Linux Audio

Check our new training course

Loading...
v6.13.7
   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
 880int
 881xfs_link(
 882	struct xfs_inode	*tdp,
 883	struct xfs_inode	*sip,
 884	struct xfs_name		*target_name)
 885{
 886	struct xfs_dir_update	du = {
 887		.dp		= tdp,
 888		.name		= target_name,
 889		.ip		= sip,
 890	};
 891	struct xfs_mount	*mp = tdp->i_mount;
 892	struct xfs_trans	*tp;
 893	int			error, nospace_error = 0;
 894	int			resblks;
 895
 896	trace_xfs_link(tdp, target_name);
 897
 898	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
 899
 900	if (xfs_is_shutdown(mp))
 901		return -EIO;
 902	if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
 903		return -EIO;
 904
 905	error = xfs_qm_dqattach(sip);
 906	if (error)
 907		goto std_return;
 908
 909	error = xfs_qm_dqattach(tdp);
 910	if (error)
 911		goto std_return;
 912
 913	error = xfs_parent_start(mp, &du.ppargs);
 
 
 
 
 
 914	if (error)
 915		goto std_return;
 916
 917	resblks = xfs_link_space_res(mp, target_name->len);
 918	error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
 919			&tp, &nospace_error);
 920	if (error)
 921		goto out_parent;
 922
 923	/*
 924	 * We don't allow reservationless or quotaless hardlinking when parent
 925	 * pointers are enabled because we can't back out if the xattrs must
 926	 * grow.
 927	 */
 928	if (du.ppargs && nospace_error) {
 929		error = nospace_error;
 930		goto error_return;
 931	}
 932
 933	/*
 934	 * If we are using project inheritance, we only allow hard link
 935	 * creation in our tree when the project IDs are the same; else
 936	 * the tree quota mechanism could be circumvented.
 937	 */
 938	if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
 939		     tdp->i_projid != sip->i_projid)) {
 940		/*
 941		 * Project quota setup skips special files which can
 942		 * leave inodes in a PROJINHERIT directory without a
 943		 * project ID set. We need to allow links to be made
 944		 * to these "project-less" inodes because userspace
 945		 * expects them to succeed after project ID setup,
 946		 * but everything else should be rejected.
 947		 */
 948		if (!special_file(VFS_I(sip)->i_mode) ||
 949		    sip->i_projid != 0) {
 950			error = -EXDEV;
 951			goto error_return;
 952		}
 953	}
 954
 955	error = xfs_dir_add_child(tp, resblks, &du);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 956	if (error)
 957		goto error_return;
 
 
 
 
 958
 959	/*
 960	 * If this is a synchronous mount, make sure that the
 961	 * link transaction goes to disk before returning to
 962	 * the user.
 963	 */
 964	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
 965		xfs_trans_set_sync(tp);
 966
 967	error = xfs_trans_commit(tp);
 968	xfs_iunlock(tdp, XFS_ILOCK_EXCL);
 969	xfs_iunlock(sip, XFS_ILOCK_EXCL);
 970	xfs_parent_finish(mp, du.ppargs);
 971	return error;
 972
 973 error_return:
 974	xfs_trans_cancel(tp);
 975	xfs_iunlock(tdp, XFS_ILOCK_EXCL);
 976	xfs_iunlock(sip, XFS_ILOCK_EXCL);
 977 out_parent:
 978	xfs_parent_finish(mp, du.ppargs);
 979 std_return:
 980	if (error == -ENOSPC && nospace_error)
 981		error = nospace_error;
 982	return error;
 983}
 984
 985/* Clear the reflink flag and the cowblocks tag if possible. */
 986static void
 987xfs_itruncate_clear_reflink_flags(
 988	struct xfs_inode	*ip)
 989{
 990	struct xfs_ifork	*dfork;
 991	struct xfs_ifork	*cfork;
 992
 993	if (!xfs_is_reflink_inode(ip))
 994		return;
 995	dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
 996	cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
 997	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
 998		ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
 999	if (cfork->if_bytes == 0)
1000		xfs_inode_clear_cowblocks_tag(ip);
1001}
1002
1003/*
1004 * Free up the underlying blocks past new_size.  The new size must be smaller
1005 * than the current size.  This routine can be used both for the attribute and
1006 * data fork, and does not modify the inode size, which is left to the caller.
1007 *
1008 * The transaction passed to this routine must have made a permanent log
1009 * reservation of at least XFS_ITRUNCATE_LOG_RES.  This routine may commit the
1010 * given transaction and start new ones, so make sure everything involved in
1011 * the transaction is tidy before calling here.  Some transaction will be
1012 * returned to the caller to be committed.  The incoming transaction must
1013 * already include the inode, and both inode locks must be held exclusively.
1014 * The inode must also be "held" within the transaction.  On return the inode
1015 * will be "held" within the returned transaction.  This routine does NOT
1016 * require any disk space to be reserved for it within the transaction.
1017 *
1018 * If we get an error, we must return with the inode locked and linked into the
1019 * current transaction. This keeps things simple for the higher level code,
1020 * because it always knows that the inode is locked and held in the transaction
1021 * that returns to it whether errors occur or not.  We don't mark the inode
1022 * dirty on error so that transactions can be easily aborted if possible.
1023 */
1024int
1025xfs_itruncate_extents_flags(
1026	struct xfs_trans	**tpp,
1027	struct xfs_inode	*ip,
1028	int			whichfork,
1029	xfs_fsize_t		new_size,
1030	int			flags)
1031{
1032	struct xfs_mount	*mp = ip->i_mount;
1033	struct xfs_trans	*tp = *tpp;
1034	xfs_fileoff_t		first_unmap_block;
 
1035	int			error = 0;
1036
1037	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1038	if (atomic_read(&VFS_I(ip)->i_count))
1039		xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1040	ASSERT(new_size <= XFS_ISIZE(ip));
1041	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1042	ASSERT(ip->i_itemp != NULL);
1043	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1044	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1045
1046	trace_xfs_itruncate_extents_start(ip, new_size);
1047
1048	flags |= xfs_bmapi_aflag(whichfork);
1049
1050	/*
1051	 * Since it is possible for space to become allocated beyond
1052	 * the end of the file (in a crash where the space is allocated
1053	 * but the inode size is not yet updated), simply remove any
1054	 * blocks which show up between the new EOF and the maximum
1055	 * possible file size.
1056	 *
1057	 * We have to free all the blocks to the bmbt maximum offset, even if
1058	 * the page cache can't scale that far.
1059	 */
1060	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1061	if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1062		WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1063		return 0;
1064	}
1065
1066	error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1067			XFS_MAX_FILEOFF);
1068	if (error)
1069		goto out;
 
 
 
 
 
 
 
 
 
1070
1071	if (whichfork == XFS_DATA_FORK) {
1072		/* Remove all pending CoW reservations. */
1073		error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1074				first_unmap_block, XFS_MAX_FILEOFF, true);
1075		if (error)
1076			goto out;
1077
1078		xfs_itruncate_clear_reflink_flags(ip);
1079	}
1080
1081	/*
1082	 * Always re-log the inode so that our permanent transaction can keep
1083	 * on rolling it forward in the log.
1084	 */
1085	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1086
1087	trace_xfs_itruncate_extents_end(ip, new_size);
1088
1089out:
1090	*tpp = tp;
1091	return error;
1092}
1093
1094/*
1095 * Mark all the buffers attached to this directory stale.  In theory we should
1096 * never be freeing a directory with any blocks at all, but this covers the
1097 * case where we've recovered a directory swap with a "temporary" directory
1098 * created by online repair and now need to dump it.
1099 */
1100STATIC void
1101xfs_inactive_dir(
1102	struct xfs_inode	*dp)
1103{
1104	struct xfs_iext_cursor	icur;
1105	struct xfs_bmbt_irec	got;
1106	struct xfs_mount	*mp = dp->i_mount;
1107	struct xfs_da_geometry	*geo = mp->m_dir_geo;
1108	struct xfs_ifork	*ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1109	xfs_fileoff_t		off;
1110
1111	/*
1112	 * Invalidate each directory block.  All directory blocks are of
1113	 * fsbcount length and alignment, so we only need to walk those same
1114	 * offsets.  We hold the only reference to this inode, so we must wait
1115	 * for the buffer locks.
1116	 */
1117	for_each_xfs_iext(ifp, &icur, &got) {
1118		for (off = round_up(got.br_startoff, geo->fsbcount);
1119		     off < got.br_startoff + got.br_blockcount;
1120		     off += geo->fsbcount) {
1121			struct xfs_buf	*bp = NULL;
1122			xfs_fsblock_t	fsbno;
1123			int		error;
1124
1125			fsbno = (off - got.br_startoff) + got.br_startblock;
1126			error = xfs_buf_incore(mp->m_ddev_targp,
1127					XFS_FSB_TO_DADDR(mp, fsbno),
1128					XFS_FSB_TO_BB(mp, geo->fsbcount),
1129					XBF_LIVESCAN, &bp);
1130			if (error)
1131				continue;
1132
1133			xfs_buf_stale(bp);
1134			xfs_buf_relse(bp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1135		}
1136	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1137}
1138
1139/*
1140 * xfs_inactive_truncate
1141 *
1142 * Called to perform a truncate when an inode becomes unlinked.
1143 */
1144STATIC int
1145xfs_inactive_truncate(
1146	struct xfs_inode *ip)
1147{
1148	struct xfs_mount	*mp = ip->i_mount;
1149	struct xfs_trans	*tp;
1150	int			error;
1151
1152	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1153	if (error) {
1154		ASSERT(xfs_is_shutdown(mp));
1155		return error;
1156	}
1157	xfs_ilock(ip, XFS_ILOCK_EXCL);
1158	xfs_trans_ijoin(tp, ip, 0);
1159
1160	/*
1161	 * Log the inode size first to prevent stale data exposure in the event
1162	 * of a system crash before the truncate completes. See the related
1163	 * comment in xfs_vn_setattr_size() for details.
1164	 */
1165	ip->i_disk_size = 0;
1166	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1167
1168	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1169	if (error)
1170		goto error_trans_cancel;
1171
1172	ASSERT(ip->i_df.if_nextents == 0);
1173
1174	error = xfs_trans_commit(tp);
1175	if (error)
1176		goto error_unlock;
1177
1178	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1179	return 0;
1180
1181error_trans_cancel:
1182	xfs_trans_cancel(tp);
1183error_unlock:
1184	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1185	return error;
1186}
1187
1188/*
1189 * xfs_inactive_ifree()
1190 *
1191 * Perform the inode free when an inode is unlinked.
1192 */
1193STATIC int
1194xfs_inactive_ifree(
1195	struct xfs_inode *ip)
1196{
1197	struct xfs_mount	*mp = ip->i_mount;
1198	struct xfs_trans	*tp;
1199	int			error;
1200
1201	/*
1202	 * We try to use a per-AG reservation for any block needed by the finobt
1203	 * tree, but as the finobt feature predates the per-AG reservation
1204	 * support a degraded file system might not have enough space for the
1205	 * reservation at mount time.  In that case try to dip into the reserved
1206	 * pool and pray.
1207	 *
1208	 * Send a warning if the reservation does happen to fail, as the inode
1209	 * now remains allocated and sits on the unlinked list until the fs is
1210	 * repaired.
1211	 */
1212	if (unlikely(mp->m_finobt_nores)) {
1213		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1214				XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1215				&tp);
1216	} else {
1217		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1218	}
1219	if (error) {
1220		if (error == -ENOSPC) {
1221			xfs_warn_ratelimited(mp,
1222			"Failed to remove inode(s) from unlinked list. "
1223			"Please free space, unmount and run xfs_repair.");
1224		} else {
1225			ASSERT(xfs_is_shutdown(mp));
1226		}
1227		return error;
1228	}
1229
1230	/*
1231	 * We do not hold the inode locked across the entire rolling transaction
1232	 * here. We only need to hold it for the first transaction that
1233	 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1234	 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1235	 * here breaks the relationship between cluster buffer invalidation and
1236	 * stale inode invalidation on cluster buffer item journal commit
1237	 * completion, and can result in leaving dirty stale inodes hanging
1238	 * around in memory.
1239	 *
1240	 * We have no need for serialising this inode operation against other
1241	 * operations - we freed the inode and hence reallocation is required
1242	 * and that will serialise on reallocating the space the deferops need
1243	 * to free. Hence we can unlock the inode on the first commit of
1244	 * the transaction rather than roll it right through the deferops. This
1245	 * avoids relogging the XFS_ISTALE inode.
1246	 *
1247	 * We check that xfs_ifree() hasn't grown an internal transaction roll
1248	 * by asserting that the inode is still locked when it returns.
1249	 */
1250	xfs_ilock(ip, XFS_ILOCK_EXCL);
1251	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1252
1253	error = xfs_ifree(tp, ip);
1254	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1255	if (error) {
1256		/*
1257		 * If we fail to free the inode, shut down.  The cancel
1258		 * might do that, we need to make sure.  Otherwise the
1259		 * inode might be lost for a long time or forever.
1260		 */
1261		if (!xfs_is_shutdown(mp)) {
1262			xfs_notice(mp, "%s: xfs_ifree returned error %d",
1263				__func__, error);
1264			xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1265		}
1266		xfs_trans_cancel(tp);
1267		return error;
1268	}
1269
1270	/*
1271	 * Credit the quota account(s). The inode is gone.
1272	 */
1273	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1274
1275	return xfs_trans_commit(tp);
1276}
1277
1278/*
1279 * Returns true if we need to update the on-disk metadata before we can free
1280 * the memory used by this inode.  Updates include freeing post-eof
1281 * preallocations; freeing COW staging extents; and marking the inode free in
1282 * the inobt if it is on the unlinked list.
1283 */
1284bool
1285xfs_inode_needs_inactive(
1286	struct xfs_inode	*ip)
1287{
1288	struct xfs_mount	*mp = ip->i_mount;
1289	struct xfs_ifork	*cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1290
1291	/*
1292	 * If the inode is already free, then there can be nothing
1293	 * to clean up here.
1294	 */
1295	if (VFS_I(ip)->i_mode == 0)
1296		return false;
1297
1298	/*
1299	 * If this is a read-only mount, don't do this (would generate I/O)
1300	 * unless we're in log recovery and cleaning the iunlinked list.
1301	 */
1302	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1303		return false;
1304
1305	/* If the log isn't running, push inodes straight to reclaim. */
1306	if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1307		return false;
1308
1309	/* Metadata inodes require explicit resource cleanup. */
1310	if (xfs_is_internal_inode(ip))
1311		return false;
1312
1313	/* Want to clean out the cow blocks if there are any. */
1314	if (cow_ifp && cow_ifp->if_bytes > 0)
1315		return true;
1316
1317	/* Unlinked files must be freed. */
1318	if (VFS_I(ip)->i_nlink == 0)
1319		return true;
1320
1321	/*
1322	 * This file isn't being freed, so check if there are post-eof blocks
1323	 * to free.
1324	 *
1325	 * Note: don't bother with iolock here since lockdep complains about
1326	 * acquiring it in reclaim context. We have the only reference to the
1327	 * inode at this point anyways.
1328	 */
1329	return xfs_can_free_eofblocks(ip);
1330}
1331
1332/*
1333 * Save health status somewhere, if we're dumping an inode with uncorrected
1334 * errors and online repair isn't running.
1335 */
1336static inline void
1337xfs_inactive_health(
1338	struct xfs_inode	*ip)
1339{
1340	struct xfs_mount	*mp = ip->i_mount;
1341	struct xfs_perag	*pag;
1342	unsigned int		sick;
1343	unsigned int		checked;
1344
1345	xfs_inode_measure_sickness(ip, &sick, &checked);
1346	if (!sick)
1347		return;
1348
1349	trace_xfs_inode_unfixed_corruption(ip, sick);
1350
1351	if (sick & XFS_SICK_INO_FORGET)
1352		return;
1353
1354	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1355	if (!pag) {
1356		/* There had better still be a perag structure! */
1357		ASSERT(0);
1358		return;
1359	}
1360
1361	xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1362	xfs_perag_put(pag);
1363}
1364
1365/*
1366 * xfs_inactive
1367 *
1368 * This is called when the vnode reference count for the vnode
1369 * goes to zero.  If the file has been unlinked, then it must
1370 * now be truncated.  Also, we clear all of the read-ahead state
1371 * kept for the inode here since the file is now closed.
1372 */
1373int
1374xfs_inactive(
1375	xfs_inode_t	*ip)
1376{
1377	struct xfs_mount	*mp;
1378	int			error = 0;
1379	int			truncate = 0;
1380
1381	/*
1382	 * If the inode is already free, then there can be nothing
1383	 * to clean up here.
1384	 */
1385	if (VFS_I(ip)->i_mode == 0) {
1386		ASSERT(ip->i_df.if_broot_bytes == 0);
1387		goto out;
1388	}
1389
1390	mp = ip->i_mount;
1391	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1392
1393	xfs_inactive_health(ip);
1394
1395	/*
1396	 * If this is a read-only mount, don't do this (would generate I/O)
1397	 * unless we're in log recovery and cleaning the iunlinked list.
1398	 */
1399	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1400		goto out;
1401
1402	/* Metadata inodes require explicit resource cleanup. */
1403	if (xfs_is_internal_inode(ip))
1404		goto out;
1405
1406	/* Try to clean out the cow blocks if there are any. */
1407	if (xfs_inode_has_cow_data(ip)) {
1408		error = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1409		if (error)
1410			goto out;
1411	}
1412
1413	if (VFS_I(ip)->i_nlink != 0) {
1414		/*
 
 
 
 
1415		 * Note: don't bother with iolock here since lockdep complains
1416		 * about acquiring it in reclaim context. We have the only
1417		 * reference to the inode at this point anyways.
1418		 */
1419		if (xfs_can_free_eofblocks(ip))
1420			error = xfs_free_eofblocks(ip);
1421
1422		goto out;
1423	}
1424
1425	if (S_ISREG(VFS_I(ip)->i_mode) &&
1426	    (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1427	     xfs_inode_has_filedata(ip)))
1428		truncate = 1;
1429
1430	if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1431		/*
1432		 * If this inode is being inactivated during a quotacheck and
1433		 * has not yet been scanned by quotacheck, we /must/ remove
1434		 * the dquots from the inode before inactivation changes the
1435		 * block and inode counts.  Most probably this is a result of
1436		 * reloading the incore iunlinked list to purge unrecovered
1437		 * unlinked inodes.
1438		 */
1439		xfs_qm_dqdetach(ip);
1440	} else {
1441		error = xfs_qm_dqattach(ip);
1442		if (error)
1443			goto out;
1444	}
1445
1446	if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1447		xfs_inactive_dir(ip);
1448		truncate = 1;
1449	}
1450
1451	if (S_ISLNK(VFS_I(ip)->i_mode))
1452		error = xfs_inactive_symlink(ip);
1453	else if (truncate)
1454		error = xfs_inactive_truncate(ip);
1455	if (error)
1456		goto out;
1457
1458	/*
1459	 * If there are attributes associated with the file then blow them away
1460	 * now.  The code calls a routine that recursively deconstructs the
1461	 * attribute fork. If also blows away the in-core attribute fork.
1462	 */
1463	if (xfs_inode_has_attr_fork(ip)) {
1464		error = xfs_attr_inactive(ip);
1465		if (error)
1466			goto out;
1467	}
1468
 
1469	ASSERT(ip->i_forkoff == 0);
1470
1471	/*
1472	 * Free the inode.
1473	 */
1474	error = xfs_inactive_ifree(ip);
1475
1476out:
1477	/*
1478	 * We're done making metadata updates for this inode, so we can release
1479	 * the attached dquots.
1480	 */
1481	xfs_qm_dqdetach(ip);
1482	return error;
1483}
1484
1485/*
1486 * Find an inode on the unlinked list. This does not take references to the
1487 * inode as we have existence guarantees by holding the AGI buffer lock and that
1488 * only unlinked, referenced inodes can be on the unlinked inode list.  If we
1489 * don't find the inode in cache, then let the caller handle the situation.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1490 */
1491struct xfs_inode *
1492xfs_iunlink_lookup(
1493	struct xfs_perag	*pag,
1494	xfs_agino_t		agino)
1495{
1496	struct xfs_inode	*ip;
1497
1498	rcu_read_lock();
1499	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1500	if (!ip) {
1501		/* Caller can handle inode not being in memory. */
1502		rcu_read_unlock();
1503		return NULL;
1504	}
1505
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1506	/*
1507	 * Inode in RCU freeing limbo should not happen.  Warn about this and
1508	 * let the caller handle the failure.
 
 
1509	 */
1510	if (WARN_ON_ONCE(!ip->i_ino)) {
1511		rcu_read_unlock();
1512		return NULL;
1513	}
1514	ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1515	rcu_read_unlock();
1516	return ip;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1517}
1518
1519/*
1520 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1521 * to @prev_agino.  Caller must hold the AGI to synchronize with other changes
1522 * to the unlinked list.
1523 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1524int
1525xfs_iunlink_reload_next(
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1526	struct xfs_trans	*tp,
 
1527	struct xfs_buf		*agibp,
1528	xfs_agino_t		prev_agino,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1529	xfs_agino_t		next_agino)
1530{
1531	struct xfs_perag	*pag = agibp->b_pag;
1532	struct xfs_mount	*mp = pag_mount(pag);
1533	struct xfs_inode	*next_ip = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1534	int			error;
1535
1536	ASSERT(next_agino != NULLAGINO);
1537
1538#ifdef DEBUG
1539	rcu_read_lock();
1540	next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1541	ASSERT(next_ip == NULL);
1542	rcu_read_unlock();
1543#endif
1544
1545	xfs_info_ratelimited(mp,
1546 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating recovery.",
1547			next_agino, pag_agno(pag));
 
 
 
 
 
1548
1549	/*
1550	 * Use an untrusted lookup just to be cautious in case the AGI has been
1551	 * corrupted and now points at a free inode.  That shouldn't happen,
1552	 * but we'd rather shut down now since we're already running in a weird
1553	 * situation.
1554	 */
1555	error = xfs_iget(mp, tp, xfs_agino_to_ino(pag, next_agino),
1556			XFS_IGET_UNTRUSTED, 0, &next_ip);
1557	if (error) {
1558		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1559		return error;
 
 
1560	}
1561
1562	/* If this is not an unlinked inode, something is very wrong. */
1563	if (VFS_I(next_ip)->i_nlink != 0) {
1564		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1565		error = -EFSCORRUPTED;
1566		goto rele;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1567	}
1568
1569	next_ip->i_prev_unlinked = prev_agino;
1570	trace_xfs_iunlink_reload_next(next_ip);
1571rele:
1572	ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
1573	if (xfs_is_quotacheck_running(mp) && next_ip)
1574		xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1575	xfs_irele(next_ip);
1576	return error;
1577}
1578
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1579/*
1580 * Look up the inode number specified and if it is not already marked XFS_ISTALE
1581 * mark it stale. We should only find clean inodes in this lookup that aren't
1582 * already stale.
1583 */
1584static void
1585xfs_ifree_mark_inode_stale(
1586	struct xfs_perag	*pag,
1587	struct xfs_inode	*free_ip,
1588	xfs_ino_t		inum)
1589{
1590	struct xfs_mount	*mp = pag_mount(pag);
1591	struct xfs_inode_log_item *iip;
1592	struct xfs_inode	*ip;
1593
1594retry:
1595	rcu_read_lock();
1596	ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1597
1598	/* Inode not in memory, nothing to do */
1599	if (!ip) {
1600		rcu_read_unlock();
1601		return;
1602	}
1603
1604	/*
1605	 * because this is an RCU protected lookup, we could find a recently
1606	 * freed or even reallocated inode during the lookup. We need to check
1607	 * under the i_flags_lock for a valid inode here. Skip it if it is not
1608	 * valid, the wrong inode or stale.
1609	 */
1610	spin_lock(&ip->i_flags_lock);
1611	if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1612		goto out_iflags_unlock;
1613
1614	/*
1615	 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
1616	 * other inodes that we did not find in the list attached to the buffer
1617	 * and are not already marked stale. If we can't lock it, back off and
1618	 * retry.
1619	 */
1620	if (ip != free_ip) {
1621		if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1622			spin_unlock(&ip->i_flags_lock);
1623			rcu_read_unlock();
1624			delay(1);
1625			goto retry;
1626		}
1627	}
1628	ip->i_flags |= XFS_ISTALE;
1629
1630	/*
1631	 * If the inode is flushing, it is already attached to the buffer.  All
1632	 * we needed to do here is mark the inode stale so buffer IO completion
1633	 * will remove it from the AIL.
1634	 */
1635	iip = ip->i_itemp;
1636	if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1637		ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1638		ASSERT(iip->ili_last_fields);
1639		goto out_iunlock;
1640	}
1641
1642	/*
1643	 * Inodes not attached to the buffer can be released immediately.
1644	 * Everything else has to go through xfs_iflush_abort() on journal
1645	 * commit as the flock synchronises removal of the inode from the
1646	 * cluster buffer against inode reclaim.
1647	 */
1648	if (!iip || list_empty(&iip->ili_item.li_bio_list))
1649		goto out_iunlock;
1650
1651	__xfs_iflags_set(ip, XFS_IFLUSHING);
1652	spin_unlock(&ip->i_flags_lock);
1653	rcu_read_unlock();
1654
1655	/* we have a dirty inode in memory that has not yet been flushed. */
1656	spin_lock(&iip->ili_lock);
1657	iip->ili_last_fields = iip->ili_fields;
1658	iip->ili_fields = 0;
1659	iip->ili_fsync_fields = 0;
1660	spin_unlock(&iip->ili_lock);
1661	ASSERT(iip->ili_last_fields);
1662
1663	if (ip != free_ip)
1664		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1665	return;
1666
1667out_iunlock:
1668	if (ip != free_ip)
1669		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1670out_iflags_unlock:
1671	spin_unlock(&ip->i_flags_lock);
1672	rcu_read_unlock();
1673}
1674
1675/*
1676 * A big issue when freeing the inode cluster is that we _cannot_ skip any
1677 * inodes that are in memory - they all must be marked stale and attached to
1678 * the cluster buffer.
1679 */
1680static int
1681xfs_ifree_cluster(
1682	struct xfs_trans	*tp,
1683	struct xfs_perag	*pag,
1684	struct xfs_inode	*free_ip,
1685	struct xfs_icluster	*xic)
1686{
1687	struct xfs_mount	*mp = free_ip->i_mount;
1688	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1689	struct xfs_buf		*bp;
1690	xfs_daddr_t		blkno;
1691	xfs_ino_t		inum = xic->first_ino;
1692	int			nbufs;
1693	int			i, j;
1694	int			ioffset;
1695	int			error;
1696
1697	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1698
1699	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1700		/*
1701		 * The allocation bitmap tells us which inodes of the chunk were
1702		 * physically allocated. Skip the cluster if an inode falls into
1703		 * a sparse region.
1704		 */
1705		ioffset = inum - xic->first_ino;
1706		if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1707			ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1708			continue;
1709		}
1710
1711		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1712					 XFS_INO_TO_AGBNO(mp, inum));
1713
1714		/*
1715		 * We obtain and lock the backing buffer first in the process
1716		 * here to ensure dirty inodes attached to the buffer remain in
1717		 * the flushing state while we mark them stale.
1718		 *
1719		 * If we scan the in-memory inodes first, then buffer IO can
1720		 * complete before we get a lock on it, and hence we may fail
1721		 * to mark all the active inodes on the buffer stale.
1722		 */
1723		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
1724				mp->m_bsize * igeo->blocks_per_cluster,
1725				XBF_UNMAPPED, &bp);
1726		if (error)
1727			return error;
1728
1729		/*
1730		 * This buffer may not have been correctly initialised as we
1731		 * didn't read it from disk. That's not important because we are
1732		 * only using to mark the buffer as stale in the log, and to
1733		 * attach stale cached inodes on it.
1734		 *
1735		 * For the inode that triggered the cluster freeing, this
1736		 * attachment may occur in xfs_inode_item_precommit() after we
1737		 * have marked this buffer stale.  If this buffer was not in
1738		 * memory before xfs_ifree_cluster() started, it will not be
1739		 * marked XBF_DONE and this will cause problems later in
1740		 * xfs_inode_item_precommit() when we trip over a (stale, !done)
1741		 * buffer to attached to the transaction.
1742		 *
1743		 * Hence we have to mark the buffer as XFS_DONE here. This is
1744		 * safe because we are also marking the buffer as XBF_STALE and
1745		 * XFS_BLI_STALE. That means it will never be dispatched for
1746		 * IO and it won't be unlocked until the cluster freeing has
1747		 * been committed to the journal and the buffer unpinned. If it
1748		 * is written, we want to know about it, and we want it to
1749		 * fail. We can acheive this by adding a write verifier to the
1750		 * buffer.
1751		 */
1752		bp->b_flags |= XBF_DONE;
1753		bp->b_ops = &xfs_inode_buf_ops;
1754
1755		/*
1756		 * Now we need to set all the cached clean inodes as XFS_ISTALE,
1757		 * too. This requires lookups, and will skip inodes that we've
1758		 * already marked XFS_ISTALE.
1759		 */
1760		for (i = 0; i < igeo->inodes_per_cluster; i++)
1761			xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
1762
1763		xfs_trans_stale_inode_buf(tp, bp);
1764		xfs_trans_binval(tp, bp);
1765	}
1766	return 0;
1767}
1768
1769/*
1770 * This is called to return an inode to the inode free list.  The inode should
1771 * already be truncated to 0 length and have no pages associated with it.  This
1772 * routine also assumes that the inode is already a part of the transaction.
1773 *
1774 * The on-disk copy of the inode will have been added to the list of unlinked
1775 * inodes in the AGI. We need to remove the inode from that list atomically with
1776 * respect to freeing it here.
 
1777 */
1778int
1779xfs_ifree(
1780	struct xfs_trans	*tp,
1781	struct xfs_inode	*ip)
1782{
1783	struct xfs_mount	*mp = ip->i_mount;
1784	struct xfs_perag	*pag;
1785	struct xfs_icluster	xic = { 0 };
1786	struct xfs_inode_log_item *iip = ip->i_itemp;
1787	int			error;
1788
1789	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1790	ASSERT(VFS_I(ip)->i_nlink == 0);
1791	ASSERT(ip->i_df.if_nextents == 0);
1792	ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1793	ASSERT(ip->i_nblocks == 0);
1794
1795	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1796
1797	error = xfs_inode_uninit(tp, pag, ip, &xic);
 
 
 
 
 
 
 
1798	if (error)
1799		goto out;
1800
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1801	if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1802		xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1803
1804	/* Don't attempt to replay owner changes for a deleted inode */
1805	spin_lock(&iip->ili_lock);
1806	iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1807	spin_unlock(&iip->ili_lock);
1808
 
 
 
 
 
 
 
1809	if (xic.deleted)
1810		error = xfs_ifree_cluster(tp, pag, ip, &xic);
1811out:
1812	xfs_perag_put(pag);
1813	return error;
1814}
1815
1816/*
1817 * This is called to unpin an inode.  The caller must have the inode locked
1818 * in at least shared mode so that the buffer cannot be subsequently pinned
1819 * once someone is waiting for it to be unpinned.
1820 */
1821static void
1822xfs_iunpin(
1823	struct xfs_inode	*ip)
1824{
1825	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1826
1827	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1828
1829	/* Give the log a push to start the unpinning I/O */
1830	xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
1831
1832}
1833
1834static void
1835__xfs_iunpin_wait(
1836	struct xfs_inode	*ip)
1837{
1838	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
1839	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1840
1841	xfs_iunpin(ip);
1842
1843	do {
1844		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1845		if (xfs_ipincount(ip))
1846			io_schedule();
1847	} while (xfs_ipincount(ip));
1848	finish_wait(wq, &wait.wq_entry);
1849}
1850
1851void
1852xfs_iunpin_wait(
1853	struct xfs_inode	*ip)
1854{
1855	if (xfs_ipincount(ip))
1856		__xfs_iunpin_wait(ip);
1857}
1858
1859/*
1860 * Removing an inode from the namespace involves removing the directory entry
1861 * and dropping the link count on the inode. Removing the directory entry can
1862 * result in locking an AGF (directory blocks were freed) and removing a link
1863 * count can result in placing the inode on an unlinked list which results in
1864 * locking an AGI.
1865 *
1866 * The big problem here is that we have an ordering constraint on AGF and AGI
1867 * locking - inode allocation locks the AGI, then can allocate a new extent for
1868 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1869 * removes the inode from the unlinked list, requiring that we lock the AGI
1870 * first, and then freeing the inode can result in an inode chunk being freed
1871 * and hence freeing disk space requiring that we lock an AGF.
1872 *
1873 * Hence the ordering that is imposed by other parts of the code is AGI before
1874 * AGF. This means we cannot remove the directory entry before we drop the inode
1875 * reference count and put it on the unlinked list as this results in a lock
1876 * order of AGF then AGI, and this can deadlock against inode allocation and
1877 * freeing. Therefore we must drop the link counts before we remove the
1878 * directory entry.
1879 *
1880 * This is still safe from a transactional point of view - it is not until we
1881 * get to xfs_defer_finish() that we have the possibility of multiple
1882 * transactions in this operation. Hence as long as we remove the directory
1883 * entry and drop the link count in the first transaction of the remove
1884 * operation, there are no transactional constraints on the ordering here.
1885 */
1886int
1887xfs_remove(
1888	struct xfs_inode	*dp,
1889	struct xfs_name		*name,
1890	struct xfs_inode	*ip)
1891{
1892	struct xfs_dir_update	du = {
1893		.dp		= dp,
1894		.name		= name,
1895		.ip		= ip,
1896	};
1897	struct xfs_mount	*mp = dp->i_mount;
1898	struct xfs_trans	*tp = NULL;
1899	int			is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1900	int			dontcare;
1901	int                     error = 0;
1902	uint			resblks;
1903
1904	trace_xfs_remove(dp, name);
1905
1906	if (xfs_is_shutdown(mp))
1907		return -EIO;
1908	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1909		return -EIO;
1910
1911	error = xfs_qm_dqattach(dp);
1912	if (error)
1913		goto std_return;
1914
1915	error = xfs_qm_dqattach(ip);
1916	if (error)
1917		goto std_return;
1918
1919	error = xfs_parent_start(mp, &du.ppargs);
1920	if (error)
1921		goto std_return;
1922
1923	/*
1924	 * We try to get the real space reservation first, allowing for
1925	 * directory btree deletion(s) implying possible bmap insert(s).  If we
1926	 * can't get the space reservation then we use 0 instead, and avoid the
1927	 * bmap btree insert(s) in the directory code by, if the bmap insert
1928	 * tries to happen, instead trimming the LAST block from the directory.
1929	 *
1930	 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
1931	 * the directory code can handle a reservationless update and we don't
1932	 * want to prevent a user from trying to free space by deleting things.
1933	 */
1934	resblks = xfs_remove_space_res(mp, name->len);
1935	error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
1936			&tp, &dontcare);
 
 
1937	if (error) {
1938		ASSERT(error != -ENOSPC);
1939		goto out_parent;
1940	}
1941
1942	error = xfs_dir_remove_child(tp, resblks, &du);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1943	if (error)
1944		goto out_trans_cancel;
1945
 
 
 
 
 
 
1946	/*
1947	 * If this is a synchronous mount, make sure that the
1948	 * remove transaction goes to disk before returning to
1949	 * the user.
1950	 */
1951	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1952		xfs_trans_set_sync(tp);
1953
1954	error = xfs_trans_commit(tp);
1955	if (error)
1956		goto out_unlock;
1957
1958	if (is_dir && xfs_inode_is_filestream(ip))
1959		xfs_filestream_deassociate(ip);
1960
1961	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1962	xfs_iunlock(dp, XFS_ILOCK_EXCL);
1963	xfs_parent_finish(mp, du.ppargs);
1964	return 0;
1965
1966 out_trans_cancel:
1967	xfs_trans_cancel(tp);
1968 out_unlock:
1969	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1970	xfs_iunlock(dp, XFS_ILOCK_EXCL);
1971 out_parent:
1972	xfs_parent_finish(mp, du.ppargs);
1973 std_return:
1974	return error;
1975}
1976
1977static inline void
1978xfs_iunlock_rename(
1979	struct xfs_inode	**i_tab,
1980	int			num_inodes)
1981{
1982	int			i;
1983
1984	for (i = num_inodes - 1; i >= 0; i--) {
1985		/* Skip duplicate inodes if src and target dps are the same */
1986		if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
1987			continue;
1988		xfs_iunlock(i_tab[i], XFS_ILOCK_EXCL);
1989	}
1990}
1991
1992/*
1993 * Enter all inodes for a rename transaction into a sorted array.
1994 */
1995#define __XFS_SORT_INODES	5
1996STATIC void
1997xfs_sort_for_rename(
1998	struct xfs_inode	*dp1,	/* in: old (source) directory inode */
1999	struct xfs_inode	*dp2,	/* in: new (target) directory inode */
2000	struct xfs_inode	*ip1,	/* in: inode of old entry */
2001	struct xfs_inode	*ip2,	/* in: inode of new entry */
2002	struct xfs_inode	*wip,	/* in: whiteout inode */
2003	struct xfs_inode	**i_tab,/* out: sorted array of inodes */
2004	int			*num_inodes)  /* in/out: inodes in array */
2005{
2006	int			i;
2007
2008	ASSERT(*num_inodes == __XFS_SORT_INODES);
2009	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2010
2011	/*
2012	 * i_tab contains a list of pointers to inodes.  We initialize
2013	 * the table here & we'll sort it.  We will then use it to
2014	 * order the acquisition of the inode locks.
2015	 *
2016	 * Note that the table may contain duplicates.  e.g., dp1 == dp2.
2017	 */
2018	i = 0;
2019	i_tab[i++] = dp1;
2020	i_tab[i++] = dp2;
2021	i_tab[i++] = ip1;
2022	if (ip2)
2023		i_tab[i++] = ip2;
2024	if (wip)
2025		i_tab[i++] = wip;
2026	*num_inodes = i;
2027
2028	xfs_sort_inodes(i_tab, *num_inodes);
 
 
 
 
 
 
 
 
 
 
 
 
2029}
2030
2031void
2032xfs_sort_inodes(
2033	struct xfs_inode	**i_tab,
2034	unsigned int		num_inodes)
2035{
2036	int			i, j;
 
 
 
 
 
2037
2038	ASSERT(num_inodes <= __XFS_SORT_INODES);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2039
2040	/*
2041	 * Sort the elements via bubble sort.  (Remember, there are at
2042	 * most 5 elements to sort, so this is adequate.)
 
2043	 */
2044	for (i = 0; i < num_inodes; i++) {
2045		for (j = 1; j < num_inodes; j++) {
2046			if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2047				swap(i_tab[j], i_tab[j - 1]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2048		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2049	}
 
 
 
 
 
 
 
 
 
 
 
2050}
2051
2052/*
2053 * xfs_rename_alloc_whiteout()
2054 *
2055 * Return a referenced, unlinked, unlocked inode that can be used as a
2056 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2057 * crash between allocating the inode and linking it into the rename transaction
2058 * recovery will free the inode and we won't leak it.
2059 */
2060static int
2061xfs_rename_alloc_whiteout(
2062	struct mnt_idmap	*idmap,
2063	struct xfs_name		*src_name,
2064	struct xfs_inode	*dp,
2065	struct xfs_inode	**wip)
2066{
2067	struct xfs_icreate_args	args = {
2068		.idmap		= idmap,
2069		.pip		= dp,
2070		.mode		= S_IFCHR | WHITEOUT_MODE,
2071		.flags		= XFS_ICREATE_TMPFILE,
2072	};
2073	struct xfs_inode	*tmpfile;
2074	struct qstr		name;
2075	int			error;
2076
2077	error = xfs_create_tmpfile(&args, &tmpfile);
 
2078	if (error)
2079		return error;
2080
2081	name.name = src_name->name;
2082	name.len = src_name->len;
2083	error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2084	if (error) {
2085		xfs_finish_inode_setup(tmpfile);
2086		xfs_irele(tmpfile);
2087		return error;
2088	}
2089
2090	/*
2091	 * Prepare the tmpfile inode as if it were created through the VFS.
2092	 * Complete the inode setup and flag it as linkable.  nlink is already
2093	 * zero, so we can skip the drop_nlink.
2094	 */
2095	xfs_setup_iops(tmpfile);
2096	xfs_finish_inode_setup(tmpfile);
2097	VFS_I(tmpfile)->i_state |= I_LINKABLE;
2098
2099	*wip = tmpfile;
2100	return 0;
2101}
2102
2103/*
2104 * xfs_rename
2105 */
2106int
2107xfs_rename(
2108	struct mnt_idmap	*idmap,
2109	struct xfs_inode	*src_dp,
2110	struct xfs_name		*src_name,
2111	struct xfs_inode	*src_ip,
2112	struct xfs_inode	*target_dp,
2113	struct xfs_name		*target_name,
2114	struct xfs_inode	*target_ip,
2115	unsigned int		flags)
2116{
2117	struct xfs_dir_update	du_src = {
2118		.dp		= src_dp,
2119		.name		= src_name,
2120		.ip		= src_ip,
2121	};
2122	struct xfs_dir_update	du_tgt = {
2123		.dp		= target_dp,
2124		.name		= target_name,
2125		.ip		= target_ip,
2126	};
2127	struct xfs_dir_update	du_wip = { };
2128	struct xfs_mount	*mp = src_dp->i_mount;
2129	struct xfs_trans	*tp;
 
2130	struct xfs_inode	*inodes[__XFS_SORT_INODES];
2131	int			i;
2132	int			num_inodes = __XFS_SORT_INODES;
2133	bool			new_parent = (src_dp != target_dp);
2134	bool			src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2135	int			spaceres;
2136	bool			retried = false;
2137	int			error, nospace_error = 0;
2138
2139	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2140
2141	if ((flags & RENAME_EXCHANGE) && !target_ip)
2142		return -EINVAL;
2143
2144	/*
2145	 * If we are doing a whiteout operation, allocate the whiteout inode
2146	 * we will be placing at the target and ensure the type is set
2147	 * appropriately.
2148	 */
2149	if (flags & RENAME_WHITEOUT) {
2150		error = xfs_rename_alloc_whiteout(idmap, src_name, target_dp,
2151				&du_wip.ip);
2152		if (error)
2153			return error;
2154
2155		/* setup target dirent info as whiteout */
2156		src_name->type = XFS_DIR3_FT_CHRDEV;
2157	}
2158
2159	xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, du_wip.ip,
2160			inodes, &num_inodes);
2161
2162	error = xfs_parent_start(mp, &du_src.ppargs);
2163	if (error)
2164		goto out_release_wip;
2165
2166	if (du_wip.ip) {
2167		error = xfs_parent_start(mp, &du_wip.ppargs);
2168		if (error)
2169			goto out_src_ppargs;
2170	}
2171
2172	if (target_ip) {
2173		error = xfs_parent_start(mp, &du_tgt.ppargs);
2174		if (error)
2175			goto out_wip_ppargs;
2176	}
2177
2178retry:
2179	nospace_error = 0;
2180	spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2181			target_name->len, du_wip.ip != NULL);
2182	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2183	if (error == -ENOSPC) {
2184		nospace_error = error;
2185		spaceres = 0;
2186		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2187				&tp);
2188	}
2189	if (error)
2190		goto out_tgt_ppargs;
2191
2192	/*
2193	 * We don't allow reservationless renaming when parent pointers are
2194	 * enabled because we can't back out if the xattrs must grow.
2195	 */
2196	if (du_src.ppargs && nospace_error) {
2197		error = nospace_error;
2198		xfs_trans_cancel(tp);
2199		goto out_tgt_ppargs;
2200	}
2201
2202	/*
2203	 * Attach the dquots to the inodes
2204	 */
2205	error = xfs_qm_vop_rename_dqattach(inodes);
2206	if (error) {
2207		xfs_trans_cancel(tp);
2208		goto out_tgt_ppargs;
2209	}
2210
2211	/*
2212	 * Lock all the participating inodes. Depending upon whether
2213	 * the target_name exists in the target directory, and
2214	 * whether the target directory is the same as the source
2215	 * directory, we can lock from 2 to 5 inodes.
2216	 */
2217	xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2218
2219	/*
2220	 * Join all the inodes to the transaction.
 
 
2221	 */
2222	xfs_trans_ijoin(tp, src_dp, 0);
2223	if (new_parent)
2224		xfs_trans_ijoin(tp, target_dp, 0);
2225	xfs_trans_ijoin(tp, src_ip, 0);
2226	if (target_ip)
2227		xfs_trans_ijoin(tp, target_ip, 0);
2228	if (du_wip.ip)
2229		xfs_trans_ijoin(tp, du_wip.ip, 0);
2230
2231	/*
2232	 * If we are using project inheritance, we only allow renames
2233	 * into our tree when the project IDs are the same; else the
2234	 * tree quota mechanism would be circumvented.
2235	 */
2236	if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2237		     target_dp->i_projid != src_ip->i_projid)) {
2238		error = -EXDEV;
2239		goto out_trans_cancel;
2240	}
2241
2242	/* RENAME_EXCHANGE is unique from here on. */
2243	if (flags & RENAME_EXCHANGE) {
2244		error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2245				spaceres);
2246		if (error)
2247			goto out_trans_cancel;
2248		goto out_commit;
2249	}
2250
2251	/*
2252	 * Try to reserve quota to handle an expansion of the target directory.
2253	 * We'll allow the rename to continue in reservationless mode if we hit
2254	 * a space usage constraint.  If we trigger reservationless mode, save
2255	 * the errno if there isn't any free space in the target directory.
2256	 */
2257	if (spaceres != 0) {
2258		error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2259				0, false);
2260		if (error == -EDQUOT || error == -ENOSPC) {
2261			if (!retried) {
2262				xfs_trans_cancel(tp);
2263				xfs_iunlock_rename(inodes, num_inodes);
2264				xfs_blockgc_free_quota(target_dp, 0);
2265				retried = true;
2266				goto retry;
2267			}
2268
2269			nospace_error = error;
2270			spaceres = 0;
2271			error = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2272		}
2273		if (error)
 
 
 
 
 
 
 
 
2274			goto out_trans_cancel;
2275	}
2276
2277	/*
2278	 * We don't allow quotaless renaming when parent pointers are enabled
2279	 * because we can't back out if the xattrs must grow.
2280	 */
2281	if (du_src.ppargs && nospace_error) {
2282		error = nospace_error;
2283		goto out_trans_cancel;
2284	}
2285
2286	/*
2287	 * Lock the AGI buffers we need to handle bumping the nlink of the
2288	 * whiteout inode off the unlinked list and to handle dropping the
2289	 * nlink of the target inode.  Per locking order rules, do this in
2290	 * increasing AG order and before directory block allocation tries to
2291	 * grab AGFs because we grab AGIs before AGFs.
2292	 *
2293	 * The (vfs) caller must ensure that if src is a directory then
2294	 * target_ip is either null or an empty directory.
2295	 */
2296	for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2297		if (inodes[i] == du_wip.ip ||
2298		    (inodes[i] == target_ip &&
2299		     (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2300			struct xfs_perag	*pag;
2301			struct xfs_buf		*bp;
2302
2303			pag = xfs_perag_get(mp,
2304					XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2305			error = xfs_read_agi(pag, tp, 0, &bp);
2306			xfs_perag_put(pag);
2307			if (error)
2308				goto out_trans_cancel;
2309		}
2310	}
2311
2312	error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2313			&du_wip);
2314	if (error)
2315		goto out_trans_cancel;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2316
2317	if (du_wip.ip) {
 
 
 
2318		/*
2319		 * Now we have a real link, clear the "I'm a tmpfile" state
2320		 * flag from the inode so it doesn't accidentally get misused in
2321		 * future.
2322		 */
2323		VFS_I(du_wip.ip)->i_state &= ~I_LINKABLE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2324	}
2325
2326out_commit:
2327	/*
2328	 * If this is a synchronous mount, make sure that the rename
2329	 * transaction goes to disk before returning to the user.
 
 
 
2330	 */
2331	if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2332		xfs_trans_set_sync(tp);
2333
2334	error = xfs_trans_commit(tp);
2335	nospace_error = 0;
2336	goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2337
2338out_trans_cancel:
2339	xfs_trans_cancel(tp);
2340out_unlock:
2341	xfs_iunlock_rename(inodes, num_inodes);
2342out_tgt_ppargs:
2343	xfs_parent_finish(mp, du_tgt.ppargs);
2344out_wip_ppargs:
2345	xfs_parent_finish(mp, du_wip.ppargs);
2346out_src_ppargs:
2347	xfs_parent_finish(mp, du_src.ppargs);
2348out_release_wip:
2349	if (du_wip.ip)
2350		xfs_irele(du_wip.ip);
2351	if (error == -ENOSPC && nospace_error)
2352		error = nospace_error;
2353	return error;
2354}
2355
2356static int
2357xfs_iflush(
2358	struct xfs_inode	*ip,
2359	struct xfs_buf		*bp)
2360{
2361	struct xfs_inode_log_item *iip = ip->i_itemp;
2362	struct xfs_dinode	*dip;
2363	struct xfs_mount	*mp = ip->i_mount;
2364	int			error;
2365
2366	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2367	ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2368	ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2369	       ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2370	ASSERT(iip->ili_item.li_buf == bp);
2371
2372	dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
2373
2374	/*
2375	 * We don't flush the inode if any of the following checks fail, but we
2376	 * do still update the log item and attach to the backing buffer as if
2377	 * the flush happened. This is a formality to facilitate predictable
2378	 * error handling as the caller will shutdown and fail the buffer.
2379	 */
2380	error = -EFSCORRUPTED;
2381	if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
2382			       mp, XFS_ERRTAG_IFLUSH_1)) {
2383		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2384			"%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2385			__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2386		goto flush_out;
2387	}
2388	if (S_ISREG(VFS_I(ip)->i_mode)) {
2389		if (XFS_TEST_ERROR(
2390		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2391		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
2392		    mp, XFS_ERRTAG_IFLUSH_3)) {
2393			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2394				"%s: Bad regular inode %llu, ptr "PTR_FMT,
2395				__func__, ip->i_ino, ip);
2396			goto flush_out;
2397		}
2398	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2399		if (XFS_TEST_ERROR(
2400		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2401		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2402		    ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
2403		    mp, XFS_ERRTAG_IFLUSH_4)) {
2404			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2405				"%s: Bad directory inode %llu, ptr "PTR_FMT,
2406				__func__, ip->i_ino, ip);
2407			goto flush_out;
2408		}
2409	}
2410	if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2411				ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
2412		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2413			"%s: detected corrupt incore inode %llu, "
2414			"total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2415			__func__, ip->i_ino,
2416			ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2417			ip->i_nblocks, ip);
2418		goto flush_out;
2419	}
2420	if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
2421				mp, XFS_ERRTAG_IFLUSH_6)) {
2422		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2423			"%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2424			__func__, ip->i_ino, ip->i_forkoff, ip);
2425		goto flush_out;
2426	}
2427
2428	/*
2429	 * Inode item log recovery for v2 inodes are dependent on the flushiter
2430	 * count for correct sequencing.  We bump the flush iteration count so
2431	 * we can detect flushes which postdate a log record during recovery.
2432	 * This is redundant as we now log every change and hence this can't
2433	 * happen but we need to still do it to ensure backwards compatibility
2434	 * with old kernels that predate logging all inode changes.
2435	 */
2436	if (!xfs_has_v3inodes(mp))
2437		ip->i_flushiter++;
2438
2439	/*
2440	 * If there are inline format data / attr forks attached to this inode,
2441	 * make sure they are not corrupt.
2442	 */
2443	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2444	    xfs_ifork_verify_local_data(ip))
2445		goto flush_out;
2446	if (xfs_inode_has_attr_fork(ip) &&
2447	    ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2448	    xfs_ifork_verify_local_attr(ip))
2449		goto flush_out;
2450
2451	/*
2452	 * Copy the dirty parts of the inode into the on-disk inode.  We always
2453	 * copy out the core of the inode, because if the inode is dirty at all
2454	 * the core must be.
2455	 */
2456	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2457
2458	/* Wrap, we never let the log put out DI_MAX_FLUSH */
2459	if (!xfs_has_v3inodes(mp)) {
2460		if (ip->i_flushiter == DI_MAX_FLUSH)
2461			ip->i_flushiter = 0;
2462	}
2463
2464	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2465	if (xfs_inode_has_attr_fork(ip))
2466		xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2467
2468	/*
2469	 * We've recorded everything logged in the inode, so we'd like to clear
2470	 * the ili_fields bits so we don't log and flush things unnecessarily.
2471	 * However, we can't stop logging all this information until the data
2472	 * we've copied into the disk buffer is written to disk.  If we did we
2473	 * might overwrite the copy of the inode in the log with all the data
2474	 * after re-logging only part of it, and in the face of a crash we
2475	 * wouldn't have all the data we need to recover.
2476	 *
2477	 * What we do is move the bits to the ili_last_fields field.  When
2478	 * logging the inode, these bits are moved back to the ili_fields field.
2479	 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2480	 * we know that the information those bits represent is permanently on
2481	 * disk.  As long as the flush completes before the inode is logged
2482	 * again, then both ili_fields and ili_last_fields will be cleared.
2483	 */
2484	error = 0;
2485flush_out:
2486	spin_lock(&iip->ili_lock);
2487	iip->ili_last_fields = iip->ili_fields;
2488	iip->ili_fields = 0;
2489	iip->ili_fsync_fields = 0;
2490	set_bit(XFS_LI_FLUSHING, &iip->ili_item.li_flags);
2491	spin_unlock(&iip->ili_lock);
2492
2493	/*
2494	 * Store the current LSN of the inode so that we can tell whether the
2495	 * item has moved in the AIL from xfs_buf_inode_iodone().
2496	 */
2497	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2498				&iip->ili_item.li_lsn);
2499
2500	/* generate the checksum. */
2501	xfs_dinode_calc_crc(mp, dip);
2502	if (error)
2503		xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2504	return error;
2505}
2506
2507/*
2508 * Non-blocking flush of dirty inode metadata into the backing buffer.
2509 *
2510 * The caller must have a reference to the inode and hold the cluster buffer
2511 * locked. The function will walk across all the inodes on the cluster buffer it
2512 * can find and lock without blocking, and flush them to the cluster buffer.
2513 *
2514 * On successful flushing of at least one inode, the caller must write out the
2515 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2516 * the caller needs to release the buffer. On failure, the filesystem will be
2517 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2518 * will be returned.
2519 */
2520int
2521xfs_iflush_cluster(
2522	struct xfs_buf		*bp)
2523{
2524	struct xfs_mount	*mp = bp->b_mount;
2525	struct xfs_log_item	*lip, *n;
2526	struct xfs_inode	*ip;
2527	struct xfs_inode_log_item *iip;
2528	int			clcount = 0;
2529	int			error = 0;
2530
2531	/*
2532	 * We must use the safe variant here as on shutdown xfs_iflush_abort()
2533	 * will remove itself from the list.
2534	 */
2535	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2536		iip = (struct xfs_inode_log_item *)lip;
2537		ip = iip->ili_inode;
2538
2539		/*
2540		 * Quick and dirty check to avoid locks if possible.
2541		 */
2542		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2543			continue;
2544		if (xfs_ipincount(ip))
2545			continue;
2546
2547		/*
2548		 * The inode is still attached to the buffer, which means it is
2549		 * dirty but reclaim might try to grab it. Check carefully for
2550		 * that, and grab the ilock while still holding the i_flags_lock
2551		 * to guarantee reclaim will not be able to reclaim this inode
2552		 * once we drop the i_flags_lock.
2553		 */
2554		spin_lock(&ip->i_flags_lock);
2555		ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2556		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2557			spin_unlock(&ip->i_flags_lock);
2558			continue;
2559		}
2560
2561		/*
2562		 * ILOCK will pin the inode against reclaim and prevent
2563		 * concurrent transactions modifying the inode while we are
2564		 * flushing the inode. If we get the lock, set the flushing
2565		 * state before we drop the i_flags_lock.
2566		 */
2567		if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2568			spin_unlock(&ip->i_flags_lock);
2569			continue;
2570		}
2571		__xfs_iflags_set(ip, XFS_IFLUSHING);
2572		spin_unlock(&ip->i_flags_lock);
2573
2574		/*
2575		 * Abort flushing this inode if we are shut down because the
2576		 * inode may not currently be in the AIL. This can occur when
2577		 * log I/O failure unpins the inode without inserting into the
2578		 * AIL, leaving a dirty/unpinned inode attached to the buffer
2579		 * that otherwise looks like it should be flushed.
2580		 */
2581		if (xlog_is_shutdown(mp->m_log)) {
2582			xfs_iunpin_wait(ip);
2583			xfs_iflush_abort(ip);
2584			xfs_iunlock(ip, XFS_ILOCK_SHARED);
2585			error = -EIO;
2586			continue;
2587		}
2588
2589		/* don't block waiting on a log force to unpin dirty inodes */
2590		if (xfs_ipincount(ip)) {
2591			xfs_iflags_clear(ip, XFS_IFLUSHING);
2592			xfs_iunlock(ip, XFS_ILOCK_SHARED);
2593			continue;
2594		}
2595
2596		if (!xfs_inode_clean(ip))
2597			error = xfs_iflush(ip, bp);
2598		else
2599			xfs_iflags_clear(ip, XFS_IFLUSHING);
2600		xfs_iunlock(ip, XFS_ILOCK_SHARED);
2601		if (error)
2602			break;
2603		clcount++;
2604	}
2605
2606	if (error) {
2607		/*
2608		 * Shutdown first so we kill the log before we release this
2609		 * buffer. If it is an INODE_ALLOC buffer and pins the tail
2610		 * of the log, failing it before the _log_ is shut down can
2611		 * result in the log tail being moved forward in the journal
2612		 * on disk because log writes can still be taking place. Hence
2613		 * unpinning the tail will allow the ICREATE intent to be
2614		 * removed from the log an recovery will fail with uninitialised
2615		 * inode cluster buffers.
2616		 */
2617		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2618		bp->b_flags |= XBF_ASYNC;
2619		xfs_buf_ioend_fail(bp);
 
2620		return error;
2621	}
2622
2623	if (!clcount)
2624		return -EAGAIN;
2625
2626	XFS_STATS_INC(mp, xs_icluster_flushcnt);
2627	XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2628	return 0;
2629
2630}
2631
2632/* Release an inode. */
2633void
2634xfs_irele(
2635	struct xfs_inode	*ip)
2636{
2637	trace_xfs_irele(ip, _RET_IP_);
2638	iput(VFS_I(ip));
2639}
2640
2641/*
2642 * Ensure all commited transactions touching the inode are written to the log.
2643 */
2644int
2645xfs_log_force_inode(
2646	struct xfs_inode	*ip)
2647{
2648	xfs_csn_t		seq = 0;
2649
2650	xfs_ilock(ip, XFS_ILOCK_SHARED);
2651	if (xfs_ipincount(ip))
2652		seq = ip->i_itemp->ili_commit_seq;
2653	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2654
2655	if (!seq)
2656		return 0;
2657	return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2658}
2659
2660/*
2661 * Grab the exclusive iolock for a data copy from src to dest, making sure to
2662 * abide vfs locking order (lowest pointer value goes first) and breaking the
2663 * layout leases before proceeding.  The loop is needed because we cannot call
2664 * the blocking break_layout() with the iolocks held, and therefore have to
2665 * back out both locks.
2666 */
2667static int
2668xfs_iolock_two_inodes_and_break_layout(
2669	struct inode		*src,
2670	struct inode		*dest)
2671{
2672	int			error;
2673
2674	if (src > dest)
2675		swap(src, dest);
2676
2677retry:
2678	/* Wait to break both inodes' layouts before we start locking. */
2679	error = break_layout(src, true);
2680	if (error)
2681		return error;
2682	if (src != dest) {
2683		error = break_layout(dest, true);
2684		if (error)
2685			return error;
2686	}
2687
2688	/* Lock one inode and make sure nobody got in and leased it. */
2689	inode_lock(src);
2690	error = break_layout(src, false);
2691	if (error) {
2692		inode_unlock(src);
2693		if (error == -EWOULDBLOCK)
2694			goto retry;
2695		return error;
2696	}
2697
2698	if (src == dest)
2699		return 0;
2700
2701	/* Lock the other inode and make sure nobody got in and leased it. */
2702	inode_lock_nested(dest, I_MUTEX_NONDIR2);
2703	error = break_layout(dest, false);
2704	if (error) {
2705		inode_unlock(src);
2706		inode_unlock(dest);
2707		if (error == -EWOULDBLOCK)
2708			goto retry;
2709		return error;
2710	}
2711
2712	return 0;
2713}
2714
2715static int
2716xfs_mmaplock_two_inodes_and_break_dax_layout(
2717	struct xfs_inode	*ip1,
2718	struct xfs_inode	*ip2)
2719{
2720	int			error;
2721	bool			retry;
2722	struct page		*page;
2723
2724	if (ip1->i_ino > ip2->i_ino)
2725		swap(ip1, ip2);
2726
2727again:
2728	retry = false;
2729	/* Lock the first inode */
2730	xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
2731	error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
2732	if (error || retry) {
2733		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2734		if (error == 0 && retry)
2735			goto again;
2736		return error;
2737	}
2738
2739	if (ip1 == ip2)
2740		return 0;
2741
2742	/* Nested lock the second inode */
2743	xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
2744	/*
2745	 * We cannot use xfs_break_dax_layouts() directly here because it may
2746	 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2747	 * for this nested lock case.
2748	 */
2749	page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
2750	if (page && page_ref_count(page) != 1) {
2751		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2752		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2753		goto again;
2754	}
2755
2756	return 0;
2757}
2758
2759/*
2760 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2761 * mmap activity.
2762 */
2763int
2764xfs_ilock2_io_mmap(
2765	struct xfs_inode	*ip1,
2766	struct xfs_inode	*ip2)
2767{
2768	int			ret;
2769
2770	ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
2771	if (ret)
2772		return ret;
2773
2774	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2775		ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2776		if (ret) {
2777			inode_unlock(VFS_I(ip2));
2778			if (ip1 != ip2)
2779				inode_unlock(VFS_I(ip1));
2780			return ret;
2781		}
2782	} else
2783		filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
2784					    VFS_I(ip2)->i_mapping);
2785
2786	return 0;
2787}
2788
2789/* Unlock both inodes to allow IO and mmap activity. */
2790void
2791xfs_iunlock2_io_mmap(
2792	struct xfs_inode	*ip1,
2793	struct xfs_inode	*ip2)
2794{
2795	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2796		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2797		if (ip1 != ip2)
2798			xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2799	} else
2800		filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
2801					      VFS_I(ip2)->i_mapping);
2802
2803	inode_unlock(VFS_I(ip2));
2804	if (ip1 != ip2)
2805		inode_unlock(VFS_I(ip1));
2806}
2807
2808/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2809void
2810xfs_iunlock2_remapping(
2811	struct xfs_inode	*ip1,
2812	struct xfs_inode	*ip2)
2813{
2814	xfs_iflags_clear(ip1, XFS_IREMAPPING);
2815
2816	if (ip1 != ip2)
2817		xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
2818	xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2819
2820	if (ip1 != ip2)
2821		inode_unlock_shared(VFS_I(ip1));
2822	inode_unlock(VFS_I(ip2));
2823}
2824
2825/*
2826 * Reload the incore inode list for this inode.  Caller should ensure that
2827 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
2828 * preventing other threads from executing.
2829 */
2830int
2831xfs_inode_reload_unlinked_bucket(
2832	struct xfs_trans	*tp,
2833	struct xfs_inode	*ip)
2834{
2835	struct xfs_mount	*mp = tp->t_mountp;
2836	struct xfs_buf		*agibp;
2837	struct xfs_agi		*agi;
2838	struct xfs_perag	*pag;
2839	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2840	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2841	xfs_agino_t		prev_agino, next_agino;
2842	unsigned int		bucket;
2843	bool			foundit = false;
2844	int			error;
2845
2846	/* Grab the first inode in the list */
2847	pag = xfs_perag_get(mp, agno);
2848	error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2849	xfs_perag_put(pag);
2850	if (error)
2851		return error;
2852
2853	/*
2854	 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2855	 * incore unlinked list pointers for this inode.  Check once more to
2856	 * see if we raced with anyone else to reload the unlinked list.
2857	 */
2858	if (!xfs_inode_unlinked_incomplete(ip)) {
2859		foundit = true;
2860		goto out_agibp;
2861	}
2862
2863	bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2864	agi = agibp->b_addr;
2865
2866	trace_xfs_inode_reload_unlinked_bucket(ip);
2867
2868	xfs_info_ratelimited(mp,
2869 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating list recovery.",
2870			agino, agno);
2871
2872	prev_agino = NULLAGINO;
2873	next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2874	while (next_agino != NULLAGINO) {
2875		struct xfs_inode	*next_ip = NULL;
2876
2877		/* Found this caller's inode, set its backlink. */
2878		if (next_agino == agino) {
2879			next_ip = ip;
2880			next_ip->i_prev_unlinked = prev_agino;
2881			foundit = true;
2882			goto next_inode;
2883		}
2884
2885		/* Try in-memory lookup first. */
2886		next_ip = xfs_iunlink_lookup(pag, next_agino);
2887		if (next_ip)
2888			goto next_inode;
2889
2890		/* Inode not in memory, try reloading it. */
2891		error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2892				next_agino);
2893		if (error)
2894			break;
2895
2896		/* Grab the reloaded inode. */
2897		next_ip = xfs_iunlink_lookup(pag, next_agino);
2898		if (!next_ip) {
2899			/* No incore inode at all?  We reloaded it... */
2900			ASSERT(next_ip != NULL);
2901			error = -EFSCORRUPTED;
2902			break;
2903		}
2904
2905next_inode:
2906		prev_agino = next_agino;
2907		next_agino = next_ip->i_next_unlinked;
2908	}
2909
2910out_agibp:
2911	xfs_trans_brelse(tp, agibp);
2912	/* Should have found this inode somewhere in the iunlinked bucket. */
2913	if (!error && !foundit)
2914		error = -EFSCORRUPTED;
2915	return error;
2916}
2917
2918/* Decide if this inode is missing its unlinked list and reload it. */
2919int
2920xfs_inode_reload_unlinked(
2921	struct xfs_inode	*ip)
2922{
2923	struct xfs_trans	*tp;
2924	int			error;
2925
2926	error = xfs_trans_alloc_empty(ip->i_mount, &tp);
2927	if (error)
2928		return error;
2929
2930	xfs_ilock(ip, XFS_ILOCK_SHARED);
2931	if (xfs_inode_unlinked_incomplete(ip))
2932		error = xfs_inode_reload_unlinked_bucket(tp, ip);
2933	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2934	xfs_trans_cancel(tp);
2935
2936	return error;
2937}
2938
2939/* Has this inode fork been zapped by repair? */
2940bool
2941xfs_ifork_zapped(
2942	const struct xfs_inode	*ip,
2943	int			whichfork)
2944{
2945	unsigned int		datamask = 0;
2946
2947	switch (whichfork) {
2948	case XFS_DATA_FORK:
2949		switch (ip->i_vnode.i_mode & S_IFMT) {
2950		case S_IFDIR:
2951			datamask = XFS_SICK_INO_DIR_ZAPPED;
2952			break;
2953		case S_IFLNK:
2954			datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
2955			break;
2956		}
2957		return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
2958	case XFS_ATTR_FORK:
2959		return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
2960	default:
2961		return false;
2962	}
2963}
2964
2965/* Compute the number of data and realtime blocks used by a file. */
2966void
2967xfs_inode_count_blocks(
2968	struct xfs_trans	*tp,
2969	struct xfs_inode	*ip,
2970	xfs_filblks_t		*dblocks,
2971	xfs_filblks_t		*rblocks)
2972{
2973	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
2974
2975	*rblocks = 0;
2976	if (XFS_IS_REALTIME_INODE(ip))
2977		xfs_bmap_count_leaves(ifp, rblocks);
2978	*dblocks = ip->i_nblocks - *rblocks;
2979}
2980
2981static void
2982xfs_wait_dax_page(
2983	struct inode		*inode)
2984{
2985	struct xfs_inode        *ip = XFS_I(inode);
2986
2987	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
2988	schedule();
2989	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
2990}
2991
2992int
2993xfs_break_dax_layouts(
2994	struct inode		*inode,
2995	bool			*retry)
2996{
2997	struct page		*page;
2998
2999	xfs_assert_ilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL);
3000
3001	page = dax_layout_busy_page(inode->i_mapping);
3002	if (!page)
3003		return 0;
3004
3005	*retry = true;
3006	return ___wait_var_event(&page->_refcount,
3007			atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
3008			0, 0, xfs_wait_dax_page(inode));
3009}
3010
3011int
3012xfs_break_layouts(
3013	struct inode		*inode,
3014	uint			*iolock,
3015	enum layout_break_reason reason)
3016{
3017	bool			retry;
3018	int			error;
3019
3020	xfs_assert_ilocked(XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3021
3022	do {
3023		retry = false;
3024		switch (reason) {
3025		case BREAK_UNMAP:
3026			error = xfs_break_dax_layouts(inode, &retry);
3027			if (error || retry)
3028				break;
3029			fallthrough;
3030		case BREAK_WRITE:
3031			error = xfs_break_leased_layouts(inode, iolock, &retry);
3032			break;
3033		default:
3034			WARN_ON_ONCE(1);
3035			error = -EINVAL;
3036		}
3037	} while (error == 0 && retry);
3038
3039	return error;
3040}
3041
3042/* Returns the size of fundamental allocation unit for a file, in bytes. */
3043unsigned int
3044xfs_inode_alloc_unitsize(
3045	struct xfs_inode	*ip)
3046{
3047	unsigned int		blocks = 1;
3048
3049	if (XFS_IS_REALTIME_INODE(ip))
3050		blocks = ip->i_mount->m_sb.sb_rextsize;
3051
3052	return XFS_FSB_TO_B(ip->i_mount, blocks);
3053}
3054
3055/* Should we always be using copy on write for file writes? */
3056bool
3057xfs_is_always_cow_inode(
3058	const struct xfs_inode	*ip)
3059{
3060	return ip->i_mount->m_always_cow && xfs_has_reflink(ip->i_mount);
3061}
v5.14.15
   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_trans_space.h"
  20#include "xfs_trans.h"
  21#include "xfs_buf_item.h"
  22#include "xfs_inode_item.h"
 
  23#include "xfs_ialloc.h"
  24#include "xfs_bmap.h"
  25#include "xfs_bmap_util.h"
  26#include "xfs_errortag.h"
  27#include "xfs_error.h"
  28#include "xfs_quota.h"
  29#include "xfs_filestream.h"
  30#include "xfs_trace.h"
  31#include "xfs_icache.h"
  32#include "xfs_symlink.h"
  33#include "xfs_trans_priv.h"
  34#include "xfs_log.h"
  35#include "xfs_bmap_btree.h"
  36#include "xfs_reflink.h"
  37#include "xfs_ag.h"
 
 
 
 
 
 
 
  38
  39kmem_zone_t *xfs_inode_zone;
  40
  41/*
  42 * Used in xfs_itruncate_extents().  This is the maximum number of extents
  43 * freed from a file in a single transaction.
  44 */
  45#define	XFS_ITRUNC_MAX_EXTENTS	2
  46
  47STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
  48STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
  49	struct xfs_inode *);
  50
  51/*
  52 * helper function to extract extent size hint from inode
  53 */
  54xfs_extlen_t
  55xfs_get_extsz_hint(
  56	struct xfs_inode	*ip)
  57{
  58	/*
  59	 * No point in aligning allocations if we need to COW to actually
  60	 * write to them.
  61	 */
  62	if (xfs_is_always_cow_inode(ip))
  63		return 0;
  64	if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
  65		return ip->i_extsize;
  66	if (XFS_IS_REALTIME_INODE(ip))
  67		return ip->i_mount->m_sb.sb_rextsize;
  68	return 0;
  69}
  70
  71/*
  72 * Helper function to extract CoW extent size hint from inode.
  73 * Between the extent size hint and the CoW extent size hint, we
  74 * return the greater of the two.  If the value is zero (automatic),
  75 * use the default size.
  76 */
  77xfs_extlen_t
  78xfs_get_cowextsz_hint(
  79	struct xfs_inode	*ip)
  80{
  81	xfs_extlen_t		a, b;
  82
  83	a = 0;
  84	if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
  85		a = ip->i_cowextsize;
  86	b = xfs_get_extsz_hint(ip);
  87
  88	a = max(a, b);
  89	if (a == 0)
  90		return XFS_DEFAULT_COWEXTSZ_HINT;
  91	return a;
  92}
  93
  94/*
  95 * These two are wrapper routines around the xfs_ilock() routine used to
  96 * centralize some grungy code.  They are used in places that wish to lock the
  97 * inode solely for reading the extents.  The reason these places can't just
  98 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
  99 * bringing in of the extents from disk for a file in b-tree format.  If the
 100 * inode is in b-tree format, then we need to lock the inode exclusively until
 101 * the extents are read in.  Locking it exclusively all the time would limit
 102 * our parallelism unnecessarily, though.  What we do instead is check to see
 103 * if the extents have been read in yet, and only lock the inode exclusively
 104 * if they have not.
 105 *
 106 * The functions return a value which should be given to the corresponding
 107 * xfs_iunlock() call.
 108 */
 109uint
 110xfs_ilock_data_map_shared(
 111	struct xfs_inode	*ip)
 112{
 113	uint			lock_mode = XFS_ILOCK_SHARED;
 114
 115	if (xfs_need_iread_extents(&ip->i_df))
 116		lock_mode = XFS_ILOCK_EXCL;
 117	xfs_ilock(ip, lock_mode);
 118	return lock_mode;
 119}
 120
 121uint
 122xfs_ilock_attr_map_shared(
 123	struct xfs_inode	*ip)
 124{
 125	uint			lock_mode = XFS_ILOCK_SHARED;
 126
 127	if (ip->i_afp && xfs_need_iread_extents(ip->i_afp))
 128		lock_mode = XFS_ILOCK_EXCL;
 129	xfs_ilock(ip, lock_mode);
 130	return lock_mode;
 131}
 132
 133/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 134 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
 135 * multi-reader locks: i_mmap_lock and the i_lock.  This routine allows
 136 * various combinations of the locks to be obtained.
 137 *
 138 * The 3 locks should always be ordered so that the IO lock is obtained first,
 139 * the mmap lock second and the ilock last in order to prevent deadlock.
 140 *
 141 * Basic locking order:
 142 *
 143 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
 144 *
 145 * mmap_lock locking order:
 146 *
 147 * i_rwsem -> page lock -> mmap_lock
 148 * mmap_lock -> i_mmap_lock -> page_lock
 149 *
 150 * The difference in mmap_lock locking order mean that we cannot hold the
 151 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
 152 * fault in pages during copy in/out (for buffered IO) or require the mmap_lock
 153 * in get_user_pages() to map the user pages into the kernel address space for
 154 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
 155 * page faults already hold the mmap_lock.
 156 *
 157 * Hence to serialise fully against both syscall and mmap based IO, we need to
 158 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
 159 * taken in places where we need to invalidate the page cache in a race
 160 * free manner (e.g. truncate, hole punch and other extent manipulation
 161 * functions).
 162 */
 163void
 164xfs_ilock(
 165	xfs_inode_t		*ip,
 166	uint			lock_flags)
 167{
 168	trace_xfs_ilock(ip, lock_flags, _RET_IP_);
 169
 170	/*
 171	 * You can't set both SHARED and EXCL for the same lock,
 172	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 173	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 174	 */
 175	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 176	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 177	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 178	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 179	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 180	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 181	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 182
 183	if (lock_flags & XFS_IOLOCK_EXCL) {
 184		down_write_nested(&VFS_I(ip)->i_rwsem,
 185				  XFS_IOLOCK_DEP(lock_flags));
 186	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 187		down_read_nested(&VFS_I(ip)->i_rwsem,
 188				 XFS_IOLOCK_DEP(lock_flags));
 189	}
 190
 191	if (lock_flags & XFS_MMAPLOCK_EXCL)
 192		mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
 193	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 194		mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
 
 
 
 195
 196	if (lock_flags & XFS_ILOCK_EXCL)
 197		mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 198	else if (lock_flags & XFS_ILOCK_SHARED)
 199		mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 200}
 201
 202/*
 203 * This is just like xfs_ilock(), except that the caller
 204 * is guaranteed not to sleep.  It returns 1 if it gets
 205 * the requested locks and 0 otherwise.  If the IO lock is
 206 * obtained but the inode lock cannot be, then the IO lock
 207 * is dropped before returning.
 208 *
 209 * ip -- the inode being locked
 210 * lock_flags -- this parameter indicates the inode's locks to be
 211 *       to be locked.  See the comment for xfs_ilock() for a list
 212 *	 of valid values.
 213 */
 214int
 215xfs_ilock_nowait(
 216	xfs_inode_t		*ip,
 217	uint			lock_flags)
 218{
 219	trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
 220
 221	/*
 222	 * You can't set both SHARED and EXCL for the same lock,
 223	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 224	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 225	 */
 226	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 227	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 228	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 229	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 230	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 231	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 232	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 233
 234	if (lock_flags & XFS_IOLOCK_EXCL) {
 235		if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
 236			goto out;
 237	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 238		if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
 239			goto out;
 240	}
 241
 242	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 243		if (!mrtryupdate(&ip->i_mmaplock))
 244			goto out_undo_iolock;
 245	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 246		if (!mrtryaccess(&ip->i_mmaplock))
 247			goto out_undo_iolock;
 248	}
 249
 250	if (lock_flags & XFS_ILOCK_EXCL) {
 251		if (!mrtryupdate(&ip->i_lock))
 252			goto out_undo_mmaplock;
 253	} else if (lock_flags & XFS_ILOCK_SHARED) {
 254		if (!mrtryaccess(&ip->i_lock))
 255			goto out_undo_mmaplock;
 256	}
 257	return 1;
 258
 259out_undo_mmaplock:
 260	if (lock_flags & XFS_MMAPLOCK_EXCL)
 261		mrunlock_excl(&ip->i_mmaplock);
 262	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 263		mrunlock_shared(&ip->i_mmaplock);
 264out_undo_iolock:
 265	if (lock_flags & XFS_IOLOCK_EXCL)
 266		up_write(&VFS_I(ip)->i_rwsem);
 267	else if (lock_flags & XFS_IOLOCK_SHARED)
 268		up_read(&VFS_I(ip)->i_rwsem);
 269out:
 270	return 0;
 271}
 272
 273/*
 274 * xfs_iunlock() is used to drop the inode locks acquired with
 275 * xfs_ilock() and xfs_ilock_nowait().  The caller must pass
 276 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
 277 * that we know which locks to drop.
 278 *
 279 * ip -- the inode being unlocked
 280 * lock_flags -- this parameter indicates the inode's locks to be
 281 *       to be unlocked.  See the comment for xfs_ilock() for a list
 282 *	 of valid values for this parameter.
 283 *
 284 */
 285void
 286xfs_iunlock(
 287	xfs_inode_t		*ip,
 288	uint			lock_flags)
 289{
 290	/*
 291	 * You can't set both SHARED and EXCL for the same lock,
 292	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 293	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 294	 */
 295	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 296	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 297	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 298	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 299	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 300	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 301	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 302	ASSERT(lock_flags != 0);
 303
 304	if (lock_flags & XFS_IOLOCK_EXCL)
 305		up_write(&VFS_I(ip)->i_rwsem);
 306	else if (lock_flags & XFS_IOLOCK_SHARED)
 307		up_read(&VFS_I(ip)->i_rwsem);
 308
 309	if (lock_flags & XFS_MMAPLOCK_EXCL)
 310		mrunlock_excl(&ip->i_mmaplock);
 311	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 312		mrunlock_shared(&ip->i_mmaplock);
 313
 314	if (lock_flags & XFS_ILOCK_EXCL)
 315		mrunlock_excl(&ip->i_lock);
 316	else if (lock_flags & XFS_ILOCK_SHARED)
 317		mrunlock_shared(&ip->i_lock);
 318
 319	trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
 320}
 321
 322/*
 323 * give up write locks.  the i/o lock cannot be held nested
 324 * if it is being demoted.
 325 */
 326void
 327xfs_ilock_demote(
 328	xfs_inode_t		*ip,
 329	uint			lock_flags)
 330{
 331	ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
 332	ASSERT((lock_flags &
 333		~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
 334
 335	if (lock_flags & XFS_ILOCK_EXCL)
 336		mrdemote(&ip->i_lock);
 337	if (lock_flags & XFS_MMAPLOCK_EXCL)
 338		mrdemote(&ip->i_mmaplock);
 339	if (lock_flags & XFS_IOLOCK_EXCL)
 340		downgrade_write(&VFS_I(ip)->i_rwsem);
 341
 342	trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
 343}
 344
 345#if defined(DEBUG) || defined(XFS_WARN)
 346int
 347xfs_isilocked(
 348	xfs_inode_t		*ip,
 349	uint			lock_flags)
 350{
 351	if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
 352		if (!(lock_flags & XFS_ILOCK_SHARED))
 353			return !!ip->i_lock.mr_writer;
 354		return rwsem_is_locked(&ip->i_lock.mr_lock);
 355	}
 356
 357	if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
 358		if (!(lock_flags & XFS_MMAPLOCK_SHARED))
 359			return !!ip->i_mmaplock.mr_writer;
 360		return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
 361	}
 362
 363	if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
 364		if (!(lock_flags & XFS_IOLOCK_SHARED))
 365			return !debug_locks ||
 366				lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
 367		return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
 368	}
 369
 370	ASSERT(0);
 371	return 0;
 372}
 373#endif
 374
 375/*
 376 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
 377 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
 378 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
 379 * errors and warnings.
 380 */
 381#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
 382static bool
 383xfs_lockdep_subclass_ok(
 384	int subclass)
 385{
 386	return subclass < MAX_LOCKDEP_SUBCLASSES;
 387}
 388#else
 389#define xfs_lockdep_subclass_ok(subclass)	(true)
 390#endif
 391
 392/*
 393 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
 394 * value. This can be called for any type of inode lock combination, including
 395 * parent locking. Care must be taken to ensure we don't overrun the subclass
 396 * storage fields in the class mask we build.
 397 */
 398static inline int
 399xfs_lock_inumorder(int lock_mode, int subclass)
 
 
 400{
 401	int	class = 0;
 402
 403	ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
 404			      XFS_ILOCK_RTSUM)));
 405	ASSERT(xfs_lockdep_subclass_ok(subclass));
 406
 407	if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
 408		ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
 409		class += subclass << XFS_IOLOCK_SHIFT;
 410	}
 411
 412	if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
 413		ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
 414		class += subclass << XFS_MMAPLOCK_SHIFT;
 415	}
 416
 417	if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
 418		ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
 419		class += subclass << XFS_ILOCK_SHIFT;
 420	}
 421
 422	return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
 423}
 424
 425/*
 426 * The following routine will lock n inodes in exclusive mode.  We assume the
 427 * caller calls us with the inodes in i_ino order.
 428 *
 429 * We need to detect deadlock where an inode that we lock is in the AIL and we
 430 * start waiting for another inode that is locked by a thread in a long running
 431 * transaction (such as truncate). This can result in deadlock since the long
 432 * running trans might need to wait for the inode we just locked in order to
 433 * push the tail and free space in the log.
 434 *
 435 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
 436 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
 437 * lock more than one at a time, lockdep will report false positives saying we
 438 * have violated locking orders.
 439 */
 440static void
 441xfs_lock_inodes(
 442	struct xfs_inode	**ips,
 443	int			inodes,
 444	uint			lock_mode)
 445{
 446	int			attempts = 0, i, j, try_lock;
 
 
 
 447	struct xfs_log_item	*lp;
 448
 449	/*
 450	 * Currently supports between 2 and 5 inodes with exclusive locking.  We
 451	 * support an arbitrary depth of locking here, but absolute limits on
 452	 * inodes depend on the type of locking and the limits placed by
 453	 * lockdep annotations in xfs_lock_inumorder.  These are all checked by
 454	 * the asserts.
 455	 */
 456	ASSERT(ips && inodes >= 2 && inodes <= 5);
 457	ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
 458			    XFS_ILOCK_EXCL));
 459	ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
 460			      XFS_ILOCK_SHARED)));
 461	ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
 462		inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
 463	ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
 464		inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
 465
 466	if (lock_mode & XFS_IOLOCK_EXCL) {
 467		ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
 468	} else if (lock_mode & XFS_MMAPLOCK_EXCL)
 469		ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
 470
 471	try_lock = 0;
 
 472	i = 0;
 473again:
 474	for (; i < inodes; i++) {
 475		ASSERT(ips[i]);
 476
 477		if (i && (ips[i] == ips[i - 1]))	/* Already locked */
 478			continue;
 479
 480		/*
 481		 * If try_lock is not set yet, make sure all locked inodes are
 482		 * not in the AIL.  If any are, set try_lock to be used later.
 483		 */
 484		if (!try_lock) {
 485			for (j = (i - 1); j >= 0 && !try_lock; j--) {
 486				lp = &ips[j]->i_itemp->ili_item;
 487				if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
 488					try_lock++;
 489			}
 490		}
 491
 492		/*
 493		 * If any of the previous locks we have locked is in the AIL,
 494		 * we must TRY to get the second and subsequent locks. If
 495		 * we can't get any, we must release all we have
 496		 * and try again.
 497		 */
 498		if (!try_lock) {
 499			xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
 500			continue;
 501		}
 502
 503		/* try_lock means we have an inode locked that is in the AIL. */
 504		ASSERT(i != 0);
 505		if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
 506			continue;
 507
 508		/*
 509		 * Unlock all previous guys and try again.  xfs_iunlock will try
 510		 * to push the tail if the inode is in the AIL.
 511		 */
 512		attempts++;
 513		for (j = i - 1; j >= 0; j--) {
 514			/*
 515			 * Check to see if we've already unlocked this one.  Not
 516			 * the first one going back, and the inode ptr is the
 517			 * same.
 518			 */
 519			if (j != (i - 1) && ips[j] == ips[j + 1])
 520				continue;
 521
 522			xfs_iunlock(ips[j], lock_mode);
 523		}
 524
 525		if ((attempts % 5) == 0) {
 526			delay(1); /* Don't just spin the CPU */
 527		}
 528		i = 0;
 529		try_lock = 0;
 530		goto again;
 531	}
 532}
 533
 534/*
 535 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
 536 * the mmaplock or the ilock, but not more than one type at a time. If we lock
 537 * more than one at a time, lockdep will report false positives saying we have
 538 * violated locking orders.  The iolock must be double-locked separately since
 539 * we use i_rwsem for that.  We now support taking one lock EXCL and the other
 540 * SHARED.
 541 */
 542void
 543xfs_lock_two_inodes(
 544	struct xfs_inode	*ip0,
 545	uint			ip0_mode,
 546	struct xfs_inode	*ip1,
 547	uint			ip1_mode)
 548{
 549	struct xfs_inode	*temp;
 550	uint			mode_temp;
 551	int			attempts = 0;
 552	struct xfs_log_item	*lp;
 553
 554	ASSERT(hweight32(ip0_mode) == 1);
 555	ASSERT(hweight32(ip1_mode) == 1);
 556	ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 557	ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 558	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 559	       !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 560	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 561	       !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 562	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 563	       !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 564	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 565	       !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 566
 567	ASSERT(ip0->i_ino != ip1->i_ino);
 568
 569	if (ip0->i_ino > ip1->i_ino) {
 570		temp = ip0;
 571		ip0 = ip1;
 572		ip1 = temp;
 573		mode_temp = ip0_mode;
 574		ip0_mode = ip1_mode;
 575		ip1_mode = mode_temp;
 576	}
 577
 578 again:
 579	xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
 580
 581	/*
 582	 * If the first lock we have locked is in the AIL, we must TRY to get
 583	 * the second lock. If we can't get it, we must release the first one
 584	 * and try again.
 585	 */
 586	lp = &ip0->i_itemp->ili_item;
 587	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
 588		if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
 589			xfs_iunlock(ip0, ip0_mode);
 590			if ((++attempts % 5) == 0)
 591				delay(1); /* Don't just spin the CPU */
 592			goto again;
 593		}
 594	} else {
 595		xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
 596	}
 597}
 598
 599uint
 600xfs_ip2xflags(
 601	struct xfs_inode	*ip)
 602{
 603	uint			flags = 0;
 604
 605	if (ip->i_diflags & XFS_DIFLAG_ANY) {
 606		if (ip->i_diflags & XFS_DIFLAG_REALTIME)
 607			flags |= FS_XFLAG_REALTIME;
 608		if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
 609			flags |= FS_XFLAG_PREALLOC;
 610		if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
 611			flags |= FS_XFLAG_IMMUTABLE;
 612		if (ip->i_diflags & XFS_DIFLAG_APPEND)
 613			flags |= FS_XFLAG_APPEND;
 614		if (ip->i_diflags & XFS_DIFLAG_SYNC)
 615			flags |= FS_XFLAG_SYNC;
 616		if (ip->i_diflags & XFS_DIFLAG_NOATIME)
 617			flags |= FS_XFLAG_NOATIME;
 618		if (ip->i_diflags & XFS_DIFLAG_NODUMP)
 619			flags |= FS_XFLAG_NODUMP;
 620		if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
 621			flags |= FS_XFLAG_RTINHERIT;
 622		if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
 623			flags |= FS_XFLAG_PROJINHERIT;
 624		if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
 625			flags |= FS_XFLAG_NOSYMLINKS;
 626		if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
 627			flags |= FS_XFLAG_EXTSIZE;
 628		if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
 629			flags |= FS_XFLAG_EXTSZINHERIT;
 630		if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
 631			flags |= FS_XFLAG_NODEFRAG;
 632		if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
 633			flags |= FS_XFLAG_FILESTREAM;
 634	}
 635
 636	if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
 637		if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
 638			flags |= FS_XFLAG_DAX;
 639		if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
 640			flags |= FS_XFLAG_COWEXTSIZE;
 641	}
 642
 643	if (XFS_IFORK_Q(ip))
 644		flags |= FS_XFLAG_HASATTR;
 645	return flags;
 646}
 647
 648/*
 649 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 650 * is allowed, otherwise it has to be an exact match. If a CI match is found,
 651 * ci_name->name will point to a the actual name (caller must free) or
 652 * will be set to NULL if an exact match is found.
 653 */
 654int
 655xfs_lookup(
 656	xfs_inode_t		*dp,
 657	struct xfs_name		*name,
 658	xfs_inode_t		**ipp,
 659	struct xfs_name		*ci_name)
 660{
 661	xfs_ino_t		inum;
 662	int			error;
 663
 664	trace_xfs_lookup(dp, name);
 665
 666	if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 
 
 667		return -EIO;
 668
 669	error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 670	if (error)
 671		goto out_unlock;
 672
 673	error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 674	if (error)
 675		goto out_free_name;
 676
 
 
 
 
 
 
 
 
 
 
 677	return 0;
 678
 
 
 679out_free_name:
 680	if (ci_name)
 681		kmem_free(ci_name->name);
 682out_unlock:
 683	*ipp = NULL;
 684	return error;
 685}
 686
 687/* Propagate di_flags from a parent inode to a child inode. */
 688static void
 689xfs_inode_inherit_flags(
 690	struct xfs_inode	*ip,
 691	const struct xfs_inode	*pip)
 692{
 693	unsigned int		di_flags = 0;
 694	xfs_failaddr_t		failaddr;
 695	umode_t			mode = VFS_I(ip)->i_mode;
 696
 697	if (S_ISDIR(mode)) {
 698		if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
 699			di_flags |= XFS_DIFLAG_RTINHERIT;
 700		if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
 701			di_flags |= XFS_DIFLAG_EXTSZINHERIT;
 702			ip->i_extsize = pip->i_extsize;
 703		}
 704		if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
 705			di_flags |= XFS_DIFLAG_PROJINHERIT;
 706	} else if (S_ISREG(mode)) {
 707		if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
 708		    xfs_sb_version_hasrealtime(&ip->i_mount->m_sb))
 709			di_flags |= XFS_DIFLAG_REALTIME;
 710		if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
 711			di_flags |= XFS_DIFLAG_EXTSIZE;
 712			ip->i_extsize = pip->i_extsize;
 713		}
 714	}
 715	if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
 716	    xfs_inherit_noatime)
 717		di_flags |= XFS_DIFLAG_NOATIME;
 718	if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
 719	    xfs_inherit_nodump)
 720		di_flags |= XFS_DIFLAG_NODUMP;
 721	if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
 722	    xfs_inherit_sync)
 723		di_flags |= XFS_DIFLAG_SYNC;
 724	if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
 725	    xfs_inherit_nosymlinks)
 726		di_flags |= XFS_DIFLAG_NOSYMLINKS;
 727	if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
 728	    xfs_inherit_nodefrag)
 729		di_flags |= XFS_DIFLAG_NODEFRAG;
 730	if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
 731		di_flags |= XFS_DIFLAG_FILESTREAM;
 732
 733	ip->i_diflags |= di_flags;
 734
 735	/*
 736	 * Inode verifiers on older kernels only check that the extent size
 737	 * hint is an integer multiple of the rt extent size on realtime files.
 738	 * They did not check the hint alignment on a directory with both
 739	 * rtinherit and extszinherit flags set.  If the misaligned hint is
 740	 * propagated from a directory into a new realtime file, new file
 741	 * allocations will fail due to math errors in the rt allocator and/or
 742	 * trip the verifiers.  Validate the hint settings in the new file so
 743	 * that we don't let broken hints propagate.
 744	 */
 745	failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
 746			VFS_I(ip)->i_mode, ip->i_diflags);
 747	if (failaddr) {
 748		ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
 749				   XFS_DIFLAG_EXTSZINHERIT);
 750		ip->i_extsize = 0;
 751	}
 752}
 753
 754/* Propagate di_flags2 from a parent inode to a child inode. */
 755static void
 756xfs_inode_inherit_flags2(
 757	struct xfs_inode	*ip,
 758	const struct xfs_inode	*pip)
 759{
 760	xfs_failaddr_t		failaddr;
 761
 762	if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
 763		ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
 764		ip->i_cowextsize = pip->i_cowextsize;
 765	}
 766	if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
 767		ip->i_diflags2 |= XFS_DIFLAG2_DAX;
 768
 769	/* Don't let invalid cowextsize hints propagate. */
 770	failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
 771			VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
 772	if (failaddr) {
 773		ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
 774		ip->i_cowextsize = 0;
 775	}
 776}
 777
 778/*
 779 * Initialise a newly allocated inode and return the in-core inode to the
 780 * caller locked exclusively.
 
 
 781 */
 782int
 783xfs_init_new_inode(
 784	struct user_namespace	*mnt_userns,
 785	struct xfs_trans	*tp,
 786	struct xfs_inode	*pip,
 787	xfs_ino_t		ino,
 788	umode_t			mode,
 789	xfs_nlink_t		nlink,
 790	dev_t			rdev,
 791	prid_t			prid,
 792	bool			init_xattrs,
 793	struct xfs_inode	**ipp)
 794{
 795	struct inode		*dir = pip ? VFS_I(pip) : NULL;
 796	struct xfs_mount	*mp = tp->t_mountp;
 797	struct xfs_inode	*ip;
 798	unsigned int		flags;
 799	int			error;
 800	struct timespec64	tv;
 801	struct inode		*inode;
 802
 803	/*
 804	 * Protect against obviously corrupt allocation btree records. Later
 805	 * xfs_iget checks will catch re-allocation of other active in-memory
 806	 * and on-disk inodes. If we don't catch reallocating the parent inode
 807	 * here we will deadlock in xfs_iget() so we have to do these checks
 808	 * first.
 809	 */
 810	if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
 811		xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
 812		return -EFSCORRUPTED;
 813	}
 814
 815	/*
 816	 * Get the in-core inode with the lock held exclusively to prevent
 817	 * others from looking at until we're done.
 818	 */
 819	error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
 820	if (error)
 821		return error;
 822
 823	ASSERT(ip != NULL);
 824	inode = VFS_I(ip);
 825	set_nlink(inode, nlink);
 826	inode->i_rdev = rdev;
 827	ip->i_projid = prid;
 828
 829	if (dir && !(dir->i_mode & S_ISGID) &&
 830	    (mp->m_flags & XFS_MOUNT_GRPID)) {
 831		inode_fsuid_set(inode, mnt_userns);
 832		inode->i_gid = dir->i_gid;
 833		inode->i_mode = mode;
 834	} else {
 835		inode_init_owner(mnt_userns, inode, dir, mode);
 836	}
 837
 838	/*
 839	 * If the group ID of the new file does not match the effective group
 840	 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
 841	 * (and only if the irix_sgid_inherit compatibility variable is set).
 842	 */
 843	if (irix_sgid_inherit &&
 844	    (inode->i_mode & S_ISGID) &&
 845	    !in_group_p(i_gid_into_mnt(mnt_userns, inode)))
 846		inode->i_mode &= ~S_ISGID;
 847
 848	ip->i_disk_size = 0;
 849	ip->i_df.if_nextents = 0;
 850	ASSERT(ip->i_nblocks == 0);
 851
 852	tv = current_time(inode);
 853	inode->i_mtime = tv;
 854	inode->i_atime = tv;
 855	inode->i_ctime = tv;
 856
 857	ip->i_extsize = 0;
 858	ip->i_diflags = 0;
 859
 860	if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
 861		inode_set_iversion(inode, 1);
 862		ip->i_cowextsize = 0;
 863		ip->i_crtime = tv;
 864	}
 865
 866	flags = XFS_ILOG_CORE;
 867	switch (mode & S_IFMT) {
 868	case S_IFIFO:
 869	case S_IFCHR:
 870	case S_IFBLK:
 871	case S_IFSOCK:
 872		ip->i_df.if_format = XFS_DINODE_FMT_DEV;
 873		flags |= XFS_ILOG_DEV;
 874		break;
 875	case S_IFREG:
 876	case S_IFDIR:
 877		if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
 878			xfs_inode_inherit_flags(ip, pip);
 879		if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
 880			xfs_inode_inherit_flags2(ip, pip);
 881		fallthrough;
 882	case S_IFLNK:
 883		ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
 884		ip->i_df.if_bytes = 0;
 885		ip->i_df.if_u1.if_root = NULL;
 886		break;
 887	default:
 888		ASSERT(0);
 889	}
 890
 891	/*
 892	 * If we need to create attributes immediately after allocating the
 893	 * inode, initialise an empty attribute fork right now. We use the
 894	 * default fork offset for attributes here as we don't know exactly what
 895	 * size or how many attributes we might be adding. We can do this
 896	 * safely here because we know the data fork is completely empty and
 897	 * this saves us from needing to run a separate transaction to set the
 898	 * fork offset in the immediate future.
 899	 */
 900	if (init_xattrs && xfs_sb_version_hasattr(&mp->m_sb)) {
 901		ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
 902		ip->i_afp = xfs_ifork_alloc(XFS_DINODE_FMT_EXTENTS, 0);
 903	}
 904
 905	/*
 906	 * Log the new values stuffed into the inode.
 907	 */
 908	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 909	xfs_trans_log_inode(tp, ip, flags);
 910
 911	/* now that we have an i_mode we can setup the inode structure */
 912	xfs_setup_inode(ip);
 913
 914	*ipp = ip;
 915	return 0;
 916}
 917
 918/*
 919 * Decrement the link count on an inode & log the change.  If this causes the
 920 * link count to go to zero, move the inode to AGI unlinked list so that it can
 921 * be freed when the last active reference goes away via xfs_inactive().
 922 */
 923static int			/* error */
 924xfs_droplink(
 925	xfs_trans_t *tp,
 926	xfs_inode_t *ip)
 927{
 928	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
 929
 930	drop_nlink(VFS_I(ip));
 931	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 
 
 
 
 
 
 
 
 
 
 
 932
 933	if (VFS_I(ip)->i_nlink)
 934		return 0;
 935
 936	return xfs_iunlink(tp, ip);
 937}
 938
 939/*
 940 * Increment the link count on an inode & log the change.
 941 */
 942static void
 943xfs_bumplink(
 944	xfs_trans_t *tp,
 945	xfs_inode_t *ip)
 946{
 947	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
 948
 949	inc_nlink(VFS_I(ip));
 950	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 951}
 952
 953int
 954xfs_create(
 955	struct user_namespace	*mnt_userns,
 956	xfs_inode_t		*dp,
 957	struct xfs_name		*name,
 958	umode_t			mode,
 959	dev_t			rdev,
 960	bool			init_xattrs,
 961	xfs_inode_t		**ipp)
 962{
 963	int			is_dir = S_ISDIR(mode);
 
 
 
 
 964	struct xfs_mount	*mp = dp->i_mount;
 965	struct xfs_inode	*ip = NULL;
 966	struct xfs_trans	*tp = NULL;
 967	int			error;
 968	bool                    unlock_dp_on_error = false;
 969	prid_t			prid;
 970	struct xfs_dquot	*udqp = NULL;
 971	struct xfs_dquot	*gdqp = NULL;
 972	struct xfs_dquot	*pdqp = NULL;
 973	struct xfs_trans_res	*tres;
 
 
 
 974	uint			resblks;
 975	xfs_ino_t		ino;
 976
 977	trace_xfs_create(dp, name);
 978
 979	if (XFS_FORCED_SHUTDOWN(mp))
 
 
 980		return -EIO;
 981
 982	prid = xfs_get_initial_prid(dp);
 983
 984	/*
 985	 * Make sure that we have allocated dquot(s) on disk.
 986	 */
 987	error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns),
 988			mapped_fsgid(mnt_userns), prid,
 989			XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
 990			&udqp, &gdqp, &pdqp);
 991	if (error)
 992		return error;
 993
 994	if (is_dir) {
 995		resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 996		tres = &M_RES(mp)->tr_mkdir;
 997	} else {
 998		resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 999		tres = &M_RES(mp)->tr_create;
1000	}
1001
 
 
 
 
1002	/*
1003	 * Initially assume that the file does not exist and
1004	 * reserve the resources for that case.  If that is not
1005	 * the case we'll drop the one we have and get a more
1006	 * appropriate transaction later.
1007	 */
1008	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1009			&tp);
1010	if (error == -ENOSPC) {
1011		/* flush outstanding delalloc blocks and retry */
1012		xfs_flush_inodes(mp);
1013		error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1014				resblks, &tp);
1015	}
1016	if (error)
1017		goto out_release_dquots;
1018
1019	xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1020	unlock_dp_on_error = true;
1021
1022	error = xfs_iext_count_may_overflow(dp, XFS_DATA_FORK,
1023			XFS_IEXT_DIR_MANIP_CNT(mp));
1024	if (error)
1025		goto out_trans_cancel;
1026
1027	/*
1028	 * A newly created regular or special file just has one directory
1029	 * entry pointing to them, but a directory also the "." entry
1030	 * pointing to itself.
1031	 */
1032	error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1033	if (!error)
1034		error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1035				is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1036	if (error)
1037		goto out_trans_cancel;
1038
1039	/*
1040	 * Now we join the directory inode to the transaction.  We do not do it
1041	 * earlier because xfs_dialloc might commit the previous transaction
1042	 * (and release all the locks).  An error from here on will result in
1043	 * the transaction cancel unlocking dp so don't do it explicitly in the
1044	 * error path.
1045	 */
1046	xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1047	unlock_dp_on_error = false;
1048
1049	error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1050					resblks - XFS_IALLOC_SPACE_RES(mp));
1051	if (error) {
1052		ASSERT(error != -ENOSPC);
1053		goto out_trans_cancel;
1054	}
1055	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1056	xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1057
1058	if (is_dir) {
1059		error = xfs_dir_init(tp, ip, dp);
1060		if (error)
1061			goto out_trans_cancel;
1062
1063		xfs_bumplink(tp, dp);
1064	}
1065
1066	/*
1067	 * If this is a synchronous mount, make sure that the
1068	 * create transaction goes to disk before returning to
1069	 * the user.
1070	 */
1071	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1072		xfs_trans_set_sync(tp);
1073
1074	/*
1075	 * Attach the dquot(s) to the inodes and modify them incore.
1076	 * These ids of the inode couldn't have changed since the new
1077	 * inode has been locked ever since it was created.
1078	 */
1079	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1080
1081	error = xfs_trans_commit(tp);
1082	if (error)
1083		goto out_release_inode;
1084
1085	xfs_qm_dqrele(udqp);
1086	xfs_qm_dqrele(gdqp);
1087	xfs_qm_dqrele(pdqp);
1088
1089	*ipp = ip;
 
 
 
1090	return 0;
1091
1092 out_trans_cancel:
1093	xfs_trans_cancel(tp);
1094 out_release_inode:
1095	/*
1096	 * Wait until after the current transaction is aborted to finish the
1097	 * setup of the inode and release the inode.  This prevents recursive
1098	 * transactions and deadlocks from xfs_inactive.
1099	 */
1100	if (ip) {
1101		xfs_finish_inode_setup(ip);
1102		xfs_irele(ip);
 
1103	}
 
 
1104 out_release_dquots:
1105	xfs_qm_dqrele(udqp);
1106	xfs_qm_dqrele(gdqp);
1107	xfs_qm_dqrele(pdqp);
1108
1109	if (unlock_dp_on_error)
1110		xfs_iunlock(dp, XFS_ILOCK_EXCL);
1111	return error;
1112}
1113
1114int
1115xfs_create_tmpfile(
1116	struct user_namespace	*mnt_userns,
1117	struct xfs_inode	*dp,
1118	umode_t			mode,
1119	struct xfs_inode	**ipp)
1120{
 
1121	struct xfs_mount	*mp = dp->i_mount;
1122	struct xfs_inode	*ip = NULL;
1123	struct xfs_trans	*tp = NULL;
1124	int			error;
1125	prid_t                  prid;
1126	struct xfs_dquot	*udqp = NULL;
1127	struct xfs_dquot	*gdqp = NULL;
1128	struct xfs_dquot	*pdqp = NULL;
1129	struct xfs_trans_res	*tres;
 
1130	uint			resblks;
1131	xfs_ino_t		ino;
 
 
1132
1133	if (XFS_FORCED_SHUTDOWN(mp))
1134		return -EIO;
1135
1136	prid = xfs_get_initial_prid(dp);
1137
1138	/*
1139	 * Make sure that we have allocated dquot(s) on disk.
1140	 */
1141	error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns),
1142			mapped_fsgid(mnt_userns), prid,
1143			XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1144			&udqp, &gdqp, &pdqp);
1145	if (error)
1146		return error;
1147
1148	resblks = XFS_IALLOC_SPACE_RES(mp);
1149	tres = &M_RES(mp)->tr_create_tmpfile;
1150
1151	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1152			&tp);
1153	if (error)
1154		goto out_release_dquots;
1155
1156	error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1157	if (!error)
1158		error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1159				0, 0, prid, false, &ip);
1160	if (error)
1161		goto out_trans_cancel;
1162
1163	if (mp->m_flags & XFS_MOUNT_WSYNC)
1164		xfs_trans_set_sync(tp);
1165
1166	/*
1167	 * Attach the dquot(s) to the inodes and modify them incore.
1168	 * These ids of the inode couldn't have changed since the new
1169	 * inode has been locked ever since it was created.
1170	 */
1171	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1172
1173	error = xfs_iunlink(tp, ip);
1174	if (error)
1175		goto out_trans_cancel;
1176
1177	error = xfs_trans_commit(tp);
1178	if (error)
1179		goto out_release_inode;
1180
1181	xfs_qm_dqrele(udqp);
1182	xfs_qm_dqrele(gdqp);
1183	xfs_qm_dqrele(pdqp);
1184
1185	*ipp = ip;
 
1186	return 0;
1187
1188 out_trans_cancel:
1189	xfs_trans_cancel(tp);
1190 out_release_inode:
1191	/*
1192	 * Wait until after the current transaction is aborted to finish the
1193	 * setup of the inode and release the inode.  This prevents recursive
1194	 * transactions and deadlocks from xfs_inactive.
1195	 */
1196	if (ip) {
 
1197		xfs_finish_inode_setup(ip);
1198		xfs_irele(ip);
1199	}
1200 out_release_dquots:
1201	xfs_qm_dqrele(udqp);
1202	xfs_qm_dqrele(gdqp);
1203	xfs_qm_dqrele(pdqp);
1204
1205	return error;
1206}
1207
1208int
1209xfs_link(
1210	xfs_inode_t		*tdp,
1211	xfs_inode_t		*sip,
1212	struct xfs_name		*target_name)
1213{
1214	xfs_mount_t		*mp = tdp->i_mount;
1215	xfs_trans_t		*tp;
1216	int			error;
 
 
 
 
 
1217	int			resblks;
1218
1219	trace_xfs_link(tdp, target_name);
1220
1221	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1222
1223	if (XFS_FORCED_SHUTDOWN(mp))
 
 
1224		return -EIO;
1225
1226	error = xfs_qm_dqattach(sip);
1227	if (error)
1228		goto std_return;
1229
1230	error = xfs_qm_dqattach(tdp);
1231	if (error)
1232		goto std_return;
1233
1234	resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1235	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1236	if (error == -ENOSPC) {
1237		resblks = 0;
1238		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1239	}
1240	if (error)
1241		goto std_return;
1242
1243	xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
 
 
 
 
1244
1245	xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1246	xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1247
1248	error = xfs_iext_count_may_overflow(tdp, XFS_DATA_FORK,
1249			XFS_IEXT_DIR_MANIP_CNT(mp));
1250	if (error)
 
1251		goto error_return;
 
1252
1253	/*
1254	 * If we are using project inheritance, we only allow hard link
1255	 * creation in our tree when the project IDs are the same; else
1256	 * the tree quota mechanism could be circumvented.
1257	 */
1258	if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1259		     tdp->i_projid != sip->i_projid)) {
1260		error = -EXDEV;
1261		goto error_return;
1262	}
1263
1264	if (!resblks) {
1265		error = xfs_dir_canenter(tp, tdp, target_name);
1266		if (error)
 
 
 
 
1267			goto error_return;
 
1268	}
1269
1270	/*
1271	 * Handle initial link state of O_TMPFILE inode
1272	 */
1273	if (VFS_I(sip)->i_nlink == 0) {
1274		struct xfs_perag	*pag;
1275
1276		pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1277		error = xfs_iunlink_remove(tp, pag, sip);
1278		xfs_perag_put(pag);
1279		if (error)
1280			goto error_return;
1281	}
1282
1283	error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1284				   resblks);
1285	if (error)
1286		goto error_return;
1287	xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1288	xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1289
1290	xfs_bumplink(tp, sip);
1291
1292	/*
1293	 * If this is a synchronous mount, make sure that the
1294	 * link transaction goes to disk before returning to
1295	 * the user.
1296	 */
1297	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1298		xfs_trans_set_sync(tp);
1299
1300	return xfs_trans_commit(tp);
 
 
 
 
1301
1302 error_return:
1303	xfs_trans_cancel(tp);
 
 
 
 
1304 std_return:
 
 
1305	return error;
1306}
1307
1308/* Clear the reflink flag and the cowblocks tag if possible. */
1309static void
1310xfs_itruncate_clear_reflink_flags(
1311	struct xfs_inode	*ip)
1312{
1313	struct xfs_ifork	*dfork;
1314	struct xfs_ifork	*cfork;
1315
1316	if (!xfs_is_reflink_inode(ip))
1317		return;
1318	dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1319	cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1320	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1321		ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1322	if (cfork->if_bytes == 0)
1323		xfs_inode_clear_cowblocks_tag(ip);
1324}
1325
1326/*
1327 * Free up the underlying blocks past new_size.  The new size must be smaller
1328 * than the current size.  This routine can be used both for the attribute and
1329 * data fork, and does not modify the inode size, which is left to the caller.
1330 *
1331 * The transaction passed to this routine must have made a permanent log
1332 * reservation of at least XFS_ITRUNCATE_LOG_RES.  This routine may commit the
1333 * given transaction and start new ones, so make sure everything involved in
1334 * the transaction is tidy before calling here.  Some transaction will be
1335 * returned to the caller to be committed.  The incoming transaction must
1336 * already include the inode, and both inode locks must be held exclusively.
1337 * The inode must also be "held" within the transaction.  On return the inode
1338 * will be "held" within the returned transaction.  This routine does NOT
1339 * require any disk space to be reserved for it within the transaction.
1340 *
1341 * If we get an error, we must return with the inode locked and linked into the
1342 * current transaction. This keeps things simple for the higher level code,
1343 * because it always knows that the inode is locked and held in the transaction
1344 * that returns to it whether errors occur or not.  We don't mark the inode
1345 * dirty on error so that transactions can be easily aborted if possible.
1346 */
1347int
1348xfs_itruncate_extents_flags(
1349	struct xfs_trans	**tpp,
1350	struct xfs_inode	*ip,
1351	int			whichfork,
1352	xfs_fsize_t		new_size,
1353	int			flags)
1354{
1355	struct xfs_mount	*mp = ip->i_mount;
1356	struct xfs_trans	*tp = *tpp;
1357	xfs_fileoff_t		first_unmap_block;
1358	xfs_filblks_t		unmap_len;
1359	int			error = 0;
1360
1361	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1362	ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1363	       xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1364	ASSERT(new_size <= XFS_ISIZE(ip));
1365	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1366	ASSERT(ip->i_itemp != NULL);
1367	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1368	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1369
1370	trace_xfs_itruncate_extents_start(ip, new_size);
1371
1372	flags |= xfs_bmapi_aflag(whichfork);
1373
1374	/*
1375	 * Since it is possible for space to become allocated beyond
1376	 * the end of the file (in a crash where the space is allocated
1377	 * but the inode size is not yet updated), simply remove any
1378	 * blocks which show up between the new EOF and the maximum
1379	 * possible file size.
1380	 *
1381	 * We have to free all the blocks to the bmbt maximum offset, even if
1382	 * the page cache can't scale that far.
1383	 */
1384	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1385	if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1386		WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1387		return 0;
1388	}
1389
1390	unmap_len = XFS_MAX_FILEOFF - first_unmap_block + 1;
1391	while (unmap_len > 0) {
1392		ASSERT(tp->t_firstblock == NULLFSBLOCK);
1393		error = __xfs_bunmapi(tp, ip, first_unmap_block, &unmap_len,
1394				flags, XFS_ITRUNC_MAX_EXTENTS);
1395		if (error)
1396			goto out;
1397
1398		/* free the just unmapped extents */
1399		error = xfs_defer_finish(&tp);
1400		if (error)
1401			goto out;
1402	}
1403
1404	if (whichfork == XFS_DATA_FORK) {
1405		/* Remove all pending CoW reservations. */
1406		error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1407				first_unmap_block, XFS_MAX_FILEOFF, true);
1408		if (error)
1409			goto out;
1410
1411		xfs_itruncate_clear_reflink_flags(ip);
1412	}
1413
1414	/*
1415	 * Always re-log the inode so that our permanent transaction can keep
1416	 * on rolling it forward in the log.
1417	 */
1418	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1419
1420	trace_xfs_itruncate_extents_end(ip, new_size);
1421
1422out:
1423	*tpp = tp;
1424	return error;
1425}
1426
1427int
1428xfs_release(
1429	xfs_inode_t	*ip)
 
 
 
 
 
 
1430{
1431	xfs_mount_t	*mp = ip->i_mount;
1432	int		error = 0;
1433
1434	if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1435		return 0;
1436
1437	/* If this is a read-only mount, don't do this (would generate I/O) */
1438	if (mp->m_flags & XFS_MOUNT_RDONLY)
1439		return 0;
1440
1441	if (!XFS_FORCED_SHUTDOWN(mp)) {
1442		int truncated;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1443
1444		/*
1445		 * If we previously truncated this file and removed old data
1446		 * in the process, we want to initiate "early" writeout on
1447		 * the last close.  This is an attempt to combat the notorious
1448		 * NULL files problem which is particularly noticeable from a
1449		 * truncate down, buffered (re-)write (delalloc), followed by
1450		 * a crash.  What we are effectively doing here is
1451		 * significantly reducing the time window where we'd otherwise
1452		 * be exposed to that problem.
1453		 */
1454		truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1455		if (truncated) {
1456			xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1457			if (ip->i_delayed_blks > 0) {
1458				error = filemap_flush(VFS_I(ip)->i_mapping);
1459				if (error)
1460					return error;
1461			}
1462		}
1463	}
1464
1465	if (VFS_I(ip)->i_nlink == 0)
1466		return 0;
1467
1468	/*
1469	 * If we can't get the iolock just skip truncating the blocks past EOF
1470	 * because we could deadlock with the mmap_lock otherwise. We'll get
1471	 * another chance to drop them once the last reference to the inode is
1472	 * dropped, so we'll never leak blocks permanently.
1473	 */
1474	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1475		return 0;
1476
1477	if (xfs_can_free_eofblocks(ip, false)) {
1478		/*
1479		 * Check if the inode is being opened, written and closed
1480		 * frequently and we have delayed allocation blocks outstanding
1481		 * (e.g. streaming writes from the NFS server), truncating the
1482		 * blocks past EOF will cause fragmentation to occur.
1483		 *
1484		 * In this case don't do the truncation, but we have to be
1485		 * careful how we detect this case. Blocks beyond EOF show up as
1486		 * i_delayed_blks even when the inode is clean, so we need to
1487		 * truncate them away first before checking for a dirty release.
1488		 * Hence on the first dirty close we will still remove the
1489		 * speculative allocation, but after that we will leave it in
1490		 * place.
1491		 */
1492		if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1493			goto out_unlock;
1494
1495		error = xfs_free_eofblocks(ip);
1496		if (error)
1497			goto out_unlock;
1498
1499		/* delalloc blocks after truncation means it really is dirty */
1500		if (ip->i_delayed_blks)
1501			xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1502	}
1503
1504out_unlock:
1505	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1506	return error;
1507}
1508
1509/*
1510 * xfs_inactive_truncate
1511 *
1512 * Called to perform a truncate when an inode becomes unlinked.
1513 */
1514STATIC int
1515xfs_inactive_truncate(
1516	struct xfs_inode *ip)
1517{
1518	struct xfs_mount	*mp = ip->i_mount;
1519	struct xfs_trans	*tp;
1520	int			error;
1521
1522	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1523	if (error) {
1524		ASSERT(XFS_FORCED_SHUTDOWN(mp));
1525		return error;
1526	}
1527	xfs_ilock(ip, XFS_ILOCK_EXCL);
1528	xfs_trans_ijoin(tp, ip, 0);
1529
1530	/*
1531	 * Log the inode size first to prevent stale data exposure in the event
1532	 * of a system crash before the truncate completes. See the related
1533	 * comment in xfs_vn_setattr_size() for details.
1534	 */
1535	ip->i_disk_size = 0;
1536	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1537
1538	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1539	if (error)
1540		goto error_trans_cancel;
1541
1542	ASSERT(ip->i_df.if_nextents == 0);
1543
1544	error = xfs_trans_commit(tp);
1545	if (error)
1546		goto error_unlock;
1547
1548	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1549	return 0;
1550
1551error_trans_cancel:
1552	xfs_trans_cancel(tp);
1553error_unlock:
1554	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1555	return error;
1556}
1557
1558/*
1559 * xfs_inactive_ifree()
1560 *
1561 * Perform the inode free when an inode is unlinked.
1562 */
1563STATIC int
1564xfs_inactive_ifree(
1565	struct xfs_inode *ip)
1566{
1567	struct xfs_mount	*mp = ip->i_mount;
1568	struct xfs_trans	*tp;
1569	int			error;
1570
1571	/*
1572	 * We try to use a per-AG reservation for any block needed by the finobt
1573	 * tree, but as the finobt feature predates the per-AG reservation
1574	 * support a degraded file system might not have enough space for the
1575	 * reservation at mount time.  In that case try to dip into the reserved
1576	 * pool and pray.
1577	 *
1578	 * Send a warning if the reservation does happen to fail, as the inode
1579	 * now remains allocated and sits on the unlinked list until the fs is
1580	 * repaired.
1581	 */
1582	if (unlikely(mp->m_finobt_nores)) {
1583		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1584				XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1585				&tp);
1586	} else {
1587		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1588	}
1589	if (error) {
1590		if (error == -ENOSPC) {
1591			xfs_warn_ratelimited(mp,
1592			"Failed to remove inode(s) from unlinked list. "
1593			"Please free space, unmount and run xfs_repair.");
1594		} else {
1595			ASSERT(XFS_FORCED_SHUTDOWN(mp));
1596		}
1597		return error;
1598	}
1599
1600	/*
1601	 * We do not hold the inode locked across the entire rolling transaction
1602	 * here. We only need to hold it for the first transaction that
1603	 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1604	 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1605	 * here breaks the relationship between cluster buffer invalidation and
1606	 * stale inode invalidation on cluster buffer item journal commit
1607	 * completion, and can result in leaving dirty stale inodes hanging
1608	 * around in memory.
1609	 *
1610	 * We have no need for serialising this inode operation against other
1611	 * operations - we freed the inode and hence reallocation is required
1612	 * and that will serialise on reallocating the space the deferops need
1613	 * to free. Hence we can unlock the inode on the first commit of
1614	 * the transaction rather than roll it right through the deferops. This
1615	 * avoids relogging the XFS_ISTALE inode.
1616	 *
1617	 * We check that xfs_ifree() hasn't grown an internal transaction roll
1618	 * by asserting that the inode is still locked when it returns.
1619	 */
1620	xfs_ilock(ip, XFS_ILOCK_EXCL);
1621	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1622
1623	error = xfs_ifree(tp, ip);
1624	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1625	if (error) {
1626		/*
1627		 * If we fail to free the inode, shut down.  The cancel
1628		 * might do that, we need to make sure.  Otherwise the
1629		 * inode might be lost for a long time or forever.
1630		 */
1631		if (!XFS_FORCED_SHUTDOWN(mp)) {
1632			xfs_notice(mp, "%s: xfs_ifree returned error %d",
1633				__func__, error);
1634			xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1635		}
1636		xfs_trans_cancel(tp);
1637		return error;
1638	}
1639
1640	/*
1641	 * Credit the quota account(s). The inode is gone.
1642	 */
1643	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1644
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1645	/*
1646	 * Just ignore errors at this point.  There is nothing we can do except
1647	 * to try to keep going. Make sure it's not a silent error.
 
 
 
 
1648	 */
1649	error = xfs_trans_commit(tp);
1650	if (error)
1651		xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1652			__func__, error);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1653
1654	return 0;
 
1655}
1656
1657/*
1658 * xfs_inactive
1659 *
1660 * This is called when the vnode reference count for the vnode
1661 * goes to zero.  If the file has been unlinked, then it must
1662 * now be truncated.  Also, we clear all of the read-ahead state
1663 * kept for the inode here since the file is now closed.
1664 */
1665void
1666xfs_inactive(
1667	xfs_inode_t	*ip)
1668{
1669	struct xfs_mount	*mp;
1670	int			error;
1671	int			truncate = 0;
1672
1673	/*
1674	 * If the inode is already free, then there can be nothing
1675	 * to clean up here.
1676	 */
1677	if (VFS_I(ip)->i_mode == 0) {
1678		ASSERT(ip->i_df.if_broot_bytes == 0);
1679		goto out;
1680	}
1681
1682	mp = ip->i_mount;
1683	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1684
1685	/* If this is a read-only mount, don't do this (would generate I/O) */
1686	if (mp->m_flags & XFS_MOUNT_RDONLY)
 
 
 
 
 
1687		goto out;
1688
1689	/* Metadata inodes require explicit resource cleanup. */
1690	if (xfs_is_metadata_inode(ip))
1691		goto out;
1692
1693	/* Try to clean out the cow blocks if there are any. */
1694	if (xfs_inode_has_cow_data(ip))
1695		xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
 
 
 
1696
1697	if (VFS_I(ip)->i_nlink != 0) {
1698		/*
1699		 * force is true because we are evicting an inode from the
1700		 * cache. Post-eof blocks must be freed, lest we end up with
1701		 * broken free space accounting.
1702		 *
1703		 * Note: don't bother with iolock here since lockdep complains
1704		 * about acquiring it in reclaim context. We have the only
1705		 * reference to the inode at this point anyways.
1706		 */
1707		if (xfs_can_free_eofblocks(ip, true))
1708			xfs_free_eofblocks(ip);
1709
1710		goto out;
1711	}
1712
1713	if (S_ISREG(VFS_I(ip)->i_mode) &&
1714	    (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1715	     ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1716		truncate = 1;
1717
1718	error = xfs_qm_dqattach(ip);
1719	if (error)
1720		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1721
1722	if (S_ISLNK(VFS_I(ip)->i_mode))
1723		error = xfs_inactive_symlink(ip);
1724	else if (truncate)
1725		error = xfs_inactive_truncate(ip);
1726	if (error)
1727		goto out;
1728
1729	/*
1730	 * If there are attributes associated with the file then blow them away
1731	 * now.  The code calls a routine that recursively deconstructs the
1732	 * attribute fork. If also blows away the in-core attribute fork.
1733	 */
1734	if (XFS_IFORK_Q(ip)) {
1735		error = xfs_attr_inactive(ip);
1736		if (error)
1737			goto out;
1738	}
1739
1740	ASSERT(!ip->i_afp);
1741	ASSERT(ip->i_forkoff == 0);
1742
1743	/*
1744	 * Free the inode.
1745	 */
1746	xfs_inactive_ifree(ip);
1747
1748out:
1749	/*
1750	 * We're done making metadata updates for this inode, so we can release
1751	 * the attached dquots.
1752	 */
1753	xfs_qm_dqdetach(ip);
 
1754}
1755
1756/*
1757 * In-Core Unlinked List Lookups
1758 * =============================
1759 *
1760 * Every inode is supposed to be reachable from some other piece of metadata
1761 * with the exception of the root directory.  Inodes with a connection to a
1762 * file descriptor but not linked from anywhere in the on-disk directory tree
1763 * are collectively known as unlinked inodes, though the filesystem itself
1764 * maintains links to these inodes so that on-disk metadata are consistent.
1765 *
1766 * XFS implements a per-AG on-disk hash table of unlinked inodes.  The AGI
1767 * header contains a number of buckets that point to an inode, and each inode
1768 * record has a pointer to the next inode in the hash chain.  This
1769 * singly-linked list causes scaling problems in the iunlink remove function
1770 * because we must walk that list to find the inode that points to the inode
1771 * being removed from the unlinked hash bucket list.
1772 *
1773 * What if we modelled the unlinked list as a collection of records capturing
1774 * "X.next_unlinked = Y" relations?  If we indexed those records on Y, we'd
1775 * have a fast way to look up unlinked list predecessors, which avoids the
1776 * slow list walk.  That's exactly what we do here (in-core) with a per-AG
1777 * rhashtable.
1778 *
1779 * Because this is a backref cache, we ignore operational failures since the
1780 * iunlink code can fall back to the slow bucket walk.  The only errors that
1781 * should bubble out are for obviously incorrect situations.
1782 *
1783 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1784 * access or have otherwise provided for concurrency control.
1785 */
1786
1787/* Capture a "X.next_unlinked = Y" relationship. */
1788struct xfs_iunlink {
1789	struct rhash_head	iu_rhash_head;
1790	xfs_agino_t		iu_agino;		/* X */
1791	xfs_agino_t		iu_next_unlinked;	/* Y */
1792};
1793
1794/* Unlinked list predecessor lookup hashtable construction */
1795static int
1796xfs_iunlink_obj_cmpfn(
1797	struct rhashtable_compare_arg	*arg,
1798	const void			*obj)
1799{
1800	const xfs_agino_t		*key = arg->key;
1801	const struct xfs_iunlink	*iu = obj;
1802
1803	if (iu->iu_next_unlinked != *key)
1804		return 1;
1805	return 0;
1806}
1807
1808static const struct rhashtable_params xfs_iunlink_hash_params = {
1809	.min_size		= XFS_AGI_UNLINKED_BUCKETS,
1810	.key_len		= sizeof(xfs_agino_t),
1811	.key_offset		= offsetof(struct xfs_iunlink,
1812					   iu_next_unlinked),
1813	.head_offset		= offsetof(struct xfs_iunlink, iu_rhash_head),
1814	.automatic_shrinking	= true,
1815	.obj_cmpfn		= xfs_iunlink_obj_cmpfn,
1816};
1817
1818/*
1819 * Return X, where X.next_unlinked == @agino.  Returns NULLAGINO if no such
1820 * relation is found.
1821 */
1822static xfs_agino_t
1823xfs_iunlink_lookup_backref(
1824	struct xfs_perag	*pag,
1825	xfs_agino_t		agino)
1826{
1827	struct xfs_iunlink	*iu;
1828
1829	iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1830			xfs_iunlink_hash_params);
1831	return iu ? iu->iu_agino : NULLAGINO;
1832}
 
 
 
1833
1834/*
1835 * Take ownership of an iunlink cache entry and insert it into the hash table.
1836 * If successful, the entry will be owned by the cache; if not, it is freed.
1837 * Either way, the caller does not own @iu after this call.
1838 */
1839static int
1840xfs_iunlink_insert_backref(
1841	struct xfs_perag	*pag,
1842	struct xfs_iunlink	*iu)
1843{
1844	int			error;
1845
1846	error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1847			&iu->iu_rhash_head, xfs_iunlink_hash_params);
1848	/*
1849	 * Fail loudly if there already was an entry because that's a sign of
1850	 * corruption of in-memory data.  Also fail loudly if we see an error
1851	 * code we didn't anticipate from the rhashtable code.  Currently we
1852	 * only anticipate ENOMEM.
1853	 */
1854	if (error) {
1855		WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1856		kmem_free(iu);
1857	}
1858	/*
1859	 * Absorb any runtime errors that aren't a result of corruption because
1860	 * this is a cache and we can always fall back to bucket list scanning.
1861	 */
1862	if (error != 0 && error != -EEXIST)
1863		error = 0;
1864	return error;
1865}
1866
1867/* Remember that @prev_agino.next_unlinked = @this_agino. */
1868static int
1869xfs_iunlink_add_backref(
1870	struct xfs_perag	*pag,
1871	xfs_agino_t		prev_agino,
1872	xfs_agino_t		this_agino)
1873{
1874	struct xfs_iunlink	*iu;
1875
1876	if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
1877		return 0;
1878
1879	iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
1880	iu->iu_agino = prev_agino;
1881	iu->iu_next_unlinked = this_agino;
1882
1883	return xfs_iunlink_insert_backref(pag, iu);
1884}
1885
1886/*
1887 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
1888 * If @next_unlinked is NULLAGINO, we drop the backref and exit.  If there
1889 * wasn't any such entry then we don't bother.
1890 */
1891static int
1892xfs_iunlink_change_backref(
1893	struct xfs_perag	*pag,
1894	xfs_agino_t		agino,
1895	xfs_agino_t		next_unlinked)
1896{
1897	struct xfs_iunlink	*iu;
1898	int			error;
1899
1900	/* Look up the old entry; if there wasn't one then exit. */
1901	iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1902			xfs_iunlink_hash_params);
1903	if (!iu)
1904		return 0;
1905
1906	/*
1907	 * Remove the entry.  This shouldn't ever return an error, but if we
1908	 * couldn't remove the old entry we don't want to add it again to the
1909	 * hash table, and if the entry disappeared on us then someone's
1910	 * violated the locking rules and we need to fail loudly.  Either way
1911	 * we cannot remove the inode because internal state is or would have
1912	 * been corrupt.
1913	 */
1914	error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
1915			&iu->iu_rhash_head, xfs_iunlink_hash_params);
1916	if (error)
1917		return error;
1918
1919	/* If there is no new next entry just free our item and return. */
1920	if (next_unlinked == NULLAGINO) {
1921		kmem_free(iu);
1922		return 0;
1923	}
1924
1925	/* Update the entry and re-add it to the hash table. */
1926	iu->iu_next_unlinked = next_unlinked;
1927	return xfs_iunlink_insert_backref(pag, iu);
1928}
1929
1930/* Set up the in-core predecessor structures. */
1931int
1932xfs_iunlink_init(
1933	struct xfs_perag	*pag)
1934{
1935	return rhashtable_init(&pag->pagi_unlinked_hash,
1936			&xfs_iunlink_hash_params);
1937}
1938
1939/* Free the in-core predecessor structures. */
1940static void
1941xfs_iunlink_free_item(
1942	void			*ptr,
1943	void			*arg)
1944{
1945	struct xfs_iunlink	*iu = ptr;
1946	bool			*freed_anything = arg;
1947
1948	*freed_anything = true;
1949	kmem_free(iu);
1950}
1951
1952void
1953xfs_iunlink_destroy(
1954	struct xfs_perag	*pag)
1955{
1956	bool			freed_anything = false;
1957
1958	rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
1959			xfs_iunlink_free_item, &freed_anything);
1960
1961	ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
1962}
1963
1964/*
1965 * Point the AGI unlinked bucket at an inode and log the results.  The caller
1966 * is responsible for validating the old value.
1967 */
1968STATIC int
1969xfs_iunlink_update_bucket(
1970	struct xfs_trans	*tp,
1971	struct xfs_perag	*pag,
1972	struct xfs_buf		*agibp,
1973	unsigned int		bucket_index,
1974	xfs_agino_t		new_agino)
1975{
1976	struct xfs_agi		*agi = agibp->b_addr;
1977	xfs_agino_t		old_value;
1978	int			offset;
1979
1980	ASSERT(xfs_verify_agino_or_null(tp->t_mountp, pag->pag_agno, new_agino));
1981
1982	old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
1983	trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
1984			old_value, new_agino);
1985
1986	/*
1987	 * We should never find the head of the list already set to the value
1988	 * passed in because either we're adding or removing ourselves from the
1989	 * head of the list.
1990	 */
1991	if (old_value == new_agino) {
1992		xfs_buf_mark_corrupt(agibp);
1993		return -EFSCORRUPTED;
1994	}
1995
1996	agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
1997	offset = offsetof(struct xfs_agi, agi_unlinked) +
1998			(sizeof(xfs_agino_t) * bucket_index);
1999	xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2000	return 0;
2001}
2002
2003/* Set an on-disk inode's next_unlinked pointer. */
2004STATIC void
2005xfs_iunlink_update_dinode(
2006	struct xfs_trans	*tp,
2007	struct xfs_perag	*pag,
2008	xfs_agino_t		agino,
2009	struct xfs_buf		*ibp,
2010	struct xfs_dinode	*dip,
2011	struct xfs_imap		*imap,
2012	xfs_agino_t		next_agino)
2013{
2014	struct xfs_mount	*mp = tp->t_mountp;
2015	int			offset;
2016
2017	ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2018
2019	trace_xfs_iunlink_update_dinode(mp, pag->pag_agno, agino,
2020			be32_to_cpu(dip->di_next_unlinked), next_agino);
2021
2022	dip->di_next_unlinked = cpu_to_be32(next_agino);
2023	offset = imap->im_boffset +
2024			offsetof(struct xfs_dinode, di_next_unlinked);
2025
2026	/* need to recalc the inode CRC if appropriate */
2027	xfs_dinode_calc_crc(mp, dip);
2028	xfs_trans_inode_buf(tp, ibp);
2029	xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2030}
2031
2032/* Set an in-core inode's unlinked pointer and return the old value. */
2033STATIC int
2034xfs_iunlink_update_inode(
2035	struct xfs_trans	*tp,
2036	struct xfs_inode	*ip,
2037	struct xfs_perag	*pag,
2038	xfs_agino_t		next_agino,
2039	xfs_agino_t		*old_next_agino)
2040{
2041	struct xfs_mount	*mp = tp->t_mountp;
2042	struct xfs_dinode	*dip;
2043	struct xfs_buf		*ibp;
2044	xfs_agino_t		old_value;
2045	int			error;
2046
2047	ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2048
2049	error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &ibp);
2050	if (error)
2051		return error;
2052	dip = xfs_buf_offset(ibp, ip->i_imap.im_boffset);
 
 
2053
2054	/* Make sure the old pointer isn't garbage. */
2055	old_value = be32_to_cpu(dip->di_next_unlinked);
2056	if (!xfs_verify_agino_or_null(mp, pag->pag_agno, old_value)) {
2057		xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__, dip,
2058				sizeof(*dip), __this_address);
2059		error = -EFSCORRUPTED;
2060		goto out;
2061	}
2062
2063	/*
2064	 * Since we're updating a linked list, we should never find that the
2065	 * current pointer is the same as the new value, unless we're
2066	 * terminating the list.
2067	 */
2068	*old_next_agino = old_value;
2069	if (old_value == next_agino) {
2070		if (next_agino != NULLAGINO) {
2071			xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__,
2072					dip, sizeof(*dip), __this_address);
2073			error = -EFSCORRUPTED;
2074		}
2075		goto out;
2076	}
2077
2078	/* Ok, update the new pointer. */
2079	xfs_iunlink_update_dinode(tp, pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
2080			ibp, dip, &ip->i_imap, next_agino);
2081	return 0;
2082out:
2083	xfs_trans_brelse(tp, ibp);
2084	return error;
2085}
2086
2087/*
2088 * This is called when the inode's link count has gone to 0 or we are creating
2089 * a tmpfile via O_TMPFILE.  The inode @ip must have nlink == 0.
2090 *
2091 * We place the on-disk inode on a list in the AGI.  It will be pulled from this
2092 * list when the inode is freed.
2093 */
2094STATIC int
2095xfs_iunlink(
2096	struct xfs_trans	*tp,
2097	struct xfs_inode	*ip)
2098{
2099	struct xfs_mount	*mp = tp->t_mountp;
2100	struct xfs_perag	*pag;
2101	struct xfs_agi		*agi;
2102	struct xfs_buf		*agibp;
2103	xfs_agino_t		next_agino;
2104	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2105	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2106	int			error;
2107
2108	ASSERT(VFS_I(ip)->i_nlink == 0);
2109	ASSERT(VFS_I(ip)->i_mode != 0);
2110	trace_xfs_iunlink(ip);
2111
2112	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2113
2114	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2115	error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2116	if (error)
2117		goto out;
2118	agi = agibp->b_addr;
2119
2120	/*
2121	 * Get the index into the agi hash table for the list this inode will
2122	 * go on.  Make sure the pointer isn't garbage and that this inode
2123	 * isn't already on the list.
2124	 */
2125	next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2126	if (next_agino == agino ||
2127	    !xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino)) {
2128		xfs_buf_mark_corrupt(agibp);
2129		error = -EFSCORRUPTED;
2130		goto out;
2131	}
2132
2133	if (next_agino != NULLAGINO) {
2134		xfs_agino_t		old_agino;
2135
2136		/*
2137		 * There is already another inode in the bucket, so point this
2138		 * inode to the current head of the list.
2139		 */
2140		error = xfs_iunlink_update_inode(tp, ip, pag, next_agino,
2141				&old_agino);
2142		if (error)
2143			goto out;
2144		ASSERT(old_agino == NULLAGINO);
2145
2146		/*
2147		 * agino has been unlinked, add a backref from the next inode
2148		 * back to agino.
2149		 */
2150		error = xfs_iunlink_add_backref(pag, agino, next_agino);
2151		if (error)
2152			goto out;
2153	}
2154
2155	/* Point the head of the list to point to this inode. */
2156	error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2157out:
2158	xfs_perag_put(pag);
 
 
 
2159	return error;
2160}
2161
2162/* Return the imap, dinode pointer, and buffer for an inode. */
2163STATIC int
2164xfs_iunlink_map_ino(
2165	struct xfs_trans	*tp,
2166	xfs_agnumber_t		agno,
2167	xfs_agino_t		agino,
2168	struct xfs_imap		*imap,
2169	struct xfs_dinode	**dipp,
2170	struct xfs_buf		**bpp)
2171{
2172	struct xfs_mount	*mp = tp->t_mountp;
2173	int			error;
2174
2175	imap->im_blkno = 0;
2176	error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2177	if (error) {
2178		xfs_warn(mp, "%s: xfs_imap returned error %d.",
2179				__func__, error);
2180		return error;
2181	}
2182
2183	error = xfs_imap_to_bp(mp, tp, imap, bpp);
2184	if (error) {
2185		xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2186				__func__, error);
2187		return error;
2188	}
2189
2190	*dipp = xfs_buf_offset(*bpp, imap->im_boffset);
2191	return 0;
2192}
2193
2194/*
2195 * Walk the unlinked chain from @head_agino until we find the inode that
2196 * points to @target_agino.  Return the inode number, map, dinode pointer,
2197 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2198 *
2199 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2200 * @agino, @imap, @dipp, and @bpp are all output parameters.
2201 *
2202 * Do not call this function if @target_agino is the head of the list.
2203 */
2204STATIC int
2205xfs_iunlink_map_prev(
2206	struct xfs_trans	*tp,
2207	struct xfs_perag	*pag,
2208	xfs_agino_t		head_agino,
2209	xfs_agino_t		target_agino,
2210	xfs_agino_t		*agino,
2211	struct xfs_imap		*imap,
2212	struct xfs_dinode	**dipp,
2213	struct xfs_buf		**bpp)
2214{
2215	struct xfs_mount	*mp = tp->t_mountp;
2216	xfs_agino_t		next_agino;
2217	int			error;
2218
2219	ASSERT(head_agino != target_agino);
2220	*bpp = NULL;
2221
2222	/* See if our backref cache can find it faster. */
2223	*agino = xfs_iunlink_lookup_backref(pag, target_agino);
2224	if (*agino != NULLAGINO) {
2225		error = xfs_iunlink_map_ino(tp, pag->pag_agno, *agino, imap,
2226				dipp, bpp);
2227		if (error)
2228			return error;
2229
2230		if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2231			return 0;
2232
2233		/*
2234		 * If we get here the cache contents were corrupt, so drop the
2235		 * buffer and fall back to walking the bucket list.
2236		 */
2237		xfs_trans_brelse(tp, *bpp);
2238		*bpp = NULL;
2239		WARN_ON_ONCE(1);
2240	}
2241
2242	trace_xfs_iunlink_map_prev_fallback(mp, pag->pag_agno);
2243
2244	/* Otherwise, walk the entire bucket until we find it. */
2245	next_agino = head_agino;
2246	while (next_agino != target_agino) {
2247		xfs_agino_t	unlinked_agino;
2248
2249		if (*bpp)
2250			xfs_trans_brelse(tp, *bpp);
2251
2252		*agino = next_agino;
2253		error = xfs_iunlink_map_ino(tp, pag->pag_agno, next_agino, imap,
2254				dipp, bpp);
2255		if (error)
2256			return error;
2257
2258		unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2259		/*
2260		 * Make sure this pointer is valid and isn't an obvious
2261		 * infinite loop.
2262		 */
2263		if (!xfs_verify_agino(mp, pag->pag_agno, unlinked_agino) ||
2264		    next_agino == unlinked_agino) {
2265			XFS_CORRUPTION_ERROR(__func__,
2266					XFS_ERRLEVEL_LOW, mp,
2267					*dipp, sizeof(**dipp));
2268			error = -EFSCORRUPTED;
2269			return error;
2270		}
2271		next_agino = unlinked_agino;
2272	}
2273
2274	return 0;
2275}
2276
2277/*
2278 * Pull the on-disk inode from the AGI unlinked list.
2279 */
2280STATIC int
2281xfs_iunlink_remove(
2282	struct xfs_trans	*tp,
2283	struct xfs_perag	*pag,
2284	struct xfs_inode	*ip)
2285{
2286	struct xfs_mount	*mp = tp->t_mountp;
2287	struct xfs_agi		*agi;
2288	struct xfs_buf		*agibp;
2289	struct xfs_buf		*last_ibp;
2290	struct xfs_dinode	*last_dip = NULL;
2291	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2292	xfs_agino_t		next_agino;
2293	xfs_agino_t		head_agino;
2294	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2295	int			error;
2296
2297	trace_xfs_iunlink_remove(ip);
2298
2299	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2300	error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2301	if (error)
2302		return error;
2303	agi = agibp->b_addr;
2304
2305	/*
2306	 * Get the index into the agi hash table for the list this inode will
2307	 * go on.  Make sure the head pointer isn't garbage.
2308	 */
2309	head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2310	if (!xfs_verify_agino(mp, pag->pag_agno, head_agino)) {
2311		XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2312				agi, sizeof(*agi));
2313		return -EFSCORRUPTED;
2314	}
2315
2316	/*
2317	 * Set our inode's next_unlinked pointer to NULL and then return
2318	 * the old pointer value so that we can update whatever was previous
2319	 * to us in the list to point to whatever was next in the list.
2320	 */
2321	error = xfs_iunlink_update_inode(tp, ip, pag, NULLAGINO, &next_agino);
2322	if (error)
2323		return error;
2324
2325	/*
2326	 * If there was a backref pointing from the next inode back to this
2327	 * one, remove it because we've removed this inode from the list.
2328	 *
2329	 * Later, if this inode was in the middle of the list we'll update
2330	 * this inode's backref to point from the next inode.
2331	 */
2332	if (next_agino != NULLAGINO) {
2333		error = xfs_iunlink_change_backref(pag, next_agino, NULLAGINO);
2334		if (error)
2335			return error;
2336	}
2337
2338	if (head_agino != agino) {
2339		struct xfs_imap	imap;
2340		xfs_agino_t	prev_agino;
2341
2342		/* We need to search the list for the inode being freed. */
2343		error = xfs_iunlink_map_prev(tp, pag, head_agino, agino,
2344				&prev_agino, &imap, &last_dip, &last_ibp);
2345		if (error)
2346			return error;
2347
2348		/* Point the previous inode on the list to the next inode. */
2349		xfs_iunlink_update_dinode(tp, pag, prev_agino, last_ibp,
2350				last_dip, &imap, next_agino);
2351
2352		/*
2353		 * Now we deal with the backref for this inode.  If this inode
2354		 * pointed at a real inode, change the backref that pointed to
2355		 * us to point to our old next.  If this inode was the end of
2356		 * the list, delete the backref that pointed to us.  Note that
2357		 * change_backref takes care of deleting the backref if
2358		 * next_agino is NULLAGINO.
2359		 */
2360		return xfs_iunlink_change_backref(agibp->b_pag, agino,
2361				next_agino);
2362	}
2363
2364	/* Point the head of the list to the next unlinked inode. */
2365	return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2366			next_agino);
2367}
2368
2369/*
2370 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2371 * mark it stale. We should only find clean inodes in this lookup that aren't
2372 * already stale.
2373 */
2374static void
2375xfs_ifree_mark_inode_stale(
2376	struct xfs_perag	*pag,
2377	struct xfs_inode	*free_ip,
2378	xfs_ino_t		inum)
2379{
2380	struct xfs_mount	*mp = pag->pag_mount;
2381	struct xfs_inode_log_item *iip;
2382	struct xfs_inode	*ip;
2383
2384retry:
2385	rcu_read_lock();
2386	ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2387
2388	/* Inode not in memory, nothing to do */
2389	if (!ip) {
2390		rcu_read_unlock();
2391		return;
2392	}
2393
2394	/*
2395	 * because this is an RCU protected lookup, we could find a recently
2396	 * freed or even reallocated inode during the lookup. We need to check
2397	 * under the i_flags_lock for a valid inode here. Skip it if it is not
2398	 * valid, the wrong inode or stale.
2399	 */
2400	spin_lock(&ip->i_flags_lock);
2401	if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2402		goto out_iflags_unlock;
2403
2404	/*
2405	 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2406	 * other inodes that we did not find in the list attached to the buffer
2407	 * and are not already marked stale. If we can't lock it, back off and
2408	 * retry.
2409	 */
2410	if (ip != free_ip) {
2411		if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2412			spin_unlock(&ip->i_flags_lock);
2413			rcu_read_unlock();
2414			delay(1);
2415			goto retry;
2416		}
2417	}
2418	ip->i_flags |= XFS_ISTALE;
2419
2420	/*
2421	 * If the inode is flushing, it is already attached to the buffer.  All
2422	 * we needed to do here is mark the inode stale so buffer IO completion
2423	 * will remove it from the AIL.
2424	 */
2425	iip = ip->i_itemp;
2426	if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2427		ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2428		ASSERT(iip->ili_last_fields);
2429		goto out_iunlock;
2430	}
2431
2432	/*
2433	 * Inodes not attached to the buffer can be released immediately.
2434	 * Everything else has to go through xfs_iflush_abort() on journal
2435	 * commit as the flock synchronises removal of the inode from the
2436	 * cluster buffer against inode reclaim.
2437	 */
2438	if (!iip || list_empty(&iip->ili_item.li_bio_list))
2439		goto out_iunlock;
2440
2441	__xfs_iflags_set(ip, XFS_IFLUSHING);
2442	spin_unlock(&ip->i_flags_lock);
2443	rcu_read_unlock();
2444
2445	/* we have a dirty inode in memory that has not yet been flushed. */
2446	spin_lock(&iip->ili_lock);
2447	iip->ili_last_fields = iip->ili_fields;
2448	iip->ili_fields = 0;
2449	iip->ili_fsync_fields = 0;
2450	spin_unlock(&iip->ili_lock);
2451	ASSERT(iip->ili_last_fields);
2452
2453	if (ip != free_ip)
2454		xfs_iunlock(ip, XFS_ILOCK_EXCL);
2455	return;
2456
2457out_iunlock:
2458	if (ip != free_ip)
2459		xfs_iunlock(ip, XFS_ILOCK_EXCL);
2460out_iflags_unlock:
2461	spin_unlock(&ip->i_flags_lock);
2462	rcu_read_unlock();
2463}
2464
2465/*
2466 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2467 * inodes that are in memory - they all must be marked stale and attached to
2468 * the cluster buffer.
2469 */
2470static int
2471xfs_ifree_cluster(
2472	struct xfs_trans	*tp,
2473	struct xfs_perag	*pag,
2474	struct xfs_inode	*free_ip,
2475	struct xfs_icluster	*xic)
2476{
2477	struct xfs_mount	*mp = free_ip->i_mount;
2478	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2479	struct xfs_buf		*bp;
2480	xfs_daddr_t		blkno;
2481	xfs_ino_t		inum = xic->first_ino;
2482	int			nbufs;
2483	int			i, j;
2484	int			ioffset;
2485	int			error;
2486
2487	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2488
2489	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2490		/*
2491		 * The allocation bitmap tells us which inodes of the chunk were
2492		 * physically allocated. Skip the cluster if an inode falls into
2493		 * a sparse region.
2494		 */
2495		ioffset = inum - xic->first_ino;
2496		if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2497			ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2498			continue;
2499		}
2500
2501		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2502					 XFS_INO_TO_AGBNO(mp, inum));
2503
2504		/*
2505		 * We obtain and lock the backing buffer first in the process
2506		 * here to ensure dirty inodes attached to the buffer remain in
2507		 * the flushing state while we mark them stale.
2508		 *
2509		 * If we scan the in-memory inodes first, then buffer IO can
2510		 * complete before we get a lock on it, and hence we may fail
2511		 * to mark all the active inodes on the buffer stale.
2512		 */
2513		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2514				mp->m_bsize * igeo->blocks_per_cluster,
2515				XBF_UNMAPPED, &bp);
2516		if (error)
2517			return error;
2518
2519		/*
2520		 * This buffer may not have been correctly initialised as we
2521		 * didn't read it from disk. That's not important because we are
2522		 * only using to mark the buffer as stale in the log, and to
2523		 * attach stale cached inodes on it. That means it will never be
2524		 * dispatched for IO. If it is, we want to know about it, and we
2525		 * want it to fail. We can acheive this by adding a write
2526		 * verifier to the buffer.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2527		 */
 
2528		bp->b_ops = &xfs_inode_buf_ops;
2529
2530		/*
2531		 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2532		 * too. This requires lookups, and will skip inodes that we've
2533		 * already marked XFS_ISTALE.
2534		 */
2535		for (i = 0; i < igeo->inodes_per_cluster; i++)
2536			xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2537
2538		xfs_trans_stale_inode_buf(tp, bp);
2539		xfs_trans_binval(tp, bp);
2540	}
2541	return 0;
2542}
2543
2544/*
2545 * This is called to return an inode to the inode free list.
2546 * The inode should already be truncated to 0 length and have
2547 * no pages associated with it.  This routine also assumes that
2548 * the inode is already a part of the transaction.
2549 *
2550 * The on-disk copy of the inode will have been added to the list
2551 * of unlinked inodes in the AGI. We need to remove the inode from
2552 * that list atomically with respect to freeing it here.
2553 */
2554int
2555xfs_ifree(
2556	struct xfs_trans	*tp,
2557	struct xfs_inode	*ip)
2558{
2559	struct xfs_mount	*mp = ip->i_mount;
2560	struct xfs_perag	*pag;
2561	struct xfs_icluster	xic = { 0 };
2562	struct xfs_inode_log_item *iip = ip->i_itemp;
2563	int			error;
2564
2565	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2566	ASSERT(VFS_I(ip)->i_nlink == 0);
2567	ASSERT(ip->i_df.if_nextents == 0);
2568	ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2569	ASSERT(ip->i_nblocks == 0);
2570
2571	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2572
2573	/*
2574	 * Pull the on-disk inode from the AGI unlinked list.
2575	 */
2576	error = xfs_iunlink_remove(tp, pag, ip);
2577	if (error)
2578		goto out;
2579
2580	error = xfs_difree(tp, pag, ip->i_ino, &xic);
2581	if (error)
2582		goto out;
2583
2584	/*
2585	 * Free any local-format data sitting around before we reset the
2586	 * data fork to extents format.  Note that the attr fork data has
2587	 * already been freed by xfs_attr_inactive.
2588	 */
2589	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2590		kmem_free(ip->i_df.if_u1.if_data);
2591		ip->i_df.if_u1.if_data = NULL;
2592		ip->i_df.if_bytes = 0;
2593	}
2594
2595	VFS_I(ip)->i_mode = 0;		/* mark incore inode as free */
2596	ip->i_diflags = 0;
2597	ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2598	ip->i_forkoff = 0;		/* mark the attr fork not in use */
2599	ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2600	if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2601		xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2602
2603	/* Don't attempt to replay owner changes for a deleted inode */
2604	spin_lock(&iip->ili_lock);
2605	iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2606	spin_unlock(&iip->ili_lock);
2607
2608	/*
2609	 * Bump the generation count so no one will be confused
2610	 * by reincarnations of this inode.
2611	 */
2612	VFS_I(ip)->i_generation++;
2613	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2614
2615	if (xic.deleted)
2616		error = xfs_ifree_cluster(tp, pag, ip, &xic);
2617out:
2618	xfs_perag_put(pag);
2619	return error;
2620}
2621
2622/*
2623 * This is called to unpin an inode.  The caller must have the inode locked
2624 * in at least shared mode so that the buffer cannot be subsequently pinned
2625 * once someone is waiting for it to be unpinned.
2626 */
2627static void
2628xfs_iunpin(
2629	struct xfs_inode	*ip)
2630{
2631	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2632
2633	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2634
2635	/* Give the log a push to start the unpinning I/O */
2636	xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2637
2638}
2639
2640static void
2641__xfs_iunpin_wait(
2642	struct xfs_inode	*ip)
2643{
2644	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2645	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2646
2647	xfs_iunpin(ip);
2648
2649	do {
2650		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2651		if (xfs_ipincount(ip))
2652			io_schedule();
2653	} while (xfs_ipincount(ip));
2654	finish_wait(wq, &wait.wq_entry);
2655}
2656
2657void
2658xfs_iunpin_wait(
2659	struct xfs_inode	*ip)
2660{
2661	if (xfs_ipincount(ip))
2662		__xfs_iunpin_wait(ip);
2663}
2664
2665/*
2666 * Removing an inode from the namespace involves removing the directory entry
2667 * and dropping the link count on the inode. Removing the directory entry can
2668 * result in locking an AGF (directory blocks were freed) and removing a link
2669 * count can result in placing the inode on an unlinked list which results in
2670 * locking an AGI.
2671 *
2672 * The big problem here is that we have an ordering constraint on AGF and AGI
2673 * locking - inode allocation locks the AGI, then can allocate a new extent for
2674 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2675 * removes the inode from the unlinked list, requiring that we lock the AGI
2676 * first, and then freeing the inode can result in an inode chunk being freed
2677 * and hence freeing disk space requiring that we lock an AGF.
2678 *
2679 * Hence the ordering that is imposed by other parts of the code is AGI before
2680 * AGF. This means we cannot remove the directory entry before we drop the inode
2681 * reference count and put it on the unlinked list as this results in a lock
2682 * order of AGF then AGI, and this can deadlock against inode allocation and
2683 * freeing. Therefore we must drop the link counts before we remove the
2684 * directory entry.
2685 *
2686 * This is still safe from a transactional point of view - it is not until we
2687 * get to xfs_defer_finish() that we have the possibility of multiple
2688 * transactions in this operation. Hence as long as we remove the directory
2689 * entry and drop the link count in the first transaction of the remove
2690 * operation, there are no transactional constraints on the ordering here.
2691 */
2692int
2693xfs_remove(
2694	xfs_inode_t             *dp,
2695	struct xfs_name		*name,
2696	xfs_inode_t		*ip)
2697{
2698	xfs_mount_t		*mp = dp->i_mount;
2699	xfs_trans_t             *tp = NULL;
 
 
 
 
 
2700	int			is_dir = S_ISDIR(VFS_I(ip)->i_mode);
 
2701	int                     error = 0;
2702	uint			resblks;
2703
2704	trace_xfs_remove(dp, name);
2705
2706	if (XFS_FORCED_SHUTDOWN(mp))
 
 
2707		return -EIO;
2708
2709	error = xfs_qm_dqattach(dp);
2710	if (error)
2711		goto std_return;
2712
2713	error = xfs_qm_dqattach(ip);
2714	if (error)
2715		goto std_return;
2716
 
 
 
 
2717	/*
2718	 * We try to get the real space reservation first,
2719	 * allowing for directory btree deletion(s) implying
2720	 * possible bmap insert(s).  If we can't get the space
2721	 * reservation then we use 0 instead, and avoid the bmap
2722	 * btree insert(s) in the directory code by, if the bmap
2723	 * insert tries to happen, instead trimming the LAST
2724	 * block from the directory.
2725	 */
2726	resblks = XFS_REMOVE_SPACE_RES(mp);
2727	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2728	if (error == -ENOSPC) {
2729		resblks = 0;
2730		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2731				&tp);
2732	}
2733	if (error) {
2734		ASSERT(error != -ENOSPC);
2735		goto std_return;
2736	}
2737
2738	xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2739
2740	xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2741	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2742
2743	/*
2744	 * If we're removing a directory perform some additional validation.
2745	 */
2746	if (is_dir) {
2747		ASSERT(VFS_I(ip)->i_nlink >= 2);
2748		if (VFS_I(ip)->i_nlink != 2) {
2749			error = -ENOTEMPTY;
2750			goto out_trans_cancel;
2751		}
2752		if (!xfs_dir_isempty(ip)) {
2753			error = -ENOTEMPTY;
2754			goto out_trans_cancel;
2755		}
2756
2757		/* Drop the link from ip's "..".  */
2758		error = xfs_droplink(tp, dp);
2759		if (error)
2760			goto out_trans_cancel;
2761
2762		/* Drop the "." link from ip to self.  */
2763		error = xfs_droplink(tp, ip);
2764		if (error)
2765			goto out_trans_cancel;
2766
2767		/*
2768		 * Point the unlinked child directory's ".." entry to the root
2769		 * directory to eliminate back-references to inodes that may
2770		 * get freed before the child directory is closed.  If the fs
2771		 * gets shrunk, this can lead to dirent inode validation errors.
2772		 */
2773		if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2774			error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2775					tp->t_mountp->m_sb.sb_rootino, 0);
2776			if (error)
2777				return error;
2778		}
2779	} else {
2780		/*
2781		 * When removing a non-directory we need to log the parent
2782		 * inode here.  For a directory this is done implicitly
2783		 * by the xfs_droplink call for the ".." entry.
2784		 */
2785		xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2786	}
2787	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2788
2789	/* Drop the link from dp to ip. */
2790	error = xfs_droplink(tp, ip);
2791	if (error)
2792		goto out_trans_cancel;
2793
2794	error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2795	if (error) {
2796		ASSERT(error != -ENOENT);
2797		goto out_trans_cancel;
2798	}
2799
2800	/*
2801	 * If this is a synchronous mount, make sure that the
2802	 * remove transaction goes to disk before returning to
2803	 * the user.
2804	 */
2805	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2806		xfs_trans_set_sync(tp);
2807
2808	error = xfs_trans_commit(tp);
2809	if (error)
2810		goto std_return;
2811
2812	if (is_dir && xfs_inode_is_filestream(ip))
2813		xfs_filestream_deassociate(ip);
2814
 
 
 
2815	return 0;
2816
2817 out_trans_cancel:
2818	xfs_trans_cancel(tp);
 
 
 
 
 
2819 std_return:
2820	return error;
2821}
2822
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2823/*
2824 * Enter all inodes for a rename transaction into a sorted array.
2825 */
2826#define __XFS_SORT_INODES	5
2827STATIC void
2828xfs_sort_for_rename(
2829	struct xfs_inode	*dp1,	/* in: old (source) directory inode */
2830	struct xfs_inode	*dp2,	/* in: new (target) directory inode */
2831	struct xfs_inode	*ip1,	/* in: inode of old entry */
2832	struct xfs_inode	*ip2,	/* in: inode of new entry */
2833	struct xfs_inode	*wip,	/* in: whiteout inode */
2834	struct xfs_inode	**i_tab,/* out: sorted array of inodes */
2835	int			*num_inodes)  /* in/out: inodes in array */
2836{
2837	int			i, j;
2838
2839	ASSERT(*num_inodes == __XFS_SORT_INODES);
2840	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2841
2842	/*
2843	 * i_tab contains a list of pointers to inodes.  We initialize
2844	 * the table here & we'll sort it.  We will then use it to
2845	 * order the acquisition of the inode locks.
2846	 *
2847	 * Note that the table may contain duplicates.  e.g., dp1 == dp2.
2848	 */
2849	i = 0;
2850	i_tab[i++] = dp1;
2851	i_tab[i++] = dp2;
2852	i_tab[i++] = ip1;
2853	if (ip2)
2854		i_tab[i++] = ip2;
2855	if (wip)
2856		i_tab[i++] = wip;
2857	*num_inodes = i;
2858
2859	/*
2860	 * Sort the elements via bubble sort.  (Remember, there are at
2861	 * most 5 elements to sort, so this is adequate.)
2862	 */
2863	for (i = 0; i < *num_inodes; i++) {
2864		for (j = 1; j < *num_inodes; j++) {
2865			if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2866				struct xfs_inode *temp = i_tab[j];
2867				i_tab[j] = i_tab[j-1];
2868				i_tab[j-1] = temp;
2869			}
2870		}
2871	}
2872}
2873
2874static int
2875xfs_finish_rename(
2876	struct xfs_trans	*tp)
 
2877{
2878	/*
2879	 * If this is a synchronous mount, make sure that the rename transaction
2880	 * goes to disk before returning to the user.
2881	 */
2882	if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2883		xfs_trans_set_sync(tp);
2884
2885	return xfs_trans_commit(tp);
2886}
2887
2888/*
2889 * xfs_cross_rename()
2890 *
2891 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2892 */
2893STATIC int
2894xfs_cross_rename(
2895	struct xfs_trans	*tp,
2896	struct xfs_inode	*dp1,
2897	struct xfs_name		*name1,
2898	struct xfs_inode	*ip1,
2899	struct xfs_inode	*dp2,
2900	struct xfs_name		*name2,
2901	struct xfs_inode	*ip2,
2902	int			spaceres)
2903{
2904	int		error = 0;
2905	int		ip1_flags = 0;
2906	int		ip2_flags = 0;
2907	int		dp2_flags = 0;
2908
2909	/* Swap inode number for dirent in first parent */
2910	error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2911	if (error)
2912		goto out_trans_abort;
2913
2914	/* Swap inode number for dirent in second parent */
2915	error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2916	if (error)
2917		goto out_trans_abort;
2918
2919	/*
2920	 * If we're renaming one or more directories across different parents,
2921	 * update the respective ".." entries (and link counts) to match the new
2922	 * parents.
2923	 */
2924	if (dp1 != dp2) {
2925		dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2926
2927		if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2928			error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2929						dp1->i_ino, spaceres);
2930			if (error)
2931				goto out_trans_abort;
2932
2933			/* transfer ip2 ".." reference to dp1 */
2934			if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2935				error = xfs_droplink(tp, dp2);
2936				if (error)
2937					goto out_trans_abort;
2938				xfs_bumplink(tp, dp1);
2939			}
2940
2941			/*
2942			 * Although ip1 isn't changed here, userspace needs
2943			 * to be warned about the change, so that applications
2944			 * relying on it (like backup ones), will properly
2945			 * notify the change
2946			 */
2947			ip1_flags |= XFS_ICHGTIME_CHG;
2948			ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2949		}
2950
2951		if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2952			error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2953						dp2->i_ino, spaceres);
2954			if (error)
2955				goto out_trans_abort;
2956
2957			/* transfer ip1 ".." reference to dp2 */
2958			if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2959				error = xfs_droplink(tp, dp1);
2960				if (error)
2961					goto out_trans_abort;
2962				xfs_bumplink(tp, dp2);
2963			}
2964
2965			/*
2966			 * Although ip2 isn't changed here, userspace needs
2967			 * to be warned about the change, so that applications
2968			 * relying on it (like backup ones), will properly
2969			 * notify the change
2970			 */
2971			ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2972			ip2_flags |= XFS_ICHGTIME_CHG;
2973		}
2974	}
2975
2976	if (ip1_flags) {
2977		xfs_trans_ichgtime(tp, ip1, ip1_flags);
2978		xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2979	}
2980	if (ip2_flags) {
2981		xfs_trans_ichgtime(tp, ip2, ip2_flags);
2982		xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2983	}
2984	if (dp2_flags) {
2985		xfs_trans_ichgtime(tp, dp2, dp2_flags);
2986		xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2987	}
2988	xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2989	xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2990	return xfs_finish_rename(tp);
2991
2992out_trans_abort:
2993	xfs_trans_cancel(tp);
2994	return error;
2995}
2996
2997/*
2998 * xfs_rename_alloc_whiteout()
2999 *
3000 * Return a referenced, unlinked, unlocked inode that can be used as a
3001 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3002 * crash between allocating the inode and linking it into the rename transaction
3003 * recovery will free the inode and we won't leak it.
3004 */
3005static int
3006xfs_rename_alloc_whiteout(
3007	struct user_namespace	*mnt_userns,
 
3008	struct xfs_inode	*dp,
3009	struct xfs_inode	**wip)
3010{
 
 
 
 
 
 
3011	struct xfs_inode	*tmpfile;
 
3012	int			error;
3013
3014	error = xfs_create_tmpfile(mnt_userns, dp, S_IFCHR | WHITEOUT_MODE,
3015				   &tmpfile);
3016	if (error)
3017		return error;
3018
 
 
 
 
 
 
 
 
 
3019	/*
3020	 * Prepare the tmpfile inode as if it were created through the VFS.
3021	 * Complete the inode setup and flag it as linkable.  nlink is already
3022	 * zero, so we can skip the drop_nlink.
3023	 */
3024	xfs_setup_iops(tmpfile);
3025	xfs_finish_inode_setup(tmpfile);
3026	VFS_I(tmpfile)->i_state |= I_LINKABLE;
3027
3028	*wip = tmpfile;
3029	return 0;
3030}
3031
3032/*
3033 * xfs_rename
3034 */
3035int
3036xfs_rename(
3037	struct user_namespace	*mnt_userns,
3038	struct xfs_inode	*src_dp,
3039	struct xfs_name		*src_name,
3040	struct xfs_inode	*src_ip,
3041	struct xfs_inode	*target_dp,
3042	struct xfs_name		*target_name,
3043	struct xfs_inode	*target_ip,
3044	unsigned int		flags)
3045{
 
 
 
 
 
 
 
 
 
 
 
3046	struct xfs_mount	*mp = src_dp->i_mount;
3047	struct xfs_trans	*tp;
3048	struct xfs_inode	*wip = NULL;		/* whiteout inode */
3049	struct xfs_inode	*inodes[__XFS_SORT_INODES];
3050	int			i;
3051	int			num_inodes = __XFS_SORT_INODES;
3052	bool			new_parent = (src_dp != target_dp);
3053	bool			src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3054	int			spaceres;
3055	int			error;
 
3056
3057	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3058
3059	if ((flags & RENAME_EXCHANGE) && !target_ip)
3060		return -EINVAL;
3061
3062	/*
3063	 * If we are doing a whiteout operation, allocate the whiteout inode
3064	 * we will be placing at the target and ensure the type is set
3065	 * appropriately.
3066	 */
3067	if (flags & RENAME_WHITEOUT) {
3068		ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3069		error = xfs_rename_alloc_whiteout(mnt_userns, target_dp, &wip);
3070		if (error)
3071			return error;
3072
3073		/* setup target dirent info as whiteout */
3074		src_name->type = XFS_DIR3_FT_CHRDEV;
3075	}
3076
3077	xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3078				inodes, &num_inodes);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3079
3080	spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
 
 
 
3081	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3082	if (error == -ENOSPC) {
 
3083		spaceres = 0;
3084		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3085				&tp);
3086	}
3087	if (error)
3088		goto out_release_wip;
 
 
 
 
 
 
 
 
 
 
3089
3090	/*
3091	 * Attach the dquots to the inodes
3092	 */
3093	error = xfs_qm_vop_rename_dqattach(inodes);
3094	if (error)
3095		goto out_trans_cancel;
 
 
3096
3097	/*
3098	 * Lock all the participating inodes. Depending upon whether
3099	 * the target_name exists in the target directory, and
3100	 * whether the target directory is the same as the source
3101	 * directory, we can lock from 2 to 4 inodes.
3102	 */
3103	xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3104
3105	/*
3106	 * Join all the inodes to the transaction. From this point on,
3107	 * we can rely on either trans_commit or trans_cancel to unlock
3108	 * them.
3109	 */
3110	xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3111	if (new_parent)
3112		xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3113	xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3114	if (target_ip)
3115		xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3116	if (wip)
3117		xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3118
3119	/*
3120	 * If we are using project inheritance, we only allow renames
3121	 * into our tree when the project IDs are the same; else the
3122	 * tree quota mechanism would be circumvented.
3123	 */
3124	if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
3125		     target_dp->i_projid != src_ip->i_projid)) {
3126		error = -EXDEV;
3127		goto out_trans_cancel;
3128	}
3129
3130	/* RENAME_EXCHANGE is unique from here on. */
3131	if (flags & RENAME_EXCHANGE)
3132		return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3133					target_dp, target_name, target_ip,
3134					spaceres);
 
 
 
3135
3136	/*
3137	 * Check for expected errors before we dirty the transaction
3138	 * so we can return an error without a transaction abort.
3139	 *
3140	 * Extent count overflow check:
3141	 *
3142	 * From the perspective of src_dp, a rename operation is essentially a
3143	 * directory entry remove operation. Hence the only place where we check
3144	 * for extent count overflow for src_dp is in
3145	 * xfs_bmap_del_extent_real(). xfs_bmap_del_extent_real() returns
3146	 * -ENOSPC when it detects a possible extent count overflow and in
3147	 * response, the higher layers of directory handling code do the
3148	 * following:
3149	 * 1. Data/Free blocks: XFS lets these blocks linger until a
3150	 *    future remove operation removes them.
3151	 * 2. Dabtree blocks: XFS swaps the blocks with the last block in the
3152	 *    Leaf space and unmaps the last block.
3153	 *
3154	 * For target_dp, there are two cases depending on whether the
3155	 * destination directory entry exists or not.
3156	 *
3157	 * When destination directory entry does not exist (i.e. target_ip ==
3158	 * NULL), extent count overflow check is performed only when transaction
3159	 * has a non-zero sized space reservation associated with it.  With a
3160	 * zero-sized space reservation, XFS allows a rename operation to
3161	 * continue only when the directory has sufficient free space in its
3162	 * data/leaf/free space blocks to hold the new entry.
3163	 *
3164	 * When destination directory entry exists (i.e. target_ip != NULL), all
3165	 * we need to do is change the inode number associated with the already
3166	 * existing entry. Hence there is no need to perform an extent count
3167	 * overflow check.
3168	 */
3169	if (target_ip == NULL) {
3170		/*
3171		 * If there's no space reservation, check the entry will
3172		 * fit before actually inserting it.
3173		 */
3174		if (!spaceres) {
3175			error = xfs_dir_canenter(tp, target_dp, target_name);
3176			if (error)
3177				goto out_trans_cancel;
3178		} else {
3179			error = xfs_iext_count_may_overflow(target_dp,
3180					XFS_DATA_FORK,
3181					XFS_IEXT_DIR_MANIP_CNT(mp));
3182			if (error)
3183				goto out_trans_cancel;
3184		}
3185	} else {
3186		/*
3187		 * If target exists and it's a directory, check that whether
3188		 * it can be destroyed.
3189		 */
3190		if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3191		    (!xfs_dir_isempty(target_ip) ||
3192		     (VFS_I(target_ip)->i_nlink > 2))) {
3193			error = -EEXIST;
3194			goto out_trans_cancel;
3195		}
 
 
 
 
 
 
 
 
3196	}
3197
3198	/*
3199	 * Lock the AGI buffers we need to handle bumping the nlink of the
3200	 * whiteout inode off the unlinked list and to handle dropping the
3201	 * nlink of the target inode.  Per locking order rules, do this in
3202	 * increasing AG order and before directory block allocation tries to
3203	 * grab AGFs because we grab AGIs before AGFs.
3204	 *
3205	 * The (vfs) caller must ensure that if src is a directory then
3206	 * target_ip is either null or an empty directory.
3207	 */
3208	for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3209		if (inodes[i] == wip ||
3210		    (inodes[i] == target_ip &&
3211		     (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3212			struct xfs_buf	*bp;
3213			xfs_agnumber_t	agno;
3214
3215			agno = XFS_INO_TO_AGNO(mp, inodes[i]->i_ino);
3216			error = xfs_read_agi(mp, tp, agno, &bp);
 
 
3217			if (error)
3218				goto out_trans_cancel;
3219		}
3220	}
3221
3222	/*
3223	 * Directory entry creation below may acquire the AGF. Remove
3224	 * the whiteout from the unlinked list first to preserve correct
3225	 * AGI/AGF locking order. This dirties the transaction so failures
3226	 * after this point will abort and log recovery will clean up the
3227	 * mess.
3228	 *
3229	 * For whiteouts, we need to bump the link count on the whiteout
3230	 * inode. After this point, we have a real link, clear the tmpfile
3231	 * state flag from the inode so it doesn't accidentally get misused
3232	 * in future.
3233	 */
3234	if (wip) {
3235		struct xfs_perag	*pag;
3236
3237		ASSERT(VFS_I(wip)->i_nlink == 0);
3238
3239		pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3240		error = xfs_iunlink_remove(tp, pag, wip);
3241		xfs_perag_put(pag);
3242		if (error)
3243			goto out_trans_cancel;
3244
3245		xfs_bumplink(tp, wip);
3246		VFS_I(wip)->i_state &= ~I_LINKABLE;
3247	}
3248
3249	/*
3250	 * Set up the target.
3251	 */
3252	if (target_ip == NULL) {
3253		/*
3254		 * If target does not exist and the rename crosses
3255		 * directories, adjust the target directory link count
3256		 * to account for the ".." reference from the new entry.
3257		 */
3258		error = xfs_dir_createname(tp, target_dp, target_name,
3259					   src_ip->i_ino, spaceres);
3260		if (error)
3261			goto out_trans_cancel;
3262
3263		xfs_trans_ichgtime(tp, target_dp,
3264					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3265
3266		if (new_parent && src_is_directory) {
3267			xfs_bumplink(tp, target_dp);
3268		}
3269	} else { /* target_ip != NULL */
3270		/*
3271		 * Link the source inode under the target name.
3272		 * If the source inode is a directory and we are moving
3273		 * it across directories, its ".." entry will be
3274		 * inconsistent until we replace that down below.
3275		 *
3276		 * In case there is already an entry with the same
3277		 * name at the destination directory, remove it first.
3278		 */
3279		error = xfs_dir_replace(tp, target_dp, target_name,
3280					src_ip->i_ino, spaceres);
3281		if (error)
3282			goto out_trans_cancel;
3283
3284		xfs_trans_ichgtime(tp, target_dp,
3285					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3286
3287		/*
3288		 * Decrement the link count on the target since the target
3289		 * dir no longer points to it.
3290		 */
3291		error = xfs_droplink(tp, target_ip);
3292		if (error)
3293			goto out_trans_cancel;
3294
3295		if (src_is_directory) {
3296			/*
3297			 * Drop the link from the old "." entry.
3298			 */
3299			error = xfs_droplink(tp, target_ip);
3300			if (error)
3301				goto out_trans_cancel;
3302		}
3303	} /* target_ip != NULL */
3304
3305	/*
3306	 * Remove the source.
3307	 */
3308	if (new_parent && src_is_directory) {
3309		/*
3310		 * Rewrite the ".." entry to point to the new
3311		 * directory.
3312		 */
3313		error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3314					target_dp->i_ino, spaceres);
3315		ASSERT(error != -EEXIST);
3316		if (error)
3317			goto out_trans_cancel;
3318	}
3319
 
3320	/*
3321	 * We always want to hit the ctime on the source inode.
3322	 *
3323	 * This isn't strictly required by the standards since the source
3324	 * inode isn't really being changed, but old unix file systems did
3325	 * it and some incremental backup programs won't work without it.
3326	 */
3327	xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3328	xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3329
3330	/*
3331	 * Adjust the link count on src_dp.  This is necessary when
3332	 * renaming a directory, either within one parent when
3333	 * the target existed, or across two parent directories.
3334	 */
3335	if (src_is_directory && (new_parent || target_ip != NULL)) {
3336
3337		/*
3338		 * Decrement link count on src_directory since the
3339		 * entry that's moved no longer points to it.
3340		 */
3341		error = xfs_droplink(tp, src_dp);
3342		if (error)
3343			goto out_trans_cancel;
3344	}
3345
3346	/*
3347	 * For whiteouts, we only need to update the source dirent with the
3348	 * inode number of the whiteout inode rather than removing it
3349	 * altogether.
3350	 */
3351	if (wip) {
3352		error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3353					spaceres);
3354	} else {
3355		/*
3356		 * NOTE: We don't need to check for extent count overflow here
3357		 * because the dir remove name code will leave the dir block in
3358		 * place if the extent count would overflow.
3359		 */
3360		error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3361					   spaceres);
3362	}
3363
3364	if (error)
3365		goto out_trans_cancel;
3366
3367	xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3368	xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3369	if (new_parent)
3370		xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3371
3372	error = xfs_finish_rename(tp);
3373	if (wip)
3374		xfs_irele(wip);
3375	return error;
3376
3377out_trans_cancel:
3378	xfs_trans_cancel(tp);
 
 
 
 
 
 
 
 
3379out_release_wip:
3380	if (wip)
3381		xfs_irele(wip);
 
 
3382	return error;
3383}
3384
3385static int
3386xfs_iflush(
3387	struct xfs_inode	*ip,
3388	struct xfs_buf		*bp)
3389{
3390	struct xfs_inode_log_item *iip = ip->i_itemp;
3391	struct xfs_dinode	*dip;
3392	struct xfs_mount	*mp = ip->i_mount;
3393	int			error;
3394
3395	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3396	ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3397	ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3398	       ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3399	ASSERT(iip->ili_item.li_buf == bp);
3400
3401	dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3402
3403	/*
3404	 * We don't flush the inode if any of the following checks fail, but we
3405	 * do still update the log item and attach to the backing buffer as if
3406	 * the flush happened. This is a formality to facilitate predictable
3407	 * error handling as the caller will shutdown and fail the buffer.
3408	 */
3409	error = -EFSCORRUPTED;
3410	if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3411			       mp, XFS_ERRTAG_IFLUSH_1)) {
3412		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3413			"%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3414			__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3415		goto flush_out;
3416	}
3417	if (S_ISREG(VFS_I(ip)->i_mode)) {
3418		if (XFS_TEST_ERROR(
3419		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3420		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3421		    mp, XFS_ERRTAG_IFLUSH_3)) {
3422			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3423				"%s: Bad regular inode %Lu, ptr "PTR_FMT,
3424				__func__, ip->i_ino, ip);
3425			goto flush_out;
3426		}
3427	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3428		if (XFS_TEST_ERROR(
3429		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3430		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3431		    ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3432		    mp, XFS_ERRTAG_IFLUSH_4)) {
3433			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3434				"%s: Bad directory inode %Lu, ptr "PTR_FMT,
3435				__func__, ip->i_ino, ip);
3436			goto flush_out;
3437		}
3438	}
3439	if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp) >
3440				ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3441		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3442			"%s: detected corrupt incore inode %Lu, "
3443			"total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3444			__func__, ip->i_ino,
3445			ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp),
3446			ip->i_nblocks, ip);
3447		goto flush_out;
3448	}
3449	if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3450				mp, XFS_ERRTAG_IFLUSH_6)) {
3451		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3452			"%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3453			__func__, ip->i_ino, ip->i_forkoff, ip);
3454		goto flush_out;
3455	}
3456
3457	/*
3458	 * Inode item log recovery for v2 inodes are dependent on the flushiter
3459	 * count for correct sequencing.  We bump the flush iteration count so
3460	 * we can detect flushes which postdate a log record during recovery.
3461	 * This is redundant as we now log every change and hence this can't
3462	 * happen but we need to still do it to ensure backwards compatibility
3463	 * with old kernels that predate logging all inode changes.
3464	 */
3465	if (!xfs_sb_version_has_v3inode(&mp->m_sb))
3466		ip->i_flushiter++;
3467
3468	/*
3469	 * If there are inline format data / attr forks attached to this inode,
3470	 * make sure they are not corrupt.
3471	 */
3472	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3473	    xfs_ifork_verify_local_data(ip))
3474		goto flush_out;
3475	if (ip->i_afp && ip->i_afp->if_format == XFS_DINODE_FMT_LOCAL &&
 
3476	    xfs_ifork_verify_local_attr(ip))
3477		goto flush_out;
3478
3479	/*
3480	 * Copy the dirty parts of the inode into the on-disk inode.  We always
3481	 * copy out the core of the inode, because if the inode is dirty at all
3482	 * the core must be.
3483	 */
3484	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3485
3486	/* Wrap, we never let the log put out DI_MAX_FLUSH */
3487	if (!xfs_sb_version_has_v3inode(&mp->m_sb)) {
3488		if (ip->i_flushiter == DI_MAX_FLUSH)
3489			ip->i_flushiter = 0;
3490	}
3491
3492	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3493	if (XFS_IFORK_Q(ip))
3494		xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3495
3496	/*
3497	 * We've recorded everything logged in the inode, so we'd like to clear
3498	 * the ili_fields bits so we don't log and flush things unnecessarily.
3499	 * However, we can't stop logging all this information until the data
3500	 * we've copied into the disk buffer is written to disk.  If we did we
3501	 * might overwrite the copy of the inode in the log with all the data
3502	 * after re-logging only part of it, and in the face of a crash we
3503	 * wouldn't have all the data we need to recover.
3504	 *
3505	 * What we do is move the bits to the ili_last_fields field.  When
3506	 * logging the inode, these bits are moved back to the ili_fields field.
3507	 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3508	 * we know that the information those bits represent is permanently on
3509	 * disk.  As long as the flush completes before the inode is logged
3510	 * again, then both ili_fields and ili_last_fields will be cleared.
3511	 */
3512	error = 0;
3513flush_out:
3514	spin_lock(&iip->ili_lock);
3515	iip->ili_last_fields = iip->ili_fields;
3516	iip->ili_fields = 0;
3517	iip->ili_fsync_fields = 0;
 
3518	spin_unlock(&iip->ili_lock);
3519
3520	/*
3521	 * Store the current LSN of the inode so that we can tell whether the
3522	 * item has moved in the AIL from xfs_buf_inode_iodone().
3523	 */
3524	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3525				&iip->ili_item.li_lsn);
3526
3527	/* generate the checksum. */
3528	xfs_dinode_calc_crc(mp, dip);
 
 
3529	return error;
3530}
3531
3532/*
3533 * Non-blocking flush of dirty inode metadata into the backing buffer.
3534 *
3535 * The caller must have a reference to the inode and hold the cluster buffer
3536 * locked. The function will walk across all the inodes on the cluster buffer it
3537 * can find and lock without blocking, and flush them to the cluster buffer.
3538 *
3539 * On successful flushing of at least one inode, the caller must write out the
3540 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3541 * the caller needs to release the buffer. On failure, the filesystem will be
3542 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3543 * will be returned.
3544 */
3545int
3546xfs_iflush_cluster(
3547	struct xfs_buf		*bp)
3548{
3549	struct xfs_mount	*mp = bp->b_mount;
3550	struct xfs_log_item	*lip, *n;
3551	struct xfs_inode	*ip;
3552	struct xfs_inode_log_item *iip;
3553	int			clcount = 0;
3554	int			error = 0;
3555
3556	/*
3557	 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3558	 * can remove itself from the list.
3559	 */
3560	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3561		iip = (struct xfs_inode_log_item *)lip;
3562		ip = iip->ili_inode;
3563
3564		/*
3565		 * Quick and dirty check to avoid locks if possible.
3566		 */
3567		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3568			continue;
3569		if (xfs_ipincount(ip))
3570			continue;
3571
3572		/*
3573		 * The inode is still attached to the buffer, which means it is
3574		 * dirty but reclaim might try to grab it. Check carefully for
3575		 * that, and grab the ilock while still holding the i_flags_lock
3576		 * to guarantee reclaim will not be able to reclaim this inode
3577		 * once we drop the i_flags_lock.
3578		 */
3579		spin_lock(&ip->i_flags_lock);
3580		ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3581		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3582			spin_unlock(&ip->i_flags_lock);
3583			continue;
3584		}
3585
3586		/*
3587		 * ILOCK will pin the inode against reclaim and prevent
3588		 * concurrent transactions modifying the inode while we are
3589		 * flushing the inode. If we get the lock, set the flushing
3590		 * state before we drop the i_flags_lock.
3591		 */
3592		if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3593			spin_unlock(&ip->i_flags_lock);
3594			continue;
3595		}
3596		__xfs_iflags_set(ip, XFS_IFLUSHING);
3597		spin_unlock(&ip->i_flags_lock);
3598
3599		/*
3600		 * Abort flushing this inode if we are shut down because the
3601		 * inode may not currently be in the AIL. This can occur when
3602		 * log I/O failure unpins the inode without inserting into the
3603		 * AIL, leaving a dirty/unpinned inode attached to the buffer
3604		 * that otherwise looks like it should be flushed.
3605		 */
3606		if (XFS_FORCED_SHUTDOWN(mp)) {
3607			xfs_iunpin_wait(ip);
3608			xfs_iflush_abort(ip);
3609			xfs_iunlock(ip, XFS_ILOCK_SHARED);
3610			error = -EIO;
3611			continue;
3612		}
3613
3614		/* don't block waiting on a log force to unpin dirty inodes */
3615		if (xfs_ipincount(ip)) {
3616			xfs_iflags_clear(ip, XFS_IFLUSHING);
3617			xfs_iunlock(ip, XFS_ILOCK_SHARED);
3618			continue;
3619		}
3620
3621		if (!xfs_inode_clean(ip))
3622			error = xfs_iflush(ip, bp);
3623		else
3624			xfs_iflags_clear(ip, XFS_IFLUSHING);
3625		xfs_iunlock(ip, XFS_ILOCK_SHARED);
3626		if (error)
3627			break;
3628		clcount++;
3629	}
3630
3631	if (error) {
 
 
 
 
 
 
 
 
 
 
 
3632		bp->b_flags |= XBF_ASYNC;
3633		xfs_buf_ioend_fail(bp);
3634		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3635		return error;
3636	}
3637
3638	if (!clcount)
3639		return -EAGAIN;
3640
3641	XFS_STATS_INC(mp, xs_icluster_flushcnt);
3642	XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3643	return 0;
3644
3645}
3646
3647/* Release an inode. */
3648void
3649xfs_irele(
3650	struct xfs_inode	*ip)
3651{
3652	trace_xfs_irele(ip, _RET_IP_);
3653	iput(VFS_I(ip));
3654}
3655
3656/*
3657 * Ensure all commited transactions touching the inode are written to the log.
3658 */
3659int
3660xfs_log_force_inode(
3661	struct xfs_inode	*ip)
3662{
3663	xfs_csn_t		seq = 0;
3664
3665	xfs_ilock(ip, XFS_ILOCK_SHARED);
3666	if (xfs_ipincount(ip))
3667		seq = ip->i_itemp->ili_commit_seq;
3668	xfs_iunlock(ip, XFS_ILOCK_SHARED);
3669
3670	if (!seq)
3671		return 0;
3672	return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3673}
3674
3675/*
3676 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3677 * abide vfs locking order (lowest pointer value goes first) and breaking the
3678 * layout leases before proceeding.  The loop is needed because we cannot call
3679 * the blocking break_layout() with the iolocks held, and therefore have to
3680 * back out both locks.
3681 */
3682static int
3683xfs_iolock_two_inodes_and_break_layout(
3684	struct inode		*src,
3685	struct inode		*dest)
3686{
3687	int			error;
3688
3689	if (src > dest)
3690		swap(src, dest);
3691
3692retry:
3693	/* Wait to break both inodes' layouts before we start locking. */
3694	error = break_layout(src, true);
3695	if (error)
3696		return error;
3697	if (src != dest) {
3698		error = break_layout(dest, true);
3699		if (error)
3700			return error;
3701	}
3702
3703	/* Lock one inode and make sure nobody got in and leased it. */
3704	inode_lock(src);
3705	error = break_layout(src, false);
3706	if (error) {
3707		inode_unlock(src);
3708		if (error == -EWOULDBLOCK)
3709			goto retry;
3710		return error;
3711	}
3712
3713	if (src == dest)
3714		return 0;
3715
3716	/* Lock the other inode and make sure nobody got in and leased it. */
3717	inode_lock_nested(dest, I_MUTEX_NONDIR2);
3718	error = break_layout(dest, false);
3719	if (error) {
3720		inode_unlock(src);
3721		inode_unlock(dest);
3722		if (error == -EWOULDBLOCK)
3723			goto retry;
3724		return error;
3725	}
3726
3727	return 0;
3728}
3729
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3730/*
3731 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3732 * mmap activity.
3733 */
3734int
3735xfs_ilock2_io_mmap(
3736	struct xfs_inode	*ip1,
3737	struct xfs_inode	*ip2)
3738{
3739	int			ret;
3740
3741	ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3742	if (ret)
3743		return ret;
3744	if (ip1 == ip2)
3745		xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3746	else
3747		xfs_lock_two_inodes(ip1, XFS_MMAPLOCK_EXCL,
3748				    ip2, XFS_MMAPLOCK_EXCL);
 
 
 
 
 
 
 
 
3749	return 0;
3750}
3751
3752/* Unlock both inodes to allow IO and mmap activity. */
3753void
3754xfs_iunlock2_io_mmap(
3755	struct xfs_inode	*ip1,
3756	struct xfs_inode	*ip2)
3757{
3758	bool			same_inode = (ip1 == ip2);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3759
 
 
3760	xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3761	if (!same_inode)
3762		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
 
3763	inode_unlock(VFS_I(ip2));
3764	if (!same_inode)
3765		inode_unlock(VFS_I(ip1));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3766}