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