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 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 int count_open_files(struct fdtable *fdt)
261 int size = fdt->max_fds;
264 /* Find the last open fd */
265 for (i = size / BITS_PER_LONG; i > 0; ) {
266 if (fdt->open_fds[--i])
269 i = (i + 1) * BITS_PER_LONG;
274 * Allocate a new files structure and copy contents from the
275 * passed in files structure.
276 * errorp will be valid only when the returned files_struct is NULL.
278 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
280 struct files_struct *newf;
281 struct file **old_fds, **new_fds;
282 int open_files, size, i;
283 struct fdtable *old_fdt, *new_fdt;
286 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
290 atomic_set(&newf->count, 1);
292 spin_lock_init(&newf->file_lock);
294 new_fdt = &newf->fdtab;
295 new_fdt->max_fds = NR_OPEN_DEFAULT;
296 new_fdt->close_on_exec = newf->close_on_exec_init;
297 new_fdt->open_fds = newf->open_fds_init;
298 new_fdt->fd = &newf->fd_array[0];
299 new_fdt->next = NULL;
301 spin_lock(&oldf->file_lock);
302 old_fdt = files_fdtable(oldf);
303 open_files = count_open_files(old_fdt);
306 * Check whether we need to allocate a larger fd array and fd set.
308 while (unlikely(open_files > new_fdt->max_fds)) {
309 spin_unlock(&oldf->file_lock);
311 if (new_fdt != &newf->fdtab)
312 __free_fdtable(new_fdt);
314 new_fdt = alloc_fdtable(open_files - 1);
320 /* beyond sysctl_nr_open; nothing to do */
321 if (unlikely(new_fdt->max_fds < open_files)) {
322 __free_fdtable(new_fdt);
328 * Reacquire the oldf lock and a pointer to its fd table
329 * who knows it may have a new bigger fd table. We need
330 * the latest pointer.
332 spin_lock(&oldf->file_lock);
333 old_fdt = files_fdtable(oldf);
334 open_files = count_open_files(old_fdt);
337 old_fds = old_fdt->fd;
338 new_fds = new_fdt->fd;
340 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
341 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
343 for (i = open_files; i != 0; i--) {
344 struct file *f = *old_fds++;
349 * The fd may be claimed in the fd bitmap but not yet
350 * instantiated in the files array if a sibling thread
351 * is partway through open(). So make sure that this
352 * fd is available to the new process.
354 __clear_open_fd(open_files - i, new_fdt);
356 rcu_assign_pointer(*new_fds++, f);
358 spin_unlock(&oldf->file_lock);
360 /* compute the remainder to be cleared */
361 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
363 /* This is long word aligned thus could use a optimized version */
364 memset(new_fds, 0, size);
366 if (new_fdt->max_fds > open_files) {
367 int left = (new_fdt->max_fds - open_files) / 8;
368 int start = open_files / BITS_PER_LONG;
370 memset(&new_fdt->open_fds[start], 0, left);
371 memset(&new_fdt->close_on_exec[start], 0, left);
374 rcu_assign_pointer(newf->fdt, new_fdt);
379 kmem_cache_free(files_cachep, newf);
384 static void close_files(struct files_struct * files)
392 * It is safe to dereference the fd table without RCU or
393 * ->file_lock because this is the last reference to the
394 * files structure. But use RCU to shut RCU-lockdep up.
397 fdt = files_fdtable(files);
401 i = j * BITS_PER_LONG;
402 if (i >= fdt->max_fds)
404 set = fdt->open_fds[j++];
407 struct file * file = xchg(&fdt->fd[i], NULL);
409 filp_close(file, files);
419 struct files_struct *get_files_struct(struct task_struct *task)
421 struct files_struct *files;
426 atomic_inc(&files->count);
432 void put_files_struct(struct files_struct *files)
436 if (atomic_dec_and_test(&files->count)) {
438 /* not really needed, since nobody can see us */
440 fdt = files_fdtable(files);
442 /* free the arrays if they are not embedded */
443 if (fdt != &files->fdtab)
445 kmem_cache_free(files_cachep, files);
449 void reset_files_struct(struct files_struct *files)
451 struct task_struct *tsk = current;
452 struct files_struct *old;
458 put_files_struct(old);
461 void exit_files(struct task_struct *tsk)
463 struct files_struct * files = tsk->files;
469 put_files_struct(files);
473 static void __devinit fdtable_defer_list_init(int cpu)
475 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
476 spin_lock_init(&fddef->lock);
477 INIT_WORK(&fddef->wq, free_fdtable_work);
481 void __init files_defer_init(void)
484 for_each_possible_cpu(i)
485 fdtable_defer_list_init(i);
486 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
490 struct files_struct init_files = {
491 .count = ATOMIC_INIT(1),
492 .fdt = &init_files.fdtab,
494 .max_fds = NR_OPEN_DEFAULT,
495 .fd = &init_files.fd_array[0],
496 .close_on_exec = init_files.close_on_exec_init,
497 .open_fds = init_files.open_fds_init,
499 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
503 * allocate a file descriptor, mark it busy.
505 int __alloc_fd(struct files_struct *files,
506 unsigned start, unsigned end, unsigned flags)
512 spin_lock(&files->file_lock);
514 fdt = files_fdtable(files);
516 if (fd < files->next_fd)
519 if (fd < fdt->max_fds)
520 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
523 * N.B. For clone tasks sharing a files structure, this test
524 * will limit the total number of files that can be opened.
530 error = expand_files(files, fd);
535 * If we needed to expand the fs array we
536 * might have blocked - try again.
541 if (start <= files->next_fd)
542 files->next_fd = fd + 1;
544 __set_open_fd(fd, fdt);
545 if (flags & O_CLOEXEC)
546 __set_close_on_exec(fd, fdt);
548 __clear_close_on_exec(fd, fdt);
552 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
553 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
554 rcu_assign_pointer(fdt->fd[fd], NULL);
559 spin_unlock(&files->file_lock);
563 int alloc_fd(unsigned start, unsigned flags)
565 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
568 int get_unused_fd_flags(unsigned flags)
570 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
572 EXPORT_SYMBOL(get_unused_fd_flags);
574 static void __put_unused_fd(struct files_struct *files, unsigned int fd)
576 struct fdtable *fdt = files_fdtable(files);
577 __clear_open_fd(fd, fdt);
578 if (fd < files->next_fd)
582 void put_unused_fd(unsigned int fd)
584 struct files_struct *files = current->files;
585 spin_lock(&files->file_lock);
586 __put_unused_fd(files, fd);
587 spin_unlock(&files->file_lock);
590 EXPORT_SYMBOL(put_unused_fd);
593 * Install a file pointer in the fd array.
595 * The VFS is full of places where we drop the files lock between
596 * setting the open_fds bitmap and installing the file in the file
597 * array. At any such point, we are vulnerable to a dup2() race
598 * installing a file in the array before us. We need to detect this and
599 * fput() the struct file we are about to overwrite in this case.
601 * It should never happen - if we allow dup2() do it, _really_ bad things
604 * NOTE: __fd_install() variant is really, really low-level; don't
605 * use it unless you are forced to by truly lousy API shoved down
606 * your throat. 'files' *MUST* be either current->files or obtained
607 * by get_files_struct(current) done by whoever had given it to you,
608 * or really bad things will happen. Normally you want to use
609 * fd_install() instead.
612 void __fd_install(struct files_struct *files, unsigned int fd,
616 spin_lock(&files->file_lock);
617 fdt = files_fdtable(files);
618 BUG_ON(fdt->fd[fd] != NULL);
619 rcu_assign_pointer(fdt->fd[fd], file);
620 spin_unlock(&files->file_lock);
623 void fd_install(unsigned int fd, struct file *file)
625 __fd_install(current->files, fd, file);
628 EXPORT_SYMBOL(fd_install);
631 * The same warnings as for __alloc_fd()/__fd_install() apply here...
633 int __close_fd(struct files_struct *files, unsigned fd)
638 spin_lock(&files->file_lock);
639 fdt = files_fdtable(files);
640 if (fd >= fdt->max_fds)
645 rcu_assign_pointer(fdt->fd[fd], NULL);
646 __clear_close_on_exec(fd, fdt);
647 __put_unused_fd(files, fd);
648 spin_unlock(&files->file_lock);
649 return filp_close(file, files);
652 spin_unlock(&files->file_lock);
656 void do_close_on_exec(struct files_struct *files)
661 /* exec unshares first */
662 BUG_ON(atomic_read(&files->count) != 1);
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 struct file *fget(unsigned int fd)
696 struct files_struct *files = current->files;
699 file = fcheck_files(files, fd);
701 /* File object ref couldn't be taken */
702 if (file->f_mode & FMODE_PATH ||
703 !atomic_long_inc_not_zero(&file->f_count))
713 struct file *fget_raw(unsigned int fd)
716 struct files_struct *files = current->files;
719 file = fcheck_files(files, fd);
721 /* File object ref couldn't be taken */
722 if (!atomic_long_inc_not_zero(&file->f_count))
730 EXPORT_SYMBOL(fget_raw);
733 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
735 * You can use this instead of fget if you satisfy all of the following
737 * 1) You must call fput_light before exiting the syscall and returning control
738 * to userspace (i.e. you cannot remember the returned struct file * after
739 * returning to userspace).
740 * 2) You must not call filp_close on the returned struct file * in between
741 * calls to fget_light and fput_light.
742 * 3) You must not clone the current task in between the calls to fget_light
745 * The fput_needed flag returned by fget_light should be passed to the
746 * corresponding fput_light.
748 struct file *fget_light(unsigned int fd, int *fput_needed)
751 struct files_struct *files = current->files;
754 if (atomic_read(&files->count) == 1) {
755 file = fcheck_files(files, fd);
756 if (file && (file->f_mode & FMODE_PATH))
760 file = fcheck_files(files, fd);
762 if (!(file->f_mode & FMODE_PATH) &&
763 atomic_long_inc_not_zero(&file->f_count))
766 /* Didn't get the reference, someone's freed */
775 struct file *fget_raw_light(unsigned int fd, int *fput_needed)
778 struct files_struct *files = current->files;
781 if (atomic_read(&files->count) == 1) {
782 file = fcheck_files(files, fd);
785 file = fcheck_files(files, fd);
787 if (atomic_long_inc_not_zero(&file->f_count))
790 /* Didn't get the reference, someone's freed */
799 void set_close_on_exec(unsigned int fd, int flag)
801 struct files_struct *files = current->files;
803 spin_lock(&files->file_lock);
804 fdt = files_fdtable(files);
806 __set_close_on_exec(fd, fdt);
808 __clear_close_on_exec(fd, fdt);
809 spin_unlock(&files->file_lock);
812 bool get_close_on_exec(unsigned int fd)
814 struct files_struct *files = current->files;
818 fdt = files_fdtable(files);
819 res = close_on_exec(fd, fdt);
824 static int do_dup2(struct files_struct *files,
825 struct file *file, unsigned fd, unsigned flags)
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 (newfd >= rlimit(RLIMIT_NOFILE))
901 spin_lock(&files->file_lock);
902 err = expand_files(files, newfd);
903 file = fcheck(oldfd);
906 if (unlikely(err < 0)) {
911 return do_dup2(files, file, newfd, flags);
916 spin_unlock(&files->file_lock);
920 SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
922 if (unlikely(newfd == oldfd)) { /* corner case */
923 struct files_struct *files = current->files;
927 if (!fcheck_files(files, oldfd))
932 return sys_dup3(oldfd, newfd, 0);
935 SYSCALL_DEFINE1(dup, unsigned int, fildes)
938 struct file *file = fget_raw(fildes);
941 ret = get_unused_fd();
943 fd_install(ret, file);
950 int f_dupfd(unsigned int from, struct file *file, unsigned flags)
953 if (from >= rlimit(RLIMIT_NOFILE))
955 err = alloc_fd(from, flags);
958 fd_install(err, file);