenharmony_ci		goto err_out;
12062306a36Sopenharmony_ci
12162306a36Sopenharmony_ci	/*
12262306a36Sopenharmony_ci	 * Fetch the pattern in the same block and check the number of
12362306a36Sopenharmony_ci	 * hugepages.
12462306a36Sopenharmony_ci	 */
12562306a36Sopenharmony_ci	if (!check_for_pattern(fp, pattern, buffer, sizeof(buffer)))
12662306a36Sopenharmony_ci		goto err_out;
12762306a36Sopenharmony_ci
12862306a36Sopenharmony_ci	snprintf(addr_pattern, MAX_LINE_LENGTH, "%s%%9ld kB", pattern);
12962306a36Sopenharmony_ci
13062306a36Sopenharmony_ci	if (sscanf(buffer, addr_pattern, &thp) != 1)
13162306a36Sopenharmony_ci		ksft_exit_fail_msg("Reading smap error\n");
13262306a36Sopenharmony_ci
13362306a36Sopenharmony_cierr_out:
13462306a36Sopenharmony_ci	fclose(fp);
13562306a36Sopenharmony_ci	return thp == (nr_hpages * (hpage_size >> 10));
13662306a36Sopenharmony_ci}
13762306a36Sopenharmony_ci
13862306a36Sopenharmony_cibool check_huge_anon(void *addr, int nr_hpages, uint64_t hpage_size)
13962306a36Sopenharmony_ci{
14062306a36Sopenharmony_ci	return __check_huge(addr, "AnonHugePages: ", nr_hpages, hpage_size);
14162306a36Sopenharmony_ci}
14262306a36Sopenharmony_ci
14362306a36Sopenharmony_cibool check_huge_file(void *addr, int nr_hpages, uint64_t hpage_size)
14462306a36Sopenharmony_ci{
14562306a36Sopenharmony_ci	return __check_huge(addr, "FilePmdMapped:", nr_hpages, hpage_size);
14662306a36Sopenharmony_ci}
14762306a36Sopenharmony_ci
14862306a36Sopenharmony_cibool check_huge_shmem(void *addr, int nr_hpages, uint64_t hpage_size)
14962306a36Sopenharmony_ci{
15062306a36Sopenharmony_ci	return __check_huge(addr, "ShmemPmdMapped:", nr_hpages, hpage_size);
15162306a36Sopenharmony_ci}
15262306a36Sopenharmony_ci
15362306a36Sopenharmony_ciint64_t allocate_transhuge(void *ptr, int pagemap_fd)
15462306a36Sopenharmony_ci{
15562306a36Sopenharmony_ci	uint64_t ent[2];
15662306a36Sopenharmony_ci
15762306a36Sopenharmony_ci	/* drop pmd */
15862306a36Sopenharmony_ci	if (mmap(ptr, HPAGE_SIZE, PROT_READ | PROT_WRITE,
15962306a36Sopenharmony_ci		 MAP_FIXED | MAP_ANONYMOUS |
16062306a36Sopenharmony_ci		 MAP_NORESERVE | MAP_PRIVATE, -1, 0) != ptr)
16162306a36Sopenharmony_ci		errx(2, "mmap transhuge");
16262306a36Sopenharmony_ci
16362306a36Sopenharmony_ci	if (madvise(ptr, HPAGE_SIZE, MADV_HUGEPAGE))
16462306a36Sopenharmony_ci		err(2, "MADV_HUGEPAGE");
16562306a36Sopenharmony_ci
16662306a36Sopenharmony_ci	/* allocate transparent huge page */
16762306a36Sopenharmony_ci	*(volatile void **)ptr = ptr;
16862306a36Sopenharmony_ci
16962306a36Sopenharmony_ci	if (pread(pagemap_fd, ent, sizeof(ent),
17062306a36Sopenharmony_ci		  (uintptr_t)ptr >> (pshift() - 3)) != sizeof(ent))
17162306a36Sopenharmony_ci		err(2, "read pagemap");
17262306a36Sopenharmony_ci
17362306a36Sopenharmony_ci	if (PAGEMAP_PRESENT(ent[0]) && PAGEMAP_PRESENT(ent[1]) &&
17462306a36Sopenharmony_ci	    PAGEMAP_PFN(ent[0]) + 1 == PAGEMAP_PFN(ent[1]) &&
17562306a36Sopenharmony_ci	    !(PAGEMAP_PFN(ent[0]) & ((1 << (HPAGE_SHIFT - pshift())) - 1)))
17662306a36Sopenharmony_ci		return PAGEMAP_PFN(ent[0]);
17762306a36Sopenharmony_ci
17862306a36Sopenharmony_ci	return -1;
17962306a36Sopenharmony_ci}
18062306a36Sopenharmony_ci
18162306a36Sopenharmony_ciunsigned long default_huge_page_size(void)
18262306a36Sopenharmony_ci{
18362306a36Sopenharmony_ci	unsigned long hps = 0;
18462306a36Sopenharmony_ci	char *line = NULL;
18562306a36Sopenharmony_ci	size_t linelen = 0;
18662306a36Sopenharmony_ci	FILE *f = fopen("/proc/meminfo", "r");
18762306a36Sopenharmony_ci
18862306a36Sopenharmony_ci	if (!f)
18962306a36Sopenharmony_ci		return 0;
19062306a36Sopenharmony_ci	while (getline(&line, &linelen, f) > 0) {
19162306a36Sopenharmony_ci		if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
19262306a36Sopenharmony_ci			hps <<= 10;
19362306a36Sopenharmony_ci			break;
19462306a36Sopenharmony_ci		}
19562306a36Sopenharmony_ci	}
19662306a36Sopenharmony_ci
19762306a36Sopenharmony_ci	free(line);
19862306a36Sopenharmony_ci	fclose(f);
19962306a36Sopenharmony_ci	return hps;
20062306a36Sopenharmony_ci}
20162306a36Sopenharmony_ci
20262306a36Sopenharmony_ciint detect_hugetlb_page_sizes(size_t sizes[], int max)
20362306a36Sopenharmony_ci{
20462306a36Sopenharmony_ci	DIR *dir = opendir("/sys/kernel/mm/hugepages/");
20562306a36Sopenharmony_ci	int count = 0;
20662306a36Sopenharmony_ci
20762306a36Sopenharmony_ci	if (!dir)
20862306a36Sopenharmony_ci		return 0;
20962306a36Sopenharmony_ci
21062306a36Sopenharmony_ci	while (count < max) {
21162306a36Sopenharmony_ci		struct dirent *entry = readdir(dir);
21262306a36Sopenharmony_ci		size_t kb;
21362306a36Sopenharmony_ci
21462306a36Sopenharmony_ci		if (!entry)
21562306a36Sopenharmony_ci			break;
21662306a36Sopenharmony_ci		if (entry->d_type != DT_DIR)
21762306a36Sopenharmony_ci			continue;
21862306a36Sopenharmony_ci		if (sscanf(entry->d_name, "hugepages-%zukB", &kb) != 1)
21962306a36Sopenharmony_ci			continue;
22062306a36Sopenharmony_ci		sizes[count++] = kb * 1024;
22162306a36Sopenharmony_ci		ksft_print_msg("[INFO] detected hugetlb page size: %zu KiB\n",
22262306a36Sopenharmony_ci			       kb);
22362306a36Sopenharmony_ci	}
22462306a36Sopenharmony_ci	closedir(dir);
22562306a36Sopenharmony_ci	return count;
22662306a36Sopenharmony_ci}
22762306a36Sopenharmony_ci
22862306a36Sopenharmony_ci/* If `ioctls' non-NULL, the allowed ioctls will be returned into the var */
22962306a36Sopenharmony_ciint uffd_register_with_ioctls(int uffd, void *addr, uint64_t len,
23062306a36Sopenharmony_ci			      bool miss, bool wp, bool minor, uint64_t *ioctls)
23162306a36Sopenharmony_ci{
23262306a36Sopenharmony_ci	struct uffdio_register uffdio_register = { 0 };
23362306a36Sopenharmony_ci	uint64_t mode = 0;
23462306a36Sopenharmony_ci	int ret = 0;
23562306a36Sopenharmony_ci
23662306a36Sopenharmony_ci	if (miss)
23762306a36Sopenharmony_ci		mode |= UFFDIO_REGISTER_MODE_MISSING;
23862306a36Sopenharmony_ci	if (wp)
23962306a36Sopenharmony_ci		mode |= UFFDIO_REGISTER_MODE_WP;
24062306a36Sopenharmony_ci	if (minor)
24162306a36Sopenharmony_ci		mode |= UFFDIO_REGISTER_MODE_MINOR;
24262306a36Sopenharmony_ci
24362306a36Sopenharmony_ci	uffdio_register.range.start = (unsigned long)addr;
24462306a36Sopenharmony_ci	uffdio_register.range.len = len;
24562306a36Sopenharmony_ci	uffdio_register.mode = mode;
24662306a36Sopenharmony_ci
24762306a36Sopenharmony_ci	if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register) == -1)
24862306a36Sopenharmony_ci		ret = -errno;
24962306a36Sopenharmony_ci	else if (ioctls)
25062306a36Sopenharmony_ci		*ioctls = uffdio_register.ioctls;
25162306a36Sopenharmony_ci
25262306a36Sopenharmony_ci	return ret;
25362306a36Sopenharmony_ci}
25462306a36Sopenharmony_ci
25562306a36Sopenharmony_ciint uffd_register(int uffd, void *addr, uint64_t len,
25662306a36Sopenharmony_ci		  bool miss, bool wp, bool minor)
25762306a36Sopenharmony_ci{
25862306a36Sopenharmony_ci	return uffd_register_with_ioctls(uffd, addr, len,
25962306a36Sopenharmony_ci					 miss, wp, minor, NULL);
26062306a36Sopenharmony_ci}
26162306a36Sopenharmony_ci
26262306a36Sopenharmony_ciint uffd_unregister(int uffd, void *addr, uint64_t len)
26362306a36Sopenharmony_ci{
26462306a36Sopenharmony_ci	struct uffdio_range range = { .start = (uintptr_t)addr, .len = len };
26562306a36Sopenharmony_ci	int ret = 0;
26662306a36Sopenharmony_ci
26762306a36Sopenharmony_ci	if (ioctl(uffd, UFFDIO_UNREGISTER, &range) == -1)
26862306a36Sopenharmony_ci		ret = -errno;
26962306a36Sopenharmony_ci
27062306a36Sopenharmony_ci	return ret;
27162306a36Sopenharmony_ci}
272