Linux Audio

Check our new training course

Loading...
v6.8
   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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  27unsigned int sysctl_nr_open __read_mostly = 1024*1024;
  28unsigned 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))
  31unsigned int sysctl_nr_open_max =
  32	__const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
  33
  34static void __free_fdtable(struct fdtable *fdt)
  35{
  36	kvfree(fdt->fd);
  37	kvfree(fdt->open_fds);
  38	kfree(fdt);
  39}
  40
  41static 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 */
  54static 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 */
  76static 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 */
 105static 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
 154out_arr:
 155	kvfree(fdt->fd);
 156out_fdt:
 157	kfree(fdt);
 158out:
 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 */
 169static 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 */
 214static 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
 221repeat:
 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
 249static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
 250{
 251	__set_bit(fd, fdt->close_on_exec);
 252}
 253
 254static 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
 260static 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
 268static 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
 274static 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 */
 301static 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 */
 316struct 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
 406out_release:
 407	kmem_cache_free(files_cachep, newf);
 408out:
 409	return NULL;
 410}
 411
 412static 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
 444void 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
 456void 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
 468struct 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
 482static 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 */
 499static 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);
 507repeat:
 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
 552out:
 553	spin_unlock(&files->file_lock);
 554	return error;
 555}
 556
 557int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
 558{
 559	return alloc_fd(0, nofile, flags);
 560}
 561
 562int get_unused_fd_flags(unsigned flags)
 563{
 564	return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
 565}
 566EXPORT_SYMBOL(get_unused_fd_flags);
 567
 568static 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
 576void 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
 584EXPORT_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
 602void fd_install(unsigned int fd, struct file *file)
 603{
 604	struct files_struct *files = current->files;
 605	struct fdtable *fdt;
 606
 607	if (WARN_ON_ONCE(unlikely(file->f_mode & FMODE_BACKING)))
 608		return;
 609
 610	rcu_read_lock_sched();
 611
 612	if (unlikely(files->resize_in_progress)) {
 613		rcu_read_unlock_sched();
 614		spin_lock(&files->file_lock);
 615		fdt = files_fdtable(files);
 616		BUG_ON(fdt->fd[fd] != NULL);
 617		rcu_assign_pointer(fdt->fd[fd], file);
 618		spin_unlock(&files->file_lock);
 619		return;
 620	}
 621	/* coupled with smp_wmb() in expand_fdtable() */
 622	smp_rmb();
 623	fdt = rcu_dereference_sched(files->fdt);
 624	BUG_ON(fdt->fd[fd] != NULL);
 625	rcu_assign_pointer(fdt->fd[fd], file);
 626	rcu_read_unlock_sched();
 627}
 628
 629EXPORT_SYMBOL(fd_install);
 630
 631/**
 632 * file_close_fd_locked - return file associated with fd
 633 * @files: file struct to retrieve file from
 634 * @fd: file descriptor to retrieve file for
 635 *
 636 * Doesn't take a separate reference count.
 637 *
 638 * Context: files_lock must be held.
 639 *
 640 * Returns: The file associated with @fd (NULL if @fd is not open)
 641 */
 642struct file *file_close_fd_locked(struct files_struct *files, unsigned fd)
 643{
 644	struct fdtable *fdt = files_fdtable(files);
 645	struct file *file;
 646
 647	lockdep_assert_held(&files->file_lock);
 648
 649	if (fd >= fdt->max_fds)
 650		return NULL;
 651
 652	fd = array_index_nospec(fd, fdt->max_fds);
 653	file = fdt->fd[fd];
 654	if (file) {
 655		rcu_assign_pointer(fdt->fd[fd], NULL);
 656		__put_unused_fd(files, fd);
 657	}
 658	return file;
 659}
 660
 661int close_fd(unsigned fd)
 662{
 663	struct files_struct *files = current->files;
 664	struct file *file;
 665
 666	spin_lock(&files->file_lock);
 667	file = file_close_fd_locked(files, fd);
 668	spin_unlock(&files->file_lock);
 669	if (!file)
 670		return -EBADF;
 671
 672	return filp_close(file, files);
 673}
 674EXPORT_SYMBOL(close_fd); /* for ksys_close() */
 675
 676/**
 677 * last_fd - return last valid index into fd table
 678 * @fdt: File descriptor table.
 679 *
 680 * Context: Either rcu read lock or files_lock must be held.
 681 *
 682 * Returns: Last valid index into fdtable.
 683 */
 684static inline unsigned last_fd(struct fdtable *fdt)
 685{
 686	return fdt->max_fds - 1;
 687}
 688
 689static inline void __range_cloexec(struct files_struct *cur_fds,
 690				   unsigned int fd, unsigned int max_fd)
 691{
 692	struct fdtable *fdt;
 693
 694	/* make sure we're using the correct maximum value */
 695	spin_lock(&cur_fds->file_lock);
 696	fdt = files_fdtable(cur_fds);
 697	max_fd = min(last_fd(fdt), max_fd);
 698	if (fd <= max_fd)
 699		bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
 700	spin_unlock(&cur_fds->file_lock);
 701}
 702
 703static inline void __range_close(struct files_struct *files, unsigned int fd,
 704				 unsigned int max_fd)
 705{
 706	struct file *file;
 707	unsigned n;
 708
 709	spin_lock(&files->file_lock);
 710	n = last_fd(files_fdtable(files));
 711	max_fd = min(max_fd, n);
 712
 713	for (; fd <= max_fd; fd++) {
 714		file = file_close_fd_locked(files, fd);
 715		if (file) {
 716			spin_unlock(&files->file_lock);
 717			filp_close(file, files);
 718			cond_resched();
 719			spin_lock(&files->file_lock);
 720		} else if (need_resched()) {
 721			spin_unlock(&files->file_lock);
 722			cond_resched();
 723			spin_lock(&files->file_lock);
 724		}
 725	}
 726	spin_unlock(&files->file_lock);
 727}
 728
 729/**
 730 * __close_range() - Close all file descriptors in a given range.
 731 *
 732 * @fd:     starting file descriptor to close
 733 * @max_fd: last file descriptor to close
 734 * @flags:  CLOSE_RANGE flags.
 735 *
 736 * This closes a range of file descriptors. All file descriptors
 737 * from @fd up to and including @max_fd are closed.
 
 738 */
 739int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
 
 740{
 741	struct task_struct *me = current;
 742	struct files_struct *cur_fds = me->files, *fds = NULL;
 743
 744	if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
 745		return -EINVAL;
 746
 747	if (fd > max_fd)
 748		return -EINVAL;
 749
 750	if (flags & CLOSE_RANGE_UNSHARE) {
 751		int ret;
 752		unsigned int max_unshare_fds = NR_OPEN_MAX;
 753
 754		/*
 755		 * If the caller requested all fds to be made cloexec we always
 756		 * copy all of the file descriptors since they still want to
 757		 * use them.
 758		 */
 759		if (!(flags & CLOSE_RANGE_CLOEXEC)) {
 760			/*
 761			 * If the requested range is greater than the current
 762			 * maximum, we're closing everything so only copy all
 763			 * file descriptors beneath the lowest file descriptor.
 764			 */
 765			rcu_read_lock();
 766			if (max_fd >= last_fd(files_fdtable(cur_fds)))
 767				max_unshare_fds = fd;
 768			rcu_read_unlock();
 769		}
 770
 771		ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
 772		if (ret)
 773			return ret;
 774
 
 
 
 775		/*
 776		 * We used to share our file descriptor table, and have now
 777		 * created a private one, make sure we're using it below.
 778		 */
 779		if (fds)
 780			swap(cur_fds, fds);
 781	}
 782
 783	if (flags & CLOSE_RANGE_CLOEXEC)
 784		__range_cloexec(cur_fds, fd, max_fd);
 785	else
 786		__range_close(cur_fds, fd, max_fd);
 787
 788	if (fds) {
 789		/*
 790		 * We're done closing the files we were supposed to. Time to install
 791		 * the new file descriptor table and drop the old one.
 792		 */
 793		task_lock(me);
 794		me->files = cur_fds;
 795		task_unlock(me);
 796		put_files_struct(fds);
 797	}
 798
 799	return 0;
 800}
 801
 802/**
 803 * file_close_fd - return file associated with fd
 804 * @fd: file descriptor to retrieve file for
 805 *
 806 * Doesn't take a separate reference count.
 807 *
 808 * Returns: The file associated with @fd (NULL if @fd is not open)
 809 */
 810struct file *file_close_fd(unsigned int fd)
 811{
 812	struct files_struct *files = current->files;
 813	struct file *file;
 814
 815	spin_lock(&files->file_lock);
 816	file = file_close_fd_locked(files, fd);
 817	spin_unlock(&files->file_lock);
 818
 819	return file;
 820}
 821
 822void do_close_on_exec(struct files_struct *files)
 823{
 824	unsigned i;
 825	struct fdtable *fdt;
 826
 827	/* exec unshares first */
 828	spin_lock(&files->file_lock);
 829	for (i = 0; ; i++) {
 830		unsigned long set;
 831		unsigned fd = i * BITS_PER_LONG;
 832		fdt = files_fdtable(files);
 833		if (fd >= fdt->max_fds)
 834			break;
 835		set = fdt->close_on_exec[i];
 836		if (!set)
 837			continue;
 838		fdt->close_on_exec[i] = 0;
 839		for ( ; set ; fd++, set >>= 1) {
 840			struct file *file;
 841			if (!(set & 1))
 842				continue;
 843			file = fdt->fd[fd];
 844			if (!file)
 845				continue;
 846			rcu_assign_pointer(fdt->fd[fd], NULL);
 847			__put_unused_fd(files, fd);
 848			spin_unlock(&files->file_lock);
 849			filp_close(file, files);
 850			cond_resched();
 851			spin_lock(&files->file_lock);
 852		}
 853
 854	}
 855	spin_unlock(&files->file_lock);
 856}
 857
 858static struct file *__get_file_rcu(struct file __rcu **f)
 859{
 860	struct file __rcu *file;
 861	struct file __rcu *file_reloaded;
 862	struct file __rcu *file_reloaded_cmp;
 863
 864	file = rcu_dereference_raw(*f);
 865	if (!file)
 866		return NULL;
 867
 868	if (unlikely(!atomic_long_inc_not_zero(&file->f_count)))
 869		return ERR_PTR(-EAGAIN);
 870
 871	file_reloaded = rcu_dereference_raw(*f);
 872
 873	/*
 874	 * Ensure that all accesses have a dependency on the load from
 875	 * rcu_dereference_raw() above so we get correct ordering
 876	 * between reuse/allocation and the pointer check below.
 877	 */
 878	file_reloaded_cmp = file_reloaded;
 879	OPTIMIZER_HIDE_VAR(file_reloaded_cmp);
 880
 881	/*
 882	 * atomic_long_inc_not_zero() above provided a full memory
 883	 * barrier when we acquired a reference.
 884	 *
 885	 * This is paired with the write barrier from assigning to the
 886	 * __rcu protected file pointer so that if that pointer still
 887	 * matches the current file, we know we have successfully
 888	 * acquired a reference to the right file.
 889	 *
 890	 * If the pointers don't match the file has been reallocated by
 891	 * SLAB_TYPESAFE_BY_RCU.
 892	 */
 893	if (file == file_reloaded_cmp)
 894		return file_reloaded;
 895
 896	fput(file);
 897	return ERR_PTR(-EAGAIN);
 898}
 899
 900/**
 901 * get_file_rcu - try go get a reference to a file under rcu
 902 * @f: the file to get a reference on
 903 *
 904 * This function tries to get a reference on @f carefully verifying that
 905 * @f hasn't been reused.
 906 *
 907 * This function should rarely have to be used and only by users who
 908 * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it.
 909 *
 910 * Return: Returns @f with the reference count increased or NULL.
 911 */
 912struct file *get_file_rcu(struct file __rcu **f)
 913{
 914	for (;;) {
 915		struct file __rcu *file;
 916
 917		file = __get_file_rcu(f);
 918		if (unlikely(!file))
 919			return NULL;
 920
 921		if (unlikely(IS_ERR(file)))
 922			continue;
 923
 924		return file;
 925	}
 926}
 927EXPORT_SYMBOL_GPL(get_file_rcu);
 928
 929/**
 930 * get_file_active - try go get a reference to a file
 931 * @f: the file to get a reference on
 932 *
 933 * In contast to get_file_rcu() the pointer itself isn't part of the
 934 * reference counting.
 935 *
 936 * This function should rarely have to be used and only by users who
 937 * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it.
 938 *
 939 * Return: Returns @f with the reference count increased or NULL.
 940 */
 941struct file *get_file_active(struct file **f)
 942{
 943	struct file __rcu *file;
 944
 945	rcu_read_lock();
 946	file = __get_file_rcu(f);
 947	rcu_read_unlock();
 948	if (IS_ERR(file))
 949		file = NULL;
 950	return file;
 951}
 952EXPORT_SYMBOL_GPL(get_file_active);
 953
 954static inline struct file *__fget_files_rcu(struct files_struct *files,
 955       unsigned int fd, fmode_t mask)
 956{
 957	for (;;) {
 958		struct file *file;
 959		struct fdtable *fdt = rcu_dereference_raw(files->fdt);
 960		struct file __rcu **fdentry;
 961		unsigned long nospec_mask;
 962
 963		/* Mask is a 0 for invalid fd's, ~0 for valid ones */
 964		nospec_mask = array_index_mask_nospec(fd, fdt->max_fds);
 965
 966		/*
 967		 * fdentry points to the 'fd' offset, or fdt->fd[0].
 968		 * Loading from fdt->fd[0] is always safe, because the
 969		 * array always exists.
 970		 */
 971		fdentry = fdt->fd + (fd & nospec_mask);
 972
 973		/* Do the load, then mask any invalid result */
 974		file = rcu_dereference_raw(*fdentry);
 975		file = (void *)(nospec_mask & (unsigned long)file);
 976		if (unlikely(!file))
 977			return NULL;
 978
 979		/*
 980		 * Ok, we have a file pointer that was valid at
 981		 * some point, but it might have become stale since.
 982		 *
 983		 * We need to confirm it by incrementing the refcount
 984		 * and then check the lookup again.
 985		 *
 986		 * atomic_long_inc_not_zero() gives us a full memory
 987		 * barrier. We only really need an 'acquire' one to
 988		 * protect the loads below, but we don't have that.
 989		 */
 990		if (unlikely(!atomic_long_inc_not_zero(&file->f_count)))
 991			continue;
 992
 993		/*
 994		 * Such a race can take two forms:
 995		 *
 996		 *  (a) the file ref already went down to zero and the
 997		 *      file hasn't been reused yet or the file count
 998		 *      isn't zero but the file has already been reused.
 999		 *
1000		 *  (b) the file table entry has changed under us.
1001		 *       Note that we don't need to re-check the 'fdt->fd'
1002		 *       pointer having changed, because it always goes
1003		 *       hand-in-hand with 'fdt'.
1004		 *
1005		 * If so, we need to put our ref and try again.
1006		 */
1007		if (unlikely(file != rcu_dereference_raw(*fdentry)) ||
1008		    unlikely(rcu_dereference_raw(files->fdt) != fdt)) {
1009			fput(file);
1010			continue;
1011		}
1012
1013		/*
1014		 * This isn't the file we're looking for or we're not
1015		 * allowed to get a reference to it.
1016		 */
1017		if (unlikely(file->f_mode & mask)) {
1018			fput(file);
1019			return NULL;
1020		}
1021
1022		/*
1023		 * Ok, we have a ref to the file, and checked that it
1024		 * still exists.
1025		 */
1026		return file;
1027	}
1028}
1029
1030static struct file *__fget_files(struct files_struct *files, unsigned int fd,
1031				 fmode_t mask)
1032{
1033	struct file *file;
1034
1035	rcu_read_lock();
1036	file = __fget_files_rcu(files, fd, mask);
1037	rcu_read_unlock();
1038
1039	return file;
1040}
1041
1042static inline struct file *__fget(unsigned int fd, fmode_t mask)
1043{
1044	return __fget_files(current->files, fd, mask);
1045}
1046
1047struct file *fget(unsigned int fd)
1048{
1049	return __fget(fd, FMODE_PATH);
1050}
1051EXPORT_SYMBOL(fget);
1052
1053struct file *fget_raw(unsigned int fd)
1054{
1055	return __fget(fd, 0);
1056}
1057EXPORT_SYMBOL(fget_raw);
1058
1059struct file *fget_task(struct task_struct *task, unsigned int fd)
1060{
1061	struct file *file = NULL;
1062
1063	task_lock(task);
1064	if (task->files)
1065		file = __fget_files(task->files, fd, 0);
1066	task_unlock(task);
1067
1068	return file;
1069}
1070
1071struct file *lookup_fdget_rcu(unsigned int fd)
1072{
1073	return __fget_files_rcu(current->files, fd, 0);
1074
1075}
1076EXPORT_SYMBOL_GPL(lookup_fdget_rcu);
1077
1078struct file *task_lookup_fdget_rcu(struct task_struct *task, unsigned int fd)
1079{
1080	/* Must be called with rcu_read_lock held */
1081	struct files_struct *files;
1082	struct file *file = NULL;
1083
1084	task_lock(task);
1085	files = task->files;
1086	if (files)
1087		file = __fget_files_rcu(files, fd, 0);
1088	task_unlock(task);
1089
1090	return file;
1091}
1092
1093struct file *task_lookup_next_fdget_rcu(struct task_struct *task, unsigned int *ret_fd)
1094{
1095	/* Must be called with rcu_read_lock held */
1096	struct files_struct *files;
1097	unsigned int fd = *ret_fd;
1098	struct file *file = NULL;
1099
1100	task_lock(task);
1101	files = task->files;
1102	if (files) {
 
1103		for (; fd < files_fdtable(files)->max_fds; fd++) {
1104			file = __fget_files_rcu(files, fd, 0);
1105			if (file)
1106				break;
1107		}
 
1108	}
1109	task_unlock(task);
1110	*ret_fd = fd;
1111	return file;
1112}
1113EXPORT_SYMBOL(task_lookup_next_fdget_rcu);
1114
1115/*
1116 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
1117 *
1118 * You can use this instead of fget if you satisfy all of the following
1119 * conditions:
1120 * 1) You must call fput_light before exiting the syscall and returning control
1121 *    to userspace (i.e. you cannot remember the returned struct file * after
1122 *    returning to userspace).
1123 * 2) You must not call filp_close on the returned struct file * in between
1124 *    calls to fget_light and fput_light.
1125 * 3) You must not clone the current task in between the calls to fget_light
1126 *    and fput_light.
1127 *
1128 * The fput_needed flag returned by fget_light should be passed to the
1129 * corresponding fput_light.
 
 
 
 
 
 
 
1130 */
1131static unsigned long __fget_light(unsigned int fd, fmode_t mask)
1132{
1133	struct files_struct *files = current->files;
1134	struct file *file;
1135
1136	/*
1137	 * If another thread is concurrently calling close_fd() followed
1138	 * by put_files_struct(), we must not observe the old table
1139	 * entry combined with the new refcount - otherwise we could
1140	 * return a file that is concurrently being freed.
1141	 *
1142	 * atomic_read_acquire() pairs with atomic_dec_and_test() in
1143	 * put_files_struct().
1144	 */
1145	if (likely(atomic_read_acquire(&files->count) == 1)) {
1146		file = files_lookup_fd_raw(files, fd);
1147		if (!file || unlikely(file->f_mode & mask))
1148			return 0;
1149		return (unsigned long)file;
1150	} else {
1151		file = __fget_files(files, fd, mask);
1152		if (!file)
1153			return 0;
1154		return FDPUT_FPUT | (unsigned long)file;
1155	}
1156}
1157unsigned long __fdget(unsigned int fd)
1158{
1159	return __fget_light(fd, FMODE_PATH);
1160}
1161EXPORT_SYMBOL(__fdget);
1162
1163unsigned long __fdget_raw(unsigned int fd)
1164{
1165	return __fget_light(fd, 0);
1166}
1167
1168/*
1169 * Try to avoid f_pos locking. We only need it if the
1170 * file is marked for FMODE_ATOMIC_POS, and it can be
1171 * accessed multiple ways.
1172 *
1173 * Always do it for directories, because pidfd_getfd()
1174 * can make a file accessible even if it otherwise would
1175 * not be, and for directories this is a correctness
1176 * issue, not a "POSIX requirement".
1177 */
1178static inline bool file_needs_f_pos_lock(struct file *file)
1179{
1180	return (file->f_mode & FMODE_ATOMIC_POS) &&
1181		(file_count(file) > 1 || file->f_op->iterate_shared);
1182}
1183
1184unsigned long __fdget_pos(unsigned int fd)
1185{
1186	unsigned long v = __fdget(fd);
1187	struct file *file = (struct file *)(v & ~3);
1188
1189	if (file && file_needs_f_pos_lock(file)) {
1190		v |= FDPUT_POS_UNLOCK;
1191		mutex_lock(&file->f_pos_lock);
1192	}
1193	return v;
1194}
1195
1196void __f_unlock_pos(struct file *f)
1197{
1198	mutex_unlock(&f->f_pos_lock);
1199}
1200
1201/*
1202 * We only lock f_pos if we have threads or if the file might be
1203 * shared with another process. In both cases we'll have an elevated
1204 * file count (done either by fdget() or by fork()).
1205 */
1206
1207void set_close_on_exec(unsigned int fd, int flag)
1208{
1209	struct files_struct *files = current->files;
1210	struct fdtable *fdt;
1211	spin_lock(&files->file_lock);
1212	fdt = files_fdtable(files);
1213	if (flag)
1214		__set_close_on_exec(fd, fdt);
1215	else
1216		__clear_close_on_exec(fd, fdt);
1217	spin_unlock(&files->file_lock);
1218}
1219
1220bool get_close_on_exec(unsigned int fd)
1221{
1222	struct files_struct *files = current->files;
1223	struct fdtable *fdt;
1224	bool res;
1225	rcu_read_lock();
1226	fdt = files_fdtable(files);
1227	res = close_on_exec(fd, fdt);
1228	rcu_read_unlock();
1229	return res;
1230}
1231
1232static int do_dup2(struct files_struct *files,
1233	struct file *file, unsigned fd, unsigned flags)
1234__releases(&files->file_lock)
1235{
1236	struct file *tofree;
1237	struct fdtable *fdt;
1238
1239	/*
1240	 * We need to detect attempts to do dup2() over allocated but still
1241	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
1242	 * extra work in their equivalent of fget() - they insert struct
1243	 * file immediately after grabbing descriptor, mark it larval if
1244	 * more work (e.g. actual opening) is needed and make sure that
1245	 * fget() treats larval files as absent.  Potentially interesting,
1246	 * but while extra work in fget() is trivial, locking implications
1247	 * and amount of surgery on open()-related paths in VFS are not.
1248	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1249	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
1250	 * scope of POSIX or SUS, since neither considers shared descriptor
1251	 * tables and this condition does not arise without those.
1252	 */
1253	fdt = files_fdtable(files);
 
1254	tofree = fdt->fd[fd];
1255	if (!tofree && fd_is_open(fd, fdt))
1256		goto Ebusy;
1257	get_file(file);
1258	rcu_assign_pointer(fdt->fd[fd], file);
1259	__set_open_fd(fd, fdt);
1260	if (flags & O_CLOEXEC)
1261		__set_close_on_exec(fd, fdt);
1262	else
1263		__clear_close_on_exec(fd, fdt);
1264	spin_unlock(&files->file_lock);
1265
1266	if (tofree)
1267		filp_close(tofree, files);
1268
1269	return fd;
1270
1271Ebusy:
1272	spin_unlock(&files->file_lock);
1273	return -EBUSY;
1274}
1275
1276int replace_fd(unsigned fd, struct file *file, unsigned flags)
1277{
1278	int err;
1279	struct files_struct *files = current->files;
1280
1281	if (!file)
1282		return close_fd(fd);
1283
1284	if (fd >= rlimit(RLIMIT_NOFILE))
1285		return -EBADF;
1286
1287	spin_lock(&files->file_lock);
1288	err = expand_files(files, fd);
1289	if (unlikely(err < 0))
1290		goto out_unlock;
1291	return do_dup2(files, file, fd, flags);
1292
1293out_unlock:
1294	spin_unlock(&files->file_lock);
1295	return err;
1296}
1297
1298/**
1299 * receive_fd() - Install received file into file descriptor table
1300 * @file: struct file that was received from another process
1301 * @ufd: __user pointer to write new fd number to
1302 * @o_flags: the O_* flags to apply to the new fd entry
1303 *
1304 * Installs a received file into the file descriptor table, with appropriate
1305 * checks and count updates. Optionally writes the fd number to userspace, if
1306 * @ufd is non-NULL.
1307 *
1308 * This helper handles its own reference counting of the incoming
1309 * struct file.
1310 *
1311 * Returns newly install fd or -ve on error.
1312 */
1313int receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1314{
1315	int new_fd;
1316	int error;
1317
1318	error = security_file_receive(file);
1319	if (error)
1320		return error;
1321
1322	new_fd = get_unused_fd_flags(o_flags);
1323	if (new_fd < 0)
1324		return new_fd;
1325
1326	if (ufd) {
1327		error = put_user(new_fd, ufd);
1328		if (error) {
1329			put_unused_fd(new_fd);
1330			return error;
1331		}
1332	}
1333
1334	fd_install(new_fd, get_file(file));
1335	__receive_sock(file);
1336	return new_fd;
1337}
1338EXPORT_SYMBOL_GPL(receive_fd);
1339
1340int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1341{
1342	int error;
1343
1344	error = security_file_receive(file);
1345	if (error)
1346		return error;
1347	error = replace_fd(new_fd, file, o_flags);
1348	if (error)
1349		return error;
1350	__receive_sock(file);
1351	return new_fd;
1352}
1353
1354static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1355{
1356	int err = -EBADF;
1357	struct file *file;
1358	struct files_struct *files = current->files;
1359
1360	if ((flags & ~O_CLOEXEC) != 0)
1361		return -EINVAL;
1362
1363	if (unlikely(oldfd == newfd))
1364		return -EINVAL;
1365
1366	if (newfd >= rlimit(RLIMIT_NOFILE))
1367		return -EBADF;
1368
1369	spin_lock(&files->file_lock);
1370	err = expand_files(files, newfd);
1371	file = files_lookup_fd_locked(files, oldfd);
1372	if (unlikely(!file))
1373		goto Ebadf;
1374	if (unlikely(err < 0)) {
1375		if (err == -EMFILE)
1376			goto Ebadf;
1377		goto out_unlock;
1378	}
1379	return do_dup2(files, file, newfd, flags);
1380
1381Ebadf:
1382	err = -EBADF;
1383out_unlock:
1384	spin_unlock(&files->file_lock);
1385	return err;
1386}
1387
1388SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1389{
1390	return ksys_dup3(oldfd, newfd, flags);
1391}
1392
1393SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1394{
1395	if (unlikely(newfd == oldfd)) { /* corner case */
1396		struct files_struct *files = current->files;
1397		struct file *f;
1398		int retval = oldfd;
1399
1400		rcu_read_lock();
1401		f = __fget_files_rcu(files, oldfd, 0);
1402		if (!f)
1403			retval = -EBADF;
1404		rcu_read_unlock();
1405		if (f)
1406			fput(f);
1407		return retval;
1408	}
1409	return ksys_dup3(oldfd, newfd, 0);
1410}
1411
1412SYSCALL_DEFINE1(dup, unsigned int, fildes)
1413{
1414	int ret = -EBADF;
1415	struct file *file = fget_raw(fildes);
1416
1417	if (file) {
1418		ret = get_unused_fd_flags(0);
1419		if (ret >= 0)
1420			fd_install(ret, file);
1421		else
1422			fput(file);
1423	}
1424	return ret;
1425}
1426
1427int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1428{
1429	unsigned long nofile = rlimit(RLIMIT_NOFILE);
1430	int err;
1431	if (from >= nofile)
1432		return -EINVAL;
1433	err = alloc_fd(from, nofile, flags);
1434	if (err >= 0) {
1435		get_file(file);
1436		fd_install(err, file);
1437	}
1438	return err;
1439}
1440
1441int iterate_fd(struct files_struct *files, unsigned n,
1442		int (*f)(const void *, struct file *, unsigned),
1443		const void *p)
1444{
1445	struct fdtable *fdt;
1446	int res = 0;
1447	if (!files)
1448		return 0;
1449	spin_lock(&files->file_lock);
1450	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1451		struct file *file;
1452		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1453		if (!file)
1454			continue;
1455		res = f(p, file, n);
1456		if (res)
1457			break;
1458	}
1459	spin_unlock(&files->file_lock);
1460	return res;
1461}
1462EXPORT_SYMBOL(iterate_fd);
v6.13.7
   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 <linux/file_ref.h>
  24#include <net/sock.h>
  25#include <linux/init_task.h>
  26
  27#include "internal.h"
  28
  29/**
  30 * __file_ref_put - Slowpath of file_ref_put()
  31 * @ref:	Pointer to the reference count
  32 * @cnt:	Current reference count
  33 *
  34 * Invoked when the reference count is outside of the valid zone.
  35 *
  36 * Return:
  37 *	True if this was the last reference with no future references
  38 *	possible. This signals the caller that it can safely schedule the
  39 *	object, which is protected by the reference counter, for
  40 *	deconstruction.
  41 *
  42 *	False if there are still active references or the put() raced
  43 *	with a concurrent get()/put() pair. Caller is not allowed to
  44 *	deconstruct the protected object.
  45 */
  46bool __file_ref_put(file_ref_t *ref, unsigned long cnt)
  47{
  48	/* Did this drop the last reference? */
  49	if (likely(cnt == FILE_REF_NOREF)) {
  50		/*
  51		 * Carefully try to set the reference count to FILE_REF_DEAD.
  52		 *
  53		 * This can fail if a concurrent get() operation has
  54		 * elevated it again or the corresponding put() even marked
  55		 * it dead already. Both are valid situations and do not
  56		 * require a retry. If this fails the caller is not
  57		 * allowed to deconstruct the object.
  58		 */
  59		if (!atomic_long_try_cmpxchg_release(&ref->refcnt, &cnt, FILE_REF_DEAD))
  60			return false;
  61
  62		/*
  63		 * The caller can safely schedule the object for
  64		 * deconstruction. Provide acquire ordering.
  65		 */
  66		smp_acquire__after_ctrl_dep();
  67		return true;
  68	}
  69
  70	/*
  71	 * If the reference count was already in the dead zone, then this
  72	 * put() operation is imbalanced. Warn, put the reference count back to
  73	 * DEAD and tell the caller to not deconstruct the object.
  74	 */
  75	if (WARN_ONCE(cnt >= FILE_REF_RELEASED, "imbalanced put on file reference count")) {
  76		atomic_long_set(&ref->refcnt, FILE_REF_DEAD);
  77		return false;
  78	}
  79
  80	/*
  81	 * This is a put() operation on a saturated refcount. Restore the
  82	 * mean saturation value and tell the caller to not deconstruct the
  83	 * object.
  84	 */
  85	if (cnt > FILE_REF_MAXREF)
  86		atomic_long_set(&ref->refcnt, FILE_REF_SATURATED);
  87	return false;
  88}
  89EXPORT_SYMBOL_GPL(__file_ref_put);
  90
  91unsigned int sysctl_nr_open __read_mostly = 1024*1024;
  92unsigned int sysctl_nr_open_min = BITS_PER_LONG;
  93/* our min() is unusable in constant expressions ;-/ */
  94#define __const_min(x, y) ((x) < (y) ? (x) : (y))
  95unsigned int sysctl_nr_open_max =
  96	__const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
  97
  98static void __free_fdtable(struct fdtable *fdt)
  99{
 100	kvfree(fdt->fd);
 101	kvfree(fdt->open_fds);
 102	kfree(fdt);
 103}
 104
 105static void free_fdtable_rcu(struct rcu_head *rcu)
 106{
 107	__free_fdtable(container_of(rcu, struct fdtable, rcu));
 108}
 109
 110#define BITBIT_NR(nr)	BITS_TO_LONGS(BITS_TO_LONGS(nr))
 111#define BITBIT_SIZE(nr)	(BITBIT_NR(nr) * sizeof(long))
 112
 113#define fdt_words(fdt) ((fdt)->max_fds / BITS_PER_LONG) // words in ->open_fds
 114/*
 115 * Copy 'count' fd bits from the old table to the new table and clear the extra
 116 * space if any.  This does not copy the file pointers.  Called with the files
 117 * spinlock held for write.
 118 */
 119static inline void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
 120			    unsigned int copy_words)
 121{
 122	unsigned int nwords = fdt_words(nfdt);
 123
 124	bitmap_copy_and_extend(nfdt->open_fds, ofdt->open_fds,
 125			copy_words * BITS_PER_LONG, nwords * BITS_PER_LONG);
 126	bitmap_copy_and_extend(nfdt->close_on_exec, ofdt->close_on_exec,
 127			copy_words * BITS_PER_LONG, nwords * BITS_PER_LONG);
 128	bitmap_copy_and_extend(nfdt->full_fds_bits, ofdt->full_fds_bits,
 129			copy_words, nwords);
 
 
 
 
 
 130}
 131
 132/*
 133 * Copy all file descriptors from the old table to the new, expanded table and
 134 * clear the extra space.  Called with the files spinlock held for write.
 135 */
 136static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 137{
 138	size_t cpy, set;
 139
 140	BUG_ON(nfdt->max_fds < ofdt->max_fds);
 141
 142	cpy = ofdt->max_fds * sizeof(struct file *);
 143	set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 144	memcpy(nfdt->fd, ofdt->fd, cpy);
 145	memset((char *)nfdt->fd + cpy, 0, set);
 146
 147	copy_fd_bitmaps(nfdt, ofdt, fdt_words(ofdt));
 148}
 149
 150/*
 151 * Note how the fdtable bitmap allocations very much have to be a multiple of
 152 * BITS_PER_LONG. This is not only because we walk those things in chunks of
 153 * 'unsigned long' in some places, but simply because that is how the Linux
 154 * kernel bitmaps are defined to work: they are not "bits in an array of bytes",
 155 * they are very much "bits in an array of unsigned long".
 
 
 
 
 
 
 
 
 156 */
 157static struct fdtable *alloc_fdtable(unsigned int slots_wanted)
 158{
 159	struct fdtable *fdt;
 160	unsigned int nr;
 161	void *data;
 162
 163	/*
 164	 * Figure out how many fds we actually want to support in this fdtable.
 165	 * Allocation steps are keyed to the size of the fdarray, since it
 166	 * grows far faster than any of the other dynamic data. We try to fit
 167	 * the fdarray into comfortable page-tuned chunks: starting at 1024B
 168	 * and growing in powers of two from there on.  Since we called only
 169	 * with slots_wanted > BITS_PER_LONG (embedded instance in files->fdtab
 170	 * already gives BITS_PER_LONG slots), the above boils down to
 171	 * 1.  use the smallest power of two large enough to give us that many
 172	 * slots.
 173	 * 2.  on 32bit skip 64 and 128 - the minimal capacity we want there is
 174	 * 256 slots (i.e. 1Kb fd array).
 175	 * 3.  on 64bit don't skip anything, 1Kb fd array means 128 slots there
 176	 * and we are never going to be asked for 64 or less.
 177	 */
 178	if (IS_ENABLED(CONFIG_32BIT) && slots_wanted < 256)
 179		nr = 256;
 180	else
 181		nr = roundup_pow_of_two(slots_wanted);
 182	/*
 183	 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
 184	 * had been set lower between the check in expand_files() and here.
 
 185	 *
 186	 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
 187	 * bitmaps handling below becomes unpleasant, to put it mildly...
 188	 */
 189	if (unlikely(nr > sysctl_nr_open)) {
 190		nr = round_down(sysctl_nr_open, BITS_PER_LONG);
 191		if (nr < slots_wanted)
 192			return ERR_PTR(-EMFILE);
 193	}
 194
 195	fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
 196	if (!fdt)
 197		goto out;
 198	fdt->max_fds = nr;
 199	data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
 200	if (!data)
 201		goto out_fdt;
 202	fdt->fd = data;
 203
 204	data = kvmalloc(max_t(size_t,
 205				 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
 206				 GFP_KERNEL_ACCOUNT);
 207	if (!data)
 208		goto out_arr;
 209	fdt->open_fds = data;
 210	data += nr / BITS_PER_BYTE;
 211	fdt->close_on_exec = data;
 212	data += nr / BITS_PER_BYTE;
 213	fdt->full_fds_bits = data;
 214
 215	return fdt;
 216
 217out_arr:
 218	kvfree(fdt->fd);
 219out_fdt:
 220	kfree(fdt);
 221out:
 222	return ERR_PTR(-ENOMEM);
 223}
 224
 225/*
 226 * Expand the file descriptor table.
 227 * This function will allocate a new fdtable and both fd array and fdset, of
 228 * the given size.
 229 * Return <0 error code on error; 0 on successful completion.
 230 * The files->file_lock should be held on entry, and will be held on exit.
 231 */
 232static int expand_fdtable(struct files_struct *files, unsigned int nr)
 233	__releases(files->file_lock)
 234	__acquires(files->file_lock)
 235{
 236	struct fdtable *new_fdt, *cur_fdt;
 237
 238	spin_unlock(&files->file_lock);
 239	new_fdt = alloc_fdtable(nr + 1);
 240
 241	/* make sure all fd_install() have seen resize_in_progress
 242	 * or have finished their rcu_read_lock_sched() section.
 243	 */
 244	if (atomic_read(&files->count) > 1)
 245		synchronize_rcu();
 246
 247	spin_lock(&files->file_lock);
 248	if (IS_ERR(new_fdt))
 249		return PTR_ERR(new_fdt);
 
 
 
 
 
 
 
 
 250	cur_fdt = files_fdtable(files);
 251	BUG_ON(nr < cur_fdt->max_fds);
 252	copy_fdtable(new_fdt, cur_fdt);
 253	rcu_assign_pointer(files->fdt, new_fdt);
 254	if (cur_fdt != &files->fdtab)
 255		call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
 256	/* coupled with smp_rmb() in fd_install() */
 257	smp_wmb();
 258	return 0;
 259}
 260
 261/*
 262 * Expand files.
 263 * This function will expand the file structures, if the requested size exceeds
 264 * the current capacity and there is room for expansion.
 265 * Return <0 error code on error; 0 on success.
 
 266 * The files->file_lock should be held on entry, and will be held on exit.
 267 */
 268static int expand_files(struct files_struct *files, unsigned int nr)
 269	__releases(files->file_lock)
 270	__acquires(files->file_lock)
 271{
 272	struct fdtable *fdt;
 273	int error;
 274
 275repeat:
 276	fdt = files_fdtable(files);
 277
 278	/* Do we need to expand? */
 279	if (nr < fdt->max_fds)
 280		return 0;
 281
 282	/* Can we expand? */
 283	if (nr >= sysctl_nr_open)
 284		return -EMFILE;
 285
 286	if (unlikely(files->resize_in_progress)) {
 287		spin_unlock(&files->file_lock);
 
 288		wait_event(files->resize_wait, !files->resize_in_progress);
 289		spin_lock(&files->file_lock);
 290		goto repeat;
 291	}
 292
 293	/* All good, so we try */
 294	files->resize_in_progress = true;
 295	error = expand_fdtable(files, nr);
 296	files->resize_in_progress = false;
 297
 298	wake_up_all(&files->resize_wait);
 299	return error;
 
 
 
 
 
 300}
 301
 302static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt,
 303				       bool set)
 304{
 305	if (set) {
 306		__set_bit(fd, fdt->close_on_exec);
 307	} else {
 308		if (test_bit(fd, fdt->close_on_exec))
 309			__clear_bit(fd, fdt->close_on_exec);
 310	}
 311}
 312
 313static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt, bool set)
 314{
 315	__set_bit(fd, fdt->open_fds);
 316	__set_close_on_exec(fd, fdt, set);
 317	fd /= BITS_PER_LONG;
 318	if (!~fdt->open_fds[fd])
 319		__set_bit(fd, fdt->full_fds_bits);
 320}
 321
 322static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
 323{
 324	__clear_bit(fd, fdt->open_fds);
 325	fd /= BITS_PER_LONG;
 326	if (test_bit(fd, fdt->full_fds_bits))
 327		__clear_bit(fd, fdt->full_fds_bits);
 328}
 329
 330static inline bool fd_is_open(unsigned int fd, const struct fdtable *fdt)
 331{
 332	return test_bit(fd, fdt->open_fds);
 
 
 
 
 
 
 
 
 
 333}
 334
 335/*
 336 * Note that a sane fdtable size always has to be a multiple of
 337 * BITS_PER_LONG, since we have bitmaps that are sized by this.
 338 *
 339 * punch_hole is optional - when close_range() is asked to unshare
 340 * and close, we don't need to copy descriptors in that range, so
 341 * a smaller cloned descriptor table might suffice if the last
 342 * currently opened descriptor falls into that range.
 
 
 
 
 343 */
 344static unsigned int sane_fdtable_size(struct fdtable *fdt, struct fd_range *punch_hole)
 345{
 346	unsigned int last = find_last_bit(fdt->open_fds, fdt->max_fds);
 347
 348	if (last == fdt->max_fds)
 349		return NR_OPEN_DEFAULT;
 350	if (punch_hole && punch_hole->to >= last && punch_hole->from <= last) {
 351		last = find_last_bit(fdt->open_fds, punch_hole->from);
 352		if (last == punch_hole->from)
 353			return NR_OPEN_DEFAULT;
 354	}
 355	return ALIGN(last + 1, BITS_PER_LONG);
 356}
 357
 358/*
 359 * Allocate a new descriptor table and copy contents from the passed in
 360 * instance.  Returns a pointer to cloned table on success, ERR_PTR()
 361 * on failure.  For 'punch_hole' see sane_fdtable_size().
 362 */
 363struct files_struct *dup_fd(struct files_struct *oldf, struct fd_range *punch_hole)
 364{
 365	struct files_struct *newf;
 366	struct file **old_fds, **new_fds;
 367	unsigned int open_files, i;
 368	struct fdtable *old_fdt, *new_fdt;
 369
 
 370	newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
 371	if (!newf)
 372		return ERR_PTR(-ENOMEM);
 373
 374	atomic_set(&newf->count, 1);
 375
 376	spin_lock_init(&newf->file_lock);
 377	newf->resize_in_progress = false;
 378	init_waitqueue_head(&newf->resize_wait);
 379	newf->next_fd = 0;
 380	new_fdt = &newf->fdtab;
 381	new_fdt->max_fds = NR_OPEN_DEFAULT;
 382	new_fdt->close_on_exec = newf->close_on_exec_init;
 383	new_fdt->open_fds = newf->open_fds_init;
 384	new_fdt->full_fds_bits = newf->full_fds_bits_init;
 385	new_fdt->fd = &newf->fd_array[0];
 386
 387	spin_lock(&oldf->file_lock);
 388	old_fdt = files_fdtable(oldf);
 389	open_files = sane_fdtable_size(old_fdt, punch_hole);
 390
 391	/*
 392	 * Check whether we need to allocate a larger fd array and fd set.
 393	 */
 394	while (unlikely(open_files > new_fdt->max_fds)) {
 395		spin_unlock(&oldf->file_lock);
 396
 397		if (new_fdt != &newf->fdtab)
 398			__free_fdtable(new_fdt);
 399
 400		new_fdt = alloc_fdtable(open_files);
 401		if (IS_ERR(new_fdt)) {
 402			kmem_cache_free(files_cachep, newf);
 403			return ERR_CAST(new_fdt);
 
 
 
 
 
 
 
 404		}
 405
 406		/*
 407		 * Reacquire the oldf lock and a pointer to its fd table
 408		 * who knows it may have a new bigger fd table. We need
 409		 * the latest pointer.
 410		 */
 411		spin_lock(&oldf->file_lock);
 412		old_fdt = files_fdtable(oldf);
 413		open_files = sane_fdtable_size(old_fdt, punch_hole);
 414	}
 415
 416	copy_fd_bitmaps(new_fdt, old_fdt, open_files / BITS_PER_LONG);
 417
 418	old_fds = old_fdt->fd;
 419	new_fds = new_fdt->fd;
 420
 421	for (i = open_files; i != 0; i--) {
 422		struct file *f = *old_fds++;
 423		if (f) {
 424			get_file(f);
 425		} else {
 426			/*
 427			 * The fd may be claimed in the fd bitmap but not yet
 428			 * instantiated in the files array if a sibling thread
 429			 * is partway through open().  So make sure that this
 430			 * fd is available to the new process.
 431			 */
 432			__clear_open_fd(open_files - i, new_fdt);
 433		}
 434		rcu_assign_pointer(*new_fds++, f);
 435	}
 436	spin_unlock(&oldf->file_lock);
 437
 438	/* clear the remainder */
 439	memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
 440
 441	rcu_assign_pointer(newf->fdt, new_fdt);
 442
 443	return newf;
 
 
 
 
 
 444}
 445
 446static struct fdtable *close_files(struct files_struct * files)
 447{
 448	/*
 449	 * It is safe to dereference the fd table without RCU or
 450	 * ->file_lock because this is the last reference to the
 451	 * files structure.
 452	 */
 453	struct fdtable *fdt = rcu_dereference_raw(files->fdt);
 454	unsigned int i, j = 0;
 455
 456	for (;;) {
 457		unsigned long set;
 458		i = j * BITS_PER_LONG;
 459		if (i >= fdt->max_fds)
 460			break;
 461		set = fdt->open_fds[j++];
 462		while (set) {
 463			if (set & 1) {
 464				struct file *file = fdt->fd[i];
 465				if (file) {
 466					filp_close(file, files);
 467					cond_resched();
 468				}
 469			}
 470			i++;
 471			set >>= 1;
 472		}
 473	}
 474
 475	return fdt;
 476}
 477
 478void put_files_struct(struct files_struct *files)
 479{
 480	if (atomic_dec_and_test(&files->count)) {
 481		struct fdtable *fdt = close_files(files);
 482
 483		/* free the arrays if they are not embedded */
 484		if (fdt != &files->fdtab)
 485			__free_fdtable(fdt);
 486		kmem_cache_free(files_cachep, files);
 487	}
 488}
 489
 490void exit_files(struct task_struct *tsk)
 491{
 492	struct files_struct * files = tsk->files;
 493
 494	if (files) {
 495		task_lock(tsk);
 496		tsk->files = NULL;
 497		task_unlock(tsk);
 498		put_files_struct(files);
 499	}
 500}
 501
 502struct files_struct init_files = {
 503	.count		= ATOMIC_INIT(1),
 504	.fdt		= &init_files.fdtab,
 505	.fdtab		= {
 506		.max_fds	= NR_OPEN_DEFAULT,
 507		.fd		= &init_files.fd_array[0],
 508		.close_on_exec	= init_files.close_on_exec_init,
 509		.open_fds	= init_files.open_fds_init,
 510		.full_fds_bits	= init_files.full_fds_bits_init,
 511	},
 512	.file_lock	= __SPIN_LOCK_UNLOCKED(init_files.file_lock),
 513	.resize_wait	= __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
 514};
 515
 516static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
 517{
 518	unsigned int maxfd = fdt->max_fds; /* always multiple of BITS_PER_LONG */
 519	unsigned int maxbit = maxfd / BITS_PER_LONG;
 520	unsigned int bitbit = start / BITS_PER_LONG;
 521	unsigned int bit;
 522
 523	/*
 524	 * Try to avoid looking at the second level bitmap
 525	 */
 526	bit = find_next_zero_bit(&fdt->open_fds[bitbit], BITS_PER_LONG,
 527				 start & (BITS_PER_LONG - 1));
 528	if (bit < BITS_PER_LONG)
 529		return bit + bitbit * BITS_PER_LONG;
 530
 531	bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
 532	if (bitbit >= maxfd)
 533		return maxfd;
 534	if (bitbit > start)
 535		start = bitbit;
 536	return find_next_zero_bit(fdt->open_fds, maxfd, start);
 537}
 538
 539/*
 540 * allocate a file descriptor, mark it busy.
 541 */
 542static int alloc_fd(unsigned start, unsigned end, unsigned flags)
 543{
 544	struct files_struct *files = current->files;
 545	unsigned int fd;
 546	int error;
 547	struct fdtable *fdt;
 548
 549	spin_lock(&files->file_lock);
 550repeat:
 551	fdt = files_fdtable(files);
 552	fd = start;
 553	if (fd < files->next_fd)
 554		fd = files->next_fd;
 555
 556	if (likely(fd < fdt->max_fds))
 557		fd = find_next_fd(fdt, fd);
 558
 559	/*
 560	 * N.B. For clone tasks sharing a files structure, this test
 561	 * will limit the total number of files that can be opened.
 562	 */
 563	error = -EMFILE;
 564	if (unlikely(fd >= end))
 565		goto out;
 566
 567	if (unlikely(fd >= fdt->max_fds)) {
 568		error = expand_files(files, fd);
 569		if (error < 0)
 570			goto out;
 571
 
 
 
 
 
 572		goto repeat;
 573	}
 574
 575	if (start <= files->next_fd)
 576		files->next_fd = fd + 1;
 577
 578	__set_open_fd(fd, fdt, flags & O_CLOEXEC);
 
 
 
 
 579	error = fd;
 
 
 
 
 
 
 
 580
 581out:
 582	spin_unlock(&files->file_lock);
 583	return error;
 584}
 585
 586int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
 587{
 588	return alloc_fd(0, nofile, flags);
 589}
 590
 591int get_unused_fd_flags(unsigned flags)
 592{
 593	return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
 594}
 595EXPORT_SYMBOL(get_unused_fd_flags);
 596
 597static void __put_unused_fd(struct files_struct *files, unsigned int fd)
 598{
 599	struct fdtable *fdt = files_fdtable(files);
 600	__clear_open_fd(fd, fdt);
 601	if (fd < files->next_fd)
 602		files->next_fd = fd;
 603}
 604
 605void put_unused_fd(unsigned int fd)
 606{
 607	struct files_struct *files = current->files;
 608	spin_lock(&files->file_lock);
 609	__put_unused_fd(files, fd);
 610	spin_unlock(&files->file_lock);
 611}
 612
 613EXPORT_SYMBOL(put_unused_fd);
 614
 615/*
 616 * Install a file pointer in the fd array.
 617 *
 618 * The VFS is full of places where we drop the files lock between
 619 * setting the open_fds bitmap and installing the file in the file
 620 * array.  At any such point, we are vulnerable to a dup2() race
 621 * installing a file in the array before us.  We need to detect this and
 622 * fput() the struct file we are about to overwrite in this case.
 623 *
 624 * It should never happen - if we allow dup2() do it, _really_ bad things
 625 * will follow.
 626 *
 627 * This consumes the "file" refcount, so callers should treat it
 628 * as if they had called fput(file).
 629 */
 630
 631void fd_install(unsigned int fd, struct file *file)
 632{
 633	struct files_struct *files = current->files;
 634	struct fdtable *fdt;
 635
 636	if (WARN_ON_ONCE(unlikely(file->f_mode & FMODE_BACKING)))
 637		return;
 638
 639	rcu_read_lock_sched();
 640
 641	if (unlikely(files->resize_in_progress)) {
 642		rcu_read_unlock_sched();
 643		spin_lock(&files->file_lock);
 644		fdt = files_fdtable(files);
 645		WARN_ON(fdt->fd[fd] != NULL);
 646		rcu_assign_pointer(fdt->fd[fd], file);
 647		spin_unlock(&files->file_lock);
 648		return;
 649	}
 650	/* coupled with smp_wmb() in expand_fdtable() */
 651	smp_rmb();
 652	fdt = rcu_dereference_sched(files->fdt);
 653	BUG_ON(fdt->fd[fd] != NULL);
 654	rcu_assign_pointer(fdt->fd[fd], file);
 655	rcu_read_unlock_sched();
 656}
 657
 658EXPORT_SYMBOL(fd_install);
 659
 660/**
 661 * file_close_fd_locked - return file associated with fd
 662 * @files: file struct to retrieve file from
 663 * @fd: file descriptor to retrieve file for
 664 *
 665 * Doesn't take a separate reference count.
 666 *
 667 * Context: files_lock must be held.
 668 *
 669 * Returns: The file associated with @fd (NULL if @fd is not open)
 670 */
 671struct file *file_close_fd_locked(struct files_struct *files, unsigned fd)
 672{
 673	struct fdtable *fdt = files_fdtable(files);
 674	struct file *file;
 675
 676	lockdep_assert_held(&files->file_lock);
 677
 678	if (fd >= fdt->max_fds)
 679		return NULL;
 680
 681	fd = array_index_nospec(fd, fdt->max_fds);
 682	file = fdt->fd[fd];
 683	if (file) {
 684		rcu_assign_pointer(fdt->fd[fd], NULL);
 685		__put_unused_fd(files, fd);
 686	}
 687	return file;
 688}
 689
 690int close_fd(unsigned fd)
 691{
 692	struct files_struct *files = current->files;
 693	struct file *file;
 694
 695	spin_lock(&files->file_lock);
 696	file = file_close_fd_locked(files, fd);
 697	spin_unlock(&files->file_lock);
 698	if (!file)
 699		return -EBADF;
 700
 701	return filp_close(file, files);
 702}
 703EXPORT_SYMBOL(close_fd);
 704
 705/**
 706 * last_fd - return last valid index into fd table
 707 * @fdt: File descriptor table.
 708 *
 709 * Context: Either rcu read lock or files_lock must be held.
 710 *
 711 * Returns: Last valid index into fdtable.
 712 */
 713static inline unsigned last_fd(struct fdtable *fdt)
 714{
 715	return fdt->max_fds - 1;
 716}
 717
 718static inline void __range_cloexec(struct files_struct *cur_fds,
 719				   unsigned int fd, unsigned int max_fd)
 720{
 721	struct fdtable *fdt;
 722
 723	/* make sure we're using the correct maximum value */
 724	spin_lock(&cur_fds->file_lock);
 725	fdt = files_fdtable(cur_fds);
 726	max_fd = min(last_fd(fdt), max_fd);
 727	if (fd <= max_fd)
 728		bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
 729	spin_unlock(&cur_fds->file_lock);
 730}
 731
 732static inline void __range_close(struct files_struct *files, unsigned int fd,
 733				 unsigned int max_fd)
 734{
 735	struct file *file;
 736	unsigned n;
 737
 738	spin_lock(&files->file_lock);
 739	n = last_fd(files_fdtable(files));
 740	max_fd = min(max_fd, n);
 741
 742	for (; fd <= max_fd; fd++) {
 743		file = file_close_fd_locked(files, fd);
 744		if (file) {
 745			spin_unlock(&files->file_lock);
 746			filp_close(file, files);
 747			cond_resched();
 748			spin_lock(&files->file_lock);
 749		} else if (need_resched()) {
 750			spin_unlock(&files->file_lock);
 751			cond_resched();
 752			spin_lock(&files->file_lock);
 753		}
 754	}
 755	spin_unlock(&files->file_lock);
 756}
 757
 758/**
 759 * sys_close_range() - Close all file descriptors in a given range.
 760 *
 761 * @fd:     starting file descriptor to close
 762 * @max_fd: last file descriptor to close
 763 * @flags:  CLOSE_RANGE flags.
 764 *
 765 * This closes a range of file descriptors. All file descriptors
 766 * from @fd up to and including @max_fd are closed.
 767 * Currently, errors to close a given file descriptor are ignored.
 768 */
 769SYSCALL_DEFINE3(close_range, unsigned int, fd, unsigned int, max_fd,
 770		unsigned int, flags)
 771{
 772	struct task_struct *me = current;
 773	struct files_struct *cur_fds = me->files, *fds = NULL;
 774
 775	if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
 776		return -EINVAL;
 777
 778	if (fd > max_fd)
 779		return -EINVAL;
 780
 781	if ((flags & CLOSE_RANGE_UNSHARE) && atomic_read(&cur_fds->count) > 1) {
 782		struct fd_range range = {fd, max_fd}, *punch_hole = &range;
 
 783
 784		/*
 785		 * If the caller requested all fds to be made cloexec we always
 786		 * copy all of the file descriptors since they still want to
 787		 * use them.
 788		 */
 789		if (flags & CLOSE_RANGE_CLOEXEC)
 790			punch_hole = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 791
 792		fds = dup_fd(cur_fds, punch_hole);
 793		if (IS_ERR(fds))
 794			return PTR_ERR(fds);
 795		/*
 796		 * We used to share our file descriptor table, and have now
 797		 * created a private one, make sure we're using it below.
 798		 */
 799		swap(cur_fds, fds);
 
 800	}
 801
 802	if (flags & CLOSE_RANGE_CLOEXEC)
 803		__range_cloexec(cur_fds, fd, max_fd);
 804	else
 805		__range_close(cur_fds, fd, max_fd);
 806
 807	if (fds) {
 808		/*
 809		 * We're done closing the files we were supposed to. Time to install
 810		 * the new file descriptor table and drop the old one.
 811		 */
 812		task_lock(me);
 813		me->files = cur_fds;
 814		task_unlock(me);
 815		put_files_struct(fds);
 816	}
 817
 818	return 0;
 819}
 820
 821/**
 822 * file_close_fd - return file associated with fd
 823 * @fd: file descriptor to retrieve file for
 824 *
 825 * Doesn't take a separate reference count.
 826 *
 827 * Returns: The file associated with @fd (NULL if @fd is not open)
 828 */
 829struct file *file_close_fd(unsigned int fd)
 830{
 831	struct files_struct *files = current->files;
 832	struct file *file;
 833
 834	spin_lock(&files->file_lock);
 835	file = file_close_fd_locked(files, fd);
 836	spin_unlock(&files->file_lock);
 837
 838	return file;
 839}
 840
 841void do_close_on_exec(struct files_struct *files)
 842{
 843	unsigned i;
 844	struct fdtable *fdt;
 845
 846	/* exec unshares first */
 847	spin_lock(&files->file_lock);
 848	for (i = 0; ; i++) {
 849		unsigned long set;
 850		unsigned fd = i * BITS_PER_LONG;
 851		fdt = files_fdtable(files);
 852		if (fd >= fdt->max_fds)
 853			break;
 854		set = fdt->close_on_exec[i];
 855		if (!set)
 856			continue;
 857		fdt->close_on_exec[i] = 0;
 858		for ( ; set ; fd++, set >>= 1) {
 859			struct file *file;
 860			if (!(set & 1))
 861				continue;
 862			file = fdt->fd[fd];
 863			if (!file)
 864				continue;
 865			rcu_assign_pointer(fdt->fd[fd], NULL);
 866			__put_unused_fd(files, fd);
 867			spin_unlock(&files->file_lock);
 868			filp_close(file, files);
 869			cond_resched();
 870			spin_lock(&files->file_lock);
 871		}
 872
 873	}
 874	spin_unlock(&files->file_lock);
 875}
 876
 877static struct file *__get_file_rcu(struct file __rcu **f)
 878{
 879	struct file __rcu *file;
 880	struct file __rcu *file_reloaded;
 881	struct file __rcu *file_reloaded_cmp;
 882
 883	file = rcu_dereference_raw(*f);
 884	if (!file)
 885		return NULL;
 886
 887	if (unlikely(!file_ref_get(&file->f_ref)))
 888		return ERR_PTR(-EAGAIN);
 889
 890	file_reloaded = rcu_dereference_raw(*f);
 891
 892	/*
 893	 * Ensure that all accesses have a dependency on the load from
 894	 * rcu_dereference_raw() above so we get correct ordering
 895	 * between reuse/allocation and the pointer check below.
 896	 */
 897	file_reloaded_cmp = file_reloaded;
 898	OPTIMIZER_HIDE_VAR(file_reloaded_cmp);
 899
 900	/*
 901	 * file_ref_get() above provided a full memory barrier when we
 902	 * acquired a reference.
 903	 *
 904	 * This is paired with the write barrier from assigning to the
 905	 * __rcu protected file pointer so that if that pointer still
 906	 * matches the current file, we know we have successfully
 907	 * acquired a reference to the right file.
 908	 *
 909	 * If the pointers don't match the file has been reallocated by
 910	 * SLAB_TYPESAFE_BY_RCU.
 911	 */
 912	if (file == file_reloaded_cmp)
 913		return file_reloaded;
 914
 915	fput(file);
 916	return ERR_PTR(-EAGAIN);
 917}
 918
 919/**
 920 * get_file_rcu - try go get a reference to a file under rcu
 921 * @f: the file to get a reference on
 922 *
 923 * This function tries to get a reference on @f carefully verifying that
 924 * @f hasn't been reused.
 925 *
 926 * This function should rarely have to be used and only by users who
 927 * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it.
 928 *
 929 * Return: Returns @f with the reference count increased or NULL.
 930 */
 931struct file *get_file_rcu(struct file __rcu **f)
 932{
 933	for (;;) {
 934		struct file __rcu *file;
 935
 936		file = __get_file_rcu(f);
 937		if (!IS_ERR(file))
 938			return file;
 
 
 
 
 
 939	}
 940}
 941EXPORT_SYMBOL_GPL(get_file_rcu);
 942
 943/**
 944 * get_file_active - try go get a reference to a file
 945 * @f: the file to get a reference on
 946 *
 947 * In contast to get_file_rcu() the pointer itself isn't part of the
 948 * reference counting.
 949 *
 950 * This function should rarely have to be used and only by users who
 951 * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it.
 952 *
 953 * Return: Returns @f with the reference count increased or NULL.
 954 */
 955struct file *get_file_active(struct file **f)
 956{
 957	struct file __rcu *file;
 958
 959	rcu_read_lock();
 960	file = __get_file_rcu(f);
 961	rcu_read_unlock();
 962	if (IS_ERR(file))
 963		file = NULL;
 964	return file;
 965}
 966EXPORT_SYMBOL_GPL(get_file_active);
 967
 968static inline struct file *__fget_files_rcu(struct files_struct *files,
 969       unsigned int fd, fmode_t mask)
 970{
 971	for (;;) {
 972		struct file *file;
 973		struct fdtable *fdt = rcu_dereference_raw(files->fdt);
 974		struct file __rcu **fdentry;
 975		unsigned long nospec_mask;
 976
 977		/* Mask is a 0 for invalid fd's, ~0 for valid ones */
 978		nospec_mask = array_index_mask_nospec(fd, fdt->max_fds);
 979
 980		/*
 981		 * fdentry points to the 'fd' offset, or fdt->fd[0].
 982		 * Loading from fdt->fd[0] is always safe, because the
 983		 * array always exists.
 984		 */
 985		fdentry = fdt->fd + (fd & nospec_mask);
 986
 987		/* Do the load, then mask any invalid result */
 988		file = rcu_dereference_raw(*fdentry);
 989		file = (void *)(nospec_mask & (unsigned long)file);
 990		if (unlikely(!file))
 991			return NULL;
 992
 993		/*
 994		 * Ok, we have a file pointer that was valid at
 995		 * some point, but it might have become stale since.
 996		 *
 997		 * We need to confirm it by incrementing the refcount
 998		 * and then check the lookup again.
 999		 *
1000		 * file_ref_get() gives us a full memory barrier. We
1001		 * only really need an 'acquire' one to protect the
1002		 * loads below, but we don't have that.
1003		 */
1004		if (unlikely(!file_ref_get(&file->f_ref)))
1005			continue;
1006
1007		/*
1008		 * Such a race can take two forms:
1009		 *
1010		 *  (a) the file ref already went down to zero and the
1011		 *      file hasn't been reused yet or the file count
1012		 *      isn't zero but the file has already been reused.
1013		 *
1014		 *  (b) the file table entry has changed under us.
1015		 *       Note that we don't need to re-check the 'fdt->fd'
1016		 *       pointer having changed, because it always goes
1017		 *       hand-in-hand with 'fdt'.
1018		 *
1019		 * If so, we need to put our ref and try again.
1020		 */
1021		if (unlikely(file != rcu_dereference_raw(*fdentry)) ||
1022		    unlikely(rcu_dereference_raw(files->fdt) != fdt)) {
1023			fput(file);
1024			continue;
1025		}
1026
1027		/*
1028		 * This isn't the file we're looking for or we're not
1029		 * allowed to get a reference to it.
1030		 */
1031		if (unlikely(file->f_mode & mask)) {
1032			fput(file);
1033			return NULL;
1034		}
1035
1036		/*
1037		 * Ok, we have a ref to the file, and checked that it
1038		 * still exists.
1039		 */
1040		return file;
1041	}
1042}
1043
1044static struct file *__fget_files(struct files_struct *files, unsigned int fd,
1045				 fmode_t mask)
1046{
1047	struct file *file;
1048
1049	rcu_read_lock();
1050	file = __fget_files_rcu(files, fd, mask);
1051	rcu_read_unlock();
1052
1053	return file;
1054}
1055
1056static inline struct file *__fget(unsigned int fd, fmode_t mask)
1057{
1058	return __fget_files(current->files, fd, mask);
1059}
1060
1061struct file *fget(unsigned int fd)
1062{
1063	return __fget(fd, FMODE_PATH);
1064}
1065EXPORT_SYMBOL(fget);
1066
1067struct file *fget_raw(unsigned int fd)
1068{
1069	return __fget(fd, 0);
1070}
1071EXPORT_SYMBOL(fget_raw);
1072
1073struct file *fget_task(struct task_struct *task, unsigned int fd)
1074{
1075	struct file *file = NULL;
1076
1077	task_lock(task);
1078	if (task->files)
1079		file = __fget_files(task->files, fd, 0);
1080	task_unlock(task);
1081
1082	return file;
1083}
1084
1085struct file *fget_task_next(struct task_struct *task, unsigned int *ret_fd)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1086{
1087	/* Must be called with rcu_read_lock held */
1088	struct files_struct *files;
1089	unsigned int fd = *ret_fd;
1090	struct file *file = NULL;
1091
1092	task_lock(task);
1093	files = task->files;
1094	if (files) {
1095		rcu_read_lock();
1096		for (; fd < files_fdtable(files)->max_fds; fd++) {
1097			file = __fget_files_rcu(files, fd, 0);
1098			if (file)
1099				break;
1100		}
1101		rcu_read_unlock();
1102	}
1103	task_unlock(task);
1104	*ret_fd = fd;
1105	return file;
1106}
1107EXPORT_SYMBOL(fget_task_next);
1108
1109/*
1110 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
1111 *
1112 * You can use this instead of fget if you satisfy all of the following
1113 * conditions:
1114 * 1) You must call fput_light before exiting the syscall and returning control
1115 *    to userspace (i.e. you cannot remember the returned struct file * after
1116 *    returning to userspace).
1117 * 2) You must not call filp_close on the returned struct file * in between
1118 *    calls to fget_light and fput_light.
1119 * 3) You must not clone the current task in between the calls to fget_light
1120 *    and fput_light.
1121 *
1122 * The fput_needed flag returned by fget_light should be passed to the
1123 * corresponding fput_light.
1124 *
1125 * (As an exception to rule 2, you can call filp_close between fget_light and
1126 * fput_light provided that you capture a real refcount with get_file before
1127 * the call to filp_close, and ensure that this real refcount is fput *after*
1128 * the fput_light call.)
1129 *
1130 * See also the documentation in rust/kernel/file.rs.
1131 */
1132static inline struct fd __fget_light(unsigned int fd, fmode_t mask)
1133{
1134	struct files_struct *files = current->files;
1135	struct file *file;
1136
1137	/*
1138	 * If another thread is concurrently calling close_fd() followed
1139	 * by put_files_struct(), we must not observe the old table
1140	 * entry combined with the new refcount - otherwise we could
1141	 * return a file that is concurrently being freed.
1142	 *
1143	 * atomic_read_acquire() pairs with atomic_dec_and_test() in
1144	 * put_files_struct().
1145	 */
1146	if (likely(atomic_read_acquire(&files->count) == 1)) {
1147		file = files_lookup_fd_raw(files, fd);
1148		if (!file || unlikely(file->f_mode & mask))
1149			return EMPTY_FD;
1150		return BORROWED_FD(file);
1151	} else {
1152		file = __fget_files(files, fd, mask);
1153		if (!file)
1154			return EMPTY_FD;
1155		return CLONED_FD(file);
1156	}
1157}
1158struct fd fdget(unsigned int fd)
1159{
1160	return __fget_light(fd, FMODE_PATH);
1161}
1162EXPORT_SYMBOL(fdget);
1163
1164struct fd fdget_raw(unsigned int fd)
1165{
1166	return __fget_light(fd, 0);
1167}
1168
1169/*
1170 * Try to avoid f_pos locking. We only need it if the
1171 * file is marked for FMODE_ATOMIC_POS, and it can be
1172 * accessed multiple ways.
1173 *
1174 * Always do it for directories, because pidfd_getfd()
1175 * can make a file accessible even if it otherwise would
1176 * not be, and for directories this is a correctness
1177 * issue, not a "POSIX requirement".
1178 */
1179static inline bool file_needs_f_pos_lock(struct file *file)
1180{
1181	return (file->f_mode & FMODE_ATOMIC_POS) &&
1182		(file_count(file) > 1 || file->f_op->iterate_shared);
1183}
1184
1185struct fd fdget_pos(unsigned int fd)
1186{
1187	struct fd f = fdget(fd);
1188	struct file *file = fd_file(f);
1189
1190	if (file && file_needs_f_pos_lock(file)) {
1191		f.word |= FDPUT_POS_UNLOCK;
1192		mutex_lock(&file->f_pos_lock);
1193	}
1194	return f;
1195}
1196
1197void __f_unlock_pos(struct file *f)
1198{
1199	mutex_unlock(&f->f_pos_lock);
1200}
1201
1202/*
1203 * We only lock f_pos if we have threads or if the file might be
1204 * shared with another process. In both cases we'll have an elevated
1205 * file count (done either by fdget() or by fork()).
1206 */
1207
1208void set_close_on_exec(unsigned int fd, int flag)
1209{
1210	struct files_struct *files = current->files;
 
1211	spin_lock(&files->file_lock);
1212	__set_close_on_exec(fd, files_fdtable(files), flag);
 
 
 
 
1213	spin_unlock(&files->file_lock);
1214}
1215
1216bool get_close_on_exec(unsigned int fd)
1217{
 
 
1218	bool res;
1219	rcu_read_lock();
1220	res = close_on_exec(fd, current->files);
 
1221	rcu_read_unlock();
1222	return res;
1223}
1224
1225static int do_dup2(struct files_struct *files,
1226	struct file *file, unsigned fd, unsigned flags)
1227__releases(&files->file_lock)
1228{
1229	struct file *tofree;
1230	struct fdtable *fdt;
1231
1232	/*
1233	 * We need to detect attempts to do dup2() over allocated but still
1234	 * not finished descriptor.  NB: OpenBSD avoids that at the price of
1235	 * extra work in their equivalent of fget() - they insert struct
1236	 * file immediately after grabbing descriptor, mark it larval if
1237	 * more work (e.g. actual opening) is needed and make sure that
1238	 * fget() treats larval files as absent.  Potentially interesting,
1239	 * but while extra work in fget() is trivial, locking implications
1240	 * and amount of surgery on open()-related paths in VFS are not.
1241	 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1242	 * deadlocks in rather amusing ways, AFAICS.  All of that is out of
1243	 * scope of POSIX or SUS, since neither considers shared descriptor
1244	 * tables and this condition does not arise without those.
1245	 */
1246	fdt = files_fdtable(files);
1247	fd = array_index_nospec(fd, fdt->max_fds);
1248	tofree = fdt->fd[fd];
1249	if (!tofree && fd_is_open(fd, fdt))
1250		goto Ebusy;
1251	get_file(file);
1252	rcu_assign_pointer(fdt->fd[fd], file);
1253	__set_open_fd(fd, fdt, flags & O_CLOEXEC);
 
 
 
 
1254	spin_unlock(&files->file_lock);
1255
1256	if (tofree)
1257		filp_close(tofree, files);
1258
1259	return fd;
1260
1261Ebusy:
1262	spin_unlock(&files->file_lock);
1263	return -EBUSY;
1264}
1265
1266int replace_fd(unsigned fd, struct file *file, unsigned flags)
1267{
1268	int err;
1269	struct files_struct *files = current->files;
1270
1271	if (!file)
1272		return close_fd(fd);
1273
1274	if (fd >= rlimit(RLIMIT_NOFILE))
1275		return -EBADF;
1276
1277	spin_lock(&files->file_lock);
1278	err = expand_files(files, fd);
1279	if (unlikely(err < 0))
1280		goto out_unlock;
1281	return do_dup2(files, file, fd, flags);
1282
1283out_unlock:
1284	spin_unlock(&files->file_lock);
1285	return err;
1286}
1287
1288/**
1289 * receive_fd() - Install received file into file descriptor table
1290 * @file: struct file that was received from another process
1291 * @ufd: __user pointer to write new fd number to
1292 * @o_flags: the O_* flags to apply to the new fd entry
1293 *
1294 * Installs a received file into the file descriptor table, with appropriate
1295 * checks and count updates. Optionally writes the fd number to userspace, if
1296 * @ufd is non-NULL.
1297 *
1298 * This helper handles its own reference counting of the incoming
1299 * struct file.
1300 *
1301 * Returns newly install fd or -ve on error.
1302 */
1303int receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1304{
1305	int new_fd;
1306	int error;
1307
1308	error = security_file_receive(file);
1309	if (error)
1310		return error;
1311
1312	new_fd = get_unused_fd_flags(o_flags);
1313	if (new_fd < 0)
1314		return new_fd;
1315
1316	if (ufd) {
1317		error = put_user(new_fd, ufd);
1318		if (error) {
1319			put_unused_fd(new_fd);
1320			return error;
1321		}
1322	}
1323
1324	fd_install(new_fd, get_file(file));
1325	__receive_sock(file);
1326	return new_fd;
1327}
1328EXPORT_SYMBOL_GPL(receive_fd);
1329
1330int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1331{
1332	int error;
1333
1334	error = security_file_receive(file);
1335	if (error)
1336		return error;
1337	error = replace_fd(new_fd, file, o_flags);
1338	if (error)
1339		return error;
1340	__receive_sock(file);
1341	return new_fd;
1342}
1343
1344static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1345{
1346	int err = -EBADF;
1347	struct file *file;
1348	struct files_struct *files = current->files;
1349
1350	if ((flags & ~O_CLOEXEC) != 0)
1351		return -EINVAL;
1352
1353	if (unlikely(oldfd == newfd))
1354		return -EINVAL;
1355
1356	if (newfd >= rlimit(RLIMIT_NOFILE))
1357		return -EBADF;
1358
1359	spin_lock(&files->file_lock);
1360	err = expand_files(files, newfd);
1361	file = files_lookup_fd_locked(files, oldfd);
1362	if (unlikely(!file))
1363		goto Ebadf;
1364	if (unlikely(err < 0)) {
1365		if (err == -EMFILE)
1366			goto Ebadf;
1367		goto out_unlock;
1368	}
1369	return do_dup2(files, file, newfd, flags);
1370
1371Ebadf:
1372	err = -EBADF;
1373out_unlock:
1374	spin_unlock(&files->file_lock);
1375	return err;
1376}
1377
1378SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1379{
1380	return ksys_dup3(oldfd, newfd, flags);
1381}
1382
1383SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1384{
1385	if (unlikely(newfd == oldfd)) { /* corner case */
1386		struct files_struct *files = current->files;
1387		struct file *f;
1388		int retval = oldfd;
1389
1390		rcu_read_lock();
1391		f = __fget_files_rcu(files, oldfd, 0);
1392		if (!f)
1393			retval = -EBADF;
1394		rcu_read_unlock();
1395		if (f)
1396			fput(f);
1397		return retval;
1398	}
1399	return ksys_dup3(oldfd, newfd, 0);
1400}
1401
1402SYSCALL_DEFINE1(dup, unsigned int, fildes)
1403{
1404	int ret = -EBADF;
1405	struct file *file = fget_raw(fildes);
1406
1407	if (file) {
1408		ret = get_unused_fd_flags(0);
1409		if (ret >= 0)
1410			fd_install(ret, file);
1411		else
1412			fput(file);
1413	}
1414	return ret;
1415}
1416
1417int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1418{
1419	unsigned long nofile = rlimit(RLIMIT_NOFILE);
1420	int err;
1421	if (from >= nofile)
1422		return -EINVAL;
1423	err = alloc_fd(from, nofile, flags);
1424	if (err >= 0) {
1425		get_file(file);
1426		fd_install(err, file);
1427	}
1428	return err;
1429}
1430
1431int iterate_fd(struct files_struct *files, unsigned n,
1432		int (*f)(const void *, struct file *, unsigned),
1433		const void *p)
1434{
1435	struct fdtable *fdt;
1436	int res = 0;
1437	if (!files)
1438		return 0;
1439	spin_lock(&files->file_lock);
1440	for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1441		struct file *file;
1442		file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1443		if (!file)
1444			continue;
1445		res = f(p, file, n);
1446		if (res)
1447			break;
1448	}
1449	spin_unlock(&files->file_lock);
1450	return res;
1451}
1452EXPORT_SYMBOL(iterate_fd);