2 * Support for RAM backed by mmaped host memory.
4 * Copyright (c) 2015 Red Hat, Inc.
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.
14 #include <linux/mman.h>
15 #else /* !CONFIG_LINUX */
17 #define MAP_SHARED_VALIDATE 0x0
18 #endif /* CONFIG_LINUX */
20 #include "qemu/osdep.h"
21 #include "qemu/mmap-alloc.h"
22 #include "qemu/host-utils.h"
24 #define HUGETLBFS_MAGIC 0x958458f6
30 size_t qemu_fd_getpagesize(int fd)
38 ret = fstatfs(fd, &fs);
39 } while (ret != 0 && errno == EINTR);
41 if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) {
46 /* SPARC Linux needs greater alignment than the pagesize */
47 return QEMU_VMALLOC_ALIGN;
51 return qemu_real_host_page_size;
54 size_t qemu_mempath_getpagesize(const char *mem_path)
62 ret = statfs(mem_path, &fs);
63 } while (ret != 0 && errno == EINTR);
66 fprintf(stderr, "Couldn't statfs() memory path: %s\n",
71 if (fs.f_type == HUGETLBFS_MAGIC) {
72 /* It's hugepage, return the huge page size */
77 /* SPARC Linux needs greater alignment than the pagesize */
78 return QEMU_VMALLOC_ALIGN;
82 return qemu_real_host_page_size;
85 void *qemu_ram_mmap(int fd,
92 int map_sync_flags = 0;
101 * Note: this always allocates at least one extra page of virtual address
102 * space, even if size is already aligned.
104 total = size + align;
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.
116 pagesize = qemu_fd_getpagesize(fd);
117 if (fd == -1 || pagesize == qemu_real_host_page_size) {
119 flags |= MAP_ANONYMOUS;
122 flags |= MAP_NORESERVE;
126 pagesize = qemu_real_host_page_size;
127 flags = MAP_PRIVATE | MAP_ANONYMOUS;
130 guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
132 if (guardptr == MAP_FAILED) {
136 assert(is_power_of_2(align));
137 /* Always align to host page size */
138 assert(align >= pagesize);
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;
147 offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
149 ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE,
150 flags | map_sync_flags, fd, 0);
152 if (ptr == MAP_FAILED && map_sync_flags) {
153 if (errno == ENOTSUP) {
154 char *proc_link, *file_name;
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);
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);
171 * if map failed with MAP_SHARED_VALIDATE | MAP_SYNC,
172 * we will remove these flags to handle compatibility.
174 ptr = mmap(guardptr + offset, size, PROT_READ | PROT_WRITE,
178 if (ptr == MAP_FAILED) {
179 munmap(guardptr, total);
184 munmap(guardptr, offset);
188 * Leave a single PROT_NONE page allocated after the RAM block, to serve as
189 * a guard page guarding against potential buffer overflows.
192 if (total > size + pagesize) {
193 munmap(ptr + size + pagesize, total - size - pagesize);
199 void qemu_ram_munmap(int fd, void *ptr, size_t size)
204 /* Unmap both the RAM block and the guard page */
205 #if defined(__powerpc64__) && defined(__linux__)
206 pagesize = qemu_fd_getpagesize(fd);
208 pagesize = qemu_real_host_page_size;
210 munmap(ptr, size + pagesize);