xref: /openbmc/linux/fs/file.c (revision 60630924bb5af8751adcecc896e7763c3783ca89)
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/kernel.h>
14  #include <linux/mm.h>
15  #include <linux/sched/signal.h>
16  #include <linux/slab.h>
17  #include <linux/file.h>
18  #include <linux/fdtable.h>
19  #include <linux/bitops.h>
20  #include <linux/spinlock.h>
21  #include <linux/rcupdate.h>
22  #include <linux/close_range.h>
23  #include <net/sock.h>
24  
25  #include "internal.h"
26  
27  unsigned int sysctl_nr_open __read_mostly = 1024*1024;
28  unsigned int sysctl_nr_open_min = BITS_PER_LONG;
29  /* our min() is unusable in constant expressions ;-/ */
30  #define __const_min(x, y) ((x) < (y) ? (x) : (y))
31  unsigned int sysctl_nr_open_max =
32  	__const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
33  
34  static void __free_fdtable(struct fdtable *fdt)
35  {
36  	kvfree(fdt->fd);
37  	kvfree(fdt->open_fds);
38  	kfree(fdt);
39  }
40  
41  static void free_fdtable_rcu(struct rcu_head *rcu)
42  {
43  	__free_fdtable(container_of(rcu, struct fdtable, rcu));
44  }
45  
46  #define BITBIT_NR(nr)	BITS_TO_LONGS(BITS_TO_LONGS(nr))
47  #define BITBIT_SIZE(nr)	(BITBIT_NR(nr) * sizeof(long))
48  
49  /*
50   * Copy 'count' fd bits from the old table to the new table and clear the extra
51   * space if any.  This does not copy the file pointers.  Called with the files
52   * spinlock held for write.
53   */
54  static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
55  			    unsigned int count)
56  {
57  	unsigned int cpy, set;
58  
59  	cpy = count / BITS_PER_BYTE;
60  	set = (nfdt->max_fds - count) / BITS_PER_BYTE;
61  	memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
62  	memset((char *)nfdt->open_fds + cpy, 0, set);
63  	memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
64  	memset((char *)nfdt->close_on_exec + cpy, 0, set);
65  
66  	cpy = BITBIT_SIZE(count);
67  	set = BITBIT_SIZE(nfdt->max_fds) - cpy;
68  	memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
69  	memset((char *)nfdt->full_fds_bits + cpy, 0, set);
70  }
71  
72  /*
73   * Copy all file descriptors from the old table to the new, expanded table and
74   * clear the extra space.  Called with the files spinlock held for write.
75   */
76  static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
77  {
78  	size_t cpy, set;
79  
80  	BUG_ON(nfdt->max_fds < ofdt->max_fds);
81  
82  	cpy = ofdt->max_fds * sizeof(struct file *);
83  	set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
84  	memcpy(nfdt->fd, ofdt->fd, cpy);
85  	memset((char *)nfdt->fd + cpy, 0, set);
86  
87  	copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
88  }
89  
90  static struct fdtable * alloc_fdtable(unsigned int nr)
91  {
92  	struct fdtable *fdt;
93  	void *data;
94  
95  	/*
96  	 * Figure out how many fds we actually want to support in this fdtable.
97  	 * Allocation steps are keyed to the size of the fdarray, since it
98  	 * grows far faster than any of the other dynamic data. We try to fit
99  	 * the fdarray into comfortable page-tuned chunks: starting at 1024B
100  	 * and growing in powers of two from there on.
101  	 */
102  	nr /= (1024 / sizeof(struct file *));
103  	nr = roundup_pow_of_two(nr + 1);
104  	nr *= (1024 / sizeof(struct file *));
105  	/*
106  	 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
107  	 * had been set lower between the check in expand_files() and here.  Deal
108  	 * with that in caller, it's cheaper that way.
109  	 *
110  	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
111  	 * bitmaps handling below becomes unpleasant, to put it mildly...
112  	 */
113  	if (unlikely(nr > sysctl_nr_open))
114  		nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
115  
116  	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
117  	if (!fdt)
118  		goto out;
119  	fdt->max_fds = nr;
120  	data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
121  	if (!data)
122  		goto out_fdt;
123  	fdt->fd = data;
124  
125  	data = kvmalloc(max_t(size_t,
126  				 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
127  				 GFP_KERNEL_ACCOUNT);
128  	if (!data)
129  		goto out_arr;
130  	fdt->open_fds = data;
131  	data += nr / BITS_PER_BYTE;
132  	fdt->close_on_exec = data;
133  	data += nr / BITS_PER_BYTE;
134  	fdt->full_fds_bits = data;
135  
136  	return fdt;
137  
138  out_arr:
139  	kvfree(fdt->fd);
140  out_fdt:
141  	kfree(fdt);
142  out:
143  	return NULL;
144  }
145  
146  /*
147   * Expand the file descriptor table.
148   * This function will allocate a new fdtable and both fd array and fdset, of
149   * the given size.
150   * Return <0 error code on error; 1 on successful completion.
151   * The files->file_lock should be held on entry, and will be held on exit.
152   */
153  static int expand_fdtable(struct files_struct *files, unsigned int nr)
154  	__releases(files->file_lock)
155  	__acquires(files->file_lock)
156  {
157  	struct fdtable *new_fdt, *cur_fdt;
158  
159  	spin_unlock(&files->file_lock);
160  	new_fdt = alloc_fdtable(nr);
161  
162  	/* make sure all fd_install() have seen resize_in_progress
163  	 * or have finished their rcu_read_lock_sched() section.
164  	 */
165  	if (atomic_read(&files->count) > 1)
166  		synchronize_rcu();
167  
168  	spin_lock(&files->file_lock);
169  	if (!new_fdt)
170  		return -ENOMEM;
171  	/*
172  	 * extremely unlikely race - sysctl_nr_open decreased between the check in
173  	 * caller and alloc_fdtable().  Cheaper to catch it here...
174  	 */
175  	if (unlikely(new_fdt->max_fds <= nr)) {
176  		__free_fdtable(new_fdt);
177  		return -EMFILE;
178  	}
179  	cur_fdt = files_fdtable(files);
180  	BUG_ON(nr < cur_fdt->max_fds);
181  	copy_fdtable(new_fdt, cur_fdt);
182  	rcu_assign_pointer(files->fdt, new_fdt);
183  	if (cur_fdt != &files->fdtab)
184  		call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
185  	/* coupled with smp_rmb() in fd_install() */
186  	smp_wmb();
187  	return 1;
188  }
189  
190  /*
191   * Expand files.
192   * This function will expand the file structures, if the requested size exceeds
193   * the current capacity and there is room for expansion.
194   * Return <0 error code on error; 0 when nothing done; 1 when files were
195   * expanded and execution may have blocked.
196   * The files->file_lock should be held on entry, and will be held on exit.
197   */
198  static int expand_files(struct files_struct *files, unsigned int nr)
199  	__releases(files->file_lock)
200  	__acquires(files->file_lock)
201  {
202  	struct fdtable *fdt;
203  	int expanded = 0;
204  
205  repeat:
206  	fdt = files_fdtable(files);
207  
208  	/* Do we need to expand? */
209  	if (nr < fdt->max_fds)
210  		return expanded;
211  
212  	/* Can we expand? */
213  	if (nr >= sysctl_nr_open)
214  		return -EMFILE;
215  
216  	if (unlikely(files->resize_in_progress)) {
217  		spin_unlock(&files->file_lock);
218  		expanded = 1;
219  		wait_event(files->resize_wait, !files->resize_in_progress);
220  		spin_lock(&files->file_lock);
221  		goto repeat;
222  	}
223  
224  	/* All good, so we try */
225  	files->resize_in_progress = true;
226  	expanded = expand_fdtable(files, nr);
227  	files->resize_in_progress = false;
228  
229  	wake_up_all(&files->resize_wait);
230  	return expanded;
231  }
232  
233  static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
234  {
235  	__set_bit(fd, fdt->close_on_exec);
236  }
237  
238  static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
239  {
240  	if (test_bit(fd, fdt->close_on_exec))
241  		__clear_bit(fd, fdt->close_on_exec);
242  }
243  
244  static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
245  {
246  	__set_bit(fd, fdt->open_fds);
247  	fd /= BITS_PER_LONG;
248  	if (!~fdt->open_fds[fd])
249  		__set_bit(fd, fdt->full_fds_bits);
250  }
251  
252  static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
253  {
254  	__clear_bit(fd, fdt->open_fds);
255  	__clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
256  }
257  
258  static unsigned int count_open_files(struct fdtable *fdt)
259  {
260  	unsigned int size = fdt->max_fds;
261  	unsigned int i;
262  
263  	/* Find the last open fd */
264  	for (i = size / BITS_PER_LONG; i > 0; ) {
265  		if (fdt->open_fds[--i])
266  			break;
267  	}
268  	i = (i + 1) * BITS_PER_LONG;
269  	return i;
270  }
271  
272  static unsigned int sane_fdtable_size(struct fdtable *fdt, unsigned int max_fds)
273  {
274  	unsigned int count;
275  
276  	count = count_open_files(fdt);
277  	if (max_fds < NR_OPEN_DEFAULT)
278  		max_fds = NR_OPEN_DEFAULT;
279  	return min(count, max_fds);
280  }
281  
282  /*
283   * Allocate a new files structure and copy contents from the
284   * passed in files structure.
285   * errorp will be valid only when the returned files_struct is NULL.
286   */
287  struct files_struct *dup_fd(struct files_struct *oldf, unsigned int max_fds, int *errorp)
288  {
289  	struct files_struct *newf;
290  	struct file **old_fds, **new_fds;
291  	unsigned int open_files, i;
292  	struct fdtable *old_fdt, *new_fdt;
293  
294  	*errorp = -ENOMEM;
295  	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
296  	if (!newf)
297  		goto out;
298  
299  	atomic_set(&newf->count, 1);
300  
301  	spin_lock_init(&newf->file_lock);
302  	newf->resize_in_progress = false;
303  	init_waitqueue_head(&newf->resize_wait);
304  	newf->next_fd = 0;
305  	new_fdt = &newf->fdtab;
306  	new_fdt->max_fds = NR_OPEN_DEFAULT;
307  	new_fdt->close_on_exec = newf->close_on_exec_init;
308  	new_fdt->open_fds = newf->open_fds_init;
309  	new_fdt->full_fds_bits = newf->full_fds_bits_init;
310  	new_fdt->fd = &newf->fd_array[0];
311  
312  	spin_lock(&oldf->file_lock);
313  	old_fdt = files_fdtable(oldf);
314  	open_files = sane_fdtable_size(old_fdt, max_fds);
315  
316  	/*
317  	 * Check whether we need to allocate a larger fd array and fd set.
318  	 */
319  	while (unlikely(open_files > new_fdt->max_fds)) {
320  		spin_unlock(&oldf->file_lock);
321  
322  		if (new_fdt != &newf->fdtab)
323  			__free_fdtable(new_fdt);
324  
325  		new_fdt = alloc_fdtable(open_files - 1);
326  		if (!new_fdt) {
327  			*errorp = -ENOMEM;
328  			goto out_release;
329  		}
330  
331  		/* beyond sysctl_nr_open; nothing to do */
332  		if (unlikely(new_fdt->max_fds < open_files)) {
333  			__free_fdtable(new_fdt);
334  			*errorp = -EMFILE;
335  			goto out_release;
336  		}
337  
338  		/*
339  		 * Reacquire the oldf lock and a pointer to its fd table
340  		 * who knows it may have a new bigger fd table. We need
341  		 * the latest pointer.
342  		 */
343  		spin_lock(&oldf->file_lock);
344  		old_fdt = files_fdtable(oldf);
345  		open_files = sane_fdtable_size(old_fdt, max_fds);
346  	}
347  
348  	copy_fd_bitmaps(new_fdt, old_fdt, open_files);
349  
350  	old_fds = old_fdt->fd;
351  	new_fds = new_fdt->fd;
352  
353  	for (i = open_files; i != 0; i--) {
354  		struct file *f = *old_fds++;
355  		if (f) {
356  			get_file(f);
357  		} else {
358  			/*
359  			 * The fd may be claimed in the fd bitmap but not yet
360  			 * instantiated in the files array if a sibling thread
361  			 * is partway through open().  So make sure that this
362  			 * fd is available to the new process.
363  			 */
364  			__clear_open_fd(open_files - i, new_fdt);
365  		}
366  		rcu_assign_pointer(*new_fds++, f);
367  	}
368  	spin_unlock(&oldf->file_lock);
369  
370  	/* clear the remainder */
371  	memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
372  
373  	rcu_assign_pointer(newf->fdt, new_fdt);
374  
375  	return newf;
376  
377  out_release:
378  	kmem_cache_free(files_cachep, newf);
379  out:
380  	return NULL;
381  }
382  
383  static struct fdtable *close_files(struct files_struct * files)
384  {
385  	/*
386  	 * It is safe to dereference the fd table without RCU or
387  	 * ->file_lock because this is the last reference to the
388  	 * files structure.
389  	 */
390  	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
391  	unsigned int i, j = 0;
392  
393  	for (;;) {
394  		unsigned long set;
395  		i = j * BITS_PER_LONG;
396  		if (i >= fdt->max_fds)
397  			break;
398  		set = fdt->open_fds[j++];
399  		while (set) {
400  			if (set & 1) {
401  				struct file * file = xchg(&fdt->fd[i], NULL);
402  				if (file) {
403  					filp_close(file, files);
404  					cond_resched();
405  				}
406  			}
407  			i++;
408  			set >>= 1;
409  		}
410  	}
411  
412  	return fdt;
413  }
414  
415  void put_files_struct(struct files_struct *files)
416  {
417  	if (atomic_dec_and_test(&files->count)) {
418  		struct fdtable *fdt = close_files(files);
419  
420  		/* free the arrays if they are not embedded */
421  		if (fdt != &files->fdtab)
422  			__free_fdtable(fdt);
423  		kmem_cache_free(files_cachep, files);
424  	}
425  }
426  
427  void exit_files(struct task_struct *tsk)
428  {
429  	struct files_struct * files = tsk->files;
430  
431  	if (files) {
432  		task_lock(tsk);
433  		tsk->files = NULL;
434  		task_unlock(tsk);
435  		put_files_struct(files);
436  	}
437  }
438  
439  struct files_struct init_files = {
440  	.count		= ATOMIC_INIT(1),
441  	.fdt		= &init_files.fdtab,
442  	.fdtab		= {
443  		.max_fds	= NR_OPEN_DEFAULT,
444  		.fd		= &init_files.fd_array[0],
445  		.close_on_exec	= init_files.close_on_exec_init,
446  		.open_fds	= init_files.open_fds_init,
447  		.full_fds_bits	= init_files.full_fds_bits_init,
448  	},
449  	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
450  	.resize_wait	= __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
451  };
452  
453  static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
454  {
455  	unsigned int maxfd = fdt->max_fds;
456  	unsigned int maxbit = maxfd / BITS_PER_LONG;
457  	unsigned int bitbit = start / BITS_PER_LONG;
458  
459  	bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
460  	if (bitbit > maxfd)
461  		return maxfd;
462  	if (bitbit > start)
463  		start = bitbit;
464  	return find_next_zero_bit(fdt->open_fds, maxfd, start);
465  }
466  
467  /*
468   * allocate a file descriptor, mark it busy.
469   */
470  static int alloc_fd(unsigned start, unsigned end, unsigned flags)
471  {
472  	struct files_struct *files = current->files;
473  	unsigned int fd;
474  	int error;
475  	struct fdtable *fdt;
476  
477  	spin_lock(&files->file_lock);
478  repeat:
479  	fdt = files_fdtable(files);
480  	fd = start;
481  	if (fd < files->next_fd)
482  		fd = files->next_fd;
483  
484  	if (fd < fdt->max_fds)
485  		fd = find_next_fd(fdt, fd);
486  
487  	/*
488  	 * N.B. For clone tasks sharing a files structure, this test
489  	 * will limit the total number of files that can be opened.
490  	 */
491  	error = -EMFILE;
492  	if (fd >= end)
493  		goto out;
494  
495  	error = expand_files(files, fd);
496  	if (error < 0)
497  		goto out;
498  
499  	/*
500  	 * If we needed to expand the fs array we
501  	 * might have blocked - try again.
502  	 */
503  	if (error)
504  		goto repeat;
505  
506  	if (start <= files->next_fd)
507  		files->next_fd = fd + 1;
508  
509  	__set_open_fd(fd, fdt);
510  	if (flags & O_CLOEXEC)
511  		__set_close_on_exec(fd, fdt);
512  	else
513  		__clear_close_on_exec(fd, fdt);
514  	error = fd;
515  #if 1
516  	/* Sanity check */
517  	if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
518  		printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
519  		rcu_assign_pointer(fdt->fd[fd], NULL);
520  	}
521  #endif
522  
523  out:
524  	spin_unlock(&files->file_lock);
525  	return error;
526  }
527  
528  int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
529  {
530  	return alloc_fd(0, nofile, flags);
531  }
532  
533  int get_unused_fd_flags(unsigned flags)
534  {
535  	return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
536  }
537  EXPORT_SYMBOL(get_unused_fd_flags);
538  
539  static void __put_unused_fd(struct files_struct *files, unsigned int fd)
540  {
541  	struct fdtable *fdt = files_fdtable(files);
542  	__clear_open_fd(fd, fdt);
543  	if (fd < files->next_fd)
544  		files->next_fd = fd;
545  }
546  
547  void put_unused_fd(unsigned int fd)
548  {
549  	struct files_struct *files = current->files;
550  	spin_lock(&files->file_lock);
551  	__put_unused_fd(files, fd);
552  	spin_unlock(&files->file_lock);
553  }
554  
555  EXPORT_SYMBOL(put_unused_fd);
556  
557  /*
558   * Install a file pointer in the fd array.
559   *
560   * The VFS is full of places where we drop the files lock between
561   * setting the open_fds bitmap and installing the file in the file
562   * array.  At any such point, we are vulnerable to a dup2() race
563   * installing a file in the array before us.  We need to detect this and
564   * fput() the struct file we are about to overwrite in this case.
565   *
566   * It should never happen - if we allow dup2() do it, _really_ bad things
567   * will follow.
568   *
569   * This consumes the "file" refcount, so callers should treat it
570   * as if they had called fput(file).
571   */
572  
573  void fd_install(unsigned int fd, struct file *file)
574  {
575  	struct files_struct *files = current->files;
576  	struct fdtable *fdt;
577  
578  	rcu_read_lock_sched();
579  
580  	if (unlikely(files->resize_in_progress)) {
581  		rcu_read_unlock_sched();
582  		spin_lock(&files->file_lock);
583  		fdt = files_fdtable(files);
584  		BUG_ON(fdt->fd[fd] != NULL);
585  		rcu_assign_pointer(fdt->fd[fd], file);
586  		spin_unlock(&files->file_lock);
587  		return;
588  	}
589  	/* coupled with smp_wmb() in expand_fdtable() */
590  	smp_rmb();
591  	fdt = rcu_dereference_sched(files->fdt);
592  	BUG_ON(fdt->fd[fd] != NULL);
593  	rcu_assign_pointer(fdt->fd[fd], file);
594  	rcu_read_unlock_sched();
595  }
596  
597  EXPORT_SYMBOL(fd_install);
598  
599  /**
600   * pick_file - return file associatd with fd
601   * @files: file struct to retrieve file from
602   * @fd: file descriptor to retrieve file for
603   *
604   * If this functions returns an EINVAL error pointer the fd was beyond the
605   * current maximum number of file descriptors for that fdtable.
606   *
607   * Returns: The file associated with @fd, on error returns an error pointer.
608   */
609  static struct file *pick_file(struct files_struct *files, unsigned fd)
610  {
611  	struct file *file;
612  	struct fdtable *fdt;
613  
614  	spin_lock(&files->file_lock);
615  	fdt = files_fdtable(files);
616  	if (fd >= fdt->max_fds) {
617  		file = ERR_PTR(-EINVAL);
618  		goto out_unlock;
619  	}
620  	file = fdt->fd[fd];
621  	if (!file) {
622  		file = ERR_PTR(-EBADF);
623  		goto out_unlock;
624  	}
625  	rcu_assign_pointer(fdt->fd[fd], NULL);
626  	__put_unused_fd(files, fd);
627  
628  out_unlock:
629  	spin_unlock(&files->file_lock);
630  	return file;
631  }
632  
633  int close_fd(unsigned fd)
634  {
635  	struct files_struct *files = current->files;
636  	struct file *file;
637  
638  	file = pick_file(files, fd);
639  	if (IS_ERR(file))
640  		return -EBADF;
641  
642  	return filp_close(file, files);
643  }
644  EXPORT_SYMBOL(close_fd); /* for ksys_close() */
645  
646  /**
647   * last_fd - return last valid index into fd table
648   * @cur_fds: files struct
649   *
650   * Context: Either rcu read lock or files_lock must be held.
651   *
652   * Returns: Last valid index into fdtable.
653   */
654  static inline unsigned last_fd(struct fdtable *fdt)
655  {
656  	return fdt->max_fds - 1;
657  }
658  
659  static inline void __range_cloexec(struct files_struct *cur_fds,
660  				   unsigned int fd, unsigned int max_fd)
661  {
662  	struct fdtable *fdt;
663  
664  	/* make sure we're using the correct maximum value */
665  	spin_lock(&cur_fds->file_lock);
666  	fdt = files_fdtable(cur_fds);
667  	max_fd = min(last_fd(fdt), max_fd);
668  	if (fd <= max_fd)
669  		bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
670  	spin_unlock(&cur_fds->file_lock);
671  }
672  
673  static inline void __range_close(struct files_struct *cur_fds, unsigned int fd,
674  				 unsigned int max_fd)
675  {
676  	while (fd <= max_fd) {
677  		struct file *file;
678  
679  		file = pick_file(cur_fds, fd++);
680  		if (!IS_ERR(file)) {
681  			/* found a valid file to close */
682  			filp_close(file, cur_fds);
683  			cond_resched();
684  			continue;
685  		}
686  
687  		/* beyond the last fd in that table */
688  		if (PTR_ERR(file) == -EINVAL)
689  			return;
690  	}
691  }
692  
693  /**
694   * __close_range() - Close all file descriptors in a given range.
695   *
696   * @fd:     starting file descriptor to close
697   * @max_fd: last file descriptor to close
698   *
699   * This closes a range of file descriptors. All file descriptors
700   * from @fd up to and including @max_fd are closed.
701   */
702  int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
703  {
704  	struct task_struct *me = current;
705  	struct files_struct *cur_fds = me->files, *fds = NULL;
706  
707  	if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
708  		return -EINVAL;
709  
710  	if (fd > max_fd)
711  		return -EINVAL;
712  
713  	if (flags & CLOSE_RANGE_UNSHARE) {
714  		int ret;
715  		unsigned int max_unshare_fds = NR_OPEN_MAX;
716  
717  		/*
718  		 * If the caller requested all fds to be made cloexec we always
719  		 * copy all of the file descriptors since they still want to
720  		 * use them.
721  		 */
722  		if (!(flags & CLOSE_RANGE_CLOEXEC)) {
723  			/*
724  			 * If the requested range is greater than the current
725  			 * maximum, we're closing everything so only copy all
726  			 * file descriptors beneath the lowest file descriptor.
727  			 */
728  			rcu_read_lock();
729  			if (max_fd >= last_fd(files_fdtable(cur_fds)))
730  				max_unshare_fds = fd;
731  			rcu_read_unlock();
732  		}
733  
734  		ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
735  		if (ret)
736  			return ret;
737  
738  		/*
739  		 * We used to share our file descriptor table, and have now
740  		 * created a private one, make sure we're using it below.
741  		 */
742  		if (fds)
743  			swap(cur_fds, fds);
744  	}
745  
746  	if (flags & CLOSE_RANGE_CLOEXEC)
747  		__range_cloexec(cur_fds, fd, max_fd);
748  	else
749  		__range_close(cur_fds, fd, max_fd);
750  
751  	if (fds) {
752  		/*
753  		 * We're done closing the files we were supposed to. Time to install
754  		 * the new file descriptor table and drop the old one.
755  		 */
756  		task_lock(me);
757  		me->files = cur_fds;
758  		task_unlock(me);
759  		put_files_struct(fds);
760  	}
761  
762  	return 0;
763  }
764  
765  /*
766   * See close_fd_get_file() below, this variant assumes current->files->file_lock
767   * is held.
768   */
769  int __close_fd_get_file(unsigned int fd, struct file **res)
770  {
771  	struct files_struct *files = current->files;
772  	struct file *file;
773  	struct fdtable *fdt;
774  
775  	fdt = files_fdtable(files);
776  	if (fd >= fdt->max_fds)
777  		goto out_err;
778  	file = fdt->fd[fd];
779  	if (!file)
780  		goto out_err;
781  	rcu_assign_pointer(fdt->fd[fd], NULL);
782  	__put_unused_fd(files, fd);
783  	get_file(file);
784  	*res = file;
785  	return 0;
786  out_err:
787  	*res = NULL;
788  	return -ENOENT;
789  }
790  
791  /*
792   * variant of close_fd that gets a ref on the file for later fput.
793   * The caller must ensure that filp_close() called on the file, and then
794   * an fput().
795   */
796  int close_fd_get_file(unsigned int fd, struct file **res)
797  {
798  	struct files_struct *files = current->files;
799  	int ret;
800  
801  	spin_lock(&files->file_lock);
802  	ret = __close_fd_get_file(fd, res);
803  	spin_unlock(&files->file_lock);
804  
805  	return ret;
806  }
807  
808  void do_close_on_exec(struct files_struct *files)
809  {
810  	unsigned i;
811  	struct fdtable *fdt;
812  
813  	/* exec unshares first */
814  	spin_lock(&files->file_lock);
815  	for (i = 0; ; i++) {
816  		unsigned long set;
817  		unsigned fd = i * BITS_PER_LONG;
818  		fdt = files_fdtable(files);
819  		if (fd >= fdt->max_fds)
820  			break;
821  		set = fdt->close_on_exec[i];
822  		if (!set)
823  			continue;
824  		fdt->close_on_exec[i] = 0;
825  		for ( ; set ; fd++, set >>= 1) {
826  			struct file *file;
827  			if (!(set & 1))
828  				continue;
829  			file = fdt->fd[fd];
830  			if (!file)
831  				continue;
832  			rcu_assign_pointer(fdt->fd[fd], NULL);
833  			__put_unused_fd(files, fd);
834  			spin_unlock(&files->file_lock);
835  			filp_close(file, files);
836  			cond_resched();
837  			spin_lock(&files->file_lock);
838  		}
839  
840  	}
841  	spin_unlock(&files->file_lock);
842  }
843  
844  static struct file *__fget_files(struct files_struct *files, unsigned int fd,
845  				 fmode_t mask, unsigned int refs)
846  {
847  	struct file *file;
848  
849  	rcu_read_lock();
850  loop:
851  	file = files_lookup_fd_rcu(files, fd);
852  	if (file) {
853  		/* File object ref couldn't be taken.
854  		 * dup2() atomicity guarantee is the reason
855  		 * we loop to catch the new file (or NULL pointer)
856  		 */
857  		if (file->f_mode & mask)
858  			file = NULL;
859  		else if (!get_file_rcu_many(file, refs))
860  			goto loop;
861  	}
862  	rcu_read_unlock();
863  
864  	return file;
865  }
866  
867  static inline struct file *__fget(unsigned int fd, fmode_t mask,
868  				  unsigned int refs)
869  {
870  	return __fget_files(current->files, fd, mask, refs);
871  }
872  
873  struct file *fget_many(unsigned int fd, unsigned int refs)
874  {
875  	return __fget(fd, FMODE_PATH, refs);
876  }
877  
878  struct file *fget(unsigned int fd)
879  {
880  	return __fget(fd, FMODE_PATH, 1);
881  }
882  EXPORT_SYMBOL(fget);
883  
884  struct file *fget_raw(unsigned int fd)
885  {
886  	return __fget(fd, 0, 1);
887  }
888  EXPORT_SYMBOL(fget_raw);
889  
890  struct file *fget_task(struct task_struct *task, unsigned int fd)
891  {
892  	struct file *file = NULL;
893  
894  	task_lock(task);
895  	if (task->files)
896  		file = __fget_files(task->files, fd, 0, 1);
897  	task_unlock(task);
898  
899  	return file;
900  }
901  
902  struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
903  {
904  	/* Must be called with rcu_read_lock held */
905  	struct files_struct *files;
906  	struct file *file = NULL;
907  
908  	task_lock(task);
909  	files = task->files;
910  	if (files)
911  		file = files_lookup_fd_rcu(files, fd);
912  	task_unlock(task);
913  
914  	return file;
915  }
916  
917  struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret_fd)
918  {
919  	/* Must be called with rcu_read_lock held */
920  	struct files_struct *files;
921  	unsigned int fd = *ret_fd;
922  	struct file *file = NULL;
923  
924  	task_lock(task);
925  	files = task->files;
926  	if (files) {
927  		for (; fd < files_fdtable(files)->max_fds; fd++) {
928  			file = files_lookup_fd_rcu(files, fd);
929  			if (file)
930  				break;
931  		}
932  	}
933  	task_unlock(task);
934  	*ret_fd = fd;
935  	return file;
936  }
937  
938  /*
939   * Lightweight file lookup - no refcnt increment if fd table isn't shared.
940   *
941   * You can use this instead of fget if you satisfy all of the following
942   * conditions:
943   * 1) You must call fput_light before exiting the syscall and returning control
944   *    to userspace (i.e. you cannot remember the returned struct file * after
945   *    returning to userspace).
946   * 2) You must not call filp_close on the returned struct file * in between
947   *    calls to fget_light and fput_light.
948   * 3) You must not clone the current task in between the calls to fget_light
949   *    and fput_light.
950   *
951   * The fput_needed flag returned by fget_light should be passed to the
952   * corresponding fput_light.
953   */
954  static unsigned long __fget_light(unsigned int fd, fmode_t mask)
955  {
956  	struct files_struct *files = current->files;
957  	struct file *file;
958  
959  	if (atomic_read(&files->count) == 1) {
960  		file = files_lookup_fd_raw(files, fd);
961  		if (!file || unlikely(file->f_mode & mask))
962  			return 0;
963  		return (unsigned long)file;
964  	} else {
965  		file = __fget(fd, mask, 1);
966  		if (!file)
967  			return 0;
968  		return FDPUT_FPUT | (unsigned long)file;
969  	}
970  }
971  unsigned long __fdget(unsigned int fd)
972  {
973  	return __fget_light(fd, FMODE_PATH);
974  }
975  EXPORT_SYMBOL(__fdget);
976  
977  unsigned long __fdget_raw(unsigned int fd)
978  {
979  	return __fget_light(fd, 0);
980  }
981  
982  unsigned long __fdget_pos(unsigned int fd)
983  {
984  	unsigned long v = __fdget(fd);
985  	struct file *file = (struct file *)(v & ~3);
986  
987  	if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
988  		if (file_count(file) > 1) {
989  			v |= FDPUT_POS_UNLOCK;
990  			mutex_lock(&file->f_pos_lock);
991  		}
992  	}
993  	return v;
994  }
995  
996  void __f_unlock_pos(struct file *f)
997  {
998  	mutex_unlock(&f->f_pos_lock);
999  }
1000  
1001  /*
1002   * We only lock f_pos if we have threads or if the file might be
1003   * shared with another process. In both cases we'll have an elevated
1004   * file count (done either by fdget() or by fork()).
1005   */
1006  
1007  void set_close_on_exec(unsigned int fd, int flag)
1008  {
1009  	struct files_struct *files = current->files;
1010  	struct fdtable *fdt;
1011  	spin_lock(&files->file_lock);
1012  	fdt = files_fdtable(files);
1013  	if (flag)
1014  		__set_close_on_exec(fd, fdt);
1015  	else
1016  		__clear_close_on_exec(fd, fdt);
1017  	spin_unlock(&files->file_lock);
1018  }
1019  
1020  bool get_close_on_exec(unsigned int fd)
1021  {
1022  	struct files_struct *files = current->files;
1023  	struct fdtable *fdt;
1024  	bool res;
1025  	rcu_read_lock();
1026  	fdt = files_fdtable(files);
1027  	res = close_on_exec(fd, fdt);
1028  	rcu_read_unlock();
1029  	return res;
1030  }
1031  
1032  static int do_dup2(struct files_struct *files,
1033  	struct file *file, unsigned fd, unsigned flags)
1034  __releases(&files->file_lock)
1035  {
1036  	struct file *tofree;
1037  	struct fdtable *fdt;
1038  
1039  	/*
1040  	 * We need to detect attempts to do dup2() over allocated but still
1041  	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
1042  	 * extra work in their equivalent of fget() - they insert struct
1043  	 * file immediately after grabbing descriptor, mark it larval if
1044  	 * more work (e.g. actual opening) is needed and make sure that
1045  	 * fget() treats larval files as absent.  Potentially interesting,
1046  	 * but while extra work in fget() is trivial, locking implications
1047  	 * and amount of surgery on open()-related paths in VFS are not.
1048  	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1049  	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
1050  	 * scope of POSIX or SUS, since neither considers shared descriptor
1051  	 * tables and this condition does not arise without those.
1052  	 */
1053  	fdt = files_fdtable(files);
1054  	tofree = fdt->fd[fd];
1055  	if (!tofree && fd_is_open(fd, fdt))
1056  		goto Ebusy;
1057  	get_file(file);
1058  	rcu_assign_pointer(fdt->fd[fd], file);
1059  	__set_open_fd(fd, fdt);
1060  	if (flags & O_CLOEXEC)
1061  		__set_close_on_exec(fd, fdt);
1062  	else
1063  		__clear_close_on_exec(fd, fdt);
1064  	spin_unlock(&files->file_lock);
1065  
1066  	if (tofree)
1067  		filp_close(tofree, files);
1068  
1069  	return fd;
1070  
1071  Ebusy:
1072  	spin_unlock(&files->file_lock);
1073  	return -EBUSY;
1074  }
1075  
1076  int replace_fd(unsigned fd, struct file *file, unsigned flags)
1077  {
1078  	int err;
1079  	struct files_struct *files = current->files;
1080  
1081  	if (!file)
1082  		return close_fd(fd);
1083  
1084  	if (fd >= rlimit(RLIMIT_NOFILE))
1085  		return -EBADF;
1086  
1087  	spin_lock(&files->file_lock);
1088  	err = expand_files(files, fd);
1089  	if (unlikely(err < 0))
1090  		goto out_unlock;
1091  	return do_dup2(files, file, fd, flags);
1092  
1093  out_unlock:
1094  	spin_unlock(&files->file_lock);
1095  	return err;
1096  }
1097  
1098  /**
1099   * __receive_fd() - Install received file into file descriptor table
1100   * @file: struct file that was received from another process
1101   * @ufd: __user pointer to write new fd number to
1102   * @o_flags: the O_* flags to apply to the new fd entry
1103   *
1104   * Installs a received file into the file descriptor table, with appropriate
1105   * checks and count updates. Optionally writes the fd number to userspace, if
1106   * @ufd is non-NULL.
1107   *
1108   * This helper handles its own reference counting of the incoming
1109   * struct file.
1110   *
1111   * Returns newly install fd or -ve on error.
1112   */
1113  int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1114  {
1115  	int new_fd;
1116  	int error;
1117  
1118  	error = security_file_receive(file);
1119  	if (error)
1120  		return error;
1121  
1122  	new_fd = get_unused_fd_flags(o_flags);
1123  	if (new_fd < 0)
1124  		return new_fd;
1125  
1126  	if (ufd) {
1127  		error = put_user(new_fd, ufd);
1128  		if (error) {
1129  			put_unused_fd(new_fd);
1130  			return error;
1131  		}
1132  	}
1133  
1134  	fd_install(new_fd, get_file(file));
1135  	__receive_sock(file);
1136  	return new_fd;
1137  }
1138  
1139  int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1140  {
1141  	int error;
1142  
1143  	error = security_file_receive(file);
1144  	if (error)
1145  		return error;
1146  	error = replace_fd(new_fd, file, o_flags);
1147  	if (error)
1148  		return error;
1149  	__receive_sock(file);
1150  	return new_fd;
1151  }
1152  
1153  int receive_fd(struct file *file, unsigned int o_flags)
1154  {
1155  	return __receive_fd(file, NULL, o_flags);
1156  }
1157  EXPORT_SYMBOL_GPL(receive_fd);
1158  
1159  static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1160  {
1161  	int err = -EBADF;
1162  	struct file *file;
1163  	struct files_struct *files = current->files;
1164  
1165  	if ((flags & ~O_CLOEXEC) != 0)
1166  		return -EINVAL;
1167  
1168  	if (unlikely(oldfd == newfd))
1169  		return -EINVAL;
1170  
1171  	if (newfd >= rlimit(RLIMIT_NOFILE))
1172  		return -EBADF;
1173  
1174  	spin_lock(&files->file_lock);
1175  	err = expand_files(files, newfd);
1176  	file = files_lookup_fd_locked(files, oldfd);
1177  	if (unlikely(!file))
1178  		goto Ebadf;
1179  	if (unlikely(err < 0)) {
1180  		if (err == -EMFILE)
1181  			goto Ebadf;
1182  		goto out_unlock;
1183  	}
1184  	return do_dup2(files, file, newfd, flags);
1185  
1186  Ebadf:
1187  	err = -EBADF;
1188  out_unlock:
1189  	spin_unlock(&files->file_lock);
1190  	return err;
1191  }
1192  
1193  SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1194  {
1195  	return ksys_dup3(oldfd, newfd, flags);
1196  }
1197  
1198  SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1199  {
1200  	if (unlikely(newfd == oldfd)) { /* corner case */
1201  		struct files_struct *files = current->files;
1202  		int retval = oldfd;
1203  
1204  		rcu_read_lock();
1205  		if (!files_lookup_fd_rcu(files, oldfd))
1206  			retval = -EBADF;
1207  		rcu_read_unlock();
1208  		return retval;
1209  	}
1210  	return ksys_dup3(oldfd, newfd, 0);
1211  }
1212  
1213  SYSCALL_DEFINE1(dup, unsigned int, fildes)
1214  {
1215  	int ret = -EBADF;
1216  	struct file *file = fget_raw(fildes);
1217  
1218  	if (file) {
1219  		ret = get_unused_fd_flags(0);
1220  		if (ret >= 0)
1221  			fd_install(ret, file);
1222  		else
1223  			fput(file);
1224  	}
1225  	return ret;
1226  }
1227  
1228  int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1229  {
1230  	unsigned long nofile = rlimit(RLIMIT_NOFILE);
1231  	int err;
1232  	if (from >= nofile)
1233  		return -EINVAL;
1234  	err = alloc_fd(from, nofile, flags);
1235  	if (err >= 0) {
1236  		get_file(file);
1237  		fd_install(err, file);
1238  	}
1239  	return err;
1240  }
1241  
1242  int iterate_fd(struct files_struct *files, unsigned n,
1243  		int (*f)(const void *, struct file *, unsigned),
1244  		const void *p)
1245  {
1246  	struct fdtable *fdt;
1247  	int res = 0;
1248  	if (!files)
1249  		return 0;
1250  	spin_lock(&files->file_lock);
1251  	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1252  		struct file *file;
1253  		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1254  		if (!file)
1255  			continue;
1256  		res = f(p, file, n);
1257  		if (res)
1258  			break;
1259  	}
1260  	spin_unlock(&files->file_lock);
1261  	return res;
1262  }
1263  EXPORT_SYMBOL(iterate_fd);
1264