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,
  79                     size_t size,
  80                     size_t align,
  81                     bool shared,
  82                     bool is_pmem)
  83 {
  84     int flags;
  85     int guardfd;
  86     size_t offset;
  87     size_t pagesize;
  88     size_t total;
  89     void *guardptr;
  90     void *ptr;
  91
  92     /*
  93      * Note: this always allocates at least one extra page of virtual address
  94      * space, even if size is already aligned.
  95      */
  96     total = size + align;
  97
  98 #if defined(__powerpc64__) && defined(__linux__)
  99     /* On ppc64 mappings in the same segment (aka slice) must share the same
 100      * page size. Since we will be re-allocating part of this segment
 101      * from the supplied fd, we should make sure to use the same page size, to
 102      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
 103      * avoid allocating backing store memory.
 104      * We do this unless we are using the system page size, in which case
 105      * anonymous memory is OK.
 106      */
 107     flags = MAP_PRIVATE;
 108     pagesize = qemu_fd_getpagesize(fd);
 109     if (fd == -1 || pagesize == getpagesize()) {
 110         guardfd = -1;
 111         flags |= MAP_ANONYMOUS;
 112     } else {
 113         guardfd = fd;
 114         flags |= MAP_NORESERVE;
 115     }
 116 #else
 117     guardfd = -1;
 118     pagesize = getpagesize();
 119     flags = MAP_PRIVATE | MAP_ANONYMOUS;
 120 #endif
 121
 122     guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
 123
 124     if (guardptr == MAP_FAILED) {
 125         return MAP_FAILED;
 126     }
 127
 128     assert(is_power_of_2(align));
 129     /* Always align to host page size */
 130     assert(align >= pagesize);
 131
 132     flags = MAP_FIXED;
 133     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 134     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 135     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 136
 137     ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE, flags, fd, 0);
 138
 139     if (ptr == MAP_FAILED) {
 140         munmap(guardptr, total);
 141         return MAP_FAILED;
 142     }
 143
 144     if (offset > 0) {
 145         munmap(guardptr, offset);
 146     }
 147
 148     /*
 149      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 150      * a guard page guarding against potential buffer overflows.
 151      */
 152     total -= offset;
 153     if (total > size + pagesize) {
 154         munmap(ptr + size + pagesize, total - size - pagesize);
 155     }
 156
 157     return ptr;
 158 }
 159
 160 void qemu_ram_munmap(int fd, void *ptr, size_t size)
 161 {
 162     size_t pagesize;
 163
 164     if (ptr) {
 165         /* Unmap both the RAM block and the guard page */
 166 #if defined(__powerpc64__) && defined(__linux__)
 167         pagesize = qemu_fd_getpagesize(fd);
 168 #else
 169         pagesize = getpagesize();
 170 #endif
 171         munmap(ptr, size + pagesize);
 172     }
 173 }