xref: /openbmc/linux/fs/file.c (revision ec5c05e5ac8bcb4a6bcd92970e15494a85400d34)
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  /*
91   * Note how the fdtable bitmap allocations very much have to be a multiple of
92   * BITS_PER_LONG. This is not only because we walk those things in chunks of
93   * 'unsigned long' in some places, but simply because that is how the Linux
94   * kernel bitmaps are defined to work: they are not "bits in an array of bytes",
95   * they are very much "bits in an array of unsigned long".
96   *
97   * The ALIGN(nr, BITS_PER_LONG) here is for clarity: since we just multiplied
98   * by that "1024/sizeof(ptr)" before, we already know there are sufficient
99   * clear low bits. Clang seems to realize that, gcc ends up being confused.
100   *
101   * On a 128-bit machine, the ALIGN() would actually matter. In the meantime,
102   * let's consider it documentation (and maybe a test-case for gcc to improve
103   * its code generation ;)
104   */
105  static struct fdtable * alloc_fdtable(unsigned int nr)
106  {
107  	struct fdtable *fdt;
108  	void *data;
109  
110  	/*
111  	 * Figure out how many fds we actually want to support in this fdtable.
112  	 * Allocation steps are keyed to the size of the fdarray, since it
113  	 * grows far faster than any of the other dynamic data. We try to fit
114  	 * the fdarray into comfortable page-tuned chunks: starting at 1024B
115  	 * and growing in powers of two from there on.
116  	 */
117  	nr /= (1024 / sizeof(struct file *));
118  	nr = roundup_pow_of_two(nr + 1);
119  	nr *= (1024 / sizeof(struct file *));
120  	nr = ALIGN(nr, BITS_PER_LONG);
121  	/*
122  	 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
123  	 * had been set lower between the check in expand_files() and here.  Deal
124  	 * with that in caller, it's cheaper that way.
125  	 *
126  	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
127  	 * bitmaps handling below becomes unpleasant, to put it mildly...
128  	 */
129  	if (unlikely(nr > sysctl_nr_open))
130  		nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
131  
132  	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
133  	if (!fdt)
134  		goto out;
135  	fdt->max_fds = nr;
136  	data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
137  	if (!data)
138  		goto out_fdt;
139  	fdt->fd = data;
140  
141  	data = kvmalloc(max_t(size_t,
142  				 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
143  				 GFP_KERNEL_ACCOUNT);
144  	if (!data)
145  		goto out_arr;
146  	fdt->open_fds = data;
147  	data += nr / BITS_PER_BYTE;
148  	fdt->close_on_exec = data;
149  	data += nr / BITS_PER_BYTE;
150  	fdt->full_fds_bits = data;
151  
152  	return fdt;
153  
154  out_arr:
155  	kvfree(fdt->fd);
156  out_fdt:
157  	kfree(fdt);
158  out:
159  	return NULL;
160  }
161  
162  /*
163   * Expand the file descriptor table.
164   * This function will allocate a new fdtable and both fd array and fdset, of
165   * the given size.
166   * Return <0 error code on error; 1 on successful completion.
167   * The files->file_lock should be held on entry, and will be held on exit.
168   */
169  static int expand_fdtable(struct files_struct *files, unsigned int nr)
170  	__releases(files->file_lock)
171  	__acquires(files->file_lock)
172  {
173  	struct fdtable *new_fdt, *cur_fdt;
174  
175  	spin_unlock(&files->file_lock);
176  	new_fdt = alloc_fdtable(nr);
177  
178  	/* make sure all fd_install() have seen resize_in_progress
179  	 * or have finished their rcu_read_lock_sched() section.
180  	 */
181  	if (atomic_read(&files->count) > 1)
182  		synchronize_rcu();
183  
184  	spin_lock(&files->file_lock);
185  	if (!new_fdt)
186  		return -ENOMEM;
187  	/*
188  	 * extremely unlikely race - sysctl_nr_open decreased between the check in
189  	 * caller and alloc_fdtable().  Cheaper to catch it here...
190  	 */
191  	if (unlikely(new_fdt->max_fds <= nr)) {
192  		__free_fdtable(new_fdt);
193  		return -EMFILE;
194  	}
195  	cur_fdt = files_fdtable(files);
196  	BUG_ON(nr < cur_fdt->max_fds);
197  	copy_fdtable(new_fdt, cur_fdt);
198  	rcu_assign_pointer(files->fdt, new_fdt);
199  	if (cur_fdt != &files->fdtab)
200  		call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
201  	/* coupled with smp_rmb() in fd_install() */
202  	smp_wmb();
203  	return 1;
204  }
205  
206  /*
207   * Expand files.
208   * This function will expand the file structures, if the requested size exceeds
209   * the current capacity and there is room for expansion.
210   * Return <0 error code on error; 0 when nothing done; 1 when files were
211   * expanded and execution may have blocked.
212   * The files->file_lock should be held on entry, and will be held on exit.
213   */
214  static int expand_files(struct files_struct *files, unsigned int nr)
215  	__releases(files->file_lock)
216  	__acquires(files->file_lock)
217  {
218  	struct fdtable *fdt;
219  	int expanded = 0;
220  
221  repeat:
222  	fdt = files_fdtable(files);
223  
224  	/* Do we need to expand? */
225  	if (nr < fdt->max_fds)
226  		return expanded;
227  
228  	/* Can we expand? */
229  	if (nr >= sysctl_nr_open)
230  		return -EMFILE;
231  
232  	if (unlikely(files->resize_in_progress)) {
233  		spin_unlock(&files->file_lock);
234  		expanded = 1;
235  		wait_event(files->resize_wait, !files->resize_in_progress);
236  		spin_lock(&files->file_lock);
237  		goto repeat;
238  	}
239  
240  	/* All good, so we try */
241  	files->resize_in_progress = true;
242  	expanded = expand_fdtable(files, nr);
243  	files->resize_in_progress = false;
244  
245  	wake_up_all(&files->resize_wait);
246  	return expanded;
247  }
248  
249  static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
250  {
251  	__set_bit(fd, fdt->close_on_exec);
252  }
253  
254  static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
255  {
256  	if (test_bit(fd, fdt->close_on_exec))
257  		__clear_bit(fd, fdt->close_on_exec);
258  }
259  
260  static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
261  {
262  	__set_bit(fd, fdt->open_fds);
263  	fd /= BITS_PER_LONG;
264  	if (!~fdt->open_fds[fd])
265  		__set_bit(fd, fdt->full_fds_bits);
266  }
267  
268  static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
269  {
270  	__clear_bit(fd, fdt->open_fds);
271  	__clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
272  }
273  
274  static unsigned int count_open_files(struct fdtable *fdt)
275  {
276  	unsigned int size = fdt->max_fds;
277  	unsigned int i;
278  
279  	/* Find the last open fd */
280  	for (i = size / BITS_PER_LONG; i > 0; ) {
281  		if (fdt->open_fds[--i])
282  			break;
283  	}
284  	i = (i + 1) * BITS_PER_LONG;
285  	return i;
286  }
287  
288  /*
289   * Note that a sane fdtable size always has to be a multiple of
290   * BITS_PER_LONG, since we have bitmaps that are sized by this.
291   *
292   * 'max_fds' will normally already be properly aligned, but it
293   * turns out that in the close_range() -> __close_range() ->
294   * unshare_fd() -> dup_fd() -> sane_fdtable_size() we can end
295   * up having a 'max_fds' value that isn't already aligned.
296   *
297   * Rather than make close_range() have to worry about this,
298   * just make that BITS_PER_LONG alignment be part of a sane
299   * fdtable size. Becuase that's really what it is.
300   */
301  static unsigned int sane_fdtable_size(struct fdtable *fdt, unsigned int max_fds)
302  {
303  	unsigned int count;
304  
305  	count = count_open_files(fdt);
306  	if (max_fds < NR_OPEN_DEFAULT)
307  		max_fds = NR_OPEN_DEFAULT;
308  	return ALIGN(min(count, max_fds), BITS_PER_LONG);
309  }
310  
311  /*
312   * Allocate a new files structure and copy contents from the
313   * passed in files structure.
314   * errorp will be valid only when the returned files_struct is NULL.
315   */
316  struct files_struct *dup_fd(struct files_struct *oldf, unsigned int max_fds, int *errorp)
317  {
318  	struct files_struct *newf;
319  	struct file **old_fds, **new_fds;
320  	unsigned int open_files, i;
321  	struct fdtable *old_fdt, *new_fdt;
322  
323  	*errorp = -ENOMEM;
324  	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
325  	if (!newf)
326  		goto out;
327  
328  	atomic_set(&newf->count, 1);
329  
330  	spin_lock_init(&newf->file_lock);
331  	newf->resize_in_progress = false;
332  	init_waitqueue_head(&newf->resize_wait);
333  	newf->next_fd = 0;
334  	new_fdt = &newf->fdtab;
335  	new_fdt->max_fds = NR_OPEN_DEFAULT;
336  	new_fdt->close_on_exec = newf->close_on_exec_init;
337  	new_fdt->open_fds = newf->open_fds_init;
338  	new_fdt->full_fds_bits = newf->full_fds_bits_init;
339  	new_fdt->fd = &newf->fd_array[0];
340  
341  	spin_lock(&oldf->file_lock);
342  	old_fdt = files_fdtable(oldf);
343  	open_files = sane_fdtable_size(old_fdt, max_fds);
344  
345  	/*
346  	 * Check whether we need to allocate a larger fd array and fd set.
347  	 */
348  	while (unlikely(open_files > new_fdt->max_fds)) {
349  		spin_unlock(&oldf->file_lock);
350  
351  		if (new_fdt != &newf->fdtab)
352  			__free_fdtable(new_fdt);
353  
354  		new_fdt = alloc_fdtable(open_files - 1);
355  		if (!new_fdt) {
356  			*errorp = -ENOMEM;
357  			goto out_release;
358  		}
359  
360  		/* beyond sysctl_nr_open; nothing to do */
361  		if (unlikely(new_fdt->max_fds < open_files)) {
362  			__free_fdtable(new_fdt);
363  			*errorp = -EMFILE;
364  			goto out_release;
365  		}
366  
367  		/*
368  		 * Reacquire the oldf lock and a pointer to its fd table
369  		 * who knows it may have a new bigger fd table. We need
370  		 * the latest pointer.
371  		 */
372  		spin_lock(&oldf->file_lock);
373  		old_fdt = files_fdtable(oldf);
374  		open_files = sane_fdtable_size(old_fdt, max_fds);
375  	}
376  
377  	copy_fd_bitmaps(new_fdt, old_fdt, open_files);
378  
379  	old_fds = old_fdt->fd;
380  	new_fds = new_fdt->fd;
381  
382  	for (i = open_files; i != 0; i--) {
383  		struct file *f = *old_fds++;
384  		if (f) {
385  			get_file(f);
386  		} else {
387  			/*
388  			 * The fd may be claimed in the fd bitmap but not yet
389  			 * instantiated in the files array if a sibling thread
390  			 * is partway through open().  So make sure that this
391  			 * fd is available to the new process.
392  			 */
393  			__clear_open_fd(open_files - i, new_fdt);
394  		}
395  		rcu_assign_pointer(*new_fds++, f);
396  	}
397  	spin_unlock(&oldf->file_lock);
398  
399  	/* clear the remainder */
400  	memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
401  
402  	rcu_assign_pointer(newf->fdt, new_fdt);
403  
404  	return newf;
405  
406  out_release:
407  	kmem_cache_free(files_cachep, newf);
408  out:
409  	return NULL;
410  }
411  
412  static struct fdtable *close_files(struct files_struct * files)
413  {
414  	/*
415  	 * It is safe to dereference the fd table without RCU or
416  	 * ->file_lock because this is the last reference to the
417  	 * files structure.
418  	 */
419  	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
420  	unsigned int i, j = 0;
421  
422  	for (;;) {
423  		unsigned long set;
424  		i = j * BITS_PER_LONG;
425  		if (i >= fdt->max_fds)
426  			break;
427  		set = fdt->open_fds[j++];
428  		while (set) {
429  			if (set & 1) {
430  				struct file * file = xchg(&fdt->fd[i], NULL);
431  				if (file) {
432  					filp_close(file, files);
433  					cond_resched();
434  				}
435  			}
436  			i++;
437  			set >>= 1;
438  		}
439  	}
440  
441  	return fdt;
442  }
443  
444  void put_files_struct(struct files_struct *files)
445  {
446  	if (atomic_dec_and_test(&files->count)) {
447  		struct fdtable *fdt = close_files(files);
448  
449  		/* free the arrays if they are not embedded */
450  		if (fdt != &files->fdtab)
451  			__free_fdtable(fdt);
452  		kmem_cache_free(files_cachep, files);
453  	}
454  }
455  
456  void exit_files(struct task_struct *tsk)
457  {
458  	struct files_struct * files = tsk->files;
459  
460  	if (files) {
461  		task_lock(tsk);
462  		tsk->files = NULL;
463  		task_unlock(tsk);
464  		put_files_struct(files);
465  	}
466  }
467  
468  struct files_struct init_files = {
469  	.count		= ATOMIC_INIT(1),
470  	.fdt		= &init_files.fdtab,
471  	.fdtab		= {
472  		.max_fds	= NR_OPEN_DEFAULT,
473  		.fd		= &init_files.fd_array[0],
474  		.close_on_exec	= init_files.close_on_exec_init,
475  		.open_fds	= init_files.open_fds_init,
476  		.full_fds_bits	= init_files.full_fds_bits_init,
477  	},
478  	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
479  	.resize_wait	= __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
480  };
481  
482  static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
483  {
484  	unsigned int maxfd = fdt->max_fds;
485  	unsigned int maxbit = maxfd / BITS_PER_LONG;
486  	unsigned int bitbit = start / BITS_PER_LONG;
487  
488  	bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
489  	if (bitbit > maxfd)
490  		return maxfd;
491  	if (bitbit > start)
492  		start = bitbit;
493  	return find_next_zero_bit(fdt->open_fds, maxfd, start);
494  }
495  
496  /*
497   * allocate a file descriptor, mark it busy.
498   */
499  static int alloc_fd(unsigned start, unsigned end, unsigned flags)
500  {
501  	struct files_struct *files = current->files;
502  	unsigned int fd;
503  	int error;
504  	struct fdtable *fdt;
505  
506  	spin_lock(&files->file_lock);
507  repeat:
508  	fdt = files_fdtable(files);
509  	fd = start;
510  	if (fd < files->next_fd)
511  		fd = files->next_fd;
512  
513  	if (fd < fdt->max_fds)
514  		fd = find_next_fd(fdt, fd);
515  
516  	/*
517  	 * N.B. For clone tasks sharing a files structure, this test
518  	 * will limit the total number of files that can be opened.
519  	 */
520  	error = -EMFILE;
521  	if (fd >= end)
522  		goto out;
523  
524  	error = expand_files(files, fd);
525  	if (error < 0)
526  		goto out;
527  
528  	/*
529  	 * If we needed to expand the fs array we
530  	 * might have blocked - try again.
531  	 */
532  	if (error)
533  		goto repeat;
534  
535  	if (start <= files->next_fd)
536  		files->next_fd = fd + 1;
537  
538  	__set_open_fd(fd, fdt);
539  	if (flags & O_CLOEXEC)
540  		__set_close_on_exec(fd, fdt);
541  	else
542  		__clear_close_on_exec(fd, fdt);
543  	error = fd;
544  #if 1
545  	/* Sanity check */
546  	if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
547  		printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
548  		rcu_assign_pointer(fdt->fd[fd], NULL);
549  	}
550  #endif
551  
552  out:
553  	spin_unlock(&files->file_lock);
554  	return error;
555  }
556  
557  int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
558  {
559  	return alloc_fd(0, nofile, flags);
560  }
561  
562  int get_unused_fd_flags(unsigned flags)
563  {
564  	return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
565  }
566  EXPORT_SYMBOL(get_unused_fd_flags);
567  
568  static void __put_unused_fd(struct files_struct *files, unsigned int fd)
569  {
570  	struct fdtable *fdt = files_fdtable(files);
571  	__clear_open_fd(fd, fdt);
572  	if (fd < files->next_fd)
573  		files->next_fd = fd;
574  }
575  
576  void put_unused_fd(unsigned int fd)
577  {
578  	struct files_struct *files = current->files;
579  	spin_lock(&files->file_lock);
580  	__put_unused_fd(files, fd);
581  	spin_unlock(&files->file_lock);
582  }
583  
584  EXPORT_SYMBOL(put_unused_fd);
585  
586  /*
587   * Install a file pointer in the fd array.
588   *
589   * The VFS is full of places where we drop the files lock between
590   * setting the open_fds bitmap and installing the file in the file
591   * array.  At any such point, we are vulnerable to a dup2() race
592   * installing a file in the array before us.  We need to detect this and
593   * fput() the struct file we are about to overwrite in this case.
594   *
595   * It should never happen - if we allow dup2() do it, _really_ bad things
596   * will follow.
597   *
598   * This consumes the "file" refcount, so callers should treat it
599   * as if they had called fput(file).
600   */
601  
602  void fd_install(unsigned int fd, struct file *file)
603  {
604  	struct files_struct *files = current->files;
605  	struct fdtable *fdt;
606  
607  	rcu_read_lock_sched();
608  
609  	if (unlikely(files->resize_in_progress)) {
610  		rcu_read_unlock_sched();
611  		spin_lock(&files->file_lock);
612  		fdt = files_fdtable(files);
613  		BUG_ON(fdt->fd[fd] != NULL);
614  		rcu_assign_pointer(fdt->fd[fd], file);
615  		spin_unlock(&files->file_lock);
616  		return;
617  	}
618  	/* coupled with smp_wmb() in expand_fdtable() */
619  	smp_rmb();
620  	fdt = rcu_dereference_sched(files->fdt);
621  	BUG_ON(fdt->fd[fd] != NULL);
622  	rcu_assign_pointer(fdt->fd[fd], file);
623  	rcu_read_unlock_sched();
624  }
625  
626  EXPORT_SYMBOL(fd_install);
627  
628  /**
629   * pick_file - return file associatd with fd
630   * @files: file struct to retrieve file from
631   * @fd: file descriptor to retrieve file for
632   *
633   * Context: files_lock must be held.
634   *
635   * Returns: The file associated with @fd (NULL if @fd is not open)
636   */
637  static struct file *pick_file(struct files_struct *files, unsigned fd)
638  {
639  	struct fdtable *fdt = files_fdtable(files);
640  	struct file *file;
641  
642  	if (fd >= fdt->max_fds)
643  		return NULL;
644  
645  	file = fdt->fd[fd];
646  	if (file) {
647  		rcu_assign_pointer(fdt->fd[fd], NULL);
648  		__put_unused_fd(files, fd);
649  	}
650  	return file;
651  }
652  
653  int close_fd(unsigned fd)
654  {
655  	struct files_struct *files = current->files;
656  	struct file *file;
657  
658  	spin_lock(&files->file_lock);
659  	file = pick_file(files, fd);
660  	spin_unlock(&files->file_lock);
661  	if (!file)
662  		return -EBADF;
663  
664  	return filp_close(file, files);
665  }
666  EXPORT_SYMBOL(close_fd); /* for ksys_close() */
667  
668  /**
669   * last_fd - return last valid index into fd table
670   * @cur_fds: files struct
671   *
672   * Context: Either rcu read lock or files_lock must be held.
673   *
674   * Returns: Last valid index into fdtable.
675   */
676  static inline unsigned last_fd(struct fdtable *fdt)
677  {
678  	return fdt->max_fds - 1;
679  }
680  
681  static inline void __range_cloexec(struct files_struct *cur_fds,
682  				   unsigned int fd, unsigned int max_fd)
683  {
684  	struct fdtable *fdt;
685  
686  	/* make sure we're using the correct maximum value */
687  	spin_lock(&cur_fds->file_lock);
688  	fdt = files_fdtable(cur_fds);
689  	max_fd = min(last_fd(fdt), max_fd);
690  	if (fd <= max_fd)
691  		bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
692  	spin_unlock(&cur_fds->file_lock);
693  }
694  
695  static inline void __range_close(struct files_struct *cur_fds, unsigned int fd,
696  				 unsigned int max_fd)
697  {
698  	unsigned n;
699  
700  	rcu_read_lock();
701  	n = last_fd(files_fdtable(cur_fds));
702  	rcu_read_unlock();
703  	max_fd = min(max_fd, n);
704  
705  	while (fd <= max_fd) {
706  		struct file *file;
707  
708  		spin_lock(&cur_fds->file_lock);
709  		file = pick_file(cur_fds, fd++);
710  		spin_unlock(&cur_fds->file_lock);
711  
712  		if (file) {
713  			/* found a valid file to close */
714  			filp_close(file, cur_fds);
715  			cond_resched();
716  		}
717  	}
718  }
719  
720  /**
721   * __close_range() - Close all file descriptors in a given range.
722   *
723   * @fd:     starting file descriptor to close
724   * @max_fd: last file descriptor to close
725   *
726   * This closes a range of file descriptors. All file descriptors
727   * from @fd up to and including @max_fd are closed.
728   */
729  int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
730  {
731  	struct task_struct *me = current;
732  	struct files_struct *cur_fds = me->files, *fds = NULL;
733  
734  	if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
735  		return -EINVAL;
736  
737  	if (fd > max_fd)
738  		return -EINVAL;
739  
740  	if (flags & CLOSE_RANGE_UNSHARE) {
741  		int ret;
742  		unsigned int max_unshare_fds = NR_OPEN_MAX;
743  
744  		/*
745  		 * If the caller requested all fds to be made cloexec we always
746  		 * copy all of the file descriptors since they still want to
747  		 * use them.
748  		 */
749  		if (!(flags & CLOSE_RANGE_CLOEXEC)) {
750  			/*
751  			 * If the requested range is greater than the current
752  			 * maximum, we're closing everything so only copy all
753  			 * file descriptors beneath the lowest file descriptor.
754  			 */
755  			rcu_read_lock();
756  			if (max_fd >= last_fd(files_fdtable(cur_fds)))
757  				max_unshare_fds = fd;
758  			rcu_read_unlock();
759  		}
760  
761  		ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
762  		if (ret)
763  			return ret;
764  
765  		/*
766  		 * We used to share our file descriptor table, and have now
767  		 * created a private one, make sure we're using it below.
768  		 */
769  		if (fds)
770  			swap(cur_fds, fds);
771  	}
772  
773  	if (flags & CLOSE_RANGE_CLOEXEC)
774  		__range_cloexec(cur_fds, fd, max_fd);
775  	else
776  		__range_close(cur_fds, fd, max_fd);
777  
778  	if (fds) {
779  		/*
780  		 * We're done closing the files we were supposed to. Time to install
781  		 * the new file descriptor table and drop the old one.
782  		 */
783  		task_lock(me);
784  		me->files = cur_fds;
785  		task_unlock(me);
786  		put_files_struct(fds);
787  	}
788  
789  	return 0;
790  }
791  
792  /*
793   * See close_fd_get_file() below, this variant assumes current->files->file_lock
794   * is held.
795   */
796  struct file *__close_fd_get_file(unsigned int fd)
797  {
798  	return pick_file(current->files, fd);
799  }
800  
801  /*
802   * variant of close_fd that gets a ref on the file for later fput.
803   * The caller must ensure that filp_close() called on the file.
804   */
805  struct file *close_fd_get_file(unsigned int fd)
806  {
807  	struct files_struct *files = current->files;
808  	struct file *file;
809  
810  	spin_lock(&files->file_lock);
811  	file = pick_file(files, fd);
812  	spin_unlock(&files->file_lock);
813  
814  	return file;
815  }
816  
817  void do_close_on_exec(struct files_struct *files)
818  {
819  	unsigned i;
820  	struct fdtable *fdt;
821  
822  	/* exec unshares first */
823  	spin_lock(&files->file_lock);
824  	for (i = 0; ; i++) {
825  		unsigned long set;
826  		unsigned fd = i * BITS_PER_LONG;
827  		fdt = files_fdtable(files);
828  		if (fd >= fdt->max_fds)
829  			break;
830  		set = fdt->close_on_exec[i];
831  		if (!set)
832  			continue;
833  		fdt->close_on_exec[i] = 0;
834  		for ( ; set ; fd++, set >>= 1) {
835  			struct file *file;
836  			if (!(set & 1))
837  				continue;
838  			file = fdt->fd[fd];
839  			if (!file)
840  				continue;
841  			rcu_assign_pointer(fdt->fd[fd], NULL);
842  			__put_unused_fd(files, fd);
843  			spin_unlock(&files->file_lock);
844  			filp_close(file, files);
845  			cond_resched();
846  			spin_lock(&files->file_lock);
847  		}
848  
849  	}
850  	spin_unlock(&files->file_lock);
851  }
852  
853  static inline struct file *__fget_files_rcu(struct files_struct *files,
854  	unsigned int fd, fmode_t mask)
855  {
856  	for (;;) {
857  		struct file *file;
858  		struct fdtable *fdt = rcu_dereference_raw(files->fdt);
859  		struct file __rcu **fdentry;
860  
861  		if (unlikely(fd >= fdt->max_fds))
862  			return NULL;
863  
864  		fdentry = fdt->fd + array_index_nospec(fd, fdt->max_fds);
865  		file = rcu_dereference_raw(*fdentry);
866  		if (unlikely(!file))
867  			return NULL;
868  
869  		if (unlikely(file->f_mode & mask))
870  			return NULL;
871  
872  		/*
873  		 * Ok, we have a file pointer. However, because we do
874  		 * this all locklessly under RCU, we may be racing with
875  		 * that file being closed.
876  		 *
877  		 * Such a race can take two forms:
878  		 *
879  		 *  (a) the file ref already went down to zero,
880  		 *      and get_file_rcu() fails. Just try again:
881  		 */
882  		if (unlikely(!get_file_rcu(file)))
883  			continue;
884  
885  		/*
886  		 *  (b) the file table entry has changed under us.
887  		 *       Note that we don't need to re-check the 'fdt->fd'
888  		 *       pointer having changed, because it always goes
889  		 *       hand-in-hand with 'fdt'.
890  		 *
891  		 * If so, we need to put our ref and try again.
892  		 */
893  		if (unlikely(rcu_dereference_raw(files->fdt) != fdt) ||
894  		    unlikely(rcu_dereference_raw(*fdentry) != file)) {
895  			fput(file);
896  			continue;
897  		}
898  
899  		/*
900  		 * Ok, we have a ref to the file, and checked that it
901  		 * still exists.
902  		 */
903  		return file;
904  	}
905  }
906  
907  static struct file *__fget_files(struct files_struct *files, unsigned int fd,
908  				 fmode_t mask)
909  {
910  	struct file *file;
911  
912  	rcu_read_lock();
913  	file = __fget_files_rcu(files, fd, mask);
914  	rcu_read_unlock();
915  
916  	return file;
917  }
918  
919  static inline struct file *__fget(unsigned int fd, fmode_t mask)
920  {
921  	return __fget_files(current->files, fd, mask);
922  }
923  
924  struct file *fget(unsigned int fd)
925  {
926  	return __fget(fd, FMODE_PATH);
927  }
928  EXPORT_SYMBOL(fget);
929  
930  struct file *fget_raw(unsigned int fd)
931  {
932  	return __fget(fd, 0);
933  }
934  EXPORT_SYMBOL(fget_raw);
935  
936  struct file *fget_task(struct task_struct *task, unsigned int fd)
937  {
938  	struct file *file = NULL;
939  
940  	task_lock(task);
941  	if (task->files)
942  		file = __fget_files(task->files, fd, 0);
943  	task_unlock(task);
944  
945  	return file;
946  }
947  
948  struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
949  {
950  	/* Must be called with rcu_read_lock held */
951  	struct files_struct *files;
952  	struct file *file = NULL;
953  
954  	task_lock(task);
955  	files = task->files;
956  	if (files)
957  		file = files_lookup_fd_rcu(files, fd);
958  	task_unlock(task);
959  
960  	return file;
961  }
962  
963  struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret_fd)
964  {
965  	/* Must be called with rcu_read_lock held */
966  	struct files_struct *files;
967  	unsigned int fd = *ret_fd;
968  	struct file *file = NULL;
969  
970  	task_lock(task);
971  	files = task->files;
972  	if (files) {
973  		for (; fd < files_fdtable(files)->max_fds; fd++) {
974  			file = files_lookup_fd_rcu(files, fd);
975  			if (file)
976  				break;
977  		}
978  	}
979  	task_unlock(task);
980  	*ret_fd = fd;
981  	return file;
982  }
983  EXPORT_SYMBOL(task_lookup_next_fd_rcu);
984  
985  /*
986   * Lightweight file lookup - no refcnt increment if fd table isn't shared.
987   *
988   * You can use this instead of fget if you satisfy all of the following
989   * conditions:
990   * 1) You must call fput_light before exiting the syscall and returning control
991   *    to userspace (i.e. you cannot remember the returned struct file * after
992   *    returning to userspace).
993   * 2) You must not call filp_close on the returned struct file * in between
994   *    calls to fget_light and fput_light.
995   * 3) You must not clone the current task in between the calls to fget_light
996   *    and fput_light.
997   *
998   * The fput_needed flag returned by fget_light should be passed to the
999   * corresponding fput_light.
1000   */
1001  static unsigned long __fget_light(unsigned int fd, fmode_t mask)
1002  {
1003  	struct files_struct *files = current->files;
1004  	struct file *file;
1005  
1006  	/*
1007  	 * If another thread is concurrently calling close_fd() followed
1008  	 * by put_files_struct(), we must not observe the old table
1009  	 * entry combined with the new refcount - otherwise we could
1010  	 * return a file that is concurrently being freed.
1011  	 *
1012  	 * atomic_read_acquire() pairs with atomic_dec_and_test() in
1013  	 * put_files_struct().
1014  	 */
1015  	if (atomic_read_acquire(&files->count) == 1) {
1016  		file = files_lookup_fd_raw(files, fd);
1017  		if (!file || unlikely(file->f_mode & mask))
1018  			return 0;
1019  		return (unsigned long)file;
1020  	} else {
1021  		file = __fget(fd, mask);
1022  		if (!file)
1023  			return 0;
1024  		return FDPUT_FPUT | (unsigned long)file;
1025  	}
1026  }
1027  unsigned long __fdget(unsigned int fd)
1028  {
1029  	return __fget_light(fd, FMODE_PATH);
1030  }
1031  EXPORT_SYMBOL(__fdget);
1032  
1033  unsigned long __fdget_raw(unsigned int fd)
1034  {
1035  	return __fget_light(fd, 0);
1036  }
1037  
1038  unsigned long __fdget_pos(unsigned int fd)
1039  {
1040  	unsigned long v = __fdget(fd);
1041  	struct file *file = (struct file *)(v & ~3);
1042  
1043  	if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
1044  		if (file_count(file) > 1) {
1045  			v |= FDPUT_POS_UNLOCK;
1046  			mutex_lock(&file->f_pos_lock);
1047  		}
1048  	}
1049  	return v;
1050  }
1051  
1052  void __f_unlock_pos(struct file *f)
1053  {
1054  	mutex_unlock(&f->f_pos_lock);
1055  }
1056  
1057  /*
1058   * We only lock f_pos if we have threads or if the file might be
1059   * shared with another process. In both cases we'll have an elevated
1060   * file count (done either by fdget() or by fork()).
1061   */
1062  
1063  void set_close_on_exec(unsigned int fd, int flag)
1064  {
1065  	struct files_struct *files = current->files;
1066  	struct fdtable *fdt;
1067  	spin_lock(&files->file_lock);
1068  	fdt = files_fdtable(files);
1069  	if (flag)
1070  		__set_close_on_exec(fd, fdt);
1071  	else
1072  		__clear_close_on_exec(fd, fdt);
1073  	spin_unlock(&files->file_lock);
1074  }
1075  
1076  bool get_close_on_exec(unsigned int fd)
1077  {
1078  	struct files_struct *files = current->files;
1079  	struct fdtable *fdt;
1080  	bool res;
1081  	rcu_read_lock();
1082  	fdt = files_fdtable(files);
1083  	res = close_on_exec(fd, fdt);
1084  	rcu_read_unlock();
1085  	return res;
1086  }
1087  
1088  static int do_dup2(struct files_struct *files,
1089  	struct file *file, unsigned fd, unsigned flags)
1090  __releases(&files->file_lock)
1091  {
1092  	struct file *tofree;
1093  	struct fdtable *fdt;
1094  
1095  	/*
1096  	 * We need to detect attempts to do dup2() over allocated but still
1097  	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
1098  	 * extra work in their equivalent of fget() - they insert struct
1099  	 * file immediately after grabbing descriptor, mark it larval if
1100  	 * more work (e.g. actual opening) is needed and make sure that
1101  	 * fget() treats larval files as absent.  Potentially interesting,
1102  	 * but while extra work in fget() is trivial, locking implications
1103  	 * and amount of surgery on open()-related paths in VFS are not.
1104  	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1105  	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
1106  	 * scope of POSIX or SUS, since neither considers shared descriptor
1107  	 * tables and this condition does not arise without those.
1108  	 */
1109  	fdt = files_fdtable(files);
1110  	tofree = fdt->fd[fd];
1111  	if (!tofree && fd_is_open(fd, fdt))
1112  		goto Ebusy;
1113  	get_file(file);
1114  	rcu_assign_pointer(fdt->fd[fd], file);
1115  	__set_open_fd(fd, fdt);
1116  	if (flags & O_CLOEXEC)
1117  		__set_close_on_exec(fd, fdt);
1118  	else
1119  		__clear_close_on_exec(fd, fdt);
1120  	spin_unlock(&files->file_lock);
1121  
1122  	if (tofree)
1123  		filp_close(tofree, files);
1124  
1125  	return fd;
1126  
1127  Ebusy:
1128  	spin_unlock(&files->file_lock);
1129  	return -EBUSY;
1130  }
1131  
1132  int replace_fd(unsigned fd, struct file *file, unsigned flags)
1133  {
1134  	int err;
1135  	struct files_struct *files = current->files;
1136  
1137  	if (!file)
1138  		return close_fd(fd);
1139  
1140  	if (fd >= rlimit(RLIMIT_NOFILE))
1141  		return -EBADF;
1142  
1143  	spin_lock(&files->file_lock);
1144  	err = expand_files(files, fd);
1145  	if (unlikely(err < 0))
1146  		goto out_unlock;
1147  	return do_dup2(files, file, fd, flags);
1148  
1149  out_unlock:
1150  	spin_unlock(&files->file_lock);
1151  	return err;
1152  }
1153  
1154  /**
1155   * __receive_fd() - Install received file into file descriptor table
1156   * @file: struct file that was received from another process
1157   * @ufd: __user pointer to write new fd number to
1158   * @o_flags: the O_* flags to apply to the new fd entry
1159   *
1160   * Installs a received file into the file descriptor table, with appropriate
1161   * checks and count updates. Optionally writes the fd number to userspace, if
1162   * @ufd is non-NULL.
1163   *
1164   * This helper handles its own reference counting of the incoming
1165   * struct file.
1166   *
1167   * Returns newly install fd or -ve on error.
1168   */
1169  int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1170  {
1171  	int new_fd;
1172  	int error;
1173  
1174  	error = security_file_receive(file);
1175  	if (error)
1176  		return error;
1177  
1178  	new_fd = get_unused_fd_flags(o_flags);
1179  	if (new_fd < 0)
1180  		return new_fd;
1181  
1182  	if (ufd) {
1183  		error = put_user(new_fd, ufd);
1184  		if (error) {
1185  			put_unused_fd(new_fd);
1186  			return error;
1187  		}
1188  	}
1189  
1190  	fd_install(new_fd, get_file(file));
1191  	__receive_sock(file);
1192  	return new_fd;
1193  }
1194  
1195  int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1196  {
1197  	int error;
1198  
1199  	error = security_file_receive(file);
1200  	if (error)
1201  		return error;
1202  	error = replace_fd(new_fd, file, o_flags);
1203  	if (error)
1204  		return error;
1205  	__receive_sock(file);
1206  	return new_fd;
1207  }
1208  
1209  int receive_fd(struct file *file, unsigned int o_flags)
1210  {
1211  	return __receive_fd(file, NULL, o_flags);
1212  }
1213  EXPORT_SYMBOL_GPL(receive_fd);
1214  
1215  static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1216  {
1217  	int err = -EBADF;
1218  	struct file *file;
1219  	struct files_struct *files = current->files;
1220  
1221  	if ((flags & ~O_CLOEXEC) != 0)
1222  		return -EINVAL;
1223  
1224  	if (unlikely(oldfd == newfd))
1225  		return -EINVAL;
1226  
1227  	if (newfd >= rlimit(RLIMIT_NOFILE))
1228  		return -EBADF;
1229  
1230  	spin_lock(&files->file_lock);
1231  	err = expand_files(files, newfd);
1232  	file = files_lookup_fd_locked(files, oldfd);
1233  	if (unlikely(!file))
1234  		goto Ebadf;
1235  	if (unlikely(err < 0)) {
1236  		if (err == -EMFILE)
1237  			goto Ebadf;
1238  		goto out_unlock;
1239  	}
1240  	return do_dup2(files, file, newfd, flags);
1241  
1242  Ebadf:
1243  	err = -EBADF;
1244  out_unlock:
1245  	spin_unlock(&files->file_lock);
1246  	return err;
1247  }
1248  
1249  SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1250  {
1251  	return ksys_dup3(oldfd, newfd, flags);
1252  }
1253  
1254  SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1255  {
1256  	if (unlikely(newfd == oldfd)) { /* corner case */
1257  		struct files_struct *files = current->files;
1258  		int retval = oldfd;
1259  
1260  		rcu_read_lock();
1261  		if (!files_lookup_fd_rcu(files, oldfd))
1262  			retval = -EBADF;
1263  		rcu_read_unlock();
1264  		return retval;
1265  	}
1266  	return ksys_dup3(oldfd, newfd, 0);
1267  }
1268  
1269  SYSCALL_DEFINE1(dup, unsigned int, fildes)
1270  {
1271  	int ret = -EBADF;
1272  	struct file *file = fget_raw(fildes);
1273  
1274  	if (file) {
1275  		ret = get_unused_fd_flags(0);
1276  		if (ret >= 0)
1277  			fd_install(ret, file);
1278  		else
1279  			fput(file);
1280  	}
1281  	return ret;
1282  }
1283  
1284  int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1285  {
1286  	unsigned long nofile = rlimit(RLIMIT_NOFILE);
1287  	int err;
1288  	if (from >= nofile)
1289  		return -EINVAL;
1290  	err = alloc_fd(from, nofile, flags);
1291  	if (err >= 0) {
1292  		get_file(file);
1293  		fd_install(err, file);
1294  	}
1295  	return err;
1296  }
1297  
1298  int iterate_fd(struct files_struct *files, unsigned n,
1299  		int (*f)(const void *, struct file *, unsigned),
1300  		const void *p)
1301  {
1302  	struct fdtable *fdt;
1303  	int res = 0;
1304  	if (!files)
1305  		return 0;
1306  	spin_lock(&files->file_lock);
1307  	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1308  		struct file *file;
1309  		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1310  		if (!file)
1311  			continue;
1312  		res = f(p, file, n);
1313  		if (res)
1314  			break;
1315  	}
1316  	spin_unlock(&files->file_lock);
1317  	return res;
1318  }
1319  EXPORT_SYMBOL(iterate_fd);
1320