1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
7 * Manage the dynamic fd arrays in the process files_struct.
10 #include <linux/syscalls.h>
11 #include <linux/export.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>
22 unsigned int sysctl_nr_open __read_mostly = 1024*1024;
23 unsigned 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))
26 unsigned int sysctl_nr_open_max =
27 __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
29 static void __free_fdtable(struct fdtable *fdt)
32 kvfree(fdt->open_fds);
36 static void free_fdtable_rcu(struct rcu_head *rcu)
38 __free_fdtable(container_of(rcu, struct fdtable, rcu));
41 #define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
42 #define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
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.
49 static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
52 unsigned int cpy, set;
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);
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);
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.
71 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
73 unsigned int cpy, set;
75 BUG_ON(nfdt->max_fds < ofdt->max_fds);
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);
82 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
85 static struct fdtable * alloc_fdtable(unsigned int nr)
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.
97 nr /= (1024 / sizeof(struct file *));
98 nr = roundup_pow_of_two(nr + 1);
99 nr *= (1024 / sizeof(struct file *));
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.
105 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
106 * bitmaps handling below becomes unpleasant, to put it mildly...
108 if (unlikely(nr > sysctl_nr_open))
109 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
111 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
115 data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
120 data = kvmalloc(max_t(size_t,
121 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
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;
142 * Expand the file descriptor table.
143 * This function will allocate a new fdtable and both fd array and fdset, of
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.
148 static int expand_fdtable(struct files_struct *files, unsigned int nr)
149 __releases(files->file_lock)
150 __acquires(files->file_lock)
152 struct fdtable *new_fdt, *cur_fdt;
154 spin_unlock(&files->file_lock);
155 new_fdt = alloc_fdtable(nr);
157 /* make sure all __fd_install() have seen resize_in_progress
158 * or have finished their rcu_read_lock_sched() section.
160 if (atomic_read(&files->count) > 1)
163 spin_lock(&files->file_lock);
167 * extremely unlikely race - sysctl_nr_open decreased between the check in
168 * caller and alloc_fdtable(). Cheaper to catch it here...
170 if (unlikely(new_fdt->max_fds <= nr)) {
171 __free_fdtable(new_fdt);
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() */
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.
193 static int expand_files(struct files_struct *files, unsigned int nr)
194 __releases(files->file_lock)
195 __acquires(files->file_lock)
201 fdt = files_fdtable(files);
203 /* Do we need to expand? */
204 if (nr < fdt->max_fds)
208 if (nr >= sysctl_nr_open)
211 if (unlikely(files->resize_in_progress)) {
212 spin_unlock(&files->file_lock);
214 wait_event(files->resize_wait, !files->resize_in_progress);
215 spin_lock(&files->file_lock);
219 /* All good, so we try */
220 files->resize_in_progress = true;
221 expanded = expand_fdtable(files, nr);
222 files->resize_in_progress = false;
224 wake_up_all(&files->resize_wait);
228 static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
230 __set_bit(fd, fdt->close_on_exec);
233 static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
235 if (test_bit(fd, fdt->close_on_exec))
236 __clear_bit(fd, fdt->close_on_exec);
239 static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
241 __set_bit(fd, fdt->open_fds);
243 if (!~fdt->open_fds[fd])
244 __set_bit(fd, fdt->full_fds_bits);
247 static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
249 __clear_bit(fd, fdt->open_fds);
250 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
253 static unsigned int count_open_files(struct fdtable *fdt)
255 unsigned int size = fdt->max_fds;
258 /* Find the last open fd */
259 for (i = size / BITS_PER_LONG; i > 0; ) {
260 if (fdt->open_fds[--i])
263 i = (i + 1) * BITS_PER_LONG;
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.
272 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
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;
280 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
284 atomic_set(&newf->count, 1);
286 spin_lock_init(&newf->file_lock);
287 newf->resize_in_progress = false;
288 init_waitqueue_head(&newf->resize_wait);
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];
297 spin_lock(&oldf->file_lock);
298 old_fdt = files_fdtable(oldf);
299 open_files = count_open_files(old_fdt);
302 * Check whether we need to allocate a larger fd array and fd set.
304 while (unlikely(open_files > new_fdt->max_fds)) {
305 spin_unlock(&oldf->file_lock);
307 if (new_fdt != &newf->fdtab)
308 __free_fdtable(new_fdt);
310 new_fdt = alloc_fdtable(open_files - 1);
316 /* beyond sysctl_nr_open; nothing to do */
317 if (unlikely(new_fdt->max_fds < open_files)) {
318 __free_fdtable(new_fdt);
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.
328 spin_lock(&oldf->file_lock);
329 old_fdt = files_fdtable(oldf);
330 open_files = count_open_files(old_fdt);
333 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
335 old_fds = old_fdt->fd;
336 new_fds = new_fdt->fd;
338 for (i = open_files; i != 0; i--) {
339 struct file *f = *old_fds++;
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.
349 __clear_open_fd(open_files - i, new_fdt);
351 rcu_assign_pointer(*new_fds++, f);
353 spin_unlock(&oldf->file_lock);
355 /* clear the remainder */
356 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
358 rcu_assign_pointer(newf->fdt, new_fdt);
363 kmem_cache_free(files_cachep, newf);
368 static struct fdtable *close_files(struct files_struct * files)
371 * It is safe to dereference the fd table without RCU or
372 * ->file_lock because this is the last reference to the
375 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
376 unsigned int i, j = 0;
380 i = j * BITS_PER_LONG;
381 if (i >= fdt->max_fds)
383 set = fdt->open_fds[j++];
386 struct file * file = xchg(&fdt->fd[i], NULL);
388 filp_close(file, files);
400 struct files_struct *get_files_struct(struct task_struct *task)
402 struct files_struct *files;
407 atomic_inc(&files->count);
413 void put_files_struct(struct files_struct *files)
415 if (atomic_dec_and_test(&files->count)) {
416 struct fdtable *fdt = close_files(files);
418 /* free the arrays if they are not embedded */
419 if (fdt != &files->fdtab)
421 kmem_cache_free(files_cachep, files);
425 void reset_files_struct(struct files_struct *files)
427 struct task_struct *tsk = current;
428 struct files_struct *old;
434 put_files_struct(old);
437 void exit_files(struct task_struct *tsk)
439 struct files_struct * files = tsk->files;
445 put_files_struct(files);
449 struct files_struct init_files = {
450 .count = ATOMIC_INIT(1),
451 .fdt = &init_files.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,
459 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
462 static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
464 unsigned int maxfd = fdt->max_fds;
465 unsigned int maxbit = maxfd / BITS_PER_LONG;
466 unsigned int bitbit = start / BITS_PER_LONG;
468 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
473 return find_next_zero_bit(fdt->open_fds, maxfd, start);
477 * allocate a file descriptor, mark it busy.
479 int __alloc_fd(struct files_struct *files,
480 unsigned start, unsigned end, unsigned flags)
486 spin_lock(&files->file_lock);
488 fdt = files_fdtable(files);
490 if (fd < files->next_fd)
493 if (fd < fdt->max_fds)
494 fd = find_next_fd(fdt, fd);
497 * N.B. For clone tasks sharing a files structure, this test
498 * will limit the total number of files that can be opened.
504 error = expand_files(files, fd);
509 * If we needed to expand the fs array we
510 * might have blocked - try again.
515 if (start <= files->next_fd)
516 files->next_fd = fd + 1;
518 __set_open_fd(fd, fdt);
519 if (flags & O_CLOEXEC)
520 __set_close_on_exec(fd, fdt);
522 __clear_close_on_exec(fd, fdt);
526 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
527 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
528 rcu_assign_pointer(fdt->fd[fd], NULL);
533 spin_unlock(&files->file_lock);
537 static int alloc_fd(unsigned start, unsigned flags)
539 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
542 int get_unused_fd_flags(unsigned flags)
544 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
546 EXPORT_SYMBOL(get_unused_fd_flags);
548 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
550 struct fdtable *fdt = files_fdtable(files);
551 __clear_open_fd(fd, fdt);
552 if (fd < files->next_fd)
556 void put_unused_fd(unsigned int fd)
558 struct files_struct *files = current->files;
559 spin_lock(&files->file_lock);
560 __put_unused_fd(files, fd);
561 spin_unlock(&files->file_lock);
564 EXPORT_SYMBOL(put_unused_fd);
567 * Install a file pointer in the fd array.
569 * The VFS is full of places where we drop the files lock between
570 * setting the open_fds bitmap and installing the file in the file
571 * array. At any such point, we are vulnerable to a dup2() race
572 * installing a file in the array before us. We need to detect this and
573 * fput() the struct file we are about to overwrite in this case.
575 * It should never happen - if we allow dup2() do it, _really_ bad things
578 * NOTE: __fd_install() variant is really, really low-level; don't
579 * use it unless you are forced to by truly lousy API shoved down
580 * your throat. 'files' *MUST* be either current->files or obtained
581 * by get_files_struct(current) done by whoever had given it to you,
582 * or really bad things will happen. Normally you want to use
583 * fd_install() instead.
586 void __fd_install(struct files_struct *files, unsigned int fd,
591 rcu_read_lock_sched();
593 if (unlikely(files->resize_in_progress)) {
594 rcu_read_unlock_sched();
595 spin_lock(&files->file_lock);
596 fdt = files_fdtable(files);
597 BUG_ON(fdt->fd[fd] != NULL);
598 rcu_assign_pointer(fdt->fd[fd], file);
599 spin_unlock(&files->file_lock);
602 /* coupled with smp_wmb() in expand_fdtable() */
604 fdt = rcu_dereference_sched(files->fdt);
605 BUG_ON(fdt->fd[fd] != NULL);
606 rcu_assign_pointer(fdt->fd[fd], file);
607 rcu_read_unlock_sched();
610 void fd_install(unsigned int fd, struct file *file)
612 __fd_install(current->files, fd, file);
615 EXPORT_SYMBOL(fd_install);
618 * The same warnings as for __alloc_fd()/__fd_install() apply here...
620 int __close_fd(struct files_struct *files, unsigned fd)
625 spin_lock(&files->file_lock);
626 fdt = files_fdtable(files);
627 if (fd >= fdt->max_fds)
632 rcu_assign_pointer(fdt->fd[fd], NULL);
633 __put_unused_fd(files, fd);
634 spin_unlock(&files->file_lock);
635 return filp_close(file, files);
638 spin_unlock(&files->file_lock);
642 void do_close_on_exec(struct files_struct *files)
647 /* exec unshares first */
648 spin_lock(&files->file_lock);
651 unsigned fd = i * BITS_PER_LONG;
652 fdt = files_fdtable(files);
653 if (fd >= fdt->max_fds)
655 set = fdt->close_on_exec[i];
658 fdt->close_on_exec[i] = 0;
659 for ( ; set ; fd++, set >>= 1) {
666 rcu_assign_pointer(fdt->fd[fd], NULL);
667 __put_unused_fd(files, fd);
668 spin_unlock(&files->file_lock);
669 filp_close(file, files);
671 spin_lock(&files->file_lock);
675 spin_unlock(&files->file_lock);
678 static struct file *__fget(unsigned int fd, fmode_t mask)
680 struct files_struct *files = current->files;
685 file = fcheck_files(files, fd);
687 /* File object ref couldn't be taken.
688 * dup2() atomicity guarantee is the reason
689 * we loop to catch the new file (or NULL pointer)
691 if (file->f_mode & mask)
693 else if (!get_file_rcu(file))
701 struct file *fget(unsigned int fd)
703 return __fget(fd, FMODE_PATH);
707 struct file *fget_raw(unsigned int fd)
709 return __fget(fd, 0);
711 EXPORT_SYMBOL(fget_raw);
714 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
716 * You can use this instead of fget if you satisfy all of the following
718 * 1) You must call fput_light before exiting the syscall and returning control
719 * to userspace (i.e. you cannot remember the returned struct file * after
720 * returning to userspace).
721 * 2) You must not call filp_close on the returned struct file * in between
722 * calls to fget_light and fput_light.
723 * 3) You must not clone the current task in between the calls to fget_light
726 * The fput_needed flag returned by fget_light should be passed to the
727 * corresponding fput_light.
729 static unsigned long __fget_light(unsigned int fd, fmode_t mask)
731 struct files_struct *files = current->files;
734 if (atomic_read(&files->count) == 1) {
735 file = __fcheck_files(files, fd);
736 if (!file || unlikely(file->f_mode & mask))
738 return (unsigned long)file;
740 file = __fget(fd, mask);
743 return FDPUT_FPUT | (unsigned long)file;
746 unsigned long __fdget(unsigned int fd)
748 return __fget_light(fd, FMODE_PATH);
750 EXPORT_SYMBOL(__fdget);
752 unsigned long __fdget_raw(unsigned int fd)
754 return __fget_light(fd, 0);
757 unsigned long __fdget_pos(unsigned int fd)
759 unsigned long v = __fdget(fd);
760 struct file *file = (struct file *)(v & ~3);
762 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
763 if (file_count(file) > 1) {
764 v |= FDPUT_POS_UNLOCK;
765 mutex_lock(&file->f_pos_lock);
771 void __f_unlock_pos(struct file *f)
773 mutex_unlock(&f->f_pos_lock);
777 * We only lock f_pos if we have threads or if the file might be
778 * shared with another process. In both cases we'll have an elevated
779 * file count (done either by fdget() or by fork()).
782 void set_close_on_exec(unsigned int fd, int flag)
784 struct files_struct *files = current->files;
786 spin_lock(&files->file_lock);
787 fdt = files_fdtable(files);
789 __set_close_on_exec(fd, fdt);
791 __clear_close_on_exec(fd, fdt);
792 spin_unlock(&files->file_lock);
795 bool get_close_on_exec(unsigned int fd)
797 struct files_struct *files = current->files;
801 fdt = files_fdtable(files);
802 res = close_on_exec(fd, fdt);
807 static int do_dup2(struct files_struct *files,
808 struct file *file, unsigned fd, unsigned flags)
809 __releases(&files->file_lock)
815 * We need to detect attempts to do dup2() over allocated but still
816 * not finished descriptor. NB: OpenBSD avoids that at the price of
817 * extra work in their equivalent of fget() - they insert struct
818 * file immediately after grabbing descriptor, mark it larval if
819 * more work (e.g. actual opening) is needed and make sure that
820 * fget() treats larval files as absent. Potentially interesting,
821 * but while extra work in fget() is trivial, locking implications
822 * and amount of surgery on open()-related paths in VFS are not.
823 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
824 * deadlocks in rather amusing ways, AFAICS. All of that is out of
825 * scope of POSIX or SUS, since neither considers shared descriptor
826 * tables and this condition does not arise without those.
828 fdt = files_fdtable(files);
829 tofree = fdt->fd[fd];
830 if (!tofree && fd_is_open(fd, fdt))
833 rcu_assign_pointer(fdt->fd[fd], file);
834 __set_open_fd(fd, fdt);
835 if (flags & O_CLOEXEC)
836 __set_close_on_exec(fd, fdt);
838 __clear_close_on_exec(fd, fdt);
839 spin_unlock(&files->file_lock);
842 filp_close(tofree, files);
847 spin_unlock(&files->file_lock);
851 int replace_fd(unsigned fd, struct file *file, unsigned flags)
854 struct files_struct *files = current->files;
857 return __close_fd(files, fd);
859 if (fd >= rlimit(RLIMIT_NOFILE))
862 spin_lock(&files->file_lock);
863 err = expand_files(files, fd);
864 if (unlikely(err < 0))
866 return do_dup2(files, file, fd, flags);
869 spin_unlock(&files->file_lock);
873 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
877 struct files_struct *files = current->files;
879 if ((flags & ~O_CLOEXEC) != 0)
882 if (unlikely(oldfd == newfd))
885 if (newfd >= rlimit(RLIMIT_NOFILE))
888 spin_lock(&files->file_lock);
889 err = expand_files(files, newfd);
890 file = fcheck(oldfd);
893 if (unlikely(err < 0)) {
898 return do_dup2(files, file, newfd, flags);
903 spin_unlock(&files->file_lock);
907 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
909 if (unlikely(newfd == oldfd)) { /* corner case */
910 struct files_struct *files = current->files;
914 if (!fcheck_files(files, oldfd))
919 return sys_dup3(oldfd, newfd, 0);
922 SYSCALL_DEFINE1(dup, unsigned int, fildes)
925 struct file *file = fget_raw(fildes);
928 ret = get_unused_fd_flags(0);
930 fd_install(ret, file);
937 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
940 if (from >= rlimit(RLIMIT_NOFILE))
942 err = alloc_fd(from, flags);
945 fd_install(err, file);
950 int iterate_fd(struct files_struct *files, unsigned n,
951 int (*f)(const void *, struct file *, unsigned),
958 spin_lock(&files->file_lock);
959 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
961 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
968 spin_unlock(&files->file_lock);
971 EXPORT_SYMBOL(iterate_fd);