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));
59 #define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
60 #define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
63 * Copy 'count' fd bits from the old table to the new table and clear the extra
64 * space if any. This does not copy the file pointers. Called with the files
65 * spinlock held for write.
67 static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
70 unsigned int cpy, set;
72 cpy = count / BITS_PER_BYTE;
73 set = (nfdt->max_fds - count) / BITS_PER_BYTE;
74 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
75 memset((char *)nfdt->open_fds + cpy, 0, set);
76 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
77 memset((char *)nfdt->close_on_exec + cpy, 0, set);
79 cpy = BITBIT_SIZE(count);
80 set = BITBIT_SIZE(nfdt->max_fds) - cpy;
81 memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
82 memset((char *)nfdt->full_fds_bits + cpy, 0, set);
86 * Copy all file descriptors from the old table to the new, expanded table and
87 * clear the extra space. Called with the files spinlock held for write.
89 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
91 unsigned int cpy, set;
93 BUG_ON(nfdt->max_fds < ofdt->max_fds);
95 cpy = ofdt->max_fds * sizeof(struct file *);
96 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
97 memcpy(nfdt->fd, ofdt->fd, cpy);
98 memset((char *)nfdt->fd + cpy, 0, set);
100 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
103 static struct fdtable * alloc_fdtable(unsigned int nr)
109 * Figure out how many fds we actually want to support in this fdtable.
110 * Allocation steps are keyed to the size of the fdarray, since it
111 * grows far faster than any of the other dynamic data. We try to fit
112 * the fdarray into comfortable page-tuned chunks: starting at 1024B
113 * and growing in powers of two from there on.
115 nr /= (1024 / sizeof(struct file *));
116 nr = roundup_pow_of_two(nr + 1);
117 nr *= (1024 / sizeof(struct file *));
119 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
120 * had been set lower between the check in expand_files() and here. Deal
121 * with that in caller, it's cheaper that way.
123 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
124 * bitmaps handling below becomes unpleasant, to put it mildly...
126 if (unlikely(nr > sysctl_nr_open))
127 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
129 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
133 data = alloc_fdmem(nr * sizeof(struct file *));
138 data = alloc_fdmem(max_t(size_t,
139 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES));
142 fdt->open_fds = data;
143 data += nr / BITS_PER_BYTE;
144 fdt->close_on_exec = data;
145 data += nr / BITS_PER_BYTE;
146 fdt->full_fds_bits = data;
159 * Expand the file descriptor table.
160 * This function will allocate a new fdtable and both fd array and fdset, of
162 * Return <0 error code on error; 1 on successful completion.
163 * The files->file_lock should be held on entry, and will be held on exit.
165 static int expand_fdtable(struct files_struct *files, int nr)
166 __releases(files->file_lock)
167 __acquires(files->file_lock)
169 struct fdtable *new_fdt, *cur_fdt;
171 spin_unlock(&files->file_lock);
172 new_fdt = alloc_fdtable(nr);
174 /* make sure all __fd_install() have seen resize_in_progress
175 * or have finished their rcu_read_lock_sched() section.
177 if (atomic_read(&files->count) > 1)
180 spin_lock(&files->file_lock);
184 * extremely unlikely race - sysctl_nr_open decreased between the check in
185 * caller and alloc_fdtable(). Cheaper to catch it here...
187 if (unlikely(new_fdt->max_fds <= nr)) {
188 __free_fdtable(new_fdt);
191 cur_fdt = files_fdtable(files);
192 BUG_ON(nr < cur_fdt->max_fds);
193 copy_fdtable(new_fdt, cur_fdt);
194 rcu_assign_pointer(files->fdt, new_fdt);
195 if (cur_fdt != &files->fdtab)
196 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
197 /* coupled with smp_rmb() in __fd_install() */
204 * This function will expand the file structures, if the requested size exceeds
205 * the current capacity and there is room for expansion.
206 * Return <0 error code on error; 0 when nothing done; 1 when files were
207 * expanded and execution may have blocked.
208 * The files->file_lock should be held on entry, and will be held on exit.
210 static int expand_files(struct files_struct *files, int nr)
211 __releases(files->file_lock)
212 __acquires(files->file_lock)
218 fdt = files_fdtable(files);
220 /* Do we need to expand? */
221 if (nr < fdt->max_fds)
225 if (nr >= sysctl_nr_open)
228 if (unlikely(files->resize_in_progress)) {
229 spin_unlock(&files->file_lock);
231 wait_event(files->resize_wait, !files->resize_in_progress);
232 spin_lock(&files->file_lock);
236 /* All good, so we try */
237 files->resize_in_progress = true;
238 expanded = expand_fdtable(files, nr);
239 files->resize_in_progress = false;
241 wake_up_all(&files->resize_wait);
245 static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
247 __set_bit(fd, fdt->close_on_exec);
250 static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
252 if (test_bit(fd, fdt->close_on_exec))
253 __clear_bit(fd, fdt->close_on_exec);
256 static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
258 __set_bit(fd, fdt->open_fds);
260 if (!~fdt->open_fds[fd])
261 __set_bit(fd, fdt->full_fds_bits);
264 static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
266 __clear_bit(fd, fdt->open_fds);
267 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
270 static int count_open_files(struct fdtable *fdt)
272 int size = fdt->max_fds;
275 /* Find the last open fd */
276 for (i = size / BITS_PER_LONG; i > 0; ) {
277 if (fdt->open_fds[--i])
280 i = (i + 1) * BITS_PER_LONG;
285 * Allocate a new files structure and copy contents from the
286 * passed in files structure.
287 * errorp will be valid only when the returned files_struct is NULL.
289 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
291 struct files_struct *newf;
292 struct file **old_fds, **new_fds;
294 struct fdtable *old_fdt, *new_fdt;
297 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
301 atomic_set(&newf->count, 1);
303 spin_lock_init(&newf->file_lock);
304 newf->resize_in_progress = false;
305 init_waitqueue_head(&newf->resize_wait);
307 new_fdt = &newf->fdtab;
308 new_fdt->max_fds = NR_OPEN_DEFAULT;
309 new_fdt->close_on_exec = newf->close_on_exec_init;
310 new_fdt->open_fds = newf->open_fds_init;
311 new_fdt->full_fds_bits = newf->full_fds_bits_init;
312 new_fdt->fd = &newf->fd_array[0];
314 spin_lock(&oldf->file_lock);
315 old_fdt = files_fdtable(oldf);
316 open_files = count_open_files(old_fdt);
319 * Check whether we need to allocate a larger fd array and fd set.
321 while (unlikely(open_files > new_fdt->max_fds)) {
322 spin_unlock(&oldf->file_lock);
324 if (new_fdt != &newf->fdtab)
325 __free_fdtable(new_fdt);
327 new_fdt = alloc_fdtable(open_files - 1);
333 /* beyond sysctl_nr_open; nothing to do */
334 if (unlikely(new_fdt->max_fds < open_files)) {
335 __free_fdtable(new_fdt);
341 * Reacquire the oldf lock and a pointer to its fd table
342 * who knows it may have a new bigger fd table. We need
343 * the latest pointer.
345 spin_lock(&oldf->file_lock);
346 old_fdt = files_fdtable(oldf);
347 open_files = count_open_files(old_fdt);
350 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
352 old_fds = old_fdt->fd;
353 new_fds = new_fdt->fd;
355 for (i = open_files; i != 0; i--) {
356 struct file *f = *old_fds++;
361 * The fd may be claimed in the fd bitmap but not yet
362 * instantiated in the files array if a sibling thread
363 * is partway through open(). So make sure that this
364 * fd is available to the new process.
366 __clear_open_fd(open_files - i, new_fdt);
368 rcu_assign_pointer(*new_fds++, f);
370 spin_unlock(&oldf->file_lock);
372 /* clear the remainder */
373 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
375 rcu_assign_pointer(newf->fdt, new_fdt);
380 kmem_cache_free(files_cachep, newf);
385 static struct fdtable *close_files(struct files_struct * files)
388 * It is safe to dereference the fd table without RCU or
389 * ->file_lock because this is the last reference to the
392 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
397 i = j * BITS_PER_LONG;
398 if (i >= fdt->max_fds)
400 set = fdt->open_fds[j++];
403 struct file * file = xchg(&fdt->fd[i], NULL);
405 filp_close(file, files);
406 cond_resched_rcu_qs();
417 struct files_struct *get_files_struct(struct task_struct *task)
419 struct files_struct *files;
424 atomic_inc(&files->count);
430 void put_files_struct(struct files_struct *files)
432 if (atomic_dec_and_test(&files->count)) {
433 struct fdtable *fdt = close_files(files);
435 /* free the arrays if they are not embedded */
436 if (fdt != &files->fdtab)
438 kmem_cache_free(files_cachep, files);
442 void reset_files_struct(struct files_struct *files)
444 struct task_struct *tsk = current;
445 struct files_struct *old;
451 put_files_struct(old);
454 void exit_files(struct task_struct *tsk)
456 struct files_struct * files = tsk->files;
462 put_files_struct(files);
466 struct files_struct init_files = {
467 .count = ATOMIC_INIT(1),
468 .fdt = &init_files.fdtab,
470 .max_fds = NR_OPEN_DEFAULT,
471 .fd = &init_files.fd_array[0],
472 .close_on_exec = init_files.close_on_exec_init,
473 .open_fds = init_files.open_fds_init,
474 .full_fds_bits = init_files.full_fds_bits_init,
476 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
479 static unsigned long find_next_fd(struct fdtable *fdt, unsigned long start)
481 unsigned long maxfd = fdt->max_fds;
482 unsigned long maxbit = maxfd / BITS_PER_LONG;
483 unsigned long bitbit = start / BITS_PER_LONG;
485 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
490 return find_next_zero_bit(fdt->open_fds, maxfd, start);
494 * allocate a file descriptor, mark it busy.
496 int __alloc_fd(struct files_struct *files,
497 unsigned start, unsigned end, unsigned flags)
503 spin_lock(&files->file_lock);
505 fdt = files_fdtable(files);
507 if (fd < files->next_fd)
510 if (fd < fdt->max_fds)
511 fd = find_next_fd(fdt, fd);
514 * N.B. For clone tasks sharing a files structure, this test
515 * will limit the total number of files that can be opened.
521 error = expand_files(files, fd);
526 * If we needed to expand the fs array we
527 * might have blocked - try again.
532 if (start <= files->next_fd)
533 files->next_fd = fd + 1;
535 __set_open_fd(fd, fdt);
536 if (flags & O_CLOEXEC)
537 __set_close_on_exec(fd, fdt);
539 __clear_close_on_exec(fd, fdt);
543 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
544 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
545 rcu_assign_pointer(fdt->fd[fd], NULL);
550 spin_unlock(&files->file_lock);
554 static int alloc_fd(unsigned start, unsigned flags)
556 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
559 int get_unused_fd_flags(unsigned flags)
561 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
563 EXPORT_SYMBOL(get_unused_fd_flags);
565 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
567 struct fdtable *fdt = files_fdtable(files);
568 __clear_open_fd(fd, fdt);
569 if (fd < files->next_fd)
573 void put_unused_fd(unsigned int fd)
575 struct files_struct *files = current->files;
576 spin_lock(&files->file_lock);
577 __put_unused_fd(files, fd);
578 spin_unlock(&files->file_lock);
581 EXPORT_SYMBOL(put_unused_fd);
584 * Install a file pointer in the fd array.
586 * The VFS is full of places where we drop the files lock between
587 * setting the open_fds bitmap and installing the file in the file
588 * array. At any such point, we are vulnerable to a dup2() race
589 * installing a file in the array before us. We need to detect this and
590 * fput() the struct file we are about to overwrite in this case.
592 * It should never happen - if we allow dup2() do it, _really_ bad things
595 * NOTE: __fd_install() variant is really, really low-level; don't
596 * use it unless you are forced to by truly lousy API shoved down
597 * your throat. 'files' *MUST* be either current->files or obtained
598 * by get_files_struct(current) done by whoever had given it to you,
599 * or really bad things will happen. Normally you want to use
600 * fd_install() instead.
603 void __fd_install(struct files_struct *files, unsigned int fd,
609 rcu_read_lock_sched();
611 while (unlikely(files->resize_in_progress)) {
612 rcu_read_unlock_sched();
613 wait_event(files->resize_wait, !files->resize_in_progress);
614 rcu_read_lock_sched();
616 /* coupled with smp_wmb() in expand_fdtable() */
618 fdt = rcu_dereference_sched(files->fdt);
619 BUG_ON(fdt->fd[fd] != NULL);
620 rcu_assign_pointer(fdt->fd[fd], file);
621 rcu_read_unlock_sched();
624 void fd_install(unsigned int fd, struct file *file)
626 __fd_install(current->files, fd, file);
629 EXPORT_SYMBOL(fd_install);
632 * The same warnings as for __alloc_fd()/__fd_install() apply here...
634 int __close_fd(struct files_struct *files, unsigned fd)
639 spin_lock(&files->file_lock);
640 fdt = files_fdtable(files);
641 if (fd >= fdt->max_fds)
646 rcu_assign_pointer(fdt->fd[fd], NULL);
647 __clear_close_on_exec(fd, fdt);
648 __put_unused_fd(files, fd);
649 spin_unlock(&files->file_lock);
650 return filp_close(file, files);
653 spin_unlock(&files->file_lock);
657 void do_close_on_exec(struct files_struct *files)
662 /* exec unshares first */
663 spin_lock(&files->file_lock);
666 unsigned fd = i * BITS_PER_LONG;
667 fdt = files_fdtable(files);
668 if (fd >= fdt->max_fds)
670 set = fdt->close_on_exec[i];
673 fdt->close_on_exec[i] = 0;
674 for ( ; set ; fd++, set >>= 1) {
681 rcu_assign_pointer(fdt->fd[fd], NULL);
682 __put_unused_fd(files, fd);
683 spin_unlock(&files->file_lock);
684 filp_close(file, files);
686 spin_lock(&files->file_lock);
690 spin_unlock(&files->file_lock);
693 static struct file *__fget(unsigned int fd, fmode_t mask)
695 struct files_struct *files = current->files;
700 file = fcheck_files(files, fd);
702 /* File object ref couldn't be taken.
703 * dup2() atomicity guarantee is the reason
704 * we loop to catch the new file (or NULL pointer)
706 if (file->f_mode & mask)
708 else if (!get_file_rcu(file))
716 struct file *fget(unsigned int fd)
718 return __fget(fd, FMODE_PATH);
722 struct file *fget_raw(unsigned int fd)
724 return __fget(fd, 0);
726 EXPORT_SYMBOL(fget_raw);
729 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
731 * You can use this instead of fget if you satisfy all of the following
733 * 1) You must call fput_light before exiting the syscall and returning control
734 * to userspace (i.e. you cannot remember the returned struct file * after
735 * returning to userspace).
736 * 2) You must not call filp_close on the returned struct file * in between
737 * calls to fget_light and fput_light.
738 * 3) You must not clone the current task in between the calls to fget_light
741 * The fput_needed flag returned by fget_light should be passed to the
742 * corresponding fput_light.
744 static unsigned long __fget_light(unsigned int fd, fmode_t mask)
746 struct files_struct *files = current->files;
749 if (atomic_read(&files->count) == 1) {
750 file = __fcheck_files(files, fd);
751 if (!file || unlikely(file->f_mode & mask))
753 return (unsigned long)file;
755 file = __fget(fd, mask);
758 return FDPUT_FPUT | (unsigned long)file;
761 unsigned long __fdget(unsigned int fd)
763 return __fget_light(fd, FMODE_PATH);
765 EXPORT_SYMBOL(__fdget);
767 unsigned long __fdget_raw(unsigned int fd)
769 return __fget_light(fd, 0);
772 unsigned long __fdget_pos(unsigned int fd)
774 unsigned long v = __fdget(fd);
775 struct file *file = (struct file *)(v & ~3);
777 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
778 if (file_count(file) > 1) {
779 v |= FDPUT_POS_UNLOCK;
780 mutex_lock(&file->f_pos_lock);
787 * We only lock f_pos if we have threads or if the file might be
788 * shared with another process. In both cases we'll have an elevated
789 * file count (done either by fdget() or by fork()).
792 void set_close_on_exec(unsigned int fd, int flag)
794 struct files_struct *files = current->files;
796 spin_lock(&files->file_lock);
797 fdt = files_fdtable(files);
799 __set_close_on_exec(fd, fdt);
801 __clear_close_on_exec(fd, fdt);
802 spin_unlock(&files->file_lock);
805 bool get_close_on_exec(unsigned int fd)
807 struct files_struct *files = current->files;
811 fdt = files_fdtable(files);
812 res = close_on_exec(fd, fdt);
817 static int do_dup2(struct files_struct *files,
818 struct file *file, unsigned fd, unsigned flags)
819 __releases(&files->file_lock)
825 * We need to detect attempts to do dup2() over allocated but still
826 * not finished descriptor. NB: OpenBSD avoids that at the price of
827 * extra work in their equivalent of fget() - they insert struct
828 * file immediately after grabbing descriptor, mark it larval if
829 * more work (e.g. actual opening) is needed and make sure that
830 * fget() treats larval files as absent. Potentially interesting,
831 * but while extra work in fget() is trivial, locking implications
832 * and amount of surgery on open()-related paths in VFS are not.
833 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
834 * deadlocks in rather amusing ways, AFAICS. All of that is out of
835 * scope of POSIX or SUS, since neither considers shared descriptor
836 * tables and this condition does not arise without those.
838 fdt = files_fdtable(files);
839 tofree = fdt->fd[fd];
840 if (!tofree && fd_is_open(fd, fdt))
843 rcu_assign_pointer(fdt->fd[fd], file);
844 __set_open_fd(fd, fdt);
845 if (flags & O_CLOEXEC)
846 __set_close_on_exec(fd, fdt);
848 __clear_close_on_exec(fd, fdt);
849 spin_unlock(&files->file_lock);
852 filp_close(tofree, files);
857 spin_unlock(&files->file_lock);
861 int replace_fd(unsigned fd, struct file *file, unsigned flags)
864 struct files_struct *files = current->files;
867 return __close_fd(files, fd);
869 if (fd >= rlimit(RLIMIT_NOFILE))
872 spin_lock(&files->file_lock);
873 err = expand_files(files, fd);
874 if (unlikely(err < 0))
876 return do_dup2(files, file, fd, flags);
879 spin_unlock(&files->file_lock);
883 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
887 struct files_struct *files = current->files;
889 if ((flags & ~O_CLOEXEC) != 0)
892 if (unlikely(oldfd == newfd))
895 if (newfd >= rlimit(RLIMIT_NOFILE))
898 spin_lock(&files->file_lock);
899 err = expand_files(files, newfd);
900 file = fcheck(oldfd);
903 if (unlikely(err < 0)) {
908 return do_dup2(files, file, newfd, flags);
913 spin_unlock(&files->file_lock);
917 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
919 if (unlikely(newfd == oldfd)) { /* corner case */
920 struct files_struct *files = current->files;
924 if (!fcheck_files(files, oldfd))
929 return sys_dup3(oldfd, newfd, 0);
932 SYSCALL_DEFINE1(dup, unsigned int, fildes)
935 struct file *file = fget_raw(fildes);
938 ret = get_unused_fd_flags(0);
940 fd_install(ret, file);
947 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
950 if (from >= rlimit(RLIMIT_NOFILE))
952 err = alloc_fd(from, flags);
955 fd_install(err, file);
960 int iterate_fd(struct files_struct *files, unsigned n,
961 int (*f)(const void *, struct file *, unsigned),
968 spin_lock(&files->file_lock);
969 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
971 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
978 spin_unlock(&files->file_lock);
981 EXPORT_SYMBOL(iterate_fd);