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 struct fdtable_defer {
28 struct work_struct wq;
32 int sysctl_nr_open __read_mostly = 1024*1024;
33 int sysctl_nr_open_min = BITS_PER_LONG;
34 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
37 * We use this list to defer free fdtables that have vmalloced
38 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
39 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
40 * this per-task structure.
42 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
44 static void *alloc_fdmem(size_t size)
47 * Very large allocations can stress page reclaim, so fall back to
48 * vmalloc() if the allocation size will be considered "large" by the VM.
50 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
51 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN);
58 static void free_fdmem(void *ptr)
60 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
63 static void __free_fdtable(struct fdtable *fdt)
66 free_fdmem(fdt->open_fds);
70 static void free_fdtable_work(struct work_struct *work)
72 struct fdtable_defer *f =
73 container_of(work, struct fdtable_defer, wq);
76 spin_lock_bh(&f->lock);
79 spin_unlock_bh(&f->lock);
81 struct fdtable *next = fdt->next;
88 static void free_fdtable_rcu(struct rcu_head *rcu)
90 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
91 struct fdtable_defer *fddef;
94 BUG_ON(fdt->max_fds <= NR_OPEN_DEFAULT);
96 if (!is_vmalloc_addr(fdt->fd) && !is_vmalloc_addr(fdt->open_fds)) {
101 fddef = &get_cpu_var(fdtable_defer_list);
102 spin_lock(&fddef->lock);
103 fdt->next = fddef->next;
105 /* vmallocs are handled from the workqueue context */
106 schedule_work(&fddef->wq);
107 spin_unlock(&fddef->lock);
108 put_cpu_var(fdtable_defer_list);
113 * Expand the fdset in the files_struct. Called with the files spinlock
116 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
118 unsigned int cpy, set;
120 BUG_ON(nfdt->max_fds < ofdt->max_fds);
122 cpy = ofdt->max_fds * sizeof(struct file *);
123 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
124 memcpy(nfdt->fd, ofdt->fd, cpy);
125 memset((char *)(nfdt->fd) + cpy, 0, set);
127 cpy = ofdt->max_fds / BITS_PER_BYTE;
128 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
129 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
130 memset((char *)(nfdt->open_fds) + cpy, 0, set);
131 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
132 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
135 static struct fdtable * alloc_fdtable(unsigned int nr)
141 * Figure out how many fds we actually want to support in this fdtable.
142 * Allocation steps are keyed to the size of the fdarray, since it
143 * grows far faster than any of the other dynamic data. We try to fit
144 * the fdarray into comfortable page-tuned chunks: starting at 1024B
145 * and growing in powers of two from there on.
147 nr /= (1024 / sizeof(struct file *));
148 nr = roundup_pow_of_two(nr + 1);
149 nr *= (1024 / sizeof(struct file *));
151 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
152 * had been set lower between the check in expand_files() and here. Deal
153 * with that in caller, it's cheaper that way.
155 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
156 * bitmaps handling below becomes unpleasant, to put it mildly...
158 if (unlikely(nr > sysctl_nr_open))
159 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
161 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
165 data = alloc_fdmem(nr * sizeof(struct file *));
170 data = alloc_fdmem(max_t(size_t,
171 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
174 fdt->open_fds = data;
175 data += nr / BITS_PER_BYTE;
176 fdt->close_on_exec = data;
190 * Expand the file descriptor table.
191 * This function will allocate a new fdtable and both fd array and fdset, of
193 * Return <0 error code on error; 1 on successful completion.
194 * The files->file_lock should be held on entry, and will be held on exit.
196 static int expand_fdtable(struct files_struct *files, int nr)
197 __releases(files->file_lock)
198 __acquires(files->file_lock)
200 struct fdtable *new_fdt, *cur_fdt;
202 spin_unlock(&files->file_lock);
203 new_fdt = alloc_fdtable(nr);
204 spin_lock(&files->file_lock);
208 * extremely unlikely race - sysctl_nr_open decreased between the check in
209 * caller and alloc_fdtable(). Cheaper to catch it here...
211 if (unlikely(new_fdt->max_fds <= nr)) {
212 __free_fdtable(new_fdt);
216 * Check again since another task may have expanded the fd table while
217 * we dropped the lock
219 cur_fdt = files_fdtable(files);
220 if (nr >= cur_fdt->max_fds) {
221 /* Continue as planned */
222 copy_fdtable(new_fdt, cur_fdt);
223 rcu_assign_pointer(files->fdt, new_fdt);
224 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
225 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
227 /* Somebody else expanded, so undo our attempt */
228 __free_fdtable(new_fdt);
235 * This function will expand the file structures, if the requested size exceeds
236 * the current capacity and there is room for expansion.
237 * Return <0 error code on error; 0 when nothing done; 1 when files were
238 * expanded and execution may have blocked.
239 * The files->file_lock should be held on entry, and will be held on exit.
241 static int expand_files(struct files_struct *files, int nr)
245 fdt = files_fdtable(files);
247 /* Do we need to expand? */
248 if (nr < fdt->max_fds)
252 if (nr >= sysctl_nr_open)
255 /* All good, so we try */
256 return expand_fdtable(files, nr);
259 static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
261 __set_bit(fd, fdt->close_on_exec);
264 static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
266 __clear_bit(fd, fdt->close_on_exec);
269 static inline void __set_open_fd(int fd, struct fdtable *fdt)
271 __set_bit(fd, fdt->open_fds);
274 static inline void __clear_open_fd(int fd, struct fdtable *fdt)
276 __clear_bit(fd, fdt->open_fds);
279 static int count_open_files(struct fdtable *fdt)
281 int size = fdt->max_fds;
284 /* Find the last open fd */
285 for (i = size / BITS_PER_LONG; i > 0; ) {
286 if (fdt->open_fds[--i])
289 i = (i + 1) * BITS_PER_LONG;
294 * Allocate a new files structure and copy contents from the
295 * passed in files structure.
296 * errorp will be valid only when the returned files_struct is NULL.
298 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
300 struct files_struct *newf;
301 struct file **old_fds, **new_fds;
302 int open_files, size, i;
303 struct fdtable *old_fdt, *new_fdt;
306 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
310 atomic_set(&newf->count, 1);
312 spin_lock_init(&newf->file_lock);
314 new_fdt = &newf->fdtab;
315 new_fdt->max_fds = NR_OPEN_DEFAULT;
316 new_fdt->close_on_exec = newf->close_on_exec_init;
317 new_fdt->open_fds = newf->open_fds_init;
318 new_fdt->fd = &newf->fd_array[0];
319 new_fdt->next = NULL;
321 spin_lock(&oldf->file_lock);
322 old_fdt = files_fdtable(oldf);
323 open_files = count_open_files(old_fdt);
326 * Check whether we need to allocate a larger fd array and fd set.
328 while (unlikely(open_files > new_fdt->max_fds)) {
329 spin_unlock(&oldf->file_lock);
331 if (new_fdt != &newf->fdtab)
332 __free_fdtable(new_fdt);
334 new_fdt = alloc_fdtable(open_files - 1);
340 /* beyond sysctl_nr_open; nothing to do */
341 if (unlikely(new_fdt->max_fds < open_files)) {
342 __free_fdtable(new_fdt);
348 * Reacquire the oldf lock and a pointer to its fd table
349 * who knows it may have a new bigger fd table. We need
350 * the latest pointer.
352 spin_lock(&oldf->file_lock);
353 old_fdt = files_fdtable(oldf);
354 open_files = count_open_files(old_fdt);
357 old_fds = old_fdt->fd;
358 new_fds = new_fdt->fd;
360 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
361 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
363 for (i = open_files; i != 0; i--) {
364 struct file *f = *old_fds++;
369 * The fd may be claimed in the fd bitmap but not yet
370 * instantiated in the files array if a sibling thread
371 * is partway through open(). So make sure that this
372 * fd is available to the new process.
374 __clear_open_fd(open_files - i, new_fdt);
376 rcu_assign_pointer(*new_fds++, f);
378 spin_unlock(&oldf->file_lock);
380 /* compute the remainder to be cleared */
381 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
383 /* This is long word aligned thus could use a optimized version */
384 memset(new_fds, 0, size);
386 if (new_fdt->max_fds > open_files) {
387 int left = (new_fdt->max_fds - open_files) / 8;
388 int start = open_files / BITS_PER_LONG;
390 memset(&new_fdt->open_fds[start], 0, left);
391 memset(&new_fdt->close_on_exec[start], 0, left);
394 rcu_assign_pointer(newf->fdt, new_fdt);
399 kmem_cache_free(files_cachep, newf);
404 static void close_files(struct files_struct * files)
412 * It is safe to dereference the fd table without RCU or
413 * ->file_lock because this is the last reference to the
414 * files structure. But use RCU to shut RCU-lockdep up.
417 fdt = files_fdtable(files);
421 i = j * BITS_PER_LONG;
422 if (i >= fdt->max_fds)
424 set = fdt->open_fds[j++];
427 struct file * file = xchg(&fdt->fd[i], NULL);
429 filp_close(file, files);
439 struct files_struct *get_files_struct(struct task_struct *task)
441 struct files_struct *files;
446 atomic_inc(&files->count);
452 void put_files_struct(struct files_struct *files)
456 if (atomic_dec_and_test(&files->count)) {
458 /* not really needed, since nobody can see us */
460 fdt = files_fdtable(files);
462 /* free the arrays if they are not embedded */
463 if (fdt != &files->fdtab)
465 kmem_cache_free(files_cachep, files);
469 void reset_files_struct(struct files_struct *files)
471 struct task_struct *tsk = current;
472 struct files_struct *old;
478 put_files_struct(old);
481 void exit_files(struct task_struct *tsk)
483 struct files_struct * files = tsk->files;
489 put_files_struct(files);
493 static void __devinit fdtable_defer_list_init(int cpu)
495 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
496 spin_lock_init(&fddef->lock);
497 INIT_WORK(&fddef->wq, free_fdtable_work);
501 void __init files_defer_init(void)
504 for_each_possible_cpu(i)
505 fdtable_defer_list_init(i);
506 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
510 struct files_struct init_files = {
511 .count = ATOMIC_INIT(1),
512 .fdt = &init_files.fdtab,
514 .max_fds = NR_OPEN_DEFAULT,
515 .fd = &init_files.fd_array[0],
516 .close_on_exec = init_files.close_on_exec_init,
517 .open_fds = init_files.open_fds_init,
519 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
523 * allocate a file descriptor, mark it busy.
525 int __alloc_fd(struct files_struct *files,
526 unsigned start, unsigned end, unsigned flags)
532 spin_lock(&files->file_lock);
534 fdt = files_fdtable(files);
536 if (fd < files->next_fd)
539 if (fd < fdt->max_fds)
540 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
543 * N.B. For clone tasks sharing a files structure, this test
544 * will limit the total number of files that can be opened.
550 error = expand_files(files, fd);
555 * If we needed to expand the fs array we
556 * might have blocked - try again.
561 if (start <= files->next_fd)
562 files->next_fd = fd + 1;
564 __set_open_fd(fd, fdt);
565 if (flags & O_CLOEXEC)
566 __set_close_on_exec(fd, fdt);
568 __clear_close_on_exec(fd, fdt);
572 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
573 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
574 rcu_assign_pointer(fdt->fd[fd], NULL);
579 spin_unlock(&files->file_lock);
583 static int alloc_fd(unsigned start, unsigned flags)
585 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
588 int get_unused_fd_flags(unsigned flags)
590 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
592 EXPORT_SYMBOL(get_unused_fd_flags);
594 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
596 struct fdtable *fdt = files_fdtable(files);
597 __clear_open_fd(fd, fdt);
598 if (fd < files->next_fd)
602 void put_unused_fd(unsigned int fd)
604 struct files_struct *files = current->files;
605 spin_lock(&files->file_lock);
606 __put_unused_fd(files, fd);
607 spin_unlock(&files->file_lock);
610 EXPORT_SYMBOL(put_unused_fd);
613 * Install a file pointer in the fd array.
615 * The VFS is full of places where we drop the files lock between
616 * setting the open_fds bitmap and installing the file in the file
617 * array. At any such point, we are vulnerable to a dup2() race
618 * installing a file in the array before us. We need to detect this and
619 * fput() the struct file we are about to overwrite in this case.
621 * It should never happen - if we allow dup2() do it, _really_ bad things
624 * NOTE: __fd_install() variant is really, really low-level; don't
625 * use it unless you are forced to by truly lousy API shoved down
626 * your throat. 'files' *MUST* be either current->files or obtained
627 * by get_files_struct(current) done by whoever had given it to you,
628 * or really bad things will happen. Normally you want to use
629 * fd_install() instead.
632 void __fd_install(struct files_struct *files, unsigned int fd,
636 spin_lock(&files->file_lock);
637 fdt = files_fdtable(files);
638 BUG_ON(fdt->fd[fd] != NULL);
639 rcu_assign_pointer(fdt->fd[fd], file);
640 spin_unlock(&files->file_lock);
643 void fd_install(unsigned int fd, struct file *file)
645 __fd_install(current->files, fd, file);
648 EXPORT_SYMBOL(fd_install);
651 * The same warnings as for __alloc_fd()/__fd_install() apply here...
653 int __close_fd(struct files_struct *files, unsigned fd)
658 spin_lock(&files->file_lock);
659 fdt = files_fdtable(files);
660 if (fd >= fdt->max_fds)
665 rcu_assign_pointer(fdt->fd[fd], NULL);
666 __clear_close_on_exec(fd, fdt);
667 __put_unused_fd(files, fd);
668 spin_unlock(&files->file_lock);
669 return filp_close(file, files);
672 spin_unlock(&files->file_lock);
676 void do_close_on_exec(struct files_struct *files)
681 /* exec unshares first */
682 spin_lock(&files->file_lock);
685 unsigned fd = i * BITS_PER_LONG;
686 fdt = files_fdtable(files);
687 if (fd >= fdt->max_fds)
689 set = fdt->close_on_exec[i];
692 fdt->close_on_exec[i] = 0;
693 for ( ; set ; fd++, set >>= 1) {
700 rcu_assign_pointer(fdt->fd[fd], NULL);
701 __put_unused_fd(files, fd);
702 spin_unlock(&files->file_lock);
703 filp_close(file, files);
705 spin_lock(&files->file_lock);
709 spin_unlock(&files->file_lock);
712 struct file *fget(unsigned int fd)
715 struct files_struct *files = current->files;
718 file = fcheck_files(files, fd);
720 /* File object ref couldn't be taken */
721 if (file->f_mode & FMODE_PATH ||
722 !atomic_long_inc_not_zero(&file->f_count))
732 struct file *fget_raw(unsigned int fd)
735 struct files_struct *files = current->files;
738 file = fcheck_files(files, fd);
740 /* File object ref couldn't be taken */
741 if (!atomic_long_inc_not_zero(&file->f_count))
749 EXPORT_SYMBOL(fget_raw);
752 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
754 * You can use this instead of fget if you satisfy all of the following
756 * 1) You must call fput_light before exiting the syscall and returning control
757 * to userspace (i.e. you cannot remember the returned struct file * after
758 * returning to userspace).
759 * 2) You must not call filp_close on the returned struct file * in between
760 * calls to fget_light and fput_light.
761 * 3) You must not clone the current task in between the calls to fget_light
764 * The fput_needed flag returned by fget_light should be passed to the
765 * corresponding fput_light.
767 struct file *fget_light(unsigned int fd, int *fput_needed)
770 struct files_struct *files = current->files;
773 if (atomic_read(&files->count) == 1) {
774 file = fcheck_files(files, fd);
775 if (file && (file->f_mode & FMODE_PATH))
779 file = fcheck_files(files, fd);
781 if (!(file->f_mode & FMODE_PATH) &&
782 atomic_long_inc_not_zero(&file->f_count))
785 /* Didn't get the reference, someone's freed */
793 EXPORT_SYMBOL(fget_light);
795 struct file *fget_raw_light(unsigned int fd, int *fput_needed)
798 struct files_struct *files = current->files;
801 if (atomic_read(&files->count) == 1) {
802 file = fcheck_files(files, fd);
805 file = fcheck_files(files, fd);
807 if (atomic_long_inc_not_zero(&file->f_count))
810 /* Didn't get the reference, someone's freed */
819 void set_close_on_exec(unsigned int fd, int flag)
821 struct files_struct *files = current->files;
823 spin_lock(&files->file_lock);
824 fdt = files_fdtable(files);
826 __set_close_on_exec(fd, fdt);
828 __clear_close_on_exec(fd, fdt);
829 spin_unlock(&files->file_lock);
832 bool get_close_on_exec(unsigned int fd)
834 struct files_struct *files = current->files;
838 fdt = files_fdtable(files);
839 res = close_on_exec(fd, fdt);
844 static int do_dup2(struct files_struct *files,
845 struct file *file, unsigned fd, unsigned flags)
851 * We need to detect attempts to do dup2() over allocated but still
852 * not finished descriptor. NB: OpenBSD avoids that at the price of
853 * extra work in their equivalent of fget() - they insert struct
854 * file immediately after grabbing descriptor, mark it larval if
855 * more work (e.g. actual opening) is needed and make sure that
856 * fget() treats larval files as absent. Potentially interesting,
857 * but while extra work in fget() is trivial, locking implications
858 * and amount of surgery on open()-related paths in VFS are not.
859 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
860 * deadlocks in rather amusing ways, AFAICS. All of that is out of
861 * scope of POSIX or SUS, since neither considers shared descriptor
862 * tables and this condition does not arise without those.
864 fdt = files_fdtable(files);
865 tofree = fdt->fd[fd];
866 if (!tofree && fd_is_open(fd, fdt))
869 rcu_assign_pointer(fdt->fd[fd], file);
870 __set_open_fd(fd, fdt);
871 if (flags & O_CLOEXEC)
872 __set_close_on_exec(fd, fdt);
874 __clear_close_on_exec(fd, fdt);
875 spin_unlock(&files->file_lock);
878 filp_close(tofree, files);
883 spin_unlock(&files->file_lock);
887 int replace_fd(unsigned fd, struct file *file, unsigned flags)
890 struct files_struct *files = current->files;
893 return __close_fd(files, fd);
895 if (fd >= rlimit(RLIMIT_NOFILE))
898 spin_lock(&files->file_lock);
899 err = expand_files(files, fd);
900 if (unlikely(err < 0))
902 return do_dup2(files, file, fd, flags);
905 spin_unlock(&files->file_lock);
909 SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
913 struct files_struct *files = current->files;
915 if ((flags & ~O_CLOEXEC) != 0)
918 if (unlikely(oldfd == newfd))
921 if (newfd >= rlimit(RLIMIT_NOFILE))
924 spin_lock(&files->file_lock);
925 err = expand_files(files, newfd);
926 file = fcheck(oldfd);
929 if (unlikely(err < 0)) {
934 return do_dup2(files, file, newfd, flags);
939 spin_unlock(&files->file_lock);
943 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
945 if (unlikely(newfd == oldfd)) { /* corner case */
946 struct files_struct *files = current->files;
950 if (!fcheck_files(files, oldfd))
955 return sys_dup3(oldfd, newfd, 0);
958 SYSCALL_DEFINE1(dup, unsigned int, fildes)
961 struct file *file = fget_raw(fildes);
964 ret = get_unused_fd();
966 fd_install(ret, file);
973 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
976 if (from >= rlimit(RLIMIT_NOFILE))
978 err = alloc_fd(from, flags);
981 fd_install(err, file);
986 int iterate_fd(struct files_struct *files, unsigned n,
987 int (*f)(const void *, struct file *, unsigned),
994 spin_lock(&files->file_lock);
995 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
997 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1000 res = f(p, file, n);
1004 spin_unlock(&files->file_lock);
1007 EXPORT_SYMBOL(iterate_fd);