qemu

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
#include <linux/magic.h>
31
#endif
32

33
QemuFsType qemu_fd_getfs(int fd)
34
{
35
#ifdef CONFIG_LINUX
36
    struct statfs fs;
37
    int ret;
38

39
    if (fd < 0) {
40
        return QEMU_FS_TYPE_UNKNOWN;
41
    }
42

43
    do {
44
        ret = fstatfs(fd, &fs);
45
    } while (ret != 0 && errno == EINTR);
46

47
    switch (fs.f_type) {
48
    case TMPFS_MAGIC:
49
        return QEMU_FS_TYPE_TMPFS;
50
    case HUGETLBFS_MAGIC:
51
        return QEMU_FS_TYPE_HUGETLBFS;
52
    default:
53
        return QEMU_FS_TYPE_UNKNOWN;
54
    }
55
#else
56
    return QEMU_FS_TYPE_UNKNOWN;
57
#endif
58
}
59

60
size_t qemu_fd_getpagesize(int fd)
61
{
62
#ifdef CONFIG_LINUX
63
    struct statfs fs;
64
    int ret;
65

66
    if (fd != -1) {
67
        do {
68
            ret = fstatfs(fd, &fs);
69
        } while (ret != 0 && errno == EINTR);
70

71
        if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
72
            return fs.f_bsize;
73
        }
74
    }
75
#ifdef __sparc__
76
    /* SPARC Linux needs greater alignment than the pagesize */
77
    return QEMU_VMALLOC_ALIGN;
78
#endif
79
#endif
80

81
    return qemu_real_host_page_size();
82
}
83

84
#define OVERCOMMIT_MEMORY_PATH "/proc/sys/vm/overcommit_memory"
85
static bool map_noreserve_effective(int fd, uint32_t qemu_map_flags)
86
{
87
#if defined(__linux__)
88
    const bool readonly = qemu_map_flags & QEMU_MAP_READONLY;
89
    const bool shared = qemu_map_flags & QEMU_MAP_SHARED;
90
    gchar *content = NULL;
91
    const char *endptr;
92
    unsigned int tmp;
93

94
    /*
95
     * hugeltb accounting is different than ordinary swap reservation:
96
     * a) Hugetlb pages from the pool are reserved for both private and
97
     *    shared mappings. For shared mappings, all mappers have to specify
98
     *    MAP_NORESERVE.
99
     * b) MAP_NORESERVE is not affected by /proc/sys/vm/overcommit_memory.
100
     */
101
    if (qemu_fd_getpagesize(fd) != qemu_real_host_page_size()) {
102
        return true;
103
    }
104

105
    /*
106
     * Accountable mappings in the kernel that can be affected by MAP_NORESEVE
107
     * are private writable mappings (see mm/mmap.c:accountable_mapping() in
108
     * Linux). For all shared or readonly mappings, MAP_NORESERVE is always
109
     * implicitly active -- no reservation; this includes shmem. The only
110
     * exception is shared anonymous memory, it is accounted like private
111
     * anonymous memory.
112
     */
113
    if (readonly || (shared && fd >= 0)) {
114
        return true;
115
    }
116

117
    /*
118
     * MAP_NORESERVE is globally ignored for applicable !hugetlb mappings when
119
     * memory overcommit is set to "never". Sparse memory regions aren't really
120
     * possible in this system configuration.
121
     *
122
     * Bail out now instead of silently committing way more memory than
123
     * currently desired by the user.
124
     */
125
    if (g_file_get_contents(OVERCOMMIT_MEMORY_PATH, &content, NULL, NULL) &&
126
        !qemu_strtoui(content, &endptr, 0, &tmp) &&
127
        (!endptr || *endptr == '\n')) {
128
        if (tmp == 2) {
129
            error_report("Skipping reservation of swap space is not supported:"
130
                         " \"" OVERCOMMIT_MEMORY_PATH "\" is \"2\"");
131
            return false;
132
        }
133
        return true;
134
    }
135
    /* this interface has been around since Linux 2.6 */
136
    error_report("Skipping reservation of swap space is not supported:"
137
                 " Could not read: \"" OVERCOMMIT_MEMORY_PATH "\"");
138
    return false;
139
#endif
140
    /*
141
     * E.g., FreeBSD used to define MAP_NORESERVE, never implemented it,
142
     * and removed it a while ago.
143
     */
144
    error_report("Skipping reservation of swap space is not supported");
145
    return false;
146
}
147

148
/*
149
 * Reserve a new memory region of the requested size to be used for mapping
150
 * from the given fd (if any).
151
 */
152
static void *mmap_reserve(size_t size, int fd)
153
{
154
    int flags = MAP_PRIVATE;
155

156
#if defined(__powerpc64__) && defined(__linux__)
157
    /*
158
     * On ppc64 mappings in the same segment (aka slice) must share the same
159
     * page size. Since we will be re-allocating part of this segment
160
     * from the supplied fd, we should make sure to use the same page size, to
161
     * this end we mmap the supplied fd.  In this case, set MAP_NORESERVE to
162
     * avoid allocating backing store memory.
163
     * We do this unless we are using the system page size, in which case
164
     * anonymous memory is OK.
165
     */
166
    if (fd == -1 || qemu_fd_getpagesize(fd) == qemu_real_host_page_size()) {
167
        fd = -1;
168
        flags |= MAP_ANONYMOUS;
169
    } else {
170
        flags |= MAP_NORESERVE;
171
    }
172
#else
173
    fd = -1;
174
    flags |= MAP_ANONYMOUS;
175
#endif
176

177
    return mmap(0, size, PROT_NONE, flags, fd, 0);
178
}
179

180
/*
181
 * Activate memory in a reserved region from the given fd (if any), to make
182
 * it accessible.
183
 */
184
static void *mmap_activate(void *ptr, size_t size, int fd,
185
                           uint32_t qemu_map_flags, off_t map_offset)
186
{
187
    const bool noreserve = qemu_map_flags & QEMU_MAP_NORESERVE;
188
    const bool readonly = qemu_map_flags & QEMU_MAP_READONLY;
189
    const bool shared = qemu_map_flags & QEMU_MAP_SHARED;
190
    const bool sync = qemu_map_flags & QEMU_MAP_SYNC;
191
    const int prot = PROT_READ | (readonly ? 0 : PROT_WRITE);
192
    int map_sync_flags = 0;
193
    int flags = MAP_FIXED;
194
    void *activated_ptr;
195

196
    if (noreserve && !map_noreserve_effective(fd, qemu_map_flags)) {
197
        return MAP_FAILED;
198
    }
199

200
    flags |= fd == -1 ? MAP_ANONYMOUS : 0;
201
    flags |= shared ? MAP_SHARED : MAP_PRIVATE;
202
    flags |= noreserve ? MAP_NORESERVE : 0;
203
    if (shared && sync) {
204
        map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
205
    }
206

207
    activated_ptr = mmap(ptr, size, prot, flags | map_sync_flags, fd,
208
                         map_offset);
209
    if (activated_ptr == MAP_FAILED && map_sync_flags) {
210
        if (errno == ENOTSUP) {
211
            char *proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
212
            char *file_name = g_malloc0(PATH_MAX);
213
            int len = readlink(proc_link, file_name, PATH_MAX - 1);
214

215
            if (len < 0) {
216
                len = 0;
217
            }
218
            file_name[len] = '\0';
219
            fprintf(stderr, "Warning: requesting persistence across crashes "
220
                    "for backend file %s failed. Proceeding without "
221
                    "persistence, data might become corrupted in case of host "
222
                    "crash.\n", file_name);
223
            g_free(proc_link);
224
            g_free(file_name);
225
            warn_report("Using non DAX backing file with 'pmem=on' option"
226
                        " is deprecated");
227
        }
228
        /*
229
         * If mmap failed with MAP_SHARED_VALIDATE | MAP_SYNC, we will try
230
         * again without these flags to handle backwards compatibility.
231
         */
232
        activated_ptr = mmap(ptr, size, prot, flags, fd, map_offset);
233
    }
234
    return activated_ptr;
235
}
236

237
static inline size_t mmap_guard_pagesize(int fd)
238
{
239
#if defined(__powerpc64__) && defined(__linux__)
240
    /* Mappings in the same segment must share the same page size */
241
    return qemu_fd_getpagesize(fd);
242
#else
243
    return qemu_real_host_page_size();
244
#endif
245
}
246

247
void *qemu_ram_mmap(int fd,
248
                    size_t size,
249
                    size_t align,
250
                    uint32_t qemu_map_flags,
251
                    off_t map_offset)
252
{
253
    const size_t guard_pagesize = mmap_guard_pagesize(fd);
254
    size_t offset, total;
255
    void *ptr, *guardptr;
256

257
    /*
258
     * Note: this always allocates at least one extra page of virtual address
259
     * space, even if size is already aligned.
260
     */
261
    total = size + align;
262

263
    guardptr = mmap_reserve(total, fd);
264
    if (guardptr == MAP_FAILED) {
265
        return MAP_FAILED;
266
    }
267

268
    assert(is_power_of_2(align));
269
    /* Always align to host page size */
270
    assert(align >= guard_pagesize);
271

272
    offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
273

274
    ptr = mmap_activate(guardptr + offset, size, fd, qemu_map_flags,
275
                        map_offset);
276
    if (ptr == MAP_FAILED) {
277
        munmap(guardptr, total);
278
        return MAP_FAILED;
279
    }
280

281
    if (offset > 0) {
282
        munmap(guardptr, offset);
283
    }
284

285
    /*
286
     * Leave a single PROT_NONE page allocated after the RAM block, to serve as
287
     * a guard page guarding against potential buffer overflows.
288
     */
289
    total -= offset;
290
    if (total > size + guard_pagesize) {
291
        munmap(ptr + size + guard_pagesize, total - size - guard_pagesize);
292
    }
293

294
    return ptr;
295
}
296

297
void qemu_ram_munmap(int fd, void *ptr, size_t size)
298
{
299
    if (ptr) {
300
        /* Unmap both the RAM block and the guard page */
301
        munmap(ptr, size + mmap_guard_pagesize(fd));
302
    }
303
}
304