blob: 8431dfde036cc75bbae793f764755beab731040d [file] [log] [blame]
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001// 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>
Olivier Deprez157378f2022-04-04 15:47:50 +020013#include <linux/kernel.h>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000014#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>
Olivier Deprez157378f2022-04-04 15:47:50 +020022#include <linux/close_range.h>
23#include <net/sock.h>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000024
25unsigned int sysctl_nr_open __read_mostly = 1024*1024;
26unsigned int sysctl_nr_open_min = BITS_PER_LONG;
27/* our min() is unusable in constant expressions ;-/ */
28#define __const_min(x, y) ((x) < (y) ? (x) : (y))
29unsigned int sysctl_nr_open_max =
30 __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
31
32static void __free_fdtable(struct fdtable *fdt)
33{
34 kvfree(fdt->fd);
35 kvfree(fdt->open_fds);
36 kfree(fdt);
37}
38
39static void free_fdtable_rcu(struct rcu_head *rcu)
40{
41 __free_fdtable(container_of(rcu, struct fdtable, rcu));
42}
43
44#define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
45#define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
46
47/*
48 * Copy 'count' fd bits from the old table to the new table and clear the extra
49 * space if any. This does not copy the file pointers. Called with the files
50 * spinlock held for write.
51 */
52static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
53 unsigned int count)
54{
55 unsigned int cpy, set;
56
57 cpy = count / BITS_PER_BYTE;
58 set = (nfdt->max_fds - count) / BITS_PER_BYTE;
59 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
60 memset((char *)nfdt->open_fds + cpy, 0, set);
61 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
62 memset((char *)nfdt->close_on_exec + cpy, 0, set);
63
64 cpy = BITBIT_SIZE(count);
65 set = BITBIT_SIZE(nfdt->max_fds) - cpy;
66 memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
67 memset((char *)nfdt->full_fds_bits + cpy, 0, set);
68}
69
70/*
71 * Copy all file descriptors from the old table to the new, expanded table and
72 * clear the extra space. Called with the files spinlock held for write.
73 */
74static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
75{
Olivier Deprez0e641232021-09-23 10:07:05 +020076 size_t cpy, set;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000077
78 BUG_ON(nfdt->max_fds < ofdt->max_fds);
79
80 cpy = ofdt->max_fds * sizeof(struct file *);
81 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
82 memcpy(nfdt->fd, ofdt->fd, cpy);
83 memset((char *)nfdt->fd + cpy, 0, set);
84
85 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
86}
87
Olivier Deprez92d4c212022-12-06 15:05:30 +010088/*
89 * Note how the fdtable bitmap allocations very much have to be a multiple of
90 * BITS_PER_LONG. This is not only because we walk those things in chunks of
91 * 'unsigned long' in some places, but simply because that is how the Linux
92 * kernel bitmaps are defined to work: they are not "bits in an array of bytes",
93 * they are very much "bits in an array of unsigned long".
94 *
95 * The ALIGN(nr, BITS_PER_LONG) here is for clarity: since we just multiplied
96 * by that "1024/sizeof(ptr)" before, we already know there are sufficient
97 * clear low bits. Clang seems to realize that, gcc ends up being confused.
98 *
99 * On a 128-bit machine, the ALIGN() would actually matter. In the meantime,
100 * let's consider it documentation (and maybe a test-case for gcc to improve
101 * its code generation ;)
102 */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000103static struct fdtable * alloc_fdtable(unsigned int nr)
104{
105 struct fdtable *fdt;
106 void *data;
107
108 /*
109 * Figure out how many fds we actually want to support in this fdtable.
110 * Allocation steps are keyed to the size of the fdarray, since it
111 * grows far faster than any of the other dynamic data. We try to fit
112 * the fdarray into comfortable page-tuned chunks: starting at 1024B
113 * and growing in powers of two from there on.
114 */
115 nr /= (1024 / sizeof(struct file *));
116 nr = roundup_pow_of_two(nr + 1);
117 nr *= (1024 / sizeof(struct file *));
Olivier Deprez92d4c212022-12-06 15:05:30 +0100118 nr = ALIGN(nr, BITS_PER_LONG);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000119 /*
120 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
121 * had been set lower between the check in expand_files() and here. Deal
122 * with that in caller, it's cheaper that way.
123 *
124 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
125 * bitmaps handling below becomes unpleasant, to put it mildly...
126 */
127 if (unlikely(nr > sysctl_nr_open))
128 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
129
130 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
131 if (!fdt)
132 goto out;
133 fdt->max_fds = nr;
134 data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
135 if (!data)
136 goto out_fdt;
137 fdt->fd = data;
138
139 data = kvmalloc(max_t(size_t,
140 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
141 GFP_KERNEL_ACCOUNT);
142 if (!data)
143 goto out_arr;
144 fdt->open_fds = data;
145 data += nr / BITS_PER_BYTE;
146 fdt->close_on_exec = data;
147 data += nr / BITS_PER_BYTE;
148 fdt->full_fds_bits = data;
149
150 return fdt;
151
152out_arr:
153 kvfree(fdt->fd);
154out_fdt:
155 kfree(fdt);
156out:
157 return NULL;
158}
159
160/*
161 * Expand the file descriptor table.
162 * This function will allocate a new fdtable and both fd array and fdset, of
163 * the given size.
164 * Return <0 error code on error; 1 on successful completion.
165 * The files->file_lock should be held on entry, and will be held on exit.
166 */
167static int expand_fdtable(struct files_struct *files, unsigned int nr)
168 __releases(files->file_lock)
169 __acquires(files->file_lock)
170{
171 struct fdtable *new_fdt, *cur_fdt;
172
173 spin_unlock(&files->file_lock);
174 new_fdt = alloc_fdtable(nr);
175
176 /* make sure all __fd_install() have seen resize_in_progress
177 * or have finished their rcu_read_lock_sched() section.
178 */
179 if (atomic_read(&files->count) > 1)
David Brazdil0f672f62019-12-10 10:32:29 +0000180 synchronize_rcu();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000181
182 spin_lock(&files->file_lock);
183 if (!new_fdt)
184 return -ENOMEM;
185 /*
186 * extremely unlikely race - sysctl_nr_open decreased between the check in
187 * caller and alloc_fdtable(). Cheaper to catch it here...
188 */
189 if (unlikely(new_fdt->max_fds <= nr)) {
190 __free_fdtable(new_fdt);
191 return -EMFILE;
192 }
193 cur_fdt = files_fdtable(files);
194 BUG_ON(nr < cur_fdt->max_fds);
195 copy_fdtable(new_fdt, cur_fdt);
196 rcu_assign_pointer(files->fdt, new_fdt);
197 if (cur_fdt != &files->fdtab)
198 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
199 /* coupled with smp_rmb() in __fd_install() */
200 smp_wmb();
201 return 1;
202}
203
204/*
205 * Expand files.
206 * This function will expand the file structures, if the requested size exceeds
207 * the current capacity and there is room for expansion.
208 * Return <0 error code on error; 0 when nothing done; 1 when files were
209 * expanded and execution may have blocked.
210 * The files->file_lock should be held on entry, and will be held on exit.
211 */
212static int expand_files(struct files_struct *files, unsigned int nr)
213 __releases(files->file_lock)
214 __acquires(files->file_lock)
215{
216 struct fdtable *fdt;
217 int expanded = 0;
218
219repeat:
220 fdt = files_fdtable(files);
221
222 /* Do we need to expand? */
223 if (nr < fdt->max_fds)
224 return expanded;
225
226 /* Can we expand? */
227 if (nr >= sysctl_nr_open)
228 return -EMFILE;
229
230 if (unlikely(files->resize_in_progress)) {
231 spin_unlock(&files->file_lock);
232 expanded = 1;
233 wait_event(files->resize_wait, !files->resize_in_progress);
234 spin_lock(&files->file_lock);
235 goto repeat;
236 }
237
238 /* All good, so we try */
239 files->resize_in_progress = true;
240 expanded = expand_fdtable(files, nr);
241 files->resize_in_progress = false;
242
243 wake_up_all(&files->resize_wait);
244 return expanded;
245}
246
247static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
248{
249 __set_bit(fd, fdt->close_on_exec);
250}
251
252static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
253{
254 if (test_bit(fd, fdt->close_on_exec))
255 __clear_bit(fd, fdt->close_on_exec);
256}
257
258static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
259{
260 __set_bit(fd, fdt->open_fds);
261 fd /= BITS_PER_LONG;
262 if (!~fdt->open_fds[fd])
263 __set_bit(fd, fdt->full_fds_bits);
264}
265
266static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
267{
268 __clear_bit(fd, fdt->open_fds);
269 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
270}
271
272static unsigned int count_open_files(struct fdtable *fdt)
273{
274 unsigned int size = fdt->max_fds;
275 unsigned int i;
276
277 /* Find the last open fd */
278 for (i = size / BITS_PER_LONG; i > 0; ) {
279 if (fdt->open_fds[--i])
280 break;
281 }
282 i = (i + 1) * BITS_PER_LONG;
283 return i;
284}
285
Olivier Deprez92d4c212022-12-06 15:05:30 +0100286/*
287 * Note that a sane fdtable size always has to be a multiple of
288 * BITS_PER_LONG, since we have bitmaps that are sized by this.
289 *
290 * 'max_fds' will normally already be properly aligned, but it
291 * turns out that in the close_range() -> __close_range() ->
292 * unshare_fd() -> dup_fd() -> sane_fdtable_size() we can end
293 * up having a 'max_fds' value that isn't already aligned.
294 *
295 * Rather than make close_range() have to worry about this,
296 * just make that BITS_PER_LONG alignment be part of a sane
297 * fdtable size. Becuase that's really what it is.
298 */
Olivier Deprez157378f2022-04-04 15:47:50 +0200299static unsigned int sane_fdtable_size(struct fdtable *fdt, unsigned int max_fds)
300{
301 unsigned int count;
302
303 count = count_open_files(fdt);
304 if (max_fds < NR_OPEN_DEFAULT)
305 max_fds = NR_OPEN_DEFAULT;
Olivier Deprez92d4c212022-12-06 15:05:30 +0100306 return ALIGN(min(count, max_fds), BITS_PER_LONG);
Olivier Deprez157378f2022-04-04 15:47:50 +0200307}
308
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000309/*
310 * Allocate a new files structure and copy contents from the
311 * passed in files structure.
312 * errorp will be valid only when the returned files_struct is NULL.
313 */
Olivier Deprez157378f2022-04-04 15:47:50 +0200314struct files_struct *dup_fd(struct files_struct *oldf, unsigned int max_fds, int *errorp)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000315{
316 struct files_struct *newf;
317 struct file **old_fds, **new_fds;
318 unsigned int open_files, i;
319 struct fdtable *old_fdt, *new_fdt;
320
321 *errorp = -ENOMEM;
322 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
323 if (!newf)
324 goto out;
325
326 atomic_set(&newf->count, 1);
327
328 spin_lock_init(&newf->file_lock);
329 newf->resize_in_progress = false;
330 init_waitqueue_head(&newf->resize_wait);
331 newf->next_fd = 0;
332 new_fdt = &newf->fdtab;
333 new_fdt->max_fds = NR_OPEN_DEFAULT;
334 new_fdt->close_on_exec = newf->close_on_exec_init;
335 new_fdt->open_fds = newf->open_fds_init;
336 new_fdt->full_fds_bits = newf->full_fds_bits_init;
337 new_fdt->fd = &newf->fd_array[0];
338
339 spin_lock(&oldf->file_lock);
340 old_fdt = files_fdtable(oldf);
Olivier Deprez157378f2022-04-04 15:47:50 +0200341 open_files = sane_fdtable_size(old_fdt, max_fds);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000342
343 /*
344 * Check whether we need to allocate a larger fd array and fd set.
345 */
346 while (unlikely(open_files > new_fdt->max_fds)) {
347 spin_unlock(&oldf->file_lock);
348
349 if (new_fdt != &newf->fdtab)
350 __free_fdtable(new_fdt);
351
352 new_fdt = alloc_fdtable(open_files - 1);
353 if (!new_fdt) {
354 *errorp = -ENOMEM;
355 goto out_release;
356 }
357
358 /* beyond sysctl_nr_open; nothing to do */
359 if (unlikely(new_fdt->max_fds < open_files)) {
360 __free_fdtable(new_fdt);
361 *errorp = -EMFILE;
362 goto out_release;
363 }
364
365 /*
366 * Reacquire the oldf lock and a pointer to its fd table
367 * who knows it may have a new bigger fd table. We need
368 * the latest pointer.
369 */
370 spin_lock(&oldf->file_lock);
371 old_fdt = files_fdtable(oldf);
Olivier Deprez157378f2022-04-04 15:47:50 +0200372 open_files = sane_fdtable_size(old_fdt, max_fds);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000373 }
374
375 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
376
377 old_fds = old_fdt->fd;
378 new_fds = new_fdt->fd;
379
380 for (i = open_files; i != 0; i--) {
381 struct file *f = *old_fds++;
382 if (f) {
383 get_file(f);
384 } else {
385 /*
386 * The fd may be claimed in the fd bitmap but not yet
387 * instantiated in the files array if a sibling thread
388 * is partway through open(). So make sure that this
389 * fd is available to the new process.
390 */
391 __clear_open_fd(open_files - i, new_fdt);
392 }
393 rcu_assign_pointer(*new_fds++, f);
394 }
395 spin_unlock(&oldf->file_lock);
396
397 /* clear the remainder */
398 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
399
400 rcu_assign_pointer(newf->fdt, new_fdt);
401
402 return newf;
403
404out_release:
405 kmem_cache_free(files_cachep, newf);
406out:
407 return NULL;
408}
409
410static struct fdtable *close_files(struct files_struct * files)
411{
412 /*
413 * It is safe to dereference the fd table without RCU or
414 * ->file_lock because this is the last reference to the
415 * files structure.
416 */
417 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
418 unsigned int i, j = 0;
419
420 for (;;) {
421 unsigned long set;
422 i = j * BITS_PER_LONG;
423 if (i >= fdt->max_fds)
424 break;
425 set = fdt->open_fds[j++];
426 while (set) {
427 if (set & 1) {
428 struct file * file = xchg(&fdt->fd[i], NULL);
429 if (file) {
430 filp_close(file, files);
431 cond_resched();
432 }
433 }
434 i++;
435 set >>= 1;
436 }
437 }
438
439 return fdt;
440}
441
442struct files_struct *get_files_struct(struct task_struct *task)
443{
444 struct files_struct *files;
445
446 task_lock(task);
447 files = task->files;
448 if (files)
449 atomic_inc(&files->count);
450 task_unlock(task);
451
452 return files;
453}
454
455void put_files_struct(struct files_struct *files)
456{
457 if (atomic_dec_and_test(&files->count)) {
458 struct fdtable *fdt = close_files(files);
459
460 /* free the arrays if they are not embedded */
461 if (fdt != &files->fdtab)
462 __free_fdtable(fdt);
463 kmem_cache_free(files_cachep, files);
464 }
465}
466
467void reset_files_struct(struct files_struct *files)
468{
469 struct task_struct *tsk = current;
470 struct files_struct *old;
471
472 old = tsk->files;
473 task_lock(tsk);
474 tsk->files = files;
475 task_unlock(tsk);
476 put_files_struct(old);
477}
478
479void exit_files(struct task_struct *tsk)
480{
481 struct files_struct * files = tsk->files;
482
483 if (files) {
484 task_lock(tsk);
485 tsk->files = NULL;
486 task_unlock(tsk);
487 put_files_struct(files);
488 }
489}
490
491struct files_struct init_files = {
492 .count = ATOMIC_INIT(1),
493 .fdt = &init_files.fdtab,
494 .fdtab = {
495 .max_fds = NR_OPEN_DEFAULT,
496 .fd = &init_files.fd_array[0],
497 .close_on_exec = init_files.close_on_exec_init,
498 .open_fds = init_files.open_fds_init,
499 .full_fds_bits = init_files.full_fds_bits_init,
500 },
501 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
David Brazdil0f672f62019-12-10 10:32:29 +0000502 .resize_wait = __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000503};
504
505static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
506{
507 unsigned int maxfd = fdt->max_fds;
508 unsigned int maxbit = maxfd / BITS_PER_LONG;
509 unsigned int bitbit = start / BITS_PER_LONG;
510
511 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
512 if (bitbit > maxfd)
513 return maxfd;
514 if (bitbit > start)
515 start = bitbit;
516 return find_next_zero_bit(fdt->open_fds, maxfd, start);
517}
518
519/*
520 * allocate a file descriptor, mark it busy.
521 */
522int __alloc_fd(struct files_struct *files,
523 unsigned start, unsigned end, unsigned flags)
524{
525 unsigned int fd;
526 int error;
527 struct fdtable *fdt;
528
529 spin_lock(&files->file_lock);
530repeat:
531 fdt = files_fdtable(files);
532 fd = start;
533 if (fd < files->next_fd)
534 fd = files->next_fd;
535
536 if (fd < fdt->max_fds)
537 fd = find_next_fd(fdt, fd);
538
539 /*
540 * N.B. For clone tasks sharing a files structure, this test
541 * will limit the total number of files that can be opened.
542 */
543 error = -EMFILE;
544 if (fd >= end)
545 goto out;
546
547 error = expand_files(files, fd);
548 if (error < 0)
549 goto out;
550
551 /*
552 * If we needed to expand the fs array we
553 * might have blocked - try again.
554 */
555 if (error)
556 goto repeat;
557
558 if (start <= files->next_fd)
559 files->next_fd = fd + 1;
560
561 __set_open_fd(fd, fdt);
562 if (flags & O_CLOEXEC)
563 __set_close_on_exec(fd, fdt);
564 else
565 __clear_close_on_exec(fd, fdt);
566 error = fd;
567#if 1
568 /* Sanity check */
569 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
570 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
571 rcu_assign_pointer(fdt->fd[fd], NULL);
572 }
573#endif
574
575out:
576 spin_unlock(&files->file_lock);
577 return error;
578}
579
580static int alloc_fd(unsigned start, unsigned flags)
581{
582 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
583}
584
Olivier Deprez157378f2022-04-04 15:47:50 +0200585int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
586{
587 return __alloc_fd(current->files, 0, nofile, flags);
588}
589
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000590int get_unused_fd_flags(unsigned flags)
591{
Olivier Deprez157378f2022-04-04 15:47:50 +0200592 return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000593}
594EXPORT_SYMBOL(get_unused_fd_flags);
595
596static void __put_unused_fd(struct files_struct *files, unsigned int fd)
597{
598 struct fdtable *fdt = files_fdtable(files);
599 __clear_open_fd(fd, fdt);
600 if (fd < files->next_fd)
601 files->next_fd = fd;
602}
603
604void put_unused_fd(unsigned int fd)
605{
606 struct files_struct *files = current->files;
607 spin_lock(&files->file_lock);
608 __put_unused_fd(files, fd);
609 spin_unlock(&files->file_lock);
610}
611
612EXPORT_SYMBOL(put_unused_fd);
613
614/*
615 * Install a file pointer in the fd array.
616 *
617 * The VFS is full of places where we drop the files lock between
618 * setting the open_fds bitmap and installing the file in the file
619 * array. At any such point, we are vulnerable to a dup2() race
620 * installing a file in the array before us. We need to detect this and
621 * fput() the struct file we are about to overwrite in this case.
622 *
623 * It should never happen - if we allow dup2() do it, _really_ bad things
624 * will follow.
625 *
626 * NOTE: __fd_install() variant is really, really low-level; don't
627 * use it unless you are forced to by truly lousy API shoved down
628 * your throat. 'files' *MUST* be either current->files or obtained
629 * by get_files_struct(current) done by whoever had given it to you,
630 * or really bad things will happen. Normally you want to use
631 * fd_install() instead.
632 */
633
634void __fd_install(struct files_struct *files, unsigned int fd,
635 struct file *file)
636{
637 struct fdtable *fdt;
638
639 rcu_read_lock_sched();
640
641 if (unlikely(files->resize_in_progress)) {
642 rcu_read_unlock_sched();
643 spin_lock(&files->file_lock);
644 fdt = files_fdtable(files);
645 BUG_ON(fdt->fd[fd] != NULL);
646 rcu_assign_pointer(fdt->fd[fd], file);
647 spin_unlock(&files->file_lock);
648 return;
649 }
650 /* coupled with smp_wmb() in expand_fdtable() */
651 smp_rmb();
652 fdt = rcu_dereference_sched(files->fdt);
653 BUG_ON(fdt->fd[fd] != NULL);
654 rcu_assign_pointer(fdt->fd[fd], file);
655 rcu_read_unlock_sched();
656}
657
Olivier Deprez157378f2022-04-04 15:47:50 +0200658/*
659 * This consumes the "file" refcount, so callers should treat it
660 * as if they had called fput(file).
661 */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000662void fd_install(unsigned int fd, struct file *file)
663{
664 __fd_install(current->files, fd, file);
665}
666
667EXPORT_SYMBOL(fd_install);
668
Olivier Deprez157378f2022-04-04 15:47:50 +0200669static struct file *pick_file(struct files_struct *files, unsigned fd)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000670{
Olivier Deprez157378f2022-04-04 15:47:50 +0200671 struct file *file = NULL;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000672 struct fdtable *fdt;
673
674 spin_lock(&files->file_lock);
675 fdt = files_fdtable(files);
676 if (fd >= fdt->max_fds)
677 goto out_unlock;
678 file = fdt->fd[fd];
679 if (!file)
680 goto out_unlock;
681 rcu_assign_pointer(fdt->fd[fd], NULL);
682 __put_unused_fd(files, fd);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000683
684out_unlock:
685 spin_unlock(&files->file_lock);
Olivier Deprez157378f2022-04-04 15:47:50 +0200686 return file;
687}
688
689/*
690 * The same warnings as for __alloc_fd()/__fd_install() apply here...
691 */
692int __close_fd(struct files_struct *files, unsigned fd)
693{
694 struct file *file;
695
696 file = pick_file(files, fd);
697 if (!file)
698 return -EBADF;
699
700 return filp_close(file, files);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000701}
702EXPORT_SYMBOL(__close_fd); /* for ksys_close() */
703
Olivier Deprez157378f2022-04-04 15:47:50 +0200704/**
705 * __close_range() - Close all file descriptors in a given range.
706 *
707 * @fd: starting file descriptor to close
708 * @max_fd: last file descriptor to close
709 *
710 * This closes a range of file descriptors. All file descriptors
711 * from @fd up to and including @max_fd are closed.
712 */
713int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
714{
715 unsigned int cur_max;
716 struct task_struct *me = current;
717 struct files_struct *cur_fds = me->files, *fds = NULL;
718
719 if (flags & ~CLOSE_RANGE_UNSHARE)
720 return -EINVAL;
721
722 if (fd > max_fd)
723 return -EINVAL;
724
725 rcu_read_lock();
726 cur_max = files_fdtable(cur_fds)->max_fds;
727 rcu_read_unlock();
728
729 /* cap to last valid index into fdtable */
730 cur_max--;
731
732 if (flags & CLOSE_RANGE_UNSHARE) {
733 int ret;
734 unsigned int max_unshare_fds = NR_OPEN_MAX;
735
736 /*
737 * If the requested range is greater than the current maximum,
738 * we're closing everything so only copy all file descriptors
739 * beneath the lowest file descriptor.
740 */
741 if (max_fd >= cur_max)
742 max_unshare_fds = fd;
743
744 ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
745 if (ret)
746 return ret;
747
748 /*
749 * We used to share our file descriptor table, and have now
750 * created a private one, make sure we're using it below.
751 */
752 if (fds)
753 swap(cur_fds, fds);
754 }
755
756 max_fd = min(max_fd, cur_max);
757 while (fd <= max_fd) {
758 struct file *file;
759
760 file = pick_file(cur_fds, fd++);
761 if (!file)
762 continue;
763
764 filp_close(file, cur_fds);
765 cond_resched();
766 }
767
768 if (fds) {
769 /*
770 * We're done closing the files we were supposed to. Time to install
771 * the new file descriptor table and drop the old one.
772 */
773 task_lock(me);
774 me->files = cur_fds;
775 task_unlock(me);
776 put_files_struct(fds);
777 }
778
779 return 0;
780}
781
David Brazdil0f672f62019-12-10 10:32:29 +0000782/*
Olivier Deprez157378f2022-04-04 15:47:50 +0200783 * variant of __close_fd that gets a ref on the file for later fput.
784 * The caller must ensure that filp_close() called on the file, and then
785 * an fput().
David Brazdil0f672f62019-12-10 10:32:29 +0000786 */
787int __close_fd_get_file(unsigned int fd, struct file **res)
788{
789 struct files_struct *files = current->files;
790 struct file *file;
791 struct fdtable *fdt;
792
793 spin_lock(&files->file_lock);
794 fdt = files_fdtable(files);
795 if (fd >= fdt->max_fds)
796 goto out_unlock;
797 file = fdt->fd[fd];
798 if (!file)
799 goto out_unlock;
800 rcu_assign_pointer(fdt->fd[fd], NULL);
801 __put_unused_fd(files, fd);
802 spin_unlock(&files->file_lock);
803 get_file(file);
804 *res = file;
Olivier Deprez157378f2022-04-04 15:47:50 +0200805 return 0;
David Brazdil0f672f62019-12-10 10:32:29 +0000806
807out_unlock:
808 spin_unlock(&files->file_lock);
809 *res = NULL;
810 return -ENOENT;
811}
812
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000813void do_close_on_exec(struct files_struct *files)
814{
815 unsigned i;
816 struct fdtable *fdt;
817
818 /* exec unshares first */
819 spin_lock(&files->file_lock);
820 for (i = 0; ; i++) {
821 unsigned long set;
822 unsigned fd = i * BITS_PER_LONG;
823 fdt = files_fdtable(files);
824 if (fd >= fdt->max_fds)
825 break;
826 set = fdt->close_on_exec[i];
827 if (!set)
828 continue;
829 fdt->close_on_exec[i] = 0;
830 for ( ; set ; fd++, set >>= 1) {
831 struct file *file;
832 if (!(set & 1))
833 continue;
834 file = fdt->fd[fd];
835 if (!file)
836 continue;
837 rcu_assign_pointer(fdt->fd[fd], NULL);
838 __put_unused_fd(files, fd);
839 spin_unlock(&files->file_lock);
840 filp_close(file, files);
841 cond_resched();
842 spin_lock(&files->file_lock);
843 }
844
845 }
846 spin_unlock(&files->file_lock);
847}
848
Olivier Deprez157378f2022-04-04 15:47:50 +0200849static inline struct file *__fget_files_rcu(struct files_struct *files,
850 unsigned int fd, fmode_t mask, unsigned int refs)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000851{
Olivier Deprez157378f2022-04-04 15:47:50 +0200852 for (;;) {
853 struct file *file;
854 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
855 struct file __rcu **fdentry;
856
857 if (unlikely(fd >= fdt->max_fds))
858 return NULL;
859
860 fdentry = fdt->fd + array_index_nospec(fd, fdt->max_fds);
861 file = rcu_dereference_raw(*fdentry);
862 if (unlikely(!file))
863 return NULL;
864
865 if (unlikely(file->f_mode & mask))
866 return NULL;
867
868 /*
869 * Ok, we have a file pointer. However, because we do
870 * this all locklessly under RCU, we may be racing with
871 * that file being closed.
872 *
873 * Such a race can take two forms:
874 *
875 * (a) the file ref already went down to zero,
876 * and get_file_rcu_many() fails. Just try
877 * again:
878 */
879 if (unlikely(!get_file_rcu_many(file, refs)))
880 continue;
881
882 /*
883 * (b) the file table entry has changed under us.
884 * Note that we don't need to re-check the 'fdt->fd'
885 * pointer having changed, because it always goes
886 * hand-in-hand with 'fdt'.
887 *
888 * If so, we need to put our refs and try again.
889 */
890 if (unlikely(rcu_dereference_raw(files->fdt) != fdt) ||
891 unlikely(rcu_dereference_raw(*fdentry) != file)) {
892 fput_many(file, refs);
893 continue;
894 }
895
896 /*
897 * Ok, we have a ref to the file, and checked that it
898 * still exists.
899 */
900 return file;
901 }
902}
903
904static struct file *__fget_files(struct files_struct *files, unsigned int fd,
905 fmode_t mask, unsigned int refs)
906{
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000907 struct file *file;
908
909 rcu_read_lock();
Olivier Deprez157378f2022-04-04 15:47:50 +0200910 file = __fget_files_rcu(files, fd, mask, refs);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000911 rcu_read_unlock();
912
913 return file;
914}
915
Olivier Deprez157378f2022-04-04 15:47:50 +0200916static inline struct file *__fget(unsigned int fd, fmode_t mask,
917 unsigned int refs)
918{
919 return __fget_files(current->files, fd, mask, refs);
920}
921
David Brazdil0f672f62019-12-10 10:32:29 +0000922struct file *fget_many(unsigned int fd, unsigned int refs)
923{
924 return __fget(fd, FMODE_PATH, refs);
925}
926
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000927struct file *fget(unsigned int fd)
928{
David Brazdil0f672f62019-12-10 10:32:29 +0000929 return __fget(fd, FMODE_PATH, 1);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000930}
931EXPORT_SYMBOL(fget);
932
933struct file *fget_raw(unsigned int fd)
934{
David Brazdil0f672f62019-12-10 10:32:29 +0000935 return __fget(fd, 0, 1);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000936}
937EXPORT_SYMBOL(fget_raw);
938
Olivier Deprez157378f2022-04-04 15:47:50 +0200939struct file *fget_task(struct task_struct *task, unsigned int fd)
940{
941 struct file *file = NULL;
942
943 task_lock(task);
944 if (task->files)
945 file = __fget_files(task->files, fd, 0, 1);
946 task_unlock(task);
947
948 return file;
949}
950
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000951/*
952 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
953 *
954 * You can use this instead of fget if you satisfy all of the following
955 * conditions:
956 * 1) You must call fput_light before exiting the syscall and returning control
957 * to userspace (i.e. you cannot remember the returned struct file * after
958 * returning to userspace).
959 * 2) You must not call filp_close on the returned struct file * in between
960 * calls to fget_light and fput_light.
961 * 3) You must not clone the current task in between the calls to fget_light
962 * and fput_light.
963 *
964 * The fput_needed flag returned by fget_light should be passed to the
965 * corresponding fput_light.
966 */
967static unsigned long __fget_light(unsigned int fd, fmode_t mask)
968{
969 struct files_struct *files = current->files;
970 struct file *file;
971
972 if (atomic_read(&files->count) == 1) {
973 file = __fcheck_files(files, fd);
974 if (!file || unlikely(file->f_mode & mask))
975 return 0;
976 return (unsigned long)file;
977 } else {
David Brazdil0f672f62019-12-10 10:32:29 +0000978 file = __fget(fd, mask, 1);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000979 if (!file)
980 return 0;
981 return FDPUT_FPUT | (unsigned long)file;
982 }
983}
984unsigned long __fdget(unsigned int fd)
985{
986 return __fget_light(fd, FMODE_PATH);
987}
988EXPORT_SYMBOL(__fdget);
989
990unsigned long __fdget_raw(unsigned int fd)
991{
992 return __fget_light(fd, 0);
993}
994
995unsigned long __fdget_pos(unsigned int fd)
996{
997 unsigned long v = __fdget(fd);
998 struct file *file = (struct file *)(v & ~3);
999
1000 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
1001 if (file_count(file) > 1) {
1002 v |= FDPUT_POS_UNLOCK;
1003 mutex_lock(&file->f_pos_lock);
1004 }
1005 }
1006 return v;
1007}
1008
1009void __f_unlock_pos(struct file *f)
1010{
1011 mutex_unlock(&f->f_pos_lock);
1012}
1013
1014/*
1015 * We only lock f_pos if we have threads or if the file might be
1016 * shared with another process. In both cases we'll have an elevated
1017 * file count (done either by fdget() or by fork()).
1018 */
1019
1020void set_close_on_exec(unsigned int fd, int flag)
1021{
1022 struct files_struct *files = current->files;
1023 struct fdtable *fdt;
1024 spin_lock(&files->file_lock);
1025 fdt = files_fdtable(files);
1026 if (flag)
1027 __set_close_on_exec(fd, fdt);
1028 else
1029 __clear_close_on_exec(fd, fdt);
1030 spin_unlock(&files->file_lock);
1031}
1032
1033bool get_close_on_exec(unsigned int fd)
1034{
1035 struct files_struct *files = current->files;
1036 struct fdtable *fdt;
1037 bool res;
1038 rcu_read_lock();
1039 fdt = files_fdtable(files);
1040 res = close_on_exec(fd, fdt);
1041 rcu_read_unlock();
1042 return res;
1043}
1044
1045static int do_dup2(struct files_struct *files,
1046 struct file *file, unsigned fd, unsigned flags)
1047__releases(&files->file_lock)
1048{
1049 struct file *tofree;
1050 struct fdtable *fdt;
1051
1052 /*
1053 * We need to detect attempts to do dup2() over allocated but still
1054 * not finished descriptor. NB: OpenBSD avoids that at the price of
1055 * extra work in their equivalent of fget() - they insert struct
1056 * file immediately after grabbing descriptor, mark it larval if
1057 * more work (e.g. actual opening) is needed and make sure that
1058 * fget() treats larval files as absent. Potentially interesting,
1059 * but while extra work in fget() is trivial, locking implications
1060 * and amount of surgery on open()-related paths in VFS are not.
1061 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1062 * deadlocks in rather amusing ways, AFAICS. All of that is out of
1063 * scope of POSIX or SUS, since neither considers shared descriptor
1064 * tables and this condition does not arise without those.
1065 */
1066 fdt = files_fdtable(files);
1067 tofree = fdt->fd[fd];
1068 if (!tofree && fd_is_open(fd, fdt))
1069 goto Ebusy;
1070 get_file(file);
1071 rcu_assign_pointer(fdt->fd[fd], file);
1072 __set_open_fd(fd, fdt);
1073 if (flags & O_CLOEXEC)
1074 __set_close_on_exec(fd, fdt);
1075 else
1076 __clear_close_on_exec(fd, fdt);
1077 spin_unlock(&files->file_lock);
1078
1079 if (tofree)
1080 filp_close(tofree, files);
1081
1082 return fd;
1083
1084Ebusy:
1085 spin_unlock(&files->file_lock);
1086 return -EBUSY;
1087}
1088
1089int replace_fd(unsigned fd, struct file *file, unsigned flags)
1090{
1091 int err;
1092 struct files_struct *files = current->files;
1093
1094 if (!file)
1095 return __close_fd(files, fd);
1096
1097 if (fd >= rlimit(RLIMIT_NOFILE))
1098 return -EBADF;
1099
1100 spin_lock(&files->file_lock);
1101 err = expand_files(files, fd);
1102 if (unlikely(err < 0))
1103 goto out_unlock;
1104 return do_dup2(files, file, fd, flags);
1105
1106out_unlock:
1107 spin_unlock(&files->file_lock);
1108 return err;
1109}
1110
Olivier Deprez157378f2022-04-04 15:47:50 +02001111/**
1112 * __receive_fd() - Install received file into file descriptor table
1113 *
1114 * @fd: fd to install into (if negative, a new fd will be allocated)
1115 * @file: struct file that was received from another process
1116 * @ufd: __user pointer to write new fd number to
1117 * @o_flags: the O_* flags to apply to the new fd entry
1118 *
1119 * Installs a received file into the file descriptor table, with appropriate
1120 * checks and count updates. Optionally writes the fd number to userspace, if
1121 * @ufd is non-NULL.
1122 *
1123 * This helper handles its own reference counting of the incoming
1124 * struct file.
1125 *
1126 * Returns newly install fd or -ve on error.
1127 */
1128int __receive_fd(int fd, struct file *file, int __user *ufd, unsigned int o_flags)
1129{
1130 int new_fd;
1131 int error;
1132
1133 error = security_file_receive(file);
1134 if (error)
1135 return error;
1136
1137 if (fd < 0) {
1138 new_fd = get_unused_fd_flags(o_flags);
1139 if (new_fd < 0)
1140 return new_fd;
1141 } else {
1142 new_fd = fd;
1143 }
1144
1145 if (ufd) {
1146 error = put_user(new_fd, ufd);
1147 if (error) {
1148 if (fd < 0)
1149 put_unused_fd(new_fd);
1150 return error;
1151 }
1152 }
1153
1154 if (fd < 0) {
1155 fd_install(new_fd, get_file(file));
1156 } else {
1157 error = replace_fd(new_fd, file, o_flags);
1158 if (error)
1159 return error;
1160 }
1161
1162 /* Bump the sock usage counts, if any. */
1163 __receive_sock(file);
1164 return new_fd;
1165}
1166
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001167static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1168{
1169 int err = -EBADF;
1170 struct file *file;
1171 struct files_struct *files = current->files;
1172
1173 if ((flags & ~O_CLOEXEC) != 0)
1174 return -EINVAL;
1175
1176 if (unlikely(oldfd == newfd))
1177 return -EINVAL;
1178
1179 if (newfd >= rlimit(RLIMIT_NOFILE))
1180 return -EBADF;
1181
1182 spin_lock(&files->file_lock);
1183 err = expand_files(files, newfd);
1184 file = fcheck(oldfd);
1185 if (unlikely(!file))
1186 goto Ebadf;
1187 if (unlikely(err < 0)) {
1188 if (err == -EMFILE)
1189 goto Ebadf;
1190 goto out_unlock;
1191 }
1192 return do_dup2(files, file, newfd, flags);
1193
1194Ebadf:
1195 err = -EBADF;
1196out_unlock:
1197 spin_unlock(&files->file_lock);
1198 return err;
1199}
1200
1201SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1202{
1203 return ksys_dup3(oldfd, newfd, flags);
1204}
1205
1206SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1207{
1208 if (unlikely(newfd == oldfd)) { /* corner case */
1209 struct files_struct *files = current->files;
1210 int retval = oldfd;
1211
1212 rcu_read_lock();
1213 if (!fcheck_files(files, oldfd))
1214 retval = -EBADF;
1215 rcu_read_unlock();
1216 return retval;
1217 }
1218 return ksys_dup3(oldfd, newfd, 0);
1219}
1220
Olivier Deprez157378f2022-04-04 15:47:50 +02001221SYSCALL_DEFINE1(dup, unsigned int, fildes)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001222{
1223 int ret = -EBADF;
1224 struct file *file = fget_raw(fildes);
1225
1226 if (file) {
1227 ret = get_unused_fd_flags(0);
1228 if (ret >= 0)
1229 fd_install(ret, file);
1230 else
1231 fput(file);
1232 }
1233 return ret;
1234}
1235
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001236int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1237{
1238 int err;
1239 if (from >= rlimit(RLIMIT_NOFILE))
1240 return -EINVAL;
1241 err = alloc_fd(from, flags);
1242 if (err >= 0) {
1243 get_file(file);
1244 fd_install(err, file);
1245 }
1246 return err;
1247}
1248
1249int iterate_fd(struct files_struct *files, unsigned n,
1250 int (*f)(const void *, struct file *, unsigned),
1251 const void *p)
1252{
1253 struct fdtable *fdt;
1254 int res = 0;
1255 if (!files)
1256 return 0;
1257 spin_lock(&files->file_lock);
1258 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1259 struct file *file;
1260 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1261 if (!file)
1262 continue;
1263 res = f(p, file, n);
1264 if (res)
1265 break;
1266 }
1267 spin_unlock(&files->file_lock);
1268 return res;
1269}
1270EXPORT_SYMBOL(iterate_fd);