mm/memfd.c

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