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 getpagesize();
  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 getpagesize();
  83 }
  84
  85 void *qemu_ram_mmap(int fd,
  86                     size_t size,
  87                     size_t align,
  88                     bool shared,
  89                     bool is_pmem)
  90 {
  91     int flags;
  92     int map_sync_flags = 0;
  93     int guardfd;
  94     size_t offset;
  95     size_t pagesize;
  96     size_t total;
  97     void *guardptr;
  98     void *ptr;
  99
 100     /*
 101      * Note: this always allocates at least one extra page of virtual address
 102      * space, even if size is already aligned.
 103      */
 104     total = size + align;
 105
 106 #if defined(__powerpc64__) && defined(__linux__)
 107     /* On ppc64 mappings in the same segment (aka slice) must share the same
 108      * page size. Since we will be re-allocating part of this segment
 109      * from the supplied fd, we should make sure to use the same page size, to
 110      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
 111      * avoid allocating backing store memory.
 112      * We do this unless we are using the system page size, in which case
 113      * anonymous memory is OK.
 114      */
 115     flags = MAP_PRIVATE;
 116     pagesize = qemu_fd_getpagesize(fd);
 117     if (fd == -1 || pagesize == getpagesize()) {
 118         guardfd = -1;
 119         flags |= MAP_ANONYMOUS;
 120     } else {
 121         guardfd = fd;
 122         flags |= MAP_NORESERVE;
 123     }
 124 #else
 125     guardfd = -1;
 126     pagesize = getpagesize();
 127     flags = MAP_PRIVATE | MAP_ANONYMOUS;
 128 #endif
 129
 130     guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
 131
 132     if (guardptr == MAP_FAILED) {
 133         return MAP_FAILED;
 134     }
 135
 136     assert(is_power_of_2(align));
 137     /* Always align to host page size */
 138     assert(align >= pagesize);
 139
 140     flags = MAP_FIXED;
 141     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 142     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 143     if (shared && is_pmem) {
 144         map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
 145     }
 146
 147     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 148
 149     ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE,
 150                flags | map_sync_flags, fd, 0);
 151
 152     if (ptr == MAP_FAILED && map_sync_flags) {
 153         if (errno == ENOTSUP) {
 154             char *proc_link, *file_name;
 155             int len;
 156             proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
 157             file_name = g_malloc0(PATH_MAX);
 158             len = readlink(proc_link, file_name, PATH_MAX - 1);
 159             if (len < 0) {
 160                 len = 0;
 161             }
 162             file_name[len] = '\0';
 163             fprintf(stderr, "Warning: requesting persistence across crashes "
 164                     "for backend file %s failed. Proceeding without "
 165                     "persistence, data might become corrupted in case of host "
 166                     "crash.\n", file_name);
 167             g_free(proc_link);
 168             g_free(file_name);
 169         }
 170         /*
 171          * if map failed with MAP_SHARED_VALIDATE | MAP_SYNC,
 172          * we will remove these flags to handle compatibility.
 173          */
 174         ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE,
 175                    flags, fd, 0);
 176     }
 177
 178     if (ptr == MAP_FAILED) {
 179         munmap(guardptr, total);
 180         return MAP_FAILED;
 181     }
 182
 183     if (offset > 0) {
 184         munmap(guardptr, offset);
 185     }
 186
 187     /*
 188      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 189      * a guard page guarding against potential buffer overflows.
 190      */
 191     total -= offset;
 192     if (total > size + pagesize) {
 193         munmap(ptr + size + pagesize, total - size - pagesize);
 194     }
 195
 196     return ptr;
 197 }
 198
 199 void qemu_ram_munmap(int fd, void *ptr, size_t size)
 200 {
 201     size_t pagesize;
 202
 203     if (ptr) {
 204         /* Unmap both the RAM block and the guard page */
 205 #if defined(__powerpc64__) && defined(__linux__)
 206         pagesize = qemu_fd_getpagesize(fd);
 207 #else
 208         pagesize = getpagesize();
 209 #endif
 210         munmap(ptr, size + pagesize);
 211     }
 212 }