Loading...
1/*
2 * linux/fs/file.c
3 *
4 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
5 *
6 * Manage the dynamic fd arrays in the process files_struct.
7 */
8
9#include <linux/syscalls.h>
10#include <linux/export.h>
11#include <linux/fs.h>
12#include <linux/mm.h>
13#include <linux/mmzone.h>
14#include <linux/time.h>
15#include <linux/sched.h>
16#include <linux/slab.h>
17#include <linux/vmalloc.h>
18#include <linux/file.h>
19#include <linux/fdtable.h>
20#include <linux/bitops.h>
21#include <linux/interrupt.h>
22#include <linux/spinlock.h>
23#include <linux/rcupdate.h>
24#include <linux/workqueue.h>
25
26unsigned int sysctl_nr_open __read_mostly = 1024*1024;
27unsigned int sysctl_nr_open_min = BITS_PER_LONG;
28/* our min() is unusable in constant expressions ;-/ */
29#define __const_min(x, y) ((x) < (y) ? (x) : (y))
30unsigned int sysctl_nr_open_max =
31 __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
32
33static void *alloc_fdmem(size_t size)
34{
35 /*
36 * Very large allocations can stress page reclaim, so fall back to
37 * vmalloc() if the allocation size will be considered "large" by the VM.
38 */
39 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
40 void *data = kmalloc(size, GFP_KERNEL_ACCOUNT |
41 __GFP_NOWARN | __GFP_NORETRY);
42 if (data != NULL)
43 return data;
44 }
45 return __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM, PAGE_KERNEL);
46}
47
48static void __free_fdtable(struct fdtable *fdt)
49{
50 kvfree(fdt->fd);
51 kvfree(fdt->open_fds);
52 kfree(fdt);
53}
54
55static void free_fdtable_rcu(struct rcu_head *rcu)
56{
57 __free_fdtable(container_of(rcu, struct fdtable, rcu));
58}
59
60#define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
61#define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
62
63/*
64 * Copy 'count' fd bits from the old table to the new table and clear the extra
65 * space if any. This does not copy the file pointers. Called with the files
66 * spinlock held for write.
67 */
68static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
69 unsigned int count)
70{
71 unsigned int cpy, set;
72
73 cpy = count / BITS_PER_BYTE;
74 set = (nfdt->max_fds - count) / BITS_PER_BYTE;
75 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
76 memset((char *)nfdt->open_fds + cpy, 0, set);
77 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
78 memset((char *)nfdt->close_on_exec + cpy, 0, set);
79
80 cpy = BITBIT_SIZE(count);
81 set = BITBIT_SIZE(nfdt->max_fds) - cpy;
82 memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
83 memset((char *)nfdt->full_fds_bits + cpy, 0, set);
84}
85
86/*
87 * Copy all file descriptors from the old table to the new, expanded table and
88 * clear the extra space. Called with the files spinlock held for write.
89 */
90static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
91{
92 unsigned int cpy, set;
93
94 BUG_ON(nfdt->max_fds < ofdt->max_fds);
95
96 cpy = ofdt->max_fds * sizeof(struct file *);
97 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
98 memcpy(nfdt->fd, ofdt->fd, cpy);
99 memset((char *)nfdt->fd + cpy, 0, set);
100
101 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
102}
103
104static struct fdtable * alloc_fdtable(unsigned int nr)
105{
106 struct fdtable *fdt;
107 void *data;
108
109 /*
110 * Figure out how many fds we actually want to support in this fdtable.
111 * Allocation steps are keyed to the size of the fdarray, since it
112 * grows far faster than any of the other dynamic data. We try to fit
113 * the fdarray into comfortable page-tuned chunks: starting at 1024B
114 * and growing in powers of two from there on.
115 */
116 nr /= (1024 / sizeof(struct file *));
117 nr = roundup_pow_of_two(nr + 1);
118 nr *= (1024 / sizeof(struct file *));
119 /*
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 = alloc_fdmem(nr * sizeof(struct file *));
135 if (!data)
136 goto out_fdt;
137 fdt->fd = data;
138
139 data = alloc_fdmem(max_t(size_t,
140 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES));
141 if (!data)
142 goto out_arr;
143 fdt->open_fds = data;
144 data += nr / BITS_PER_BYTE;
145 fdt->close_on_exec = data;
146 data += nr / BITS_PER_BYTE;
147 fdt->full_fds_bits = data;
148
149 return fdt;
150
151out_arr:
152 kvfree(fdt->fd);
153out_fdt:
154 kfree(fdt);
155out:
156 return NULL;
157}
158
159/*
160 * Expand the file descriptor table.
161 * This function will allocate a new fdtable and both fd array and fdset, of
162 * the given size.
163 * Return <0 error code on error; 1 on successful completion.
164 * The files->file_lock should be held on entry, and will be held on exit.
165 */
166static int expand_fdtable(struct files_struct *files, unsigned int nr)
167 __releases(files->file_lock)
168 __acquires(files->file_lock)
169{
170 struct fdtable *new_fdt, *cur_fdt;
171
172 spin_unlock(&files->file_lock);
173 new_fdt = alloc_fdtable(nr);
174
175 /* make sure all __fd_install() have seen resize_in_progress
176 * or have finished their rcu_read_lock_sched() section.
177 */
178 if (atomic_read(&files->count) > 1)
179 synchronize_sched();
180
181 spin_lock(&files->file_lock);
182 if (!new_fdt)
183 return -ENOMEM;
184 /*
185 * extremely unlikely race - sysctl_nr_open decreased between the check in
186 * caller and alloc_fdtable(). Cheaper to catch it here...
187 */
188 if (unlikely(new_fdt->max_fds <= nr)) {
189 __free_fdtable(new_fdt);
190 return -EMFILE;
191 }
192 cur_fdt = files_fdtable(files);
193 BUG_ON(nr < cur_fdt->max_fds);
194 copy_fdtable(new_fdt, cur_fdt);
195 rcu_assign_pointer(files->fdt, new_fdt);
196 if (cur_fdt != &files->fdtab)
197 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
198 /* coupled with smp_rmb() in __fd_install() */
199 smp_wmb();
200 return 1;
201}
202
203/*
204 * Expand files.
205 * This function will expand the file structures, if the requested size exceeds
206 * the current capacity and there is room for expansion.
207 * Return <0 error code on error; 0 when nothing done; 1 when files were
208 * expanded and execution may have blocked.
209 * The files->file_lock should be held on entry, and will be held on exit.
210 */
211static int expand_files(struct files_struct *files, unsigned int nr)
212 __releases(files->file_lock)
213 __acquires(files->file_lock)
214{
215 struct fdtable *fdt;
216 int expanded = 0;
217
218repeat:
219 fdt = files_fdtable(files);
220
221 /* Do we need to expand? */
222 if (nr < fdt->max_fds)
223 return expanded;
224
225 /* Can we expand? */
226 if (nr >= sysctl_nr_open)
227 return -EMFILE;
228
229 if (unlikely(files->resize_in_progress)) {
230 spin_unlock(&files->file_lock);
231 expanded = 1;
232 wait_event(files->resize_wait, !files->resize_in_progress);
233 spin_lock(&files->file_lock);
234 goto repeat;
235 }
236
237 /* All good, so we try */
238 files->resize_in_progress = true;
239 expanded = expand_fdtable(files, nr);
240 files->resize_in_progress = false;
241
242 wake_up_all(&files->resize_wait);
243 return expanded;
244}
245
246static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
247{
248 __set_bit(fd, fdt->close_on_exec);
249}
250
251static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
252{
253 if (test_bit(fd, fdt->close_on_exec))
254 __clear_bit(fd, fdt->close_on_exec);
255}
256
257static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
258{
259 __set_bit(fd, fdt->open_fds);
260 fd /= BITS_PER_LONG;
261 if (!~fdt->open_fds[fd])
262 __set_bit(fd, fdt->full_fds_bits);
263}
264
265static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
266{
267 __clear_bit(fd, fdt->open_fds);
268 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
269}
270
271static unsigned int count_open_files(struct fdtable *fdt)
272{
273 unsigned int size = fdt->max_fds;
274 unsigned int i;
275
276 /* Find the last open fd */
277 for (i = size / BITS_PER_LONG; i > 0; ) {
278 if (fdt->open_fds[--i])
279 break;
280 }
281 i = (i + 1) * BITS_PER_LONG;
282 return i;
283}
284
285/*
286 * Allocate a new files structure and copy contents from the
287 * passed in files structure.
288 * errorp will be valid only when the returned files_struct is NULL.
289 */
290struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
291{
292 struct files_struct *newf;
293 struct file **old_fds, **new_fds;
294 unsigned int open_files, i;
295 struct fdtable *old_fdt, *new_fdt;
296
297 *errorp = -ENOMEM;
298 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
299 if (!newf)
300 goto out;
301
302 atomic_set(&newf->count, 1);
303
304 spin_lock_init(&newf->file_lock);
305 newf->resize_in_progress = false;
306 init_waitqueue_head(&newf->resize_wait);
307 newf->next_fd = 0;
308 new_fdt = &newf->fdtab;
309 new_fdt->max_fds = NR_OPEN_DEFAULT;
310 new_fdt->close_on_exec = newf->close_on_exec_init;
311 new_fdt->open_fds = newf->open_fds_init;
312 new_fdt->full_fds_bits = newf->full_fds_bits_init;
313 new_fdt->fd = &newf->fd_array[0];
314
315 spin_lock(&oldf->file_lock);
316 old_fdt = files_fdtable(oldf);
317 open_files = count_open_files(old_fdt);
318
319 /*
320 * Check whether we need to allocate a larger fd array and fd set.
321 */
322 while (unlikely(open_files > new_fdt->max_fds)) {
323 spin_unlock(&oldf->file_lock);
324
325 if (new_fdt != &newf->fdtab)
326 __free_fdtable(new_fdt);
327
328 new_fdt = alloc_fdtable(open_files - 1);
329 if (!new_fdt) {
330 *errorp = -ENOMEM;
331 goto out_release;
332 }
333
334 /* beyond sysctl_nr_open; nothing to do */
335 if (unlikely(new_fdt->max_fds < open_files)) {
336 __free_fdtable(new_fdt);
337 *errorp = -EMFILE;
338 goto out_release;
339 }
340
341 /*
342 * Reacquire the oldf lock and a pointer to its fd table
343 * who knows it may have a new bigger fd table. We need
344 * the latest pointer.
345 */
346 spin_lock(&oldf->file_lock);
347 old_fdt = files_fdtable(oldf);
348 open_files = count_open_files(old_fdt);
349 }
350
351 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
352
353 old_fds = old_fdt->fd;
354 new_fds = new_fdt->fd;
355
356 for (i = open_files; i != 0; i--) {
357 struct file *f = *old_fds++;
358 if (f) {
359 get_file(f);
360 } else {
361 /*
362 * The fd may be claimed in the fd bitmap but not yet
363 * instantiated in the files array if a sibling thread
364 * is partway through open(). So make sure that this
365 * fd is available to the new process.
366 */
367 __clear_open_fd(open_files - i, new_fdt);
368 }
369 rcu_assign_pointer(*new_fds++, f);
370 }
371 spin_unlock(&oldf->file_lock);
372
373 /* clear the remainder */
374 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
375
376 rcu_assign_pointer(newf->fdt, new_fdt);
377
378 return newf;
379
380out_release:
381 kmem_cache_free(files_cachep, newf);
382out:
383 return NULL;
384}
385
386static struct fdtable *close_files(struct files_struct * files)
387{
388 /*
389 * It is safe to dereference the fd table without RCU or
390 * ->file_lock because this is the last reference to the
391 * files structure.
392 */
393 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
394 unsigned int i, j = 0;
395
396 for (;;) {
397 unsigned long set;
398 i = j * BITS_PER_LONG;
399 if (i >= fdt->max_fds)
400 break;
401 set = fdt->open_fds[j++];
402 while (set) {
403 if (set & 1) {
404 struct file * file = xchg(&fdt->fd[i], NULL);
405 if (file) {
406 filp_close(file, files);
407 cond_resched_rcu_qs();
408 }
409 }
410 i++;
411 set >>= 1;
412 }
413 }
414
415 return fdt;
416}
417
418struct files_struct *get_files_struct(struct task_struct *task)
419{
420 struct files_struct *files;
421
422 task_lock(task);
423 files = task->files;
424 if (files)
425 atomic_inc(&files->count);
426 task_unlock(task);
427
428 return files;
429}
430
431void put_files_struct(struct files_struct *files)
432{
433 if (atomic_dec_and_test(&files->count)) {
434 struct fdtable *fdt = close_files(files);
435
436 /* free the arrays if they are not embedded */
437 if (fdt != &files->fdtab)
438 __free_fdtable(fdt);
439 kmem_cache_free(files_cachep, files);
440 }
441}
442
443void reset_files_struct(struct files_struct *files)
444{
445 struct task_struct *tsk = current;
446 struct files_struct *old;
447
448 old = tsk->files;
449 task_lock(tsk);
450 tsk->files = files;
451 task_unlock(tsk);
452 put_files_struct(old);
453}
454
455void exit_files(struct task_struct *tsk)
456{
457 struct files_struct * files = tsk->files;
458
459 if (files) {
460 task_lock(tsk);
461 tsk->files = NULL;
462 task_unlock(tsk);
463 put_files_struct(files);
464 }
465}
466
467struct files_struct init_files = {
468 .count = ATOMIC_INIT(1),
469 .fdt = &init_files.fdtab,
470 .fdtab = {
471 .max_fds = NR_OPEN_DEFAULT,
472 .fd = &init_files.fd_array[0],
473 .close_on_exec = init_files.close_on_exec_init,
474 .open_fds = init_files.open_fds_init,
475 .full_fds_bits = init_files.full_fds_bits_init,
476 },
477 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
478};
479
480static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
481{
482 unsigned int maxfd = fdt->max_fds;
483 unsigned int maxbit = maxfd / BITS_PER_LONG;
484 unsigned int bitbit = start / BITS_PER_LONG;
485
486 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
487 if (bitbit > maxfd)
488 return maxfd;
489 if (bitbit > start)
490 start = bitbit;
491 return find_next_zero_bit(fdt->open_fds, maxfd, start);
492}
493
494/*
495 * allocate a file descriptor, mark it busy.
496 */
497int __alloc_fd(struct files_struct *files,
498 unsigned start, unsigned end, unsigned flags)
499{
500 unsigned int fd;
501 int error;
502 struct fdtable *fdt;
503
504 spin_lock(&files->file_lock);
505repeat:
506 fdt = files_fdtable(files);
507 fd = start;
508 if (fd < files->next_fd)
509 fd = files->next_fd;
510
511 if (fd < fdt->max_fds)
512 fd = find_next_fd(fdt, fd);
513
514 /*
515 * N.B. For clone tasks sharing a files structure, this test
516 * will limit the total number of files that can be opened.
517 */
518 error = -EMFILE;
519 if (fd >= end)
520 goto out;
521
522 error = expand_files(files, fd);
523 if (error < 0)
524 goto out;
525
526 /*
527 * If we needed to expand the fs array we
528 * might have blocked - try again.
529 */
530 if (error)
531 goto repeat;
532
533 if (start <= files->next_fd)
534 files->next_fd = fd + 1;
535
536 __set_open_fd(fd, fdt);
537 if (flags & O_CLOEXEC)
538 __set_close_on_exec(fd, fdt);
539 else
540 __clear_close_on_exec(fd, fdt);
541 error = fd;
542#if 1
543 /* Sanity check */
544 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
545 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
546 rcu_assign_pointer(fdt->fd[fd], NULL);
547 }
548#endif
549
550out:
551 spin_unlock(&files->file_lock);
552 return error;
553}
554
555static int alloc_fd(unsigned start, unsigned flags)
556{
557 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
558}
559
560int get_unused_fd_flags(unsigned flags)
561{
562 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
563}
564EXPORT_SYMBOL(get_unused_fd_flags);
565
566static void __put_unused_fd(struct files_struct *files, unsigned int fd)
567{
568 struct fdtable *fdt = files_fdtable(files);
569 __clear_open_fd(fd, fdt);
570 if (fd < files->next_fd)
571 files->next_fd = fd;
572}
573
574void put_unused_fd(unsigned int fd)
575{
576 struct files_struct *files = current->files;
577 spin_lock(&files->file_lock);
578 __put_unused_fd(files, fd);
579 spin_unlock(&files->file_lock);
580}
581
582EXPORT_SYMBOL(put_unused_fd);
583
584/*
585 * Install a file pointer in the fd array.
586 *
587 * The VFS is full of places where we drop the files lock between
588 * setting the open_fds bitmap and installing the file in the file
589 * array. At any such point, we are vulnerable to a dup2() race
590 * installing a file in the array before us. We need to detect this and
591 * fput() the struct file we are about to overwrite in this case.
592 *
593 * It should never happen - if we allow dup2() do it, _really_ bad things
594 * will follow.
595 *
596 * NOTE: __fd_install() variant is really, really low-level; don't
597 * use it unless you are forced to by truly lousy API shoved down
598 * your throat. 'files' *MUST* be either current->files or obtained
599 * by get_files_struct(current) done by whoever had given it to you,
600 * or really bad things will happen. Normally you want to use
601 * fd_install() instead.
602 */
603
604void __fd_install(struct files_struct *files, unsigned int fd,
605 struct file *file)
606{
607 struct fdtable *fdt;
608
609 might_sleep();
610 rcu_read_lock_sched();
611
612 while (unlikely(files->resize_in_progress)) {
613 rcu_read_unlock_sched();
614 wait_event(files->resize_wait, !files->resize_in_progress);
615 rcu_read_lock_sched();
616 }
617 /* coupled with smp_wmb() in expand_fdtable() */
618 smp_rmb();
619 fdt = rcu_dereference_sched(files->fdt);
620 BUG_ON(fdt->fd[fd] != NULL);
621 rcu_assign_pointer(fdt->fd[fd], file);
622 rcu_read_unlock_sched();
623}
624
625void fd_install(unsigned int fd, struct file *file)
626{
627 __fd_install(current->files, fd, file);
628}
629
630EXPORT_SYMBOL(fd_install);
631
632/*
633 * The same warnings as for __alloc_fd()/__fd_install() apply here...
634 */
635int __close_fd(struct files_struct *files, unsigned fd)
636{
637 struct file *file;
638 struct fdtable *fdt;
639
640 spin_lock(&files->file_lock);
641 fdt = files_fdtable(files);
642 if (fd >= fdt->max_fds)
643 goto out_unlock;
644 file = fdt->fd[fd];
645 if (!file)
646 goto out_unlock;
647 rcu_assign_pointer(fdt->fd[fd], NULL);
648 __clear_close_on_exec(fd, fdt);
649 __put_unused_fd(files, fd);
650 spin_unlock(&files->file_lock);
651 return filp_close(file, files);
652
653out_unlock:
654 spin_unlock(&files->file_lock);
655 return -EBADF;
656}
657
658void do_close_on_exec(struct files_struct *files)
659{
660 unsigned i;
661 struct fdtable *fdt;
662
663 /* exec unshares first */
664 spin_lock(&files->file_lock);
665 for (i = 0; ; i++) {
666 unsigned long set;
667 unsigned fd = i * BITS_PER_LONG;
668 fdt = files_fdtable(files);
669 if (fd >= fdt->max_fds)
670 break;
671 set = fdt->close_on_exec[i];
672 if (!set)
673 continue;
674 fdt->close_on_exec[i] = 0;
675 for ( ; set ; fd++, set >>= 1) {
676 struct file *file;
677 if (!(set & 1))
678 continue;
679 file = fdt->fd[fd];
680 if (!file)
681 continue;
682 rcu_assign_pointer(fdt->fd[fd], NULL);
683 __put_unused_fd(files, fd);
684 spin_unlock(&files->file_lock);
685 filp_close(file, files);
686 cond_resched();
687 spin_lock(&files->file_lock);
688 }
689
690 }
691 spin_unlock(&files->file_lock);
692}
693
694static struct file *__fget(unsigned int fd, fmode_t mask)
695{
696 struct files_struct *files = current->files;
697 struct file *file;
698
699 rcu_read_lock();
700loop:
701 file = fcheck_files(files, fd);
702 if (file) {
703 /* File object ref couldn't be taken.
704 * dup2() atomicity guarantee is the reason
705 * we loop to catch the new file (or NULL pointer)
706 */
707 if (file->f_mode & mask)
708 file = NULL;
709 else if (!get_file_rcu(file))
710 goto loop;
711 }
712 rcu_read_unlock();
713
714 return file;
715}
716
717struct file *fget(unsigned int fd)
718{
719 return __fget(fd, FMODE_PATH);
720}
721EXPORT_SYMBOL(fget);
722
723struct file *fget_raw(unsigned int fd)
724{
725 return __fget(fd, 0);
726}
727EXPORT_SYMBOL(fget_raw);
728
729/*
730 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
731 *
732 * You can use this instead of fget if you satisfy all of the following
733 * conditions:
734 * 1) You must call fput_light before exiting the syscall and returning control
735 * to userspace (i.e. you cannot remember the returned struct file * after
736 * returning to userspace).
737 * 2) You must not call filp_close on the returned struct file * in between
738 * calls to fget_light and fput_light.
739 * 3) You must not clone the current task in between the calls to fget_light
740 * and fput_light.
741 *
742 * The fput_needed flag returned by fget_light should be passed to the
743 * corresponding fput_light.
744 */
745static unsigned long __fget_light(unsigned int fd, fmode_t mask)
746{
747 struct files_struct *files = current->files;
748 struct file *file;
749
750 if (atomic_read(&files->count) == 1) {
751 file = __fcheck_files(files, fd);
752 if (!file || unlikely(file->f_mode & mask))
753 return 0;
754 return (unsigned long)file;
755 } else {
756 file = __fget(fd, mask);
757 if (!file)
758 return 0;
759 return FDPUT_FPUT | (unsigned long)file;
760 }
761}
762unsigned long __fdget(unsigned int fd)
763{
764 return __fget_light(fd, FMODE_PATH);
765}
766EXPORT_SYMBOL(__fdget);
767
768unsigned long __fdget_raw(unsigned int fd)
769{
770 return __fget_light(fd, 0);
771}
772
773unsigned long __fdget_pos(unsigned int fd)
774{
775 unsigned long v = __fdget(fd);
776 struct file *file = (struct file *)(v & ~3);
777
778 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
779 if (file_count(file) > 1) {
780 v |= FDPUT_POS_UNLOCK;
781 mutex_lock(&file->f_pos_lock);
782 }
783 }
784 return v;
785}
786
787void __f_unlock_pos(struct file *f)
788{
789 mutex_unlock(&f->f_pos_lock);
790}
791
792/*
793 * We only lock f_pos if we have threads or if the file might be
794 * shared with another process. In both cases we'll have an elevated
795 * file count (done either by fdget() or by fork()).
796 */
797
798void set_close_on_exec(unsigned int fd, int flag)
799{
800 struct files_struct *files = current->files;
801 struct fdtable *fdt;
802 spin_lock(&files->file_lock);
803 fdt = files_fdtable(files);
804 if (flag)
805 __set_close_on_exec(fd, fdt);
806 else
807 __clear_close_on_exec(fd, fdt);
808 spin_unlock(&files->file_lock);
809}
810
811bool get_close_on_exec(unsigned int fd)
812{
813 struct files_struct *files = current->files;
814 struct fdtable *fdt;
815 bool res;
816 rcu_read_lock();
817 fdt = files_fdtable(files);
818 res = close_on_exec(fd, fdt);
819 rcu_read_unlock();
820 return res;
821}
822
823static int do_dup2(struct files_struct *files,
824 struct file *file, unsigned fd, unsigned flags)
825__releases(&files->file_lock)
826{
827 struct file *tofree;
828 struct fdtable *fdt;
829
830 /*
831 * We need to detect attempts to do dup2() over allocated but still
832 * not finished descriptor. NB: OpenBSD avoids that at the price of
833 * extra work in their equivalent of fget() - they insert struct
834 * file immediately after grabbing descriptor, mark it larval if
835 * more work (e.g. actual opening) is needed and make sure that
836 * fget() treats larval files as absent. Potentially interesting,
837 * but while extra work in fget() is trivial, locking implications
838 * and amount of surgery on open()-related paths in VFS are not.
839 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
840 * deadlocks in rather amusing ways, AFAICS. All of that is out of
841 * scope of POSIX or SUS, since neither considers shared descriptor
842 * tables and this condition does not arise without those.
843 */
844 fdt = files_fdtable(files);
845 tofree = fdt->fd[fd];
846 if (!tofree && fd_is_open(fd, fdt))
847 goto Ebusy;
848 get_file(file);
849 rcu_assign_pointer(fdt->fd[fd], file);
850 __set_open_fd(fd, fdt);
851 if (flags & O_CLOEXEC)
852 __set_close_on_exec(fd, fdt);
853 else
854 __clear_close_on_exec(fd, fdt);
855 spin_unlock(&files->file_lock);
856
857 if (tofree)
858 filp_close(tofree, files);
859
860 return fd;
861
862Ebusy:
863 spin_unlock(&files->file_lock);
864 return -EBUSY;
865}
866
867int replace_fd(unsigned fd, struct file *file, unsigned flags)
868{
869 int err;
870 struct files_struct *files = current->files;
871
872 if (!file)
873 return __close_fd(files, fd);
874
875 if (fd >= rlimit(RLIMIT_NOFILE))
876 return -EBADF;
877
878 spin_lock(&files->file_lock);
879 err = expand_files(files, fd);
880 if (unlikely(err < 0))
881 goto out_unlock;
882 return do_dup2(files, file, fd, flags);
883
884out_unlock:
885 spin_unlock(&files->file_lock);
886 return err;
887}
888
889SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
890{
891 int err = -EBADF;
892 struct file *file;
893 struct files_struct *files = current->files;
894
895 if ((flags & ~O_CLOEXEC) != 0)
896 return -EINVAL;
897
898 if (unlikely(oldfd == newfd))
899 return -EINVAL;
900
901 if (newfd >= rlimit(RLIMIT_NOFILE))
902 return -EBADF;
903
904 spin_lock(&files->file_lock);
905 err = expand_files(files, newfd);
906 file = fcheck(oldfd);
907 if (unlikely(!file))
908 goto Ebadf;
909 if (unlikely(err < 0)) {
910 if (err == -EMFILE)
911 goto Ebadf;
912 goto out_unlock;
913 }
914 return do_dup2(files, file, newfd, flags);
915
916Ebadf:
917 err = -EBADF;
918out_unlock:
919 spin_unlock(&files->file_lock);
920 return err;
921}
922
923SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
924{
925 if (unlikely(newfd == oldfd)) { /* corner case */
926 struct files_struct *files = current->files;
927 int retval = oldfd;
928
929 rcu_read_lock();
930 if (!fcheck_files(files, oldfd))
931 retval = -EBADF;
932 rcu_read_unlock();
933 return retval;
934 }
935 return sys_dup3(oldfd, newfd, 0);
936}
937
938SYSCALL_DEFINE1(dup, unsigned int, fildes)
939{
940 int ret = -EBADF;
941 struct file *file = fget_raw(fildes);
942
943 if (file) {
944 ret = get_unused_fd_flags(0);
945 if (ret >= 0)
946 fd_install(ret, file);
947 else
948 fput(file);
949 }
950 return ret;
951}
952
953int f_dupfd(unsigned int from, struct file *file, unsigned flags)
954{
955 int err;
956 if (from >= rlimit(RLIMIT_NOFILE))
957 return -EINVAL;
958 err = alloc_fd(from, flags);
959 if (err >= 0) {
960 get_file(file);
961 fd_install(err, file);
962 }
963 return err;
964}
965
966int iterate_fd(struct files_struct *files, unsigned n,
967 int (*f)(const void *, struct file *, unsigned),
968 const void *p)
969{
970 struct fdtable *fdt;
971 int res = 0;
972 if (!files)
973 return 0;
974 spin_lock(&files->file_lock);
975 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
976 struct file *file;
977 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
978 if (!file)
979 continue;
980 res = f(p, file, n);
981 if (res)
982 break;
983 }
984 spin_unlock(&files->file_lock);
985 return res;
986}
987EXPORT_SYMBOL(iterate_fd);
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(fdt->fd);
37 kvfree(fdt->open_fds);
38 kfree(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, ofdt->max_fds);
88}
89
90static struct fdtable * alloc_fdtable(unsigned int nr)
91{
92 struct fdtable *fdt;
93 void *data;
94
95 /*
96 * Figure out how many fds we actually want to support in this fdtable.
97 * Allocation steps are keyed to the size of the fdarray, since it
98 * grows far faster than any of the other dynamic data. We try to fit
99 * the fdarray into comfortable page-tuned chunks: starting at 1024B
100 * and growing in powers of two from there on.
101 */
102 nr /= (1024 / sizeof(struct file *));
103 nr = roundup_pow_of_two(nr + 1);
104 nr *= (1024 / sizeof(struct file *));
105 /*
106 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
107 * had been set lower between the check in expand_files() and here. Deal
108 * with that in caller, it's cheaper that way.
109 *
110 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
111 * bitmaps handling below becomes unpleasant, to put it mildly...
112 */
113 if (unlikely(nr > sysctl_nr_open))
114 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
115
116 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
117 if (!fdt)
118 goto out;
119 fdt->max_fds = nr;
120 data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
121 if (!data)
122 goto out_fdt;
123 fdt->fd = data;
124
125 data = kvmalloc(max_t(size_t,
126 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
127 GFP_KERNEL_ACCOUNT);
128 if (!data)
129 goto out_arr;
130 fdt->open_fds = data;
131 data += nr / BITS_PER_BYTE;
132 fdt->close_on_exec = data;
133 data += nr / BITS_PER_BYTE;
134 fdt->full_fds_bits = data;
135
136 return fdt;
137
138out_arr:
139 kvfree(fdt->fd);
140out_fdt:
141 kfree(fdt);
142out:
143 return NULL;
144}
145
146/*
147 * Expand the file descriptor table.
148 * This function will allocate a new fdtable and both fd array and fdset, of
149 * the given size.
150 * Return <0 error code on error; 1 on successful completion.
151 * The files->file_lock should be held on entry, and will be held on exit.
152 */
153static int expand_fdtable(struct files_struct *files, unsigned int nr)
154 __releases(files->file_lock)
155 __acquires(files->file_lock)
156{
157 struct fdtable *new_fdt, *cur_fdt;
158
159 spin_unlock(&files->file_lock);
160 new_fdt = alloc_fdtable(nr);
161
162 /* make sure all fd_install() have seen resize_in_progress
163 * or have finished their rcu_read_lock_sched() section.
164 */
165 if (atomic_read(&files->count) > 1)
166 synchronize_rcu();
167
168 spin_lock(&files->file_lock);
169 if (!new_fdt)
170 return -ENOMEM;
171 /*
172 * extremely unlikely race - sysctl_nr_open decreased between the check in
173 * caller and alloc_fdtable(). Cheaper to catch it here...
174 */
175 if (unlikely(new_fdt->max_fds <= nr)) {
176 __free_fdtable(new_fdt);
177 return -EMFILE;
178 }
179 cur_fdt = files_fdtable(files);
180 BUG_ON(nr < cur_fdt->max_fds);
181 copy_fdtable(new_fdt, cur_fdt);
182 rcu_assign_pointer(files->fdt, new_fdt);
183 if (cur_fdt != &files->fdtab)
184 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
185 /* coupled with smp_rmb() in fd_install() */
186 smp_wmb();
187 return 1;
188}
189
190/*
191 * Expand files.
192 * This function will expand the file structures, if the requested size exceeds
193 * the current capacity and there is room for expansion.
194 * Return <0 error code on error; 0 when nothing done; 1 when files were
195 * expanded and execution may have blocked.
196 * The files->file_lock should be held on entry, and will be held on exit.
197 */
198static int expand_files(struct files_struct *files, unsigned int nr)
199 __releases(files->file_lock)
200 __acquires(files->file_lock)
201{
202 struct fdtable *fdt;
203 int expanded = 0;
204
205repeat:
206 fdt = files_fdtable(files);
207
208 /* Do we need to expand? */
209 if (nr < fdt->max_fds)
210 return expanded;
211
212 /* Can we expand? */
213 if (nr >= sysctl_nr_open)
214 return -EMFILE;
215
216 if (unlikely(files->resize_in_progress)) {
217 spin_unlock(&files->file_lock);
218 expanded = 1;
219 wait_event(files->resize_wait, !files->resize_in_progress);
220 spin_lock(&files->file_lock);
221 goto repeat;
222 }
223
224 /* All good, so we try */
225 files->resize_in_progress = true;
226 expanded = expand_fdtable(files, nr);
227 files->resize_in_progress = false;
228
229 wake_up_all(&files->resize_wait);
230 return expanded;
231}
232
233static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
234{
235 __set_bit(fd, fdt->close_on_exec);
236}
237
238static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
239{
240 if (test_bit(fd, fdt->close_on_exec))
241 __clear_bit(fd, fdt->close_on_exec);
242}
243
244static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
245{
246 __set_bit(fd, fdt->open_fds);
247 fd /= BITS_PER_LONG;
248 if (!~fdt->open_fds[fd])
249 __set_bit(fd, fdt->full_fds_bits);
250}
251
252static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
253{
254 __clear_bit(fd, fdt->open_fds);
255 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
256}
257
258static unsigned int count_open_files(struct fdtable *fdt)
259{
260 unsigned int size = fdt->max_fds;
261 unsigned int i;
262
263 /* Find the last open fd */
264 for (i = size / BITS_PER_LONG; i > 0; ) {
265 if (fdt->open_fds[--i])
266 break;
267 }
268 i = (i + 1) * BITS_PER_LONG;
269 return i;
270}
271
272static unsigned int sane_fdtable_size(struct fdtable *fdt, unsigned int max_fds)
273{
274 unsigned int count;
275
276 count = count_open_files(fdt);
277 if (max_fds < NR_OPEN_DEFAULT)
278 max_fds = NR_OPEN_DEFAULT;
279 return min(count, max_fds);
280}
281
282/*
283 * Allocate a new files structure and copy contents from the
284 * passed in files structure.
285 * errorp will be valid only when the returned files_struct is NULL.
286 */
287struct files_struct *dup_fd(struct files_struct *oldf, unsigned int max_fds, int *errorp)
288{
289 struct files_struct *newf;
290 struct file **old_fds, **new_fds;
291 unsigned int open_files, i;
292 struct fdtable *old_fdt, *new_fdt;
293
294 *errorp = -ENOMEM;
295 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
296 if (!newf)
297 goto out;
298
299 atomic_set(&newf->count, 1);
300
301 spin_lock_init(&newf->file_lock);
302 newf->resize_in_progress = false;
303 init_waitqueue_head(&newf->resize_wait);
304 newf->next_fd = 0;
305 new_fdt = &newf->fdtab;
306 new_fdt->max_fds = NR_OPEN_DEFAULT;
307 new_fdt->close_on_exec = newf->close_on_exec_init;
308 new_fdt->open_fds = newf->open_fds_init;
309 new_fdt->full_fds_bits = newf->full_fds_bits_init;
310 new_fdt->fd = &newf->fd_array[0];
311
312 spin_lock(&oldf->file_lock);
313 old_fdt = files_fdtable(oldf);
314 open_files = sane_fdtable_size(old_fdt, max_fds);
315
316 /*
317 * Check whether we need to allocate a larger fd array and fd set.
318 */
319 while (unlikely(open_files > new_fdt->max_fds)) {
320 spin_unlock(&oldf->file_lock);
321
322 if (new_fdt != &newf->fdtab)
323 __free_fdtable(new_fdt);
324
325 new_fdt = alloc_fdtable(open_files - 1);
326 if (!new_fdt) {
327 *errorp = -ENOMEM;
328 goto out_release;
329 }
330
331 /* beyond sysctl_nr_open; nothing to do */
332 if (unlikely(new_fdt->max_fds < open_files)) {
333 __free_fdtable(new_fdt);
334 *errorp = -EMFILE;
335 goto out_release;
336 }
337
338 /*
339 * Reacquire the oldf lock and a pointer to its fd table
340 * who knows it may have a new bigger fd table. We need
341 * the latest pointer.
342 */
343 spin_lock(&oldf->file_lock);
344 old_fdt = files_fdtable(oldf);
345 open_files = sane_fdtable_size(old_fdt, max_fds);
346 }
347
348 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
349
350 old_fds = old_fdt->fd;
351 new_fds = new_fdt->fd;
352
353 for (i = open_files; i != 0; i--) {
354 struct file *f = *old_fds++;
355 if (f) {
356 get_file(f);
357 } else {
358 /*
359 * The fd may be claimed in the fd bitmap but not yet
360 * instantiated in the files array if a sibling thread
361 * is partway through open(). So make sure that this
362 * fd is available to the new process.
363 */
364 __clear_open_fd(open_files - i, new_fdt);
365 }
366 rcu_assign_pointer(*new_fds++, f);
367 }
368 spin_unlock(&oldf->file_lock);
369
370 /* clear the remainder */
371 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
372
373 rcu_assign_pointer(newf->fdt, new_fdt);
374
375 return newf;
376
377out_release:
378 kmem_cache_free(files_cachep, newf);
379out:
380 return NULL;
381}
382
383static struct fdtable *close_files(struct files_struct * files)
384{
385 /*
386 * It is safe to dereference the fd table without RCU or
387 * ->file_lock because this is the last reference to the
388 * files structure.
389 */
390 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
391 unsigned int i, j = 0;
392
393 for (;;) {
394 unsigned long set;
395 i = j * BITS_PER_LONG;
396 if (i >= fdt->max_fds)
397 break;
398 set = fdt->open_fds[j++];
399 while (set) {
400 if (set & 1) {
401 struct file * file = xchg(&fdt->fd[i], NULL);
402 if (file) {
403 filp_close(file, files);
404 cond_resched();
405 }
406 }
407 i++;
408 set >>= 1;
409 }
410 }
411
412 return fdt;
413}
414
415void put_files_struct(struct files_struct *files)
416{
417 if (atomic_dec_and_test(&files->count)) {
418 struct fdtable *fdt = close_files(files);
419
420 /* free the arrays if they are not embedded */
421 if (fdt != &files->fdtab)
422 __free_fdtable(fdt);
423 kmem_cache_free(files_cachep, files);
424 }
425}
426
427void exit_files(struct task_struct *tsk)
428{
429 struct files_struct * files = tsk->files;
430
431 if (files) {
432 task_lock(tsk);
433 tsk->files = NULL;
434 task_unlock(tsk);
435 put_files_struct(files);
436 }
437}
438
439struct files_struct init_files = {
440 .count = ATOMIC_INIT(1),
441 .fdt = &init_files.fdtab,
442 .fdtab = {
443 .max_fds = NR_OPEN_DEFAULT,
444 .fd = &init_files.fd_array[0],
445 .close_on_exec = init_files.close_on_exec_init,
446 .open_fds = init_files.open_fds_init,
447 .full_fds_bits = init_files.full_fds_bits_init,
448 },
449 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
450 .resize_wait = __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
451};
452
453static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
454{
455 unsigned int maxfd = fdt->max_fds;
456 unsigned int maxbit = maxfd / BITS_PER_LONG;
457 unsigned int bitbit = start / BITS_PER_LONG;
458
459 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
460 if (bitbit > maxfd)
461 return maxfd;
462 if (bitbit > start)
463 start = bitbit;
464 return find_next_zero_bit(fdt->open_fds, maxfd, start);
465}
466
467/*
468 * allocate a file descriptor, mark it busy.
469 */
470static int alloc_fd(unsigned start, unsigned end, unsigned flags)
471{
472 struct files_struct *files = current->files;
473 unsigned int fd;
474 int error;
475 struct fdtable *fdt;
476
477 spin_lock(&files->file_lock);
478repeat:
479 fdt = files_fdtable(files);
480 fd = start;
481 if (fd < files->next_fd)
482 fd = files->next_fd;
483
484 if (fd < fdt->max_fds)
485 fd = find_next_fd(fdt, fd);
486
487 /*
488 * N.B. For clone tasks sharing a files structure, this test
489 * will limit the total number of files that can be opened.
490 */
491 error = -EMFILE;
492 if (fd >= end)
493 goto out;
494
495 error = expand_files(files, fd);
496 if (error < 0)
497 goto out;
498
499 /*
500 * If we needed to expand the fs array we
501 * might have blocked - try again.
502 */
503 if (error)
504 goto repeat;
505
506 if (start <= files->next_fd)
507 files->next_fd = fd + 1;
508
509 __set_open_fd(fd, fdt);
510 if (flags & O_CLOEXEC)
511 __set_close_on_exec(fd, fdt);
512 else
513 __clear_close_on_exec(fd, fdt);
514 error = fd;
515#if 1
516 /* Sanity check */
517 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
518 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
519 rcu_assign_pointer(fdt->fd[fd], NULL);
520 }
521#endif
522
523out:
524 spin_unlock(&files->file_lock);
525 return error;
526}
527
528int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
529{
530 return alloc_fd(0, nofile, flags);
531}
532
533int get_unused_fd_flags(unsigned flags)
534{
535 return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
536}
537EXPORT_SYMBOL(get_unused_fd_flags);
538
539static void __put_unused_fd(struct files_struct *files, unsigned int fd)
540{
541 struct fdtable *fdt = files_fdtable(files);
542 __clear_open_fd(fd, fdt);
543 if (fd < files->next_fd)
544 files->next_fd = fd;
545}
546
547void put_unused_fd(unsigned int fd)
548{
549 struct files_struct *files = current->files;
550 spin_lock(&files->file_lock);
551 __put_unused_fd(files, fd);
552 spin_unlock(&files->file_lock);
553}
554
555EXPORT_SYMBOL(put_unused_fd);
556
557/*
558 * Install a file pointer in the fd array.
559 *
560 * The VFS is full of places where we drop the files lock between
561 * setting the open_fds bitmap and installing the file in the file
562 * array. At any such point, we are vulnerable to a dup2() race
563 * installing a file in the array before us. We need to detect this and
564 * fput() the struct file we are about to overwrite in this case.
565 *
566 * It should never happen - if we allow dup2() do it, _really_ bad things
567 * will follow.
568 *
569 * This consumes the "file" refcount, so callers should treat it
570 * as if they had called fput(file).
571 */
572
573void fd_install(unsigned int fd, struct file *file)
574{
575 struct files_struct *files = current->files;
576 struct fdtable *fdt;
577
578 rcu_read_lock_sched();
579
580 if (unlikely(files->resize_in_progress)) {
581 rcu_read_unlock_sched();
582 spin_lock(&files->file_lock);
583 fdt = files_fdtable(files);
584 BUG_ON(fdt->fd[fd] != NULL);
585 rcu_assign_pointer(fdt->fd[fd], file);
586 spin_unlock(&files->file_lock);
587 return;
588 }
589 /* coupled with smp_wmb() in expand_fdtable() */
590 smp_rmb();
591 fdt = rcu_dereference_sched(files->fdt);
592 BUG_ON(fdt->fd[fd] != NULL);
593 rcu_assign_pointer(fdt->fd[fd], file);
594 rcu_read_unlock_sched();
595}
596
597EXPORT_SYMBOL(fd_install);
598
599static struct file *pick_file(struct files_struct *files, unsigned fd)
600{
601 struct file *file = NULL;
602 struct fdtable *fdt;
603
604 spin_lock(&files->file_lock);
605 fdt = files_fdtable(files);
606 if (fd >= fdt->max_fds)
607 goto out_unlock;
608 file = fdt->fd[fd];
609 if (!file)
610 goto out_unlock;
611 rcu_assign_pointer(fdt->fd[fd], NULL);
612 __put_unused_fd(files, fd);
613
614out_unlock:
615 spin_unlock(&files->file_lock);
616 return file;
617}
618
619int close_fd(unsigned fd)
620{
621 struct files_struct *files = current->files;
622 struct file *file;
623
624 file = pick_file(files, fd);
625 if (!file)
626 return -EBADF;
627
628 return filp_close(file, files);
629}
630EXPORT_SYMBOL(close_fd); /* for ksys_close() */
631
632/**
633 * last_fd - return last valid index into fd table
634 * @cur_fds: files struct
635 *
636 * Context: Either rcu read lock or files_lock must be held.
637 *
638 * Returns: Last valid index into fdtable.
639 */
640static inline unsigned last_fd(struct fdtable *fdt)
641{
642 return fdt->max_fds - 1;
643}
644
645static inline void __range_cloexec(struct files_struct *cur_fds,
646 unsigned int fd, unsigned int max_fd)
647{
648 struct fdtable *fdt;
649
650 /* make sure we're using the correct maximum value */
651 spin_lock(&cur_fds->file_lock);
652 fdt = files_fdtable(cur_fds);
653 max_fd = min(last_fd(fdt), max_fd);
654 if (fd <= max_fd)
655 bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
656 spin_unlock(&cur_fds->file_lock);
657}
658
659static inline void __range_close(struct files_struct *cur_fds, unsigned int fd,
660 unsigned int max_fd)
661{
662 while (fd <= max_fd) {
663 struct file *file;
664
665 file = pick_file(cur_fds, fd++);
666 if (!file)
667 continue;
668
669 filp_close(file, cur_fds);
670 cond_resched();
671 }
672}
673
674/**
675 * __close_range() - Close all file descriptors in a given range.
676 *
677 * @fd: starting file descriptor to close
678 * @max_fd: last file descriptor to close
679 *
680 * This closes a range of file descriptors. All file descriptors
681 * from @fd up to and including @max_fd are closed.
682 */
683int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
684{
685 unsigned int cur_max;
686 struct task_struct *me = current;
687 struct files_struct *cur_fds = me->files, *fds = NULL;
688
689 if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
690 return -EINVAL;
691
692 if (fd > max_fd)
693 return -EINVAL;
694
695 rcu_read_lock();
696 cur_max = files_fdtable(cur_fds)->max_fds;
697 rcu_read_unlock();
698
699 /* cap to last valid index into fdtable */
700 cur_max--;
701
702 if (flags & CLOSE_RANGE_UNSHARE) {
703 int ret;
704 unsigned int max_unshare_fds = NR_OPEN_MAX;
705
706 /*
707 * If the requested range is greater than the current maximum,
708 * we're closing everything so only copy all file descriptors
709 * beneath the lowest file descriptor.
710 * If the caller requested all fds to be made cloexec copy all
711 * of the file descriptors since they still want to use them.
712 */
713 if (!(flags & CLOSE_RANGE_CLOEXEC) && (max_fd >= cur_max))
714 max_unshare_fds = fd;
715
716 ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
717 if (ret)
718 return ret;
719
720 /*
721 * We used to share our file descriptor table, and have now
722 * created a private one, make sure we're using it below.
723 */
724 if (fds)
725 swap(cur_fds, fds);
726 }
727
728 max_fd = min(max_fd, cur_max);
729
730 if (flags & CLOSE_RANGE_CLOEXEC)
731 __range_cloexec(cur_fds, fd, max_fd);
732 else
733 __range_close(cur_fds, fd, max_fd);
734
735 if (fds) {
736 /*
737 * We're done closing the files we were supposed to. Time to install
738 * the new file descriptor table and drop the old one.
739 */
740 task_lock(me);
741 me->files = cur_fds;
742 task_unlock(me);
743 put_files_struct(fds);
744 }
745
746 return 0;
747}
748
749/*
750 * See close_fd_get_file() below, this variant assumes current->files->file_lock
751 * is held.
752 */
753int __close_fd_get_file(unsigned int fd, struct file **res)
754{
755 struct files_struct *files = current->files;
756 struct file *file;
757 struct fdtable *fdt;
758
759 fdt = files_fdtable(files);
760 if (fd >= fdt->max_fds)
761 goto out_err;
762 file = fdt->fd[fd];
763 if (!file)
764 goto out_err;
765 rcu_assign_pointer(fdt->fd[fd], NULL);
766 __put_unused_fd(files, fd);
767 get_file(file);
768 *res = file;
769 return 0;
770out_err:
771 *res = NULL;
772 return -ENOENT;
773}
774
775/*
776 * variant of close_fd that gets a ref on the file for later fput.
777 * The caller must ensure that filp_close() called on the file, and then
778 * an fput().
779 */
780int close_fd_get_file(unsigned int fd, struct file **res)
781{
782 struct files_struct *files = current->files;
783 int ret;
784
785 spin_lock(&files->file_lock);
786 ret = __close_fd_get_file(fd, res);
787 spin_unlock(&files->file_lock);
788
789 return ret;
790}
791
792void do_close_on_exec(struct files_struct *files)
793{
794 unsigned i;
795 struct fdtable *fdt;
796
797 /* exec unshares first */
798 spin_lock(&files->file_lock);
799 for (i = 0; ; i++) {
800 unsigned long set;
801 unsigned fd = i * BITS_PER_LONG;
802 fdt = files_fdtable(files);
803 if (fd >= fdt->max_fds)
804 break;
805 set = fdt->close_on_exec[i];
806 if (!set)
807 continue;
808 fdt->close_on_exec[i] = 0;
809 for ( ; set ; fd++, set >>= 1) {
810 struct file *file;
811 if (!(set & 1))
812 continue;
813 file = fdt->fd[fd];
814 if (!file)
815 continue;
816 rcu_assign_pointer(fdt->fd[fd], NULL);
817 __put_unused_fd(files, fd);
818 spin_unlock(&files->file_lock);
819 filp_close(file, files);
820 cond_resched();
821 spin_lock(&files->file_lock);
822 }
823
824 }
825 spin_unlock(&files->file_lock);
826}
827
828static struct file *__fget_files(struct files_struct *files, unsigned int fd,
829 fmode_t mask, unsigned int refs)
830{
831 struct file *file;
832
833 rcu_read_lock();
834loop:
835 file = files_lookup_fd_rcu(files, fd);
836 if (file) {
837 /* File object ref couldn't be taken.
838 * dup2() atomicity guarantee is the reason
839 * we loop to catch the new file (or NULL pointer)
840 */
841 if (file->f_mode & mask)
842 file = NULL;
843 else if (!get_file_rcu_many(file, refs))
844 goto loop;
845 }
846 rcu_read_unlock();
847
848 return file;
849}
850
851static inline struct file *__fget(unsigned int fd, fmode_t mask,
852 unsigned int refs)
853{
854 return __fget_files(current->files, fd, mask, refs);
855}
856
857struct file *fget_many(unsigned int fd, unsigned int refs)
858{
859 return __fget(fd, FMODE_PATH, refs);
860}
861
862struct file *fget(unsigned int fd)
863{
864 return __fget(fd, FMODE_PATH, 1);
865}
866EXPORT_SYMBOL(fget);
867
868struct file *fget_raw(unsigned int fd)
869{
870 return __fget(fd, 0, 1);
871}
872EXPORT_SYMBOL(fget_raw);
873
874struct file *fget_task(struct task_struct *task, unsigned int fd)
875{
876 struct file *file = NULL;
877
878 task_lock(task);
879 if (task->files)
880 file = __fget_files(task->files, fd, 0, 1);
881 task_unlock(task);
882
883 return file;
884}
885
886struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
887{
888 /* Must be called with rcu_read_lock held */
889 struct files_struct *files;
890 struct file *file = NULL;
891
892 task_lock(task);
893 files = task->files;
894 if (files)
895 file = files_lookup_fd_rcu(files, fd);
896 task_unlock(task);
897
898 return file;
899}
900
901struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret_fd)
902{
903 /* Must be called with rcu_read_lock held */
904 struct files_struct *files;
905 unsigned int fd = *ret_fd;
906 struct file *file = NULL;
907
908 task_lock(task);
909 files = task->files;
910 if (files) {
911 for (; fd < files_fdtable(files)->max_fds; fd++) {
912 file = files_lookup_fd_rcu(files, fd);
913 if (file)
914 break;
915 }
916 }
917 task_unlock(task);
918 *ret_fd = fd;
919 return file;
920}
921
922/*
923 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
924 *
925 * You can use this instead of fget if you satisfy all of the following
926 * conditions:
927 * 1) You must call fput_light before exiting the syscall and returning control
928 * to userspace (i.e. you cannot remember the returned struct file * after
929 * returning to userspace).
930 * 2) You must not call filp_close on the returned struct file * in between
931 * calls to fget_light and fput_light.
932 * 3) You must not clone the current task in between the calls to fget_light
933 * and fput_light.
934 *
935 * The fput_needed flag returned by fget_light should be passed to the
936 * corresponding fput_light.
937 */
938static unsigned long __fget_light(unsigned int fd, fmode_t mask)
939{
940 struct files_struct *files = current->files;
941 struct file *file;
942
943 if (atomic_read(&files->count) == 1) {
944 file = files_lookup_fd_raw(files, fd);
945 if (!file || unlikely(file->f_mode & mask))
946 return 0;
947 return (unsigned long)file;
948 } else {
949 file = __fget(fd, mask, 1);
950 if (!file)
951 return 0;
952 return FDPUT_FPUT | (unsigned long)file;
953 }
954}
955unsigned long __fdget(unsigned int fd)
956{
957 return __fget_light(fd, FMODE_PATH);
958}
959EXPORT_SYMBOL(__fdget);
960
961unsigned long __fdget_raw(unsigned int fd)
962{
963 return __fget_light(fd, 0);
964}
965
966unsigned long __fdget_pos(unsigned int fd)
967{
968 unsigned long v = __fdget(fd);
969 struct file *file = (struct file *)(v & ~3);
970
971 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
972 if (file_count(file) > 1) {
973 v |= FDPUT_POS_UNLOCK;
974 mutex_lock(&file->f_pos_lock);
975 }
976 }
977 return v;
978}
979
980void __f_unlock_pos(struct file *f)
981{
982 mutex_unlock(&f->f_pos_lock);
983}
984
985/*
986 * We only lock f_pos if we have threads or if the file might be
987 * shared with another process. In both cases we'll have an elevated
988 * file count (done either by fdget() or by fork()).
989 */
990
991void set_close_on_exec(unsigned int fd, int flag)
992{
993 struct files_struct *files = current->files;
994 struct fdtable *fdt;
995 spin_lock(&files->file_lock);
996 fdt = files_fdtable(files);
997 if (flag)
998 __set_close_on_exec(fd, fdt);
999 else
1000 __clear_close_on_exec(fd, fdt);
1001 spin_unlock(&files->file_lock);
1002}
1003
1004bool get_close_on_exec(unsigned int fd)
1005{
1006 struct files_struct *files = current->files;
1007 struct fdtable *fdt;
1008 bool res;
1009 rcu_read_lock();
1010 fdt = files_fdtable(files);
1011 res = close_on_exec(fd, fdt);
1012 rcu_read_unlock();
1013 return res;
1014}
1015
1016static int do_dup2(struct files_struct *files,
1017 struct file *file, unsigned fd, unsigned flags)
1018__releases(&files->file_lock)
1019{
1020 struct file *tofree;
1021 struct fdtable *fdt;
1022
1023 /*
1024 * We need to detect attempts to do dup2() over allocated but still
1025 * not finished descriptor. NB: OpenBSD avoids that at the price of
1026 * extra work in their equivalent of fget() - they insert struct
1027 * file immediately after grabbing descriptor, mark it larval if
1028 * more work (e.g. actual opening) is needed and make sure that
1029 * fget() treats larval files as absent. Potentially interesting,
1030 * but while extra work in fget() is trivial, locking implications
1031 * and amount of surgery on open()-related paths in VFS are not.
1032 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1033 * deadlocks in rather amusing ways, AFAICS. All of that is out of
1034 * scope of POSIX or SUS, since neither considers shared descriptor
1035 * tables and this condition does not arise without those.
1036 */
1037 fdt = files_fdtable(files);
1038 tofree = fdt->fd[fd];
1039 if (!tofree && fd_is_open(fd, fdt))
1040 goto Ebusy;
1041 get_file(file);
1042 rcu_assign_pointer(fdt->fd[fd], file);
1043 __set_open_fd(fd, fdt);
1044 if (flags & O_CLOEXEC)
1045 __set_close_on_exec(fd, fdt);
1046 else
1047 __clear_close_on_exec(fd, fdt);
1048 spin_unlock(&files->file_lock);
1049
1050 if (tofree)
1051 filp_close(tofree, files);
1052
1053 return fd;
1054
1055Ebusy:
1056 spin_unlock(&files->file_lock);
1057 return -EBUSY;
1058}
1059
1060int replace_fd(unsigned fd, struct file *file, unsigned flags)
1061{
1062 int err;
1063 struct files_struct *files = current->files;
1064
1065 if (!file)
1066 return close_fd(fd);
1067
1068 if (fd >= rlimit(RLIMIT_NOFILE))
1069 return -EBADF;
1070
1071 spin_lock(&files->file_lock);
1072 err = expand_files(files, fd);
1073 if (unlikely(err < 0))
1074 goto out_unlock;
1075 return do_dup2(files, file, fd, flags);
1076
1077out_unlock:
1078 spin_unlock(&files->file_lock);
1079 return err;
1080}
1081
1082/**
1083 * __receive_fd() - Install received file into file descriptor table
1084 * @file: struct file that was received from another process
1085 * @ufd: __user pointer to write new fd number to
1086 * @o_flags: the O_* flags to apply to the new fd entry
1087 *
1088 * Installs a received file into the file descriptor table, with appropriate
1089 * checks and count updates. Optionally writes the fd number to userspace, if
1090 * @ufd is non-NULL.
1091 *
1092 * This helper handles its own reference counting of the incoming
1093 * struct file.
1094 *
1095 * Returns newly install fd or -ve on error.
1096 */
1097int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1098{
1099 int new_fd;
1100 int error;
1101
1102 error = security_file_receive(file);
1103 if (error)
1104 return error;
1105
1106 new_fd = get_unused_fd_flags(o_flags);
1107 if (new_fd < 0)
1108 return new_fd;
1109
1110 if (ufd) {
1111 error = put_user(new_fd, ufd);
1112 if (error) {
1113 put_unused_fd(new_fd);
1114 return error;
1115 }
1116 }
1117
1118 fd_install(new_fd, get_file(file));
1119 __receive_sock(file);
1120 return new_fd;
1121}
1122
1123int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1124{
1125 int error;
1126
1127 error = security_file_receive(file);
1128 if (error)
1129 return error;
1130 error = replace_fd(new_fd, file, o_flags);
1131 if (error)
1132 return error;
1133 __receive_sock(file);
1134 return new_fd;
1135}
1136
1137static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1138{
1139 int err = -EBADF;
1140 struct file *file;
1141 struct files_struct *files = current->files;
1142
1143 if ((flags & ~O_CLOEXEC) != 0)
1144 return -EINVAL;
1145
1146 if (unlikely(oldfd == newfd))
1147 return -EINVAL;
1148
1149 if (newfd >= rlimit(RLIMIT_NOFILE))
1150 return -EBADF;
1151
1152 spin_lock(&files->file_lock);
1153 err = expand_files(files, newfd);
1154 file = files_lookup_fd_locked(files, oldfd);
1155 if (unlikely(!file))
1156 goto Ebadf;
1157 if (unlikely(err < 0)) {
1158 if (err == -EMFILE)
1159 goto Ebadf;
1160 goto out_unlock;
1161 }
1162 return do_dup2(files, file, newfd, flags);
1163
1164Ebadf:
1165 err = -EBADF;
1166out_unlock:
1167 spin_unlock(&files->file_lock);
1168 return err;
1169}
1170
1171SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1172{
1173 return ksys_dup3(oldfd, newfd, flags);
1174}
1175
1176SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1177{
1178 if (unlikely(newfd == oldfd)) { /* corner case */
1179 struct files_struct *files = current->files;
1180 int retval = oldfd;
1181
1182 rcu_read_lock();
1183 if (!files_lookup_fd_rcu(files, oldfd))
1184 retval = -EBADF;
1185 rcu_read_unlock();
1186 return retval;
1187 }
1188 return ksys_dup3(oldfd, newfd, 0);
1189}
1190
1191SYSCALL_DEFINE1(dup, unsigned int, fildes)
1192{
1193 int ret = -EBADF;
1194 struct file *file = fget_raw(fildes);
1195
1196 if (file) {
1197 ret = get_unused_fd_flags(0);
1198 if (ret >= 0)
1199 fd_install(ret, file);
1200 else
1201 fput(file);
1202 }
1203 return ret;
1204}
1205
1206int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1207{
1208 unsigned long nofile = rlimit(RLIMIT_NOFILE);
1209 int err;
1210 if (from >= nofile)
1211 return -EINVAL;
1212 err = alloc_fd(from, nofile, flags);
1213 if (err >= 0) {
1214 get_file(file);
1215 fd_install(err, file);
1216 }
1217 return err;
1218}
1219
1220int iterate_fd(struct files_struct *files, unsigned n,
1221 int (*f)(const void *, struct file *, unsigned),
1222 const void *p)
1223{
1224 struct fdtable *fdt;
1225 int res = 0;
1226 if (!files)
1227 return 0;
1228 spin_lock(&files->file_lock);
1229 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1230 struct file *file;
1231 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1232 if (!file)
1233 continue;
1234 res = f(p, file, n);
1235 if (res)
1236 break;
1237 }
1238 spin_unlock(&files->file_lock);
1239 return res;
1240}
1241EXPORT_SYMBOL(iterate_fd);