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 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 *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_ACCOUNT |
41 __GFP_NOWARN | __GFP_NORETRY);
45 return __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM, PAGE_KERNEL);
48 static void __free_fdtable(struct fdtable *fdt)
51 kvfree(fdt->open_fds);
55 static void free_fdtable_rcu(struct rcu_head *rcu)
57 __free_fdtable(container_of(rcu, struct fdtable, rcu));
60 #define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
61 #define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
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.
68 static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
71 unsigned int cpy, set;
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);
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);
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.
90 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
92 unsigned int cpy, set;
94 BUG_ON(nfdt->max_fds < ofdt->max_fds);
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);
101 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
104 static struct fdtable * alloc_fdtable(unsigned int nr)
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.
116 nr /= (1024 / sizeof(struct file *));
117 nr = roundup_pow_of_two(nr + 1);
118 nr *= (1024 / sizeof(struct file *));
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.
124 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
125 * bitmaps handling below becomes unpleasant, to put it mildly...
127 if (unlikely(nr > sysctl_nr_open))
128 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
130 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
134 data = alloc_fdmem(nr * sizeof(struct file *));
139 data = alloc_fdmem(max_t(size_t,
140 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES));
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;
160 * Expand the file descriptor table.
161 * This function will allocate a new fdtable and both fd array and fdset, of
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.
166 static int expand_fdtable(struct files_struct *files, unsigned int nr)
167 __releases(files->file_lock)
168 __acquires(files->file_lock)
170 struct fdtable *new_fdt, *cur_fdt;
172 spin_unlock(&files->file_lock);
173 new_fdt = alloc_fdtable(nr);
175 /* make sure all __fd_install() have seen resize_in_progress
176 * or have finished their rcu_read_lock_sched() section.
178 if (atomic_read(&files->count) > 1)
181 spin_lock(&files->file_lock);
185 * extremely unlikely race - sysctl_nr_open decreased between the check in
186 * caller and alloc_fdtable(). Cheaper to catch it here...
188 if (unlikely(new_fdt->max_fds <= nr)) {
189 __free_fdtable(new_fdt);
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() */
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.
211 static int expand_files(struct files_struct *files, unsigned int nr)
212 __releases(files->file_lock)
213 __acquires(files->file_lock)
219 fdt = files_fdtable(files);
221 /* Do we need to expand? */
222 if (nr < fdt->max_fds)
226 if (nr >= sysctl_nr_open)
229 if (unlikely(files->resize_in_progress)) {
230 spin_unlock(&files->file_lock);
232 wait_event(files->resize_wait, !files->resize_in_progress);
233 spin_lock(&files->file_lock);
237 /* All good, so we try */
238 files->resize_in_progress = true;
239 expanded = expand_fdtable(files, nr);
240 files->resize_in_progress = false;
242 wake_up_all(&files->resize_wait);
246 static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
248 __set_bit(fd, fdt->close_on_exec);
251 static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
253 if (test_bit(fd, fdt->close_on_exec))
254 __clear_bit(fd, fdt->close_on_exec);
257 static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
259 __set_bit(fd, fdt->open_fds);
261 if (!~fdt->open_fds[fd])
262 __set_bit(fd, fdt->full_fds_bits);
265 static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
267 __clear_bit(fd, fdt->open_fds);
268 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
271 static unsigned int count_open_files(struct fdtable *fdt)
273 unsigned int size = fdt->max_fds;
276 /* Find the last open fd */
277 for (i = size / BITS_PER_LONG; i > 0; ) {
278 if (fdt->open_fds[--i])
281 i = (i + 1) * BITS_PER_LONG;
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.
290 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
292 struct files_struct *newf;
293 struct file **old_fds, **new_fds;
294 unsigned int open_files, i;
295 struct fdtable *old_fdt, *new_fdt;
298 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
302 atomic_set(&newf->count, 1);
304 spin_lock_init(&newf->file_lock);
305 newf->resize_in_progress = false;
306 init_waitqueue_head(&newf->resize_wait);
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];
315 spin_lock(&oldf->file_lock);
316 old_fdt = files_fdtable(oldf);
317 open_files = count_open_files(old_fdt);
320 * Check whether we need to allocate a larger fd array and fd set.
322 while (unlikely(open_files > new_fdt->max_fds)) {
323 spin_unlock(&oldf->file_lock);
325 if (new_fdt != &newf->fdtab)
326 __free_fdtable(new_fdt);
328 new_fdt = alloc_fdtable(open_files - 1);
334 /* beyond sysctl_nr_open; nothing to do */
335 if (unlikely(new_fdt->max_fds < open_files)) {
336 __free_fdtable(new_fdt);
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.
346 spin_lock(&oldf->file_lock);
347 old_fdt = files_fdtable(oldf);
348 open_files = count_open_files(old_fdt);
351 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
353 old_fds = old_fdt->fd;
354 new_fds = new_fdt->fd;
356 for (i = open_files; i != 0; i--) {
357 struct file *f = *old_fds++;
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.
367 __clear_open_fd(open_files - i, new_fdt);
369 rcu_assign_pointer(*new_fds++, f);
371 spin_unlock(&oldf->file_lock);
373 /* clear the remainder */
374 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
376 rcu_assign_pointer(newf->fdt, new_fdt);
381 kmem_cache_free(files_cachep, newf);
386 static struct fdtable *close_files(struct files_struct * files)
389 * It is safe to dereference the fd table without RCU or
390 * ->file_lock because this is the last reference to the
393 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
394 unsigned int i, j = 0;
398 i = j * BITS_PER_LONG;
399 if (i >= fdt->max_fds)
401 set = fdt->open_fds[j++];
404 struct file * file = xchg(&fdt->fd[i], NULL);
406 filp_close(file, files);
407 cond_resched_rcu_qs();
418 struct files_struct *get_files_struct(struct task_struct *task)
420 struct files_struct *files;
425 atomic_inc(&files->count);
431 void put_files_struct(struct files_struct *files)
433 if (atomic_dec_and_test(&files->count)) {
434 struct fdtable *fdt = close_files(files);
436 /* free the arrays if they are not embedded */
437 if (fdt != &files->fdtab)
439 kmem_cache_free(files_cachep, files);
443 void reset_files_struct(struct files_struct *files)
445 struct task_struct *tsk = current;
446 struct files_struct *old;
452 put_files_struct(old);
455 void exit_files(struct task_struct *tsk)
457 struct files_struct * files = tsk->files;
463 put_files_struct(files);
467 struct files_struct init_files = {
468 .count = ATOMIC_INIT(1),
469 .fdt = &init_files.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,
477 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
480 static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
482 unsigned int maxfd = fdt->max_fds;
483 unsigned int maxbit = maxfd / BITS_PER_LONG;
484 unsigned int bitbit = start / BITS_PER_LONG;
486 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
491 return find_next_zero_bit(fdt->open_fds, maxfd, start);
495 * allocate a file descriptor, mark it busy.
497 int __alloc_fd(struct files_struct *files,
498 unsigned start, unsigned end, unsigned flags)
504 spin_lock(&files->file_lock);
506 fdt = files_fdtable(files);
508 if (fd < files->next_fd)
511 if (fd < fdt->max_fds)
512 fd = find_next_fd(fdt, fd);
515 * N.B. For clone tasks sharing a files structure, this test
516 * will limit the total number of files that can be opened.
522 error = expand_files(files, fd);
527 * If we needed to expand the fs array we
528 * might have blocked - try again.
533 if (start <= files->next_fd)
534 files->next_fd = fd + 1;
536 __set_open_fd(fd, fdt);
537 if (flags & O_CLOEXEC)
538 __set_close_on_exec(fd, fdt);
540 __clear_close_on_exec(fd, fdt);
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);
551 spin_unlock(&files->file_lock);
555 static int alloc_fd(unsigned start, unsigned flags)
557 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
560 int get_unused_fd_flags(unsigned flags)
562 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
564 EXPORT_SYMBOL(get_unused_fd_flags);
566 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
568 struct fdtable *fdt = files_fdtable(files);
569 __clear_open_fd(fd, fdt);
570 if (fd < files->next_fd)
574 void put_unused_fd(unsigned int fd)
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);
582 EXPORT_SYMBOL(put_unused_fd);
585 * Install a file pointer in the fd array.
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.
593 * It should never happen - if we allow dup2() do it, _really_ bad things
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.
604 void __fd_install(struct files_struct *files, unsigned int fd,
610 rcu_read_lock_sched();
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();
617 /* coupled with smp_wmb() in expand_fdtable() */
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();
625 void fd_install(unsigned int fd, struct file *file)
627 __fd_install(current->files, fd, file);
630 EXPORT_SYMBOL(fd_install);
633 * The same warnings as for __alloc_fd()/__fd_install() apply here...
635 int __close_fd(struct files_struct *files, unsigned fd)
640 spin_lock(&files->file_lock);
641 fdt = files_fdtable(files);
642 if (fd >= fdt->max_fds)
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);
654 spin_unlock(&files->file_lock);
658 void do_close_on_exec(struct files_struct *files)
663 /* exec unshares first */
664 spin_lock(&files->file_lock);
667 unsigned fd = i * BITS_PER_LONG;
668 fdt = files_fdtable(files);
669 if (fd >= fdt->max_fds)
671 set = fdt->close_on_exec[i];
674 fdt->close_on_exec[i] = 0;
675 for ( ; set ; fd++, set >>= 1) {
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);
687 spin_lock(&files->file_lock);
691 spin_unlock(&files->file_lock);
694 static struct file *__fget(unsigned int fd, fmode_t mask)
696 struct files_struct *files = current->files;
701 file = fcheck_files(files, fd);
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)
707 if (file->f_mode & mask)
709 else if (!get_file_rcu(file))
717 struct file *fget(unsigned int fd)
719 return __fget(fd, FMODE_PATH);
723 struct file *fget_raw(unsigned int fd)
725 return __fget(fd, 0);
727 EXPORT_SYMBOL(fget_raw);
730 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
732 * You can use this instead of fget if you satisfy all of the following
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
742 * The fput_needed flag returned by fget_light should be passed to the
743 * corresponding fput_light.
745 static unsigned long __fget_light(unsigned int fd, fmode_t mask)
747 struct files_struct *files = current->files;
750 if (atomic_read(&files->count) == 1) {
751 file = __fcheck_files(files, fd);
752 if (!file || unlikely(file->f_mode & mask))
754 return (unsigned long)file;
756 file = __fget(fd, mask);
759 return FDPUT_FPUT | (unsigned long)file;
762 unsigned long __fdget(unsigned int fd)
764 return __fget_light(fd, FMODE_PATH);
766 EXPORT_SYMBOL(__fdget);
768 unsigned long __fdget_raw(unsigned int fd)
770 return __fget_light(fd, 0);
773 unsigned long __fdget_pos(unsigned int fd)
775 unsigned long v = __fdget(fd);
776 struct file *file = (struct file *)(v & ~3);
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);
787 void __f_unlock_pos(struct file *f)
789 mutex_unlock(&f->f_pos_lock);
793 * We only lock f_pos if we have threads or if the file might be
794 * shared with another process. In both cases we'll have an elevated
795 * file count (done either by fdget() or by fork()).
798 void set_close_on_exec(unsigned int fd, int flag)
800 struct files_struct *files = current->files;
802 spin_lock(&files->file_lock);
803 fdt = files_fdtable(files);
805 __set_close_on_exec(fd, fdt);
807 __clear_close_on_exec(fd, fdt);
808 spin_unlock(&files->file_lock);
811 bool get_close_on_exec(unsigned int fd)
813 struct files_struct *files = current->files;
817 fdt = files_fdtable(files);
818 res = close_on_exec(fd, fdt);
823 static int do_dup2(struct files_struct *files,
824 struct file *file, unsigned fd, unsigned flags)
825 __releases(&files->file_lock)
831 * We need to detect attempts to do dup2() over allocated but still
832 * not finished descriptor. NB: OpenBSD avoids that at the price of
833 * extra work in their equivalent of fget() - they insert struct
834 * file immediately after grabbing descriptor, mark it larval if
835 * more work (e.g. actual opening) is needed and make sure that
836 * fget() treats larval files as absent. Potentially interesting,
837 * but while extra work in fget() is trivial, locking implications
838 * and amount of surgery on open()-related paths in VFS are not.
839 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
840 * deadlocks in rather amusing ways, AFAICS. All of that is out of
841 * scope of POSIX or SUS, since neither considers shared descriptor
842 * tables and this condition does not arise without those.
844 fdt = files_fdtable(files);
845 tofree = fdt->fd[fd];
846 if (!tofree && fd_is_open(fd, fdt))
849 rcu_assign_pointer(fdt->fd[fd], file);
850 __set_open_fd(fd, fdt);
851 if (flags & O_CLOEXEC)
852 __set_close_on_exec(fd, fdt);
854 __clear_close_on_exec(fd, fdt);
855 spin_unlock(&files->file_lock);
858 filp_close(tofree, files);
863 spin_unlock(&files->file_lock);
867 int replace_fd(unsigned fd, struct file *file, unsigned flags)
870 struct files_struct *files = current->files;
873 return __close_fd(files, fd);
875 if (fd >= rlimit(RLIMIT_NOFILE))
878 spin_lock(&files->file_lock);
879 err = expand_files(files, fd);
880 if (unlikely(err < 0))
882 return do_dup2(files, file, fd, flags);
885 spin_unlock(&files->file_lock);
889 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
893 struct files_struct *files = current->files;
895 if ((flags & ~O_CLOEXEC) != 0)
898 if (unlikely(oldfd == newfd))
901 if (newfd >= rlimit(RLIMIT_NOFILE))
904 spin_lock(&files->file_lock);
905 err = expand_files(files, newfd);
906 file = fcheck(oldfd);
909 if (unlikely(err < 0)) {
914 return do_dup2(files, file, newfd, flags);
919 spin_unlock(&files->file_lock);
923 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
925 if (unlikely(newfd == oldfd)) { /* corner case */
926 struct files_struct *files = current->files;
930 if (!fcheck_files(files, oldfd))
935 return sys_dup3(oldfd, newfd, 0);
938 SYSCALL_DEFINE1(dup, unsigned int, fildes)
941 struct file *file = fget_raw(fildes);
944 ret = get_unused_fd_flags(0);
946 fd_install(ret, file);
953 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
956 if (from >= rlimit(RLIMIT_NOFILE))
958 err = alloc_fd(from, flags);
961 fd_install(err, file);
966 int iterate_fd(struct files_struct *files, unsigned n,
967 int (*f)(const void *, struct file *, unsigned),
974 spin_lock(&files->file_lock);
975 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
977 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
984 spin_unlock(&files->file_lock);
987 EXPORT_SYMBOL(iterate_fd);