1/*
  2 *  linux/fs/file.c
  3 *
  4 *  Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
  5 *
  6 *  Manage the dynamic fd arrays in the process files_struct.
  7 */
  8
  9#include <linux/syscalls.h>
 10#include <linux/export.h>
 11#include <linux/fs.h>
 12#include <linux/mm.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>
 25
 
 
 
 
 
 
 26int sysctl_nr_open __read_mostly = 1024*1024;
 27int 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))
 30int sysctl_nr_open_max = __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) &
 31			 -BITS_PER_LONG;
 
 
 
 
 
 32
 33static void *alloc_fdmem(size_t size)
 34{
 35	/*
 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.
 38	 */
 39	if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
 40		void *data = kmalloc(size, GFP_KERNEL_ACCOUNT |
 41				     __GFP_NOWARN | __GFP_NORETRY);
 42		if (data != NULL)
 43			return data;
 44	}
 45	return __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM, PAGE_KERNEL);
 
 
 
 
 
 46}
 47
 48static void __free_fdtable(struct fdtable *fdt)
 49{
 50	kvfree(fdt->fd);
 51	kvfree(fdt->open_fds);
 52	kfree(fdt);
 53}
 54
 55static void free_fdtable_rcu(struct rcu_head *rcu)
 56{
 57	__free_fdtable(container_of(rcu, struct fdtable, rcu));
 58}
 
 59
 60#define BITBIT_NR(nr)	BITS_TO_LONGS(BITS_TO_LONGS(nr))
 61#define BITBIT_SIZE(nr)	(BITBIT_NR(nr) * sizeof(long))
 
 
 
 
 62
 63/*
 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.
 67 */
 68static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
 69			    unsigned int count)
 70{
 71	unsigned int cpy, set;
 
 72
 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);
 79
 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);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 84}
 85
 86/*
 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.
 89 */
 90static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 91{
 92	unsigned int cpy, set;
 93
 94	BUG_ON(nfdt->max_fds < ofdt->max_fds);
 95
 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);
100
101	copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
 
 
 
 
 
102}
103
104static struct fdtable * alloc_fdtable(unsigned int nr)
105{
106	struct fdtable *fdt;
107	void *data;
108
109	/*
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.
115	 */
116	nr /= (1024 / sizeof(struct file *));
117	nr = roundup_pow_of_two(nr + 1);
118	nr *= (1024 / sizeof(struct file *));
119	/*
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.
123	 *
124	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
125	 * bitmaps handling below becomes unpleasant, to put it mildly...
126	 */
127	if (unlikely(nr > sysctl_nr_open))
128		nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
129
130	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
131	if (!fdt)
132		goto out;
133	fdt->max_fds = nr;
134	data = alloc_fdmem(nr * sizeof(struct file *));
135	if (!data)
136		goto out_fdt;
137	fdt->fd = data;
138
139	data = alloc_fdmem(max_t(size_t,
140				 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES));
141	if (!data)
142		goto out_arr;
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;
148
149	return fdt;
150
151out_arr:
152	kvfree(fdt->fd);
153out_fdt:
154	kfree(fdt);
155out:
156	return NULL;
157}
158
159/*
160 * Expand the file descriptor table.
161 * This function will allocate a new fdtable and both fd array and fdset, of
162 * the given size.
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.
165 */
166static int expand_fdtable(struct files_struct *files, int nr)
167	__releases(files->file_lock)
168	__acquires(files->file_lock)
169{
170	struct fdtable *new_fdt, *cur_fdt;
171
172	spin_unlock(&files->file_lock);
173	new_fdt = alloc_fdtable(nr);
174
175	/* make sure all __fd_install() have seen resize_in_progress
176	 * or have finished their rcu_read_lock_sched() section.
177	 */
178	if (atomic_read(&files->count) > 1)
179		synchronize_sched();
180
181	spin_lock(&files->file_lock);
182	if (!new_fdt)
183		return -ENOMEM;
184	/*
185	 * extremely unlikely race - sysctl_nr_open decreased between the check in
186	 * caller and alloc_fdtable().  Cheaper to catch it here...
187	 */
188	if (unlikely(new_fdt->max_fds <= nr)) {
189		__free_fdtable(new_fdt);
190		return -EMFILE;
191	}
 
 
 
 
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() */
199	smp_wmb();
 
 
 
200	return 1;
201}
202
203/*
204 * Expand files.
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.
210 */
211static int expand_files(struct files_struct *files, int nr)
212	__releases(files->file_lock)
213	__acquires(files->file_lock)
214{
215	struct fdtable *fdt;
216	int expanded = 0;
217
218repeat:
219	fdt = files_fdtable(files);
220
 
 
 
 
 
 
 
221	/* Do we need to expand? */
222	if (nr < fdt->max_fds)
223		return expanded;
224
225	/* Can we expand? */
226	if (nr >= sysctl_nr_open)
227		return -EMFILE;
228
229	if (unlikely(files->resize_in_progress)) {
230		spin_unlock(&files->file_lock);
231		expanded = 1;
232		wait_event(files->resize_wait, !files->resize_in_progress);
233		spin_lock(&files->file_lock);
234		goto repeat;
235	}
236
237	/* All good, so we try */
238	files->resize_in_progress = true;
239	expanded = expand_fdtable(files, nr);
240	files->resize_in_progress = false;
241
242	wake_up_all(&files->resize_wait);
243	return expanded;
244}
245
246static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
247{
248	__set_bit(fd, fdt->close_on_exec);
249}
250
251static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
252{
253	if (test_bit(fd, fdt->close_on_exec))
254		__clear_bit(fd, fdt->close_on_exec);
255}
256
257static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
258{
259	__set_bit(fd, fdt->open_fds);
260	fd /= BITS_PER_LONG;
261	if (!~fdt->open_fds[fd])
262		__set_bit(fd, fdt->full_fds_bits);
263}
264
265static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
266{
267	__clear_bit(fd, fdt->open_fds);
268	__clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
269}
270
271static int count_open_files(struct fdtable *fdt)
272{
273	int size = fdt->max_fds;
274	int i;
275
276	/* Find the last open fd */
277	for (i = size / BITS_PER_LONG; i > 0; ) {
278		if (fdt->open_fds[--i])
279			break;
280	}
281	i = (i + 1) * BITS_PER_LONG;
282	return i;
283}
284
285/*
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.
289 */
290struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
291{
292	struct files_struct *newf;
293	struct file **old_fds, **new_fds;
294	int open_files, i;
295	struct fdtable *old_fdt, *new_fdt;
296
297	*errorp = -ENOMEM;
298	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
299	if (!newf)
300		goto out;
301
302	atomic_set(&newf->count, 1);
303
304	spin_lock_init(&newf->file_lock);
305	newf->resize_in_progress = false;
306	init_waitqueue_head(&newf->resize_wait);
307	newf->next_fd = 0;
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];
 
314
315	spin_lock(&oldf->file_lock);
316	old_fdt = files_fdtable(oldf);
317	open_files = count_open_files(old_fdt);
318
319	/*
320	 * Check whether we need to allocate a larger fd array and fd set.
321	 */
322	while (unlikely(open_files > new_fdt->max_fds)) {
323		spin_unlock(&oldf->file_lock);
324
325		if (new_fdt != &newf->fdtab)
326			__free_fdtable(new_fdt);
327
328		new_fdt = alloc_fdtable(open_files - 1);
329		if (!new_fdt) {
330			*errorp = -ENOMEM;
331			goto out_release;
332		}
333
334		/* beyond sysctl_nr_open; nothing to do */
335		if (unlikely(new_fdt->max_fds < open_files)) {
336			__free_fdtable(new_fdt);
337			*errorp = -EMFILE;
338			goto out_release;
339		}
340
341		/*
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.
345		 */
346		spin_lock(&oldf->file_lock);
347		old_fdt = files_fdtable(oldf);
348		open_files = count_open_files(old_fdt);
349	}
350
351	copy_fd_bitmaps(new_fdt, old_fdt, open_files);
352
353	old_fds = old_fdt->fd;
354	new_fds = new_fdt->fd;
355
 
 
 
 
 
356	for (i = open_files; i != 0; i--) {
357		struct file *f = *old_fds++;
358		if (f) {
359			get_file(f);
360		} else {
361			/*
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.
366			 */
367			__clear_open_fd(open_files - i, new_fdt);
368		}
369		rcu_assign_pointer(*new_fds++, f);
370	}
371	spin_unlock(&oldf->file_lock);
372
373	/* clear the remainder */
374	memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
 
 
 
 
 
 
 
 
 
 
 
375
376	rcu_assign_pointer(newf->fdt, new_fdt);
377
378	return newf;
379
380out_release:
381	kmem_cache_free(files_cachep, newf);
382out:
383	return NULL;
384}
385
386static struct fdtable *close_files(struct files_struct * files)
387{
388	/*
389	 * It is safe to dereference the fd table without RCU or
390	 * ->file_lock because this is the last reference to the
391	 * files structure.
392	 */
393	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
394	int i, j = 0;
395
396	for (;;) {
397		unsigned long set;
398		i = j * BITS_PER_LONG;
399		if (i >= fdt->max_fds)
400			break;
401		set = fdt->open_fds[j++];
402		while (set) {
403			if (set & 1) {
404				struct file * file = xchg(&fdt->fd[i], NULL);
405				if (file) {
406					filp_close(file, files);
407					cond_resched_rcu_qs();
408				}
409			}
410			i++;
411			set >>= 1;
412		}
413	}
414
415	return fdt;
416}
417
418struct files_struct *get_files_struct(struct task_struct *task)
419{
420	struct files_struct *files;
421
422	task_lock(task);
423	files = task->files;
424	if (files)
425		atomic_inc(&files->count);
426	task_unlock(task);
427
428	return files;
429}
430
431void put_files_struct(struct files_struct *files)
432{
433	if (atomic_dec_and_test(&files->count)) {
434		struct fdtable *fdt = close_files(files);
435
436		/* free the arrays if they are not embedded */
437		if (fdt != &files->fdtab)
438			__free_fdtable(fdt);
439		kmem_cache_free(files_cachep, files);
440	}
441}
442
443void reset_files_struct(struct files_struct *files)
444{
445	struct task_struct *tsk = current;
446	struct files_struct *old;
447
448	old = tsk->files;
449	task_lock(tsk);
450	tsk->files = files;
451	task_unlock(tsk);
452	put_files_struct(old);
453}
454
455void exit_files(struct task_struct *tsk)
456{
457	struct files_struct * files = tsk->files;
458
459	if (files) {
460		task_lock(tsk);
461		tsk->files = NULL;
462		task_unlock(tsk);
463		put_files_struct(files);
464	}
465}
466
467struct files_struct init_files = {
468	.count		= ATOMIC_INIT(1),
469	.fdt		= &init_files.fdtab,
470	.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,
476	},
477	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
478};
479
480static unsigned long find_next_fd(struct fdtable *fdt, unsigned long start)
481{
482	unsigned long maxfd = fdt->max_fds;
483	unsigned long maxbit = maxfd / BITS_PER_LONG;
484	unsigned long bitbit = start / BITS_PER_LONG;
485
486	bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
487	if (bitbit > maxfd)
488		return maxfd;
489	if (bitbit > start)
490		start = bitbit;
491	return find_next_zero_bit(fdt->open_fds, maxfd, start);
492}
493
494/*
495 * allocate a file descriptor, mark it busy.
496 */
497int __alloc_fd(struct files_struct *files,
498	       unsigned start, unsigned end, unsigned flags)
499{
 
500	unsigned int fd;
501	int error;
502	struct fdtable *fdt;
503
504	spin_lock(&files->file_lock);
505repeat:
506	fdt = files_fdtable(files);
507	fd = start;
508	if (fd < files->next_fd)
509		fd = files->next_fd;
510
511	if (fd < fdt->max_fds)
512		fd = find_next_fd(fdt, fd);
513
514	/*
515	 * N.B. For clone tasks sharing a files structure, this test
516	 * will limit the total number of files that can be opened.
517	 */
518	error = -EMFILE;
519	if (fd >= end)
520		goto out;
521
522	error = expand_files(files, fd);
523	if (error < 0)
524		goto out;
525
526	/*
527	 * If we needed to expand the fs array we
528	 * might have blocked - try again.
529	 */
530	if (error)
531		goto repeat;
532
533	if (start <= files->next_fd)
534		files->next_fd = fd + 1;
535
536	__set_open_fd(fd, fdt);
537	if (flags & O_CLOEXEC)
538		__set_close_on_exec(fd, fdt);
539	else
540		__clear_close_on_exec(fd, fdt);
541	error = fd;
542#if 1
543	/* Sanity check */
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);
547	}
548#endif
549
550out:
551	spin_unlock(&files->file_lock);
552	return error;
553}
554
555static int alloc_fd(unsigned start, unsigned flags)
556{
557	return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
558}
559
560int get_unused_fd_flags(unsigned flags)
561{
562	return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
563}
564EXPORT_SYMBOL(get_unused_fd_flags);
565
566static void __put_unused_fd(struct files_struct *files, unsigned int fd)
567{
568	struct fdtable *fdt = files_fdtable(files);
569	__clear_open_fd(fd, fdt);
570	if (fd < files->next_fd)
571		files->next_fd = fd;
572}
573
574void put_unused_fd(unsigned int fd)
575{
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);
580}
581
582EXPORT_SYMBOL(put_unused_fd);
583
584/*
585 * Install a file pointer in the fd array.
586 *
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.
592 *
593 * It should never happen - if we allow dup2() do it, _really_ bad things
594 * will follow.
595 *
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.
602 */
603
604void __fd_install(struct files_struct *files, unsigned int fd,
605		struct file *file)
606{
607	struct fdtable *fdt;
608
609	might_sleep();
610	rcu_read_lock_sched();
611
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();
616	}
617	/* coupled with smp_wmb() in expand_fdtable() */
618	smp_rmb();
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();
623}
624
625void fd_install(unsigned int fd, struct file *file)
626{
627	__fd_install(current->files, fd, file);
628}
629
630EXPORT_SYMBOL(fd_install);
631
632/*
633 * The same warnings as for __alloc_fd()/__fd_install() apply here...
634 */
635int __close_fd(struct files_struct *files, unsigned fd)
636{
637	struct file *file;
638	struct fdtable *fdt;
639
640	spin_lock(&files->file_lock);
641	fdt = files_fdtable(files);
642	if (fd >= fdt->max_fds)
643		goto out_unlock;
644	file = fdt->fd[fd];
645	if (!file)
646		goto out_unlock;
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);
652
653out_unlock:
654	spin_unlock(&files->file_lock);
655	return -EBADF;
656}
657
658void do_close_on_exec(struct files_struct *files)
659{
660	unsigned i;
661	struct fdtable *fdt;
662
663	/* exec unshares first */
664	spin_lock(&files->file_lock);
665	for (i = 0; ; i++) {
666		unsigned long set;
667		unsigned fd = i * BITS_PER_LONG;
668		fdt = files_fdtable(files);
669		if (fd >= fdt->max_fds)
670			break;
671		set = fdt->close_on_exec[i];
672		if (!set)
673			continue;
674		fdt->close_on_exec[i] = 0;
675		for ( ; set ; fd++, set >>= 1) {
676			struct file *file;
677			if (!(set & 1))
678				continue;
679			file = fdt->fd[fd];
680			if (!file)
681				continue;
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);
686			cond_resched();
687			spin_lock(&files->file_lock);
688		}
689
690	}
691	spin_unlock(&files->file_lock);
692}
693
694static struct file *__fget(unsigned int fd, fmode_t mask)
695{
696	struct files_struct *files = current->files;
697	struct file *file;
698
699	rcu_read_lock();
700loop:
701	file = fcheck_files(files, fd);
702	if (file) {
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)
706		 */
707		if (file->f_mode & mask)
708			file = NULL;
709		else if (!get_file_rcu(file))
710			goto loop;
711	}
712	rcu_read_unlock();
713
714	return file;
715}
716
717struct file *fget(unsigned int fd)
718{
719	return __fget(fd, FMODE_PATH);
720}
721EXPORT_SYMBOL(fget);
722
723struct file *fget_raw(unsigned int fd)
724{
725	return __fget(fd, 0);
726}
727EXPORT_SYMBOL(fget_raw);
728
729/*
730 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
731 *
732 * You can use this instead of fget if you satisfy all of the following
733 * conditions:
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
740 *    and fput_light.
741 *
742 * The fput_needed flag returned by fget_light should be passed to the
743 * corresponding fput_light.
744 */
745static unsigned long __fget_light(unsigned int fd, fmode_t mask)
746{
747	struct files_struct *files = current->files;
748	struct file *file;
749
750	if (atomic_read(&files->count) == 1) {
751		file = __fcheck_files(files, fd);
752		if (!file || unlikely(file->f_mode & mask))
753			return 0;
754		return (unsigned long)file;
755	} else {
756		file = __fget(fd, mask);
757		if (!file)
758			return 0;
759		return FDPUT_FPUT | (unsigned long)file;
760	}
761}
762unsigned long __fdget(unsigned int fd)
763{
764	return __fget_light(fd, FMODE_PATH);
765}
766EXPORT_SYMBOL(__fdget);
767
768unsigned long __fdget_raw(unsigned int fd)
769{
770	return __fget_light(fd, 0);
771}
772
773unsigned long __fdget_pos(unsigned int fd)
774{
775	unsigned long v = __fdget(fd);
776	struct file *file = (struct file *)(v & ~3);
777
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);
782		}
783	}
784	return v;
785}
786
787/*
788 * We only lock f_pos if we have threads or if the file might be
789 * shared with another process. In both cases we'll have an elevated
790 * file count (done either by fdget() or by fork()).
791 */
792
793void set_close_on_exec(unsigned int fd, int flag)
794{
795	struct files_struct *files = current->files;
796	struct fdtable *fdt;
797	spin_lock(&files->file_lock);
798	fdt = files_fdtable(files);
799	if (flag)
800		__set_close_on_exec(fd, fdt);
801	else
802		__clear_close_on_exec(fd, fdt);
803	spin_unlock(&files->file_lock);
804}
805
806bool get_close_on_exec(unsigned int fd)
807{
808	struct files_struct *files = current->files;
809	struct fdtable *fdt;
810	bool res;
811	rcu_read_lock();
812	fdt = files_fdtable(files);
813	res = close_on_exec(fd, fdt);
814	rcu_read_unlock();
815	return res;
816}
817
818static int do_dup2(struct files_struct *files,
819	struct file *file, unsigned fd, unsigned flags)
820__releases(&files->file_lock)
821{
822	struct file *tofree;
823	struct fdtable *fdt;
824
825	/*
826	 * We need to detect attempts to do dup2() over allocated but still
827	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
828	 * extra work in their equivalent of fget() - they insert struct
829	 * file immediately after grabbing descriptor, mark it larval if
830	 * more work (e.g. actual opening) is needed and make sure that
831	 * fget() treats larval files as absent.  Potentially interesting,
832	 * but while extra work in fget() is trivial, locking implications
833	 * and amount of surgery on open()-related paths in VFS are not.
834	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
835	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
836	 * scope of POSIX or SUS, since neither considers shared descriptor
837	 * tables and this condition does not arise without those.
838	 */
839	fdt = files_fdtable(files);
840	tofree = fdt->fd[fd];
841	if (!tofree && fd_is_open(fd, fdt))
842		goto Ebusy;
843	get_file(file);
844	rcu_assign_pointer(fdt->fd[fd], file);
845	__set_open_fd(fd, fdt);
846	if (flags & O_CLOEXEC)
847		__set_close_on_exec(fd, fdt);
848	else
849		__clear_close_on_exec(fd, fdt);
850	spin_unlock(&files->file_lock);
851
852	if (tofree)
853		filp_close(tofree, files);
854
855	return fd;
856
857Ebusy:
858	spin_unlock(&files->file_lock);
859	return -EBUSY;
860}
861
862int replace_fd(unsigned fd, struct file *file, unsigned flags)
863{
864	int err;
865	struct files_struct *files = current->files;
866
867	if (!file)
868		return __close_fd(files, fd);
869
870	if (fd >= rlimit(RLIMIT_NOFILE))
871		return -EBADF;
872
873	spin_lock(&files->file_lock);
874	err = expand_files(files, fd);
875	if (unlikely(err < 0))
876		goto out_unlock;
877	return do_dup2(files, file, fd, flags);
878
879out_unlock:
880	spin_unlock(&files->file_lock);
881	return err;
882}
883
884SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
885{
886	int err = -EBADF;
887	struct file *file;
888	struct files_struct *files = current->files;
889
890	if ((flags & ~O_CLOEXEC) != 0)
891		return -EINVAL;
892
893	if (unlikely(oldfd == newfd))
894		return -EINVAL;
895
896	if (newfd >= rlimit(RLIMIT_NOFILE))
897		return -EBADF;
898
899	spin_lock(&files->file_lock);
900	err = expand_files(files, newfd);
901	file = fcheck(oldfd);
902	if (unlikely(!file))
903		goto Ebadf;
904	if (unlikely(err < 0)) {
905		if (err == -EMFILE)
906			goto Ebadf;
907		goto out_unlock;
908	}
909	return do_dup2(files, file, newfd, flags);
910
911Ebadf:
912	err = -EBADF;
913out_unlock:
914	spin_unlock(&files->file_lock);
915	return err;
916}
917
918SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
919{
920	if (unlikely(newfd == oldfd)) { /* corner case */
921		struct files_struct *files = current->files;
922		int retval = oldfd;
923
924		rcu_read_lock();
925		if (!fcheck_files(files, oldfd))
926			retval = -EBADF;
927		rcu_read_unlock();
928		return retval;
929	}
930	return sys_dup3(oldfd, newfd, 0);
931}
932
933SYSCALL_DEFINE1(dup, unsigned int, fildes)
934{
935	int ret = -EBADF;
936	struct file *file = fget_raw(fildes);
937
938	if (file) {
939		ret = get_unused_fd_flags(0);
940		if (ret >= 0)
941			fd_install(ret, file);
942		else
943			fput(file);
944	}
945	return ret;
946}
947
948int f_dupfd(unsigned int from, struct file *file, unsigned flags)
949{
950	int err;
951	if (from >= rlimit(RLIMIT_NOFILE))
952		return -EINVAL;
953	err = alloc_fd(from, flags);
954	if (err >= 0) {
955		get_file(file);
956		fd_install(err, file);
957	}
958	return err;
959}
960
961int iterate_fd(struct files_struct *files, unsigned n,
962		int (*f)(const void *, struct file *, unsigned),
963		const void *p)
964{
965	struct fdtable *fdt;
966	int res = 0;
967	if (!files)
968		return 0;
969	spin_lock(&files->file_lock);
970	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
971		struct file *file;
972		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
973		if (!file)
974			continue;
975		res = f(p, file, n);
976		if (res)
977			break;
978	}
979	spin_unlock(&files->file_lock);
980	return res;
981}
982EXPORT_SYMBOL(iterate_fd);