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 static inline size_t mmap_guard_pagesize(int fd)
  86 {
  87 #if defined(__powerpc64__) && defined(__linux__)
  88     /* Mappings in the same segment must share the same page size */
  89     return qemu_fd_getpagesize(fd);
  90 #else
  91     return qemu_real_host_page_size;
  92 #endif
  93 }
  94
  95 void *qemu_ram_mmap(int fd,
  96                     size_t size,
  97                     size_t align,
  98                     bool readonly,
  99                     bool shared,
 100                     bool is_pmem,
 101                     off_t map_offset)
 102 {
 103     const size_t guard_pagesize = mmap_guard_pagesize(fd);
 104     int prot;
 105     int flags;
 106     int map_sync_flags = 0;
 107     int guardfd;
 108     size_t offset;
 109     size_t total;
 110     void *guardptr;
 111     void *ptr;
 112
 113     /*
 114      * Note: this always allocates at least one extra page of virtual address
 115      * space, even if size is already aligned.
 116      */
 117     total = size + align;
 118
 119 #if defined(__powerpc64__) && defined(__linux__)
 120     /* On ppc64 mappings in the same segment (aka slice) must share the same
 121      * page size. Since we will be re-allocating part of this segment
 122      * from the supplied fd, we should make sure to use the same page size, to
 123      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
 124      * avoid allocating backing store memory.
 125      * We do this unless we are using the system page size, in which case
 126      * anonymous memory is OK.
 127      */
 128     flags = MAP_PRIVATE;
 129     if (fd == -1 || guard_pagesize == qemu_real_host_page_size) {
 130         guardfd = -1;
 131         flags |= MAP_ANONYMOUS;
 132     } else {
 133         guardfd = fd;
 134         flags |= MAP_NORESERVE;
 135     }
 136 #else
 137     guardfd = -1;
 138     flags = MAP_PRIVATE | MAP_ANONYMOUS;
 139 #endif
 140
 141     guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
 142
 143     if (guardptr == MAP_FAILED) {
 144         return MAP_FAILED;
 145     }
 146
 147     assert(is_power_of_2(align));
 148     /* Always align to host page size */
 149     assert(align >= guard_pagesize);
 150
 151     flags = MAP_FIXED;
 152     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 153     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 154     if (shared && is_pmem) {
 155         map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
 156     }
 157
 158     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 159
 160     prot = PROT_READ | (readonly ? 0 : PROT_WRITE);
 161
 162     ptr = mmap(guardptr + offset, size, prot,
 163                flags | map_sync_flags, fd, map_offset);
 164
 165     if (ptr == MAP_FAILED && map_sync_flags) {
 166         if (errno == ENOTSUP) {
 167             char *proc_link, *file_name;
 168             int len;
 169             proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
 170             file_name = g_malloc0(PATH_MAX);
 171             len = readlink(proc_link, file_name, PATH_MAX - 1);
 172             if (len < 0) {
 173                 len = 0;
 174             }
 175             file_name[len] = '\0';
 176             fprintf(stderr, "Warning: requesting persistence across crashes "
 177                     "for backend file %s failed. Proceeding without "
 178                     "persistence, data might become corrupted in case of host "
 179                     "crash.\n", file_name);
 180             g_free(proc_link);
 181             g_free(file_name);
 182         }
 183         /*
 184          * if map failed with MAP_SHARED_VALIDATE | MAP_SYNC,
 185          * we will remove these flags to handle compatibility.
 186          */
 187         ptr = mmap(guardptr + offset, size, prot, flags, fd, map_offset);
 188     }
 189
 190     if (ptr == MAP_FAILED) {
 191         munmap(guardptr, total);
 192         return MAP_FAILED;
 193     }
 194
 195     if (offset > 0) {
 196         munmap(guardptr, offset);
 197     }
 198
 199     /*
 200      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 201      * a guard page guarding against potential buffer overflows.
 202      */
 203     total -= offset;
 204     if (total > size + guard_pagesize) {
 205         munmap(ptr + size + guard_pagesize, total - size - guard_pagesize);
 206     }
 207
 208     return ptr;
 209 }
 210
 211 void qemu_ram_munmap(int fd, void *ptr, size_t size)
 212 {
 213     if (ptr) {
 214         /* Unmap both the RAM block and the guard page */
 215         munmap(ptr, size + mmap_guard_pagesize(fd));
 216     }
 217 }