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/error-report.h" 24 25 #define HUGETLBFS_MAGIC 0x958458f6 26 27 #ifdef CONFIG_LINUX 28 #include <sys/vfs.h> 29 #endif 30 31 size_t qemu_fd_getpagesize(int fd) 32 { 33 #ifdef CONFIG_LINUX 34 struct statfs fs; 35 int ret; 36 37 if (fd != -1) { 38 do { 39 ret = fstatfs(fd, &fs); 40 } while (ret != 0 && errno == EINTR); 41 42 if (ret == 0 && fs.f_type == HUGETLBFS_MAGIC) { 43 return fs.f_bsize; 44 } 45 } 46 #ifdef __sparc__ 47 /* SPARC Linux needs greater alignment than the pagesize */ 48 return QEMU_VMALLOC_ALIGN; 49 #endif 50 #endif 51 52 return qemu_real_host_page_size; 53 } 54 55 size_t qemu_mempath_getpagesize(const char *mem_path) 56 { 57 #ifdef CONFIG_LINUX 58 struct statfs fs; 59 int ret; 60 61 if (mem_path) { 62 do { 63 ret = statfs(mem_path, &fs); 64 } while (ret != 0 && errno == EINTR); 65 66 if (ret != 0) { 67 fprintf(stderr, "Couldn't statfs() memory path: %s\n", 68 strerror(errno)); 69 exit(1); 70 } 71 72 if (fs.f_type == HUGETLBFS_MAGIC) { 73 /* It's hugepage, return the huge page size */ 74 return fs.f_bsize; 75 } 76 } 77 #ifdef __sparc__ 78 /* SPARC Linux needs greater alignment than the pagesize */ 79 return QEMU_VMALLOC_ALIGN; 80 #endif 81 #endif 82 83 return qemu_real_host_page_size; 84 } 85 86 /* 87 * Reserve a new memory region of the requested size to be used for mapping 88 * from the given fd (if any). 89 */ 90 static void *mmap_reserve(size_t size, int fd) 91 { 92 int flags = MAP_PRIVATE; 93 94 #if defined(__powerpc64__) && defined(__linux__) 95 /* 96 * On ppc64 mappings in the same segment (aka slice) must share the same 97 * page size. Since we will be re-allocating part of this segment 98 * from the supplied fd, we should make sure to use the same page size, to 99 * this end we mmap the supplied fd. In this case, set MAP_NORESERVE to 100 * avoid allocating backing store memory. 101 * We do this unless we are using the system page size, in which case 102 * anonymous memory is OK. 103 */ 104 if (fd == -1 || qemu_fd_getpagesize(fd) == qemu_real_host_page_size) { 105 fd = -1; 106 flags |= MAP_ANONYMOUS; 107 } else { 108 flags |= MAP_NORESERVE; 109 } 110 #else 111 fd = -1; 112 flags |= MAP_ANONYMOUS; 113 #endif 114 115 return mmap(0, size, PROT_NONE, flags, fd, 0); 116 } 117 118 /* 119 * Activate memory in a reserved region from the given fd (if any), to make 120 * it accessible. 121 */ 122 static void *mmap_activate(void *ptr, size_t size, int fd, 123 uint32_t qemu_map_flags, off_t map_offset) 124 { 125 const bool noreserve = qemu_map_flags & QEMU_MAP_NORESERVE; 126 const bool readonly = qemu_map_flags & QEMU_MAP_READONLY; 127 const bool shared = qemu_map_flags & QEMU_MAP_SHARED; 128 const bool sync = qemu_map_flags & QEMU_MAP_SYNC; 129 const int prot = PROT_READ | (readonly ? 0 : PROT_WRITE); 130 int map_sync_flags = 0; 131 int flags = MAP_FIXED; 132 void *activated_ptr; 133 134 if (noreserve) { 135 error_report("Skipping reservation of swap space is not supported"); 136 return MAP_FAILED; 137 } 138 139 flags |= fd == -1 ? MAP_ANONYMOUS : 0; 140 flags |= shared ? MAP_SHARED : MAP_PRIVATE; 141 if (shared && sync) { 142 map_sync_flags = MAP_SYNC | MAP_SHARED_VALIDATE; 143 } 144 145 activated_ptr = mmap(ptr, size, prot, flags | map_sync_flags, fd, 146 map_offset); 147 if (activated_ptr == MAP_FAILED && map_sync_flags) { 148 if (errno == ENOTSUP) { 149 char *proc_link = g_strdup_printf("/proc/self/fd/%d", fd); 150 char *file_name = g_malloc0(PATH_MAX); 151 int len = readlink(proc_link, file_name, PATH_MAX - 1); 152 153 if (len < 0) { 154 len = 0; 155 } 156 file_name[len] = '\0'; 157 fprintf(stderr, "Warning: requesting persistence across crashes " 158 "for backend file %s failed. Proceeding without " 159 "persistence, data might become corrupted in case of host " 160 "crash.\n", file_name); 161 g_free(proc_link); 162 g_free(file_name); 163 } 164 /* 165 * If mmap failed with MAP_SHARED_VALIDATE | MAP_SYNC, we will try 166 * again without these flags to handle backwards compatibility. 167 */ 168 activated_ptr = mmap(ptr, size, prot, flags, fd, map_offset); 169 } 170 return activated_ptr; 171 } 172 173 static inline size_t mmap_guard_pagesize(int fd) 174 { 175 #if defined(__powerpc64__) && defined(__linux__) 176 /* Mappings in the same segment must share the same page size */ 177 return qemu_fd_getpagesize(fd); 178 #else 179 return qemu_real_host_page_size; 180 #endif 181 } 182 183 void *qemu_ram_mmap(int fd, 184 size_t size, 185 size_t align, 186 uint32_t qemu_map_flags, 187 off_t map_offset) 188 { 189 const size_t guard_pagesize = mmap_guard_pagesize(fd); 190 size_t offset, total; 191 void *ptr, *guardptr; 192 193 /* 194 * Note: this always allocates at least one extra page of virtual address 195 * space, even if size is already aligned. 196 */ 197 total = size + align; 198 199 guardptr = mmap_reserve(total, fd); 200 if (guardptr == MAP_FAILED) { 201 return MAP_FAILED; 202 } 203 204 assert(is_power_of_2(align)); 205 /* Always align to host page size */ 206 assert(align >= guard_pagesize); 207 208 offset = QEMU_ALIGN_UP((uintptr_t)guardptr, align) - (uintptr_t)guardptr; 209 210 ptr = mmap_activate(guardptr + offset, size, fd, qemu_map_flags, 211 map_offset); 212 if (ptr == MAP_FAILED) { 213 munmap(guardptr, total); 214 return MAP_FAILED; 215 } 216 217 if (offset > 0) { 218 munmap(guardptr, offset); 219 } 220 221 /* 222 * Leave a single PROT_NONE page allocated after the RAM block, to serve as 223 * a guard page guarding against potential buffer overflows. 224 */ 225 total -= offset; 226 if (total > size + guard_pagesize) { 227 munmap(ptr + size + guard_pagesize, total - size - guard_pagesize); 228 } 229 230 return ptr; 231 } 232 233 void qemu_ram_munmap(int fd, void *ptr, size_t size) 234 { 235 if (ptr) { 236 /* Unmap both the RAM block and the guard page */ 237 munmap(ptr, size + mmap_guard_pagesize(fd)); 238 } 239 } 240