util/mmap-alloc.c

   1 /*
   2  * Support for RAM backed by mmaped host memory.
   3  *
   4  * Copyright (c) 2015 Red Hat, Inc.
   5  *
   6  * Authors:
   7  *  Michael S. Tsirkin <mst@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or
  10  * later.  See the COPYING file in the top-level directory.
  11  */
  12 #include <qemu/mmap-alloc.h>
  13 #include <sys/types.h>
  14 #include <sys/mman.h>
  15 #include <assert.h>
  16
  17 #define HUGETLBFS_MAGIC       0x958458f6
  18
  19 #ifdef CONFIG_LINUX
  20 #include <sys/vfs.h>
  21 #endif
  22
  23 size_t qemu_fd_getpagesize(int fd)
  24 {
  25 #ifdef CONFIG_LINUX
  26     struct statfs fs;
  27     int ret;
  28
  29     if (fd != -1) {
  30         do {
  31             ret = fstatfs(fd, &fs);
  32         } while (ret != 0 && errno == EINTR);
  33
  34         if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
  35             return fs.f_bsize;
  36         }
  37     }
  38 #endif
  39
  40     return getpagesize();
  41 }
  42
  43 void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared)
  44 {
  45     /*
  46      * Note: this always allocates at least one extra page of virtual address
  47      * space, even if size is already aligned.
  48      */
  49     size_t total = size + align;
  50 #if defined(__powerpc64__) && defined(__linux__)
  51     /* On ppc64 mappings in the same segment (aka slice) must share the same
  52      * page size. Since we will be re-allocating part of this segment
  53      * from the supplied fd, we should make sure to use the same page size, to
  54      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
  55      * avoid allocating backing store memory.
  56      * We do this unless we are using the system page size, in which case
  57      * anonymous memory is OK.
  58      */
  59     int anonfd = fd == -1 || qemu_fd_getpagesize(fd) == getpagesize() ? -1 : fd;
  60     int flags = anonfd == -1 ? MAP_ANONYMOUS : MAP_NORESERVE;
  61     void *ptr = mmap(0, total, PROT_NONE, flags | MAP_PRIVATE, anonfd, 0);
  62 #else
  63     void *ptr = mmap(0, total, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
  64 #endif
  65     size_t offset = QEMU_ALIGN_UP((uintptr_t)ptr, align) - (uintptr_t)ptr;
  66     void *ptr1;
  67
  68     if (ptr == MAP_FAILED) {
  69         return MAP_FAILED;
  70     }
  71
  72     /* Make sure align is a power of 2 */
  73     assert(!(align & (align - 1)));
  74     /* Always align to host page size */
  75     assert(align >= getpagesize());
  76
  77     ptr1 = mmap(ptr + offset, size, PROT_READ | PROT_WRITE,
  78                 MAP_FIXED |
  79                 (fd == -1 ? MAP_ANONYMOUS : 0) |
  80                 (shared ? MAP_SHARED : MAP_PRIVATE),
  81                 fd, 0);
  82     if (ptr1 == MAP_FAILED) {
  83         munmap(ptr, total);
  84         return MAP_FAILED;
  85     }
  86
  87     ptr += offset;
  88     total -= offset;
  89
  90     if (offset > 0) {
  91         munmap(ptr - offset, offset);
  92     }
  93
  94     /*
  95      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
  96      * a guard page guarding against potential buffer overflows.
  97      */
  98     if (total > size + getpagesize()) {
  99         munmap(ptr + size + getpagesize(), total - size - getpagesize());
 100     }
 101
 102     return ptr;
 103 }
 104
 105 void qemu_ram_munmap(void *ptr, size_t size)
 106 {
 107     if (ptr) {
 108         /* Unmap both the RAM block and the guard page */
 109         munmap(ptr, size + getpagesize());
 110     }
 111 }