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

source code of linux/fs/file.c