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 #include "qemu/host-utils.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 #ifdef __sparc__
  39     /* SPARC Linux needs greater alignment than the pagesize */
  40     return QEMU_VMALLOC_ALIGN;
  41 #endif
  42 #endif
  43
  44     return getpagesize();
  45 }
  46
  47 size_t qemu_mempath_getpagesize(const char *mem_path)
  48 {
  49 #ifdef CONFIG_LINUX
  50     struct statfs fs;
  51     int ret;
  52
  53     if (mem_path) {
  54         do {
  55             ret = statfs(mem_path, &fs);
  56         } while (ret != 0 && errno == EINTR);
  57
  58         if (ret != 0) {
  59             fprintf(stderr, "Couldn't statfs() memory path: %s\n",
  60                     strerror(errno));
  61             exit(1);
  62         }
  63
  64         if (fs.f_type == HUGETLBFS_MAGIC) {
  65             /* It's hugepage, return the huge page size */
  66             return fs.f_bsize;
  67         }
  68     }
  69 #ifdef __sparc__
  70     /* SPARC Linux needs greater alignment than the pagesize */
  71     return QEMU_VMALLOC_ALIGN;
  72 #endif
  73 #endif
  74
  75     return getpagesize();
  76 }
  77
  78 void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared)
  79 {
  80     int flags;
  81     int guardfd;
  82     size_t offset;
  83     size_t total;
  84     void *guardptr;
  85     void *ptr;
  86
  87     /*
  88      * Note: this always allocates at least one extra page of virtual address
  89      * space, even if size is already aligned.
  90      */
  91     total = size + align;
  92
  93 #if defined(__powerpc64__) && defined(__linux__)
  94     /* On ppc64 mappings in the same segment (aka slice) must share the same
  95      * page size. Since we will be re-allocating part of this segment
  96      * from the supplied fd, we should make sure to use the same page size, to
  97      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
  98      * avoid allocating backing store memory.
  99      * We do this unless we are using the system page size, in which case
 100      * anonymous memory is OK.
 101      */
 102     flags = MAP_PRIVATE;
 103     if (fd == -1 || qemu_fd_getpagesize(fd) == getpagesize()) {
 104         guardfd = -1;
 105         flags |= MAP_ANONYMOUS;
 106     } else {
 107         guardfd = fd;
 108         flags |= MAP_NORESERVE;
 109     }
 110 #else
 111     guardfd = -1;
 112     flags = MAP_PRIVATE | MAP_ANONYMOUS;
 113 #endif
 114
 115     guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
 116
 117     if (guardptr == MAP_FAILED) {
 118         return MAP_FAILED;
 119     }
 120
 121     assert(is_power_of_2(align));
 122     /* Always align to host page size */
 123     assert(align >= getpagesize());
 124
 125     flags = MAP_FIXED;
 126     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 127     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 128     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 129
 130     ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE, flags, fd, 0);
 131
 132     if (ptr == MAP_FAILED) {
 133         munmap(guardptr, total);
 134         return MAP_FAILED;
 135     }
 136
 137     if (offset > 0) {
 138         munmap(guardptr, offset);
 139     }
 140
 141     /*
 142      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 143      * a guard page guarding against potential buffer overflows.
 144      */
 145     total -= offset;
 146     if (total > size + getpagesize()) {
 147         munmap(ptr + size + getpagesize(), total - size - getpagesize());
 148     }
 149
 150     return ptr;
 151 }
 152
 153 void qemu_ram_munmap(void *ptr, size_t size)
 154 {
 155     if (ptr) {
 156         /* Unmap both the RAM block and the guard page */
 157         munmap(ptr, size + getpagesize());
 158     }
 159 }