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