| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * linux/fs/file.c |
| 4 | * |
| 5 | * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes |
| 6 | * |
| 7 | * Manage the dynamic fd arrays in the process files_struct. |
| 8 | */ |
| 9 | |
| 10 | #include <linux/syscalls.h> |
| 11 | #include <linux/export.h> |
| 12 | #include <linux/fs.h> |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/mm.h> |
| 15 | #include <linux/sched/signal.h> |
| 16 | #include <linux/slab.h> |
| 17 | #include <linux/file.h> |
| 18 | #include <linux/fdtable.h> |
| 19 | #include <linux/bitops.h> |
| 20 | #include <linux/spinlock.h> |
| 21 | #include <linux/rcupdate.h> |
| 22 | #include <linux/close_range.h> |
| 23 | #include <linux/file_ref.h> |
| 24 | #include <net/sock.h> |
| 25 | #include <linux/init_task.h> |
| 26 | |
| 27 | #include "internal.h" |
| 28 | |
| 29 | static noinline bool __file_ref_put_badval(file_ref_t *ref, unsigned long cnt) |
| 30 | { |
| 31 | /* |
| 32 | * If the reference count was already in the dead zone, then this |
| 33 | * put() operation is imbalanced. Warn, put the reference count back to |
| 34 | * DEAD and tell the caller to not deconstruct the object. |
| 35 | */ |
| 36 | if (WARN_ONCE(cnt >= FILE_REF_RELEASED, "imbalanced put on file reference count" )) { |
| 37 | atomic_long_set(v: &ref->refcnt, FILE_REF_DEAD); |
| 38 | return false; |
| 39 | } |
| 40 | |
| 41 | /* |
| 42 | * This is a put() operation on a saturated refcount. Restore the |
| 43 | * mean saturation value and tell the caller to not deconstruct the |
| 44 | * object. |
| 45 | */ |
| 46 | if (cnt > FILE_REF_MAXREF) |
| 47 | atomic_long_set(v: &ref->refcnt, FILE_REF_SATURATED); |
| 48 | return false; |
| 49 | } |
| 50 | |
| 51 | /** |
| 52 | * __file_ref_put - Slowpath of file_ref_put() |
| 53 | * @ref: Pointer to the reference count |
| 54 | * @cnt: Current reference count |
| 55 | * |
| 56 | * Invoked when the reference count is outside of the valid zone. |
| 57 | * |
| 58 | * Return: |
| 59 | * True if this was the last reference with no future references |
| 60 | * possible. This signals the caller that it can safely schedule the |
| 61 | * object, which is protected by the reference counter, for |
| 62 | * deconstruction. |
| 63 | * |
| 64 | * False if there are still active references or the put() raced |
| 65 | * with a concurrent get()/put() pair. Caller is not allowed to |
| 66 | * deconstruct the protected object. |
| 67 | */ |
| 68 | bool __file_ref_put(file_ref_t *ref, unsigned long cnt) |
| 69 | { |
| 70 | /* Did this drop the last reference? */ |
| 71 | if (likely(cnt == FILE_REF_NOREF)) { |
| 72 | /* |
| 73 | * Carefully try to set the reference count to FILE_REF_DEAD. |
| 74 | * |
| 75 | * This can fail if a concurrent get() operation has |
| 76 | * elevated it again or the corresponding put() even marked |
| 77 | * it dead already. Both are valid situations and do not |
| 78 | * require a retry. If this fails the caller is not |
| 79 | * allowed to deconstruct the object. |
| 80 | */ |
| 81 | if (!atomic_long_try_cmpxchg_release(v: &ref->refcnt, old: &cnt, FILE_REF_DEAD)) |
| 82 | return false; |
| 83 | |
| 84 | /* |
| 85 | * The caller can safely schedule the object for |
| 86 | * deconstruction. Provide acquire ordering. |
| 87 | */ |
| 88 | smp_acquire__after_ctrl_dep(); |
| 89 | return true; |
| 90 | } |
| 91 | |
| 92 | return __file_ref_put_badval(ref, cnt); |
| 93 | } |
| 94 | EXPORT_SYMBOL_GPL(__file_ref_put); |
| 95 | |
| 96 | unsigned int sysctl_nr_open __read_mostly = 1024*1024; |
| 97 | unsigned int sysctl_nr_open_min = BITS_PER_LONG; |
| 98 | /* our min() is unusable in constant expressions ;-/ */ |
| 99 | #define __const_min(x, y) ((x) < (y) ? (x) : (y)) |
| 100 | unsigned int sysctl_nr_open_max = |
| 101 | __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG; |
| 102 | |
| 103 | static void __free_fdtable(struct fdtable *fdt) |
| 104 | { |
| 105 | kvfree(addr: fdt->fd); |
| 106 | kvfree(addr: fdt->open_fds); |
| 107 | kfree(objp: fdt); |
| 108 | } |
| 109 | |
| 110 | static void free_fdtable_rcu(struct rcu_head *rcu) |
| 111 | { |
| 112 | __free_fdtable(container_of(rcu, struct fdtable, rcu)); |
| 113 | } |
| 114 | |
| 115 | #define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr)) |
| 116 | #define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long)) |
| 117 | |
| 118 | #define fdt_words(fdt) ((fdt)->max_fds / BITS_PER_LONG) // words in ->open_fds |
| 119 | /* |
| 120 | * Copy 'count' fd bits from the old table to the new table and clear the extra |
| 121 | * space if any. This does not copy the file pointers. Called with the files |
| 122 | * spinlock held for write. |
| 123 | */ |
| 124 | static inline void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt, |
| 125 | unsigned int copy_words) |
| 126 | { |
| 127 | unsigned int nwords = fdt_words(nfdt); |
| 128 | |
| 129 | bitmap_copy_and_extend(to: nfdt->open_fds, from: ofdt->open_fds, |
| 130 | count: copy_words * BITS_PER_LONG, size: nwords * BITS_PER_LONG); |
| 131 | bitmap_copy_and_extend(to: nfdt->close_on_exec, from: ofdt->close_on_exec, |
| 132 | count: copy_words * BITS_PER_LONG, size: nwords * BITS_PER_LONG); |
| 133 | bitmap_copy_and_extend(to: nfdt->full_fds_bits, from: ofdt->full_fds_bits, |
| 134 | count: copy_words, size: nwords); |
| 135 | } |
| 136 | |
| 137 | /* |
| 138 | * Copy all file descriptors from the old table to the new, expanded table and |
| 139 | * clear the extra space. Called with the files spinlock held for write. |
| 140 | */ |
| 141 | static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt) |
| 142 | { |
| 143 | size_t cpy, set; |
| 144 | |
| 145 | BUG_ON(nfdt->max_fds < ofdt->max_fds); |
| 146 | |
| 147 | cpy = ofdt->max_fds * sizeof(struct file *); |
| 148 | set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *); |
| 149 | memcpy(to: nfdt->fd, from: ofdt->fd, len: cpy); |
| 150 | memset(s: (char *)nfdt->fd + cpy, c: 0, n: set); |
| 151 | |
| 152 | copy_fd_bitmaps(nfdt, ofdt, fdt_words(ofdt)); |
| 153 | } |
| 154 | |
| 155 | /* |
| 156 | * Note how the fdtable bitmap allocations very much have to be a multiple of |
| 157 | * BITS_PER_LONG. This is not only because we walk those things in chunks of |
| 158 | * 'unsigned long' in some places, but simply because that is how the Linux |
| 159 | * kernel bitmaps are defined to work: they are not "bits in an array of bytes", |
| 160 | * they are very much "bits in an array of unsigned long". |
| 161 | */ |
| 162 | static struct fdtable *alloc_fdtable(unsigned int slots_wanted) |
| 163 | { |
| 164 | struct fdtable *fdt; |
| 165 | unsigned int nr; |
| 166 | void *data; |
| 167 | |
| 168 | /* |
| 169 | * Figure out how many fds we actually want to support in this fdtable. |
| 170 | * Allocation steps are keyed to the size of the fdarray, since it |
| 171 | * grows far faster than any of the other dynamic data. We try to fit |
| 172 | * the fdarray into comfortable page-tuned chunks: starting at 1024B |
| 173 | * and growing in powers of two from there on. Since we called only |
| 174 | * with slots_wanted > BITS_PER_LONG (embedded instance in files->fdtab |
| 175 | * already gives BITS_PER_LONG slots), the above boils down to |
| 176 | * 1. use the smallest power of two large enough to give us that many |
| 177 | * slots. |
| 178 | * 2. on 32bit skip 64 and 128 - the minimal capacity we want there is |
| 179 | * 256 slots (i.e. 1Kb fd array). |
| 180 | * 3. on 64bit don't skip anything, 1Kb fd array means 128 slots there |
| 181 | * and we are never going to be asked for 64 or less. |
| 182 | */ |
| 183 | if (IS_ENABLED(CONFIG_32BIT) && slots_wanted < 256) |
| 184 | nr = 256; |
| 185 | else |
| 186 | nr = roundup_pow_of_two(slots_wanted); |
| 187 | /* |
| 188 | * Note that this can drive nr *below* what we had passed if sysctl_nr_open |
| 189 | * had been set lower between the check in expand_files() and here. |
| 190 | * |
| 191 | * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise |
| 192 | * bitmaps handling below becomes unpleasant, to put it mildly... |
| 193 | */ |
| 194 | if (unlikely(nr > sysctl_nr_open)) { |
| 195 | nr = round_down(sysctl_nr_open, BITS_PER_LONG); |
| 196 | if (nr < slots_wanted) |
| 197 | return ERR_PTR(error: -EMFILE); |
| 198 | } |
| 199 | |
| 200 | /* |
| 201 | * Check if the allocation size would exceed INT_MAX. kvmalloc_array() |
| 202 | * and kvmalloc() will warn if the allocation size is greater than |
| 203 | * INT_MAX, as filp_cachep objects are not __GFP_NOWARN. |
| 204 | * |
| 205 | * This can happen when sysctl_nr_open is set to a very high value and |
| 206 | * a process tries to use a file descriptor near that limit. For example, |
| 207 | * if sysctl_nr_open is set to 1073741816 (0x3ffffff8) - which is what |
| 208 | * systemd typically sets it to - then trying to use a file descriptor |
| 209 | * close to that value will require allocating a file descriptor table |
| 210 | * that exceeds 8GB in size. |
| 211 | */ |
| 212 | if (unlikely(nr > INT_MAX / sizeof(struct file *))) |
| 213 | return ERR_PTR(error: -EMFILE); |
| 214 | |
| 215 | fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT); |
| 216 | if (!fdt) |
| 217 | goto out; |
| 218 | fdt->max_fds = nr; |
| 219 | data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT); |
| 220 | if (!data) |
| 221 | goto out_fdt; |
| 222 | fdt->fd = data; |
| 223 | |
| 224 | data = kvmalloc(max_t(size_t, |
| 225 | 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES), |
| 226 | GFP_KERNEL_ACCOUNT); |
| 227 | if (!data) |
| 228 | goto out_arr; |
| 229 | fdt->open_fds = data; |
| 230 | data += nr / BITS_PER_BYTE; |
| 231 | fdt->close_on_exec = data; |
| 232 | data += nr / BITS_PER_BYTE; |
| 233 | fdt->full_fds_bits = data; |
| 234 | |
| 235 | return fdt; |
| 236 | |
| 237 | out_arr: |
| 238 | kvfree(addr: fdt->fd); |
| 239 | out_fdt: |
| 240 | kfree(objp: fdt); |
| 241 | out: |
| 242 | return ERR_PTR(error: -ENOMEM); |
| 243 | } |
| 244 | |
| 245 | /* |
| 246 | * Expand the file descriptor table. |
| 247 | * This function will allocate a new fdtable and both fd array and fdset, of |
| 248 | * the given size. |
| 249 | * Return <0 error code on error; 0 on successful completion. |
| 250 | * The files->file_lock should be held on entry, and will be held on exit. |
| 251 | */ |
| 252 | static int expand_fdtable(struct files_struct *files, unsigned int nr) |
| 253 | __releases(files->file_lock) |
| 254 | __acquires(files->file_lock) |
| 255 | { |
| 256 | struct fdtable *new_fdt, *cur_fdt; |
| 257 | |
| 258 | spin_unlock(lock: &files->file_lock); |
| 259 | new_fdt = alloc_fdtable(slots_wanted: nr + 1); |
| 260 | |
| 261 | /* make sure all fd_install() have seen resize_in_progress |
| 262 | * or have finished their rcu_read_lock_sched() section. |
| 263 | */ |
| 264 | if (atomic_read(v: &files->count) > 1) |
| 265 | synchronize_rcu(); |
| 266 | |
| 267 | spin_lock(lock: &files->file_lock); |
| 268 | if (IS_ERR(ptr: new_fdt)) |
| 269 | return PTR_ERR(ptr: new_fdt); |
| 270 | cur_fdt = files_fdtable(files); |
| 271 | BUG_ON(nr < cur_fdt->max_fds); |
| 272 | copy_fdtable(nfdt: new_fdt, ofdt: cur_fdt); |
| 273 | rcu_assign_pointer(files->fdt, new_fdt); |
| 274 | if (cur_fdt != &files->fdtab) |
| 275 | call_rcu(head: &cur_fdt->rcu, func: free_fdtable_rcu); |
| 276 | /* coupled with smp_rmb() in fd_install() */ |
| 277 | smp_wmb(); |
| 278 | return 0; |
| 279 | } |
| 280 | |
| 281 | /* |
| 282 | * Expand files. |
| 283 | * This function will expand the file structures, if the requested size exceeds |
| 284 | * the current capacity and there is room for expansion. |
| 285 | * Return <0 error code on error; 0 on success. |
| 286 | * The files->file_lock should be held on entry, and will be held on exit. |
| 287 | */ |
| 288 | static int expand_files(struct files_struct *files, unsigned int nr) |
| 289 | __releases(files->file_lock) |
| 290 | __acquires(files->file_lock) |
| 291 | { |
| 292 | struct fdtable *fdt; |
| 293 | int error; |
| 294 | |
| 295 | repeat: |
| 296 | fdt = files_fdtable(files); |
| 297 | |
| 298 | /* Do we need to expand? */ |
| 299 | if (nr < fdt->max_fds) |
| 300 | return 0; |
| 301 | |
| 302 | if (unlikely(files->resize_in_progress)) { |
| 303 | spin_unlock(lock: &files->file_lock); |
| 304 | wait_event(files->resize_wait, !files->resize_in_progress); |
| 305 | spin_lock(lock: &files->file_lock); |
| 306 | goto repeat; |
| 307 | } |
| 308 | |
| 309 | /* Can we expand? */ |
| 310 | if (unlikely(nr >= sysctl_nr_open)) |
| 311 | return -EMFILE; |
| 312 | |
| 313 | /* All good, so we try */ |
| 314 | files->resize_in_progress = true; |
| 315 | error = expand_fdtable(files, nr); |
| 316 | files->resize_in_progress = false; |
| 317 | |
| 318 | wake_up_all(&files->resize_wait); |
| 319 | return error; |
| 320 | } |
| 321 | |
| 322 | static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt, |
| 323 | bool set) |
| 324 | { |
| 325 | if (set) { |
| 326 | __set_bit(fd, fdt->close_on_exec); |
| 327 | } else { |
| 328 | if (test_bit(fd, fdt->close_on_exec)) |
| 329 | __clear_bit(fd, fdt->close_on_exec); |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt, bool set) |
| 334 | { |
| 335 | __set_bit(fd, fdt->open_fds); |
| 336 | __set_close_on_exec(fd, fdt, set); |
| 337 | fd /= BITS_PER_LONG; |
| 338 | if (!~fdt->open_fds[fd]) |
| 339 | __set_bit(fd, fdt->full_fds_bits); |
| 340 | } |
| 341 | |
| 342 | static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt) |
| 343 | { |
| 344 | __clear_bit(fd, fdt->open_fds); |
| 345 | fd /= BITS_PER_LONG; |
| 346 | if (test_bit(fd, fdt->full_fds_bits)) |
| 347 | __clear_bit(fd, fdt->full_fds_bits); |
| 348 | } |
| 349 | |
| 350 | static inline bool fd_is_open(unsigned int fd, const struct fdtable *fdt) |
| 351 | { |
| 352 | return test_bit(fd, fdt->open_fds); |
| 353 | } |
| 354 | |
| 355 | /* |
| 356 | * Note that a sane fdtable size always has to be a multiple of |
| 357 | * BITS_PER_LONG, since we have bitmaps that are sized by this. |
| 358 | * |
| 359 | * punch_hole is optional - when close_range() is asked to unshare |
| 360 | * and close, we don't need to copy descriptors in that range, so |
| 361 | * a smaller cloned descriptor table might suffice if the last |
| 362 | * currently opened descriptor falls into that range. |
| 363 | */ |
| 364 | static unsigned int sane_fdtable_size(struct fdtable *fdt, struct fd_range *punch_hole) |
| 365 | { |
| 366 | unsigned int last = find_last_bit(addr: fdt->open_fds, size: fdt->max_fds); |
| 367 | |
| 368 | if (last == fdt->max_fds) |
| 369 | return NR_OPEN_DEFAULT; |
| 370 | if (punch_hole && punch_hole->to >= last && punch_hole->from <= last) { |
| 371 | last = find_last_bit(addr: fdt->open_fds, size: punch_hole->from); |
| 372 | if (last == punch_hole->from) |
| 373 | return NR_OPEN_DEFAULT; |
| 374 | } |
| 375 | return ALIGN(last + 1, BITS_PER_LONG); |
| 376 | } |
| 377 | |
| 378 | /* |
| 379 | * Allocate a new descriptor table and copy contents from the passed in |
| 380 | * instance. Returns a pointer to cloned table on success, ERR_PTR() |
| 381 | * on failure. For 'punch_hole' see sane_fdtable_size(). |
| 382 | */ |
| 383 | struct files_struct *dup_fd(struct files_struct *oldf, struct fd_range *punch_hole) |
| 384 | { |
| 385 | struct files_struct *newf; |
| 386 | struct file **old_fds, **new_fds; |
| 387 | unsigned int open_files, i; |
| 388 | struct fdtable *old_fdt, *new_fdt; |
| 389 | |
| 390 | newf = kmem_cache_alloc(files_cachep, GFP_KERNEL); |
| 391 | if (!newf) |
| 392 | return ERR_PTR(error: -ENOMEM); |
| 393 | |
| 394 | atomic_set(v: &newf->count, i: 1); |
| 395 | |
| 396 | spin_lock_init(&newf->file_lock); |
| 397 | newf->resize_in_progress = false; |
| 398 | init_waitqueue_head(&newf->resize_wait); |
| 399 | newf->next_fd = 0; |
| 400 | new_fdt = &newf->fdtab; |
| 401 | new_fdt->max_fds = NR_OPEN_DEFAULT; |
| 402 | new_fdt->close_on_exec = newf->close_on_exec_init; |
| 403 | new_fdt->open_fds = newf->open_fds_init; |
| 404 | new_fdt->full_fds_bits = newf->full_fds_bits_init; |
| 405 | new_fdt->fd = &newf->fd_array[0]; |
| 406 | |
| 407 | spin_lock(lock: &oldf->file_lock); |
| 408 | old_fdt = files_fdtable(oldf); |
| 409 | open_files = sane_fdtable_size(fdt: old_fdt, punch_hole); |
| 410 | |
| 411 | /* |
| 412 | * Check whether we need to allocate a larger fd array and fd set. |
| 413 | */ |
| 414 | while (unlikely(open_files > new_fdt->max_fds)) { |
| 415 | spin_unlock(lock: &oldf->file_lock); |
| 416 | |
| 417 | if (new_fdt != &newf->fdtab) |
| 418 | __free_fdtable(fdt: new_fdt); |
| 419 | |
| 420 | new_fdt = alloc_fdtable(slots_wanted: open_files); |
| 421 | if (IS_ERR(ptr: new_fdt)) { |
| 422 | kmem_cache_free(s: files_cachep, objp: newf); |
| 423 | return ERR_CAST(ptr: new_fdt); |
| 424 | } |
| 425 | |
| 426 | /* |
| 427 | * Reacquire the oldf lock and a pointer to its fd table |
| 428 | * who knows it may have a new bigger fd table. We need |
| 429 | * the latest pointer. |
| 430 | */ |
| 431 | spin_lock(lock: &oldf->file_lock); |
| 432 | old_fdt = files_fdtable(oldf); |
| 433 | open_files = sane_fdtable_size(fdt: old_fdt, punch_hole); |
| 434 | } |
| 435 | |
| 436 | copy_fd_bitmaps(nfdt: new_fdt, ofdt: old_fdt, copy_words: open_files / BITS_PER_LONG); |
| 437 | |
| 438 | old_fds = old_fdt->fd; |
| 439 | new_fds = new_fdt->fd; |
| 440 | |
| 441 | /* |
| 442 | * We may be racing against fd allocation from other threads using this |
| 443 | * files_struct, despite holding ->file_lock. |
| 444 | * |
| 445 | * alloc_fd() might have already claimed a slot, while fd_install() |
| 446 | * did not populate it yet. Note the latter operates locklessly, so |
| 447 | * the file can show up as we are walking the array below. |
| 448 | * |
| 449 | * At the same time we know no files will disappear as all other |
| 450 | * operations take the lock. |
| 451 | * |
| 452 | * Instead of trying to placate userspace racing with itself, we |
| 453 | * ref the file if we see it and mark the fd slot as unused otherwise. |
| 454 | */ |
| 455 | for (i = open_files; i != 0; i--) { |
| 456 | struct file *f = rcu_dereference_raw(*old_fds++); |
| 457 | if (f) { |
| 458 | get_file(f); |
| 459 | } else { |
| 460 | __clear_open_fd(fd: open_files - i, fdt: new_fdt); |
| 461 | } |
| 462 | rcu_assign_pointer(*new_fds++, f); |
| 463 | } |
| 464 | spin_unlock(lock: &oldf->file_lock); |
| 465 | |
| 466 | /* clear the remainder */ |
| 467 | memset(s: new_fds, c: 0, n: (new_fdt->max_fds - open_files) * sizeof(struct file *)); |
| 468 | |
| 469 | rcu_assign_pointer(newf->fdt, new_fdt); |
| 470 | |
| 471 | return newf; |
| 472 | } |
| 473 | |
| 474 | static struct fdtable *close_files(struct files_struct * files) |
| 475 | { |
| 476 | /* |
| 477 | * It is safe to dereference the fd table without RCU or |
| 478 | * ->file_lock because this is the last reference to the |
| 479 | * files structure. |
| 480 | */ |
| 481 | struct fdtable *fdt = rcu_dereference_raw(files->fdt); |
| 482 | unsigned int i, j = 0; |
| 483 | |
| 484 | for (;;) { |
| 485 | unsigned long set; |
| 486 | i = j * BITS_PER_LONG; |
| 487 | if (i >= fdt->max_fds) |
| 488 | break; |
| 489 | set = fdt->open_fds[j++]; |
| 490 | while (set) { |
| 491 | if (set & 1) { |
| 492 | struct file *file = fdt->fd[i]; |
| 493 | if (file) { |
| 494 | filp_close(file, id: files); |
| 495 | cond_resched(); |
| 496 | } |
| 497 | } |
| 498 | i++; |
| 499 | set >>= 1; |
| 500 | } |
| 501 | } |
| 502 | |
| 503 | return fdt; |
| 504 | } |
| 505 | |
| 506 | void put_files_struct(struct files_struct *files) |
| 507 | { |
| 508 | if (atomic_dec_and_test(v: &files->count)) { |
| 509 | struct fdtable *fdt = close_files(files); |
| 510 | |
| 511 | /* free the arrays if they are not embedded */ |
| 512 | if (fdt != &files->fdtab) |
| 513 | __free_fdtable(fdt); |
| 514 | kmem_cache_free(s: files_cachep, objp: files); |
| 515 | } |
| 516 | } |
| 517 | |
| 518 | void exit_files(struct task_struct *tsk) |
| 519 | { |
| 520 | struct files_struct * files = tsk->files; |
| 521 | |
| 522 | if (files) { |
| 523 | task_lock(p: tsk); |
| 524 | tsk->files = NULL; |
| 525 | task_unlock(p: tsk); |
| 526 | put_files_struct(files); |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | struct files_struct init_files = { |
| 531 | .count = ATOMIC_INIT(1), |
| 532 | .fdt = &init_files.fdtab, |
| 533 | .fdtab = { |
| 534 | .max_fds = NR_OPEN_DEFAULT, |
| 535 | .fd = &init_files.fd_array[0], |
| 536 | .close_on_exec = init_files.close_on_exec_init, |
| 537 | .open_fds = init_files.open_fds_init, |
| 538 | .full_fds_bits = init_files.full_fds_bits_init, |
| 539 | }, |
| 540 | .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock), |
| 541 | .resize_wait = __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait), |
| 542 | }; |
| 543 | |
| 544 | static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start) |
| 545 | { |
| 546 | unsigned int maxfd = fdt->max_fds; /* always multiple of BITS_PER_LONG */ |
| 547 | unsigned int maxbit = maxfd / BITS_PER_LONG; |
| 548 | unsigned int bitbit = start / BITS_PER_LONG; |
| 549 | unsigned int bit; |
| 550 | |
| 551 | /* |
| 552 | * Try to avoid looking at the second level bitmap |
| 553 | */ |
| 554 | bit = find_next_zero_bit(addr: &fdt->open_fds[bitbit], BITS_PER_LONG, |
| 555 | offset: start & (BITS_PER_LONG - 1)); |
| 556 | if (bit < BITS_PER_LONG) |
| 557 | return bit + bitbit * BITS_PER_LONG; |
| 558 | |
| 559 | bitbit = find_next_zero_bit(addr: fdt->full_fds_bits, size: maxbit, offset: bitbit) * BITS_PER_LONG; |
| 560 | if (bitbit >= maxfd) |
| 561 | return maxfd; |
| 562 | if (bitbit > start) |
| 563 | start = bitbit; |
| 564 | return find_next_zero_bit(addr: fdt->open_fds, size: maxfd, offset: start); |
| 565 | } |
| 566 | |
| 567 | /* |
| 568 | * allocate a file descriptor, mark it busy. |
| 569 | */ |
| 570 | static int alloc_fd(unsigned start, unsigned end, unsigned flags) |
| 571 | { |
| 572 | struct files_struct *files = current->files; |
| 573 | unsigned int fd; |
| 574 | int error; |
| 575 | struct fdtable *fdt; |
| 576 | |
| 577 | spin_lock(lock: &files->file_lock); |
| 578 | repeat: |
| 579 | fdt = files_fdtable(files); |
| 580 | fd = start; |
| 581 | if (fd < files->next_fd) |
| 582 | fd = files->next_fd; |
| 583 | |
| 584 | if (likely(fd < fdt->max_fds)) |
| 585 | fd = find_next_fd(fdt, start: fd); |
| 586 | |
| 587 | /* |
| 588 | * N.B. For clone tasks sharing a files structure, this test |
| 589 | * will limit the total number of files that can be opened. |
| 590 | */ |
| 591 | error = -EMFILE; |
| 592 | if (unlikely(fd >= end)) |
| 593 | goto out; |
| 594 | |
| 595 | if (unlikely(fd >= fdt->max_fds)) { |
| 596 | error = expand_files(files, nr: fd); |
| 597 | if (error < 0) |
| 598 | goto out; |
| 599 | |
| 600 | goto repeat; |
| 601 | } |
| 602 | |
| 603 | if (start <= files->next_fd) |
| 604 | files->next_fd = fd + 1; |
| 605 | |
| 606 | __set_open_fd(fd, fdt, set: flags & O_CLOEXEC); |
| 607 | error = fd; |
| 608 | VFS_BUG_ON(rcu_access_pointer(fdt->fd[fd]) != NULL); |
| 609 | |
| 610 | out: |
| 611 | spin_unlock(lock: &files->file_lock); |
| 612 | return error; |
| 613 | } |
| 614 | |
| 615 | int __get_unused_fd_flags(unsigned flags, unsigned long nofile) |
| 616 | { |
| 617 | return alloc_fd(start: 0, end: nofile, flags); |
| 618 | } |
| 619 | |
| 620 | int get_unused_fd_flags(unsigned flags) |
| 621 | { |
| 622 | return __get_unused_fd_flags(flags, nofile: rlimit(RLIMIT_NOFILE)); |
| 623 | } |
| 624 | EXPORT_SYMBOL(get_unused_fd_flags); |
| 625 | |
| 626 | static void __put_unused_fd(struct files_struct *files, unsigned int fd) |
| 627 | { |
| 628 | struct fdtable *fdt = files_fdtable(files); |
| 629 | __clear_open_fd(fd, fdt); |
| 630 | if (fd < files->next_fd) |
| 631 | files->next_fd = fd; |
| 632 | } |
| 633 | |
| 634 | void put_unused_fd(unsigned int fd) |
| 635 | { |
| 636 | struct files_struct *files = current->files; |
| 637 | spin_lock(lock: &files->file_lock); |
| 638 | __put_unused_fd(files, fd); |
| 639 | spin_unlock(lock: &files->file_lock); |
| 640 | } |
| 641 | |
| 642 | EXPORT_SYMBOL(put_unused_fd); |
| 643 | |
| 644 | /** |
| 645 | * fd_install - install a file pointer in the fd array |
| 646 | * @fd: file descriptor to install the file in |
| 647 | * @file: the file to install |
| 648 | * |
| 649 | * This consumes the "file" refcount, so callers should treat it |
| 650 | * as if they had called fput(file). |
| 651 | */ |
| 652 | void fd_install(unsigned int fd, struct file *file) |
| 653 | { |
| 654 | struct files_struct *files = current->files; |
| 655 | struct fdtable *fdt; |
| 656 | |
| 657 | if (WARN_ON_ONCE(unlikely(file->f_mode & FMODE_BACKING))) |
| 658 | return; |
| 659 | |
| 660 | rcu_read_lock_sched(); |
| 661 | |
| 662 | if (unlikely(files->resize_in_progress)) { |
| 663 | rcu_read_unlock_sched(); |
| 664 | spin_lock(lock: &files->file_lock); |
| 665 | fdt = files_fdtable(files); |
| 666 | VFS_BUG_ON(rcu_access_pointer(fdt->fd[fd]) != NULL); |
| 667 | rcu_assign_pointer(fdt->fd[fd], file); |
| 668 | spin_unlock(lock: &files->file_lock); |
| 669 | return; |
| 670 | } |
| 671 | /* coupled with smp_wmb() in expand_fdtable() */ |
| 672 | smp_rmb(); |
| 673 | fdt = rcu_dereference_sched(files->fdt); |
| 674 | VFS_BUG_ON(rcu_access_pointer(fdt->fd[fd]) != NULL); |
| 675 | rcu_assign_pointer(fdt->fd[fd], file); |
| 676 | rcu_read_unlock_sched(); |
| 677 | } |
| 678 | |
| 679 | EXPORT_SYMBOL(fd_install); |
| 680 | |
| 681 | /** |
| 682 | * file_close_fd_locked - return file associated with fd |
| 683 | * @files: file struct to retrieve file from |
| 684 | * @fd: file descriptor to retrieve file for |
| 685 | * |
| 686 | * Doesn't take a separate reference count. |
| 687 | * |
| 688 | * Context: files_lock must be held. |
| 689 | * |
| 690 | * Returns: The file associated with @fd (NULL if @fd is not open) |
| 691 | */ |
| 692 | struct file *file_close_fd_locked(struct files_struct *files, unsigned fd) |
| 693 | { |
| 694 | struct fdtable *fdt = files_fdtable(files); |
| 695 | struct file *file; |
| 696 | |
| 697 | lockdep_assert_held(&files->file_lock); |
| 698 | |
| 699 | if (fd >= fdt->max_fds) |
| 700 | return NULL; |
| 701 | |
| 702 | fd = array_index_nospec(fd, fdt->max_fds); |
| 703 | file = rcu_dereference_raw(fdt->fd[fd]); |
| 704 | if (file) { |
| 705 | rcu_assign_pointer(fdt->fd[fd], NULL); |
| 706 | __put_unused_fd(files, fd); |
| 707 | } |
| 708 | return file; |
| 709 | } |
| 710 | |
| 711 | int close_fd(unsigned fd) |
| 712 | { |
| 713 | struct files_struct *files = current->files; |
| 714 | struct file *file; |
| 715 | |
| 716 | spin_lock(lock: &files->file_lock); |
| 717 | file = file_close_fd_locked(files, fd); |
| 718 | spin_unlock(lock: &files->file_lock); |
| 719 | if (!file) |
| 720 | return -EBADF; |
| 721 | |
| 722 | return filp_close(file, id: files); |
| 723 | } |
| 724 | EXPORT_SYMBOL(close_fd); |
| 725 | |
| 726 | /** |
| 727 | * last_fd - return last valid index into fd table |
| 728 | * @fdt: File descriptor table. |
| 729 | * |
| 730 | * Context: Either rcu read lock or files_lock must be held. |
| 731 | * |
| 732 | * Returns: Last valid index into fdtable. |
| 733 | */ |
| 734 | static inline unsigned last_fd(struct fdtable *fdt) |
| 735 | { |
| 736 | return fdt->max_fds - 1; |
| 737 | } |
| 738 | |
| 739 | static inline void __range_cloexec(struct files_struct *cur_fds, |
| 740 | unsigned int fd, unsigned int max_fd) |
| 741 | { |
| 742 | struct fdtable *fdt; |
| 743 | |
| 744 | /* make sure we're using the correct maximum value */ |
| 745 | spin_lock(lock: &cur_fds->file_lock); |
| 746 | fdt = files_fdtable(cur_fds); |
| 747 | max_fd = min(last_fd(fdt), max_fd); |
| 748 | if (fd <= max_fd) |
| 749 | bitmap_set(map: fdt->close_on_exec, start: fd, nbits: max_fd - fd + 1); |
| 750 | spin_unlock(lock: &cur_fds->file_lock); |
| 751 | } |
| 752 | |
| 753 | static inline void __range_close(struct files_struct *files, unsigned int fd, |
| 754 | unsigned int max_fd) |
| 755 | { |
| 756 | struct file *file; |
| 757 | unsigned n; |
| 758 | |
| 759 | spin_lock(lock: &files->file_lock); |
| 760 | n = last_fd(files_fdtable(files)); |
| 761 | max_fd = min(max_fd, n); |
| 762 | |
| 763 | for (; fd <= max_fd; fd++) { |
| 764 | file = file_close_fd_locked(files, fd); |
| 765 | if (file) { |
| 766 | spin_unlock(lock: &files->file_lock); |
| 767 | filp_close(file, id: files); |
| 768 | cond_resched(); |
| 769 | spin_lock(lock: &files->file_lock); |
| 770 | } else if (need_resched()) { |
| 771 | spin_unlock(lock: &files->file_lock); |
| 772 | cond_resched(); |
| 773 | spin_lock(lock: &files->file_lock); |
| 774 | } |
| 775 | } |
| 776 | spin_unlock(lock: &files->file_lock); |
| 777 | } |
| 778 | |
| 779 | /** |
| 780 | * sys_close_range() - Close all file descriptors in a given range. |
| 781 | * |
| 782 | * @fd: starting file descriptor to close |
| 783 | * @max_fd: last file descriptor to close |
| 784 | * @flags: CLOSE_RANGE flags. |
| 785 | * |
| 786 | * This closes a range of file descriptors. All file descriptors |
| 787 | * from @fd up to and including @max_fd are closed. |
| 788 | * Currently, errors to close a given file descriptor are ignored. |
| 789 | */ |
| 790 | SYSCALL_DEFINE3(close_range, unsigned int, fd, unsigned int, max_fd, |
| 791 | unsigned int, flags) |
| 792 | { |
| 793 | struct task_struct *me = current; |
| 794 | struct files_struct *cur_fds = me->files, *fds = NULL; |
| 795 | |
| 796 | if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC)) |
| 797 | return -EINVAL; |
| 798 | |
| 799 | if (fd > max_fd) |
| 800 | return -EINVAL; |
| 801 | |
| 802 | if ((flags & CLOSE_RANGE_UNSHARE) && atomic_read(v: &cur_fds->count) > 1) { |
| 803 | struct fd_range range = {fd, max_fd}, *punch_hole = ⦥ |
| 804 | |
| 805 | /* |
| 806 | * If the caller requested all fds to be made cloexec we always |
| 807 | * copy all of the file descriptors since they still want to |
| 808 | * use them. |
| 809 | */ |
| 810 | if (flags & CLOSE_RANGE_CLOEXEC) |
| 811 | punch_hole = NULL; |
| 812 | |
| 813 | fds = dup_fd(oldf: cur_fds, punch_hole); |
| 814 | if (IS_ERR(ptr: fds)) |
| 815 | return PTR_ERR(ptr: fds); |
| 816 | /* |
| 817 | * We used to share our file descriptor table, and have now |
| 818 | * created a private one, make sure we're using it below. |
| 819 | */ |
| 820 | swap(cur_fds, fds); |
| 821 | } |
| 822 | |
| 823 | if (flags & CLOSE_RANGE_CLOEXEC) |
| 824 | __range_cloexec(cur_fds, fd, max_fd); |
| 825 | else |
| 826 | __range_close(files: cur_fds, fd, max_fd); |
| 827 | |
| 828 | if (fds) { |
| 829 | /* |
| 830 | * We're done closing the files we were supposed to. Time to install |
| 831 | * the new file descriptor table and drop the old one. |
| 832 | */ |
| 833 | task_lock(p: me); |
| 834 | me->files = cur_fds; |
| 835 | task_unlock(p: me); |
| 836 | put_files_struct(files: fds); |
| 837 | } |
| 838 | |
| 839 | return 0; |
| 840 | } |
| 841 | |
| 842 | /** |
| 843 | * file_close_fd - return file associated with fd |
| 844 | * @fd: file descriptor to retrieve file for |
| 845 | * |
| 846 | * Doesn't take a separate reference count. |
| 847 | * |
| 848 | * Returns: The file associated with @fd (NULL if @fd is not open) |
| 849 | */ |
| 850 | struct file *file_close_fd(unsigned int fd) |
| 851 | { |
| 852 | struct files_struct *files = current->files; |
| 853 | struct file *file; |
| 854 | |
| 855 | spin_lock(lock: &files->file_lock); |
| 856 | file = file_close_fd_locked(files, fd); |
| 857 | spin_unlock(lock: &files->file_lock); |
| 858 | |
| 859 | return file; |
| 860 | } |
| 861 | |
| 862 | void do_close_on_exec(struct files_struct *files) |
| 863 | { |
| 864 | unsigned i; |
| 865 | struct fdtable *fdt; |
| 866 | |
| 867 | /* exec unshares first */ |
| 868 | spin_lock(lock: &files->file_lock); |
| 869 | for (i = 0; ; i++) { |
| 870 | unsigned long set; |
| 871 | unsigned fd = i * BITS_PER_LONG; |
| 872 | fdt = files_fdtable(files); |
| 873 | if (fd >= fdt->max_fds) |
| 874 | break; |
| 875 | set = fdt->close_on_exec[i]; |
| 876 | if (!set) |
| 877 | continue; |
| 878 | fdt->close_on_exec[i] = 0; |
| 879 | for ( ; set ; fd++, set >>= 1) { |
| 880 | struct file *file; |
| 881 | if (!(set & 1)) |
| 882 | continue; |
| 883 | file = fdt->fd[fd]; |
| 884 | if (!file) |
| 885 | continue; |
| 886 | rcu_assign_pointer(fdt->fd[fd], NULL); |
| 887 | __put_unused_fd(files, fd); |
| 888 | spin_unlock(lock: &files->file_lock); |
| 889 | filp_close(file, id: files); |
| 890 | cond_resched(); |
| 891 | spin_lock(lock: &files->file_lock); |
| 892 | } |
| 893 | |
| 894 | } |
| 895 | spin_unlock(lock: &files->file_lock); |
| 896 | } |
| 897 | |
| 898 | static struct file *__get_file_rcu(struct file __rcu **f) |
| 899 | { |
| 900 | struct file __rcu *file; |
| 901 | struct file __rcu *file_reloaded; |
| 902 | struct file __rcu *file_reloaded_cmp; |
| 903 | |
| 904 | file = rcu_dereference_raw(*f); |
| 905 | if (!file) |
| 906 | return NULL; |
| 907 | |
| 908 | if (unlikely(!file_ref_get(&file->f_ref))) |
| 909 | return ERR_PTR(error: -EAGAIN); |
| 910 | |
| 911 | file_reloaded = rcu_dereference_raw(*f); |
| 912 | |
| 913 | /* |
| 914 | * Ensure that all accesses have a dependency on the load from |
| 915 | * rcu_dereference_raw() above so we get correct ordering |
| 916 | * between reuse/allocation and the pointer check below. |
| 917 | */ |
| 918 | file_reloaded_cmp = file_reloaded; |
| 919 | OPTIMIZER_HIDE_VAR(file_reloaded_cmp); |
| 920 | |
| 921 | /* |
| 922 | * file_ref_get() above provided a full memory barrier when we |
| 923 | * acquired a reference. |
| 924 | * |
| 925 | * This is paired with the write barrier from assigning to the |
| 926 | * __rcu protected file pointer so that if that pointer still |
| 927 | * matches the current file, we know we have successfully |
| 928 | * acquired a reference to the right file. |
| 929 | * |
| 930 | * If the pointers don't match the file has been reallocated by |
| 931 | * SLAB_TYPESAFE_BY_RCU. |
| 932 | */ |
| 933 | if (file == file_reloaded_cmp) |
| 934 | return file_reloaded; |
| 935 | |
| 936 | fput(file); |
| 937 | return ERR_PTR(error: -EAGAIN); |
| 938 | } |
| 939 | |
| 940 | /** |
| 941 | * get_file_rcu - try go get a reference to a file under rcu |
| 942 | * @f: the file to get a reference on |
| 943 | * |
| 944 | * This function tries to get a reference on @f carefully verifying that |
| 945 | * @f hasn't been reused. |
| 946 | * |
| 947 | * This function should rarely have to be used and only by users who |
| 948 | * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it. |
| 949 | * |
| 950 | * Return: Returns @f with the reference count increased or NULL. |
| 951 | */ |
| 952 | struct file *get_file_rcu(struct file __rcu **f) |
| 953 | { |
| 954 | for (;;) { |
| 955 | struct file __rcu *file; |
| 956 | |
| 957 | file = __get_file_rcu(f); |
| 958 | if (!IS_ERR(ptr: file)) |
| 959 | return file; |
| 960 | } |
| 961 | } |
| 962 | EXPORT_SYMBOL_GPL(get_file_rcu); |
| 963 | |
| 964 | /** |
| 965 | * get_file_active - try go get a reference to a file |
| 966 | * @f: the file to get a reference on |
| 967 | * |
| 968 | * In contast to get_file_rcu() the pointer itself isn't part of the |
| 969 | * reference counting. |
| 970 | * |
| 971 | * This function should rarely have to be used and only by users who |
| 972 | * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it. |
| 973 | * |
| 974 | * Return: Returns @f with the reference count increased or NULL. |
| 975 | */ |
| 976 | struct file *get_file_active(struct file **f) |
| 977 | { |
| 978 | struct file __rcu *file; |
| 979 | |
| 980 | rcu_read_lock(); |
| 981 | file = __get_file_rcu(f); |
| 982 | rcu_read_unlock(); |
| 983 | if (IS_ERR(ptr: file)) |
| 984 | file = NULL; |
| 985 | return file; |
| 986 | } |
| 987 | EXPORT_SYMBOL_GPL(get_file_active); |
| 988 | |
| 989 | static inline struct file *__fget_files_rcu(struct files_struct *files, |
| 990 | unsigned int fd, fmode_t mask) |
| 991 | { |
| 992 | for (;;) { |
| 993 | struct file *file; |
| 994 | struct fdtable *fdt = rcu_dereference_raw(files->fdt); |
| 995 | struct file __rcu **fdentry; |
| 996 | unsigned long nospec_mask; |
| 997 | |
| 998 | /* Mask is a 0 for invalid fd's, ~0 for valid ones */ |
| 999 | nospec_mask = array_index_mask_nospec(fd, fdt->max_fds); |
| 1000 | |
| 1001 | /* |
| 1002 | * fdentry points to the 'fd' offset, or fdt->fd[0]. |
| 1003 | * Loading from fdt->fd[0] is always safe, because the |
| 1004 | * array always exists. |
| 1005 | */ |
| 1006 | fdentry = fdt->fd + (fd & nospec_mask); |
| 1007 | |
| 1008 | /* Do the load, then mask any invalid result */ |
| 1009 | file = rcu_dereference_raw(*fdentry); |
| 1010 | file = (void *)(nospec_mask & (unsigned long)file); |
| 1011 | if (unlikely(!file)) |
| 1012 | return NULL; |
| 1013 | |
| 1014 | /* |
| 1015 | * Ok, we have a file pointer that was valid at |
| 1016 | * some point, but it might have become stale since. |
| 1017 | * |
| 1018 | * We need to confirm it by incrementing the refcount |
| 1019 | * and then check the lookup again. |
| 1020 | * |
| 1021 | * file_ref_get() gives us a full memory barrier. We |
| 1022 | * only really need an 'acquire' one to protect the |
| 1023 | * loads below, but we don't have that. |
| 1024 | */ |
| 1025 | if (unlikely(!file_ref_get(&file->f_ref))) |
| 1026 | continue; |
| 1027 | |
| 1028 | /* |
| 1029 | * Such a race can take two forms: |
| 1030 | * |
| 1031 | * (a) the file ref already went down to zero and the |
| 1032 | * file hasn't been reused yet or the file count |
| 1033 | * isn't zero but the file has already been reused. |
| 1034 | * |
| 1035 | * (b) the file table entry has changed under us. |
| 1036 | * Note that we don't need to re-check the 'fdt->fd' |
| 1037 | * pointer having changed, because it always goes |
| 1038 | * hand-in-hand with 'fdt'. |
| 1039 | * |
| 1040 | * If so, we need to put our ref and try again. |
| 1041 | */ |
| 1042 | if (unlikely(file != rcu_dereference_raw(*fdentry)) || |
| 1043 | unlikely(rcu_dereference_raw(files->fdt) != fdt)) { |
| 1044 | fput(file); |
| 1045 | continue; |
| 1046 | } |
| 1047 | |
| 1048 | /* |
| 1049 | * This isn't the file we're looking for or we're not |
| 1050 | * allowed to get a reference to it. |
| 1051 | */ |
| 1052 | if (unlikely(file->f_mode & mask)) { |
| 1053 | fput(file); |
| 1054 | return NULL; |
| 1055 | } |
| 1056 | |
| 1057 | /* |
| 1058 | * Ok, we have a ref to the file, and checked that it |
| 1059 | * still exists. |
| 1060 | */ |
| 1061 | return file; |
| 1062 | } |
| 1063 | } |
| 1064 | |
| 1065 | static struct file *__fget_files(struct files_struct *files, unsigned int fd, |
| 1066 | fmode_t mask) |
| 1067 | { |
| 1068 | struct file *file; |
| 1069 | |
| 1070 | rcu_read_lock(); |
| 1071 | file = __fget_files_rcu(files, fd, mask); |
| 1072 | rcu_read_unlock(); |
| 1073 | |
| 1074 | return file; |
| 1075 | } |
| 1076 | |
| 1077 | static inline struct file *__fget(unsigned int fd, fmode_t mask) |
| 1078 | { |
| 1079 | return __fget_files(current->files, fd, mask); |
| 1080 | } |
| 1081 | |
| 1082 | struct file *fget(unsigned int fd) |
| 1083 | { |
| 1084 | return __fget(fd, FMODE_PATH); |
| 1085 | } |
| 1086 | EXPORT_SYMBOL(fget); |
| 1087 | |
| 1088 | struct file *fget_raw(unsigned int fd) |
| 1089 | { |
| 1090 | return __fget(fd, mask: 0); |
| 1091 | } |
| 1092 | EXPORT_SYMBOL(fget_raw); |
| 1093 | |
| 1094 | struct file *fget_task(struct task_struct *task, unsigned int fd) |
| 1095 | { |
| 1096 | struct file *file = NULL; |
| 1097 | |
| 1098 | task_lock(p: task); |
| 1099 | if (task->files) |
| 1100 | file = __fget_files(files: task->files, fd, mask: 0); |
| 1101 | task_unlock(p: task); |
| 1102 | |
| 1103 | return file; |
| 1104 | } |
| 1105 | |
| 1106 | struct file *fget_task_next(struct task_struct *task, unsigned int *ret_fd) |
| 1107 | { |
| 1108 | /* Must be called with rcu_read_lock held */ |
| 1109 | struct files_struct *files; |
| 1110 | unsigned int fd = *ret_fd; |
| 1111 | struct file *file = NULL; |
| 1112 | |
| 1113 | task_lock(p: task); |
| 1114 | files = task->files; |
| 1115 | if (files) { |
| 1116 | rcu_read_lock(); |
| 1117 | for (; fd < files_fdtable(files)->max_fds; fd++) { |
| 1118 | file = __fget_files_rcu(files, fd, mask: 0); |
| 1119 | if (file) |
| 1120 | break; |
| 1121 | } |
| 1122 | rcu_read_unlock(); |
| 1123 | } |
| 1124 | task_unlock(p: task); |
| 1125 | *ret_fd = fd; |
| 1126 | return file; |
| 1127 | } |
| 1128 | EXPORT_SYMBOL(fget_task_next); |
| 1129 | |
| 1130 | /* |
| 1131 | * Lightweight file lookup - no refcnt increment if fd table isn't shared. |
| 1132 | * |
| 1133 | * You can use this instead of fget if you satisfy all of the following |
| 1134 | * conditions: |
| 1135 | * 1) You must call fput_light before exiting the syscall and returning control |
| 1136 | * to userspace (i.e. you cannot remember the returned struct file * after |
| 1137 | * returning to userspace). |
| 1138 | * 2) You must not call filp_close on the returned struct file * in between |
| 1139 | * calls to fget_light and fput_light. |
| 1140 | * 3) You must not clone the current task in between the calls to fget_light |
| 1141 | * and fput_light. |
| 1142 | * |
| 1143 | * The fput_needed flag returned by fget_light should be passed to the |
| 1144 | * corresponding fput_light. |
| 1145 | * |
| 1146 | * (As an exception to rule 2, you can call filp_close between fget_light and |
| 1147 | * fput_light provided that you capture a real refcount with get_file before |
| 1148 | * the call to filp_close, and ensure that this real refcount is fput *after* |
| 1149 | * the fput_light call.) |
| 1150 | * |
| 1151 | * See also the documentation in rust/kernel/file.rs. |
| 1152 | */ |
| 1153 | static inline struct fd __fget_light(unsigned int fd, fmode_t mask) |
| 1154 | { |
| 1155 | struct files_struct *files = current->files; |
| 1156 | struct file *file; |
| 1157 | |
| 1158 | /* |
| 1159 | * If another thread is concurrently calling close_fd() followed |
| 1160 | * by put_files_struct(), we must not observe the old table |
| 1161 | * entry combined with the new refcount - otherwise we could |
| 1162 | * return a file that is concurrently being freed. |
| 1163 | * |
| 1164 | * atomic_read_acquire() pairs with atomic_dec_and_test() in |
| 1165 | * put_files_struct(). |
| 1166 | */ |
| 1167 | if (likely(atomic_read_acquire(&files->count) == 1)) { |
| 1168 | file = files_lookup_fd_raw(files, fd); |
| 1169 | if (!file || unlikely(file->f_mode & mask)) |
| 1170 | return EMPTY_FD; |
| 1171 | return BORROWED_FD(file); |
| 1172 | } else { |
| 1173 | file = __fget_files(files, fd, mask); |
| 1174 | if (!file) |
| 1175 | return EMPTY_FD; |
| 1176 | return CLONED_FD(file); |
| 1177 | } |
| 1178 | } |
| 1179 | struct fd fdget(unsigned int fd) |
| 1180 | { |
| 1181 | return __fget_light(fd, FMODE_PATH); |
| 1182 | } |
| 1183 | EXPORT_SYMBOL(fdget); |
| 1184 | |
| 1185 | struct fd fdget_raw(unsigned int fd) |
| 1186 | { |
| 1187 | return __fget_light(fd, mask: 0); |
| 1188 | } |
| 1189 | |
| 1190 | /* |
| 1191 | * Try to avoid f_pos locking. We only need it if the |
| 1192 | * file is marked for FMODE_ATOMIC_POS, and it can be |
| 1193 | * accessed multiple ways. |
| 1194 | * |
| 1195 | * Always do it for directories, because pidfd_getfd() |
| 1196 | * can make a file accessible even if it otherwise would |
| 1197 | * not be, and for directories this is a correctness |
| 1198 | * issue, not a "POSIX requirement". |
| 1199 | */ |
| 1200 | static inline bool file_needs_f_pos_lock(struct file *file) |
| 1201 | { |
| 1202 | if (!(file->f_mode & FMODE_ATOMIC_POS)) |
| 1203 | return false; |
| 1204 | if (__file_ref_read_raw(ref: &file->f_ref) != FILE_REF_ONEREF) |
| 1205 | return true; |
| 1206 | if (file->f_op->iterate_shared) |
| 1207 | return true; |
| 1208 | return false; |
| 1209 | } |
| 1210 | |
| 1211 | bool file_seek_cur_needs_f_lock(struct file *file) |
| 1212 | { |
| 1213 | if (!(file->f_mode & FMODE_ATOMIC_POS) && !file->f_op->iterate_shared) |
| 1214 | return false; |
| 1215 | |
| 1216 | /* |
| 1217 | * Note that we are not guaranteed to be called after fdget_pos() on |
| 1218 | * this file obj, in which case the caller is expected to provide the |
| 1219 | * appropriate locking. |
| 1220 | */ |
| 1221 | |
| 1222 | return true; |
| 1223 | } |
| 1224 | |
| 1225 | struct fd fdget_pos(unsigned int fd) |
| 1226 | { |
| 1227 | struct fd f = fdget(fd); |
| 1228 | struct file *file = fd_file(f); |
| 1229 | |
| 1230 | if (likely(file) && file_needs_f_pos_lock(file)) { |
| 1231 | f.word |= FDPUT_POS_UNLOCK; |
| 1232 | mutex_lock(lock: &file->f_pos_lock); |
| 1233 | } |
| 1234 | return f; |
| 1235 | } |
| 1236 | |
| 1237 | void __f_unlock_pos(struct file *f) |
| 1238 | { |
| 1239 | mutex_unlock(lock: &f->f_pos_lock); |
| 1240 | } |
| 1241 | |
| 1242 | /* |
| 1243 | * We only lock f_pos if we have threads or if the file might be |
| 1244 | * shared with another process. In both cases we'll have an elevated |
| 1245 | * file count (done either by fdget() or by fork()). |
| 1246 | */ |
| 1247 | |
| 1248 | void set_close_on_exec(unsigned int fd, int flag) |
| 1249 | { |
| 1250 | struct files_struct *files = current->files; |
| 1251 | spin_lock(lock: &files->file_lock); |
| 1252 | __set_close_on_exec(fd, files_fdtable(files), set: flag); |
| 1253 | spin_unlock(lock: &files->file_lock); |
| 1254 | } |
| 1255 | |
| 1256 | bool get_close_on_exec(unsigned int fd) |
| 1257 | { |
| 1258 | bool res; |
| 1259 | rcu_read_lock(); |
| 1260 | res = close_on_exec(fd, current->files); |
| 1261 | rcu_read_unlock(); |
| 1262 | return res; |
| 1263 | } |
| 1264 | |
| 1265 | static int do_dup2(struct files_struct *files, |
| 1266 | struct file *file, unsigned fd, unsigned flags) |
| 1267 | __releases(&files->file_lock) |
| 1268 | { |
| 1269 | struct file *tofree; |
| 1270 | struct fdtable *fdt; |
| 1271 | |
| 1272 | /* |
| 1273 | * dup2() is expected to close the file installed in the target fd slot |
| 1274 | * (if any). However, userspace hand-picking a fd may be racing against |
| 1275 | * its own threads which happened to allocate it in open() et al but did |
| 1276 | * not populate it yet. |
| 1277 | * |
| 1278 | * Broadly speaking we may be racing against the following: |
| 1279 | * fd = get_unused_fd_flags(); // fd slot reserved, ->fd[fd] == NULL |
| 1280 | * file = hard_work_goes_here(); |
| 1281 | * fd_install(fd, file); // only now ->fd[fd] == file |
| 1282 | * |
| 1283 | * It is an invariant that a successfully allocated fd has a NULL entry |
| 1284 | * in the array until the matching fd_install(). |
| 1285 | * |
| 1286 | * If we fit the window, we have the fd to populate, yet no target file |
| 1287 | * to close. Trying to ignore it and install our new file would violate |
| 1288 | * the invariant and make fd_install() overwrite our file. |
| 1289 | * |
| 1290 | * Things can be done(tm) to handle this. However, the issue does not |
| 1291 | * concern legitimate programs and we only need to make sure the kernel |
| 1292 | * does not trip over it. |
| 1293 | * |
| 1294 | * The simplest way out is to return an error if we find ourselves here. |
| 1295 | * |
| 1296 | * POSIX is silent on the issue, we return -EBUSY. |
| 1297 | */ |
| 1298 | fdt = files_fdtable(files); |
| 1299 | fd = array_index_nospec(fd, fdt->max_fds); |
| 1300 | tofree = rcu_dereference_raw(fdt->fd[fd]); |
| 1301 | if (!tofree && fd_is_open(fd, fdt)) |
| 1302 | goto Ebusy; |
| 1303 | get_file(f: file); |
| 1304 | rcu_assign_pointer(fdt->fd[fd], file); |
| 1305 | __set_open_fd(fd, fdt, set: flags & O_CLOEXEC); |
| 1306 | spin_unlock(lock: &files->file_lock); |
| 1307 | |
| 1308 | if (tofree) |
| 1309 | filp_close(tofree, id: files); |
| 1310 | |
| 1311 | return fd; |
| 1312 | |
| 1313 | Ebusy: |
| 1314 | spin_unlock(lock: &files->file_lock); |
| 1315 | return -EBUSY; |
| 1316 | } |
| 1317 | |
| 1318 | int replace_fd(unsigned fd, struct file *file, unsigned flags) |
| 1319 | { |
| 1320 | int err; |
| 1321 | struct files_struct *files = current->files; |
| 1322 | |
| 1323 | if (!file) |
| 1324 | return close_fd(fd); |
| 1325 | |
| 1326 | if (fd >= rlimit(RLIMIT_NOFILE)) |
| 1327 | return -EBADF; |
| 1328 | |
| 1329 | spin_lock(lock: &files->file_lock); |
| 1330 | err = expand_files(files, nr: fd); |
| 1331 | if (unlikely(err < 0)) |
| 1332 | goto out_unlock; |
| 1333 | err = do_dup2(files, file, fd, flags); |
| 1334 | if (err < 0) |
| 1335 | return err; |
| 1336 | return 0; |
| 1337 | |
| 1338 | out_unlock: |
| 1339 | spin_unlock(lock: &files->file_lock); |
| 1340 | return err; |
| 1341 | } |
| 1342 | |
| 1343 | /** |
| 1344 | * receive_fd() - Install received file into file descriptor table |
| 1345 | * @file: struct file that was received from another process |
| 1346 | * @ufd: __user pointer to write new fd number to |
| 1347 | * @o_flags: the O_* flags to apply to the new fd entry |
| 1348 | * |
| 1349 | * Installs a received file into the file descriptor table, with appropriate |
| 1350 | * checks and count updates. Optionally writes the fd number to userspace, if |
| 1351 | * @ufd is non-NULL. |
| 1352 | * |
| 1353 | * This helper handles its own reference counting of the incoming |
| 1354 | * struct file. |
| 1355 | * |
| 1356 | * Returns newly install fd or -ve on error. |
| 1357 | */ |
| 1358 | int receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) |
| 1359 | { |
| 1360 | int new_fd; |
| 1361 | int error; |
| 1362 | |
| 1363 | error = security_file_receive(file); |
| 1364 | if (error) |
| 1365 | return error; |
| 1366 | |
| 1367 | new_fd = get_unused_fd_flags(o_flags); |
| 1368 | if (new_fd < 0) |
| 1369 | return new_fd; |
| 1370 | |
| 1371 | if (ufd) { |
| 1372 | error = put_user(new_fd, ufd); |
| 1373 | if (error) { |
| 1374 | put_unused_fd(new_fd); |
| 1375 | return error; |
| 1376 | } |
| 1377 | } |
| 1378 | |
| 1379 | fd_install(new_fd, get_file(f: file)); |
| 1380 | __receive_sock(file); |
| 1381 | return new_fd; |
| 1382 | } |
| 1383 | EXPORT_SYMBOL_GPL(receive_fd); |
| 1384 | |
| 1385 | int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags) |
| 1386 | { |
| 1387 | int error; |
| 1388 | |
| 1389 | error = security_file_receive(file); |
| 1390 | if (error) |
| 1391 | return error; |
| 1392 | error = replace_fd(fd: new_fd, file, flags: o_flags); |
| 1393 | if (error) |
| 1394 | return error; |
| 1395 | __receive_sock(file); |
| 1396 | return new_fd; |
| 1397 | } |
| 1398 | |
| 1399 | static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags) |
| 1400 | { |
| 1401 | int err = -EBADF; |
| 1402 | struct file *file; |
| 1403 | struct files_struct *files = current->files; |
| 1404 | |
| 1405 | if ((flags & ~O_CLOEXEC) != 0) |
| 1406 | return -EINVAL; |
| 1407 | |
| 1408 | if (unlikely(oldfd == newfd)) |
| 1409 | return -EINVAL; |
| 1410 | |
| 1411 | if (newfd >= rlimit(RLIMIT_NOFILE)) |
| 1412 | return -EBADF; |
| 1413 | |
| 1414 | spin_lock(lock: &files->file_lock); |
| 1415 | err = expand_files(files, nr: newfd); |
| 1416 | file = files_lookup_fd_locked(files, fd: oldfd); |
| 1417 | if (unlikely(!file)) |
| 1418 | goto Ebadf; |
| 1419 | if (unlikely(err < 0)) { |
| 1420 | if (err == -EMFILE) |
| 1421 | goto Ebadf; |
| 1422 | goto out_unlock; |
| 1423 | } |
| 1424 | return do_dup2(files, file, fd: newfd, flags); |
| 1425 | |
| 1426 | Ebadf: |
| 1427 | err = -EBADF; |
| 1428 | out_unlock: |
| 1429 | spin_unlock(lock: &files->file_lock); |
| 1430 | return err; |
| 1431 | } |
| 1432 | |
| 1433 | SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags) |
| 1434 | { |
| 1435 | return ksys_dup3(oldfd, newfd, flags); |
| 1436 | } |
| 1437 | |
| 1438 | SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd) |
| 1439 | { |
| 1440 | if (unlikely(newfd == oldfd)) { /* corner case */ |
| 1441 | struct files_struct *files = current->files; |
| 1442 | struct file *f; |
| 1443 | int retval = oldfd; |
| 1444 | |
| 1445 | rcu_read_lock(); |
| 1446 | f = __fget_files_rcu(files, fd: oldfd, mask: 0); |
| 1447 | if (!f) |
| 1448 | retval = -EBADF; |
| 1449 | rcu_read_unlock(); |
| 1450 | if (f) |
| 1451 | fput(f); |
| 1452 | return retval; |
| 1453 | } |
| 1454 | return ksys_dup3(oldfd, newfd, flags: 0); |
| 1455 | } |
| 1456 | |
| 1457 | SYSCALL_DEFINE1(dup, unsigned int, fildes) |
| 1458 | { |
| 1459 | int ret = -EBADF; |
| 1460 | struct file *file = fget_raw(fildes); |
| 1461 | |
| 1462 | if (file) { |
| 1463 | ret = get_unused_fd_flags(0); |
| 1464 | if (ret >= 0) |
| 1465 | fd_install(ret, file); |
| 1466 | else |
| 1467 | fput(file); |
| 1468 | } |
| 1469 | return ret; |
| 1470 | } |
| 1471 | |
| 1472 | int f_dupfd(unsigned int from, struct file *file, unsigned flags) |
| 1473 | { |
| 1474 | unsigned long nofile = rlimit(RLIMIT_NOFILE); |
| 1475 | int err; |
| 1476 | if (from >= nofile) |
| 1477 | return -EINVAL; |
| 1478 | err = alloc_fd(start: from, end: nofile, flags); |
| 1479 | if (err >= 0) { |
| 1480 | get_file(f: file); |
| 1481 | fd_install(err, file); |
| 1482 | } |
| 1483 | return err; |
| 1484 | } |
| 1485 | |
| 1486 | int iterate_fd(struct files_struct *files, unsigned n, |
| 1487 | int (*f)(const void *, struct file *, unsigned), |
| 1488 | const void *p) |
| 1489 | { |
| 1490 | struct fdtable *fdt; |
| 1491 | int res = 0; |
| 1492 | if (!files) |
| 1493 | return 0; |
| 1494 | spin_lock(lock: &files->file_lock); |
| 1495 | for (fdt = files_fdtable(files); n < fdt->max_fds; n++) { |
| 1496 | struct file *file; |
| 1497 | file = rcu_dereference_check_fdtable(files, fdt->fd[n]); |
| 1498 | if (!file) |
| 1499 | continue; |
| 1500 | res = f(p, file, n); |
| 1501 | if (res) |
| 1502 | break; |
| 1503 | } |
| 1504 | spin_unlock(lock: &files->file_lock); |
| 1505 | return res; |
| 1506 | } |
| 1507 | EXPORT_SYMBOL(iterate_fd); |
| 1508 | |