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 #ifdef CONFIG_LINUX
  14 #include <linux/mman.h>
  15 #else  /* !CONFIG_LINUX */
  16 #define MAP_SYNC              0x0
  17 #define MAP_SHARED_VALIDATE   0x0
  18 #endif /* CONFIG_LINUX */
  19
  20 #include "qemu/osdep.h"
  21 #include "qemu/mmap-alloc.h"
  22 #include "qemu/host-utils.h"
  23
  24 #define HUGETLBFS_MAGIC       0x958458f6
  25
  26 #ifdef CONFIG_LINUX
  27 #include <sys/vfs.h>
  28 #endif
  29
  30 size_t qemu_fd_getpagesize(int fd)
  31 {
  32 #ifdef CONFIG_LINUX
  33     struct statfs fs;
  34     int ret;
  35
  36     if (fd != -1) {
  37         do {
  38             ret = fstatfs(fd, &fs);
  39         } while (ret != 0 && errno == EINTR);
  40
  41         if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
  42             return fs.f_bsize;
  43         }
  44     }
  45 #ifdef __sparc__
  46     /* SPARC Linux needs greater alignment than the pagesize */
  47     return QEMU_VMALLOC_ALIGN;
  48 #endif
  49 #endif
  50
  51     return qemu_real_host_page_size;
  52 }
  53
  54 size_t qemu_mempath_getpagesize(const char *mem_path)
  55 {
  56 #ifdef CONFIG_LINUX
  57     struct statfs fs;
  58     int ret;
  59
  60     if (mem_path) {
  61         do {
  62             ret = statfs(mem_path, &fs);
  63         } while (ret != 0 && errno == EINTR);
  64
  65         if (ret != 0) {
  66             fprintf(stderr, "Couldn't statfs() memory path: %s\n",
  67                     strerror(errno));
  68             exit(1);
  69         }
  70
  71         if (fs.f_type == HUGETLBFS_MAGIC) {
  72             /* It's hugepage, return the huge page size */
  73             return fs.f_bsize;
  74         }
  75     }
  76 #ifdef __sparc__
  77     /* SPARC Linux needs greater alignment than the pagesize */
  78     return QEMU_VMALLOC_ALIGN;
  79 #endif
  80 #endif
  81
  82     return qemu_real_host_page_size;
  83 }
  84
  85 void *qemu_ram_mmap(int fd,
  86                     size_t size,
  87                     size_t align,
  88                     bool readonly,
  89                     bool shared,
  90                     bool is_pmem,
  91                     off_t map_offset)
  92 {
  93     int prot;
  94     int flags;
  95     int map_sync_flags = 0;
  96     int guardfd;
  97     size_t offset;
  98     size_t pagesize;
  99     size_t total;
 100     void *guardptr;
 101     void *ptr;
 102
 103     /*
 104      * Note: this always allocates at least one extra page of virtual address
 105      * space, even if size is already aligned.
 106      */
 107     total = size + align;
 108
 109 #if defined(__powerpc64__) && defined(__linux__)
 110     /* On ppc64 mappings in the same segment (aka slice) must share the same
 111      * page size. Since we will be re-allocating part of this segment
 112      * from the supplied fd, we should make sure to use the same page size, to
 113      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
 114      * avoid allocating backing store memory.
 115      * We do this unless we are using the system page size, in which case
 116      * anonymous memory is OK.
 117      */
 118     flags = MAP_PRIVATE;
 119     pagesize = qemu_fd_getpagesize(fd);
 120     if (fd == -1 || pagesize == qemu_real_host_page_size) {
 121         guardfd = -1;
 122         flags |= MAP_ANONYMOUS;
 123     } else {
 124         guardfd = fd;
 125         flags |= MAP_NORESERVE;
 126     }
 127 #else
 128     guardfd = -1;
 129     pagesize = qemu_real_host_page_size;
 130     flags = MAP_PRIVATE | MAP_ANONYMOUS;
 131 #endif
 132
 133     guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
 134
 135     if (guardptr == MAP_FAILED) {
 136         return MAP_FAILED;
 137     }
 138
 139     assert(is_power_of_2(align));
 140     /* Always align to host page size */
 141     assert(align >= pagesize);
 142
 143     flags = MAP_FIXED;
 144     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 145     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 146     if (shared && is_pmem) {
 147         map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
 148     }
 149
 150     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 151
 152     prot = PROT_READ | (readonly ? 0 : PROT_WRITE);
 153
 154     ptr = mmap(guardptr + offset, size, prot,
 155                flags | map_sync_flags, fd, map_offset);
 156
 157     if (ptr == MAP_FAILED && map_sync_flags) {
 158         if (errno == ENOTSUP) {
 159             char *proc_link, *file_name;
 160             int len;
 161             proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
 162             file_name = g_malloc0(PATH_MAX);
 163             len = readlink(proc_link, file_name, PATH_MAX - 1);
 164             if (len < 0) {
 165                 len = 0;
 166             }
 167             file_name[len] = '\0';
 168             fprintf(stderr, "Warning: requesting persistence across crashes "
 169                     "for backend file %s failed. Proceeding without "
 170                     "persistence, data might become corrupted in case of host "
 171                     "crash.\n", file_name);
 172             g_free(proc_link);
 173             g_free(file_name);
 174         }
 175         /*
 176          * if map failed with MAP_SHARED_VALIDATE | MAP_SYNC,
 177          * we will remove these flags to handle compatibility.
 178          */
 179         ptr = mmap(guardptr + offset, size, prot, flags, fd, map_offset);
 180     }
 181
 182     if (ptr == MAP_FAILED) {
 183         munmap(guardptr, total);
 184         return MAP_FAILED;
 185     }
 186
 187     if (offset > 0) {
 188         munmap(guardptr, offset);
 189     }
 190
 191     /*
 192      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 193      * a guard page guarding against potential buffer overflows.
 194      */
 195     total -= offset;
 196     if (total > size + pagesize) {
 197         munmap(ptr + size + pagesize, total - size - pagesize);
 198     }
 199
 200     return ptr;
 201 }
 202
 203 void qemu_ram_munmap(int fd, void *ptr, size_t size)
 204 {
 205     size_t pagesize;
 206
 207     if (ptr) {
 208         /* Unmap both the RAM block and the guard page */
 209 #if defined(__powerpc64__) && defined(__linux__)
 210         pagesize = qemu_fd_getpagesize(fd);
 211 #else
 212         pagesize = qemu_real_host_page_size;
 213 #endif
 214         munmap(ptr, size + pagesize);
 215     }
 216 }