Quelle  hugetlb-madvise.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * hugepage-madvise:
 *
 * Basic functional testing of madvise MADV_DONTNEED and MADV_REMOVE
 * on hugetlb mappings.
 *
 * Before running this test, make sure the administrator has pre-allocated
 * at least MIN_FREE_PAGES hugetlb pages and they are free.  In addition,
 * the test takes an argument that is the path to a file in a hugetlbfs
 * filesystem.  Therefore, a hugetlbfs filesystem must be mounted on some
 * directory.
 */

#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/mman.h>
#include <fcntl.h>
#include "vm_util.h"
#include "../kselftest.h"

#define MIN_FREE_PAGES 20
#define NR_HUGE_PAGES 10 /* common number of pages to map/allocate */

#define validate_free_pages(exp_free)     \
 do {        \
  int fhp = get_free_hugepages();    \
  if (fhp != (exp_free)) {    \
   printf("Unexpected number of free huge " \
    "pages line %d\n", __LINE__);  \
   exit(1);     \
  }       \
 } while (0)

unsigned long huge_page_size;
unsigned long base_page_size;

void write_fault_pages(void *addr, unsigned long nr_pages)
{
 unsigned long i;

 for (i = 0; i < nr_pages; i++)
  *((unsigned long *)(addr + (i * huge_page_size))) = i;
}

void read_fault_pages(void *addr, unsigned long nr_pages)
{
 unsigned long i;

 for (i = 0; i < nr_pages; i++) {
  unsigned long *addr2 =
   ((unsigned long *)(addr + (i * huge_page_size)));
  /* Prevent the compiler from optimizing out the entire loop: */
  FORCE_READ(*addr2);
 }
}

int main(int argc, char **argv)
{
 unsigned long free_hugepages;
 void *addr, *addr2;
 int fd;
 int ret;

 huge_page_size = default_huge_page_size();
 if (!huge_page_size) {
  printf("Unable to determine huge page size, exiting!\n");
  exit(1);
 }
 base_page_size = sysconf(_SC_PAGE_SIZE);
 if (!huge_page_size) {
  printf("Unable to determine base page size, exiting!\n");
  exit(1);
 }

 free_hugepages = get_free_hugepages();
 if (free_hugepages < MIN_FREE_PAGES) {
  printf("Not enough free huge pages to test, exiting!\n");
  exit(KSFT_SKIP);
 }

 fd = memfd_create(argv[0], MFD_HUGETLB);
 if (fd < 0) {
  perror("memfd_create() failed");
  exit(1);
 }

 /*
 * Test validity of MADV_DONTNEED addr and length arguments.  mmap
 * size is NR_HUGE_PAGES + 2.  One page at the beginning and end of
 * the mapping will be unmapped so we KNOW there is nothing mapped
 * there.
 */
 addr = mmap(NULL, (NR_HUGE_PAGES + 2) * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
   -1, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }
 if (munmap(addr, huge_page_size) ||
   munmap(addr + (NR_HUGE_PAGES + 1) * huge_page_size,
    huge_page_size)) {
  perror("munmap");
  exit(1);
 }
 addr = addr + huge_page_size;

 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* addr before mapping should fail */
 ret = madvise(addr - base_page_size, NR_HUGE_PAGES * huge_page_size,
  MADV_DONTNEED);
 if (!ret) {
  printf("Unexpected success of madvise call with invalid addr line %d\n",
    __LINE__);
   exit(1);
 }

 /* addr + length after mapping should fail */
 ret = madvise(addr, (NR_HUGE_PAGES * huge_page_size) + base_page_size,
  MADV_DONTNEED);
 if (!ret) {
  printf("Unexpected success of madvise call with invalid length line %d\n",
    __LINE__);
   exit(1);
 }

 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);

 /*
 * Test alignment of MADV_DONTNEED addr and length arguments
 */
 addr = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
   -1, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* addr is not huge page size aligned and should fail */
 ret = madvise(addr + base_page_size,
   NR_HUGE_PAGES * huge_page_size - base_page_size,
   MADV_DONTNEED);
 if (!ret) {
  printf("Unexpected success of madvise call with unaligned start address %d\n",
    __LINE__);
   exit(1);
 }

 /* addr + length should be aligned down to huge page size */
 if (madvise(addr,
   ((NR_HUGE_PAGES - 1) * huge_page_size) + base_page_size,
   MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }

 /* should free all but last page in mapping */
 validate_free_pages(free_hugepages - 1);

 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);
 validate_free_pages(free_hugepages);

 /*
 * Test MADV_DONTNEED on anonymous private mapping
 */
 addr = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
   -1, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }

 /* should free all pages in mapping */
 validate_free_pages(free_hugepages);

 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);

 /*
 * Test MADV_DONTNEED on private mapping of hugetlb file
 */
 if (fallocate(fd, 0, 0, NR_HUGE_PAGES * huge_page_size)) {
  perror("fallocate");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 addr = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_PRIVATE, fd, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }

 /* read should not consume any pages */
 read_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* madvise should not free any pages */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* writes should allocate private pages */
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - (2 * NR_HUGE_PAGES));

 /* madvise should free private pages */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* writes should allocate private pages */
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - (2 * NR_HUGE_PAGES));

 /*
 * The fallocate below certainly should free the pages associated
 * with the file.  However, pages in the private mapping are also
 * freed.  This is not the 'correct' behavior, but is expected
 * because this is how it has worked since the initial hugetlb
 * implementation.
 */
 if (fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
     0, NR_HUGE_PAGES * huge_page_size)) {
  perror("fallocate");
  exit(1);
 }
 validate_free_pages(free_hugepages);

 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);

 /*
 * Test MADV_DONTNEED on shared mapping of hugetlb file
 */
 if (fallocate(fd, 0, 0, NR_HUGE_PAGES * huge_page_size)) {
  perror("fallocate");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 addr = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_SHARED, fd, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }

 /* write should not consume any pages */
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* madvise should not free any pages */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /*
 * Test MADV_REMOVE on shared mapping of hugetlb file
 *
 * madvise is same as hole punch and should free all pages.
 */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_REMOVE)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages);
 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);

 /*
 * Test MADV_REMOVE on shared and private mapping of hugetlb file
 */
 if (fallocate(fd, 0, 0, NR_HUGE_PAGES * huge_page_size)) {
  perror("fallocate");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 addr = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_SHARED, fd, 0);
 if (addr == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }

 /* shared write should not consume any additional pages */
 write_fault_pages(addr, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 addr2 = mmap(NULL, NR_HUGE_PAGES * huge_page_size,
   PROT_READ | PROT_WRITE,
   MAP_PRIVATE, fd, 0);
 if (addr2 == MAP_FAILED) {
  perror("mmap");
  exit(1);
 }

 /* private read should not consume any pages */
 read_fault_pages(addr2, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* private write should consume additional pages */
 write_fault_pages(addr2, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - (2 * NR_HUGE_PAGES));

 /* madvise of shared mapping should not free any pages */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages - (2 * NR_HUGE_PAGES));

 /* madvise of private mapping should free private pages */
 if (madvise(addr2, NR_HUGE_PAGES * huge_page_size, MADV_DONTNEED)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages - NR_HUGE_PAGES);

 /* private write should consume additional pages again */
 write_fault_pages(addr2, NR_HUGE_PAGES);
 validate_free_pages(free_hugepages - (2 * NR_HUGE_PAGES));

 /*
 * madvise should free both file and private pages although this is
 * not correct.  private pages should not be freed, but this is
 * expected.  See comment associated with FALLOC_FL_PUNCH_HOLE call.
 */
 if (madvise(addr, NR_HUGE_PAGES * huge_page_size, MADV_REMOVE)) {
  perror("madvise");
  exit(1);
 }
 validate_free_pages(free_hugepages);

 (void)munmap(addr, NR_HUGE_PAGES * huge_page_size);
 (void)munmap(addr2, NR_HUGE_PAGES * huge_page_size);

 close(fd);
 return 0;
}