xfs_inode.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include <linux/iversion.h>
   7
   8#include "xfs.h"
   9#include "xfs_fs.h"
  10#include "xfs_shared.h"
  11#include "xfs_format.h"
  12#include "xfs_log_format.h"
  13#include "xfs_trans_resv.h"
  14#include "xfs_mount.h"
  15#include "xfs_defer.h"
  16#include "xfs_inode.h"
  17#include "xfs_dir2.h"
  18#include "xfs_attr.h"
  19#include "xfs_bit.h"
  20#include "xfs_trans_space.h"
  21#include "xfs_trans.h"
  22#include "xfs_buf_item.h"
  23#include "xfs_inode_item.h"
  24#include "xfs_iunlink_item.h"
  25#include "xfs_ialloc.h"
  26#include "xfs_bmap.h"
  27#include "xfs_bmap_util.h"
  28#include "xfs_errortag.h"
  29#include "xfs_error.h"
  30#include "xfs_quota.h"
  31#include "xfs_filestream.h"
  32#include "xfs_trace.h"
  33#include "xfs_icache.h"
  34#include "xfs_symlink.h"
  35#include "xfs_trans_priv.h"
  36#include "xfs_log.h"
  37#include "xfs_bmap_btree.h"
  38#include "xfs_reflink.h"
  39#include "xfs_ag.h"
  40#include "xfs_log_priv.h"
  41#include "xfs_health.h"
  42#include "xfs_pnfs.h"
  43#include "xfs_parent.h"
  44#include "xfs_xattr.h"
  45#include "xfs_inode_util.h"
  46#include "xfs_metafile.h"
  47
  48struct kmem_cache *xfs_inode_cache;
  49
  50/*
  51 * These two are wrapper routines around the xfs_ilock() routine used to
  52 * centralize some grungy code.  They are used in places that wish to lock the
  53 * inode solely for reading the extents.  The reason these places can't just
  54 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
  55 * bringing in of the extents from disk for a file in b-tree format.  If the
  56 * inode is in b-tree format, then we need to lock the inode exclusively until
  57 * the extents are read in.  Locking it exclusively all the time would limit
  58 * our parallelism unnecessarily, though.  What we do instead is check to see
  59 * if the extents have been read in yet, and only lock the inode exclusively
  60 * if they have not.
  61 *
  62 * The functions return a value which should be given to the corresponding
  63 * xfs_iunlock() call.
  64 */
  65uint
  66xfs_ilock_data_map_shared(
  67	struct xfs_inode	*ip)
  68{
  69	uint			lock_mode = XFS_ILOCK_SHARED;
  70
  71	if (xfs_need_iread_extents(&ip->i_df))
  72		lock_mode = XFS_ILOCK_EXCL;
  73	xfs_ilock(ip, lock_mode);
  74	return lock_mode;
  75}
  76
  77uint
  78xfs_ilock_attr_map_shared(
  79	struct xfs_inode	*ip)
  80{
  81	uint			lock_mode = XFS_ILOCK_SHARED;
  82
  83	if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
  84		lock_mode = XFS_ILOCK_EXCL;
  85	xfs_ilock(ip, lock_mode);
  86	return lock_mode;
  87}
  88
  89/*
  90 * You can't set both SHARED and EXCL for the same lock,
  91 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
  92 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
  93 * to set in lock_flags.
  94 */
  95static inline void
  96xfs_lock_flags_assert(
  97	uint		lock_flags)
  98{
  99	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 100		(XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 101	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 102		(XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 103	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 104		(XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 105	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 106	ASSERT(lock_flags != 0);
 107}
 108
 109/*
 110 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
 111 * multi-reader locks: invalidate_lock and the i_lock.  This routine allows
 112 * various combinations of the locks to be obtained.
 113 *
 114 * The 3 locks should always be ordered so that the IO lock is obtained first,
 115 * the mmap lock second and the ilock last in order to prevent deadlock.
 116 *
 117 * Basic locking order:
 118 *
 119 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
 120 *
 121 * mmap_lock locking order:
 122 *
 123 * i_rwsem -> page lock -> mmap_lock
 124 * mmap_lock -> invalidate_lock -> page_lock
 125 *
 126 * The difference in mmap_lock locking order mean that we cannot hold the
 127 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
 128 * can fault in pages during copy in/out (for buffered IO) or require the
 129 * mmap_lock in get_user_pages() to map the user pages into the kernel address
 130 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
 131 * fault because page faults already hold the mmap_lock.
 132 *
 133 * Hence to serialise fully against both syscall and mmap based IO, we need to
 134 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
 135 * both taken in places where we need to invalidate the page cache in a race
 136 * free manner (e.g. truncate, hole punch and other extent manipulation
 137 * functions).
 138 */
 139void
 140xfs_ilock(
 141	xfs_inode_t		*ip,
 142	uint			lock_flags)
 143{
 144	trace_xfs_ilock(ip, lock_flags, _RET_IP_);
 145
 146	xfs_lock_flags_assert(lock_flags);
 147
 148	if (lock_flags & XFS_IOLOCK_EXCL) {
 149		down_write_nested(&VFS_I(ip)->i_rwsem,
 150				  XFS_IOLOCK_DEP(lock_flags));
 151	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 152		down_read_nested(&VFS_I(ip)->i_rwsem,
 153				 XFS_IOLOCK_DEP(lock_flags));
 154	}
 155
 156	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 157		down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
 158				  XFS_MMAPLOCK_DEP(lock_flags));
 159	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 160		down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
 161				 XFS_MMAPLOCK_DEP(lock_flags));
 162	}
 163
 164	if (lock_flags & XFS_ILOCK_EXCL)
 165		down_write_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 166	else if (lock_flags & XFS_ILOCK_SHARED)
 167		down_read_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 168}
 169
 170/*
 171 * This is just like xfs_ilock(), except that the caller
 172 * is guaranteed not to sleep.  It returns 1 if it gets
 173 * the requested locks and 0 otherwise.  If the IO lock is
 174 * obtained but the inode lock cannot be, then the IO lock
 175 * is dropped before returning.
 176 *
 177 * ip -- the inode being locked
 178 * lock_flags -- this parameter indicates the inode's locks to be
 179 *       to be locked.  See the comment for xfs_ilock() for a list
 180 *	 of valid values.
 181 */
 182int
 183xfs_ilock_nowait(
 184	xfs_inode_t		*ip,
 185	uint			lock_flags)
 186{
 187	trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
 188
 189	xfs_lock_flags_assert(lock_flags);
 190
 191	if (lock_flags & XFS_IOLOCK_EXCL) {
 192		if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
 193			goto out;
 194	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 195		if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
 196			goto out;
 197	}
 198
 199	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 200		if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
 201			goto out_undo_iolock;
 202	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 203		if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
 204			goto out_undo_iolock;
 205	}
 206
 207	if (lock_flags & XFS_ILOCK_EXCL) {
 208		if (!down_write_trylock(&ip->i_lock))
 209			goto out_undo_mmaplock;
 210	} else if (lock_flags & XFS_ILOCK_SHARED) {
 211		if (!down_read_trylock(&ip->i_lock))
 212			goto out_undo_mmaplock;
 213	}
 214	return 1;
 215
 216out_undo_mmaplock:
 217	if (lock_flags & XFS_MMAPLOCK_EXCL)
 218		up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 219	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 220		up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
 221out_undo_iolock:
 222	if (lock_flags & XFS_IOLOCK_EXCL)
 223		up_write(&VFS_I(ip)->i_rwsem);
 224	else if (lock_flags & XFS_IOLOCK_SHARED)
 225		up_read(&VFS_I(ip)->i_rwsem);
 226out:
 227	return 0;
 228}
 229
 230/*
 231 * xfs_iunlock() is used to drop the inode locks acquired with
 232 * xfs_ilock() and xfs_ilock_nowait().  The caller must pass
 233 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
 234 * that we know which locks to drop.
 235 *
 236 * ip -- the inode being unlocked
 237 * lock_flags -- this parameter indicates the inode's locks to be
 238 *       to be unlocked.  See the comment for xfs_ilock() for a list
 239 *	 of valid values for this parameter.
 240 *
 241 */
 242void
 243xfs_iunlock(
 244	xfs_inode_t		*ip,
 245	uint			lock_flags)
 246{
 247	xfs_lock_flags_assert(lock_flags);
 248
 249	if (lock_flags & XFS_IOLOCK_EXCL)
 250		up_write(&VFS_I(ip)->i_rwsem);
 251	else if (lock_flags & XFS_IOLOCK_SHARED)
 252		up_read(&VFS_I(ip)->i_rwsem);
 253
 254	if (lock_flags & XFS_MMAPLOCK_EXCL)
 255		up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 256	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 257		up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
 258
 259	if (lock_flags & XFS_ILOCK_EXCL)
 260		up_write(&ip->i_lock);
 261	else if (lock_flags & XFS_ILOCK_SHARED)
 262		up_read(&ip->i_lock);
 263
 264	trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
 265}
 266
 267/*
 268 * give up write locks.  the i/o lock cannot be held nested
 269 * if it is being demoted.
 270 */
 271void
 272xfs_ilock_demote(
 273	xfs_inode_t		*ip,
 274	uint			lock_flags)
 275{
 276	ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
 277	ASSERT((lock_flags &
 278		~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
 279
 280	if (lock_flags & XFS_ILOCK_EXCL)
 281		downgrade_write(&ip->i_lock);
 282	if (lock_flags & XFS_MMAPLOCK_EXCL)
 283		downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 284	if (lock_flags & XFS_IOLOCK_EXCL)
 285		downgrade_write(&VFS_I(ip)->i_rwsem);
 286
 287	trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
 288}
 289
 290void
 291xfs_assert_ilocked(
 292	struct xfs_inode	*ip,
 293	uint			lock_flags)
 294{
 295	/*
 296	 * Sometimes we assert the ILOCK is held exclusively, but we're in
 297	 * a workqueue, so lockdep doesn't know we're the owner.
 298	 */
 299	if (lock_flags & XFS_ILOCK_SHARED)
 300		rwsem_assert_held(&ip->i_lock);
 301	else if (lock_flags & XFS_ILOCK_EXCL)
 302		rwsem_assert_held_write_nolockdep(&ip->i_lock);
 303
 304	if (lock_flags & XFS_MMAPLOCK_SHARED)
 305		rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
 306	else if (lock_flags & XFS_MMAPLOCK_EXCL)
 307		rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 308
 309	if (lock_flags & XFS_IOLOCK_SHARED)
 310		rwsem_assert_held(&VFS_I(ip)->i_rwsem);
 311	else if (lock_flags & XFS_IOLOCK_EXCL)
 312		rwsem_assert_held_write(&VFS_I(ip)->i_rwsem);
 313}
 314
 315/*
 316 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
 317 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
 318 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
 319 * errors and warnings.
 320 */
 321#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
 322static bool
 323xfs_lockdep_subclass_ok(
 324	int subclass)
 325{
 326	return subclass < MAX_LOCKDEP_SUBCLASSES;
 327}
 328#else
 329#define xfs_lockdep_subclass_ok(subclass)	(true)
 330#endif
 331
 332/*
 333 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
 334 * value. This can be called for any type of inode lock combination, including
 335 * parent locking. Care must be taken to ensure we don't overrun the subclass
 336 * storage fields in the class mask we build.
 337 */
 338static inline uint
 339xfs_lock_inumorder(
 340	uint	lock_mode,
 341	uint	subclass)
 342{
 343	uint	class = 0;
 344
 345	ASSERT(!(lock_mode & XFS_ILOCK_PARENT));
 346	ASSERT(xfs_lockdep_subclass_ok(subclass));
 347
 348	if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
 349		ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
 350		class += subclass << XFS_IOLOCK_SHIFT;
 351	}
 352
 353	if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
 354		ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
 355		class += subclass << XFS_MMAPLOCK_SHIFT;
 356	}
 357
 358	if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
 359		ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
 360		class += subclass << XFS_ILOCK_SHIFT;
 361	}
 362
 363	return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
 364}
 365
 366/*
 367 * The following routine will lock n inodes in exclusive mode.  We assume the
 368 * caller calls us with the inodes in i_ino order.
 369 *
 370 * We need to detect deadlock where an inode that we lock is in the AIL and we
 371 * start waiting for another inode that is locked by a thread in a long running
 372 * transaction (such as truncate). This can result in deadlock since the long
 373 * running trans might need to wait for the inode we just locked in order to
 374 * push the tail and free space in the log.
 375 *
 376 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
 377 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
 378 * lock more than one at a time, lockdep will report false positives saying we
 379 * have violated locking orders.
 380 */
 381void
 382xfs_lock_inodes(
 383	struct xfs_inode	**ips,
 384	int			inodes,
 385	uint			lock_mode)
 386{
 387	int			attempts = 0;
 388	uint			i;
 389	int			j;
 390	bool			try_lock;
 391	struct xfs_log_item	*lp;
 392
 393	/*
 394	 * Currently supports between 2 and 5 inodes with exclusive locking.  We
 395	 * support an arbitrary depth of locking here, but absolute limits on
 396	 * inodes depend on the type of locking and the limits placed by
 397	 * lockdep annotations in xfs_lock_inumorder.  These are all checked by
 398	 * the asserts.
 399	 */
 400	ASSERT(ips && inodes >= 2 && inodes <= 5);
 401	ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
 402			    XFS_ILOCK_EXCL));
 403	ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
 404			      XFS_ILOCK_SHARED)));
 405	ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
 406		inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
 407	ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
 408		inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
 409
 410	if (lock_mode & XFS_IOLOCK_EXCL) {
 411		ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
 412	} else if (lock_mode & XFS_MMAPLOCK_EXCL)
 413		ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
 414
 415again:
 416	try_lock = false;
 417	i = 0;
 418	for (; i < inodes; i++) {
 419		ASSERT(ips[i]);
 420
 421		if (i && (ips[i] == ips[i - 1]))	/* Already locked */
 422			continue;
 423
 424		/*
 425		 * If try_lock is not set yet, make sure all locked inodes are
 426		 * not in the AIL.  If any are, set try_lock to be used later.
 427		 */
 428		if (!try_lock) {
 429			for (j = (i - 1); j >= 0 && !try_lock; j--) {
 430				lp = &ips[j]->i_itemp->ili_item;
 431				if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
 432					try_lock = true;
 433			}
 434		}
 435
 436		/*
 437		 * If any of the previous locks we have locked is in the AIL,
 438		 * we must TRY to get the second and subsequent locks. If
 439		 * we can't get any, we must release all we have
 440		 * and try again.
 441		 */
 442		if (!try_lock) {
 443			xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
 444			continue;
 445		}
 446
 447		/* try_lock means we have an inode locked that is in the AIL. */
 448		ASSERT(i != 0);
 449		if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
 450			continue;
 451
 452		/*
 453		 * Unlock all previous guys and try again.  xfs_iunlock will try
 454		 * to push the tail if the inode is in the AIL.
 455		 */
 456		attempts++;
 457		for (j = i - 1; j >= 0; j--) {
 458			/*
 459			 * Check to see if we've already unlocked this one.  Not
 460			 * the first one going back, and the inode ptr is the
 461			 * same.
 462			 */
 463			if (j != (i - 1) && ips[j] == ips[j + 1])
 464				continue;
 465
 466			xfs_iunlock(ips[j], lock_mode);
 467		}
 468
 469		if ((attempts % 5) == 0) {
 470			delay(1); /* Don't just spin the CPU */
 471		}
 472		goto again;
 473	}
 474}
 475
 476/*
 477 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
 478 * mmaplock must be double-locked separately since we use i_rwsem and
 479 * invalidate_lock for that. We now support taking one lock EXCL and the
 480 * other SHARED.
 481 */
 482void
 483xfs_lock_two_inodes(
 484	struct xfs_inode	*ip0,
 485	uint			ip0_mode,
 486	struct xfs_inode	*ip1,
 487	uint			ip1_mode)
 488{
 489	int			attempts = 0;
 490	struct xfs_log_item	*lp;
 491
 492	ASSERT(hweight32(ip0_mode) == 1);
 493	ASSERT(hweight32(ip1_mode) == 1);
 494	ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 495	ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 496	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
 497	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
 498	ASSERT(ip0->i_ino != ip1->i_ino);
 499
 500	if (ip0->i_ino > ip1->i_ino) {
 501		swap(ip0, ip1);
 502		swap(ip0_mode, ip1_mode);
 503	}
 504
 505 again:
 506	xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
 507
 508	/*
 509	 * If the first lock we have locked is in the AIL, we must TRY to get
 510	 * the second lock. If we can't get it, we must release the first one
 511	 * and try again.
 512	 */
 513	lp = &ip0->i_itemp->ili_item;
 514	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
 515		if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
 516			xfs_iunlock(ip0, ip0_mode);
 517			if ((++attempts % 5) == 0)
 518				delay(1); /* Don't just spin the CPU */
 519			goto again;
 520		}
 521	} else {
 522		xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
 523	}
 524}
 525
 526/*
 527 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 528 * is allowed, otherwise it has to be an exact match. If a CI match is found,
 529 * ci_name->name will point to a the actual name (caller must free) or
 530 * will be set to NULL if an exact match is found.
 531 */
 532int
 533xfs_lookup(
 534	struct xfs_inode	*dp,
 535	const struct xfs_name	*name,
 536	struct xfs_inode	**ipp,
 537	struct xfs_name		*ci_name)
 538{
 539	xfs_ino_t		inum;
 540	int			error;
 541
 542	trace_xfs_lookup(dp, name);
 543
 544	if (xfs_is_shutdown(dp->i_mount))
 545		return -EIO;
 546	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
 547		return -EIO;
 548
 549	error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 550	if (error)
 551		goto out_unlock;
 552
 553	error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 554	if (error)
 555		goto out_free_name;
 556
 557	/*
 558	 * Fail if a directory entry in the regular directory tree points to
 559	 * a metadata file.
 560	 */
 561	if (XFS_IS_CORRUPT(dp->i_mount, xfs_is_metadir_inode(*ipp))) {
 562		xfs_fs_mark_sick(dp->i_mount, XFS_SICK_FS_METADIR);
 563		error = -EFSCORRUPTED;
 564		goto out_irele;
 565	}
 566
 567	return 0;
 568
 569out_irele:
 570	xfs_irele(*ipp);
 571out_free_name:
 572	if (ci_name)
 573		kfree(ci_name->name);
 574out_unlock:
 575	*ipp = NULL;
 576	return error;
 577}
 578
 579/*
 580 * Initialise a newly allocated inode and return the in-core inode to the
 581 * caller locked exclusively.
 582 *
 583 * Caller is responsible for unlocking the inode manually upon return
 584 */
 585int
 586xfs_icreate(
 587	struct xfs_trans	*tp,
 588	xfs_ino_t		ino,
 589	const struct xfs_icreate_args *args,
 590	struct xfs_inode	**ipp)
 591{
 592	struct xfs_mount	*mp = tp->t_mountp;
 593	struct xfs_inode	*ip = NULL;
 594	int			error;
 595
 596	/*
 597	 * Get the in-core inode with the lock held exclusively to prevent
 598	 * others from looking at until we're done.
 599	 */
 600	error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
 601	if (error)
 602		return error;
 603
 604	ASSERT(ip != NULL);
 605	xfs_trans_ijoin(tp, ip, 0);
 606	xfs_inode_init(tp, args, ip);
 607
 608	/* now that we have an i_mode we can setup the inode structure */
 609	xfs_setup_inode(ip);
 610
 611	*ipp = ip;
 612	return 0;
 613}
 614
 615/* Return dquots for the ids that will be assigned to a new file. */
 616int
 617xfs_icreate_dqalloc(
 618	const struct xfs_icreate_args	*args,
 619	struct xfs_dquot		**udqpp,
 620	struct xfs_dquot		**gdqpp,
 621	struct xfs_dquot		**pdqpp)
 622{
 623	struct inode			*dir = VFS_I(args->pip);
 624	kuid_t				uid = GLOBAL_ROOT_UID;
 625	kgid_t				gid = GLOBAL_ROOT_GID;
 626	prid_t				prid = 0;
 627	unsigned int			flags = XFS_QMOPT_QUOTALL;
 628
 629	if (args->idmap) {
 630		/*
 631		 * The uid/gid computation code must match what the VFS uses to
 632		 * assign i_[ug]id.  INHERIT adjusts the gid computation for
 633		 * setgid/grpid systems.
 634		 */
 635		uid = mapped_fsuid(args->idmap, i_user_ns(dir));
 636		gid = mapped_fsgid(args->idmap, i_user_ns(dir));
 637		prid = xfs_get_initial_prid(args->pip);
 638		flags |= XFS_QMOPT_INHERIT;
 639	}
 640
 641	*udqpp = *gdqpp = *pdqpp = NULL;
 642
 643	return xfs_qm_vop_dqalloc(args->pip, uid, gid, prid, flags, udqpp,
 644			gdqpp, pdqpp);
 645}
 646
 647int
 648xfs_create(
 649	const struct xfs_icreate_args *args,
 650	struct xfs_name		*name,
 651	struct xfs_inode	**ipp)
 652{
 653	struct xfs_inode	*dp = args->pip;
 654	struct xfs_dir_update	du = {
 655		.dp		= dp,
 656		.name		= name,
 657	};
 658	struct xfs_mount	*mp = dp->i_mount;
 659	struct xfs_trans	*tp = NULL;
 660	struct xfs_dquot	*udqp;
 661	struct xfs_dquot	*gdqp;
 662	struct xfs_dquot	*pdqp;
 663	struct xfs_trans_res	*tres;
 664	xfs_ino_t		ino;
 665	bool			unlock_dp_on_error = false;
 666	bool			is_dir = S_ISDIR(args->mode);
 667	uint			resblks;
 668	int			error;
 669
 670	trace_xfs_create(dp, name);
 671
 672	if (xfs_is_shutdown(mp))
 673		return -EIO;
 674	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
 675		return -EIO;
 676
 677	/* Make sure that we have allocated dquot(s) on disk. */
 678	error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
 679	if (error)
 680		return error;
 681
 682	if (is_dir) {
 683		resblks = xfs_mkdir_space_res(mp, name->len);
 684		tres = &M_RES(mp)->tr_mkdir;
 685	} else {
 686		resblks = xfs_create_space_res(mp, name->len);
 687		tres = &M_RES(mp)->tr_create;
 688	}
 689
 690	error = xfs_parent_start(mp, &du.ppargs);
 691	if (error)
 692		goto out_release_dquots;
 693
 694	/*
 695	 * Initially assume that the file does not exist and
 696	 * reserve the resources for that case.  If that is not
 697	 * the case we'll drop the one we have and get a more
 698	 * appropriate transaction later.
 699	 */
 700	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
 701			&tp);
 702	if (error == -ENOSPC) {
 703		/* flush outstanding delalloc blocks and retry */
 704		xfs_flush_inodes(mp);
 705		error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
 706				resblks, &tp);
 707	}
 708	if (error)
 709		goto out_parent;
 710
 711	xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 712	unlock_dp_on_error = true;
 713
 714	/*
 715	 * A newly created regular or special file just has one directory
 716	 * entry pointing to them, but a directory also the "." entry
 717	 * pointing to itself.
 718	 */
 719	error = xfs_dialloc(&tp, args, &ino);
 720	if (!error)
 721		error = xfs_icreate(tp, ino, args, &du.ip);
 722	if (error)
 723		goto out_trans_cancel;
 724
 725	/*
 726	 * Now we join the directory inode to the transaction.  We do not do it
 727	 * earlier because xfs_dialloc might commit the previous transaction
 728	 * (and release all the locks).  An error from here on will result in
 729	 * the transaction cancel unlocking dp so don't do it explicitly in the
 730	 * error path.
 731	 */
 732	xfs_trans_ijoin(tp, dp, 0);
 733
 734	error = xfs_dir_create_child(tp, resblks, &du);
 735	if (error)
 736		goto out_trans_cancel;
 737
 738	/*
 739	 * If this is a synchronous mount, make sure that the
 740	 * create transaction goes to disk before returning to
 741	 * the user.
 742	 */
 743	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
 744		xfs_trans_set_sync(tp);
 745
 746	/*
 747	 * Attach the dquot(s) to the inodes and modify them incore.
 748	 * These ids of the inode couldn't have changed since the new
 749	 * inode has been locked ever since it was created.
 750	 */
 751	xfs_qm_vop_create_dqattach(tp, du.ip, udqp, gdqp, pdqp);
 752
 753	error = xfs_trans_commit(tp);
 754	if (error)
 755		goto out_release_inode;
 756
 757	xfs_qm_dqrele(udqp);
 758	xfs_qm_dqrele(gdqp);
 759	xfs_qm_dqrele(pdqp);
 760
 761	*ipp = du.ip;
 762	xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
 763	xfs_iunlock(dp, XFS_ILOCK_EXCL);
 764	xfs_parent_finish(mp, du.ppargs);
 765	return 0;
 766
 767 out_trans_cancel:
 768	xfs_trans_cancel(tp);
 769 out_release_inode:
 770	/*
 771	 * Wait until after the current transaction is aborted to finish the
 772	 * setup of the inode and release the inode.  This prevents recursive
 773	 * transactions and deadlocks from xfs_inactive.
 774	 */
 775	if (du.ip) {
 776		xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
 777		xfs_finish_inode_setup(du.ip);
 778		xfs_irele(du.ip);
 779	}
 780 out_parent:
 781	xfs_parent_finish(mp, du.ppargs);
 782 out_release_dquots:
 783	xfs_qm_dqrele(udqp);
 784	xfs_qm_dqrele(gdqp);
 785	xfs_qm_dqrele(pdqp);
 786
 787	if (unlock_dp_on_error)
 788		xfs_iunlock(dp, XFS_ILOCK_EXCL);
 789	return error;
 790}
 791
 792int
 793xfs_create_tmpfile(
 794	const struct xfs_icreate_args *args,
 795	struct xfs_inode	**ipp)
 796{
 797	struct xfs_inode	*dp = args->pip;
 798	struct xfs_mount	*mp = dp->i_mount;
 799	struct xfs_inode	*ip = NULL;
 800	struct xfs_trans	*tp = NULL;
 801	struct xfs_dquot	*udqp;
 802	struct xfs_dquot	*gdqp;
 803	struct xfs_dquot	*pdqp;
 804	struct xfs_trans_res	*tres;
 805	xfs_ino_t		ino;
 806	uint			resblks;
 807	int			error;
 808
 809	ASSERT(args->flags & XFS_ICREATE_TMPFILE);
 810
 811	if (xfs_is_shutdown(mp))
 812		return -EIO;
 813
 814	/* Make sure that we have allocated dquot(s) on disk. */
 815	error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
 816	if (error)
 817		return error;
 818
 819	resblks = XFS_IALLOC_SPACE_RES(mp);
 820	tres = &M_RES(mp)->tr_create_tmpfile;
 821
 822	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
 823			&tp);
 824	if (error)
 825		goto out_release_dquots;
 826
 827	error = xfs_dialloc(&tp, args, &ino);
 828	if (!error)
 829		error = xfs_icreate(tp, ino, args, &ip);
 830	if (error)
 831		goto out_trans_cancel;
 832
 833	if (xfs_has_wsync(mp))
 834		xfs_trans_set_sync(tp);
 835
 836	/*
 837	 * Attach the dquot(s) to the inodes and modify them incore.
 838	 * These ids of the inode couldn't have changed since the new
 839	 * inode has been locked ever since it was created.
 840	 */
 841	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
 842
 843	error = xfs_iunlink(tp, ip);
 844	if (error)
 845		goto out_trans_cancel;
 846
 847	error = xfs_trans_commit(tp);
 848	if (error)
 849		goto out_release_inode;
 850
 851	xfs_qm_dqrele(udqp);
 852	xfs_qm_dqrele(gdqp);
 853	xfs_qm_dqrele(pdqp);
 854
 855	*ipp = ip;
 856	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 857	return 0;
 858
 859 out_trans_cancel:
 860	xfs_trans_cancel(tp);
 861 out_release_inode:
 862	/*
 863	 * Wait until after the current transaction is aborted to finish the
 864	 * setup of the inode and release the inode.  This prevents recursive
 865	 * transactions and deadlocks from xfs_inactive.
 866	 */
 867	if (ip) {
 868		xfs_iunlock(ip, XFS_ILOCK_EXCL);
 869		xfs_finish_inode_setup(ip);
 870		xfs_irele(ip);
 871	}
 872 out_release_dquots:
 873	xfs_qm_dqrele(udqp);
 874	xfs_qm_dqrele(gdqp);
 875	xfs_qm_dqrele(pdqp);
 876
 877	return error;
 878}
 879
 880static inline int
 881xfs_projid_differ(
 882	struct xfs_inode	*tdp,
 883	struct xfs_inode	*sip)
 884{
 885	/*
 886	 * If we are using project inheritance, we only allow hard link/renames
 887	 * creation in our tree when the project IDs are the same; else
 888	 * the tree quota mechanism could be circumvented.
 889	 */
 890	if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
 891		     tdp->i_projid != sip->i_projid)) {
 892		/*
 893		 * Project quota setup skips special files which can
 894		 * leave inodes in a PROJINHERIT directory without a
 895		 * project ID set. We need to allow links to be made
 896		 * to these "project-less" inodes because userspace
 897		 * expects them to succeed after project ID setup,
 898		 * but everything else should be rejected.
 899		 */
 900		if (!special_file(VFS_I(sip)->i_mode) ||
 901		    sip->i_projid != 0) {
 902			return -EXDEV;
 903		}
 904	}
 905
 906	return 0;
 907}
 908
 909int
 910xfs_link(
 911	struct xfs_inode	*tdp,
 912	struct xfs_inode	*sip,
 913	struct xfs_name		*target_name)
 914{
 915	struct xfs_dir_update	du = {
 916		.dp		= tdp,
 917		.name		= target_name,
 918		.ip		= sip,
 919	};
 920	struct xfs_mount	*mp = tdp->i_mount;
 921	struct xfs_trans	*tp;
 922	int			error, nospace_error = 0;
 923	int			resblks;
 924
 925	trace_xfs_link(tdp, target_name);
 926
 927	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
 928
 929	if (xfs_is_shutdown(mp))
 930		return -EIO;
 931	if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
 932		return -EIO;
 933
 934	error = xfs_qm_dqattach(sip);
 935	if (error)
 936		goto std_return;
 937
 938	error = xfs_qm_dqattach(tdp);
 939	if (error)
 940		goto std_return;
 941
 942	error = xfs_parent_start(mp, &du.ppargs);
 943	if (error)
 944		goto std_return;
 945
 946	resblks = xfs_link_space_res(mp, target_name->len);
 947	error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
 948			&tp, &nospace_error);
 949	if (error)
 950		goto out_parent;
 951
 952	/*
 953	 * We don't allow reservationless or quotaless hardlinking when parent
 954	 * pointers are enabled because we can't back out if the xattrs must
 955	 * grow.
 956	 */
 957	if (du.ppargs && nospace_error) {
 958		error = nospace_error;
 959		goto error_return;
 960	}
 961
 962	error = xfs_projid_differ(tdp, sip);
 963	if (error)
 964		goto error_return;
 965
 966	error = xfs_dir_add_child(tp, resblks, &du);
 967	if (error)
 968		goto error_return;
 969
 970	/*
 971	 * If this is a synchronous mount, make sure that the
 972	 * link transaction goes to disk before returning to
 973	 * the user.
 974	 */
 975	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
 976		xfs_trans_set_sync(tp);
 977
 978	error = xfs_trans_commit(tp);
 979	xfs_iunlock(tdp, XFS_ILOCK_EXCL);
 980	xfs_iunlock(sip, XFS_ILOCK_EXCL);
 981	xfs_parent_finish(mp, du.ppargs);
 982	return error;
 983
 984 error_return:
 985	xfs_trans_cancel(tp);
 986	xfs_iunlock(tdp, XFS_ILOCK_EXCL);
 987	xfs_iunlock(sip, XFS_ILOCK_EXCL);
 988 out_parent:
 989	xfs_parent_finish(mp, du.ppargs);
 990 std_return:
 991	if (error == -ENOSPC && nospace_error)
 992		error = nospace_error;
 993	return error;
 994}
 995
 996/* Clear the reflink flag and the cowblocks tag if possible. */
 997static void
 998xfs_itruncate_clear_reflink_flags(
 999	struct xfs_inode	*ip)
1000{
1001	struct xfs_ifork	*dfork;
1002	struct xfs_ifork	*cfork;
1003
1004	if (!xfs_is_reflink_inode(ip))
1005		return;
1006	dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1007	cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
1008	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1009		ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1010	if (cfork->if_bytes == 0)
1011		xfs_inode_clear_cowblocks_tag(ip);
1012}
1013
1014/*
1015 * Free up the underlying blocks past new_size.  The new size must be smaller
1016 * than the current size.  This routine can be used both for the attribute and
1017 * data fork, and does not modify the inode size, which is left to the caller.
1018 *
1019 * The transaction passed to this routine must have made a permanent log
1020 * reservation of at least XFS_ITRUNCATE_LOG_RES.  This routine may commit the
1021 * given transaction and start new ones, so make sure everything involved in
1022 * the transaction is tidy before calling here.  Some transaction will be
1023 * returned to the caller to be committed.  The incoming transaction must
1024 * already include the inode, and both inode locks must be held exclusively.
1025 * The inode must also be "held" within the transaction.  On return the inode
1026 * will be "held" within the returned transaction.  This routine does NOT
1027 * require any disk space to be reserved for it within the transaction.
1028 *
1029 * If we get an error, we must return with the inode locked and linked into the
1030 * current transaction. This keeps things simple for the higher level code,
1031 * because it always knows that the inode is locked and held in the transaction
1032 * that returns to it whether errors occur or not.  We don't mark the inode
1033 * dirty on error so that transactions can be easily aborted if possible.
1034 */
1035int
1036xfs_itruncate_extents_flags(
1037	struct xfs_trans	**tpp,
1038	struct xfs_inode	*ip,
1039	int			whichfork,
1040	xfs_fsize_t		new_size,
1041	int			flags)
1042{
1043	struct xfs_mount	*mp = ip->i_mount;
1044	struct xfs_trans	*tp = *tpp;
1045	xfs_fileoff_t		first_unmap_block;
1046	int			error = 0;
1047
1048	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1049	if (icount_read(VFS_I(ip)))
1050		xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1051	if (whichfork == XFS_DATA_FORK)
1052		ASSERT(new_size <= XFS_ISIZE(ip));
1053	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1054	ASSERT(ip->i_itemp != NULL);
1055	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1056	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1057
1058	trace_xfs_itruncate_extents_start(ip, new_size);
1059
1060	flags |= xfs_bmapi_aflag(whichfork);
1061
1062	/*
1063	 * Since it is possible for space to become allocated beyond
1064	 * the end of the file (in a crash where the space is allocated
1065	 * but the inode size is not yet updated), simply remove any
1066	 * blocks which show up between the new EOF and the maximum
1067	 * possible file size.
1068	 *
1069	 * We have to free all the blocks to the bmbt maximum offset, even if
1070	 * the page cache can't scale that far.
1071	 */
1072	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1073	if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1074		WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1075		return 0;
1076	}
1077
1078	error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1079			XFS_MAX_FILEOFF);
1080	if (error)
1081		goto out;
1082
1083	if (whichfork == XFS_DATA_FORK) {
1084		/* Remove all pending CoW reservations. */
1085		error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1086				first_unmap_block, XFS_MAX_FILEOFF, true);
1087		if (error)
1088			goto out;
1089
1090		xfs_itruncate_clear_reflink_flags(ip);
1091	}
1092
1093	/*
1094	 * Always re-log the inode so that our permanent transaction can keep
1095	 * on rolling it forward in the log.
1096	 */
1097	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1098
1099	trace_xfs_itruncate_extents_end(ip, new_size);
1100
1101out:
1102	*tpp = tp;
1103	return error;
1104}
1105
1106/*
1107 * Mark all the buffers attached to this directory stale.  In theory we should
1108 * never be freeing a directory with any blocks at all, but this covers the
1109 * case where we've recovered a directory swap with a "temporary" directory
1110 * created by online repair and now need to dump it.
1111 */
1112STATIC void
1113xfs_inactive_dir(
1114	struct xfs_inode	*dp)
1115{
1116	struct xfs_iext_cursor	icur;
1117	struct xfs_bmbt_irec	got;
1118	struct xfs_mount	*mp = dp->i_mount;
1119	struct xfs_da_geometry	*geo = mp->m_dir_geo;
1120	struct xfs_ifork	*ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1121	xfs_fileoff_t		off;
1122
1123	/*
1124	 * Invalidate each directory block.  All directory blocks are of
1125	 * fsbcount length and alignment, so we only need to walk those same
1126	 * offsets.  We hold the only reference to this inode, so we must wait
1127	 * for the buffer locks.
1128	 */
1129	for_each_xfs_iext(ifp, &icur, &got) {
1130		for (off = round_up(got.br_startoff, geo->fsbcount);
1131		     off < got.br_startoff + got.br_blockcount;
1132		     off += geo->fsbcount) {
1133			struct xfs_buf	*bp = NULL;
1134			xfs_fsblock_t	fsbno;
1135			int		error;
1136
1137			fsbno = (off - got.br_startoff) + got.br_startblock;
1138			error = xfs_buf_incore(mp->m_ddev_targp,
1139					XFS_FSB_TO_DADDR(mp, fsbno),
1140					XFS_FSB_TO_BB(mp, geo->fsbcount),
1141					XBF_LIVESCAN, &bp);
1142			if (error)
1143				continue;
1144
1145			xfs_buf_stale(bp);
1146			xfs_buf_relse(bp);
1147		}
1148	}
1149}
1150
1151/*
1152 * xfs_inactive_truncate
1153 *
1154 * Called to perform a truncate when an inode becomes unlinked.
1155 */
1156STATIC int
1157xfs_inactive_truncate(
1158	struct xfs_inode *ip)
1159{
1160	struct xfs_mount	*mp = ip->i_mount;
1161	struct xfs_trans	*tp;
1162	int			error;
1163
1164	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1165	if (error) {
1166		ASSERT(xfs_is_shutdown(mp));
1167		return error;
1168	}
1169	xfs_ilock(ip, XFS_ILOCK_EXCL);
1170	xfs_trans_ijoin(tp, ip, 0);
1171
1172	/*
1173	 * Log the inode size first to prevent stale data exposure in the event
1174	 * of a system crash before the truncate completes. See the related
1175	 * comment in xfs_vn_setattr_size() for details.
1176	 */
1177	ip->i_disk_size = 0;
1178	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1179
1180	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1181	if (error)
1182		goto error_trans_cancel;
1183
1184	ASSERT(ip->i_df.if_nextents == 0);
1185
1186	error = xfs_trans_commit(tp);
1187	if (error)
1188		goto error_unlock;
1189
1190	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1191	return 0;
1192
1193error_trans_cancel:
1194	xfs_trans_cancel(tp);
1195error_unlock:
1196	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1197	return error;
1198}
1199
1200/*
1201 * xfs_inactive_ifree()
1202 *
1203 * Perform the inode free when an inode is unlinked.
1204 */
1205STATIC int
1206xfs_inactive_ifree(
1207	struct xfs_inode *ip)
1208{
1209	struct xfs_mount	*mp = ip->i_mount;
1210	struct xfs_trans	*tp;
1211	int			error;
1212
1213	/*
1214	 * We try to use a per-AG reservation for any block needed by the finobt
1215	 * tree, but as the finobt feature predates the per-AG reservation
1216	 * support a degraded file system might not have enough space for the
1217	 * reservation at mount time.  In that case try to dip into the reserved
1218	 * pool and pray.
1219	 *
1220	 * Send a warning if the reservation does happen to fail, as the inode
1221	 * now remains allocated and sits on the unlinked list until the fs is
1222	 * repaired.
1223	 */
1224	if (unlikely(mp->m_finobt_nores)) {
1225		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1226				XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1227				&tp);
1228	} else {
1229		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1230	}
1231	if (error) {
1232		if (error == -ENOSPC) {
1233			xfs_warn_ratelimited(mp,
1234			"Failed to remove inode(s) from unlinked list. "
1235			"Please free space, unmount and run xfs_repair.");
1236		} else {
1237			ASSERT(xfs_is_shutdown(mp));
1238		}
1239		return error;
1240	}
1241
1242	/*
1243	 * We do not hold the inode locked across the entire rolling transaction
1244	 * here. We only need to hold it for the first transaction that
1245	 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1246	 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1247	 * here breaks the relationship between cluster buffer invalidation and
1248	 * stale inode invalidation on cluster buffer item journal commit
1249	 * completion, and can result in leaving dirty stale inodes hanging
1250	 * around in memory.
1251	 *
1252	 * We have no need for serialising this inode operation against other
1253	 * operations - we freed the inode and hence reallocation is required
1254	 * and that will serialise on reallocating the space the deferops need
1255	 * to free. Hence we can unlock the inode on the first commit of
1256	 * the transaction rather than roll it right through the deferops. This
1257	 * avoids relogging the XFS_ISTALE inode.
1258	 *
1259	 * We check that xfs_ifree() hasn't grown an internal transaction roll
1260	 * by asserting that the inode is still locked when it returns.
1261	 */
1262	xfs_ilock(ip, XFS_ILOCK_EXCL);
1263	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1264
1265	error = xfs_ifree(tp, ip);
1266	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1267	if (error) {
1268		/*
1269		 * If we fail to free the inode, shut down.  The cancel
1270		 * might do that, we need to make sure.  Otherwise the
1271		 * inode might be lost for a long time or forever.
1272		 */
1273		if (!xfs_is_shutdown(mp)) {
1274			xfs_notice(mp, "%s: xfs_ifree returned error %d",
1275				__func__, error);
1276			xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1277		}
1278		xfs_trans_cancel(tp);
1279		return error;
1280	}
1281
1282	/*
1283	 * Credit the quota account(s). The inode is gone.
1284	 */
1285	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1286
1287	return xfs_trans_commit(tp);
1288}
1289
1290/*
1291 * Returns true if we need to update the on-disk metadata before we can free
1292 * the memory used by this inode.  Updates include freeing post-eof
1293 * preallocations; freeing COW staging extents; and marking the inode free in
1294 * the inobt if it is on the unlinked list.
1295 */
1296bool
1297xfs_inode_needs_inactive(
1298	struct xfs_inode	*ip)
1299{
1300	struct xfs_mount	*mp = ip->i_mount;
1301	struct xfs_ifork	*cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1302
1303	/*
1304	 * If the inode is already free, then there can be nothing
1305	 * to clean up here.
1306	 */
1307	if (VFS_I(ip)->i_mode == 0)
1308		return false;
1309
1310	/*
1311	 * If this is a read-only mount, don't do this (would generate I/O)
1312	 * unless we're in log recovery and cleaning the iunlinked list.
1313	 */
1314	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1315		return false;
1316
1317	/* If the log isn't running, push inodes straight to reclaim. */
1318	if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1319		return false;
1320
1321	/* Metadata inodes require explicit resource cleanup. */
1322	if (xfs_is_internal_inode(ip))
1323		return false;
1324
1325	/* Want to clean out the cow blocks if there are any. */
1326	if (cow_ifp && cow_ifp->if_bytes > 0)
1327		return true;
1328
1329	/* Unlinked files must be freed. */
1330	if (VFS_I(ip)->i_nlink == 0)
1331		return true;
1332
1333	/*
1334	 * This file isn't being freed, so check if there are post-eof blocks
1335	 * to free.
1336	 *
1337	 * Note: don't bother with iolock here since lockdep complains about
1338	 * acquiring it in reclaim context. We have the only reference to the
1339	 * inode at this point anyways.
1340	 */
1341	return xfs_can_free_eofblocks(ip);
1342}
1343
1344/*
1345 * Save health status somewhere, if we're dumping an inode with uncorrected
1346 * errors and online repair isn't running.
1347 */
1348static inline void
1349xfs_inactive_health(
1350	struct xfs_inode	*ip)
1351{
1352	struct xfs_mount	*mp = ip->i_mount;
1353	struct xfs_perag	*pag;
1354	unsigned int		sick;
1355	unsigned int		checked;
1356
1357	xfs_inode_measure_sickness(ip, &sick, &checked);
1358	if (!sick)
1359		return;
1360
1361	trace_xfs_inode_unfixed_corruption(ip, sick);
1362
1363	if (sick & XFS_SICK_INO_FORGET)
1364		return;
1365
1366	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1367	if (!pag) {
1368		/* There had better still be a perag structure! */
1369		ASSERT(0);
1370		return;
1371	}
1372
1373	xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1374	xfs_perag_put(pag);
1375}
1376
1377/*
1378 * xfs_inactive
1379 *
1380 * This is called when the vnode reference count for the vnode
1381 * goes to zero.  If the file has been unlinked, then it must
1382 * now be truncated.  Also, we clear all of the read-ahead state
1383 * kept for the inode here since the file is now closed.
1384 */
1385int
1386xfs_inactive(
1387	xfs_inode_t	*ip)
1388{
1389	struct xfs_mount	*mp;
1390	int			error = 0;
1391	int			truncate = 0;
1392
1393	/*
1394	 * If the inode is already free, then there can be nothing
1395	 * to clean up here.
1396	 */
1397	if (VFS_I(ip)->i_mode == 0) {
1398		ASSERT(ip->i_df.if_broot_bytes == 0);
1399		goto out;
1400	}
1401
1402	mp = ip->i_mount;
1403	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1404
1405	xfs_inactive_health(ip);
1406
1407	/*
1408	 * If this is a read-only mount, don't do this (would generate I/O)
1409	 * unless we're in log recovery and cleaning the iunlinked list.
1410	 */
1411	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1412		goto out;
1413
1414	/* Metadata inodes require explicit resource cleanup. */
1415	if (xfs_is_internal_inode(ip))
1416		goto out;
1417
1418	/* Try to clean out the cow blocks if there are any. */
1419	if (xfs_inode_has_cow_data(ip)) {
1420		error = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1421		if (error)
1422			goto out;
1423	}
1424
1425	if (VFS_I(ip)->i_nlink != 0) {
1426		/*
1427		 * Note: don't bother with iolock here since lockdep complains
1428		 * about acquiring it in reclaim context. We have the only
1429		 * reference to the inode at this point anyways.
1430		 */
1431		if (xfs_can_free_eofblocks(ip))
1432			error = xfs_free_eofblocks(ip);
1433
1434		goto out;
1435	}
1436
1437	if (S_ISREG(VFS_I(ip)->i_mode) &&
1438	    (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1439	     xfs_inode_has_filedata(ip)))
1440		truncate = 1;
1441
1442	if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1443		/*
1444		 * If this inode is being inactivated during a quotacheck and
1445		 * has not yet been scanned by quotacheck, we /must/ remove
1446		 * the dquots from the inode before inactivation changes the
1447		 * block and inode counts.  Most probably this is a result of
1448		 * reloading the incore iunlinked list to purge unrecovered
1449		 * unlinked inodes.
1450		 */
1451		xfs_qm_dqdetach(ip);
1452	} else {
1453		error = xfs_qm_dqattach(ip);
1454		if (error)
1455			goto out;
1456	}
1457
1458	if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1459		xfs_inactive_dir(ip);
1460		truncate = 1;
1461	}
1462
1463	if (S_ISLNK(VFS_I(ip)->i_mode))
1464		error = xfs_inactive_symlink(ip);
1465	else if (truncate)
1466		error = xfs_inactive_truncate(ip);
1467	if (error)
1468		goto out;
1469
1470	/*
1471	 * If there are attributes associated with the file then blow them away
1472	 * now.  The code calls a routine that recursively deconstructs the
1473	 * attribute fork. If also blows away the in-core attribute fork.
1474	 */
1475	if (xfs_inode_has_attr_fork(ip)) {
1476		error = xfs_attr_inactive(ip);
1477		if (error)
1478			goto out;
1479	}
1480
1481	ASSERT(ip->i_forkoff == 0);
1482
1483	/*
1484	 * Free the inode.
1485	 */
1486	error = xfs_inactive_ifree(ip);
1487
1488out:
1489	/*
1490	 * We're done making metadata updates for this inode, so we can release
1491	 * the attached dquots.
1492	 */
1493	xfs_qm_dqdetach(ip);
1494	return error;
1495}
1496
1497/*
1498 * Find an inode on the unlinked list. This does not take references to the
1499 * inode as we have existence guarantees by holding the AGI buffer lock and that
1500 * only unlinked, referenced inodes can be on the unlinked inode list.  If we
1501 * don't find the inode in cache, then let the caller handle the situation.
1502 */
1503struct xfs_inode *
1504xfs_iunlink_lookup(
1505	struct xfs_perag	*pag,
1506	xfs_agino_t		agino)
1507{
1508	struct xfs_inode	*ip;
1509
1510	rcu_read_lock();
1511	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1512	if (!ip) {
1513		/* Caller can handle inode not being in memory. */
1514		rcu_read_unlock();
1515		return NULL;
1516	}
1517
1518	/*
1519	 * Inode in RCU freeing limbo should not happen.  Warn about this and
1520	 * let the caller handle the failure.
1521	 */
1522	if (WARN_ON_ONCE(!ip->i_ino)) {
1523		rcu_read_unlock();
1524		return NULL;
1525	}
1526	ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1527	rcu_read_unlock();
1528	return ip;
1529}
1530
1531/*
1532 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1533 * to @prev_agino.  Caller must hold the AGI to synchronize with other changes
1534 * to the unlinked list.
1535 */
1536int
1537xfs_iunlink_reload_next(
1538	struct xfs_trans	*tp,
1539	struct xfs_buf		*agibp,
1540	xfs_agino_t		prev_agino,
1541	xfs_agino_t		next_agino)
1542{
1543	struct xfs_perag	*pag = agibp->b_pag;
1544	struct xfs_mount	*mp = pag_mount(pag);
1545	struct xfs_inode	*next_ip = NULL;
1546	int			error;
1547
1548	ASSERT(next_agino != NULLAGINO);
1549
1550#ifdef DEBUG
1551	rcu_read_lock();
1552	next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1553	ASSERT(next_ip == NULL);
1554	rcu_read_unlock();
1555#endif
1556
1557	xfs_info_ratelimited(mp,
1558 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating recovery.",
1559			next_agino, pag_agno(pag));
1560
1561	/*
1562	 * Use an untrusted lookup just to be cautious in case the AGI has been
1563	 * corrupted and now points at a free inode.  That shouldn't happen,
1564	 * but we'd rather shut down now since we're already running in a weird
1565	 * situation.
1566	 */
1567	error = xfs_iget(mp, tp, xfs_agino_to_ino(pag, next_agino),
1568			XFS_IGET_UNTRUSTED, 0, &next_ip);
1569	if (error) {
1570		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1571		return error;
1572	}
1573
1574	/* If this is not an unlinked inode, something is very wrong. */
1575	if (VFS_I(next_ip)->i_nlink != 0) {
1576		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1577		error = -EFSCORRUPTED;
1578		goto rele;
1579	}
1580
1581	next_ip->i_prev_unlinked = prev_agino;
1582	trace_xfs_iunlink_reload_next(next_ip);
1583rele:
1584	ASSERT(!(inode_state_read_once(VFS_I(next_ip)) & I_DONTCACHE));
1585	if (xfs_is_quotacheck_running(mp) && next_ip)
1586		xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1587	xfs_irele(next_ip);
1588	return error;
1589}
1590
1591/*
1592 * Look up the inode number specified and if it is not already marked XFS_ISTALE
1593 * mark it stale. We should only find clean inodes in this lookup that aren't
1594 * already stale.
1595 */
1596static void
1597xfs_ifree_mark_inode_stale(
1598	struct xfs_perag	*pag,
1599	struct xfs_inode	*free_ip,
1600	xfs_ino_t		inum)
1601{
1602	struct xfs_mount	*mp = pag_mount(pag);
1603	struct xfs_inode_log_item *iip;
1604	struct xfs_inode	*ip;
1605
1606retry:
1607	rcu_read_lock();
1608	ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1609
1610	/* Inode not in memory, nothing to do */
1611	if (!ip) {
1612		rcu_read_unlock();
1613		return;
1614	}
1615
1616	/*
1617	 * because this is an RCU protected lookup, we could find a recently
1618	 * freed or even reallocated inode during the lookup. We need to check
1619	 * under the i_flags_lock for a valid inode here. Skip it if it is not
1620	 * valid, the wrong inode or stale.
1621	 */
1622	spin_lock(&ip->i_flags_lock);
1623	if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1624		goto out_iflags_unlock;
1625
1626	/*
1627	 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
1628	 * other inodes that we did not find in the list attached to the buffer
1629	 * and are not already marked stale. If we can't lock it, back off and
1630	 * retry.
1631	 */
1632	if (ip != free_ip) {
1633		if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1634			spin_unlock(&ip->i_flags_lock);
1635			rcu_read_unlock();
1636			delay(1);
1637			goto retry;
1638		}
1639	}
1640	ip->i_flags |= XFS_ISTALE;
1641
1642	/*
1643	 * If the inode is flushing, it is already attached to the buffer.  All
1644	 * we needed to do here is mark the inode stale so buffer IO completion
1645	 * will remove it from the AIL.
1646	 */
1647	iip = ip->i_itemp;
1648	if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1649		ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1650		ASSERT(iip->ili_last_fields || xlog_is_shutdown(mp->m_log));
1651		goto out_iunlock;
1652	}
1653
1654	/*
1655	 * Inodes not attached to the buffer can be released immediately.
1656	 * Everything else has to go through xfs_iflush_abort() on journal
1657	 * commit as the flock synchronises removal of the inode from the
1658	 * cluster buffer against inode reclaim.
1659	 */
1660	if (!iip || list_empty(&iip->ili_item.li_bio_list))
1661		goto out_iunlock;
1662
1663	__xfs_iflags_set(ip, XFS_IFLUSHING);
1664	spin_unlock(&ip->i_flags_lock);
1665	rcu_read_unlock();
1666
1667	/* we have a dirty inode in memory that has not yet been flushed. */
1668	spin_lock(&iip->ili_lock);
1669	iip->ili_last_fields = iip->ili_fields;
1670	iip->ili_fields = 0;
1671	spin_unlock(&iip->ili_lock);
1672	ASSERT(iip->ili_last_fields);
1673
1674	if (ip != free_ip)
1675		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1676	return;
1677
1678out_iunlock:
1679	if (ip != free_ip)
1680		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1681out_iflags_unlock:
1682	spin_unlock(&ip->i_flags_lock);
1683	rcu_read_unlock();
1684}
1685
1686/*
1687 * A big issue when freeing the inode cluster is that we _cannot_ skip any
1688 * inodes that are in memory - they all must be marked stale and attached to
1689 * the cluster buffer.
1690 */
1691static int
1692xfs_ifree_cluster(
1693	struct xfs_trans	*tp,
1694	struct xfs_perag	*pag,
1695	struct xfs_inode	*free_ip,
1696	struct xfs_icluster	*xic)
1697{
1698	struct xfs_mount	*mp = free_ip->i_mount;
1699	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1700	struct xfs_buf		*bp;
1701	xfs_daddr_t		blkno;
1702	xfs_ino_t		inum = xic->first_ino;
1703	int			nbufs;
1704	int			i, j;
1705	int			ioffset;
1706	int			error;
1707
1708	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1709
1710	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1711		/*
1712		 * The allocation bitmap tells us which inodes of the chunk were
1713		 * physically allocated. Skip the cluster if an inode falls into
1714		 * a sparse region.
1715		 */
1716		ioffset = inum - xic->first_ino;
1717		if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1718			ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1719			continue;
1720		}
1721
1722		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1723					 XFS_INO_TO_AGBNO(mp, inum));
1724
1725		/*
1726		 * We obtain and lock the backing buffer first in the process
1727		 * here to ensure dirty inodes attached to the buffer remain in
1728		 * the flushing state while we mark them stale.
1729		 *
1730		 * If we scan the in-memory inodes first, then buffer IO can
1731		 * complete before we get a lock on it, and hence we may fail
1732		 * to mark all the active inodes on the buffer stale.
1733		 */
1734		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
1735				mp->m_bsize * igeo->blocks_per_cluster, 0, &bp);
1736		if (error)
1737			return error;
1738
1739		/*
1740		 * This buffer may not have been correctly initialised as we
1741		 * didn't read it from disk. That's not important because we are
1742		 * only using to mark the buffer as stale in the log, and to
1743		 * attach stale cached inodes on it.
1744		 *
1745		 * For the inode that triggered the cluster freeing, this
1746		 * attachment may occur in xfs_inode_item_precommit() after we
1747		 * have marked this buffer stale.  If this buffer was not in
1748		 * memory before xfs_ifree_cluster() started, it will not be
1749		 * marked XBF_DONE and this will cause problems later in
1750		 * xfs_inode_item_precommit() when we trip over a (stale, !done)
1751		 * buffer to attached to the transaction.
1752		 *
1753		 * Hence we have to mark the buffer as XFS_DONE here. This is
1754		 * safe because we are also marking the buffer as XBF_STALE and
1755		 * XFS_BLI_STALE. That means it will never be dispatched for
1756		 * IO and it won't be unlocked until the cluster freeing has
1757		 * been committed to the journal and the buffer unpinned. If it
1758		 * is written, we want to know about it, and we want it to
1759		 * fail. We can acheive this by adding a write verifier to the
1760		 * buffer.
1761		 */
1762		bp->b_flags |= XBF_DONE;
1763		bp->b_ops = &xfs_inode_buf_ops;
1764
1765		/*
1766		 * Now we need to set all the cached clean inodes as XFS_ISTALE,
1767		 * too. This requires lookups, and will skip inodes that we've
1768		 * already marked XFS_ISTALE.
1769		 */
1770		for (i = 0; i < igeo->inodes_per_cluster; i++)
1771			xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
1772
1773		xfs_trans_stale_inode_buf(tp, bp);
1774		xfs_trans_binval(tp, bp);
1775	}
1776	return 0;
1777}
1778
1779/*
1780 * This is called to return an inode to the inode free list.  The inode should
1781 * already be truncated to 0 length and have no pages associated with it.  This
1782 * routine also assumes that the inode is already a part of the transaction.
1783 *
1784 * The on-disk copy of the inode will have been added to the list of unlinked
1785 * inodes in the AGI. We need to remove the inode from that list atomically with
1786 * respect to freeing it here.
1787 */
1788int
1789xfs_ifree(
1790	struct xfs_trans	*tp,
1791	struct xfs_inode	*ip)
1792{
1793	struct xfs_mount	*mp = ip->i_mount;
1794	struct xfs_perag	*pag;
1795	struct xfs_icluster	xic = { 0 };
1796	struct xfs_inode_log_item *iip = ip->i_itemp;
1797	int			error;
1798
1799	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1800	ASSERT(VFS_I(ip)->i_nlink == 0);
1801	ASSERT(ip->i_df.if_nextents == 0);
1802	ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1803	ASSERT(ip->i_nblocks == 0);
1804
1805	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1806
1807	error = xfs_inode_uninit(tp, pag, ip, &xic);
1808	if (error)
1809		goto out;
1810
1811	if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1812		xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1813
1814	/* Don't attempt to replay owner changes for a deleted inode */
1815	spin_lock(&iip->ili_lock);
1816	iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1817	spin_unlock(&iip->ili_lock);
1818
1819	if (xic.deleted)
1820		error = xfs_ifree_cluster(tp, pag, ip, &xic);
1821out:
1822	xfs_perag_put(pag);
1823	return error;
1824}
1825
1826/*
1827 * This is called to unpin an inode.  The caller must have the inode locked
1828 * in at least shared mode so that the buffer cannot be subsequently pinned
1829 * once someone is waiting for it to be unpinned.
1830 */
1831static void
1832xfs_iunpin(
1833	struct xfs_inode	*ip)
1834{
1835	struct xfs_inode_log_item *iip = ip->i_itemp;
1836	xfs_csn_t		seq = 0;
1837
1838	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1839	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1840
1841	spin_lock(&iip->ili_lock);
1842	seq = iip->ili_commit_seq;
1843	spin_unlock(&iip->ili_lock);
1844	if (!seq)
1845		return;
1846
1847	/* Give the log a push to start the unpinning I/O */
1848	xfs_log_force_seq(ip->i_mount, seq, 0, NULL);
1849
1850}
1851
1852static void
1853__xfs_iunpin_wait(
1854	struct xfs_inode	*ip)
1855{
1856	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
1857	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1858
1859	xfs_iunpin(ip);
1860
1861	do {
1862		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1863		if (xfs_ipincount(ip))
1864			io_schedule();
1865	} while (xfs_ipincount(ip));
1866	finish_wait(wq, &wait.wq_entry);
1867}
1868
1869void
1870xfs_iunpin_wait(
1871	struct xfs_inode	*ip)
1872{
1873	if (xfs_ipincount(ip))
1874		__xfs_iunpin_wait(ip);
1875}
1876
1877/*
1878 * Removing an inode from the namespace involves removing the directory entry
1879 * and dropping the link count on the inode. Removing the directory entry can
1880 * result in locking an AGF (directory blocks were freed) and removing a link
1881 * count can result in placing the inode on an unlinked list which results in
1882 * locking an AGI.
1883 *
1884 * The big problem here is that we have an ordering constraint on AGF and AGI
1885 * locking - inode allocation locks the AGI, then can allocate a new extent for
1886 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1887 * removes the inode from the unlinked list, requiring that we lock the AGI
1888 * first, and then freeing the inode can result in an inode chunk being freed
1889 * and hence freeing disk space requiring that we lock an AGF.
1890 *
1891 * Hence the ordering that is imposed by other parts of the code is AGI before
1892 * AGF. This means we cannot remove the directory entry before we drop the inode
1893 * reference count and put it on the unlinked list as this results in a lock
1894 * order of AGF then AGI, and this can deadlock against inode allocation and
1895 * freeing. Therefore we must drop the link counts before we remove the
1896 * directory entry.
1897 *
1898 * This is still safe from a transactional point of view - it is not until we
1899 * get to xfs_defer_finish() that we have the possibility of multiple
1900 * transactions in this operation. Hence as long as we remove the directory
1901 * entry and drop the link count in the first transaction of the remove
1902 * operation, there are no transactional constraints on the ordering here.
1903 */
1904int
1905xfs_remove(
1906	struct xfs_inode	*dp,
1907	struct xfs_name		*name,
1908	struct xfs_inode	*ip)
1909{
1910	struct xfs_dir_update	du = {
1911		.dp		= dp,
1912		.name		= name,
1913		.ip		= ip,
1914	};
1915	struct xfs_mount	*mp = dp->i_mount;
1916	struct xfs_trans	*tp = NULL;
1917	int			is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1918	int			dontcare;
1919	int                     error = 0;
1920	uint			resblks;
1921
1922	trace_xfs_remove(dp, name);
1923
1924	if (xfs_is_shutdown(mp))
1925		return -EIO;
1926	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1927		return -EIO;
1928
1929	error = xfs_qm_dqattach(dp);
1930	if (error)
1931		goto std_return;
1932
1933	error = xfs_qm_dqattach(ip);
1934	if (error)
1935		goto std_return;
1936
1937	error = xfs_parent_start(mp, &du.ppargs);
1938	if (error)
1939		goto std_return;
1940
1941	/*
1942	 * We try to get the real space reservation first, allowing for
1943	 * directory btree deletion(s) implying possible bmap insert(s).  If we
1944	 * can't get the space reservation then we use 0 instead, and avoid the
1945	 * bmap btree insert(s) in the directory code by, if the bmap insert
1946	 * tries to happen, instead trimming the LAST block from the directory.
1947	 *
1948	 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
1949	 * the directory code can handle a reservationless update and we don't
1950	 * want to prevent a user from trying to free space by deleting things.
1951	 */
1952	resblks = xfs_remove_space_res(mp, name->len);
1953	error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
1954			&tp, &dontcare);
1955	if (error) {
1956		ASSERT(error != -ENOSPC);
1957		goto out_parent;
1958	}
1959
1960	error = xfs_dir_remove_child(tp, resblks, &du);
1961	if (error)
1962		goto out_trans_cancel;
1963
1964	/*
1965	 * If this is a synchronous mount, make sure that the
1966	 * remove transaction goes to disk before returning to
1967	 * the user.
1968	 */
1969	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1970		xfs_trans_set_sync(tp);
1971
1972	error = xfs_trans_commit(tp);
1973	if (error)
1974		goto out_unlock;
1975
1976	if (is_dir && xfs_inode_is_filestream(ip))
1977		xfs_filestream_deassociate(ip);
1978
1979	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1980	xfs_iunlock(dp, XFS_ILOCK_EXCL);
1981	xfs_parent_finish(mp, du.ppargs);
1982	return 0;
1983
1984 out_trans_cancel:
1985	xfs_trans_cancel(tp);
1986 out_unlock:
1987	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1988	xfs_iunlock(dp, XFS_ILOCK_EXCL);
1989 out_parent:
1990	xfs_parent_finish(mp, du.ppargs);
1991 std_return:
1992	return error;
1993}
1994
1995static inline void
1996xfs_iunlock_rename(
1997	struct xfs_inode	**i_tab,
1998	int			num_inodes)
1999{
2000	int			i;
2001
2002	for (i = num_inodes - 1; i >= 0; i--) {
2003		/* Skip duplicate inodes if src and target dps are the same */
2004		if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
2005			continue;
2006		xfs_iunlock(i_tab[i], XFS_ILOCK_EXCL);
2007	}
2008}
2009
2010/*
2011 * Enter all inodes for a rename transaction into a sorted array.
2012 */
2013#define __XFS_SORT_INODES	5
2014STATIC void
2015xfs_sort_for_rename(
2016	struct xfs_inode	*dp1,	/* in: old (source) directory inode */
2017	struct xfs_inode	*dp2,	/* in: new (target) directory inode */
2018	struct xfs_inode	*ip1,	/* in: inode of old entry */
2019	struct xfs_inode	*ip2,	/* in: inode of new entry */
2020	struct xfs_inode	*wip,	/* in: whiteout inode */
2021	struct xfs_inode	**i_tab,/* out: sorted array of inodes */
2022	int			*num_inodes)  /* in/out: inodes in array */
2023{
2024	int			i;
2025
2026	ASSERT(*num_inodes == __XFS_SORT_INODES);
2027	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2028
2029	/*
2030	 * i_tab contains a list of pointers to inodes.  We initialize
2031	 * the table here & we'll sort it.  We will then use it to
2032	 * order the acquisition of the inode locks.
2033	 *
2034	 * Note that the table may contain duplicates.  e.g., dp1 == dp2.
2035	 */
2036	i = 0;
2037	i_tab[i++] = dp1;
2038	i_tab[i++] = dp2;
2039	i_tab[i++] = ip1;
2040	if (ip2)
2041		i_tab[i++] = ip2;
2042	if (wip)
2043		i_tab[i++] = wip;
2044	*num_inodes = i;
2045
2046	xfs_sort_inodes(i_tab, *num_inodes);
2047}
2048
2049void
2050xfs_sort_inodes(
2051	struct xfs_inode	**i_tab,
2052	unsigned int		num_inodes)
2053{
2054	int			i, j;
2055
2056	ASSERT(num_inodes <= __XFS_SORT_INODES);
2057
2058	/*
2059	 * Sort the elements via bubble sort.  (Remember, there are at
2060	 * most 5 elements to sort, so this is adequate.)
2061	 */
2062	for (i = 0; i < num_inodes; i++) {
2063		for (j = 1; j < num_inodes; j++) {
2064			if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2065				swap(i_tab[j], i_tab[j - 1]);
2066		}
2067	}
2068}
2069
2070/*
2071 * xfs_rename_alloc_whiteout()
2072 *
2073 * Return a referenced, unlinked, unlocked inode that can be used as a
2074 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2075 * crash between allocating the inode and linking it into the rename transaction
2076 * recovery will free the inode and we won't leak it.
2077 */
2078static int
2079xfs_rename_alloc_whiteout(
2080	struct mnt_idmap	*idmap,
2081	struct xfs_name		*src_name,
2082	struct xfs_inode	*dp,
2083	struct xfs_inode	**wip)
2084{
2085	struct xfs_icreate_args	args = {
2086		.idmap		= idmap,
2087		.pip		= dp,
2088		.mode		= S_IFCHR | WHITEOUT_MODE,
2089		.flags		= XFS_ICREATE_TMPFILE,
2090	};
2091	struct xfs_inode	*tmpfile;
2092	struct qstr		name;
2093	int			error;
2094
2095	error = xfs_create_tmpfile(&args, &tmpfile);
2096	if (error)
2097		return error;
2098
2099	name.name = src_name->name;
2100	name.len = src_name->len;
2101	error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2102	if (error) {
2103		xfs_finish_inode_setup(tmpfile);
2104		xfs_irele(tmpfile);
2105		return error;
2106	}
2107
2108	/*
2109	 * Prepare the tmpfile inode as if it were created through the VFS.
2110	 * Complete the inode setup and flag it as linkable.  nlink is already
2111	 * zero, so we can skip the drop_nlink.
2112	 */
2113	xfs_setup_iops(tmpfile);
2114	xfs_finish_inode_setup(tmpfile);
2115	inode_state_set_raw(VFS_I(tmpfile), I_LINKABLE);
2116
2117	*wip = tmpfile;
2118	return 0;
2119}
2120
2121/*
2122 * xfs_rename
2123 */
2124int
2125xfs_rename(
2126	struct mnt_idmap	*idmap,
2127	struct xfs_inode	*src_dp,
2128	struct xfs_name		*src_name,
2129	struct xfs_inode	*src_ip,
2130	struct xfs_inode	*target_dp,
2131	struct xfs_name		*target_name,
2132	struct xfs_inode	*target_ip,
2133	unsigned int		flags)
2134{
2135	struct xfs_dir_update	du_src = {
2136		.dp		= src_dp,
2137		.name		= src_name,
2138		.ip		= src_ip,
2139	};
2140	struct xfs_dir_update	du_tgt = {
2141		.dp		= target_dp,
2142		.name		= target_name,
2143		.ip		= target_ip,
2144	};
2145	struct xfs_dir_update	du_wip = { };
2146	struct xfs_mount	*mp = src_dp->i_mount;
2147	struct xfs_trans	*tp;
2148	struct xfs_inode	*inodes[__XFS_SORT_INODES];
2149	int			i;
2150	int			num_inodes = __XFS_SORT_INODES;
2151	bool			new_parent = (src_dp != target_dp);
2152	bool			src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2153	int			spaceres;
2154	bool			retried = false;
2155	int			error, nospace_error = 0;
2156
2157	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2158
2159	if ((flags & RENAME_EXCHANGE) && !target_ip)
2160		return -EINVAL;
2161
2162	/*
2163	 * If we are doing a whiteout operation, allocate the whiteout inode
2164	 * we will be placing at the target and ensure the type is set
2165	 * appropriately.
2166	 */
2167	if (flags & RENAME_WHITEOUT) {
2168		error = xfs_rename_alloc_whiteout(idmap, src_name, target_dp,
2169				&du_wip.ip);
2170		if (error)
2171			return error;
2172
2173		/* setup target dirent info as whiteout */
2174		src_name->type = XFS_DIR3_FT_CHRDEV;
2175	}
2176
2177	xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, du_wip.ip,
2178			inodes, &num_inodes);
2179
2180	error = xfs_parent_start(mp, &du_src.ppargs);
2181	if (error)
2182		goto out_release_wip;
2183
2184	if (du_wip.ip) {
2185		error = xfs_parent_start(mp, &du_wip.ppargs);
2186		if (error)
2187			goto out_src_ppargs;
2188	}
2189
2190	if (target_ip) {
2191		error = xfs_parent_start(mp, &du_tgt.ppargs);
2192		if (error)
2193			goto out_wip_ppargs;
2194	}
2195
2196retry:
2197	nospace_error = 0;
2198	spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2199			target_name->len, du_wip.ip != NULL);
2200	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2201	if (error == -ENOSPC) {
2202		nospace_error = error;
2203		spaceres = 0;
2204		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2205				&tp);
2206	}
2207	if (error)
2208		goto out_tgt_ppargs;
2209
2210	/*
2211	 * We don't allow reservationless renaming when parent pointers are
2212	 * enabled because we can't back out if the xattrs must grow.
2213	 */
2214	if (du_src.ppargs && nospace_error) {
2215		error = nospace_error;
2216		xfs_trans_cancel(tp);
2217		goto out_tgt_ppargs;
2218	}
2219
2220	/*
2221	 * Attach the dquots to the inodes
2222	 */
2223	error = xfs_qm_vop_rename_dqattach(inodes);
2224	if (error) {
2225		xfs_trans_cancel(tp);
2226		goto out_tgt_ppargs;
2227	}
2228
2229	/*
2230	 * Lock all the participating inodes. Depending upon whether
2231	 * the target_name exists in the target directory, and
2232	 * whether the target directory is the same as the source
2233	 * directory, we can lock from 2 to 5 inodes.
2234	 */
2235	xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2236
2237	/*
2238	 * Join all the inodes to the transaction.
2239	 */
2240	xfs_trans_ijoin(tp, src_dp, 0);
2241	if (new_parent)
2242		xfs_trans_ijoin(tp, target_dp, 0);
2243	xfs_trans_ijoin(tp, src_ip, 0);
2244	if (target_ip)
2245		xfs_trans_ijoin(tp, target_ip, 0);
2246	if (du_wip.ip)
2247		xfs_trans_ijoin(tp, du_wip.ip, 0);
2248
2249	error = xfs_projid_differ(target_dp, src_ip);
2250	if (error)
2251		goto out_trans_cancel;
2252
2253	/* RENAME_EXCHANGE is unique from here on. */
2254	if (flags & RENAME_EXCHANGE) {
2255		error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2256				spaceres);
2257		if (error)
2258			goto out_trans_cancel;
2259		goto out_commit;
2260	}
2261
2262	/*
2263	 * Try to reserve quota to handle an expansion of the target directory.
2264	 * We'll allow the rename to continue in reservationless mode if we hit
2265	 * a space usage constraint.  If we trigger reservationless mode, save
2266	 * the errno if there isn't any free space in the target directory.
2267	 */
2268	if (spaceres != 0) {
2269		error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2270				0, false);
2271		if (error == -EDQUOT || error == -ENOSPC) {
2272			if (!retried) {
2273				xfs_trans_cancel(tp);
2274				xfs_iunlock_rename(inodes, num_inodes);
2275				xfs_blockgc_free_quota(target_dp, 0);
2276				retried = true;
2277				goto retry;
2278			}
2279
2280			nospace_error = error;
2281			spaceres = 0;
2282			error = 0;
2283		}
2284		if (error)
2285			goto out_trans_cancel;
2286	}
2287
2288	/*
2289	 * We don't allow quotaless renaming when parent pointers are enabled
2290	 * because we can't back out if the xattrs must grow.
2291	 */
2292	if (du_src.ppargs && nospace_error) {
2293		error = nospace_error;
2294		goto out_trans_cancel;
2295	}
2296
2297	/*
2298	 * Lock the AGI buffers we need to handle bumping the nlink of the
2299	 * whiteout inode off the unlinked list and to handle dropping the
2300	 * nlink of the target inode.  Per locking order rules, do this in
2301	 * increasing AG order and before directory block allocation tries to
2302	 * grab AGFs because we grab AGIs before AGFs.
2303	 *
2304	 * The (vfs) caller must ensure that if src is a directory then
2305	 * target_ip is either null or an empty directory.
2306	 */
2307	for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2308		if (inodes[i] == du_wip.ip ||
2309		    (inodes[i] == target_ip &&
2310		     (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2311			struct xfs_perag	*pag;
2312			struct xfs_buf		*bp;
2313
2314			pag = xfs_perag_get(mp,
2315					XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2316			error = xfs_read_agi(pag, tp, 0, &bp);
2317			xfs_perag_put(pag);
2318			if (error)
2319				goto out_trans_cancel;
2320		}
2321	}
2322
2323	error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2324			&du_wip);
2325	if (error)
2326		goto out_trans_cancel;
2327
2328	if (du_wip.ip) {
2329		/*
2330		 * Now we have a real link, clear the "I'm a tmpfile" state
2331		 * flag from the inode so it doesn't accidentally get misused in
2332		 * future.
2333		 */
2334		inode_state_clear_raw(VFS_I(du_wip.ip), I_LINKABLE);
2335	}
2336
2337out_commit:
2338	/*
2339	 * If this is a synchronous mount, make sure that the rename
2340	 * transaction goes to disk before returning to the user.
2341	 */
2342	if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2343		xfs_trans_set_sync(tp);
2344
2345	error = xfs_trans_commit(tp);
2346	nospace_error = 0;
2347	goto out_unlock;
2348
2349out_trans_cancel:
2350	xfs_trans_cancel(tp);
2351out_unlock:
2352	xfs_iunlock_rename(inodes, num_inodes);
2353out_tgt_ppargs:
2354	xfs_parent_finish(mp, du_tgt.ppargs);
2355out_wip_ppargs:
2356	xfs_parent_finish(mp, du_wip.ppargs);
2357out_src_ppargs:
2358	xfs_parent_finish(mp, du_src.ppargs);
2359out_release_wip:
2360	if (du_wip.ip)
2361		xfs_irele(du_wip.ip);
2362	if (error == -ENOSPC && nospace_error)
2363		error = nospace_error;
2364	return error;
2365}
2366
2367static int
2368xfs_iflush(
2369	struct xfs_inode	*ip,
2370	struct xfs_buf		*bp)
2371{
2372	struct xfs_inode_log_item *iip = ip->i_itemp;
2373	struct xfs_dinode	*dip;
2374	struct xfs_mount	*mp = ip->i_mount;
2375	int			error;
2376
2377	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2378	ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2379	ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2380	       ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2381	ASSERT(iip->ili_item.li_buf == bp);
2382
2383	dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
2384
2385	/*
2386	 * We don't flush the inode if any of the following checks fail, but we
2387	 * do still update the log item and attach to the backing buffer as if
2388	 * the flush happened. This is a formality to facilitate predictable
2389	 * error handling as the caller will shutdown and fail the buffer.
2390	 */
2391	error = -EFSCORRUPTED;
2392	if (dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC) ||
2393	    XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_1)) {
2394		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2395			"%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2396			__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2397		goto flush_out;
2398	}
2399	if (ip->i_df.if_format == XFS_DINODE_FMT_META_BTREE) {
2400		if (!S_ISREG(VFS_I(ip)->i_mode) ||
2401		    !(ip->i_diflags2 & XFS_DIFLAG2_METADATA)) {
2402			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2403				"%s: Bad %s meta btree inode %Lu, ptr "PTR_FMT,
2404				__func__, xfs_metafile_type_str(ip->i_metatype),
2405				ip->i_ino, ip);
2406			goto flush_out;
2407		}
2408	} else if (S_ISREG(VFS_I(ip)->i_mode)) {
2409		if ((ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2410		     ip->i_df.if_format != XFS_DINODE_FMT_BTREE) ||
2411		    XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_3)) {
2412			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2413				"%s: Bad regular inode %llu, ptr "PTR_FMT,
2414				__func__, ip->i_ino, ip);
2415			goto flush_out;
2416		}
2417	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2418		if ((ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2419		     ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2420		     ip->i_df.if_format != XFS_DINODE_FMT_LOCAL) ||
2421		    XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_4)) {
2422			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2423				"%s: Bad directory inode %llu, ptr "PTR_FMT,
2424				__func__, ip->i_ino, ip);
2425			goto flush_out;
2426		}
2427	}
2428	if (ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2429	    ip->i_nblocks || XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_5)) {
2430		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2431			"%s: detected corrupt incore inode %llu, "
2432			"total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2433			__func__, ip->i_ino,
2434			ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2435			ip->i_nblocks, ip);
2436		goto flush_out;
2437	}
2438	if (ip->i_forkoff > mp->m_sb.sb_inodesize ||
2439	    XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_6)) {
2440		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2441			"%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2442			__func__, ip->i_ino, ip->i_forkoff, ip);
2443		goto flush_out;
2444	}
2445
2446	if (xfs_inode_has_attr_fork(ip) &&
2447	    ip->i_af.if_format == XFS_DINODE_FMT_META_BTREE) {
2448		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2449			"%s: meta btree in inode %Lu attr fork, ptr "PTR_FMT,
2450			__func__, ip->i_ino, ip);
2451		goto flush_out;
2452	}
2453
2454	/*
2455	 * Inode item log recovery for v2 inodes are dependent on the flushiter
2456	 * count for correct sequencing.  We bump the flush iteration count so
2457	 * we can detect flushes which postdate a log record during recovery.
2458	 * This is redundant as we now log every change and hence this can't
2459	 * happen but we need to still do it to ensure backwards compatibility
2460	 * with old kernels that predate logging all inode changes.
2461	 */
2462	if (!xfs_has_v3inodes(mp))
2463		ip->i_flushiter++;
2464
2465	/*
2466	 * If there are inline format data / attr forks attached to this inode,
2467	 * make sure they are not corrupt.
2468	 */
2469	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2470	    xfs_ifork_verify_local_data(ip))
2471		goto flush_out;
2472	if (xfs_inode_has_attr_fork(ip) &&
2473	    ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2474	    xfs_ifork_verify_local_attr(ip))
2475		goto flush_out;
2476
2477	/*
2478	 * Copy the dirty parts of the inode into the on-disk inode.  We always
2479	 * copy out the core of the inode, because if the inode is dirty at all
2480	 * the core must be.
2481	 */
2482	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2483
2484	/* Wrap, we never let the log put out DI_MAX_FLUSH */
2485	if (!xfs_has_v3inodes(mp)) {
2486		if (ip->i_flushiter == DI_MAX_FLUSH)
2487			ip->i_flushiter = 0;
2488	}
2489
2490	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2491	if (xfs_inode_has_attr_fork(ip))
2492		xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2493
2494	/*
2495	 * We've recorded everything logged in the inode, so we'd like to clear
2496	 * the ili_fields bits so we don't log and flush things unnecessarily.
2497	 * However, we can't stop logging all this information until the data
2498	 * we've copied into the disk buffer is written to disk.  If we did we
2499	 * might overwrite the copy of the inode in the log with all the data
2500	 * after re-logging only part of it, and in the face of a crash we
2501	 * wouldn't have all the data we need to recover.
2502	 *
2503	 * What we do is move the bits to the ili_last_fields field.  When
2504	 * logging the inode, these bits are moved back to the ili_fields field.
2505	 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2506	 * we know that the information those bits represent is permanently on
2507	 * disk.  As long as the flush completes before the inode is logged
2508	 * again, then both ili_fields and ili_last_fields will be cleared.
2509	 */
2510	error = 0;
2511flush_out:
2512	spin_lock(&iip->ili_lock);
2513	iip->ili_last_fields = iip->ili_fields;
2514	iip->ili_fields = 0;
2515	set_bit(XFS_LI_FLUSHING, &iip->ili_item.li_flags);
2516	spin_unlock(&iip->ili_lock);
2517
2518	/*
2519	 * Store the current LSN of the inode so that we can tell whether the
2520	 * item has moved in the AIL from xfs_buf_inode_iodone().
2521	 */
2522	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2523				&iip->ili_item.li_lsn);
2524
2525	/* generate the checksum. */
2526	xfs_dinode_calc_crc(mp, dip);
2527	if (error)
2528		xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2529	return error;
2530}
2531
2532/*
2533 * Non-blocking flush of dirty inode metadata into the backing buffer.
2534 *
2535 * The caller must have a reference to the inode and hold the cluster buffer
2536 * locked. The function will walk across all the inodes on the cluster buffer it
2537 * can find and lock without blocking, and flush them to the cluster buffer.
2538 *
2539 * On successful flushing of at least one inode, the caller must write out the
2540 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2541 * the caller needs to release the buffer. On failure, the filesystem will be
2542 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2543 * will be returned.
2544 */
2545int
2546xfs_iflush_cluster(
2547	struct xfs_buf		*bp)
2548{
2549	struct xfs_mount	*mp = bp->b_mount;
2550	struct xfs_log_item	*lip, *n;
2551	struct xfs_inode	*ip;
2552	struct xfs_inode_log_item *iip;
2553	int			clcount = 0;
2554	int			error = 0;
2555
2556	/*
2557	 * We must use the safe variant here as on shutdown xfs_iflush_abort()
2558	 * will remove itself from the list.
2559	 */
2560	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2561		iip = (struct xfs_inode_log_item *)lip;
2562		ip = iip->ili_inode;
2563
2564		/*
2565		 * Quick and dirty check to avoid locks if possible.
2566		 */
2567		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2568			continue;
2569		if (xfs_ipincount(ip))
2570			continue;
2571
2572		/*
2573		 * The inode is still attached to the buffer, which means it is
2574		 * dirty but reclaim might try to grab it. Check carefully for
2575		 * that, and grab the ilock while still holding the i_flags_lock
2576		 * to guarantee reclaim will not be able to reclaim this inode
2577		 * once we drop the i_flags_lock.
2578		 */
2579		spin_lock(&ip->i_flags_lock);
2580		ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2581		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2582			spin_unlock(&ip->i_flags_lock);
2583			continue;
2584		}
2585
2586		/*
2587		 * ILOCK will pin the inode against reclaim and prevent
2588		 * concurrent transactions modifying the inode while we are
2589		 * flushing the inode. If we get the lock, set the flushing
2590		 * state before we drop the i_flags_lock.
2591		 */
2592		if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2593			spin_unlock(&ip->i_flags_lock);
2594			continue;
2595		}
2596		__xfs_iflags_set(ip, XFS_IFLUSHING);
2597		spin_unlock(&ip->i_flags_lock);
2598
2599		/*
2600		 * Abort flushing this inode if we are shut down because the
2601		 * inode may not currently be in the AIL. This can occur when
2602		 * log I/O failure unpins the inode without inserting into the
2603		 * AIL, leaving a dirty/unpinned inode attached to the buffer
2604		 * that otherwise looks like it should be flushed.
2605		 */
2606		if (xlog_is_shutdown(mp->m_log)) {
2607			xfs_iunpin_wait(ip);
2608			xfs_iflush_abort(ip);
2609			xfs_iunlock(ip, XFS_ILOCK_SHARED);
2610			error = -EIO;
2611			continue;
2612		}
2613
2614		/* don't block waiting on a log force to unpin dirty inodes */
2615		if (xfs_ipincount(ip)) {
2616			xfs_iflags_clear(ip, XFS_IFLUSHING);
2617			xfs_iunlock(ip, XFS_ILOCK_SHARED);
2618			continue;
2619		}
2620
2621		if (!xfs_inode_clean(ip))
2622			error = xfs_iflush(ip, bp);
2623		else
2624			xfs_iflags_clear(ip, XFS_IFLUSHING);
2625		xfs_iunlock(ip, XFS_ILOCK_SHARED);
2626		if (error)
2627			break;
2628		clcount++;
2629	}
2630
2631	if (error) {
2632		/*
2633		 * Shutdown first so we kill the log before we release this
2634		 * buffer. If it is an INODE_ALLOC buffer and pins the tail
2635		 * of the log, failing it before the _log_ is shut down can
2636		 * result in the log tail being moved forward in the journal
2637		 * on disk because log writes can still be taking place. Hence
2638		 * unpinning the tail will allow the ICREATE intent to be
2639		 * removed from the log an recovery will fail with uninitialised
2640		 * inode cluster buffers.
2641		 */
2642		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2643		bp->b_flags |= XBF_ASYNC;
2644		xfs_buf_ioend_fail(bp);
2645		return error;
2646	}
2647
2648	if (!clcount)
2649		return -EAGAIN;
2650
2651	XFS_STATS_INC(mp, xs_icluster_flushcnt);
2652	XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2653	return 0;
2654
2655}
2656
2657/* Release an inode. */
2658void
2659xfs_irele(
2660	struct xfs_inode	*ip)
2661{
2662	trace_xfs_irele(ip, _RET_IP_);
2663	iput(VFS_I(ip));
2664}
2665
2666/*
2667 * Ensure all commited transactions touching the inode are written to the log.
2668 */
2669int
2670xfs_log_force_inode(
2671	struct xfs_inode	*ip)
2672{
2673	struct xfs_inode_log_item *iip = ip->i_itemp;
2674	xfs_csn_t		seq = 0;
2675
2676	if (!iip)
2677		return 0;
2678
2679	spin_lock(&iip->ili_lock);
2680	seq = iip->ili_commit_seq;
2681	spin_unlock(&iip->ili_lock);
2682
2683	if (!seq)
2684		return 0;
2685	return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2686}
2687
2688/*
2689 * Grab the exclusive iolock for a data copy from src to dest, making sure to
2690 * abide vfs locking order (lowest pointer value goes first) and breaking the
2691 * layout leases before proceeding.  The loop is needed because we cannot call
2692 * the blocking break_layout() with the iolocks held, and therefore have to
2693 * back out both locks.
2694 */
2695static int
2696xfs_iolock_two_inodes_and_break_layout(
2697	struct inode		*src,
2698	struct inode		*dest)
2699{
2700	int			error;
2701
2702	if (src > dest)
2703		swap(src, dest);
2704
2705retry:
2706	/* Wait to break both inodes' layouts before we start locking. */
2707	error = break_layout(src, true);
2708	if (error)
2709		return error;
2710	if (src != dest) {
2711		error = break_layout(dest, true);
2712		if (error)
2713			return error;
2714	}
2715
2716	/* Lock one inode and make sure nobody got in and leased it. */
2717	inode_lock(src);
2718	error = break_layout(src, false);
2719	if (error) {
2720		inode_unlock(src);
2721		if (error == -EWOULDBLOCK)
2722			goto retry;
2723		return error;
2724	}
2725
2726	if (src == dest)
2727		return 0;
2728
2729	/* Lock the other inode and make sure nobody got in and leased it. */
2730	inode_lock_nested(dest, I_MUTEX_NONDIR2);
2731	error = break_layout(dest, false);
2732	if (error) {
2733		inode_unlock(src);
2734		inode_unlock(dest);
2735		if (error == -EWOULDBLOCK)
2736			goto retry;
2737		return error;
2738	}
2739
2740	return 0;
2741}
2742
2743static int
2744xfs_mmaplock_two_inodes_and_break_dax_layout(
2745	struct xfs_inode	*ip1,
2746	struct xfs_inode	*ip2)
2747{
2748	int			error;
2749
2750	if (ip1->i_ino > ip2->i_ino)
2751		swap(ip1, ip2);
2752
2753again:
2754	/* Lock the first inode */
2755	xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
2756	error = xfs_break_dax_layouts(VFS_I(ip1));
2757	if (error) {
2758		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2759		return error;
2760	}
2761
2762	if (ip1 == ip2)
2763		return 0;
2764
2765	/* Nested lock the second inode */
2766	xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
2767	/*
2768	 * We cannot use xfs_break_dax_layouts() directly here because it may
2769	 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2770	 * for this nested lock case.
2771	 */
2772	error = dax_break_layout(VFS_I(ip2), 0, -1, NULL);
2773	if (error) {
2774		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2775		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2776		goto again;
2777	}
2778
2779	return 0;
2780}
2781
2782/*
2783 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2784 * mmap activity.
2785 */
2786int
2787xfs_ilock2_io_mmap(
2788	struct xfs_inode	*ip1,
2789	struct xfs_inode	*ip2)
2790{
2791	int			ret;
2792
2793	ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
2794	if (ret)
2795		return ret;
2796
2797	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2798		ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2799		if (ret) {
2800			inode_unlock(VFS_I(ip2));
2801			if (ip1 != ip2)
2802				inode_unlock(VFS_I(ip1));
2803			return ret;
2804		}
2805	} else
2806		filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
2807					    VFS_I(ip2)->i_mapping);
2808
2809	return 0;
2810}
2811
2812/* Unlock both inodes to allow IO and mmap activity. */
2813void
2814xfs_iunlock2_io_mmap(
2815	struct xfs_inode	*ip1,
2816	struct xfs_inode	*ip2)
2817{
2818	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2819		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2820		if (ip1 != ip2)
2821			xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2822	} else
2823		filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
2824					      VFS_I(ip2)->i_mapping);
2825
2826	inode_unlock(VFS_I(ip2));
2827	if (ip1 != ip2)
2828		inode_unlock(VFS_I(ip1));
2829}
2830
2831/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2832void
2833xfs_iunlock2_remapping(
2834	struct xfs_inode	*ip1,
2835	struct xfs_inode	*ip2)
2836{
2837	xfs_iflags_clear(ip1, XFS_IREMAPPING);
2838
2839	if (ip1 != ip2)
2840		xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
2841	xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2842
2843	if (ip1 != ip2)
2844		inode_unlock_shared(VFS_I(ip1));
2845	inode_unlock(VFS_I(ip2));
2846}
2847
2848/*
2849 * Reload the incore inode list for this inode.  Caller should ensure that
2850 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
2851 * preventing other threads from executing.
2852 */
2853int
2854xfs_inode_reload_unlinked_bucket(
2855	struct xfs_trans	*tp,
2856	struct xfs_inode	*ip)
2857{
2858	struct xfs_mount	*mp = tp->t_mountp;
2859	struct xfs_buf		*agibp;
2860	struct xfs_agi		*agi;
2861	struct xfs_perag	*pag;
2862	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2863	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2864	xfs_agino_t		prev_agino, next_agino;
2865	unsigned int		bucket;
2866	bool			foundit = false;
2867	int			error;
2868
2869	/* Grab the first inode in the list */
2870	pag = xfs_perag_get(mp, agno);
2871	error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2872	xfs_perag_put(pag);
2873	if (error)
2874		return error;
2875
2876	/*
2877	 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2878	 * incore unlinked list pointers for this inode.  Check once more to
2879	 * see if we raced with anyone else to reload the unlinked list.
2880	 */
2881	if (!xfs_inode_unlinked_incomplete(ip)) {
2882		foundit = true;
2883		goto out_agibp;
2884	}
2885
2886	bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2887	agi = agibp->b_addr;
2888
2889	trace_xfs_inode_reload_unlinked_bucket(ip);
2890
2891	xfs_info_ratelimited(mp,
2892 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating list recovery.",
2893			agino, agno);
2894
2895	prev_agino = NULLAGINO;
2896	next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2897	while (next_agino != NULLAGINO) {
2898		struct xfs_inode	*next_ip = NULL;
2899
2900		/* Found this caller's inode, set its backlink. */
2901		if (next_agino == agino) {
2902			next_ip = ip;
2903			next_ip->i_prev_unlinked = prev_agino;
2904			foundit = true;
2905			goto next_inode;
2906		}
2907
2908		/* Try in-memory lookup first. */
2909		next_ip = xfs_iunlink_lookup(pag, next_agino);
2910		if (next_ip)
2911			goto next_inode;
2912
2913		/* Inode not in memory, try reloading it. */
2914		error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2915				next_agino);
2916		if (error)
2917			break;
2918
2919		/* Grab the reloaded inode. */
2920		next_ip = xfs_iunlink_lookup(pag, next_agino);
2921		if (!next_ip) {
2922			/* No incore inode at all?  We reloaded it... */
2923			ASSERT(next_ip != NULL);
2924			error = -EFSCORRUPTED;
2925			break;
2926		}
2927
2928next_inode:
2929		prev_agino = next_agino;
2930		next_agino = next_ip->i_next_unlinked;
2931	}
2932
2933out_agibp:
2934	xfs_trans_brelse(tp, agibp);
2935	/* Should have found this inode somewhere in the iunlinked bucket. */
2936	if (!error && !foundit)
2937		error = -EFSCORRUPTED;
2938	return error;
2939}
2940
2941/* Decide if this inode is missing its unlinked list and reload it. */
2942int
2943xfs_inode_reload_unlinked(
2944	struct xfs_inode	*ip)
2945{
2946	struct xfs_trans	*tp;
2947	int			error = 0;
2948
2949	tp = xfs_trans_alloc_empty(ip->i_mount);
2950	xfs_ilock(ip, XFS_ILOCK_SHARED);
2951	if (xfs_inode_unlinked_incomplete(ip))
2952		error = xfs_inode_reload_unlinked_bucket(tp, ip);
2953	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2954	xfs_trans_cancel(tp);
2955
2956	return error;
2957}
2958
2959/* Has this inode fork been zapped by repair? */
2960bool
2961xfs_ifork_zapped(
2962	const struct xfs_inode	*ip,
2963	int			whichfork)
2964{
2965	unsigned int		datamask = 0;
2966
2967	switch (whichfork) {
2968	case XFS_DATA_FORK:
2969		switch (ip->i_vnode.i_mode & S_IFMT) {
2970		case S_IFDIR:
2971			datamask = XFS_SICK_INO_DIR_ZAPPED;
2972			break;
2973		case S_IFLNK:
2974			datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
2975			break;
2976		}
2977		return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
2978	case XFS_ATTR_FORK:
2979		return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
2980	default:
2981		return false;
2982	}
2983}
2984
2985/* Compute the number of data and realtime blocks used by a file. */
2986void
2987xfs_inode_count_blocks(
2988	struct xfs_trans	*tp,
2989	struct xfs_inode	*ip,
2990	xfs_filblks_t		*dblocks,
2991	xfs_filblks_t		*rblocks)
2992{
2993	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
2994
2995	*rblocks = 0;
2996	if (XFS_IS_REALTIME_INODE(ip))
2997		xfs_bmap_count_leaves(ifp, rblocks);
2998	*dblocks = ip->i_nblocks - *rblocks;
2999}
3000
3001static void
3002xfs_wait_dax_page(
3003	struct inode		*inode)
3004{
3005	struct xfs_inode        *ip = XFS_I(inode);
3006
3007	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
3008	schedule();
3009	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
3010}
3011
3012int
3013xfs_break_dax_layouts(
3014	struct inode		*inode)
3015{
3016	xfs_assert_ilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL);
3017
3018	return dax_break_layout_inode(inode, xfs_wait_dax_page);
3019}
3020
3021int
3022xfs_break_layouts(
3023	struct inode		*inode,
3024	uint			*iolock,
3025	enum layout_break_reason reason)
3026{
3027	bool			retry;
3028	int			error;
3029
3030	xfs_assert_ilocked(XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3031
3032	do {
3033		retry = false;
3034		switch (reason) {
3035		case BREAK_UNMAP:
3036			error = xfs_break_dax_layouts(inode);
3037			if (error)
3038				break;
3039			fallthrough;
3040		case BREAK_WRITE:
3041			error = xfs_break_leased_layouts(inode, iolock, &retry);
3042			break;
3043		default:
3044			WARN_ON_ONCE(1);
3045			error = -EINVAL;
3046		}
3047	} while (error == 0 && retry);
3048
3049	return error;
3050}
3051
3052/* Returns the size of fundamental allocation unit for a file, in bytes. */
3053unsigned int
3054xfs_inode_alloc_unitsize(
3055	struct xfs_inode	*ip)
3056{
3057	unsigned int		blocks = 1;
3058
3059	if (XFS_IS_REALTIME_INODE(ip))
3060		blocks = ip->i_mount->m_sb.sb_rextsize;
3061
3062	return XFS_FSB_TO_B(ip->i_mount, blocks);
3063}
3064
3065/* Should we always be using copy on write for file writes? */
3066bool
3067xfs_is_always_cow_inode(
3068	const struct xfs_inode	*ip)
3069{
3070	return xfs_is_zoned_inode(ip) ||
3071		(ip->i_mount->m_always_cow && xfs_has_reflink(ip->i_mount));
3072}