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