1 /* 1 /* 2 * memfd_create system call and file sealing s 2 * memfd_create system call and file sealing support 3 * 3 * 4 * Code was originally included in shmem.c, an 4 * Code was originally included in shmem.c, and broken out to facilitate 5 * use by hugetlbfs as well as tmpfs. 5 * use by hugetlbfs as well as tmpfs. 6 * 6 * 7 * This file is released under the GPL. 7 * This file is released under the GPL. 8 */ 8 */ 9 9 10 #include <linux/fs.h> 10 #include <linux/fs.h> 11 #include <linux/vfs.h> 11 #include <linux/vfs.h> 12 #include <linux/pagemap.h> 12 #include <linux/pagemap.h> 13 #include <linux/file.h> 13 #include <linux/file.h> 14 #include <linux/mm.h> 14 #include <linux/mm.h> 15 #include <linux/sched/signal.h> 15 #include <linux/sched/signal.h> 16 #include <linux/khugepaged.h> 16 #include <linux/khugepaged.h> 17 #include <linux/syscalls.h> 17 #include <linux/syscalls.h> 18 #include <linux/hugetlb.h> 18 #include <linux/hugetlb.h> 19 #include <linux/shmem_fs.h> 19 #include <linux/shmem_fs.h> 20 #include <linux/memfd.h> 20 #include <linux/memfd.h> 21 #include <linux/pid_namespace.h> << 22 #include <uapi/linux/memfd.h> 21 #include <uapi/linux/memfd.h> 23 22 24 /* 23 /* 25 * We need a tag: a new tag would expand every !! 24 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes, 26 * so reuse a tag which we firmly believe is n 25 * so reuse a tag which we firmly believe is never set or cleared on tmpfs 27 * or hugetlbfs because they are memory only f 26 * or hugetlbfs because they are memory only filesystems. 28 */ 27 */ 29 #define MEMFD_TAG_PINNED PAGECACHE_TAG_ 28 #define MEMFD_TAG_PINNED PAGECACHE_TAG_TOWRITE 30 #define LAST_SCAN 4 /* abo 29 #define LAST_SCAN 4 /* about 150ms max */ 31 30 32 static bool memfd_folio_has_extra_refs(struct !! 31 static void memfd_tag_pins(struct address_space *mapping) 33 { 32 { 34 return folio_ref_count(folio) - folio_ !! 33 struct radix_tree_iter iter; 35 folio_nr_pages(folio); !! 34 void __rcu **slot; 36 } !! 35 pgoff_t start; 37 !! 36 struct page *page; 38 static void memfd_tag_pins(struct xa_state *xa << 39 { << 40 struct folio *folio; << 41 int latency = 0; << 42 37 43 lru_add_drain(); 38 lru_add_drain(); >> 39 start = 0; >> 40 rcu_read_lock(); 44 41 45 xas_lock_irq(xas); !! 42 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) { 46 xas_for_each(xas, folio, ULONG_MAX) { !! 43 page = radix_tree_deref_slot(slot); 47 if (!xa_is_value(folio) && mem !! 44 if (!page || radix_tree_exception(page)) { 48 xas_set_mark(xas, MEMF !! 45 if (radix_tree_deref_retry(page)) { 49 !! 46 slot = radix_tree_iter_retry(&iter); 50 if (++latency < XA_CHECK_SCHED !! 47 continue; 51 continue; << 52 latency = 0; << 53 << 54 xas_pause(xas); << 55 xas_unlock_irq(xas); << 56 cond_resched(); << 57 xas_lock_irq(xas); << 58 } << 59 xas_unlock_irq(xas); << 60 } << 61 << 62 /* << 63 * This is a helper function used by memfd_pin << 64 * It is mainly called to allocate a folio in << 65 * (memfd_pin_folios()) cannot find a folio in << 66 * index in the mapping. << 67 */ << 68 struct folio *memfd_alloc_folio(struct file *m << 69 { << 70 #ifdef CONFIG_HUGETLB_PAGE << 71 struct folio *folio; << 72 gfp_t gfp_mask; << 73 int err; << 74 << 75 if (is_file_hugepages(memfd)) { << 76 /* << 77 * The folio would most likely << 78 * therefore, we have zone mem << 79 * alloc from. Also, the folio << 80 * amount of time, so it is no << 81 */ << 82 struct hstate *h = hstate_file << 83 << 84 gfp_mask = htlb_alloc_mask(h); << 85 gfp_mask &= ~(__GFP_HIGHMEM | << 86 idx >>= huge_page_order(h); << 87 << 88 folio = alloc_hugetlb_folio_re << 89 << 90 << 91 << 92 if (folio) { << 93 err = hugetlb_add_to_p << 94 << 95 << 96 if (err) { << 97 folio_put(foli << 98 return ERR_PTR << 99 } 48 } 100 folio_unlock(folio); !! 49 } else if (page_count(page) - page_mapcount(page) > 1) { 101 return folio; !! 50 xa_lock_irq(&mapping->i_pages); >> 51 radix_tree_tag_set(&mapping->i_pages, iter.index, >> 52 MEMFD_TAG_PINNED); >> 53 xa_unlock_irq(&mapping->i_pages); >> 54 } >> 55 >> 56 if (need_resched()) { >> 57 slot = radix_tree_iter_resume(slot, &iter); >> 58 cond_resched_rcu(); 102 } 59 } 103 return ERR_PTR(-ENOMEM); << 104 } 60 } 105 #endif !! 61 rcu_read_unlock(); 106 return shmem_read_folio(memfd->f_mappi << 107 } 62 } 108 63 109 /* 64 /* 110 * Setting SEAL_WRITE requires us to verify th 65 * Setting SEAL_WRITE requires us to verify there's no pending writer. However, 111 * via get_user_pages(), drivers might have so 66 * via get_user_pages(), drivers might have some pending I/O without any active 112 * user-space mappings (eg., direct-IO, AIO). !! 67 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages 113 * and see whether it has an elevated ref-coun 68 * and see whether it has an elevated ref-count. If so, we tag them and wait for 114 * them to be dropped. 69 * them to be dropped. 115 * The caller must guarantee that no new user 70 * The caller must guarantee that no new user will acquire writable references 116 * to those folios to avoid races. !! 71 * to those pages to avoid races. 117 */ 72 */ 118 static int memfd_wait_for_pins(struct address_ 73 static int memfd_wait_for_pins(struct address_space *mapping) 119 { 74 { 120 XA_STATE(xas, &mapping->i_pages, 0); !! 75 struct radix_tree_iter iter; 121 struct folio *folio; !! 76 void __rcu **slot; >> 77 pgoff_t start; >> 78 struct page *page; 122 int error, scan; 79 int error, scan; 123 80 124 memfd_tag_pins(&xas); !! 81 memfd_tag_pins(mapping); 125 82 126 error = 0; 83 error = 0; 127 for (scan = 0; scan <= LAST_SCAN; scan 84 for (scan = 0; scan <= LAST_SCAN; scan++) { 128 int latency = 0; !! 85 if (!radix_tree_tagged(&mapping->i_pages, MEMFD_TAG_PINNED)) 129 << 130 if (!xas_marked(&xas, MEMFD_TA << 131 break; 86 break; 132 87 133 if (!scan) 88 if (!scan) 134 lru_add_drain_all(); 89 lru_add_drain_all(); 135 else if (schedule_timeout_kill 90 else if (schedule_timeout_killable((HZ << scan) / 200)) 136 scan = LAST_SCAN; 91 scan = LAST_SCAN; 137 92 138 xas_set(&xas, 0); !! 93 start = 0; 139 xas_lock_irq(&xas); !! 94 rcu_read_lock(); 140 xas_for_each_marked(&xas, foli !! 95 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, 141 bool clear = true; !! 96 start, MEMFD_TAG_PINNED) { >> 97 >> 98 page = radix_tree_deref_slot(slot); >> 99 if (radix_tree_exception(page)) { >> 100 if (radix_tree_deref_retry(page)) { >> 101 slot = radix_tree_iter_retry(&iter); >> 102 continue; >> 103 } >> 104 >> 105 page = NULL; >> 106 } >> 107 >> 108 if (page && >> 109 page_count(page) - page_mapcount(page) != 1) { >> 110 if (scan < LAST_SCAN) >> 111 goto continue_resched; 142 112 143 if (!xa_is_value(folio << 144 memfd_folio_has_ex << 145 /* 113 /* 146 * On the last 114 * On the last scan, we clean up all those tags 147 * we inserted 115 * we inserted; but make a note that we still 148 * found folio !! 116 * found pages pinned. 149 */ 117 */ 150 if (scan == LA !! 118 error = -EBUSY; 151 error << 152 else << 153 clear << 154 } 119 } 155 if (clear) << 156 xas_clear_mark << 157 << 158 if (++latency < XA_CHE << 159 continue; << 160 latency = 0; << 161 120 162 xas_pause(&xas); !! 121 xa_lock_irq(&mapping->i_pages); 163 xas_unlock_irq(&xas); !! 122 radix_tree_tag_clear(&mapping->i_pages, 164 cond_resched(); !! 123 iter.index, MEMFD_TAG_PINNED); 165 xas_lock_irq(&xas); !! 124 xa_unlock_irq(&mapping->i_pages); >> 125 continue_resched: >> 126 if (need_resched()) { >> 127 slot = radix_tree_iter_resume(slot, &iter); >> 128 cond_resched_rcu(); >> 129 } 166 } 130 } 167 xas_unlock_irq(&xas); !! 131 rcu_read_unlock(); 168 } 132 } 169 133 170 return error; 134 return error; 171 } 135 } 172 136 173 static unsigned int *memfd_file_seals_ptr(stru 137 static unsigned int *memfd_file_seals_ptr(struct file *file) 174 { 138 { 175 if (shmem_file(file)) 139 if (shmem_file(file)) 176 return &SHMEM_I(file_inode(fil 140 return &SHMEM_I(file_inode(file))->seals; 177 141 178 #ifdef CONFIG_HUGETLBFS 142 #ifdef CONFIG_HUGETLBFS 179 if (is_file_hugepages(file)) 143 if (is_file_hugepages(file)) 180 return &HUGETLBFS_I(file_inode 144 return &HUGETLBFS_I(file_inode(file))->seals; 181 #endif 145 #endif 182 146 183 return NULL; 147 return NULL; 184 } 148 } 185 149 186 #define F_ALL_SEALS (F_SEAL_SEAL | \ 150 #define F_ALL_SEALS (F_SEAL_SEAL | \ 187 F_SEAL_EXEC | \ << 188 F_SEAL_SHRINK | \ 151 F_SEAL_SHRINK | \ 189 F_SEAL_GROW | \ 152 F_SEAL_GROW | \ 190 F_SEAL_WRITE | \ !! 153 F_SEAL_WRITE) 191 F_SEAL_FUTURE_WRITE) << 192 154 193 static int memfd_add_seals(struct file *file, 155 static int memfd_add_seals(struct file *file, unsigned int seals) 194 { 156 { 195 struct inode *inode = file_inode(file) 157 struct inode *inode = file_inode(file); 196 unsigned int *file_seals; 158 unsigned int *file_seals; 197 int error; 159 int error; 198 160 199 /* 161 /* 200 * SEALING 162 * SEALING 201 * Sealing allows multiple parties to 163 * Sealing allows multiple parties to share a tmpfs or hugetlbfs file 202 * but restrict access to a specific s 164 * but restrict access to a specific subset of file operations. Seals 203 * can only be added, but never remove 165 * can only be added, but never removed. This way, mutually untrusted 204 * parties can share common memory reg 166 * parties can share common memory regions with a well-defined policy. 205 * A malicious peer can thus never per 167 * A malicious peer can thus never perform unwanted operations on a 206 * shared object. 168 * shared object. 207 * 169 * 208 * Seals are only supported on special 170 * Seals are only supported on special tmpfs or hugetlbfs files and 209 * always affect the whole underlying 171 * always affect the whole underlying inode. Once a seal is set, it 210 * may prevent some kinds of access to 172 * may prevent some kinds of access to the file. Currently, the 211 * following seals are defined: 173 * following seals are defined: 212 * SEAL_SEAL: Prevent further seals 174 * SEAL_SEAL: Prevent further seals from being set on this file 213 * SEAL_SHRINK: Prevent the file fro 175 * SEAL_SHRINK: Prevent the file from shrinking 214 * SEAL_GROW: Prevent the file from 176 * SEAL_GROW: Prevent the file from growing 215 * SEAL_WRITE: Prevent write access 177 * SEAL_WRITE: Prevent write access to the file 216 * SEAL_EXEC: Prevent modification o << 217 * 178 * 218 * As we don't require any trust relat 179 * As we don't require any trust relationship between two parties, we 219 * must prevent seals from being remov 180 * must prevent seals from being removed. Therefore, sealing a file 220 * only adds a given set of seals to t 181 * only adds a given set of seals to the file, it never touches 221 * existing seals. Furthermore, the "s 182 * existing seals. Furthermore, the "setting seals"-operation can be 222 * sealed itself, which basically prev 183 * sealed itself, which basically prevents any further seal from being 223 * added. 184 * added. 224 * 185 * 225 * Semantics of sealing are only defin 186 * Semantics of sealing are only defined on volatile files. Only 226 * anonymous tmpfs and hugetlbfs files 187 * anonymous tmpfs and hugetlbfs files support sealing. More 227 * importantly, seals are never writte 188 * importantly, seals are never written to disk. Therefore, there's 228 * no plan to support it on other file 189 * no plan to support it on other file types. 229 */ 190 */ 230 191 231 if (!(file->f_mode & FMODE_WRITE)) 192 if (!(file->f_mode & FMODE_WRITE)) 232 return -EPERM; 193 return -EPERM; 233 if (seals & ~(unsigned int)F_ALL_SEALS 194 if (seals & ~(unsigned int)F_ALL_SEALS) 234 return -EINVAL; 195 return -EINVAL; 235 196 236 inode_lock(inode); 197 inode_lock(inode); 237 198 238 file_seals = memfd_file_seals_ptr(file 199 file_seals = memfd_file_seals_ptr(file); 239 if (!file_seals) { 200 if (!file_seals) { 240 error = -EINVAL; 201 error = -EINVAL; 241 goto unlock; 202 goto unlock; 242 } 203 } 243 204 244 if (*file_seals & F_SEAL_SEAL) { 205 if (*file_seals & F_SEAL_SEAL) { 245 error = -EPERM; 206 error = -EPERM; 246 goto unlock; 207 goto unlock; 247 } 208 } 248 209 249 if ((seals & F_SEAL_WRITE) && !(*file_ 210 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) { 250 error = mapping_deny_writable( 211 error = mapping_deny_writable(file->f_mapping); 251 if (error) 212 if (error) 252 goto unlock; 213 goto unlock; 253 214 254 error = memfd_wait_for_pins(fi 215 error = memfd_wait_for_pins(file->f_mapping); 255 if (error) { 216 if (error) { 256 mapping_allow_writable 217 mapping_allow_writable(file->f_mapping); 257 goto unlock; 218 goto unlock; 258 } 219 } 259 } 220 } 260 221 261 /* << 262 * SEAL_EXEC implys SEAL_WRITE, making << 263 */ << 264 if (seals & F_SEAL_EXEC && inode->i_mo << 265 seals |= F_SEAL_SHRINK|F_SEAL_ << 266 << 267 *file_seals |= seals; 222 *file_seals |= seals; 268 error = 0; 223 error = 0; 269 224 270 unlock: 225 unlock: 271 inode_unlock(inode); 226 inode_unlock(inode); 272 return error; 227 return error; 273 } 228 } 274 229 275 static int memfd_get_seals(struct file *file) 230 static int memfd_get_seals(struct file *file) 276 { 231 { 277 unsigned int *seals = memfd_file_seals 232 unsigned int *seals = memfd_file_seals_ptr(file); 278 233 279 return seals ? *seals : -EINVAL; 234 return seals ? *seals : -EINVAL; 280 } 235 } 281 236 282 long memfd_fcntl(struct file *file, unsigned i !! 237 long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg) 283 { 238 { 284 long error; 239 long error; 285 240 286 switch (cmd) { 241 switch (cmd) { 287 case F_ADD_SEALS: 242 case F_ADD_SEALS: >> 243 /* disallow upper 32bit */ >> 244 if (arg > UINT_MAX) >> 245 return -EINVAL; >> 246 288 error = memfd_add_seals(file, 247 error = memfd_add_seals(file, arg); 289 break; 248 break; 290 case F_GET_SEALS: 249 case F_GET_SEALS: 291 error = memfd_get_seals(file); 250 error = memfd_get_seals(file); 292 break; 251 break; 293 default: 252 default: 294 error = -EINVAL; 253 error = -EINVAL; 295 break; 254 break; 296 } 255 } 297 256 298 return error; 257 return error; 299 } 258 } 300 259 301 #define MFD_NAME_PREFIX "memfd:" 260 #define MFD_NAME_PREFIX "memfd:" 302 #define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_P 261 #define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1) 303 #define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_ 262 #define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN) 304 263 305 #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW !! 264 #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB) 306 << 307 static int check_sysctl_memfd_noexec(unsigned << 308 { << 309 #ifdef CONFIG_SYSCTL << 310 struct pid_namespace *ns = task_active << 311 int sysctl = pidns_memfd_noexec_scope( << 312 << 313 if (!(*flags & (MFD_EXEC | MFD_NOEXEC_ << 314 if (sysctl >= MEMFD_NOEXEC_SCO << 315 *flags |= MFD_NOEXEC_S << 316 else << 317 *flags |= MFD_EXEC; << 318 } << 319 << 320 if (!(*flags & MFD_NOEXEC_SEAL) && sys << 321 pr_err_ratelimited( << 322 "%s[%d]: memfd_create( << 323 current->comm, task_pi << 324 return -EACCES; << 325 } << 326 #endif << 327 return 0; << 328 } << 329 265 330 SYSCALL_DEFINE2(memfd_create, 266 SYSCALL_DEFINE2(memfd_create, 331 const char __user *, uname, 267 const char __user *, uname, 332 unsigned int, flags) 268 unsigned int, flags) 333 { 269 { 334 unsigned int *file_seals; 270 unsigned int *file_seals; 335 struct file *file; 271 struct file *file; 336 int fd, error; 272 int fd, error; 337 char *name; 273 char *name; 338 long len; 274 long len; 339 275 340 if (!(flags & MFD_HUGETLB)) { 276 if (!(flags & MFD_HUGETLB)) { 341 if (flags & ~(unsigned int)MFD 277 if (flags & ~(unsigned int)MFD_ALL_FLAGS) 342 return -EINVAL; 278 return -EINVAL; 343 } else { 279 } else { 344 /* Allow huge page size encodi 280 /* Allow huge page size encoding in flags. */ 345 if (flags & ~(unsigned int)(MF 281 if (flags & ~(unsigned int)(MFD_ALL_FLAGS | 346 (MFD_HUGE_MASK 282 (MFD_HUGE_MASK << MFD_HUGE_SHIFT))) 347 return -EINVAL; 283 return -EINVAL; 348 } 284 } 349 285 350 /* Invalid if both EXEC and NOEXEC_SEA << 351 if ((flags & MFD_EXEC) && (flags & MFD << 352 return -EINVAL; << 353 << 354 error = check_sysctl_memfd_noexec(&fla << 355 if (error < 0) << 356 return error; << 357 << 358 /* length includes terminating zero */ 286 /* length includes terminating zero */ 359 len = strnlen_user(uname, MFD_NAME_MAX 287 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1); 360 if (len <= 0) 288 if (len <= 0) 361 return -EFAULT; 289 return -EFAULT; 362 if (len > MFD_NAME_MAX_LEN + 1) 290 if (len > MFD_NAME_MAX_LEN + 1) 363 return -EINVAL; 291 return -EINVAL; 364 292 365 name = kmalloc(len + MFD_NAME_PREFIX_L 293 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL); 366 if (!name) 294 if (!name) 367 return -ENOMEM; 295 return -ENOMEM; 368 296 369 strcpy(name, MFD_NAME_PREFIX); 297 strcpy(name, MFD_NAME_PREFIX); 370 if (copy_from_user(&name[MFD_NAME_PREF 298 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) { 371 error = -EFAULT; 299 error = -EFAULT; 372 goto err_name; 300 goto err_name; 373 } 301 } 374 302 375 /* terminating-zero may have changed a 303 /* terminating-zero may have changed after strnlen_user() returned */ 376 if (name[len + MFD_NAME_PREFIX_LEN - 1 304 if (name[len + MFD_NAME_PREFIX_LEN - 1]) { 377 error = -EFAULT; 305 error = -EFAULT; 378 goto err_name; 306 goto err_name; 379 } 307 } 380 308 381 fd = get_unused_fd_flags((flags & MFD_ 309 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0); 382 if (fd < 0) { 310 if (fd < 0) { 383 error = fd; 311 error = fd; 384 goto err_name; 312 goto err_name; 385 } 313 } 386 314 387 if (flags & MFD_HUGETLB) { 315 if (flags & MFD_HUGETLB) { 388 file = hugetlb_file_setup(name !! 316 struct user_struct *user = NULL; >> 317 >> 318 file = hugetlb_file_setup(name, 0, VM_NORESERVE, &user, 389 HUGETL 319 HUGETLB_ANONHUGE_INODE, 390 (flags 320 (flags >> MFD_HUGE_SHIFT) & 391 MFD_HU 321 MFD_HUGE_MASK); 392 } else 322 } else 393 file = shmem_file_setup(name, 323 file = shmem_file_setup(name, 0, VM_NORESERVE); 394 if (IS_ERR(file)) { 324 if (IS_ERR(file)) { 395 error = PTR_ERR(file); 325 error = PTR_ERR(file); 396 goto err_fd; 326 goto err_fd; 397 } 327 } 398 file->f_mode |= FMODE_LSEEK | FMODE_PR 328 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE; 399 file->f_flags |= O_LARGEFILE; !! 329 file->f_flags |= O_RDWR | O_LARGEFILE; 400 << 401 if (flags & MFD_NOEXEC_SEAL) { << 402 struct inode *inode = file_ino << 403 330 404 inode->i_mode &= ~0111; !! 331 if (flags & MFD_ALLOW_SEALING) { 405 file_seals = memfd_file_seals_ << 406 if (file_seals) { << 407 *file_seals &= ~F_SEAL << 408 *file_seals |= F_SEAL_ << 409 } << 410 } else if (flags & MFD_ALLOW_SEALING) << 411 /* MFD_EXEC and MFD_ALLOW_SEAL << 412 file_seals = memfd_file_seals_ 332 file_seals = memfd_file_seals_ptr(file); 413 if (file_seals) !! 333 *file_seals &= ~F_SEAL_SEAL; 414 *file_seals &= ~F_SEAL << 415 } 334 } 416 335 417 fd_install(fd, file); 336 fd_install(fd, file); 418 kfree(name); 337 kfree(name); 419 return fd; 338 return fd; 420 339 421 err_fd: 340 err_fd: 422 put_unused_fd(fd); 341 put_unused_fd(fd); 423 err_name: 342 err_name: 424 kfree(name); 343 kfree(name); 425 return error; 344 return error; 426 } 345 } 427 346
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.