2 * Support for RAM backed by mmaped host memory.
4 * Copyright (c) 2015 Red Hat, Inc.
7 * Michael S. Tsirkin <mst@redhat.com>
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,
95 int map_sync_flags = 0;
104 * Note: this always allocates at least one extra page of virtual address
105 * space, even if size is already aligned.
107 total = size + align;
109 #if defined(__powerpc64__) && defined(__linux__)
110 /* On ppc64 mappings in the same segment (aka slice) must share the same
111 * page size. Since we will be re-allocating part of this segment
112 * from the supplied fd, we should make sure to use the same page size, to
113 * this end we mmap the supplied fd. In this case, set MAP_NORESERVE to
114 * avoid allocating backing store memory.
115 * We do this unless we are using the system page size, in which case
116 * anonymous memory is OK.
119 pagesize = qemu_fd_getpagesize(fd);
120 if (fd == -1 || pagesize == qemu_real_host_page_size) {
122 flags |= MAP_ANONYMOUS;
125 flags |= MAP_NORESERVE;
129 pagesize = qemu_real_host_page_size;
130 flags = MAP_PRIVATE | MAP_ANONYMOUS;
133 guardptr = mmap(0, total, PROT_NONE, flags, guardfd, 0);
135 if (guardptr == MAP_FAILED) {
139 assert(is_power_of_2(align));
140 /* Always align to host page size */
141 assert(align >= pagesize);
144 flags |= fd == -1 ? MAP_ANONYMOUS : 0;
145 flags |= shared ? MAP_SHARED : MAP_PRIVATE;
146 if (shared && is_pmem) {
147 map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE;
150 offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr;
152 prot = PROT_READ | (readonly ? 0 : PROT_WRITE);
154 ptr = mmap(guardptr + offset, size, prot,
155 flags | map_sync_flags, fd, map_offset);
157 if (ptr == MAP_FAILED && map_sync_flags) {
158 if (errno == ENOTSUP) {
159 char *proc_link, *file_name;
161 proc_link = g_strdup_printf("/proc/self/fd/%d", fd);
162 file_name = g_malloc0(PATH_MAX);
163 len = readlink(proc_link, file_name, PATH_MAX - 1);
167 file_name[len] = '\0';
168 fprintf(stderr, "Warning: requesting persistence across crashes "
169 "for backend file %s failed. Proceeding without "
170 "persistence, data might become corrupted in case of host "
171 "crash.\n", file_name);
176 * if map failed with MAP_SHARED_VALIDATE | MAP_SYNC,
177 * we will remove these flags to handle compatibility.
179 ptr = mmap(guardptr + offset, size, prot, flags, fd, map_offset);
182 if (ptr == MAP_FAILED) {
183 munmap(guardptr, total);
188 munmap(guardptr, offset);
192 * Leave a single PROT_NONE page allocated after the RAM block, to serve as
193 * a guard page guarding against potential buffer overflows.
196 if (total > size + pagesize) {
197 munmap(ptr + size + pagesize, total - size - pagesize);
203 void qemu_ram_munmap(int fd, void *ptr, size_t size)
208 /* Unmap both the RAM block and the guard page */
209 #if defined(__powerpc64__) && defined(__linux__)
210 pagesize = qemu_fd_getpagesize(fd);
212 pagesize = qemu_real_host_page_size;
214 munmap(ptr, size + pagesize);