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 #include "qemu/cutils.h"
  24 #include "qemu/error-report.h"
  25
  26 #define HUGETLBFS_MAGIC       0x958458f6
  27
  28 #ifdef CONFIG_LINUX
  29 #include <sys/vfs.h>
  30 #endif
  31
  32 size_t qemu_fd_getpagesize(int fd)
  33 {
  34 #ifdef CONFIG_LINUX
  35     struct statfs fs;
  36     int ret;
  37
  38     if (fd != -1) {
  39         do {
  40             ret = fstatfs(fd, &fs);
  41         } while (ret != 0 && errno == EINTR);
  42
  43         if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
  44             return fs.f_bsize;
  45         }
  46     }
  47 #ifdef __sparc__
  48     /* SPARC Linux needs greater alignment than the pagesize */
  49     return QEMU_VMALLOC_ALIGN;
  50 #endif
  51 #endif
  52
  53     return qemu_real_host_page_size;
  54 }
  55
  56 size_t qemu_mempath_getpagesize(const char *mem_path)
  57 {
  58 #ifdef CONFIG_LINUX
  59     struct statfs fs;
  60     int ret;
  61
  62     if (mem_path) {
  63         do {
  64             ret = statfs(mem_path, &fs);
  65         } while (ret != 0 && errno == EINTR);
  66
  67         if (ret != 0) {
  68             fprintf(stderr, "Couldn't statfs() memory path: %s\n",
  69                     strerror(errno));
  70             exit(1);
  71         }
  72
  73         if (fs.f_type == HUGETLBFS_MAGIC) {
  74             /* It's hugepage, return the huge page size */
  75             return fs.f_bsize;
  76         }
  77     }
  78 #ifdef __sparc__
  79     /* SPARC Linux needs greater alignment than the pagesize */
  80     return QEMU_VMALLOC_ALIGN;
  81 #endif
  82 #endif
  83
  84     return qemu_real_host_page_size;
  85 }
  86
  87 #define OVERCOMMIT_MEMORY_PATH "/proc/sys/vm/overcommit_memory"
  88 static bool map_noreserve_effective(int fd, uint32_t qemu_map_flags)
  89 {
  90 #if defined(__linux__)
  91     const bool readonly = qemu_map_flags & QEMU_MAP_READONLY;
  92     const bool shared = qemu_map_flags & QEMU_MAP_SHARED;
  93     gchar *content = NULL;
  94     const char *endptr;
  95     unsigned int tmp;
  96
  97     /*
  98      * hugeltb accounting is different than ordinary swap reservation:
  99      * a) Hugetlb pages from the pool are reserved for both private and
 100      *    shared mappings. For shared mappings, all mappers have to specify
 101      *    MAP_NORESERVE.
 102      * b) MAP_NORESERVE is not affected by /proc/sys/vm/overcommit_memory.
 103      */
 104     if (qemu_fd_getpagesize(fd) != qemu_real_host_page_size) {
 105         return true;
 106     }
 107
 108     /*
 109      * Accountable mappings in the kernel that can be affected by MAP_NORESEVE
 110      * are private writable mappings (see mm/mmap.c:accountable_mapping() in
 111      * Linux). For all shared or readonly mappings, MAP_NORESERVE is always
 112      * implicitly active -- no reservation; this includes shmem. The only
 113      * exception is shared anonymous memory, it is accounted like private
 114      * anonymous memory.
 115      */
 116     if (readonly || (shared && fd >= 0)) {
 117         return true;
 118     }
 119
 120     /*
 121      * MAP_NORESERVE is globally ignored for applicable !hugetlb mappings when
 122      * memory overcommit is set to "never". Sparse memory regions aren't really
 123      * possible in this system configuration.
 124      *
 125      * Bail out now instead of silently committing way more memory than
 126      * currently desired by the user.
 127      */
 128     if (g_file_get_contents(OVERCOMMIT_MEMORY_PATH, &content, NULL, NULL) &&
 129         !qemu_strtoui(content, &endptr, 0, &tmp) &&
 130         (!endptr || *endptr == '\n')) {
 131         if (tmp == 2) {
 132             error_report("Skipping reservation of swap space is not supported:"
 133                          " \"" OVERCOMMIT_MEMORY_PATH "\" is \"2\"");
 134             return false;
 135         }
 136         return true;
 137     }
 138     /* this interface has been around since Linux 2.6 */
 139     error_report("Skipping reservation of swap space is not supported:"
 140                  " Could not read: \"" OVERCOMMIT_MEMORY_PATH "\"");
 141     return false;
 142 #endif
 143     /*
 144      * E.g., FreeBSD used to define MAP_NORESERVE, never implemented it,
 145      * and removed it a while ago.
 146      */
 147     error_report("Skipping reservation of swap space is not supported");
 148     return false;
 149 }
 150
 151 /*
 152  * Reserve a new memory region of the requested size to be used for mapping
 153  * from the given fd (if any).
 154  */
 155 static void *mmap_reserve(size_t size, int fd)
 156 {
 157     int flags = MAP_PRIVATE;
 158
 159 #if defined(__powerpc64__) && defined(__linux__)
 160     /*
 161      * On ppc64 mappings in the same segment (aka slice) must share the same
 162      * page size. Since we will be re-allocating part of this segment
 163      * from the supplied fd, we should make sure to use the same page size, to
 164      * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
 165      * avoid allocating backing store memory.
 166      * We do this unless we are using the system page size, in which case
 167      * anonymous memory is OK.
 168      */
 169     if (fd == -1 || qemu_fd_getpagesize(fd) == qemu_real_host_page_size) {
 170         fd = -1;
 171         flags |= MAP_ANONYMOUS;
 172     } else {
 173         flags |= MAP_NORESERVE;
 174     }
 175 #else
 176     fd = -1;
 177     flags |= MAP_ANONYMOUS;
 178 #endif
 179
 180     return mmap(0, size, PROT_NONE, flags, fd, 0);
 181 }
 182
 183 /*
 184  * Activate memory in a reserved region from the given fd (if any), to make
 185  * it accessible.
 186  */
 187 static void *mmap_activate(void *ptr, size_t size, int fd,
 188                            uint32_t qemu_map_flags, off_t map_offset)
 189 {
 190     const bool noreserve = qemu_map_flags & QEMU_MAP_NORESERVE;
 191     const bool readonly = qemu_map_flags & QEMU_MAP_READONLY;
 192     const bool shared = qemu_map_flags & QEMU_MAP_SHARED;
 193     const bool sync = qemu_map_flags & QEMU_MAP_SYNC;
 194     const int prot = PROT_READ | (readonly ? 0 : PROT_WRITE);
 195     int map_sync_flags = 0;
 196     int flags = MAP_FIXED;
 197     void *activated_ptr;
 198
 199     if (noreserve && !map_noreserve_effective(fd, qemu_map_flags)) {
 200         return MAP_FAILED;
 201     }
 202
 203     flags |= fd == -1 ? MAP_ANONYMOUS : 0;
 204     flags |= shared ? MAP_SHARED : MAP_PRIVATE;
 205     flags |= noreserve ? MAP_NORESERVE : 0;
 206     if (shared && sync) {
 207         map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
 208     }
 209
 210     activated_ptr = mmap(ptr, size, prot, flags | map_sync_flags, fd,
 211                          map_offset);
 212     if (activated_ptr == MAP_FAILED && map_sync_flags) {
 213         if (errno == ENOTSUP) {
 214             char *proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
 215             char *file_name = g_malloc0(PATH_MAX);
 216             int len = readlink(proc_link, file_name, PATH_MAX - 1);
 217
 218             if (len < 0) {
 219                 len = 0;
 220             }
 221             file_name[len] = '\0';
 222             fprintf(stderr, "Warning: requesting persistence across crashes "
 223                     "for backend file %s failed. Proceeding without "
 224                     "persistence, data might become corrupted in case of host "
 225                     "crash.\n", file_name);
 226             g_free(proc_link);
 227             g_free(file_name);
 228             warn_report("Using non DAX backing file with 'pmem=on' option"
 229                         " is deprecated");
 230         }
 231         /*
 232          * If mmap failed with MAP_SHARED_VALIDATE | MAP_SYNC, we will try
 233          * again without these flags to handle backwards compatibility.
 234          */
 235         activated_ptr = mmap(ptr, size, prot, flags, fd, map_offset);
 236     }
 237     return activated_ptr;
 238 }
 239
 240 static inline size_t mmap_guard_pagesize(int fd)
 241 {
 242 #if defined(__powerpc64__) && defined(__linux__)
 243     /* Mappings in the same segment must share the same page size */
 244     return qemu_fd_getpagesize(fd);
 245 #else
 246     return qemu_real_host_page_size;
 247 #endif
 248 }
 249
 250 void *qemu_ram_mmap(int fd,
 251                     size_t size,
 252                     size_t align,
 253                     uint32_t qemu_map_flags,
 254                     off_t map_offset)
 255 {
 256     const size_t guard_pagesize = mmap_guard_pagesize(fd);
 257     size_t offset, total;
 258     void *ptr, *guardptr;
 259
 260     /*
 261      * Note: this always allocates at least one extra page of virtual address
 262      * space, even if size is already aligned.
 263      */
 264     total = size + align;
 265
 266     guardptr = mmap_reserve(total, fd);
 267     if (guardptr == MAP_FAILED) {
 268         return MAP_FAILED;
 269     }
 270
 271     assert(is_power_of_2(align));
 272     /* Always align to host page size */
 273     assert(align >= guard_pagesize);
 274
 275     offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
 276
 277     ptr = mmap_activate(guardptr + offset, size, fd, qemu_map_flags,
 278                         map_offset);
 279     if (ptr == MAP_FAILED) {
 280         munmap(guardptr, total);
 281         return MAP_FAILED;
 282     }
 283
 284     if (offset > 0) {
 285         munmap(guardptr, offset);
 286     }
 287
 288     /*
 289      * Leave a single PROT_NONE page allocated after the RAM block, to serve as
 290      * a guard page guarding against potential buffer overflows.
 291      */
 292     total -= offset;
 293     if (total > size + guard_pagesize) {
 294         munmap(ptr + size + guard_pagesize, total - size - guard_pagesize);
 295     }
 296
 297     return ptr;
 298 }
 299
 300 void qemu_ram_munmap(int fd, void *ptr, size_t size)
 301 {
 302     if (ptr) {
 303         /* Unmap both the RAM block and the guard page */
 304         munmap(ptr, size + mmap_guard_pagesize(fd));
 305     }
 306 }