4 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
6 * Manage the dynamic fd arrays in the process files_struct.
9 #include <linux/syscalls.h>
10 #include <linux/export.h>
13 #include <linux/mmzone.h>
14 #include <linux/time.h>
15 #include <linux/sched/signal.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>
26 unsigned int sysctl_nr_open __read_mostly = 1024*1024;
27 unsigned 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))
30 unsigned int sysctl_nr_open_max =
31 __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
33 static void __free_fdtable(struct fdtable *fdt)
36 kvfree(fdt->open_fds);
40 static void free_fdtable_rcu(struct rcu_head *rcu)
42 __free_fdtable(container_of(rcu, struct fdtable, rcu));
45 #define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
46 #define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
49 * Copy 'count' fd bits from the old table to the new table and clear the extra
50 * space if any. This does not copy the file pointers. Called with the files
51 * spinlock held for write.
53 static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
56 unsigned int cpy, set;
58 cpy = count / BITS_PER_BYTE;
59 set = (nfdt->max_fds - count) / BITS_PER_BYTE;
60 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
61 memset((char *)nfdt->open_fds + cpy, 0, set);
62 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
63 memset((char *)nfdt->close_on_exec + cpy, 0, set);
65 cpy = BITBIT_SIZE(count);
66 set = BITBIT_SIZE(nfdt->max_fds) - cpy;
67 memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
68 memset((char *)nfdt->full_fds_bits + cpy, 0, set);
72 * Copy all file descriptors from the old table to the new, expanded table and
73 * clear the extra space. Called with the files spinlock held for write.
75 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
77 unsigned int cpy, set;
79 BUG_ON(nfdt->max_fds < ofdt->max_fds);
81 cpy = ofdt->max_fds * sizeof(struct file *);
82 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
83 memcpy(nfdt->fd, ofdt->fd, cpy);
84 memset((char *)nfdt->fd + cpy, 0, set);
86 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
89 static struct fdtable * alloc_fdtable(unsigned int nr)
95 * Figure out how many fds we actually want to support in this fdtable.
96 * Allocation steps are keyed to the size of the fdarray, since it
97 * grows far faster than any of the other dynamic data. We try to fit
98 * the fdarray into comfortable page-tuned chunks: starting at 1024B
99 * and growing in powers of two from there on.
101 nr /= (1024 / sizeof(struct file *));
102 nr = roundup_pow_of_two(nr + 1);
103 nr *= (1024 / sizeof(struct file *));
105 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
106 * had been set lower between the check in expand_files() and here. Deal
107 * with that in caller, it's cheaper that way.
109 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
110 * bitmaps handling below becomes unpleasant, to put it mildly...
112 if (unlikely(nr > sysctl_nr_open))
113 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
115 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
119 data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
124 data = kvmalloc(max_t(size_t,
125 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
129 fdt->open_fds = data;
130 data += nr / BITS_PER_BYTE;
131 fdt->close_on_exec = data;
132 data += nr / BITS_PER_BYTE;
133 fdt->full_fds_bits = data;
146 * Expand the file descriptor table.
147 * This function will allocate a new fdtable and both fd array and fdset, of
149 * Return <0 error code on error; 1 on successful completion.
150 * The files->file_lock should be held on entry, and will be held on exit.
152 static int expand_fdtable(struct files_struct *files, unsigned int nr)
153 __releases(files->file_lock)
154 __acquires(files->file_lock)
156 struct fdtable *new_fdt, *cur_fdt;
158 spin_unlock(&files->file_lock);
159 new_fdt = alloc_fdtable(nr);
161 /* make sure all __fd_install() have seen resize_in_progress
162 * or have finished their rcu_read_lock_sched() section.
164 if (atomic_read(&files->count) > 1)
167 spin_lock(&files->file_lock);
171 * extremely unlikely race - sysctl_nr_open decreased between the check in
172 * caller and alloc_fdtable(). Cheaper to catch it here...
174 if (unlikely(new_fdt->max_fds <= nr)) {
175 __free_fdtable(new_fdt);
178 cur_fdt = files_fdtable(files);
179 BUG_ON(nr < cur_fdt->max_fds);
180 copy_fdtable(new_fdt, cur_fdt);
181 rcu_assign_pointer(files->fdt, new_fdt);
182 if (cur_fdt != &files->fdtab)
183 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
184 /* coupled with smp_rmb() in __fd_install() */
191 * This function will expand the file structures, if the requested size exceeds
192 * the current capacity and there is room for expansion.
193 * Return <0 error code on error; 0 when nothing done; 1 when files were
194 * expanded and execution may have blocked.
195 * The files->file_lock should be held on entry, and will be held on exit.
197 static int expand_files(struct files_struct *files, unsigned int nr)
198 __releases(files->file_lock)
199 __acquires(files->file_lock)
205 fdt = files_fdtable(files);
207 /* Do we need to expand? */
208 if (nr < fdt->max_fds)
212 if (nr >= sysctl_nr_open)
215 if (unlikely(files->resize_in_progress)) {
216 spin_unlock(&files->file_lock);
218 wait_event(files->resize_wait, !files->resize_in_progress);
219 spin_lock(&files->file_lock);
223 /* All good, so we try */
224 files->resize_in_progress = true;
225 expanded = expand_fdtable(files, nr);
226 files->resize_in_progress = false;
228 wake_up_all(&files->resize_wait);
232 static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
234 __set_bit(fd, fdt->close_on_exec);
237 static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
239 if (test_bit(fd, fdt->close_on_exec))
240 __clear_bit(fd, fdt->close_on_exec);
243 static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
245 __set_bit(fd, fdt->open_fds);
247 if (!~fdt->open_fds[fd])
248 __set_bit(fd, fdt->full_fds_bits);
251 static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
253 __clear_bit(fd, fdt->open_fds);
254 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
257 static unsigned int count_open_files(struct fdtable *fdt)
259 unsigned int size = fdt->max_fds;
262 /* Find the last open fd */
263 for (i = size / BITS_PER_LONG; i > 0; ) {
264 if (fdt->open_fds[--i])
267 i = (i + 1) * BITS_PER_LONG;
272 * Allocate a new files structure and copy contents from the
273 * passed in files structure.
274 * errorp will be valid only when the returned files_struct is NULL.
276 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
278 struct files_struct *newf;
279 struct file **old_fds, **new_fds;
280 unsigned int open_files, i;
281 struct fdtable *old_fdt, *new_fdt;
284 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
288 atomic_set(&newf->count, 1);
290 spin_lock_init(&newf->file_lock);
291 newf->resize_in_progress = false;
292 init_waitqueue_head(&newf->resize_wait);
294 new_fdt = &newf->fdtab;
295 new_fdt->max_fds = NR_OPEN_DEFAULT;
296 new_fdt->close_on_exec = newf->close_on_exec_init;
297 new_fdt->open_fds = newf->open_fds_init;
298 new_fdt->full_fds_bits = newf->full_fds_bits_init;
299 new_fdt->fd = &newf->fd_array[0];
301 spin_lock(&oldf->file_lock);
302 old_fdt = files_fdtable(oldf);
303 open_files = count_open_files(old_fdt);
306 * Check whether we need to allocate a larger fd array and fd set.
308 while (unlikely(open_files > new_fdt->max_fds)) {
309 spin_unlock(&oldf->file_lock);
311 if (new_fdt != &newf->fdtab)
312 __free_fdtable(new_fdt);
314 new_fdt = alloc_fdtable(open_files - 1);
320 /* beyond sysctl_nr_open; nothing to do */
321 if (unlikely(new_fdt->max_fds < open_files)) {
322 __free_fdtable(new_fdt);
328 * Reacquire the oldf lock and a pointer to its fd table
329 * who knows it may have a new bigger fd table. We need
330 * the latest pointer.
332 spin_lock(&oldf->file_lock);
333 old_fdt = files_fdtable(oldf);
334 open_files = count_open_files(old_fdt);
337 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
339 old_fds = old_fdt->fd;
340 new_fds = new_fdt->fd;
342 for (i = open_files; i != 0; i--) {
343 struct file *f = *old_fds++;
348 * The fd may be claimed in the fd bitmap but not yet
349 * instantiated in the files array if a sibling thread
350 * is partway through open(). So make sure that this
351 * fd is available to the new process.
353 __clear_open_fd(open_files - i, new_fdt);
355 rcu_assign_pointer(*new_fds++, f);
357 spin_unlock(&oldf->file_lock);
359 /* clear the remainder */
360 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
362 rcu_assign_pointer(newf->fdt, new_fdt);
367 kmem_cache_free(files_cachep, newf);
372 static struct fdtable *close_files(struct files_struct * files)
375 * It is safe to dereference the fd table without RCU or
376 * ->file_lock because this is the last reference to the
379 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
380 unsigned int i, j = 0;
384 i = j * BITS_PER_LONG;
385 if (i >= fdt->max_fds)
387 set = fdt->open_fds[j++];
390 struct file * file = xchg(&fdt->fd[i], NULL);
392 filp_close(file, files);
393 cond_resched_rcu_qs();
404 struct files_struct *get_files_struct(struct task_struct *task)
406 struct files_struct *files;
411 atomic_inc(&files->count);
417 void put_files_struct(struct files_struct *files)
419 if (atomic_dec_and_test(&files->count)) {
420 struct fdtable *fdt = close_files(files);
422 /* free the arrays if they are not embedded */
423 if (fdt != &files->fdtab)
425 kmem_cache_free(files_cachep, files);
429 void reset_files_struct(struct files_struct *files)
431 struct task_struct *tsk = current;
432 struct files_struct *old;
438 put_files_struct(old);
441 void exit_files(struct task_struct *tsk)
443 struct files_struct * files = tsk->files;
449 put_files_struct(files);
453 struct files_struct init_files = {
454 .count = ATOMIC_INIT(1),
455 .fdt = &init_files.fdtab,
457 .max_fds = NR_OPEN_DEFAULT,
458 .fd = &init_files.fd_array[0],
459 .close_on_exec = init_files.close_on_exec_init,
460 .open_fds = init_files.open_fds_init,
461 .full_fds_bits = init_files.full_fds_bits_init,
463 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
466 static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
468 unsigned int maxfd = fdt->max_fds;
469 unsigned int maxbit = maxfd / BITS_PER_LONG;
470 unsigned int bitbit = start / BITS_PER_LONG;
472 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
477 return find_next_zero_bit(fdt->open_fds, maxfd, start);
481 * allocate a file descriptor, mark it busy.
483 int __alloc_fd(struct files_struct *files,
484 unsigned start, unsigned end, unsigned flags)
490 spin_lock(&files->file_lock);
492 fdt = files_fdtable(files);
494 if (fd < files->next_fd)
497 if (fd < fdt->max_fds)
498 fd = find_next_fd(fdt, fd);
501 * N.B. For clone tasks sharing a files structure, this test
502 * will limit the total number of files that can be opened.
508 error = expand_files(files, fd);
513 * If we needed to expand the fs array we
514 * might have blocked - try again.
519 if (start <= files->next_fd)
520 files->next_fd = fd + 1;
522 __set_open_fd(fd, fdt);
523 if (flags & O_CLOEXEC)
524 __set_close_on_exec(fd, fdt);
526 __clear_close_on_exec(fd, fdt);
530 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
531 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
532 rcu_assign_pointer(fdt->fd[fd], NULL);
537 spin_unlock(&files->file_lock);
541 static int alloc_fd(unsigned start, unsigned flags)
543 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
546 int get_unused_fd_flags(unsigned flags)
548 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
550 EXPORT_SYMBOL(get_unused_fd_flags);
552 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
554 struct fdtable *fdt = files_fdtable(files);
555 __clear_open_fd(fd, fdt);
556 if (fd < files->next_fd)
560 void put_unused_fd(unsigned int fd)
562 struct files_struct *files = current->files;
563 spin_lock(&files->file_lock);
564 __put_unused_fd(files, fd);
565 spin_unlock(&files->file_lock);
568 EXPORT_SYMBOL(put_unused_fd);
571 * Install a file pointer in the fd array.
573 * The VFS is full of places where we drop the files lock between
574 * setting the open_fds bitmap and installing the file in the file
575 * array. At any such point, we are vulnerable to a dup2() race
576 * installing a file in the array before us. We need to detect this and
577 * fput() the struct file we are about to overwrite in this case.
579 * It should never happen - if we allow dup2() do it, _really_ bad things
582 * NOTE: __fd_install() variant is really, really low-level; don't
583 * use it unless you are forced to by truly lousy API shoved down
584 * your throat. 'files' *MUST* be either current->files or obtained
585 * by get_files_struct(current) done by whoever had given it to you,
586 * or really bad things will happen. Normally you want to use
587 * fd_install() instead.
590 void __fd_install(struct files_struct *files, unsigned int fd,
596 rcu_read_lock_sched();
598 while (unlikely(files->resize_in_progress)) {
599 rcu_read_unlock_sched();
600 wait_event(files->resize_wait, !files->resize_in_progress);
601 rcu_read_lock_sched();
603 /* coupled with smp_wmb() in expand_fdtable() */
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();
611 void fd_install(unsigned int fd, struct file *file)
613 __fd_install(current->files, fd, file);
616 EXPORT_SYMBOL(fd_install);
619 * The same warnings as for __alloc_fd()/__fd_install() apply here...
621 int __close_fd(struct files_struct *files, unsigned fd)
626 spin_lock(&files->file_lock);
627 fdt = files_fdtable(files);
628 if (fd >= fdt->max_fds)
633 rcu_assign_pointer(fdt->fd[fd], NULL);
634 __clear_close_on_exec(fd, fdt);
635 __put_unused_fd(files, fd);
636 spin_unlock(&files->file_lock);
637 return filp_close(file, files);
640 spin_unlock(&files->file_lock);
644 void do_close_on_exec(struct files_struct *files)
649 /* exec unshares first */
650 spin_lock(&files->file_lock);
653 unsigned fd = i * BITS_PER_LONG;
654 fdt = files_fdtable(files);
655 if (fd >= fdt->max_fds)
657 set = fdt->close_on_exec[i];
660 fdt->close_on_exec[i] = 0;
661 for ( ; set ; fd++, set >>= 1) {
668 rcu_assign_pointer(fdt->fd[fd], NULL);
669 __put_unused_fd(files, fd);
670 spin_unlock(&files->file_lock);
671 filp_close(file, files);
673 spin_lock(&files->file_lock);
677 spin_unlock(&files->file_lock);
680 static struct file *__fget(unsigned int fd, fmode_t mask)
682 struct files_struct *files = current->files;
687 file = fcheck_files(files, fd);
689 /* File object ref couldn't be taken.
690 * dup2() atomicity guarantee is the reason
691 * we loop to catch the new file (or NULL pointer)
693 if (file->f_mode & mask)
695 else if (!get_file_rcu(file))
703 struct file *fget(unsigned int fd)
705 return __fget(fd, FMODE_PATH);
709 struct file *fget_raw(unsigned int fd)
711 return __fget(fd, 0);
713 EXPORT_SYMBOL(fget_raw);
716 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
718 * You can use this instead of fget if you satisfy all of the following
720 * 1) You must call fput_light before exiting the syscall and returning control
721 * to userspace (i.e. you cannot remember the returned struct file * after
722 * returning to userspace).
723 * 2) You must not call filp_close on the returned struct file * in between
724 * calls to fget_light and fput_light.
725 * 3) You must not clone the current task in between the calls to fget_light
728 * The fput_needed flag returned by fget_light should be passed to the
729 * corresponding fput_light.
731 static unsigned long __fget_light(unsigned int fd, fmode_t mask)
733 struct files_struct *files = current->files;
736 if (atomic_read(&files->count) == 1) {
737 file = __fcheck_files(files, fd);
738 if (!file || unlikely(file->f_mode & mask))
740 return (unsigned long)file;
742 file = __fget(fd, mask);
745 return FDPUT_FPUT | (unsigned long)file;
748 unsigned long __fdget(unsigned int fd)
750 return __fget_light(fd, FMODE_PATH);
752 EXPORT_SYMBOL(__fdget);
754 unsigned long __fdget_raw(unsigned int fd)
756 return __fget_light(fd, 0);
759 unsigned long __fdget_pos(unsigned int fd)
761 unsigned long v = __fdget(fd);
762 struct file *file = (struct file *)(v & ~3);
764 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
765 if (file_count(file) > 1) {
766 v |= FDPUT_POS_UNLOCK;
767 mutex_lock(&file->f_pos_lock);
773 void __f_unlock_pos(struct file *f)
775 mutex_unlock(&f->f_pos_lock);
779 * We only lock f_pos if we have threads or if the file might be
780 * shared with another process. In both cases we'll have an elevated
781 * file count (done either by fdget() or by fork()).
784 void set_close_on_exec(unsigned int fd, int flag)
786 struct files_struct *files = current->files;
788 spin_lock(&files->file_lock);
789 fdt = files_fdtable(files);
791 __set_close_on_exec(fd, fdt);
793 __clear_close_on_exec(fd, fdt);
794 spin_unlock(&files->file_lock);
797 bool get_close_on_exec(unsigned int fd)
799 struct files_struct *files = current->files;
803 fdt = files_fdtable(files);
804 res = close_on_exec(fd, fdt);
809 static int do_dup2(struct files_struct *files,
810 struct file *file, unsigned fd, unsigned flags)
811 __releases(&files->file_lock)
817 * We need to detect attempts to do dup2() over allocated but still
818 * not finished descriptor. NB: OpenBSD avoids that at the price of
819 * extra work in their equivalent of fget() - they insert struct
820 * file immediately after grabbing descriptor, mark it larval if
821 * more work (e.g. actual opening) is needed and make sure that
822 * fget() treats larval files as absent. Potentially interesting,
823 * but while extra work in fget() is trivial, locking implications
824 * and amount of surgery on open()-related paths in VFS are not.
825 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
826 * deadlocks in rather amusing ways, AFAICS. All of that is out of
827 * scope of POSIX or SUS, since neither considers shared descriptor
828 * tables and this condition does not arise without those.
830 fdt = files_fdtable(files);
831 tofree = fdt->fd[fd];
832 if (!tofree && fd_is_open(fd, fdt))
835 rcu_assign_pointer(fdt->fd[fd], file);
836 __set_open_fd(fd, fdt);
837 if (flags & O_CLOEXEC)
838 __set_close_on_exec(fd, fdt);
840 __clear_close_on_exec(fd, fdt);
841 spin_unlock(&files->file_lock);
844 filp_close(tofree, files);
849 spin_unlock(&files->file_lock);
853 int replace_fd(unsigned fd, struct file *file, unsigned flags)
856 struct files_struct *files = current->files;
859 return __close_fd(files, fd);
861 if (fd >= rlimit(RLIMIT_NOFILE))
864 spin_lock(&files->file_lock);
865 err = expand_files(files, fd);
866 if (unlikely(err < 0))
868 return do_dup2(files, file, fd, flags);
871 spin_unlock(&files->file_lock);
875 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
879 struct files_struct *files = current->files;
881 if ((flags & ~O_CLOEXEC) != 0)
884 if (unlikely(oldfd == newfd))
887 if (newfd >= rlimit(RLIMIT_NOFILE))
890 spin_lock(&files->file_lock);
891 err = expand_files(files, newfd);
892 file = fcheck(oldfd);
895 if (unlikely(err < 0)) {
900 return do_dup2(files, file, newfd, flags);
905 spin_unlock(&files->file_lock);
909 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
911 if (unlikely(newfd == oldfd)) { /* corner case */
912 struct files_struct *files = current->files;
916 if (!fcheck_files(files, oldfd))
921 return sys_dup3(oldfd, newfd, 0);
924 SYSCALL_DEFINE1(dup, unsigned int, fildes)
927 struct file *file = fget_raw(fildes);
930 ret = get_unused_fd_flags(0);
932 fd_install(ret, file);
939 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
942 if (from >= rlimit(RLIMIT_NOFILE))
944 err = alloc_fd(from, flags);
947 fd_install(err, file);
952 int iterate_fd(struct files_struct *files, unsigned n,
953 int (*f)(const void *, struct file *, unsigned),
960 spin_lock(&files->file_lock);
961 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
963 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
970 spin_unlock(&files->file_lock);
973 EXPORT_SYMBOL(iterate_fd);