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