Linux Audio

Check our new training course

Loading...
v3.15
 
  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 max() is unusable in constant expressions ;-/ */
 29#define __const_max(x, y) ((x) < (y) ? (x) : (y))
 30int sysctl_nr_open_max = __const_max(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|__GFP_NOWARN|__GFP_NORETRY);
 41		if (data != NULL)
 42			return data;
 43	}
 44	return vmalloc(size);
 45}
 46
 47static void free_fdmem(void *ptr)
 48{
 49	is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
 50}
 51
 52static void __free_fdtable(struct fdtable *fdt)
 53{
 54	free_fdmem(fdt->fd);
 55	free_fdmem(fdt->open_fds);
 56	kfree(fdt);
 57}
 58
 59static void free_fdtable_rcu(struct rcu_head *rcu)
 60{
 61	__free_fdtable(container_of(rcu, struct fdtable, rcu));
 62}
 63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 64/*
 65 * Expand the fdset in the files_struct.  Called with the files spinlock
 66 * held for write.
 67 */
 68static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 69{
 70	unsigned int cpy, set;
 71
 72	BUG_ON(nfdt->max_fds < ofdt->max_fds);
 73
 74	cpy = ofdt->max_fds * sizeof(struct file *);
 75	set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 76	memcpy(nfdt->fd, ofdt->fd, cpy);
 77	memset((char *)(nfdt->fd) + cpy, 0, set);
 78
 79	cpy = ofdt->max_fds / BITS_PER_BYTE;
 80	set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
 81	memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
 82	memset((char *)(nfdt->open_fds) + cpy, 0, set);
 83	memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
 84	memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
 85}
 86
 87static struct fdtable * alloc_fdtable(unsigned int nr)
 88{
 89	struct fdtable *fdt;
 90	void *data;
 91
 92	/*
 93	 * Figure out how many fds we actually want to support in this fdtable.
 94	 * Allocation steps are keyed to the size of the fdarray, since it
 95	 * grows far faster than any of the other dynamic data. We try to fit
 96	 * the fdarray into comfortable page-tuned chunks: starting at 1024B
 97	 * and growing in powers of two from there on.
 98	 */
 99	nr /= (1024 / sizeof(struct file *));
100	nr = roundup_pow_of_two(nr + 1);
101	nr *= (1024 / sizeof(struct file *));
102	/*
103	 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
104	 * had been set lower between the check in expand_files() and here.  Deal
105	 * with that in caller, it's cheaper that way.
106	 *
107	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
108	 * bitmaps handling below becomes unpleasant, to put it mildly...
109	 */
110	if (unlikely(nr > sysctl_nr_open))
111		nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
112
113	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
114	if (!fdt)
115		goto out;
116	fdt->max_fds = nr;
117	data = alloc_fdmem(nr * sizeof(struct file *));
118	if (!data)
119		goto out_fdt;
120	fdt->fd = data;
121
122	data = alloc_fdmem(max_t(size_t,
123				 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
 
124	if (!data)
125		goto out_arr;
126	fdt->open_fds = data;
127	data += nr / BITS_PER_BYTE;
128	fdt->close_on_exec = data;
 
 
129
130	return fdt;
131
132out_arr:
133	free_fdmem(fdt->fd);
134out_fdt:
135	kfree(fdt);
136out:
137	return NULL;
138}
139
140/*
141 * Expand the file descriptor table.
142 * This function will allocate a new fdtable and both fd array and fdset, of
143 * the given size.
144 * Return <0 error code on error; 1 on successful completion.
145 * The files->file_lock should be held on entry, and will be held on exit.
146 */
147static int expand_fdtable(struct files_struct *files, int nr)
148	__releases(files->file_lock)
149	__acquires(files->file_lock)
150{
151	struct fdtable *new_fdt, *cur_fdt;
152
153	spin_unlock(&files->file_lock);
154	new_fdt = alloc_fdtable(nr);
 
 
 
 
 
 
 
155	spin_lock(&files->file_lock);
156	if (!new_fdt)
157		return -ENOMEM;
158	/*
159	 * extremely unlikely race - sysctl_nr_open decreased between the check in
160	 * caller and alloc_fdtable().  Cheaper to catch it here...
161	 */
162	if (unlikely(new_fdt->max_fds <= nr)) {
163		__free_fdtable(new_fdt);
164		return -EMFILE;
165	}
166	/*
167	 * Check again since another task may have expanded the fd table while
168	 * we dropped the lock
169	 */
170	cur_fdt = files_fdtable(files);
171	if (nr >= cur_fdt->max_fds) {
172		/* Continue as planned */
173		copy_fdtable(new_fdt, cur_fdt);
174		rcu_assign_pointer(files->fdt, new_fdt);
175		if (cur_fdt != &files->fdtab)
176			call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
177	} else {
178		/* Somebody else expanded, so undo our attempt */
179		__free_fdtable(new_fdt);
180	}
181	return 1;
182}
183
184/*
185 * Expand files.
186 * This function will expand the file structures, if the requested size exceeds
187 * the current capacity and there is room for expansion.
188 * Return <0 error code on error; 0 when nothing done; 1 when files were
189 * expanded and execution may have blocked.
190 * The files->file_lock should be held on entry, and will be held on exit.
191 */
192static int expand_files(struct files_struct *files, int nr)
 
 
193{
194	struct fdtable *fdt;
 
195
 
196	fdt = files_fdtable(files);
197
198	/* Do we need to expand? */
199	if (nr < fdt->max_fds)
200		return 0;
201
202	/* Can we expand? */
203	if (nr >= sysctl_nr_open)
204		return -EMFILE;
205
 
 
 
 
 
 
 
 
206	/* All good, so we try */
207	return expand_fdtable(files, nr);
 
 
 
 
 
208}
209
210static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
211{
212	__set_bit(fd, fdt->close_on_exec);
213}
214
215static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
216{
217	__clear_bit(fd, fdt->close_on_exec);
 
218}
219
220static inline void __set_open_fd(int fd, struct fdtable *fdt)
221{
222	__set_bit(fd, fdt->open_fds);
 
 
 
223}
224
225static inline void __clear_open_fd(int fd, struct fdtable *fdt)
226{
227	__clear_bit(fd, fdt->open_fds);
 
228}
229
230static int count_open_files(struct fdtable *fdt)
231{
232	int size = fdt->max_fds;
233	int i;
234
235	/* Find the last open fd */
236	for (i = size / BITS_PER_LONG; i > 0; ) {
237		if (fdt->open_fds[--i])
238			break;
239	}
240	i = (i + 1) * BITS_PER_LONG;
241	return i;
242}
243
244/*
245 * Allocate a new files structure and copy contents from the
246 * passed in files structure.
247 * errorp will be valid only when the returned files_struct is NULL.
248 */
249struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
250{
251	struct files_struct *newf;
252	struct file **old_fds, **new_fds;
253	int open_files, size, i;
254	struct fdtable *old_fdt, *new_fdt;
255
256	*errorp = -ENOMEM;
257	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
258	if (!newf)
259		goto out;
260
261	atomic_set(&newf->count, 1);
262
263	spin_lock_init(&newf->file_lock);
 
 
264	newf->next_fd = 0;
265	new_fdt = &newf->fdtab;
266	new_fdt->max_fds = NR_OPEN_DEFAULT;
267	new_fdt->close_on_exec = newf->close_on_exec_init;
268	new_fdt->open_fds = newf->open_fds_init;
 
269	new_fdt->fd = &newf->fd_array[0];
270
271	spin_lock(&oldf->file_lock);
272	old_fdt = files_fdtable(oldf);
273	open_files = count_open_files(old_fdt);
274
275	/*
276	 * Check whether we need to allocate a larger fd array and fd set.
277	 */
278	while (unlikely(open_files > new_fdt->max_fds)) {
279		spin_unlock(&oldf->file_lock);
280
281		if (new_fdt != &newf->fdtab)
282			__free_fdtable(new_fdt);
283
284		new_fdt = alloc_fdtable(open_files - 1);
285		if (!new_fdt) {
286			*errorp = -ENOMEM;
287			goto out_release;
288		}
289
290		/* beyond sysctl_nr_open; nothing to do */
291		if (unlikely(new_fdt->max_fds < open_files)) {
292			__free_fdtable(new_fdt);
293			*errorp = -EMFILE;
294			goto out_release;
295		}
296
297		/*
298		 * Reacquire the oldf lock and a pointer to its fd table
299		 * who knows it may have a new bigger fd table. We need
300		 * the latest pointer.
301		 */
302		spin_lock(&oldf->file_lock);
303		old_fdt = files_fdtable(oldf);
304		open_files = count_open_files(old_fdt);
305	}
306
 
 
307	old_fds = old_fdt->fd;
308	new_fds = new_fdt->fd;
309
310	memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
311	memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
312
313	for (i = open_files; i != 0; i--) {
314		struct file *f = *old_fds++;
315		if (f) {
316			get_file(f);
317		} else {
318			/*
319			 * The fd may be claimed in the fd bitmap but not yet
320			 * instantiated in the files array if a sibling thread
321			 * is partway through open().  So make sure that this
322			 * fd is available to the new process.
323			 */
324			__clear_open_fd(open_files - i, new_fdt);
325		}
326		rcu_assign_pointer(*new_fds++, f);
327	}
328	spin_unlock(&oldf->file_lock);
329
330	/* compute the remainder to be cleared */
331	size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
332
333	/* This is long word aligned thus could use a optimized version */
334	memset(new_fds, 0, size);
335
336	if (new_fdt->max_fds > open_files) {
337		int left = (new_fdt->max_fds - open_files) / 8;
338		int start = open_files / BITS_PER_LONG;
339
340		memset(&new_fdt->open_fds[start], 0, left);
341		memset(&new_fdt->close_on_exec[start], 0, left);
342	}
343
344	rcu_assign_pointer(newf->fdt, new_fdt);
345
346	return newf;
347
348out_release:
349	kmem_cache_free(files_cachep, newf);
350out:
351	return NULL;
352}
353
354static struct fdtable *close_files(struct files_struct * files)
355{
356	/*
357	 * It is safe to dereference the fd table without RCU or
358	 * ->file_lock because this is the last reference to the
359	 * files structure.
360	 */
361	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
362	int i, j = 0;
363
364	for (;;) {
365		unsigned long set;
366		i = j * BITS_PER_LONG;
367		if (i >= fdt->max_fds)
368			break;
369		set = fdt->open_fds[j++];
370		while (set) {
371			if (set & 1) {
372				struct file * file = xchg(&fdt->fd[i], NULL);
373				if (file) {
374					filp_close(file, files);
375					cond_resched();
376				}
377			}
378			i++;
379			set >>= 1;
380		}
381	}
382
383	return fdt;
384}
385
386struct files_struct *get_files_struct(struct task_struct *task)
387{
388	struct files_struct *files;
389
390	task_lock(task);
391	files = task->files;
392	if (files)
393		atomic_inc(&files->count);
394	task_unlock(task);
395
396	return files;
397}
398
399void put_files_struct(struct files_struct *files)
400{
401	if (atomic_dec_and_test(&files->count)) {
402		struct fdtable *fdt = close_files(files);
403
404		/* free the arrays if they are not embedded */
405		if (fdt != &files->fdtab)
406			__free_fdtable(fdt);
407		kmem_cache_free(files_cachep, files);
408	}
409}
410
411void reset_files_struct(struct files_struct *files)
412{
413	struct task_struct *tsk = current;
414	struct files_struct *old;
415
416	old = tsk->files;
417	task_lock(tsk);
418	tsk->files = files;
419	task_unlock(tsk);
420	put_files_struct(old);
421}
422
423void exit_files(struct task_struct *tsk)
424{
425	struct files_struct * files = tsk->files;
426
427	if (files) {
428		task_lock(tsk);
429		tsk->files = NULL;
430		task_unlock(tsk);
431		put_files_struct(files);
432	}
433}
434
435struct files_struct init_files = {
436	.count		= ATOMIC_INIT(1),
437	.fdt		= &init_files.fdtab,
438	.fdtab		= {
439		.max_fds	= NR_OPEN_DEFAULT,
440		.fd		= &init_files.fd_array[0],
441		.close_on_exec	= init_files.close_on_exec_init,
442		.open_fds	= init_files.open_fds_init,
 
443	},
444	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
 
445};
446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
447/*
448 * allocate a file descriptor, mark it busy.
449 */
450int __alloc_fd(struct files_struct *files,
451	       unsigned start, unsigned end, unsigned flags)
452{
453	unsigned int fd;
454	int error;
455	struct fdtable *fdt;
456
457	spin_lock(&files->file_lock);
458repeat:
459	fdt = files_fdtable(files);
460	fd = start;
461	if (fd < files->next_fd)
462		fd = files->next_fd;
463
464	if (fd < fdt->max_fds)
465		fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
466
467	/*
468	 * N.B. For clone tasks sharing a files structure, this test
469	 * will limit the total number of files that can be opened.
470	 */
471	error = -EMFILE;
472	if (fd >= end)
473		goto out;
474
475	error = expand_files(files, fd);
476	if (error < 0)
477		goto out;
478
479	/*
480	 * If we needed to expand the fs array we
481	 * might have blocked - try again.
482	 */
483	if (error)
484		goto repeat;
485
486	if (start <= files->next_fd)
487		files->next_fd = fd + 1;
488
489	__set_open_fd(fd, fdt);
490	if (flags & O_CLOEXEC)
491		__set_close_on_exec(fd, fdt);
492	else
493		__clear_close_on_exec(fd, fdt);
494	error = fd;
495#if 1
496	/* Sanity check */
497	if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
498		printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
499		rcu_assign_pointer(fdt->fd[fd], NULL);
500	}
501#endif
502
503out:
504	spin_unlock(&files->file_lock);
505	return error;
506}
507
508static int alloc_fd(unsigned start, unsigned flags)
509{
510	return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
511}
512
513int get_unused_fd_flags(unsigned flags)
514{
515	return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
516}
517EXPORT_SYMBOL(get_unused_fd_flags);
518
519static void __put_unused_fd(struct files_struct *files, unsigned int fd)
520{
521	struct fdtable *fdt = files_fdtable(files);
522	__clear_open_fd(fd, fdt);
523	if (fd < files->next_fd)
524		files->next_fd = fd;
525}
526
527void put_unused_fd(unsigned int fd)
528{
529	struct files_struct *files = current->files;
530	spin_lock(&files->file_lock);
531	__put_unused_fd(files, fd);
532	spin_unlock(&files->file_lock);
533}
534
535EXPORT_SYMBOL(put_unused_fd);
536
537/*
538 * Install a file pointer in the fd array.
539 *
540 * The VFS is full of places where we drop the files lock between
541 * setting the open_fds bitmap and installing the file in the file
542 * array.  At any such point, we are vulnerable to a dup2() race
543 * installing a file in the array before us.  We need to detect this and
544 * fput() the struct file we are about to overwrite in this case.
545 *
546 * It should never happen - if we allow dup2() do it, _really_ bad things
547 * will follow.
548 *
549 * NOTE: __fd_install() variant is really, really low-level; don't
550 * use it unless you are forced to by truly lousy API shoved down
551 * your throat.  'files' *MUST* be either current->files or obtained
552 * by get_files_struct(current) done by whoever had given it to you,
553 * or really bad things will happen.  Normally you want to use
554 * fd_install() instead.
555 */
556
557void __fd_install(struct files_struct *files, unsigned int fd,
558		struct file *file)
559{
560	struct fdtable *fdt;
561	spin_lock(&files->file_lock);
562	fdt = files_fdtable(files);
 
 
 
 
 
 
 
 
 
 
 
 
 
563	BUG_ON(fdt->fd[fd] != NULL);
564	rcu_assign_pointer(fdt->fd[fd], file);
565	spin_unlock(&files->file_lock);
566}
567
568void fd_install(unsigned int fd, struct file *file)
569{
570	__fd_install(current->files, fd, file);
571}
572
573EXPORT_SYMBOL(fd_install);
574
575/*
576 * The same warnings as for __alloc_fd()/__fd_install() apply here...
577 */
578int __close_fd(struct files_struct *files, unsigned fd)
579{
580	struct file *file;
581	struct fdtable *fdt;
582
583	spin_lock(&files->file_lock);
584	fdt = files_fdtable(files);
585	if (fd >= fdt->max_fds)
586		goto out_unlock;
587	file = fdt->fd[fd];
588	if (!file)
589		goto out_unlock;
590	rcu_assign_pointer(fdt->fd[fd], NULL);
591	__clear_close_on_exec(fd, fdt);
592	__put_unused_fd(files, fd);
593	spin_unlock(&files->file_lock);
594	return filp_close(file, files);
595
596out_unlock:
597	spin_unlock(&files->file_lock);
598	return -EBADF;
599}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
600
601void do_close_on_exec(struct files_struct *files)
602{
603	unsigned i;
604	struct fdtable *fdt;
605
606	/* exec unshares first */
607	spin_lock(&files->file_lock);
608	for (i = 0; ; i++) {
609		unsigned long set;
610		unsigned fd = i * BITS_PER_LONG;
611		fdt = files_fdtable(files);
612		if (fd >= fdt->max_fds)
613			break;
614		set = fdt->close_on_exec[i];
615		if (!set)
616			continue;
617		fdt->close_on_exec[i] = 0;
618		for ( ; set ; fd++, set >>= 1) {
619			struct file *file;
620			if (!(set & 1))
621				continue;
622			file = fdt->fd[fd];
623			if (!file)
624				continue;
625			rcu_assign_pointer(fdt->fd[fd], NULL);
626			__put_unused_fd(files, fd);
627			spin_unlock(&files->file_lock);
628			filp_close(file, files);
629			cond_resched();
630			spin_lock(&files->file_lock);
631		}
632
633	}
634	spin_unlock(&files->file_lock);
635}
636
637static struct file *__fget(unsigned int fd, fmode_t mask)
638{
639	struct files_struct *files = current->files;
640	struct file *file;
641
642	rcu_read_lock();
 
643	file = fcheck_files(files, fd);
644	if (file) {
645		/* File object ref couldn't be taken */
646		if ((file->f_mode & mask) ||
647		    !atomic_long_inc_not_zero(&file->f_count))
 
 
648			file = NULL;
 
 
649	}
650	rcu_read_unlock();
651
652	return file;
653}
654
 
 
 
 
 
655struct file *fget(unsigned int fd)
656{
657	return __fget(fd, FMODE_PATH);
658}
659EXPORT_SYMBOL(fget);
660
661struct file *fget_raw(unsigned int fd)
662{
663	return __fget(fd, 0);
664}
665EXPORT_SYMBOL(fget_raw);
666
667/*
668 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
669 *
670 * You can use this instead of fget if you satisfy all of the following
671 * conditions:
672 * 1) You must call fput_light before exiting the syscall and returning control
673 *    to userspace (i.e. you cannot remember the returned struct file * after
674 *    returning to userspace).
675 * 2) You must not call filp_close on the returned struct file * in between
676 *    calls to fget_light and fput_light.
677 * 3) You must not clone the current task in between the calls to fget_light
678 *    and fput_light.
679 *
680 * The fput_needed flag returned by fget_light should be passed to the
681 * corresponding fput_light.
682 */
683static unsigned long __fget_light(unsigned int fd, fmode_t mask)
684{
685	struct files_struct *files = current->files;
686	struct file *file;
687
688	if (atomic_read(&files->count) == 1) {
689		file = __fcheck_files(files, fd);
690		if (!file || unlikely(file->f_mode & mask))
691			return 0;
692		return (unsigned long)file;
693	} else {
694		file = __fget(fd, mask);
695		if (!file)
696			return 0;
697		return FDPUT_FPUT | (unsigned long)file;
698	}
699}
700unsigned long __fdget(unsigned int fd)
701{
702	return __fget_light(fd, FMODE_PATH);
703}
704EXPORT_SYMBOL(__fdget);
705
706unsigned long __fdget_raw(unsigned int fd)
707{
708	return __fget_light(fd, 0);
709}
710
711unsigned long __fdget_pos(unsigned int fd)
712{
713	unsigned long v = __fdget(fd);
714	struct file *file = (struct file *)(v & ~3);
715
716	if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
717		if (file_count(file) > 1) {
718			v |= FDPUT_POS_UNLOCK;
719			mutex_lock(&file->f_pos_lock);
720		}
721	}
722	return v;
723}
724
 
 
 
 
 
725/*
726 * We only lock f_pos if we have threads or if the file might be
727 * shared with another process. In both cases we'll have an elevated
728 * file count (done either by fdget() or by fork()).
729 */
730
731void set_close_on_exec(unsigned int fd, int flag)
732{
733	struct files_struct *files = current->files;
734	struct fdtable *fdt;
735	spin_lock(&files->file_lock);
736	fdt = files_fdtable(files);
737	if (flag)
738		__set_close_on_exec(fd, fdt);
739	else
740		__clear_close_on_exec(fd, fdt);
741	spin_unlock(&files->file_lock);
742}
743
744bool get_close_on_exec(unsigned int fd)
745{
746	struct files_struct *files = current->files;
747	struct fdtable *fdt;
748	bool res;
749	rcu_read_lock();
750	fdt = files_fdtable(files);
751	res = close_on_exec(fd, fdt);
752	rcu_read_unlock();
753	return res;
754}
755
756static int do_dup2(struct files_struct *files,
757	struct file *file, unsigned fd, unsigned flags)
 
758{
759	struct file *tofree;
760	struct fdtable *fdt;
761
762	/*
763	 * We need to detect attempts to do dup2() over allocated but still
764	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
765	 * extra work in their equivalent of fget() - they insert struct
766	 * file immediately after grabbing descriptor, mark it larval if
767	 * more work (e.g. actual opening) is needed and make sure that
768	 * fget() treats larval files as absent.  Potentially interesting,
769	 * but while extra work in fget() is trivial, locking implications
770	 * and amount of surgery on open()-related paths in VFS are not.
771	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
772	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
773	 * scope of POSIX or SUS, since neither considers shared descriptor
774	 * tables and this condition does not arise without those.
775	 */
776	fdt = files_fdtable(files);
777	tofree = fdt->fd[fd];
778	if (!tofree && fd_is_open(fd, fdt))
779		goto Ebusy;
780	get_file(file);
781	rcu_assign_pointer(fdt->fd[fd], file);
782	__set_open_fd(fd, fdt);
783	if (flags & O_CLOEXEC)
784		__set_close_on_exec(fd, fdt);
785	else
786		__clear_close_on_exec(fd, fdt);
787	spin_unlock(&files->file_lock);
788
789	if (tofree)
790		filp_close(tofree, files);
791
792	return fd;
793
794Ebusy:
795	spin_unlock(&files->file_lock);
796	return -EBUSY;
797}
798
799int replace_fd(unsigned fd, struct file *file, unsigned flags)
800{
801	int err;
802	struct files_struct *files = current->files;
803
804	if (!file)
805		return __close_fd(files, fd);
806
807	if (fd >= rlimit(RLIMIT_NOFILE))
808		return -EBADF;
809
810	spin_lock(&files->file_lock);
811	err = expand_files(files, fd);
812	if (unlikely(err < 0))
813		goto out_unlock;
814	return do_dup2(files, file, fd, flags);
815
816out_unlock:
817	spin_unlock(&files->file_lock);
818	return err;
819}
820
821SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
822{
823	int err = -EBADF;
824	struct file *file;
825	struct files_struct *files = current->files;
826
827	if ((flags & ~O_CLOEXEC) != 0)
828		return -EINVAL;
829
830	if (unlikely(oldfd == newfd))
831		return -EINVAL;
832
833	if (newfd >= rlimit(RLIMIT_NOFILE))
834		return -EBADF;
835
836	spin_lock(&files->file_lock);
837	err = expand_files(files, newfd);
838	file = fcheck(oldfd);
839	if (unlikely(!file))
840		goto Ebadf;
841	if (unlikely(err < 0)) {
842		if (err == -EMFILE)
843			goto Ebadf;
844		goto out_unlock;
845	}
846	return do_dup2(files, file, newfd, flags);
847
848Ebadf:
849	err = -EBADF;
850out_unlock:
851	spin_unlock(&files->file_lock);
852	return err;
853}
854
 
 
 
 
 
855SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
856{
857	if (unlikely(newfd == oldfd)) { /* corner case */
858		struct files_struct *files = current->files;
859		int retval = oldfd;
860
861		rcu_read_lock();
862		if (!fcheck_files(files, oldfd))
863			retval = -EBADF;
864		rcu_read_unlock();
865		return retval;
866	}
867	return sys_dup3(oldfd, newfd, 0);
868}
869
870SYSCALL_DEFINE1(dup, unsigned int, fildes)
871{
872	int ret = -EBADF;
873	struct file *file = fget_raw(fildes);
874
875	if (file) {
876		ret = get_unused_fd();
877		if (ret >= 0)
878			fd_install(ret, file);
879		else
880			fput(file);
881	}
882	return ret;
 
 
 
 
 
883}
884
885int f_dupfd(unsigned int from, struct file *file, unsigned flags)
886{
887	int err;
888	if (from >= rlimit(RLIMIT_NOFILE))
889		return -EINVAL;
890	err = alloc_fd(from, flags);
891	if (err >= 0) {
892		get_file(file);
893		fd_install(err, file);
894	}
895	return err;
896}
897
898int iterate_fd(struct files_struct *files, unsigned n,
899		int (*f)(const void *, struct file *, unsigned),
900		const void *p)
901{
902	struct fdtable *fdt;
903	int res = 0;
904	if (!files)
905		return 0;
906	spin_lock(&files->file_lock);
907	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
908		struct file *file;
909		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
910		if (!file)
911			continue;
912		res = f(p, file, n);
913		if (res)
914			break;
915	}
916	spin_unlock(&files->file_lock);
917	return res;
918}
919EXPORT_SYMBOL(iterate_fd);
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  linux/fs/file.c
   4 *
   5 *  Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
   6 *
   7 *  Manage the dynamic fd arrays in the process files_struct.
   8 */
   9
  10#include <linux/syscalls.h>
  11#include <linux/export.h>
  12#include <linux/fs.h>
  13#include <linux/mm.h>
  14#include <linux/sched/signal.h>
 
 
  15#include <linux/slab.h>
 
  16#include <linux/file.h>
  17#include <linux/fdtable.h>
  18#include <linux/bitops.h>
 
  19#include <linux/spinlock.h>
  20#include <linux/rcupdate.h>
 
  21
  22unsigned int sysctl_nr_open __read_mostly = 1024*1024;
  23unsigned int sysctl_nr_open_min = BITS_PER_LONG;
  24/* our min() is unusable in constant expressions ;-/ */
  25#define __const_min(x, y) ((x) < (y) ? (x) : (y))
  26unsigned int sysctl_nr_open_max =
  27	__const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  28
  29static void __free_fdtable(struct fdtable *fdt)
  30{
  31	kvfree(fdt->fd);
  32	kvfree(fdt->open_fds);
  33	kfree(fdt);
  34}
  35
  36static void free_fdtable_rcu(struct rcu_head *rcu)
  37{
  38	__free_fdtable(container_of(rcu, struct fdtable, rcu));
  39}
  40
  41#define BITBIT_NR(nr)	BITS_TO_LONGS(BITS_TO_LONGS(nr))
  42#define BITBIT_SIZE(nr)	(BITBIT_NR(nr) * sizeof(long))
  43
  44/*
  45 * Copy 'count' fd bits from the old table to the new table and clear the extra
  46 * space if any.  This does not copy the file pointers.  Called with the files
  47 * spinlock held for write.
  48 */
  49static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
  50			    unsigned int count)
  51{
  52	unsigned int cpy, set;
  53
  54	cpy = count / BITS_PER_BYTE;
  55	set = (nfdt->max_fds - count) / BITS_PER_BYTE;
  56	memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
  57	memset((char *)nfdt->open_fds + cpy, 0, set);
  58	memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
  59	memset((char *)nfdt->close_on_exec + cpy, 0, set);
  60
  61	cpy = BITBIT_SIZE(count);
  62	set = BITBIT_SIZE(nfdt->max_fds) - cpy;
  63	memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
  64	memset((char *)nfdt->full_fds_bits + cpy, 0, set);
  65}
  66
  67/*
  68 * Copy all file descriptors from the old table to the new, expanded table and
  69 * clear the extra space.  Called with the files spinlock held for write.
  70 */
  71static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
  72{
  73	unsigned int cpy, set;
  74
  75	BUG_ON(nfdt->max_fds < ofdt->max_fds);
  76
  77	cpy = ofdt->max_fds * sizeof(struct file *);
  78	set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
  79	memcpy(nfdt->fd, ofdt->fd, cpy);
  80	memset((char *)nfdt->fd + cpy, 0, set);
  81
  82	copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
 
 
 
 
 
  83}
  84
  85static struct fdtable * alloc_fdtable(unsigned int nr)
  86{
  87	struct fdtable *fdt;
  88	void *data;
  89
  90	/*
  91	 * Figure out how many fds we actually want to support in this fdtable.
  92	 * Allocation steps are keyed to the size of the fdarray, since it
  93	 * grows far faster than any of the other dynamic data. We try to fit
  94	 * the fdarray into comfortable page-tuned chunks: starting at 1024B
  95	 * and growing in powers of two from there on.
  96	 */
  97	nr /= (1024 / sizeof(struct file *));
  98	nr = roundup_pow_of_two(nr + 1);
  99	nr *= (1024 / sizeof(struct file *));
 100	/*
 101	 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
 102	 * had been set lower between the check in expand_files() and here.  Deal
 103	 * with that in caller, it's cheaper that way.
 104	 *
 105	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
 106	 * bitmaps handling below becomes unpleasant, to put it mildly...
 107	 */
 108	if (unlikely(nr > sysctl_nr_open))
 109		nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
 110
 111	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
 112	if (!fdt)
 113		goto out;
 114	fdt->max_fds = nr;
 115	data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
 116	if (!data)
 117		goto out_fdt;
 118	fdt->fd = data;
 119
 120	data = kvmalloc(max_t(size_t,
 121				 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
 122				 GFP_KERNEL_ACCOUNT);
 123	if (!data)
 124		goto out_arr;
 125	fdt->open_fds = data;
 126	data += nr / BITS_PER_BYTE;
 127	fdt->close_on_exec = data;
 128	data += nr / BITS_PER_BYTE;
 129	fdt->full_fds_bits = data;
 130
 131	return fdt;
 132
 133out_arr:
 134	kvfree(fdt->fd);
 135out_fdt:
 136	kfree(fdt);
 137out:
 138	return NULL;
 139}
 140
 141/*
 142 * Expand the file descriptor table.
 143 * This function will allocate a new fdtable and both fd array and fdset, of
 144 * the given size.
 145 * Return <0 error code on error; 1 on successful completion.
 146 * The files->file_lock should be held on entry, and will be held on exit.
 147 */
 148static int expand_fdtable(struct files_struct *files, unsigned int nr)
 149	__releases(files->file_lock)
 150	__acquires(files->file_lock)
 151{
 152	struct fdtable *new_fdt, *cur_fdt;
 153
 154	spin_unlock(&files->file_lock);
 155	new_fdt = alloc_fdtable(nr);
 156
 157	/* make sure all __fd_install() have seen resize_in_progress
 158	 * or have finished their rcu_read_lock_sched() section.
 159	 */
 160	if (atomic_read(&files->count) > 1)
 161		synchronize_rcu();
 162
 163	spin_lock(&files->file_lock);
 164	if (!new_fdt)
 165		return -ENOMEM;
 166	/*
 167	 * extremely unlikely race - sysctl_nr_open decreased between the check in
 168	 * caller and alloc_fdtable().  Cheaper to catch it here...
 169	 */
 170	if (unlikely(new_fdt->max_fds <= nr)) {
 171		__free_fdtable(new_fdt);
 172		return -EMFILE;
 173	}
 
 
 
 
 174	cur_fdt = files_fdtable(files);
 175	BUG_ON(nr < cur_fdt->max_fds);
 176	copy_fdtable(new_fdt, cur_fdt);
 177	rcu_assign_pointer(files->fdt, new_fdt);
 178	if (cur_fdt != &files->fdtab)
 179		call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
 180	/* coupled with smp_rmb() in __fd_install() */
 181	smp_wmb();
 
 
 
 182	return 1;
 183}
 184
 185/*
 186 * Expand files.
 187 * This function will expand the file structures, if the requested size exceeds
 188 * the current capacity and there is room for expansion.
 189 * Return <0 error code on error; 0 when nothing done; 1 when files were
 190 * expanded and execution may have blocked.
 191 * The files->file_lock should be held on entry, and will be held on exit.
 192 */
 193static int expand_files(struct files_struct *files, unsigned int nr)
 194	__releases(files->file_lock)
 195	__acquires(files->file_lock)
 196{
 197	struct fdtable *fdt;
 198	int expanded = 0;
 199
 200repeat:
 201	fdt = files_fdtable(files);
 202
 203	/* Do we need to expand? */
 204	if (nr < fdt->max_fds)
 205		return expanded;
 206
 207	/* Can we expand? */
 208	if (nr >= sysctl_nr_open)
 209		return -EMFILE;
 210
 211	if (unlikely(files->resize_in_progress)) {
 212		spin_unlock(&files->file_lock);
 213		expanded = 1;
 214		wait_event(files->resize_wait, !files->resize_in_progress);
 215		spin_lock(&files->file_lock);
 216		goto repeat;
 217	}
 218
 219	/* All good, so we try */
 220	files->resize_in_progress = true;
 221	expanded = expand_fdtable(files, nr);
 222	files->resize_in_progress = false;
 223
 224	wake_up_all(&files->resize_wait);
 225	return expanded;
 226}
 227
 228static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
 229{
 230	__set_bit(fd, fdt->close_on_exec);
 231}
 232
 233static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
 234{
 235	if (test_bit(fd, fdt->close_on_exec))
 236		__clear_bit(fd, fdt->close_on_exec);
 237}
 238
 239static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
 240{
 241	__set_bit(fd, fdt->open_fds);
 242	fd /= BITS_PER_LONG;
 243	if (!~fdt->open_fds[fd])
 244		__set_bit(fd, fdt->full_fds_bits);
 245}
 246
 247static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
 248{
 249	__clear_bit(fd, fdt->open_fds);
 250	__clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
 251}
 252
 253static unsigned int count_open_files(struct fdtable *fdt)
 254{
 255	unsigned int size = fdt->max_fds;
 256	unsigned int i;
 257
 258	/* Find the last open fd */
 259	for (i = size / BITS_PER_LONG; i > 0; ) {
 260		if (fdt->open_fds[--i])
 261			break;
 262	}
 263	i = (i + 1) * BITS_PER_LONG;
 264	return i;
 265}
 266
 267/*
 268 * Allocate a new files structure and copy contents from the
 269 * passed in files structure.
 270 * errorp will be valid only when the returned files_struct is NULL.
 271 */
 272struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
 273{
 274	struct files_struct *newf;
 275	struct file **old_fds, **new_fds;
 276	unsigned int open_files, i;
 277	struct fdtable *old_fdt, *new_fdt;
 278
 279	*errorp = -ENOMEM;
 280	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
 281	if (!newf)
 282		goto out;
 283
 284	atomic_set(&newf->count, 1);
 285
 286	spin_lock_init(&newf->file_lock);
 287	newf->resize_in_progress = false;
 288	init_waitqueue_head(&newf->resize_wait);
 289	newf->next_fd = 0;
 290	new_fdt = &newf->fdtab;
 291	new_fdt->max_fds = NR_OPEN_DEFAULT;
 292	new_fdt->close_on_exec = newf->close_on_exec_init;
 293	new_fdt->open_fds = newf->open_fds_init;
 294	new_fdt->full_fds_bits = newf->full_fds_bits_init;
 295	new_fdt->fd = &newf->fd_array[0];
 296
 297	spin_lock(&oldf->file_lock);
 298	old_fdt = files_fdtable(oldf);
 299	open_files = count_open_files(old_fdt);
 300
 301	/*
 302	 * Check whether we need to allocate a larger fd array and fd set.
 303	 */
 304	while (unlikely(open_files > new_fdt->max_fds)) {
 305		spin_unlock(&oldf->file_lock);
 306
 307		if (new_fdt != &newf->fdtab)
 308			__free_fdtable(new_fdt);
 309
 310		new_fdt = alloc_fdtable(open_files - 1);
 311		if (!new_fdt) {
 312			*errorp = -ENOMEM;
 313			goto out_release;
 314		}
 315
 316		/* beyond sysctl_nr_open; nothing to do */
 317		if (unlikely(new_fdt->max_fds < open_files)) {
 318			__free_fdtable(new_fdt);
 319			*errorp = -EMFILE;
 320			goto out_release;
 321		}
 322
 323		/*
 324		 * Reacquire the oldf lock and a pointer to its fd table
 325		 * who knows it may have a new bigger fd table. We need
 326		 * the latest pointer.
 327		 */
 328		spin_lock(&oldf->file_lock);
 329		old_fdt = files_fdtable(oldf);
 330		open_files = count_open_files(old_fdt);
 331	}
 332
 333	copy_fd_bitmaps(new_fdt, old_fdt, open_files);
 334
 335	old_fds = old_fdt->fd;
 336	new_fds = new_fdt->fd;
 337
 
 
 
 338	for (i = open_files; i != 0; i--) {
 339		struct file *f = *old_fds++;
 340		if (f) {
 341			get_file(f);
 342		} else {
 343			/*
 344			 * The fd may be claimed in the fd bitmap but not yet
 345			 * instantiated in the files array if a sibling thread
 346			 * is partway through open().  So make sure that this
 347			 * fd is available to the new process.
 348			 */
 349			__clear_open_fd(open_files - i, new_fdt);
 350		}
 351		rcu_assign_pointer(*new_fds++, f);
 352	}
 353	spin_unlock(&oldf->file_lock);
 354
 355	/* clear the remainder */
 356	memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
 
 
 
 
 
 
 
 
 
 
 
 357
 358	rcu_assign_pointer(newf->fdt, new_fdt);
 359
 360	return newf;
 361
 362out_release:
 363	kmem_cache_free(files_cachep, newf);
 364out:
 365	return NULL;
 366}
 367
 368static struct fdtable *close_files(struct files_struct * files)
 369{
 370	/*
 371	 * It is safe to dereference the fd table without RCU or
 372	 * ->file_lock because this is the last reference to the
 373	 * files structure.
 374	 */
 375	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
 376	unsigned int i, j = 0;
 377
 378	for (;;) {
 379		unsigned long set;
 380		i = j * BITS_PER_LONG;
 381		if (i >= fdt->max_fds)
 382			break;
 383		set = fdt->open_fds[j++];
 384		while (set) {
 385			if (set & 1) {
 386				struct file * file = xchg(&fdt->fd[i], NULL);
 387				if (file) {
 388					filp_close(file, files);
 389					cond_resched();
 390				}
 391			}
 392			i++;
 393			set >>= 1;
 394		}
 395	}
 396
 397	return fdt;
 398}
 399
 400struct files_struct *get_files_struct(struct task_struct *task)
 401{
 402	struct files_struct *files;
 403
 404	task_lock(task);
 405	files = task->files;
 406	if (files)
 407		atomic_inc(&files->count);
 408	task_unlock(task);
 409
 410	return files;
 411}
 412
 413void put_files_struct(struct files_struct *files)
 414{
 415	if (atomic_dec_and_test(&files->count)) {
 416		struct fdtable *fdt = close_files(files);
 417
 418		/* free the arrays if they are not embedded */
 419		if (fdt != &files->fdtab)
 420			__free_fdtable(fdt);
 421		kmem_cache_free(files_cachep, files);
 422	}
 423}
 424
 425void reset_files_struct(struct files_struct *files)
 426{
 427	struct task_struct *tsk = current;
 428	struct files_struct *old;
 429
 430	old = tsk->files;
 431	task_lock(tsk);
 432	tsk->files = files;
 433	task_unlock(tsk);
 434	put_files_struct(old);
 435}
 436
 437void exit_files(struct task_struct *tsk)
 438{
 439	struct files_struct * files = tsk->files;
 440
 441	if (files) {
 442		task_lock(tsk);
 443		tsk->files = NULL;
 444		task_unlock(tsk);
 445		put_files_struct(files);
 446	}
 447}
 448
 449struct files_struct init_files = {
 450	.count		= ATOMIC_INIT(1),
 451	.fdt		= &init_files.fdtab,
 452	.fdtab		= {
 453		.max_fds	= NR_OPEN_DEFAULT,
 454		.fd		= &init_files.fd_array[0],
 455		.close_on_exec	= init_files.close_on_exec_init,
 456		.open_fds	= init_files.open_fds_init,
 457		.full_fds_bits	= init_files.full_fds_bits_init,
 458	},
 459	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
 460	.resize_wait	= __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
 461};
 462
 463static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
 464{
 465	unsigned int maxfd = fdt->max_fds;
 466	unsigned int maxbit = maxfd / BITS_PER_LONG;
 467	unsigned int bitbit = start / BITS_PER_LONG;
 468
 469	bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
 470	if (bitbit > maxfd)
 471		return maxfd;
 472	if (bitbit > start)
 473		start = bitbit;
 474	return find_next_zero_bit(fdt->open_fds, maxfd, start);
 475}
 476
 477/*
 478 * allocate a file descriptor, mark it busy.
 479 */
 480int __alloc_fd(struct files_struct *files,
 481	       unsigned start, unsigned end, unsigned flags)
 482{
 483	unsigned int fd;
 484	int error;
 485	struct fdtable *fdt;
 486
 487	spin_lock(&files->file_lock);
 488repeat:
 489	fdt = files_fdtable(files);
 490	fd = start;
 491	if (fd < files->next_fd)
 492		fd = files->next_fd;
 493
 494	if (fd < fdt->max_fds)
 495		fd = find_next_fd(fdt, fd);
 496
 497	/*
 498	 * N.B. For clone tasks sharing a files structure, this test
 499	 * will limit the total number of files that can be opened.
 500	 */
 501	error = -EMFILE;
 502	if (fd >= end)
 503		goto out;
 504
 505	error = expand_files(files, fd);
 506	if (error < 0)
 507		goto out;
 508
 509	/*
 510	 * If we needed to expand the fs array we
 511	 * might have blocked - try again.
 512	 */
 513	if (error)
 514		goto repeat;
 515
 516	if (start <= files->next_fd)
 517		files->next_fd = fd + 1;
 518
 519	__set_open_fd(fd, fdt);
 520	if (flags & O_CLOEXEC)
 521		__set_close_on_exec(fd, fdt);
 522	else
 523		__clear_close_on_exec(fd, fdt);
 524	error = fd;
 525#if 1
 526	/* Sanity check */
 527	if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
 528		printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
 529		rcu_assign_pointer(fdt->fd[fd], NULL);
 530	}
 531#endif
 532
 533out:
 534	spin_unlock(&files->file_lock);
 535	return error;
 536}
 537
 538static int alloc_fd(unsigned start, unsigned flags)
 539{
 540	return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
 541}
 542
 543int get_unused_fd_flags(unsigned flags)
 544{
 545	return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
 546}
 547EXPORT_SYMBOL(get_unused_fd_flags);
 548
 549static void __put_unused_fd(struct files_struct *files, unsigned int fd)
 550{
 551	struct fdtable *fdt = files_fdtable(files);
 552	__clear_open_fd(fd, fdt);
 553	if (fd < files->next_fd)
 554		files->next_fd = fd;
 555}
 556
 557void put_unused_fd(unsigned int fd)
 558{
 559	struct files_struct *files = current->files;
 560	spin_lock(&files->file_lock);
 561	__put_unused_fd(files, fd);
 562	spin_unlock(&files->file_lock);
 563}
 564
 565EXPORT_SYMBOL(put_unused_fd);
 566
 567/*
 568 * Install a file pointer in the fd array.
 569 *
 570 * The VFS is full of places where we drop the files lock between
 571 * setting the open_fds bitmap and installing the file in the file
 572 * array.  At any such point, we are vulnerable to a dup2() race
 573 * installing a file in the array before us.  We need to detect this and
 574 * fput() the struct file we are about to overwrite in this case.
 575 *
 576 * It should never happen - if we allow dup2() do it, _really_ bad things
 577 * will follow.
 578 *
 579 * NOTE: __fd_install() variant is really, really low-level; don't
 580 * use it unless you are forced to by truly lousy API shoved down
 581 * your throat.  'files' *MUST* be either current->files or obtained
 582 * by get_files_struct(current) done by whoever had given it to you,
 583 * or really bad things will happen.  Normally you want to use
 584 * fd_install() instead.
 585 */
 586
 587void __fd_install(struct files_struct *files, unsigned int fd,
 588		struct file *file)
 589{
 590	struct fdtable *fdt;
 591
 592	rcu_read_lock_sched();
 593
 594	if (unlikely(files->resize_in_progress)) {
 595		rcu_read_unlock_sched();
 596		spin_lock(&files->file_lock);
 597		fdt = files_fdtable(files);
 598		BUG_ON(fdt->fd[fd] != NULL);
 599		rcu_assign_pointer(fdt->fd[fd], file);
 600		spin_unlock(&files->file_lock);
 601		return;
 602	}
 603	/* coupled with smp_wmb() in expand_fdtable() */
 604	smp_rmb();
 605	fdt = rcu_dereference_sched(files->fdt);
 606	BUG_ON(fdt->fd[fd] != NULL);
 607	rcu_assign_pointer(fdt->fd[fd], file);
 608	rcu_read_unlock_sched();
 609}
 610
 611void fd_install(unsigned int fd, struct file *file)
 612{
 613	__fd_install(current->files, fd, file);
 614}
 615
 616EXPORT_SYMBOL(fd_install);
 617
 618/*
 619 * The same warnings as for __alloc_fd()/__fd_install() apply here...
 620 */
 621int __close_fd(struct files_struct *files, unsigned fd)
 622{
 623	struct file *file;
 624	struct fdtable *fdt;
 625
 626	spin_lock(&files->file_lock);
 627	fdt = files_fdtable(files);
 628	if (fd >= fdt->max_fds)
 629		goto out_unlock;
 630	file = fdt->fd[fd];
 631	if (!file)
 632		goto out_unlock;
 633	rcu_assign_pointer(fdt->fd[fd], NULL);
 
 634	__put_unused_fd(files, fd);
 635	spin_unlock(&files->file_lock);
 636	return filp_close(file, files);
 637
 638out_unlock:
 639	spin_unlock(&files->file_lock);
 640	return -EBADF;
 641}
 642EXPORT_SYMBOL(__close_fd); /* for ksys_close() */
 643
 644/*
 645 * variant of __close_fd that gets a ref on the file for later fput
 646 */
 647int __close_fd_get_file(unsigned int fd, struct file **res)
 648{
 649	struct files_struct *files = current->files;
 650	struct file *file;
 651	struct fdtable *fdt;
 652
 653	spin_lock(&files->file_lock);
 654	fdt = files_fdtable(files);
 655	if (fd >= fdt->max_fds)
 656		goto out_unlock;
 657	file = fdt->fd[fd];
 658	if (!file)
 659		goto out_unlock;
 660	rcu_assign_pointer(fdt->fd[fd], NULL);
 661	__put_unused_fd(files, fd);
 662	spin_unlock(&files->file_lock);
 663	get_file(file);
 664	*res = file;
 665	return filp_close(file, files);
 666
 667out_unlock:
 668	spin_unlock(&files->file_lock);
 669	*res = NULL;
 670	return -ENOENT;
 671}
 672
 673void do_close_on_exec(struct files_struct *files)
 674{
 675	unsigned i;
 676	struct fdtable *fdt;
 677
 678	/* exec unshares first */
 679	spin_lock(&files->file_lock);
 680	for (i = 0; ; i++) {
 681		unsigned long set;
 682		unsigned fd = i * BITS_PER_LONG;
 683		fdt = files_fdtable(files);
 684		if (fd >= fdt->max_fds)
 685			break;
 686		set = fdt->close_on_exec[i];
 687		if (!set)
 688			continue;
 689		fdt->close_on_exec[i] = 0;
 690		for ( ; set ; fd++, set >>= 1) {
 691			struct file *file;
 692			if (!(set & 1))
 693				continue;
 694			file = fdt->fd[fd];
 695			if (!file)
 696				continue;
 697			rcu_assign_pointer(fdt->fd[fd], NULL);
 698			__put_unused_fd(files, fd);
 699			spin_unlock(&files->file_lock);
 700			filp_close(file, files);
 701			cond_resched();
 702			spin_lock(&files->file_lock);
 703		}
 704
 705	}
 706	spin_unlock(&files->file_lock);
 707}
 708
 709static struct file *__fget(unsigned int fd, fmode_t mask, unsigned int refs)
 710{
 711	struct files_struct *files = current->files;
 712	struct file *file;
 713
 714	rcu_read_lock();
 715loop:
 716	file = fcheck_files(files, fd);
 717	if (file) {
 718		/* File object ref couldn't be taken.
 719		 * dup2() atomicity guarantee is the reason
 720		 * we loop to catch the new file (or NULL pointer)
 721		 */
 722		if (file->f_mode & mask)
 723			file = NULL;
 724		else if (!get_file_rcu_many(file, refs))
 725			goto loop;
 726	}
 727	rcu_read_unlock();
 728
 729	return file;
 730}
 731
 732struct file *fget_many(unsigned int fd, unsigned int refs)
 733{
 734	return __fget(fd, FMODE_PATH, refs);
 735}
 736
 737struct file *fget(unsigned int fd)
 738{
 739	return __fget(fd, FMODE_PATH, 1);
 740}
 741EXPORT_SYMBOL(fget);
 742
 743struct file *fget_raw(unsigned int fd)
 744{
 745	return __fget(fd, 0, 1);
 746}
 747EXPORT_SYMBOL(fget_raw);
 748
 749/*
 750 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
 751 *
 752 * You can use this instead of fget if you satisfy all of the following
 753 * conditions:
 754 * 1) You must call fput_light before exiting the syscall and returning control
 755 *    to userspace (i.e. you cannot remember the returned struct file * after
 756 *    returning to userspace).
 757 * 2) You must not call filp_close on the returned struct file * in between
 758 *    calls to fget_light and fput_light.
 759 * 3) You must not clone the current task in between the calls to fget_light
 760 *    and fput_light.
 761 *
 762 * The fput_needed flag returned by fget_light should be passed to the
 763 * corresponding fput_light.
 764 */
 765static unsigned long __fget_light(unsigned int fd, fmode_t mask)
 766{
 767	struct files_struct *files = current->files;
 768	struct file *file;
 769
 770	if (atomic_read(&files->count) == 1) {
 771		file = __fcheck_files(files, fd);
 772		if (!file || unlikely(file->f_mode & mask))
 773			return 0;
 774		return (unsigned long)file;
 775	} else {
 776		file = __fget(fd, mask, 1);
 777		if (!file)
 778			return 0;
 779		return FDPUT_FPUT | (unsigned long)file;
 780	}
 781}
 782unsigned long __fdget(unsigned int fd)
 783{
 784	return __fget_light(fd, FMODE_PATH);
 785}
 786EXPORT_SYMBOL(__fdget);
 787
 788unsigned long __fdget_raw(unsigned int fd)
 789{
 790	return __fget_light(fd, 0);
 791}
 792
 793unsigned long __fdget_pos(unsigned int fd)
 794{
 795	unsigned long v = __fdget(fd);
 796	struct file *file = (struct file *)(v & ~3);
 797
 798	if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
 799		if (file_count(file) > 1) {
 800			v |= FDPUT_POS_UNLOCK;
 801			mutex_lock(&file->f_pos_lock);
 802		}
 803	}
 804	return v;
 805}
 806
 807void __f_unlock_pos(struct file *f)
 808{
 809	mutex_unlock(&f->f_pos_lock);
 810}
 811
 812/*
 813 * We only lock f_pos if we have threads or if the file might be
 814 * shared with another process. In both cases we'll have an elevated
 815 * file count (done either by fdget() or by fork()).
 816 */
 817
 818void set_close_on_exec(unsigned int fd, int flag)
 819{
 820	struct files_struct *files = current->files;
 821	struct fdtable *fdt;
 822	spin_lock(&files->file_lock);
 823	fdt = files_fdtable(files);
 824	if (flag)
 825		__set_close_on_exec(fd, fdt);
 826	else
 827		__clear_close_on_exec(fd, fdt);
 828	spin_unlock(&files->file_lock);
 829}
 830
 831bool get_close_on_exec(unsigned int fd)
 832{
 833	struct files_struct *files = current->files;
 834	struct fdtable *fdt;
 835	bool res;
 836	rcu_read_lock();
 837	fdt = files_fdtable(files);
 838	res = close_on_exec(fd, fdt);
 839	rcu_read_unlock();
 840	return res;
 841}
 842
 843static int do_dup2(struct files_struct *files,
 844	struct file *file, unsigned fd, unsigned flags)
 845__releases(&files->file_lock)
 846{
 847	struct file *tofree;
 848	struct fdtable *fdt;
 849
 850	/*
 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.
 863	 */
 864	fdt = files_fdtable(files);
 865	tofree = fdt->fd[fd];
 866	if (!tofree && fd_is_open(fd, fdt))
 867		goto Ebusy;
 868	get_file(file);
 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);
 873	else
 874		__clear_close_on_exec(fd, fdt);
 875	spin_unlock(&files->file_lock);
 876
 877	if (tofree)
 878		filp_close(tofree, files);
 879
 880	return fd;
 881
 882Ebusy:
 883	spin_unlock(&files->file_lock);
 884	return -EBUSY;
 885}
 886
 887int replace_fd(unsigned fd, struct file *file, unsigned flags)
 888{
 889	int err;
 890	struct files_struct *files = current->files;
 891
 892	if (!file)
 893		return __close_fd(files, fd);
 894
 895	if (fd >= rlimit(RLIMIT_NOFILE))
 896		return -EBADF;
 897
 898	spin_lock(&files->file_lock);
 899	err = expand_files(files, fd);
 900	if (unlikely(err < 0))
 901		goto out_unlock;
 902	return do_dup2(files, file, fd, flags);
 903
 904out_unlock:
 905	spin_unlock(&files->file_lock);
 906	return err;
 907}
 908
 909static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
 910{
 911	int err = -EBADF;
 912	struct file *file;
 913	struct files_struct *files = current->files;
 914
 915	if ((flags & ~O_CLOEXEC) != 0)
 916		return -EINVAL;
 917
 918	if (unlikely(oldfd == newfd))
 919		return -EINVAL;
 920
 921	if (newfd >= rlimit(RLIMIT_NOFILE))
 922		return -EBADF;
 923
 924	spin_lock(&files->file_lock);
 925	err = expand_files(files, newfd);
 926	file = fcheck(oldfd);
 927	if (unlikely(!file))
 928		goto Ebadf;
 929	if (unlikely(err < 0)) {
 930		if (err == -EMFILE)
 931			goto Ebadf;
 932		goto out_unlock;
 933	}
 934	return do_dup2(files, file, newfd, flags);
 935
 936Ebadf:
 937	err = -EBADF;
 938out_unlock:
 939	spin_unlock(&files->file_lock);
 940	return err;
 941}
 942
 943SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
 944{
 945	return ksys_dup3(oldfd, newfd, flags);
 946}
 947
 948SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
 949{
 950	if (unlikely(newfd == oldfd)) { /* corner case */
 951		struct files_struct *files = current->files;
 952		int retval = oldfd;
 953
 954		rcu_read_lock();
 955		if (!fcheck_files(files, oldfd))
 956			retval = -EBADF;
 957		rcu_read_unlock();
 958		return retval;
 959	}
 960	return ksys_dup3(oldfd, newfd, 0);
 961}
 962
 963int ksys_dup(unsigned int fildes)
 964{
 965	int ret = -EBADF;
 966	struct file *file = fget_raw(fildes);
 967
 968	if (file) {
 969		ret = get_unused_fd_flags(0);
 970		if (ret >= 0)
 971			fd_install(ret, file);
 972		else
 973			fput(file);
 974	}
 975	return ret;
 976}
 977
 978SYSCALL_DEFINE1(dup, unsigned int, fildes)
 979{
 980	return ksys_dup(fildes);
 981}
 982
 983int f_dupfd(unsigned int from, struct file *file, unsigned flags)
 984{
 985	int err;
 986	if (from >= rlimit(RLIMIT_NOFILE))
 987		return -EINVAL;
 988	err = alloc_fd(from, flags);
 989	if (err >= 0) {
 990		get_file(file);
 991		fd_install(err, file);
 992	}
 993	return err;
 994}
 995
 996int iterate_fd(struct files_struct *files, unsigned n,
 997		int (*f)(const void *, struct file *, unsigned),
 998		const void *p)
 999{
1000	struct fdtable *fdt;
1001	int res = 0;
1002	if (!files)
1003		return 0;
1004	spin_lock(&files->file_lock);
1005	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1006		struct file *file;
1007		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1008		if (!file)
1009			continue;
1010		res = f(p, file, n);
1011		if (res)
1012			break;
1013	}
1014	spin_unlock(&files->file_lock);
1015	return res;
1016}
1017EXPORT_SYMBOL(iterate_fd);