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 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
30 static void *alloc_fdmem(size_t size)
33 * Very large allocations can stress page reclaim, so fall back to
34 * vmalloc() if the allocation size will be considered "large" by the VM.
36 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
37 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN);
44 static void free_fdmem(void *ptr)
46 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
49 static void __free_fdtable(struct fdtable *fdt)
52 free_fdmem(fdt->open_fds);
56 static void free_fdtable_rcu(struct rcu_head *rcu)
58 __free_fdtable(container_of(rcu, struct fdtable, rcu));
62 * Expand the fdset in the files_struct. Called with the files spinlock
65 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
67 unsigned int cpy, set;
69 BUG_ON(nfdt->max_fds < ofdt->max_fds);
71 cpy = ofdt->max_fds * sizeof(struct file *);
72 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
73 memcpy(nfdt->fd, ofdt->fd, cpy);
74 memset((char *)(nfdt->fd) + cpy, 0, set);
76 cpy = ofdt->max_fds / BITS_PER_BYTE;
77 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
78 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
79 memset((char *)(nfdt->open_fds) + cpy, 0, set);
80 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
81 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
84 static struct fdtable * alloc_fdtable(unsigned int nr)
90 * Figure out how many fds we actually want to support in this fdtable.
91 * Allocation steps are keyed to the size of the fdarray, since it
92 * grows far faster than any of the other dynamic data. We try to fit
93 * the fdarray into comfortable page-tuned chunks: starting at 1024B
94 * and growing in powers of two from there on.
96 nr /= (1024 / sizeof(struct file *));
97 nr = roundup_pow_of_two(nr + 1);
98 nr *= (1024 / sizeof(struct file *));
100 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
101 * had been set lower between the check in expand_files() and here. Deal
102 * with that in caller, it's cheaper that way.
104 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
105 * bitmaps handling below becomes unpleasant, to put it mildly...
107 if (unlikely(nr > sysctl_nr_open))
108 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
110 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
114 data = alloc_fdmem(nr * sizeof(struct file *));
119 data = alloc_fdmem(max_t(size_t,
120 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
123 fdt->open_fds = data;
124 data += nr / BITS_PER_BYTE;
125 fdt->close_on_exec = data;
138 * Expand the file descriptor table.
139 * This function will allocate a new fdtable and both fd array and fdset, of
141 * Return <0 error code on error; 1 on successful completion.
142 * The files->file_lock should be held on entry, and will be held on exit.
144 static int expand_fdtable(struct files_struct *files, int nr)
145 __releases(files->file_lock)
146 __acquires(files->file_lock)
148 struct fdtable *new_fdt, *cur_fdt;
150 spin_unlock(&files->file_lock);
151 new_fdt = alloc_fdtable(nr);
152 spin_lock(&files->file_lock);
156 * extremely unlikely race - sysctl_nr_open decreased between the check in
157 * caller and alloc_fdtable(). Cheaper to catch it here...
159 if (unlikely(new_fdt->max_fds <= nr)) {
160 __free_fdtable(new_fdt);
164 * Check again since another task may have expanded the fd table while
165 * we dropped the lock
167 cur_fdt = files_fdtable(files);
168 if (nr >= cur_fdt->max_fds) {
169 /* Continue as planned */
170 copy_fdtable(new_fdt, cur_fdt);
171 rcu_assign_pointer(files->fdt, new_fdt);
172 if (cur_fdt != &files->fdtab)
173 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
175 /* Somebody else expanded, so undo our attempt */
176 __free_fdtable(new_fdt);
183 * This function will expand the file structures, if the requested size exceeds
184 * the current capacity and there is room for expansion.
185 * Return <0 error code on error; 0 when nothing done; 1 when files were
186 * expanded and execution may have blocked.
187 * The files->file_lock should be held on entry, and will be held on exit.
189 static int expand_files(struct files_struct *files, int nr)
193 fdt = files_fdtable(files);
195 /* Do we need to expand? */
196 if (nr < fdt->max_fds)
200 if (nr >= sysctl_nr_open)
203 /* All good, so we try */
204 return expand_fdtable(files, nr);
207 static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
209 __set_bit(fd, fdt->close_on_exec);
212 static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
214 __clear_bit(fd, fdt->close_on_exec);
217 static inline void __set_open_fd(int fd, struct fdtable *fdt)
219 __set_bit(fd, fdt->open_fds);
222 static inline void __clear_open_fd(int fd, struct fdtable *fdt)
224 __clear_bit(fd, fdt->open_fds);
227 static int count_open_files(struct fdtable *fdt)
229 int size = fdt->max_fds;
232 /* Find the last open fd */
233 for (i = size / BITS_PER_LONG; i > 0; ) {
234 if (fdt->open_fds[--i])
237 i = (i + 1) * BITS_PER_LONG;
242 * Allocate a new files structure and copy contents from the
243 * passed in files structure.
244 * errorp will be valid only when the returned files_struct is NULL.
246 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
248 struct files_struct *newf;
249 struct file **old_fds, **new_fds;
250 int open_files, size, i;
251 struct fdtable *old_fdt, *new_fdt;
254 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
258 atomic_set(&newf->count, 1);
260 spin_lock_init(&newf->file_lock);
262 new_fdt = &newf->fdtab;
263 new_fdt->max_fds = NR_OPEN_DEFAULT;
264 new_fdt->close_on_exec = newf->close_on_exec_init;
265 new_fdt->open_fds = newf->open_fds_init;
266 new_fdt->fd = &newf->fd_array[0];
268 spin_lock(&oldf->file_lock);
269 old_fdt = files_fdtable(oldf);
270 open_files = count_open_files(old_fdt);
273 * Check whether we need to allocate a larger fd array and fd set.
275 while (unlikely(open_files > new_fdt->max_fds)) {
276 spin_unlock(&oldf->file_lock);
278 if (new_fdt != &newf->fdtab)
279 __free_fdtable(new_fdt);
281 new_fdt = alloc_fdtable(open_files - 1);
287 /* beyond sysctl_nr_open; nothing to do */
288 if (unlikely(new_fdt->max_fds < open_files)) {
289 __free_fdtable(new_fdt);
295 * Reacquire the oldf lock and a pointer to its fd table
296 * who knows it may have a new bigger fd table. We need
297 * the latest pointer.
299 spin_lock(&oldf->file_lock);
300 old_fdt = files_fdtable(oldf);
301 open_files = count_open_files(old_fdt);
304 old_fds = old_fdt->fd;
305 new_fds = new_fdt->fd;
307 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
308 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
310 for (i = open_files; i != 0; i--) {
311 struct file *f = *old_fds++;
316 * The fd may be claimed in the fd bitmap but not yet
317 * instantiated in the files array if a sibling thread
318 * is partway through open(). So make sure that this
319 * fd is available to the new process.
321 __clear_open_fd(open_files - i, new_fdt);
323 rcu_assign_pointer(*new_fds++, f);
325 spin_unlock(&oldf->file_lock);
327 /* compute the remainder to be cleared */
328 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
330 /* This is long word aligned thus could use a optimized version */
331 memset(new_fds, 0, size);
333 if (new_fdt->max_fds > open_files) {
334 int left = (new_fdt->max_fds - open_files) / 8;
335 int start = open_files / BITS_PER_LONG;
337 memset(&new_fdt->open_fds[start], 0, left);
338 memset(&new_fdt->close_on_exec[start], 0, left);
341 rcu_assign_pointer(newf->fdt, new_fdt);
346 kmem_cache_free(files_cachep, newf);
351 static struct fdtable *close_files(struct files_struct * files)
354 * It is safe to dereference the fd table without RCU or
355 * ->file_lock because this is the last reference to the
358 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
363 i = j * BITS_PER_LONG;
364 if (i >= fdt->max_fds)
366 set = fdt->open_fds[j++];
369 struct file * file = xchg(&fdt->fd[i], NULL);
371 filp_close(file, files);
383 struct files_struct *get_files_struct(struct task_struct *task)
385 struct files_struct *files;
390 atomic_inc(&files->count);
396 void put_files_struct(struct files_struct *files)
398 if (atomic_dec_and_test(&files->count)) {
399 struct fdtable *fdt = close_files(files);
401 /* free the arrays if they are not embedded */
402 if (fdt != &files->fdtab)
404 kmem_cache_free(files_cachep, files);
408 void reset_files_struct(struct files_struct *files)
410 struct task_struct *tsk = current;
411 struct files_struct *old;
417 put_files_struct(old);
420 void exit_files(struct task_struct *tsk)
422 struct files_struct * files = tsk->files;
428 put_files_struct(files);
432 void __init files_defer_init(void)
434 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
438 struct files_struct init_files = {
439 .count = ATOMIC_INIT(1),
440 .fdt = &init_files.fdtab,
442 .max_fds = NR_OPEN_DEFAULT,
443 .fd = &init_files.fd_array[0],
444 .close_on_exec = init_files.close_on_exec_init,
445 .open_fds = init_files.open_fds_init,
447 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
451 * allocate a file descriptor, mark it busy.
453 int __alloc_fd(struct files_struct *files,
454 unsigned start, unsigned end, unsigned flags)
460 spin_lock(&files->file_lock);
462 fdt = files_fdtable(files);
464 if (fd < files->next_fd)
467 if (fd < fdt->max_fds)
468 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
471 * N.B. For clone tasks sharing a files structure, this test
472 * will limit the total number of files that can be opened.
478 error = expand_files(files, fd);
483 * If we needed to expand the fs array we
484 * might have blocked - try again.
489 if (start <= files->next_fd)
490 files->next_fd = fd + 1;
492 __set_open_fd(fd, fdt);
493 if (flags & O_CLOEXEC)
494 __set_close_on_exec(fd, fdt);
496 __clear_close_on_exec(fd, fdt);
500 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
501 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
502 rcu_assign_pointer(fdt->fd[fd], NULL);
507 spin_unlock(&files->file_lock);
511 static int alloc_fd(unsigned start, unsigned flags)
513 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
516 int get_unused_fd_flags(unsigned flags)
518 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
520 EXPORT_SYMBOL(get_unused_fd_flags);
522 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
524 struct fdtable *fdt = files_fdtable(files);
525 __clear_open_fd(fd, fdt);
526 if (fd < files->next_fd)
530 void put_unused_fd(unsigned int fd)
532 struct files_struct *files = current->files;
533 spin_lock(&files->file_lock);
534 __put_unused_fd(files, fd);
535 spin_unlock(&files->file_lock);
538 EXPORT_SYMBOL(put_unused_fd);
541 * Install a file pointer in the fd array.
543 * The VFS is full of places where we drop the files lock between
544 * setting the open_fds bitmap and installing the file in the file
545 * array. At any such point, we are vulnerable to a dup2() race
546 * installing a file in the array before us. We need to detect this and
547 * fput() the struct file we are about to overwrite in this case.
549 * It should never happen - if we allow dup2() do it, _really_ bad things
552 * NOTE: __fd_install() variant is really, really low-level; don't
553 * use it unless you are forced to by truly lousy API shoved down
554 * your throat. 'files' *MUST* be either current->files or obtained
555 * by get_files_struct(current) done by whoever had given it to you,
556 * or really bad things will happen. Normally you want to use
557 * fd_install() instead.
560 void __fd_install(struct files_struct *files, unsigned int fd,
564 spin_lock(&files->file_lock);
565 fdt = files_fdtable(files);
566 BUG_ON(fdt->fd[fd] != NULL);
567 rcu_assign_pointer(fdt->fd[fd], file);
568 spin_unlock(&files->file_lock);
571 void fd_install(unsigned int fd, struct file *file)
573 __fd_install(current->files, fd, file);
576 EXPORT_SYMBOL(fd_install);
579 * The same warnings as for __alloc_fd()/__fd_install() apply here...
581 int __close_fd(struct files_struct *files, unsigned fd)
586 spin_lock(&files->file_lock);
587 fdt = files_fdtable(files);
588 if (fd >= fdt->max_fds)
593 rcu_assign_pointer(fdt->fd[fd], NULL);
594 __clear_close_on_exec(fd, fdt);
595 __put_unused_fd(files, fd);
596 spin_unlock(&files->file_lock);
597 return filp_close(file, files);
600 spin_unlock(&files->file_lock);
604 void do_close_on_exec(struct files_struct *files)
609 /* exec unshares first */
610 spin_lock(&files->file_lock);
613 unsigned fd = i * BITS_PER_LONG;
614 fdt = files_fdtable(files);
615 if (fd >= fdt->max_fds)
617 set = fdt->close_on_exec[i];
620 fdt->close_on_exec[i] = 0;
621 for ( ; set ; fd++, set >>= 1) {
628 rcu_assign_pointer(fdt->fd[fd], NULL);
629 __put_unused_fd(files, fd);
630 spin_unlock(&files->file_lock);
631 filp_close(file, files);
633 spin_lock(&files->file_lock);
637 spin_unlock(&files->file_lock);
640 static struct file *__fget(unsigned int fd, fmode_t mask)
642 struct files_struct *files = current->files;
646 file = fcheck_files(files, fd);
648 /* File object ref couldn't be taken */
649 if ((file->f_mode & mask) ||
650 !atomic_long_inc_not_zero(&file->f_count))
658 struct file *fget(unsigned int fd)
660 return __fget(fd, FMODE_PATH);
664 struct file *fget_raw(unsigned int fd)
666 return __fget(fd, 0);
668 EXPORT_SYMBOL(fget_raw);
671 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
673 * You can use this instead of fget if you satisfy all of the following
675 * 1) You must call fput_light before exiting the syscall and returning control
676 * to userspace (i.e. you cannot remember the returned struct file * after
677 * returning to userspace).
678 * 2) You must not call filp_close on the returned struct file * in between
679 * calls to fget_light and fput_light.
680 * 3) You must not clone the current task in between the calls to fget_light
683 * The fput_needed flag returned by fget_light should be passed to the
684 * corresponding fput_light.
686 struct file *__fget_light(unsigned int fd, fmode_t mask, int *fput_needed)
688 struct files_struct *files = current->files;
692 if (atomic_read(&files->count) == 1) {
693 file = __fcheck_files(files, fd);
694 if (file && (file->f_mode & mask))
697 file = __fget(fd, mask);
704 struct file *fget_light(unsigned int fd, int *fput_needed)
706 return __fget_light(fd, FMODE_PATH, fput_needed);
708 EXPORT_SYMBOL(fget_light);
710 struct file *fget_raw_light(unsigned int fd, int *fput_needed)
712 return __fget_light(fd, 0, fput_needed);
715 void set_close_on_exec(unsigned int fd, int flag)
717 struct files_struct *files = current->files;
719 spin_lock(&files->file_lock);
720 fdt = files_fdtable(files);
722 __set_close_on_exec(fd, fdt);
724 __clear_close_on_exec(fd, fdt);
725 spin_unlock(&files->file_lock);
728 bool get_close_on_exec(unsigned int fd)
730 struct files_struct *files = current->files;
734 fdt = files_fdtable(files);
735 res = close_on_exec(fd, fdt);
740 static int do_dup2(struct files_struct *files,
741 struct file *file, unsigned fd, unsigned flags)
747 * We need to detect attempts to do dup2() over allocated but still
748 * not finished descriptor. NB: OpenBSD avoids that at the price of
749 * extra work in their equivalent of fget() - they insert struct
750 * file immediately after grabbing descriptor, mark it larval if
751 * more work (e.g. actual opening) is needed and make sure that
752 * fget() treats larval files as absent. Potentially interesting,
753 * but while extra work in fget() is trivial, locking implications
754 * and amount of surgery on open()-related paths in VFS are not.
755 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
756 * deadlocks in rather amusing ways, AFAICS. All of that is out of
757 * scope of POSIX or SUS, since neither considers shared descriptor
758 * tables and this condition does not arise without those.
760 fdt = files_fdtable(files);
761 tofree = fdt->fd[fd];
762 if (!tofree && fd_is_open(fd, fdt))
765 rcu_assign_pointer(fdt->fd[fd], file);
766 __set_open_fd(fd, fdt);
767 if (flags & O_CLOEXEC)
768 __set_close_on_exec(fd, fdt);
770 __clear_close_on_exec(fd, fdt);
771 spin_unlock(&files->file_lock);
774 filp_close(tofree, files);
779 spin_unlock(&files->file_lock);
783 int replace_fd(unsigned fd, struct file *file, unsigned flags)
786 struct files_struct *files = current->files;
789 return __close_fd(files, fd);
791 if (fd >= rlimit(RLIMIT_NOFILE))
794 spin_lock(&files->file_lock);
795 err = expand_files(files, fd);
796 if (unlikely(err < 0))
798 return do_dup2(files, file, fd, flags);
801 spin_unlock(&files->file_lock);
805 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
809 struct files_struct *files = current->files;
811 if ((flags & ~O_CLOEXEC) != 0)
814 if (unlikely(oldfd == newfd))
817 if (newfd >= rlimit(RLIMIT_NOFILE))
820 spin_lock(&files->file_lock);
821 err = expand_files(files, newfd);
822 file = fcheck(oldfd);
825 if (unlikely(err < 0)) {
830 return do_dup2(files, file, newfd, flags);
835 spin_unlock(&files->file_lock);
839 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
841 if (unlikely(newfd == oldfd)) { /* corner case */
842 struct files_struct *files = current->files;
846 if (!fcheck_files(files, oldfd))
851 return sys_dup3(oldfd, newfd, 0);
854 SYSCALL_DEFINE1(dup, unsigned int, fildes)
857 struct file *file = fget_raw(fildes);
860 ret = get_unused_fd();
862 fd_install(ret, file);
869 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
872 if (from >= rlimit(RLIMIT_NOFILE))
874 err = alloc_fd(from, flags);
877 fd_install(err, file);
882 int iterate_fd(struct files_struct *files, unsigned n,
883 int (*f)(const void *, struct file *, unsigned),
890 spin_lock(&files->file_lock);
891 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
893 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
900 spin_unlock(&files->file_lock);
903 EXPORT_SYMBOL(iterate_fd);