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.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 int sysctl_nr_open __read_mostly = 1024*1024;
27 int sysctl_nr_open_min = BITS_PER_LONG;
28 /* our max() is unusable in constant expressions ;-/ */
29 #define __const_max(x, y) ((x) < (y) ? (x) : (y))
30 int sysctl_nr_open_max = __const_max(INT_MAX, ~(size_t)0/sizeof(void *)) &
33 static void *alloc_fdmem(size_t size)
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.
39 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
40 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY);
47 static void __free_fdtable(struct fdtable *fdt)
50 kvfree(fdt->open_fds);
54 static void free_fdtable_rcu(struct rcu_head *rcu)
56 __free_fdtable(container_of(rcu, struct fdtable, rcu));
60 * Expand the fdset in the files_struct. Called with the files spinlock
63 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
65 unsigned int cpy, set;
67 BUG_ON(nfdt->max_fds < ofdt->max_fds);
69 cpy = ofdt->max_fds * sizeof(struct file *);
70 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
71 memcpy(nfdt->fd, ofdt->fd, cpy);
72 memset((char *)(nfdt->fd) + cpy, 0, set);
74 cpy = ofdt->max_fds / BITS_PER_BYTE;
75 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
76 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
77 memset((char *)(nfdt->open_fds) + cpy, 0, set);
78 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
79 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
82 static struct fdtable * alloc_fdtable(unsigned int nr)
88 * Figure out how many fds we actually want to support in this fdtable.
89 * Allocation steps are keyed to the size of the fdarray, since it
90 * grows far faster than any of the other dynamic data. We try to fit
91 * the fdarray into comfortable page-tuned chunks: starting at 1024B
92 * and growing in powers of two from there on.
94 nr /= (1024 / sizeof(struct file *));
95 nr = roundup_pow_of_two(nr + 1);
96 nr *= (1024 / sizeof(struct file *));
98 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
99 * had been set lower between the check in expand_files() and here. Deal
100 * with that in caller, it's cheaper that way.
102 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
103 * bitmaps handling below becomes unpleasant, to put it mildly...
105 if (unlikely(nr > sysctl_nr_open))
106 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
108 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
112 data = alloc_fdmem(nr * sizeof(struct file *));
117 data = alloc_fdmem(max_t(size_t,
118 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
121 fdt->open_fds = data;
122 data += nr / BITS_PER_BYTE;
123 fdt->close_on_exec = data;
136 * Expand the file descriptor table.
137 * This function will allocate a new fdtable and both fd array and fdset, of
139 * Return <0 error code on error; 1 on successful completion.
140 * The files->file_lock should be held on entry, and will be held on exit.
142 static int expand_fdtable(struct files_struct *files, int nr)
143 __releases(files->file_lock)
144 __acquires(files->file_lock)
146 struct fdtable *new_fdt, *cur_fdt;
148 spin_unlock(&files->file_lock);
149 new_fdt = alloc_fdtable(nr);
150 spin_lock(&files->file_lock);
154 * extremely unlikely race - sysctl_nr_open decreased between the check in
155 * caller and alloc_fdtable(). Cheaper to catch it here...
157 if (unlikely(new_fdt->max_fds <= nr)) {
158 __free_fdtable(new_fdt);
162 * Check again since another task may have expanded the fd table while
163 * we dropped the lock
165 cur_fdt = files_fdtable(files);
166 if (nr >= cur_fdt->max_fds) {
167 /* Continue as planned */
168 copy_fdtable(new_fdt, cur_fdt);
169 rcu_assign_pointer(files->fdt, new_fdt);
170 if (cur_fdt != &files->fdtab)
171 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
173 /* Somebody else expanded, so undo our attempt */
174 __free_fdtable(new_fdt);
181 * This function will expand the file structures, if the requested size exceeds
182 * the current capacity and there is room for expansion.
183 * Return <0 error code on error; 0 when nothing done; 1 when files were
184 * expanded and execution may have blocked.
185 * The files->file_lock should be held on entry, and will be held on exit.
187 static int expand_files(struct files_struct *files, int nr)
191 fdt = files_fdtable(files);
193 /* Do we need to expand? */
194 if (nr < fdt->max_fds)
198 if (nr >= sysctl_nr_open)
201 /* All good, so we try */
202 return expand_fdtable(files, nr);
205 static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
207 __set_bit(fd, fdt->close_on_exec);
210 static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
212 __clear_bit(fd, fdt->close_on_exec);
215 static inline void __set_open_fd(int fd, struct fdtable *fdt)
217 __set_bit(fd, fdt->open_fds);
220 static inline void __clear_open_fd(int fd, struct fdtable *fdt)
222 __clear_bit(fd, fdt->open_fds);
225 static int count_open_files(struct fdtable *fdt)
227 int size = fdt->max_fds;
230 /* Find the last open fd */
231 for (i = size / BITS_PER_LONG; i > 0; ) {
232 if (fdt->open_fds[--i])
235 i = (i + 1) * BITS_PER_LONG;
240 * Allocate a new files structure and copy contents from the
241 * passed in files structure.
242 * errorp will be valid only when the returned files_struct is NULL.
244 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
246 struct files_struct *newf;
247 struct file **old_fds, **new_fds;
248 int open_files, size, i;
249 struct fdtable *old_fdt, *new_fdt;
252 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
256 atomic_set(&newf->count, 1);
258 spin_lock_init(&newf->file_lock);
260 new_fdt = &newf->fdtab;
261 new_fdt->max_fds = NR_OPEN_DEFAULT;
262 new_fdt->close_on_exec = newf->close_on_exec_init;
263 new_fdt->open_fds = newf->open_fds_init;
264 new_fdt->fd = &newf->fd_array[0];
266 spin_lock(&oldf->file_lock);
267 old_fdt = files_fdtable(oldf);
268 open_files = count_open_files(old_fdt);
271 * Check whether we need to allocate a larger fd array and fd set.
273 while (unlikely(open_files > new_fdt->max_fds)) {
274 spin_unlock(&oldf->file_lock);
276 if (new_fdt != &newf->fdtab)
277 __free_fdtable(new_fdt);
279 new_fdt = alloc_fdtable(open_files - 1);
285 /* beyond sysctl_nr_open; nothing to do */
286 if (unlikely(new_fdt->max_fds < open_files)) {
287 __free_fdtable(new_fdt);
293 * Reacquire the oldf lock and a pointer to its fd table
294 * who knows it may have a new bigger fd table. We need
295 * the latest pointer.
297 spin_lock(&oldf->file_lock);
298 old_fdt = files_fdtable(oldf);
299 open_files = count_open_files(old_fdt);
302 old_fds = old_fdt->fd;
303 new_fds = new_fdt->fd;
305 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
306 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
308 for (i = open_files; i != 0; i--) {
309 struct file *f = *old_fds++;
314 * The fd may be claimed in the fd bitmap but not yet
315 * instantiated in the files array if a sibling thread
316 * is partway through open(). So make sure that this
317 * fd is available to the new process.
319 __clear_open_fd(open_files - i, new_fdt);
321 rcu_assign_pointer(*new_fds++, f);
323 spin_unlock(&oldf->file_lock);
325 /* compute the remainder to be cleared */
326 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
328 /* This is long word aligned thus could use a optimized version */
329 memset(new_fds, 0, size);
331 if (new_fdt->max_fds > open_files) {
332 int left = (new_fdt->max_fds - open_files) / 8;
333 int start = open_files / BITS_PER_LONG;
335 memset(&new_fdt->open_fds[start], 0, left);
336 memset(&new_fdt->close_on_exec[start], 0, left);
339 rcu_assign_pointer(newf->fdt, new_fdt);
344 kmem_cache_free(files_cachep, newf);
349 static struct fdtable *close_files(struct files_struct * files)
352 * It is safe to dereference the fd table without RCU or
353 * ->file_lock because this is the last reference to the
356 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
361 i = j * BITS_PER_LONG;
362 if (i >= fdt->max_fds)
364 set = fdt->open_fds[j++];
367 struct file * file = xchg(&fdt->fd[i], NULL);
369 filp_close(file, files);
370 cond_resched_rcu_qs();
381 struct files_struct *get_files_struct(struct task_struct *task)
383 struct files_struct *files;
388 atomic_inc(&files->count);
394 void put_files_struct(struct files_struct *files)
396 if (atomic_dec_and_test(&files->count)) {
397 struct fdtable *fdt = close_files(files);
399 /* free the arrays if they are not embedded */
400 if (fdt != &files->fdtab)
402 kmem_cache_free(files_cachep, files);
406 void reset_files_struct(struct files_struct *files)
408 struct task_struct *tsk = current;
409 struct files_struct *old;
415 put_files_struct(old);
418 void exit_files(struct task_struct *tsk)
420 struct files_struct * files = tsk->files;
426 put_files_struct(files);
430 struct files_struct init_files = {
431 .count = ATOMIC_INIT(1),
432 .fdt = &init_files.fdtab,
434 .max_fds = NR_OPEN_DEFAULT,
435 .fd = &init_files.fd_array[0],
436 .close_on_exec = init_files.close_on_exec_init,
437 .open_fds = init_files.open_fds_init,
439 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
443 * allocate a file descriptor, mark it busy.
445 int __alloc_fd(struct files_struct *files,
446 unsigned start, unsigned end, unsigned flags)
452 spin_lock(&files->file_lock);
454 fdt = files_fdtable(files);
456 if (fd < files->next_fd)
459 if (fd < fdt->max_fds)
460 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
463 * N.B. For clone tasks sharing a files structure, this test
464 * will limit the total number of files that can be opened.
470 error = expand_files(files, fd);
475 * If we needed to expand the fs array we
476 * might have blocked - try again.
481 if (start <= files->next_fd)
482 files->next_fd = fd + 1;
484 __set_open_fd(fd, fdt);
485 if (flags & O_CLOEXEC)
486 __set_close_on_exec(fd, fdt);
488 __clear_close_on_exec(fd, fdt);
492 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
493 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
494 rcu_assign_pointer(fdt->fd[fd], NULL);
499 spin_unlock(&files->file_lock);
503 static int alloc_fd(unsigned start, unsigned flags)
505 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
508 int get_unused_fd_flags(unsigned flags)
510 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
512 EXPORT_SYMBOL(get_unused_fd_flags);
514 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
516 struct fdtable *fdt = files_fdtable(files);
517 __clear_open_fd(fd, fdt);
518 if (fd < files->next_fd)
522 void put_unused_fd(unsigned int fd)
524 struct files_struct *files = current->files;
525 spin_lock(&files->file_lock);
526 __put_unused_fd(files, fd);
527 spin_unlock(&files->file_lock);
530 EXPORT_SYMBOL(put_unused_fd);
533 * Install a file pointer in the fd array.
535 * The VFS is full of places where we drop the files lock between
536 * setting the open_fds bitmap and installing the file in the file
537 * array. At any such point, we are vulnerable to a dup2() race
538 * installing a file in the array before us. We need to detect this and
539 * fput() the struct file we are about to overwrite in this case.
541 * It should never happen - if we allow dup2() do it, _really_ bad things
544 * NOTE: __fd_install() variant is really, really low-level; don't
545 * use it unless you are forced to by truly lousy API shoved down
546 * your throat. 'files' *MUST* be either current->files or obtained
547 * by get_files_struct(current) done by whoever had given it to you,
548 * or really bad things will happen. Normally you want to use
549 * fd_install() instead.
552 void __fd_install(struct files_struct *files, unsigned int fd,
556 spin_lock(&files->file_lock);
557 fdt = files_fdtable(files);
558 BUG_ON(fdt->fd[fd] != NULL);
559 rcu_assign_pointer(fdt->fd[fd], file);
560 spin_unlock(&files->file_lock);
563 void fd_install(unsigned int fd, struct file *file)
565 __fd_install(current->files, fd, file);
568 EXPORT_SYMBOL(fd_install);
571 * The same warnings as for __alloc_fd()/__fd_install() apply here...
573 int __close_fd(struct files_struct *files, unsigned fd)
578 spin_lock(&files->file_lock);
579 fdt = files_fdtable(files);
580 if (fd >= fdt->max_fds)
585 rcu_assign_pointer(fdt->fd[fd], NULL);
586 __clear_close_on_exec(fd, fdt);
587 __put_unused_fd(files, fd);
588 spin_unlock(&files->file_lock);
589 return filp_close(file, files);
592 spin_unlock(&files->file_lock);
596 void do_close_on_exec(struct files_struct *files)
601 /* exec unshares first */
602 spin_lock(&files->file_lock);
605 unsigned fd = i * BITS_PER_LONG;
606 fdt = files_fdtable(files);
607 if (fd >= fdt->max_fds)
609 set = fdt->close_on_exec[i];
612 fdt->close_on_exec[i] = 0;
613 for ( ; set ; fd++, set >>= 1) {
620 rcu_assign_pointer(fdt->fd[fd], NULL);
621 __put_unused_fd(files, fd);
622 spin_unlock(&files->file_lock);
623 filp_close(file, files);
625 spin_lock(&files->file_lock);
629 spin_unlock(&files->file_lock);
632 static struct file *__fget(unsigned int fd, fmode_t mask)
634 struct files_struct *files = current->files;
638 file = fcheck_files(files, fd);
640 /* File object ref couldn't be taken */
641 if ((file->f_mode & mask) || !get_file_rcu(file))
649 struct file *fget(unsigned int fd)
651 return __fget(fd, FMODE_PATH);
655 struct file *fget_raw(unsigned int fd)
657 return __fget(fd, 0);
659 EXPORT_SYMBOL(fget_raw);
662 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
664 * You can use this instead of fget if you satisfy all of the following
666 * 1) You must call fput_light before exiting the syscall and returning control
667 * to userspace (i.e. you cannot remember the returned struct file * after
668 * returning to userspace).
669 * 2) You must not call filp_close on the returned struct file * in between
670 * calls to fget_light and fput_light.
671 * 3) You must not clone the current task in between the calls to fget_light
674 * The fput_needed flag returned by fget_light should be passed to the
675 * corresponding fput_light.
677 static unsigned long __fget_light(unsigned int fd, fmode_t mask)
679 struct files_struct *files = current->files;
682 if (atomic_read(&files->count) == 1) {
683 file = __fcheck_files(files, fd);
684 if (!file || unlikely(file->f_mode & mask))
686 return (unsigned long)file;
688 file = __fget(fd, mask);
691 return FDPUT_FPUT | (unsigned long)file;
694 unsigned long __fdget(unsigned int fd)
696 return __fget_light(fd, FMODE_PATH);
698 EXPORT_SYMBOL(__fdget);
700 unsigned long __fdget_raw(unsigned int fd)
702 return __fget_light(fd, 0);
705 unsigned long __fdget_pos(unsigned int fd)
707 unsigned long v = __fdget(fd);
708 struct file *file = (struct file *)(v & ~3);
710 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
711 if (file_count(file) > 1) {
712 v |= FDPUT_POS_UNLOCK;
713 mutex_lock(&file->f_pos_lock);
720 * We only lock f_pos if we have threads or if the file might be
721 * shared with another process. In both cases we'll have an elevated
722 * file count (done either by fdget() or by fork()).
725 void set_close_on_exec(unsigned int fd, int flag)
727 struct files_struct *files = current->files;
729 spin_lock(&files->file_lock);
730 fdt = files_fdtable(files);
732 __set_close_on_exec(fd, fdt);
734 __clear_close_on_exec(fd, fdt);
735 spin_unlock(&files->file_lock);
738 bool get_close_on_exec(unsigned int fd)
740 struct files_struct *files = current->files;
744 fdt = files_fdtable(files);
745 res = close_on_exec(fd, fdt);
750 static int do_dup2(struct files_struct *files,
751 struct file *file, unsigned fd, unsigned flags)
752 __releases(&files->file_lock)
758 * We need to detect attempts to do dup2() over allocated but still
759 * not finished descriptor. NB: OpenBSD avoids that at the price of
760 * extra work in their equivalent of fget() - they insert struct
761 * file immediately after grabbing descriptor, mark it larval if
762 * more work (e.g. actual opening) is needed and make sure that
763 * fget() treats larval files as absent. Potentially interesting,
764 * but while extra work in fget() is trivial, locking implications
765 * and amount of surgery on open()-related paths in VFS are not.
766 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
767 * deadlocks in rather amusing ways, AFAICS. All of that is out of
768 * scope of POSIX or SUS, since neither considers shared descriptor
769 * tables and this condition does not arise without those.
771 fdt = files_fdtable(files);
772 tofree = fdt->fd[fd];
773 if (!tofree && fd_is_open(fd, fdt))
776 rcu_assign_pointer(fdt->fd[fd], file);
777 __set_open_fd(fd, fdt);
778 if (flags & O_CLOEXEC)
779 __set_close_on_exec(fd, fdt);
781 __clear_close_on_exec(fd, fdt);
782 spin_unlock(&files->file_lock);
785 filp_close(tofree, files);
790 spin_unlock(&files->file_lock);
794 int replace_fd(unsigned fd, struct file *file, unsigned flags)
797 struct files_struct *files = current->files;
800 return __close_fd(files, fd);
802 if (fd >= rlimit(RLIMIT_NOFILE))
805 spin_lock(&files->file_lock);
806 err = expand_files(files, fd);
807 if (unlikely(err < 0))
809 return do_dup2(files, file, fd, flags);
812 spin_unlock(&files->file_lock);
816 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
820 struct files_struct *files = current->files;
822 if ((flags & ~O_CLOEXEC) != 0)
825 if (unlikely(oldfd == newfd))
828 if (newfd >= rlimit(RLIMIT_NOFILE))
831 spin_lock(&files->file_lock);
832 err = expand_files(files, newfd);
833 file = fcheck(oldfd);
836 if (unlikely(err < 0)) {
841 return do_dup2(files, file, newfd, flags);
846 spin_unlock(&files->file_lock);
850 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
852 if (unlikely(newfd == oldfd)) { /* corner case */
853 struct files_struct *files = current->files;
857 if (!fcheck_files(files, oldfd))
862 return sys_dup3(oldfd, newfd, 0);
865 SYSCALL_DEFINE1(dup, unsigned int, fildes)
868 struct file *file = fget_raw(fildes);
871 ret = get_unused_fd_flags(0);
873 fd_install(ret, file);
880 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
883 if (from >= rlimit(RLIMIT_NOFILE))
885 err = alloc_fd(from, flags);
888 fd_install(err, file);
893 int iterate_fd(struct files_struct *files, unsigned n,
894 int (*f)(const void *, struct file *, unsigned),
901 spin_lock(&files->file_lock);
902 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
904 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
911 spin_unlock(&files->file_lock);
914 EXPORT_SYMBOL(iterate_fd);