Linux Audio

Check our new training course

Loading...
v5.4
   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_sb.h"
  15#include "xfs_mount.h"
  16#include "xfs_defer.h"
  17#include "xfs_inode.h"
  18#include "xfs_dir2.h"
  19#include "xfs_attr.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_ialloc.h"
  25#include "xfs_bmap.h"
  26#include "xfs_bmap_util.h"
  27#include "xfs_errortag.h"
  28#include "xfs_error.h"
  29#include "xfs_quota.h"
  30#include "xfs_filestream.h"
  31#include "xfs_trace.h"
  32#include "xfs_icache.h"
  33#include "xfs_symlink.h"
  34#include "xfs_trans_priv.h"
  35#include "xfs_log.h"
  36#include "xfs_bmap_btree.h"
  37#include "xfs_reflink.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_iflush_int(struct xfs_inode *, struct xfs_buf *);
  48STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
  49STATIC int xfs_iunlink_remove(struct xfs_trans *, 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	if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
  59		return ip->i_d.di_extsize;
 
 
 
 
 
 
  60	if (XFS_IS_REALTIME_INODE(ip))
  61		return ip->i_mount->m_sb.sb_rextsize;
  62	return 0;
  63}
  64
  65/*
  66 * Helper function to extract CoW extent size hint from inode.
  67 * Between the extent size hint and the CoW extent size hint, we
  68 * return the greater of the two.  If the value is zero (automatic),
  69 * use the default size.
  70 */
  71xfs_extlen_t
  72xfs_get_cowextsz_hint(
  73	struct xfs_inode	*ip)
  74{
  75	xfs_extlen_t		a, b;
  76
  77	a = 0;
  78	if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
  79		a = ip->i_d.di_cowextsize;
  80	b = xfs_get_extsz_hint(ip);
  81
  82	a = max(a, b);
  83	if (a == 0)
  84		return XFS_DEFAULT_COWEXTSZ_HINT;
  85	return a;
  86}
  87
  88/*
  89 * These two are wrapper routines around the xfs_ilock() routine used to
  90 * centralize some grungy code.  They are used in places that wish to lock the
  91 * inode solely for reading the extents.  The reason these places can't just
  92 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
  93 * bringing in of the extents from disk for a file in b-tree format.  If the
  94 * inode is in b-tree format, then we need to lock the inode exclusively until
  95 * the extents are read in.  Locking it exclusively all the time would limit
  96 * our parallelism unnecessarily, though.  What we do instead is check to see
  97 * if the extents have been read in yet, and only lock the inode exclusively
  98 * if they have not.
  99 *
 100 * The functions return a value which should be given to the corresponding
 101 * xfs_iunlock() call.
 102 */
 103uint
 104xfs_ilock_data_map_shared(
 105	struct xfs_inode	*ip)
 106{
 107	uint			lock_mode = XFS_ILOCK_SHARED;
 108
 109	if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
 110	    (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
 111		lock_mode = XFS_ILOCK_EXCL;
 112	xfs_ilock(ip, lock_mode);
 113	return lock_mode;
 114}
 115
 116uint
 117xfs_ilock_attr_map_shared(
 118	struct xfs_inode	*ip)
 119{
 120	uint			lock_mode = XFS_ILOCK_SHARED;
 121
 122	if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
 123	    (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
 124		lock_mode = XFS_ILOCK_EXCL;
 125	xfs_ilock(ip, lock_mode);
 126	return lock_mode;
 127}
 128
 129/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 130 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
 131 * multi-reader locks: i_mmap_lock and the i_lock.  This routine allows
 132 * various combinations of the locks to be obtained.
 133 *
 134 * The 3 locks should always be ordered so that the IO lock is obtained first,
 135 * the mmap lock second and the ilock last in order to prevent deadlock.
 136 *
 137 * Basic locking order:
 138 *
 139 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
 140 *
 141 * mmap_sem locking order:
 142 *
 143 * i_rwsem -> page lock -> mmap_sem
 144 * mmap_sem -> i_mmap_lock -> page_lock
 145 *
 146 * The difference in mmap_sem locking order mean that we cannot hold the
 147 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
 148 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
 149 * in get_user_pages() to map the user pages into the kernel address space for
 150 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
 151 * page faults already hold the mmap_sem.
 152 *
 153 * Hence to serialise fully against both syscall and mmap based IO, we need to
 154 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
 155 * taken in places where we need to invalidate the page cache in a race
 156 * free manner (e.g. truncate, hole punch and other extent manipulation
 157 * functions).
 158 */
 159void
 160xfs_ilock(
 161	xfs_inode_t		*ip,
 162	uint			lock_flags)
 163{
 164	trace_xfs_ilock(ip, lock_flags, _RET_IP_);
 165
 166	/*
 167	 * You can't set both SHARED and EXCL for the same lock,
 168	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 169	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 170	 */
 171	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 172	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 173	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 174	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 175	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 176	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 177	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 178
 179	if (lock_flags & XFS_IOLOCK_EXCL) {
 180		down_write_nested(&VFS_I(ip)->i_rwsem,
 181				  XFS_IOLOCK_DEP(lock_flags));
 182	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 183		down_read_nested(&VFS_I(ip)->i_rwsem,
 184				 XFS_IOLOCK_DEP(lock_flags));
 185	}
 186
 187	if (lock_flags & XFS_MMAPLOCK_EXCL)
 188		mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
 189	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 190		mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
 
 
 
 191
 192	if (lock_flags & XFS_ILOCK_EXCL)
 193		mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 194	else if (lock_flags & XFS_ILOCK_SHARED)
 195		mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 196}
 197
 198/*
 199 * This is just like xfs_ilock(), except that the caller
 200 * is guaranteed not to sleep.  It returns 1 if it gets
 201 * the requested locks and 0 otherwise.  If the IO lock is
 202 * obtained but the inode lock cannot be, then the IO lock
 203 * is dropped before returning.
 204 *
 205 * ip -- the inode being locked
 206 * lock_flags -- this parameter indicates the inode's locks to be
 207 *       to be locked.  See the comment for xfs_ilock() for a list
 208 *	 of valid values.
 209 */
 210int
 211xfs_ilock_nowait(
 212	xfs_inode_t		*ip,
 213	uint			lock_flags)
 214{
 215	trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
 216
 217	/*
 218	 * You can't set both SHARED and EXCL for the same lock,
 219	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 220	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 221	 */
 222	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 223	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 224	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 225	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 226	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 227	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 228	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 229
 230	if (lock_flags & XFS_IOLOCK_EXCL) {
 231		if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
 232			goto out;
 233	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 234		if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
 235			goto out;
 236	}
 237
 238	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 239		if (!mrtryupdate(&ip->i_mmaplock))
 240			goto out_undo_iolock;
 241	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 242		if (!mrtryaccess(&ip->i_mmaplock))
 243			goto out_undo_iolock;
 244	}
 245
 246	if (lock_flags & XFS_ILOCK_EXCL) {
 247		if (!mrtryupdate(&ip->i_lock))
 248			goto out_undo_mmaplock;
 249	} else if (lock_flags & XFS_ILOCK_SHARED) {
 250		if (!mrtryaccess(&ip->i_lock))
 251			goto out_undo_mmaplock;
 252	}
 253	return 1;
 254
 255out_undo_mmaplock:
 256	if (lock_flags & XFS_MMAPLOCK_EXCL)
 257		mrunlock_excl(&ip->i_mmaplock);
 258	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 259		mrunlock_shared(&ip->i_mmaplock);
 260out_undo_iolock:
 261	if (lock_flags & XFS_IOLOCK_EXCL)
 262		up_write(&VFS_I(ip)->i_rwsem);
 263	else if (lock_flags & XFS_IOLOCK_SHARED)
 264		up_read(&VFS_I(ip)->i_rwsem);
 265out:
 266	return 0;
 267}
 268
 269/*
 270 * xfs_iunlock() is used to drop the inode locks acquired with
 271 * xfs_ilock() and xfs_ilock_nowait().  The caller must pass
 272 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
 273 * that we know which locks to drop.
 274 *
 275 * ip -- the inode being unlocked
 276 * lock_flags -- this parameter indicates the inode's locks to be
 277 *       to be unlocked.  See the comment for xfs_ilock() for a list
 278 *	 of valid values for this parameter.
 279 *
 280 */
 281void
 282xfs_iunlock(
 283	xfs_inode_t		*ip,
 284	uint			lock_flags)
 285{
 286	/*
 287	 * You can't set both SHARED and EXCL for the same lock,
 288	 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
 289	 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
 290	 */
 291	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 292	       (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 293	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 294	       (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 295	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 296	       (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 297	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 298	ASSERT(lock_flags != 0);
 299
 300	if (lock_flags & XFS_IOLOCK_EXCL)
 301		up_write(&VFS_I(ip)->i_rwsem);
 302	else if (lock_flags & XFS_IOLOCK_SHARED)
 303		up_read(&VFS_I(ip)->i_rwsem);
 304
 305	if (lock_flags & XFS_MMAPLOCK_EXCL)
 306		mrunlock_excl(&ip->i_mmaplock);
 307	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 308		mrunlock_shared(&ip->i_mmaplock);
 309
 310	if (lock_flags & XFS_ILOCK_EXCL)
 311		mrunlock_excl(&ip->i_lock);
 312	else if (lock_flags & XFS_ILOCK_SHARED)
 313		mrunlock_shared(&ip->i_lock);
 314
 315	trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
 316}
 317
 318/*
 319 * give up write locks.  the i/o lock cannot be held nested
 320 * if it is being demoted.
 321 */
 322void
 323xfs_ilock_demote(
 324	xfs_inode_t		*ip,
 325	uint			lock_flags)
 326{
 327	ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
 328	ASSERT((lock_flags &
 329		~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
 330
 331	if (lock_flags & XFS_ILOCK_EXCL)
 332		mrdemote(&ip->i_lock);
 333	if (lock_flags & XFS_MMAPLOCK_EXCL)
 334		mrdemote(&ip->i_mmaplock);
 335	if (lock_flags & XFS_IOLOCK_EXCL)
 336		downgrade_write(&VFS_I(ip)->i_rwsem);
 337
 338	trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
 339}
 340
 341#if defined(DEBUG) || defined(XFS_WARN)
 342int
 343xfs_isilocked(
 344	xfs_inode_t		*ip,
 345	uint			lock_flags)
 346{
 347	if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
 348		if (!(lock_flags & XFS_ILOCK_SHARED))
 349			return !!ip->i_lock.mr_writer;
 350		return rwsem_is_locked(&ip->i_lock.mr_lock);
 351	}
 352
 353	if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
 354		if (!(lock_flags & XFS_MMAPLOCK_SHARED))
 355			return !!ip->i_mmaplock.mr_writer;
 356		return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
 357	}
 358
 359	if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
 360		if (!(lock_flags & XFS_IOLOCK_SHARED))
 361			return !debug_locks ||
 362				lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
 363		return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
 364	}
 365
 366	ASSERT(0);
 367	return 0;
 368}
 369#endif
 370
 371/*
 372 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
 373 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
 374 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
 375 * errors and warnings.
 376 */
 377#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
 378static bool
 379xfs_lockdep_subclass_ok(
 380	int subclass)
 381{
 382	return subclass < MAX_LOCKDEP_SUBCLASSES;
 383}
 384#else
 385#define xfs_lockdep_subclass_ok(subclass)	(true)
 386#endif
 387
 388/*
 389 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
 390 * value. This can be called for any type of inode lock combination, including
 391 * parent locking. Care must be taken to ensure we don't overrun the subclass
 392 * storage fields in the class mask we build.
 393 */
 394static inline int
 395xfs_lock_inumorder(int lock_mode, int subclass)
 
 
 396{
 397	int	class = 0;
 398
 399	ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
 400			      XFS_ILOCK_RTSUM)));
 401	ASSERT(xfs_lockdep_subclass_ok(subclass));
 402
 403	if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
 404		ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
 405		class += subclass << XFS_IOLOCK_SHIFT;
 406	}
 407
 408	if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
 409		ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
 410		class += subclass << XFS_MMAPLOCK_SHIFT;
 411	}
 412
 413	if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
 414		ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
 415		class += subclass << XFS_ILOCK_SHIFT;
 416	}
 417
 418	return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
 419}
 420
 421/*
 422 * The following routine will lock n inodes in exclusive mode.  We assume the
 423 * caller calls us with the inodes in i_ino order.
 424 *
 425 * We need to detect deadlock where an inode that we lock is in the AIL and we
 426 * start waiting for another inode that is locked by a thread in a long running
 427 * transaction (such as truncate). This can result in deadlock since the long
 428 * running trans might need to wait for the inode we just locked in order to
 429 * push the tail and free space in the log.
 430 *
 431 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
 432 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
 433 * lock more than one at a time, lockdep will report false positives saying we
 434 * have violated locking orders.
 435 */
 436static void
 437xfs_lock_inodes(
 438	struct xfs_inode	**ips,
 439	int			inodes,
 440	uint			lock_mode)
 441{
 442	int			attempts = 0, i, j, try_lock;
 
 
 
 443	struct xfs_log_item	*lp;
 444
 445	/*
 446	 * Currently supports between 2 and 5 inodes with exclusive locking.  We
 447	 * support an arbitrary depth of locking here, but absolute limits on
 448	 * inodes depend on the the type of locking and the limits placed by
 449	 * lockdep annotations in xfs_lock_inumorder.  These are all checked by
 450	 * the asserts.
 451	 */
 452	ASSERT(ips && inodes >= 2 && inodes <= 5);
 453	ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
 454			    XFS_ILOCK_EXCL));
 455	ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
 456			      XFS_ILOCK_SHARED)));
 457	ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
 458		inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
 459	ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
 460		inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
 461
 462	if (lock_mode & XFS_IOLOCK_EXCL) {
 463		ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
 464	} else if (lock_mode & XFS_MMAPLOCK_EXCL)
 465		ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
 466
 467	try_lock = 0;
 468	i = 0;
 469again:
 
 
 470	for (; i < inodes; i++) {
 471		ASSERT(ips[i]);
 472
 473		if (i && (ips[i] == ips[i - 1]))	/* Already locked */
 474			continue;
 475
 476		/*
 477		 * If try_lock is not set yet, make sure all locked inodes are
 478		 * not in the AIL.  If any are, set try_lock to be used later.
 479		 */
 480		if (!try_lock) {
 481			for (j = (i - 1); j >= 0 && !try_lock; j--) {
 482				lp = &ips[j]->i_itemp->ili_item;
 483				if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
 484					try_lock++;
 485			}
 486		}
 487
 488		/*
 489		 * If any of the previous locks we have locked is in the AIL,
 490		 * we must TRY to get the second and subsequent locks. If
 491		 * we can't get any, we must release all we have
 492		 * and try again.
 493		 */
 494		if (!try_lock) {
 495			xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
 496			continue;
 497		}
 498
 499		/* try_lock means we have an inode locked that is in the AIL. */
 500		ASSERT(i != 0);
 501		if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
 502			continue;
 503
 504		/*
 505		 * Unlock all previous guys and try again.  xfs_iunlock will try
 506		 * to push the tail if the inode is in the AIL.
 507		 */
 508		attempts++;
 509		for (j = i - 1; j >= 0; j--) {
 510			/*
 511			 * Check to see if we've already unlocked this one.  Not
 512			 * the first one going back, and the inode ptr is the
 513			 * same.
 514			 */
 515			if (j != (i - 1) && ips[j] == ips[j + 1])
 516				continue;
 517
 518			xfs_iunlock(ips[j], lock_mode);
 519		}
 520
 521		if ((attempts % 5) == 0) {
 522			delay(1); /* Don't just spin the CPU */
 523		}
 524		i = 0;
 525		try_lock = 0;
 526		goto again;
 527	}
 528}
 529
 530/*
 531 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
 532 * the mmaplock or the ilock, but not more than one type at a time. If we lock
 533 * more than one at a time, lockdep will report false positives saying we have
 534 * violated locking orders.  The iolock must be double-locked separately since
 535 * we use i_rwsem for that.  We now support taking one lock EXCL and the other
 536 * SHARED.
 537 */
 538void
 539xfs_lock_two_inodes(
 540	struct xfs_inode	*ip0,
 541	uint			ip0_mode,
 542	struct xfs_inode	*ip1,
 543	uint			ip1_mode)
 544{
 545	struct xfs_inode	*temp;
 546	uint			mode_temp;
 547	int			attempts = 0;
 548	struct xfs_log_item	*lp;
 549
 550	ASSERT(hweight32(ip0_mode) == 1);
 551	ASSERT(hweight32(ip1_mode) == 1);
 552	ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 553	ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 554	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 555	       !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 556	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 557	       !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 558	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 559	       !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 560	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
 561	       !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
 562
 563	ASSERT(ip0->i_ino != ip1->i_ino);
 564
 565	if (ip0->i_ino > ip1->i_ino) {
 566		temp = ip0;
 567		ip0 = ip1;
 568		ip1 = temp;
 569		mode_temp = ip0_mode;
 570		ip0_mode = ip1_mode;
 571		ip1_mode = mode_temp;
 572	}
 573
 574 again:
 575	xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
 576
 577	/*
 578	 * If the first lock we have locked is in the AIL, we must TRY to get
 579	 * the second lock. If we can't get it, we must release the first one
 580	 * and try again.
 581	 */
 582	lp = &ip0->i_itemp->ili_item;
 583	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
 584		if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
 585			xfs_iunlock(ip0, ip0_mode);
 586			if ((++attempts % 5) == 0)
 587				delay(1); /* Don't just spin the CPU */
 588			goto again;
 589		}
 590	} else {
 591		xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
 592	}
 593}
 594
 595void
 596__xfs_iflock(
 597	struct xfs_inode	*ip)
 598{
 599	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
 600	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
 601
 602	do {
 603		prepare_to_wait_exclusive(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
 604		if (xfs_isiflocked(ip))
 605			io_schedule();
 606	} while (!xfs_iflock_nowait(ip));
 607
 608	finish_wait(wq, &wait.wq_entry);
 609}
 610
 611STATIC uint
 612_xfs_dic2xflags(
 613	uint16_t		di_flags,
 614	uint64_t		di_flags2,
 615	bool			has_attr)
 616{
 617	uint			flags = 0;
 618
 619	if (di_flags & XFS_DIFLAG_ANY) {
 620		if (di_flags & XFS_DIFLAG_REALTIME)
 621			flags |= FS_XFLAG_REALTIME;
 622		if (di_flags & XFS_DIFLAG_PREALLOC)
 623			flags |= FS_XFLAG_PREALLOC;
 624		if (di_flags & XFS_DIFLAG_IMMUTABLE)
 625			flags |= FS_XFLAG_IMMUTABLE;
 626		if (di_flags & XFS_DIFLAG_APPEND)
 627			flags |= FS_XFLAG_APPEND;
 628		if (di_flags & XFS_DIFLAG_SYNC)
 629			flags |= FS_XFLAG_SYNC;
 630		if (di_flags & XFS_DIFLAG_NOATIME)
 631			flags |= FS_XFLAG_NOATIME;
 632		if (di_flags & XFS_DIFLAG_NODUMP)
 633			flags |= FS_XFLAG_NODUMP;
 634		if (di_flags & XFS_DIFLAG_RTINHERIT)
 635			flags |= FS_XFLAG_RTINHERIT;
 636		if (di_flags & XFS_DIFLAG_PROJINHERIT)
 637			flags |= FS_XFLAG_PROJINHERIT;
 638		if (di_flags & XFS_DIFLAG_NOSYMLINKS)
 639			flags |= FS_XFLAG_NOSYMLINKS;
 640		if (di_flags & XFS_DIFLAG_EXTSIZE)
 641			flags |= FS_XFLAG_EXTSIZE;
 642		if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
 643			flags |= FS_XFLAG_EXTSZINHERIT;
 644		if (di_flags & XFS_DIFLAG_NODEFRAG)
 645			flags |= FS_XFLAG_NODEFRAG;
 646		if (di_flags & XFS_DIFLAG_FILESTREAM)
 647			flags |= FS_XFLAG_FILESTREAM;
 648	}
 649
 650	if (di_flags2 & XFS_DIFLAG2_ANY) {
 651		if (di_flags2 & XFS_DIFLAG2_DAX)
 652			flags |= FS_XFLAG_DAX;
 653		if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
 654			flags |= FS_XFLAG_COWEXTSIZE;
 655	}
 656
 657	if (has_attr)
 658		flags |= FS_XFLAG_HASATTR;
 659
 660	return flags;
 661}
 662
 663uint
 664xfs_ip2xflags(
 665	struct xfs_inode	*ip)
 666{
 667	struct xfs_icdinode	*dic = &ip->i_d;
 668
 669	return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
 670}
 671
 672/*
 673 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 674 * is allowed, otherwise it has to be an exact match. If a CI match is found,
 675 * ci_name->name will point to a the actual name (caller must free) or
 676 * will be set to NULL if an exact match is found.
 677 */
 678int
 679xfs_lookup(
 680	xfs_inode_t		*dp,
 681	struct xfs_name		*name,
 682	xfs_inode_t		**ipp,
 683	struct xfs_name		*ci_name)
 684{
 685	xfs_ino_t		inum;
 686	int			error;
 687
 688	trace_xfs_lookup(dp, name);
 689
 690	if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 
 
 691		return -EIO;
 692
 693	error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 694	if (error)
 695		goto out_unlock;
 696
 697	error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 698	if (error)
 699		goto out_free_name;
 700
 701	return 0;
 702
 703out_free_name:
 704	if (ci_name)
 705		kmem_free(ci_name->name);
 706out_unlock:
 707	*ipp = NULL;
 708	return error;
 709}
 710
 711/*
 712 * Allocate an inode on disk and return a copy of its in-core version.
 713 * The in-core inode is locked exclusively.  Set mode, nlink, and rdev
 714 * appropriately within the inode.  The uid and gid for the inode are
 715 * set according to the contents of the given cred structure.
 716 *
 717 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
 718 * has a free inode available, call xfs_iget() to obtain the in-core
 719 * version of the allocated inode.  Finally, fill in the inode and
 720 * log its initial contents.  In this case, ialloc_context would be
 721 * set to NULL.
 722 *
 723 * If xfs_dialloc() does not have an available inode, it will replenish
 724 * its supply by doing an allocation. Since we can only do one
 725 * allocation within a transaction without deadlocks, we must commit
 726 * the current transaction before returning the inode itself.
 727 * In this case, therefore, we will set ialloc_context and return.
 728 * The caller should then commit the current transaction, start a new
 729 * transaction, and call xfs_ialloc() again to actually get the inode.
 730 *
 731 * To ensure that some other process does not grab the inode that
 732 * was allocated during the first call to xfs_ialloc(), this routine
 733 * also returns the [locked] bp pointing to the head of the freelist
 734 * as ialloc_context.  The caller should hold this buffer across
 735 * the commit and pass it back into this routine on the second call.
 736 *
 737 * If we are allocating quota inodes, we do not have a parent inode
 738 * to attach to or associate with (i.e. pip == NULL) because they
 739 * are not linked into the directory structure - they are attached
 740 * directly to the superblock - and so have no parent.
 741 */
 742static int
 743xfs_ialloc(
 744	xfs_trans_t	*tp,
 745	xfs_inode_t	*pip,
 746	umode_t		mode,
 747	xfs_nlink_t	nlink,
 748	dev_t		rdev,
 749	prid_t		prid,
 750	xfs_buf_t	**ialloc_context,
 751	xfs_inode_t	**ipp)
 752{
 753	struct xfs_mount *mp = tp->t_mountp;
 754	xfs_ino_t	ino;
 755	xfs_inode_t	*ip;
 756	uint		flags;
 757	int		error;
 758	struct timespec64 tv;
 759	struct inode	*inode;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 760
 761	/*
 762	 * Call the space management code to pick
 763	 * the on-disk inode to be allocated.
 764	 */
 765	error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode,
 766			    ialloc_context, &ino);
 767	if (error)
 768		return error;
 769	if (*ialloc_context || ino == NULLFSINO) {
 770		*ipp = NULL;
 771		return 0;
 772	}
 773	ASSERT(*ialloc_context == NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 774
 775	/*
 776	 * Protect against obviously corrupt allocation btree records. Later
 777	 * xfs_iget checks will catch re-allocation of other active in-memory
 778	 * and on-disk inodes. If we don't catch reallocating the parent inode
 779	 * here we will deadlock in xfs_iget() so we have to do these checks
 780	 * first.
 781	 */
 782	if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
 783		xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
 
 
 784		return -EFSCORRUPTED;
 785	}
 786
 787	/*
 788	 * Get the in-core inode with the lock held exclusively.
 789	 * This is because we're setting fields here we need
 790	 * to prevent others from looking at until we're done.
 791	 */
 792	error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
 793			 XFS_ILOCK_EXCL, &ip);
 794	if (error)
 795		return error;
 
 796	ASSERT(ip != NULL);
 797	inode = VFS_I(ip);
 798
 799	/*
 800	 * We always convert v1 inodes to v2 now - we only support filesystems
 801	 * with >= v2 inode capability, so there is no reason for ever leaving
 802	 * an inode in v1 format.
 803	 */
 804	if (ip->i_d.di_version == 1)
 805		ip->i_d.di_version = 2;
 806
 807	inode->i_mode = mode;
 808	set_nlink(inode, nlink);
 809	ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
 810	ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
 811	inode->i_rdev = rdev;
 812	xfs_set_projid(ip, prid);
 813
 814	if (pip && XFS_INHERIT_GID(pip)) {
 815		ip->i_d.di_gid = pip->i_d.di_gid;
 816		if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
 817			inode->i_mode |= S_ISGID;
 
 
 818	}
 819
 820	/*
 821	 * If the group ID of the new file does not match the effective group
 822	 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
 823	 * (and only if the irix_sgid_inherit compatibility variable is set).
 824	 */
 825	if ((irix_sgid_inherit) &&
 826	    (inode->i_mode & S_ISGID) &&
 827	    (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
 828		inode->i_mode &= ~S_ISGID;
 829
 830	ip->i_d.di_size = 0;
 831	ip->i_d.di_nextents = 0;
 832	ASSERT(ip->i_d.di_nblocks == 0);
 833
 834	tv = current_time(inode);
 835	inode->i_mtime = tv;
 836	inode->i_atime = tv;
 837	inode->i_ctime = tv;
 838
 839	ip->i_d.di_extsize = 0;
 840	ip->i_d.di_dmevmask = 0;
 841	ip->i_d.di_dmstate = 0;
 842	ip->i_d.di_flags = 0;
 843
 844	if (ip->i_d.di_version == 3) {
 
 
 
 845		inode_set_iversion(inode, 1);
 846		ip->i_d.di_flags2 = 0;
 847		ip->i_d.di_cowextsize = 0;
 848		ip->i_d.di_crtime.t_sec = (int32_t)tv.tv_sec;
 849		ip->i_d.di_crtime.t_nsec = (int32_t)tv.tv_nsec;
 850	}
 851
 852
 853	flags = XFS_ILOG_CORE;
 854	switch (mode & S_IFMT) {
 855	case S_IFIFO:
 856	case S_IFCHR:
 857	case S_IFBLK:
 858	case S_IFSOCK:
 859		ip->i_d.di_format = XFS_DINODE_FMT_DEV;
 860		ip->i_df.if_flags = 0;
 861		flags |= XFS_ILOG_DEV;
 862		break;
 863	case S_IFREG:
 864	case S_IFDIR:
 865		if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
 866			uint		di_flags = 0;
 867
 868			if (S_ISDIR(mode)) {
 869				if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
 870					di_flags |= XFS_DIFLAG_RTINHERIT;
 871				if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
 872					di_flags |= XFS_DIFLAG_EXTSZINHERIT;
 873					ip->i_d.di_extsize = pip->i_d.di_extsize;
 874				}
 875				if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 876					di_flags |= XFS_DIFLAG_PROJINHERIT;
 877			} else if (S_ISREG(mode)) {
 878				if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
 879					di_flags |= XFS_DIFLAG_REALTIME;
 880				if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
 881					di_flags |= XFS_DIFLAG_EXTSIZE;
 882					ip->i_d.di_extsize = pip->i_d.di_extsize;
 883				}
 884			}
 885			if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
 886			    xfs_inherit_noatime)
 887				di_flags |= XFS_DIFLAG_NOATIME;
 888			if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
 889			    xfs_inherit_nodump)
 890				di_flags |= XFS_DIFLAG_NODUMP;
 891			if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
 892			    xfs_inherit_sync)
 893				di_flags |= XFS_DIFLAG_SYNC;
 894			if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
 895			    xfs_inherit_nosymlinks)
 896				di_flags |= XFS_DIFLAG_NOSYMLINKS;
 897			if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
 898			    xfs_inherit_nodefrag)
 899				di_flags |= XFS_DIFLAG_NODEFRAG;
 900			if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
 901				di_flags |= XFS_DIFLAG_FILESTREAM;
 902
 903			ip->i_d.di_flags |= di_flags;
 904		}
 905		if (pip &&
 906		    (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
 907		    pip->i_d.di_version == 3 &&
 908		    ip->i_d.di_version == 3) {
 909			uint64_t	di_flags2 = 0;
 910
 911			if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
 912				di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
 913				ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
 914			}
 915			if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
 916				di_flags2 |= XFS_DIFLAG2_DAX;
 917
 918			ip->i_d.di_flags2 |= di_flags2;
 919		}
 920		/* FALLTHROUGH */
 921	case S_IFLNK:
 922		ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
 923		ip->i_df.if_flags = XFS_IFEXTENTS;
 924		ip->i_df.if_bytes = 0;
 925		ip->i_df.if_u1.if_root = NULL;
 926		break;
 927	default:
 928		ASSERT(0);
 929	}
 
 930	/*
 931	 * Attribute fork settings for new inode.
 932	 */
 933	ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
 934	ip->i_d.di_anextents = 0;
 
 
 
 
 
 
 
 
 935
 936	/*
 937	 * Log the new values stuffed into the inode.
 938	 */
 939	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 940	xfs_trans_log_inode(tp, ip, flags);
 941
 942	/* now that we have an i_mode we can setup the inode structure */
 943	xfs_setup_inode(ip);
 944
 945	*ipp = ip;
 946	return 0;
 947}
 948
 949/*
 950 * Allocates a new inode from disk and return a pointer to the
 951 * incore copy. This routine will internally commit the current
 952 * transaction and allocate a new one if the Space Manager needed
 953 * to do an allocation to replenish the inode free-list.
 954 *
 955 * This routine is designed to be called from xfs_create and
 956 * xfs_create_dir.
 957 *
 958 */
 959int
 960xfs_dir_ialloc(
 961	xfs_trans_t	**tpp,		/* input: current transaction;
 962					   output: may be a new transaction. */
 963	xfs_inode_t	*dp,		/* directory within whose allocate
 964					   the inode. */
 965	umode_t		mode,
 966	xfs_nlink_t	nlink,
 967	dev_t		rdev,
 968	prid_t		prid,		/* project id */
 969	xfs_inode_t	**ipp)		/* pointer to inode; it will be
 970					   locked. */
 971{
 972	xfs_trans_t	*tp;
 973	xfs_inode_t	*ip;
 974	xfs_buf_t	*ialloc_context = NULL;
 975	int		code;
 976	void		*dqinfo;
 977	uint		tflags;
 978
 979	tp = *tpp;
 980	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
 981
 982	/*
 983	 * xfs_ialloc will return a pointer to an incore inode if
 984	 * the Space Manager has an available inode on the free
 985	 * list. Otherwise, it will do an allocation and replenish
 986	 * the freelist.  Since we can only do one allocation per
 987	 * transaction without deadlocks, we will need to commit the
 988	 * current transaction and start a new one.  We will then
 989	 * need to call xfs_ialloc again to get the inode.
 990	 *
 991	 * If xfs_ialloc did an allocation to replenish the freelist,
 992	 * it returns the bp containing the head of the freelist as
 993	 * ialloc_context. We will hold a lock on it across the
 994	 * transaction commit so that no other process can steal
 995	 * the inode(s) that we've just allocated.
 996	 */
 997	code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, &ialloc_context,
 998			&ip);
 999
1000	/*
1001	 * Return an error if we were unable to allocate a new inode.
1002	 * This should only happen if we run out of space on disk or
1003	 * encounter a disk error.
1004	 */
1005	if (code) {
1006		*ipp = NULL;
1007		return code;
1008	}
1009	if (!ialloc_context && !ip) {
1010		*ipp = NULL;
1011		return -ENOSPC;
1012	}
1013
1014	/*
1015	 * If the AGI buffer is non-NULL, then we were unable to get an
1016	 * inode in one operation.  We need to commit the current
1017	 * transaction and call xfs_ialloc() again.  It is guaranteed
1018	 * to succeed the second time.
1019	 */
1020	if (ialloc_context) {
1021		/*
1022		 * Normally, xfs_trans_commit releases all the locks.
1023		 * We call bhold to hang on to the ialloc_context across
1024		 * the commit.  Holding this buffer prevents any other
1025		 * processes from doing any allocations in this
1026		 * allocation group.
1027		 */
1028		xfs_trans_bhold(tp, ialloc_context);
1029
1030		/*
1031		 * We want the quota changes to be associated with the next
1032		 * transaction, NOT this one. So, detach the dqinfo from this
1033		 * and attach it to the next transaction.
1034		 */
1035		dqinfo = NULL;
1036		tflags = 0;
1037		if (tp->t_dqinfo) {
1038			dqinfo = (void *)tp->t_dqinfo;
1039			tp->t_dqinfo = NULL;
1040			tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1041			tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1042		}
1043
1044		code = xfs_trans_roll(&tp);
1045
1046		/*
1047		 * Re-attach the quota info that we detached from prev trx.
1048		 */
1049		if (dqinfo) {
1050			tp->t_dqinfo = dqinfo;
1051			tp->t_flags |= tflags;
1052		}
1053
1054		if (code) {
1055			xfs_buf_relse(ialloc_context);
1056			*tpp = tp;
1057			*ipp = NULL;
1058			return code;
1059		}
1060		xfs_trans_bjoin(tp, ialloc_context);
1061
1062		/*
1063		 * Call ialloc again. Since we've locked out all
1064		 * other allocations in this allocation group,
1065		 * this call should always succeed.
1066		 */
1067		code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1068				  &ialloc_context, &ip);
1069
1070		/*
1071		 * If we get an error at this point, return to the caller
1072		 * so that the current transaction can be aborted.
1073		 */
1074		if (code) {
1075			*tpp = tp;
1076			*ipp = NULL;
1077			return code;
1078		}
1079		ASSERT(!ialloc_context && ip);
1080
1081	}
1082
1083	*ipp = ip;
1084	*tpp = tp;
1085
1086	return 0;
1087}
1088
1089/*
1090 * Decrement the link count on an inode & log the change.  If this causes the
1091 * link count to go to zero, move the inode to AGI unlinked list so that it can
1092 * be freed when the last active reference goes away via xfs_inactive().
1093 */
1094static int			/* error */
1095xfs_droplink(
1096	xfs_trans_t *tp,
1097	xfs_inode_t *ip)
1098{
 
 
 
 
 
 
 
1099	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1100
1101	drop_nlink(VFS_I(ip));
1102	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1103
1104	if (VFS_I(ip)->i_nlink)
1105		return 0;
1106
1107	return xfs_iunlink(tp, ip);
1108}
1109
1110/*
1111 * Increment the link count on an inode & log the change.
1112 */
1113static void
1114xfs_bumplink(
1115	xfs_trans_t *tp,
1116	xfs_inode_t *ip)
1117{
1118	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1119
1120	ASSERT(ip->i_d.di_version > 1);
1121	inc_nlink(VFS_I(ip));
1122	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1123}
1124
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1125int
1126xfs_create(
 
1127	xfs_inode_t		*dp,
1128	struct xfs_name		*name,
1129	umode_t			mode,
1130	dev_t			rdev,
 
1131	xfs_inode_t		**ipp)
1132{
1133	int			is_dir = S_ISDIR(mode);
1134	struct xfs_mount	*mp = dp->i_mount;
1135	struct xfs_inode	*ip = NULL;
1136	struct xfs_trans	*tp = NULL;
1137	int			error;
1138	bool                    unlock_dp_on_error = false;
1139	prid_t			prid;
1140	struct xfs_dquot	*udqp = NULL;
1141	struct xfs_dquot	*gdqp = NULL;
1142	struct xfs_dquot	*pdqp = NULL;
1143	struct xfs_trans_res	*tres;
1144	uint			resblks;
 
1145
1146	trace_xfs_create(dp, name);
1147
1148	if (XFS_FORCED_SHUTDOWN(mp))
 
 
1149		return -EIO;
1150
1151	prid = xfs_get_initial_prid(dp);
1152
1153	/*
1154	 * Make sure that we have allocated dquot(s) on disk.
1155	 */
1156	error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1157					xfs_kgid_to_gid(current_fsgid()), prid,
1158					XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1159					&udqp, &gdqp, &pdqp);
1160	if (error)
1161		return error;
1162
1163	if (is_dir) {
1164		resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1165		tres = &M_RES(mp)->tr_mkdir;
1166	} else {
1167		resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1168		tres = &M_RES(mp)->tr_create;
1169	}
1170
1171	/*
1172	 * Initially assume that the file does not exist and
1173	 * reserve the resources for that case.  If that is not
1174	 * the case we'll drop the one we have and get a more
1175	 * appropriate transaction later.
1176	 */
1177	error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
 
1178	if (error == -ENOSPC) {
1179		/* flush outstanding delalloc blocks and retry */
1180		xfs_flush_inodes(mp);
1181		error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
 
1182	}
1183	if (error)
1184		goto out_release_inode;
1185
1186	xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1187	unlock_dp_on_error = true;
1188
1189	/*
1190	 * Reserve disk quota and the inode.
1191	 */
1192	error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1193						pdqp, resblks, 1, 0);
1194	if (error)
1195		goto out_trans_cancel;
1196
1197	/*
1198	 * A newly created regular or special file just has one directory
1199	 * entry pointing to them, but a directory also the "." entry
1200	 * pointing to itself.
1201	 */
1202	error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev, prid, &ip);
 
 
 
1203	if (error)
1204		goto out_trans_cancel;
1205
1206	/*
1207	 * Now we join the directory inode to the transaction.  We do not do it
1208	 * earlier because xfs_dir_ialloc might commit the previous transaction
1209	 * (and release all the locks).  An error from here on will result in
1210	 * the transaction cancel unlocking dp so don't do it explicitly in the
1211	 * error path.
1212	 */
1213	xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1214	unlock_dp_on_error = false;
1215
1216	error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1217				   resblks ?
1218					resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1219	if (error) {
1220		ASSERT(error != -ENOSPC);
1221		goto out_trans_cancel;
1222	}
1223	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1224	xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1225
1226	if (is_dir) {
1227		error = xfs_dir_init(tp, ip, dp);
1228		if (error)
1229			goto out_trans_cancel;
1230
1231		xfs_bumplink(tp, dp);
1232	}
1233
1234	/*
 
 
 
 
 
 
1235	 * If this is a synchronous mount, make sure that the
1236	 * create transaction goes to disk before returning to
1237	 * the user.
1238	 */
1239	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1240		xfs_trans_set_sync(tp);
1241
1242	/*
1243	 * Attach the dquot(s) to the inodes and modify them incore.
1244	 * These ids of the inode couldn't have changed since the new
1245	 * inode has been locked ever since it was created.
1246	 */
1247	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1248
1249	error = xfs_trans_commit(tp);
1250	if (error)
1251		goto out_release_inode;
1252
1253	xfs_qm_dqrele(udqp);
1254	xfs_qm_dqrele(gdqp);
1255	xfs_qm_dqrele(pdqp);
1256
1257	*ipp = ip;
1258	return 0;
1259
1260 out_trans_cancel:
1261	xfs_trans_cancel(tp);
1262 out_release_inode:
1263	/*
1264	 * Wait until after the current transaction is aborted to finish the
1265	 * setup of the inode and release the inode.  This prevents recursive
1266	 * transactions and deadlocks from xfs_inactive.
1267	 */
1268	if (ip) {
1269		xfs_finish_inode_setup(ip);
1270		xfs_irele(ip);
1271	}
1272
1273	xfs_qm_dqrele(udqp);
1274	xfs_qm_dqrele(gdqp);
1275	xfs_qm_dqrele(pdqp);
1276
1277	if (unlock_dp_on_error)
1278		xfs_iunlock(dp, XFS_ILOCK_EXCL);
1279	return error;
1280}
1281
1282int
1283xfs_create_tmpfile(
 
1284	struct xfs_inode	*dp,
1285	umode_t			mode,
1286	struct xfs_inode	**ipp)
1287{
1288	struct xfs_mount	*mp = dp->i_mount;
1289	struct xfs_inode	*ip = NULL;
1290	struct xfs_trans	*tp = NULL;
1291	int			error;
1292	prid_t                  prid;
1293	struct xfs_dquot	*udqp = NULL;
1294	struct xfs_dquot	*gdqp = NULL;
1295	struct xfs_dquot	*pdqp = NULL;
1296	struct xfs_trans_res	*tres;
1297	uint			resblks;
 
1298
1299	if (XFS_FORCED_SHUTDOWN(mp))
1300		return -EIO;
1301
1302	prid = xfs_get_initial_prid(dp);
1303
1304	/*
1305	 * Make sure that we have allocated dquot(s) on disk.
1306	 */
1307	error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1308				xfs_kgid_to_gid(current_fsgid()), prid,
1309				XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1310				&udqp, &gdqp, &pdqp);
1311	if (error)
1312		return error;
1313
1314	resblks = XFS_IALLOC_SPACE_RES(mp);
1315	tres = &M_RES(mp)->tr_create_tmpfile;
1316
1317	error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1318	if (error)
1319		goto out_release_inode;
1320
1321	error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1322						pdqp, resblks, 1, 0);
1323	if (error)
1324		goto out_trans_cancel;
1325
1326	error = xfs_dir_ialloc(&tp, dp, mode, 0, 0, prid, &ip);
 
 
 
1327	if (error)
1328		goto out_trans_cancel;
1329
1330	if (mp->m_flags & XFS_MOUNT_WSYNC)
1331		xfs_trans_set_sync(tp);
1332
1333	/*
1334	 * Attach the dquot(s) to the inodes and modify them incore.
1335	 * These ids of the inode couldn't have changed since the new
1336	 * inode has been locked ever since it was created.
1337	 */
1338	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1339
1340	error = xfs_iunlink(tp, ip);
1341	if (error)
1342		goto out_trans_cancel;
1343
1344	error = xfs_trans_commit(tp);
1345	if (error)
1346		goto out_release_inode;
1347
1348	xfs_qm_dqrele(udqp);
1349	xfs_qm_dqrele(gdqp);
1350	xfs_qm_dqrele(pdqp);
1351
1352	*ipp = ip;
1353	return 0;
1354
1355 out_trans_cancel:
1356	xfs_trans_cancel(tp);
1357 out_release_inode:
1358	/*
1359	 * Wait until after the current transaction is aborted to finish the
1360	 * setup of the inode and release the inode.  This prevents recursive
1361	 * transactions and deadlocks from xfs_inactive.
1362	 */
1363	if (ip) {
1364		xfs_finish_inode_setup(ip);
1365		xfs_irele(ip);
1366	}
1367
1368	xfs_qm_dqrele(udqp);
1369	xfs_qm_dqrele(gdqp);
1370	xfs_qm_dqrele(pdqp);
1371
1372	return error;
1373}
1374
1375int
1376xfs_link(
1377	xfs_inode_t		*tdp,
1378	xfs_inode_t		*sip,
1379	struct xfs_name		*target_name)
1380{
1381	xfs_mount_t		*mp = tdp->i_mount;
1382	xfs_trans_t		*tp;
1383	int			error;
1384	int			resblks;
1385
1386	trace_xfs_link(tdp, target_name);
1387
1388	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1389
1390	if (XFS_FORCED_SHUTDOWN(mp))
 
 
1391		return -EIO;
1392
1393	error = xfs_qm_dqattach(sip);
1394	if (error)
1395		goto std_return;
1396
1397	error = xfs_qm_dqattach(tdp);
1398	if (error)
1399		goto std_return;
1400
1401	resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1402	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1403	if (error == -ENOSPC) {
1404		resblks = 0;
1405		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1406	}
1407	if (error)
1408		goto std_return;
1409
1410	xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1411
1412	xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1413	xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1414
1415	/*
1416	 * If we are using project inheritance, we only allow hard link
1417	 * creation in our tree when the project IDs are the same; else
1418	 * the tree quota mechanism could be circumvented.
1419	 */
1420	if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1421		     (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1422		error = -EXDEV;
1423		goto error_return;
 
 
 
 
 
 
 
 
 
 
 
1424	}
1425
1426	if (!resblks) {
1427		error = xfs_dir_canenter(tp, tdp, target_name);
1428		if (error)
1429			goto error_return;
1430	}
1431
1432	/*
1433	 * Handle initial link state of O_TMPFILE inode
1434	 */
1435	if (VFS_I(sip)->i_nlink == 0) {
1436		error = xfs_iunlink_remove(tp, sip);
 
 
 
 
1437		if (error)
1438			goto error_return;
1439	}
1440
1441	error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1442				   resblks);
1443	if (error)
1444		goto error_return;
1445	xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1446	xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1447
1448	xfs_bumplink(tp, sip);
 
1449
1450	/*
1451	 * If this is a synchronous mount, make sure that the
1452	 * link transaction goes to disk before returning to
1453	 * the user.
1454	 */
1455	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1456		xfs_trans_set_sync(tp);
1457
1458	return xfs_trans_commit(tp);
1459
1460 error_return:
1461	xfs_trans_cancel(tp);
1462 std_return:
 
 
1463	return error;
1464}
1465
1466/* Clear the reflink flag and the cowblocks tag if possible. */
1467static void
1468xfs_itruncate_clear_reflink_flags(
1469	struct xfs_inode	*ip)
1470{
1471	struct xfs_ifork	*dfork;
1472	struct xfs_ifork	*cfork;
1473
1474	if (!xfs_is_reflink_inode(ip))
1475		return;
1476	dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1477	cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1478	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1479		ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1480	if (cfork->if_bytes == 0)
1481		xfs_inode_clear_cowblocks_tag(ip);
1482}
1483
1484/*
1485 * Free up the underlying blocks past new_size.  The new size must be smaller
1486 * than the current size.  This routine can be used both for the attribute and
1487 * data fork, and does not modify the inode size, which is left to the caller.
1488 *
1489 * The transaction passed to this routine must have made a permanent log
1490 * reservation of at least XFS_ITRUNCATE_LOG_RES.  This routine may commit the
1491 * given transaction and start new ones, so make sure everything involved in
1492 * the transaction is tidy before calling here.  Some transaction will be
1493 * returned to the caller to be committed.  The incoming transaction must
1494 * already include the inode, and both inode locks must be held exclusively.
1495 * The inode must also be "held" within the transaction.  On return the inode
1496 * will be "held" within the returned transaction.  This routine does NOT
1497 * require any disk space to be reserved for it within the transaction.
1498 *
1499 * If we get an error, we must return with the inode locked and linked into the
1500 * current transaction. This keeps things simple for the higher level code,
1501 * because it always knows that the inode is locked and held in the transaction
1502 * that returns to it whether errors occur or not.  We don't mark the inode
1503 * dirty on error so that transactions can be easily aborted if possible.
1504 */
1505int
1506xfs_itruncate_extents_flags(
1507	struct xfs_trans	**tpp,
1508	struct xfs_inode	*ip,
1509	int			whichfork,
1510	xfs_fsize_t		new_size,
1511	int			flags)
1512{
1513	struct xfs_mount	*mp = ip->i_mount;
1514	struct xfs_trans	*tp = *tpp;
1515	xfs_fileoff_t		first_unmap_block;
1516	xfs_fileoff_t		last_block;
1517	xfs_filblks_t		unmap_len;
1518	int			error = 0;
1519	int			done = 0;
1520
1521	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1522	ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1523	       xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1524	ASSERT(new_size <= XFS_ISIZE(ip));
1525	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1526	ASSERT(ip->i_itemp != NULL);
1527	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1528	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1529
1530	trace_xfs_itruncate_extents_start(ip, new_size);
1531
1532	flags |= xfs_bmapi_aflag(whichfork);
1533
1534	/*
1535	 * Since it is possible for space to become allocated beyond
1536	 * the end of the file (in a crash where the space is allocated
1537	 * but the inode size is not yet updated), simply remove any
1538	 * blocks which show up between the new EOF and the maximum
1539	 * possible file size.  If the first block to be removed is
1540	 * beyond the maximum file size (ie it is the same as last_block),
1541	 * then there is nothing to do.
 
1542	 */
1543	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1544	last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1545	if (first_unmap_block == last_block)
1546		return 0;
1547
1548	ASSERT(first_unmap_block < last_block);
1549	unmap_len = last_block - first_unmap_block + 1;
1550	while (!done) {
1551		ASSERT(tp->t_firstblock == NULLFSBLOCK);
1552		error = xfs_bunmapi(tp, ip, first_unmap_block, unmap_len, flags,
1553				    XFS_ITRUNC_MAX_EXTENTS, &done);
1554		if (error)
1555			goto out;
1556
1557		/*
1558		 * Duplicate the transaction that has the permanent
1559		 * reservation and commit the old transaction.
1560		 */
1561		error = xfs_defer_finish(&tp);
1562		if (error)
1563			goto out;
1564
1565		error = xfs_trans_roll_inode(&tp, ip);
1566		if (error)
1567			goto out;
1568	}
1569
 
 
 
 
 
1570	if (whichfork == XFS_DATA_FORK) {
1571		/* Remove all pending CoW reservations. */
1572		error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1573				first_unmap_block, last_block, true);
1574		if (error)
1575			goto out;
1576
1577		xfs_itruncate_clear_reflink_flags(ip);
1578	}
1579
1580	/*
1581	 * Always re-log the inode so that our permanent transaction can keep
1582	 * on rolling it forward in the log.
1583	 */
1584	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1585
1586	trace_xfs_itruncate_extents_end(ip, new_size);
1587
1588out:
1589	*tpp = tp;
1590	return error;
1591}
1592
1593int
1594xfs_release(
1595	xfs_inode_t	*ip)
1596{
1597	xfs_mount_t	*mp = ip->i_mount;
1598	int		error;
1599
1600	if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1601		return 0;
1602
1603	/* If this is a read-only mount, don't do this (would generate I/O) */
1604	if (mp->m_flags & XFS_MOUNT_RDONLY)
1605		return 0;
1606
1607	if (!XFS_FORCED_SHUTDOWN(mp)) {
1608		int truncated;
1609
1610		/*
1611		 * If we previously truncated this file and removed old data
1612		 * in the process, we want to initiate "early" writeout on
1613		 * the last close.  This is an attempt to combat the notorious
1614		 * NULL files problem which is particularly noticeable from a
1615		 * truncate down, buffered (re-)write (delalloc), followed by
1616		 * a crash.  What we are effectively doing here is
1617		 * significantly reducing the time window where we'd otherwise
1618		 * be exposed to that problem.
1619		 */
1620		truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1621		if (truncated) {
1622			xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1623			if (ip->i_delayed_blks > 0) {
1624				error = filemap_flush(VFS_I(ip)->i_mapping);
1625				if (error)
1626					return error;
1627			}
1628		}
1629	}
1630
1631	if (VFS_I(ip)->i_nlink == 0)
1632		return 0;
1633
1634	if (xfs_can_free_eofblocks(ip, false)) {
 
 
 
 
 
 
 
1635
 
1636		/*
1637		 * Check if the inode is being opened, written and closed
1638		 * frequently and we have delayed allocation blocks outstanding
1639		 * (e.g. streaming writes from the NFS server), truncating the
1640		 * blocks past EOF will cause fragmentation to occur.
1641		 *
1642		 * In this case don't do the truncation, but we have to be
1643		 * careful how we detect this case. Blocks beyond EOF show up as
1644		 * i_delayed_blks even when the inode is clean, so we need to
1645		 * truncate them away first before checking for a dirty release.
1646		 * Hence on the first dirty close we will still remove the
1647		 * speculative allocation, but after that we will leave it in
1648		 * place.
1649		 */
1650		if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1651			return 0;
1652		/*
1653		 * If we can't get the iolock just skip truncating the blocks
1654		 * past EOF because we could deadlock with the mmap_sem
1655		 * otherwise. We'll get another chance to drop them once the
1656		 * last reference to the inode is dropped, so we'll never leak
1657		 * blocks permanently.
1658		 */
1659		if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1660			error = xfs_free_eofblocks(ip);
1661			xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1662			if (error)
1663				return error;
1664		}
1665
1666		/* delalloc blocks after truncation means it really is dirty */
1667		if (ip->i_delayed_blks)
1668			xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1669	}
1670	return 0;
 
 
 
1671}
1672
1673/*
1674 * xfs_inactive_truncate
1675 *
1676 * Called to perform a truncate when an inode becomes unlinked.
1677 */
1678STATIC int
1679xfs_inactive_truncate(
1680	struct xfs_inode *ip)
1681{
1682	struct xfs_mount	*mp = ip->i_mount;
1683	struct xfs_trans	*tp;
1684	int			error;
1685
1686	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1687	if (error) {
1688		ASSERT(XFS_FORCED_SHUTDOWN(mp));
1689		return error;
1690	}
1691	xfs_ilock(ip, XFS_ILOCK_EXCL);
1692	xfs_trans_ijoin(tp, ip, 0);
1693
1694	/*
1695	 * Log the inode size first to prevent stale data exposure in the event
1696	 * of a system crash before the truncate completes. See the related
1697	 * comment in xfs_vn_setattr_size() for details.
1698	 */
1699	ip->i_d.di_size = 0;
1700	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1701
1702	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1703	if (error)
1704		goto error_trans_cancel;
1705
1706	ASSERT(ip->i_d.di_nextents == 0);
1707
1708	error = xfs_trans_commit(tp);
1709	if (error)
1710		goto error_unlock;
1711
1712	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1713	return 0;
1714
1715error_trans_cancel:
1716	xfs_trans_cancel(tp);
1717error_unlock:
1718	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1719	return error;
1720}
1721
1722/*
1723 * xfs_inactive_ifree()
1724 *
1725 * Perform the inode free when an inode is unlinked.
1726 */
1727STATIC int
1728xfs_inactive_ifree(
1729	struct xfs_inode *ip)
1730{
1731	struct xfs_mount	*mp = ip->i_mount;
1732	struct xfs_trans	*tp;
1733	int			error;
1734
1735	/*
1736	 * We try to use a per-AG reservation for any block needed by the finobt
1737	 * tree, but as the finobt feature predates the per-AG reservation
1738	 * support a degraded file system might not have enough space for the
1739	 * reservation at mount time.  In that case try to dip into the reserved
1740	 * pool and pray.
1741	 *
1742	 * Send a warning if the reservation does happen to fail, as the inode
1743	 * now remains allocated and sits on the unlinked list until the fs is
1744	 * repaired.
1745	 */
1746	if (unlikely(mp->m_finobt_nores)) {
1747		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1748				XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1749				&tp);
1750	} else {
1751		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1752	}
1753	if (error) {
1754		if (error == -ENOSPC) {
1755			xfs_warn_ratelimited(mp,
1756			"Failed to remove inode(s) from unlinked list. "
1757			"Please free space, unmount and run xfs_repair.");
1758		} else {
1759			ASSERT(XFS_FORCED_SHUTDOWN(mp));
1760		}
1761		return error;
1762	}
1763
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1764	xfs_ilock(ip, XFS_ILOCK_EXCL);
1765	xfs_trans_ijoin(tp, ip, 0);
1766
1767	error = xfs_ifree(tp, ip);
 
1768	if (error) {
1769		/*
1770		 * If we fail to free the inode, shut down.  The cancel
1771		 * might do that, we need to make sure.  Otherwise the
1772		 * inode might be lost for a long time or forever.
1773		 */
1774		if (!XFS_FORCED_SHUTDOWN(mp)) {
1775			xfs_notice(mp, "%s: xfs_ifree returned error %d",
1776				__func__, error);
1777			xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1778		}
1779		xfs_trans_cancel(tp);
1780		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1781		return error;
1782	}
1783
1784	/*
1785	 * Credit the quota account(s). The inode is gone.
1786	 */
1787	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1788
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1789	/*
1790	 * Just ignore errors at this point.  There is nothing we can do except
1791	 * to try to keep going. Make sure it's not a silent error.
1792	 */
1793	error = xfs_trans_commit(tp);
1794	if (error)
1795		xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1796			__func__, error);
1797
1798	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1799	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1800}
1801
1802/*
1803 * xfs_inactive
1804 *
1805 * This is called when the vnode reference count for the vnode
1806 * goes to zero.  If the file has been unlinked, then it must
1807 * now be truncated.  Also, we clear all of the read-ahead state
1808 * kept for the inode here since the file is now closed.
1809 */
1810void
1811xfs_inactive(
1812	xfs_inode_t	*ip)
1813{
1814	struct xfs_mount	*mp;
1815	int			error;
1816	int			truncate = 0;
1817
1818	/*
1819	 * If the inode is already free, then there can be nothing
1820	 * to clean up here.
1821	 */
1822	if (VFS_I(ip)->i_mode == 0) {
1823		ASSERT(ip->i_df.if_broot_bytes == 0);
1824		return;
1825	}
1826
1827	mp = ip->i_mount;
1828	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1829
1830	/* If this is a read-only mount, don't do this (would generate I/O) */
1831	if (mp->m_flags & XFS_MOUNT_RDONLY)
1832		return;
 
 
 
 
 
 
 
 
 
1833
1834	/* Try to clean out the cow blocks if there are any. */
1835	if (xfs_inode_has_cow_data(ip))
1836		xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1837
1838	if (VFS_I(ip)->i_nlink != 0) {
1839		/*
1840		 * force is true because we are evicting an inode from the
1841		 * cache. Post-eof blocks must be freed, lest we end up with
1842		 * broken free space accounting.
1843		 *
1844		 * Note: don't bother with iolock here since lockdep complains
1845		 * about acquiring it in reclaim context. We have the only
1846		 * reference to the inode at this point anyways.
1847		 */
1848		if (xfs_can_free_eofblocks(ip, true))
1849			xfs_free_eofblocks(ip);
1850
1851		return;
1852	}
1853
1854	if (S_ISREG(VFS_I(ip)->i_mode) &&
1855	    (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1856	     ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1857		truncate = 1;
1858
1859	error = xfs_qm_dqattach(ip);
1860	if (error)
1861		return;
 
 
 
 
 
 
 
 
 
 
 
 
1862
1863	if (S_ISLNK(VFS_I(ip)->i_mode))
1864		error = xfs_inactive_symlink(ip);
1865	else if (truncate)
1866		error = xfs_inactive_truncate(ip);
1867	if (error)
1868		return;
1869
1870	/*
1871	 * If there are attributes associated with the file then blow them away
1872	 * now.  The code calls a routine that recursively deconstructs the
1873	 * attribute fork. If also blows away the in-core attribute fork.
1874	 */
1875	if (XFS_IFORK_Q(ip)) {
1876		error = xfs_attr_inactive(ip);
1877		if (error)
1878			return;
1879	}
1880
1881	ASSERT(!ip->i_afp);
1882	ASSERT(ip->i_d.di_anextents == 0);
1883	ASSERT(ip->i_d.di_forkoff == 0);
1884
1885	/*
1886	 * Free the inode.
1887	 */
1888	error = xfs_inactive_ifree(ip);
1889	if (error)
1890		return;
1891
 
1892	/*
1893	 * Release the dquots held by inode, if any.
 
1894	 */
1895	xfs_qm_dqdetach(ip);
 
1896}
1897
1898/*
1899 * In-Core Unlinked List Lookups
1900 * =============================
1901 *
1902 * Every inode is supposed to be reachable from some other piece of metadata
1903 * with the exception of the root directory.  Inodes with a connection to a
1904 * file descriptor but not linked from anywhere in the on-disk directory tree
1905 * are collectively known as unlinked inodes, though the filesystem itself
1906 * maintains links to these inodes so that on-disk metadata are consistent.
1907 *
1908 * XFS implements a per-AG on-disk hash table of unlinked inodes.  The AGI
1909 * header contains a number of buckets that point to an inode, and each inode
1910 * record has a pointer to the next inode in the hash chain.  This
1911 * singly-linked list causes scaling problems in the iunlink remove function
1912 * because we must walk that list to find the inode that points to the inode
1913 * being removed from the unlinked hash bucket list.
1914 *
1915 * What if we modelled the unlinked list as a collection of records capturing
1916 * "X.next_unlinked = Y" relations?  If we indexed those records on Y, we'd
1917 * have a fast way to look up unlinked list predecessors, which avoids the
1918 * slow list walk.  That's exactly what we do here (in-core) with a per-AG
1919 * rhashtable.
1920 *
1921 * Because this is a backref cache, we ignore operational failures since the
1922 * iunlink code can fall back to the slow bucket walk.  The only errors that
1923 * should bubble out are for obviously incorrect situations.
1924 *
1925 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1926 * access or have otherwise provided for concurrency control.
1927 */
1928
1929/* Capture a "X.next_unlinked = Y" relationship. */
1930struct xfs_iunlink {
1931	struct rhash_head	iu_rhash_head;
1932	xfs_agino_t		iu_agino;		/* X */
1933	xfs_agino_t		iu_next_unlinked;	/* Y */
1934};
1935
1936/* Unlinked list predecessor lookup hashtable construction */
1937static int
1938xfs_iunlink_obj_cmpfn(
1939	struct rhashtable_compare_arg	*arg,
1940	const void			*obj)
1941{
1942	const xfs_agino_t		*key = arg->key;
1943	const struct xfs_iunlink	*iu = obj;
1944
1945	if (iu->iu_next_unlinked != *key)
1946		return 1;
1947	return 0;
1948}
1949
1950static const struct rhashtable_params xfs_iunlink_hash_params = {
1951	.min_size		= XFS_AGI_UNLINKED_BUCKETS,
1952	.key_len		= sizeof(xfs_agino_t),
1953	.key_offset		= offsetof(struct xfs_iunlink,
1954					   iu_next_unlinked),
1955	.head_offset		= offsetof(struct xfs_iunlink, iu_rhash_head),
1956	.automatic_shrinking	= true,
1957	.obj_cmpfn		= xfs_iunlink_obj_cmpfn,
1958};
1959
1960/*
1961 * Return X, where X.next_unlinked == @agino.  Returns NULLAGINO if no such
1962 * relation is found.
 
 
1963 */
1964static xfs_agino_t
1965xfs_iunlink_lookup_backref(
1966	struct xfs_perag	*pag,
1967	xfs_agino_t		agino)
1968{
1969	struct xfs_iunlink	*iu;
1970
1971	iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1972			xfs_iunlink_hash_params);
1973	return iu ? iu->iu_agino : NULLAGINO;
1974}
1975
1976/*
1977 * Take ownership of an iunlink cache entry and insert it into the hash table.
1978 * If successful, the entry will be owned by the cache; if not, it is freed.
1979 * Either way, the caller does not own @iu after this call.
1980 */
1981static int
1982xfs_iunlink_insert_backref(
1983	struct xfs_perag	*pag,
1984	struct xfs_iunlink	*iu)
1985{
1986	int			error;
1987
1988	error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1989			&iu->iu_rhash_head, xfs_iunlink_hash_params);
1990	/*
1991	 * Fail loudly if there already was an entry because that's a sign of
1992	 * corruption of in-memory data.  Also fail loudly if we see an error
1993	 * code we didn't anticipate from the rhashtable code.  Currently we
1994	 * only anticipate ENOMEM.
1995	 */
1996	if (error) {
1997		WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1998		kmem_free(iu);
1999	}
2000	/*
2001	 * Absorb any runtime errors that aren't a result of corruption because
2002	 * this is a cache and we can always fall back to bucket list scanning.
2003	 */
2004	if (error != 0 && error != -EEXIST)
2005		error = 0;
2006	return error;
2007}
2008
2009/* Remember that @prev_agino.next_unlinked = @this_agino. */
2010static int
2011xfs_iunlink_add_backref(
2012	struct xfs_perag	*pag,
2013	xfs_agino_t		prev_agino,
2014	xfs_agino_t		this_agino)
2015{
2016	struct xfs_iunlink	*iu;
2017
2018	if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
2019		return 0;
2020
2021	iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
2022	iu->iu_agino = prev_agino;
2023	iu->iu_next_unlinked = this_agino;
2024
2025	return xfs_iunlink_insert_backref(pag, iu);
2026}
2027
2028/*
2029 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
2030 * If @next_unlinked is NULLAGINO, we drop the backref and exit.  If there
2031 * wasn't any such entry then we don't bother.
2032 */
2033static int
2034xfs_iunlink_change_backref(
2035	struct xfs_perag	*pag,
2036	xfs_agino_t		agino,
2037	xfs_agino_t		next_unlinked)
2038{
2039	struct xfs_iunlink	*iu;
2040	int			error;
2041
2042	/* Look up the old entry; if there wasn't one then exit. */
2043	iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
2044			xfs_iunlink_hash_params);
2045	if (!iu)
2046		return 0;
2047
2048	/*
2049	 * Remove the entry.  This shouldn't ever return an error, but if we
2050	 * couldn't remove the old entry we don't want to add it again to the
2051	 * hash table, and if the entry disappeared on us then someone's
2052	 * violated the locking rules and we need to fail loudly.  Either way
2053	 * we cannot remove the inode because internal state is or would have
2054	 * been corrupt.
2055	 */
2056	error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
2057			&iu->iu_rhash_head, xfs_iunlink_hash_params);
2058	if (error)
2059		return error;
2060
2061	/* If there is no new next entry just free our item and return. */
2062	if (next_unlinked == NULLAGINO) {
2063		kmem_free(iu);
2064		return 0;
2065	}
2066
2067	/* Update the entry and re-add it to the hash table. */
2068	iu->iu_next_unlinked = next_unlinked;
2069	return xfs_iunlink_insert_backref(pag, iu);
2070}
2071
2072/* Set up the in-core predecessor structures. */
2073int
2074xfs_iunlink_init(
2075	struct xfs_perag	*pag)
2076{
2077	return rhashtable_init(&pag->pagi_unlinked_hash,
2078			&xfs_iunlink_hash_params);
2079}
2080
2081/* Free the in-core predecessor structures. */
2082static void
2083xfs_iunlink_free_item(
2084	void			*ptr,
2085	void			*arg)
2086{
2087	struct xfs_iunlink	*iu = ptr;
2088	bool			*freed_anything = arg;
2089
2090	*freed_anything = true;
2091	kmem_free(iu);
2092}
2093
2094void
2095xfs_iunlink_destroy(
2096	struct xfs_perag	*pag)
2097{
2098	bool			freed_anything = false;
2099
2100	rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
2101			xfs_iunlink_free_item, &freed_anything);
2102
2103	ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
2104}
2105
2106/*
2107 * Point the AGI unlinked bucket at an inode and log the results.  The caller
2108 * is responsible for validating the old value.
2109 */
2110STATIC int
2111xfs_iunlink_update_bucket(
2112	struct xfs_trans	*tp,
2113	xfs_agnumber_t		agno,
2114	struct xfs_buf		*agibp,
2115	unsigned int		bucket_index,
2116	xfs_agino_t		new_agino)
2117{
2118	struct xfs_agi		*agi = XFS_BUF_TO_AGI(agibp);
2119	xfs_agino_t		old_value;
2120	int			offset;
2121
2122	ASSERT(xfs_verify_agino_or_null(tp->t_mountp, agno, new_agino));
2123
2124	old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2125	trace_xfs_iunlink_update_bucket(tp->t_mountp, agno, bucket_index,
2126			old_value, new_agino);
2127
2128	/*
2129	 * We should never find the head of the list already set to the value
2130	 * passed in because either we're adding or removing ourselves from the
2131	 * head of the list.
2132	 */
2133	if (old_value == new_agino)
 
 
2134		return -EFSCORRUPTED;
 
2135
2136	agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2137	offset = offsetof(struct xfs_agi, agi_unlinked) +
2138			(sizeof(xfs_agino_t) * bucket_index);
2139	xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2140	return 0;
2141}
2142
2143/* Set an on-disk inode's next_unlinked pointer. */
2144STATIC void
2145xfs_iunlink_update_dinode(
 
 
 
 
2146	struct xfs_trans	*tp,
2147	xfs_agnumber_t		agno,
2148	xfs_agino_t		agino,
2149	struct xfs_buf		*ibp,
2150	struct xfs_dinode	*dip,
2151	struct xfs_imap		*imap,
2152	xfs_agino_t		next_agino)
2153{
2154	struct xfs_mount	*mp = tp->t_mountp;
2155	int			offset;
2156
2157	ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2158
2159	trace_xfs_iunlink_update_dinode(mp, agno, agino,
2160			be32_to_cpu(dip->di_next_unlinked), next_agino);
2161
2162	dip->di_next_unlinked = cpu_to_be32(next_agino);
2163	offset = imap->im_boffset +
2164			offsetof(struct xfs_dinode, di_next_unlinked);
2165
2166	/* need to recalc the inode CRC if appropriate */
2167	xfs_dinode_calc_crc(mp, dip);
2168	xfs_trans_inode_buf(tp, ibp);
2169	xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2170	xfs_inobp_check(mp, ibp);
2171}
2172
2173/* Set an in-core inode's unlinked pointer and return the old value. */
2174STATIC int
2175xfs_iunlink_update_inode(
2176	struct xfs_trans	*tp,
2177	struct xfs_inode	*ip,
2178	xfs_agnumber_t		agno,
2179	xfs_agino_t		next_agino,
2180	xfs_agino_t		*old_next_agino)
2181{
2182	struct xfs_mount	*mp = tp->t_mountp;
2183	struct xfs_dinode	*dip;
2184	struct xfs_buf		*ibp;
2185	xfs_agino_t		old_value;
2186	int			error;
2187
2188	ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
 
 
2189
2190	error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp, 0, 0);
2191	if (error)
 
 
 
 
 
 
 
 
2192		return error;
2193
2194	/* Make sure the old pointer isn't garbage. */
2195	old_value = be32_to_cpu(dip->di_next_unlinked);
2196	if (!xfs_verify_agino_or_null(mp, agno, old_value)) {
2197		error = -EFSCORRUPTED;
2198		goto out;
2199	}
2200
2201	/*
2202	 * Since we're updating a linked list, we should never find that the
2203	 * current pointer is the same as the new value, unless we're
2204	 * terminating the list.
2205	 */
2206	*old_next_agino = old_value;
2207	if (old_value == next_agino) {
2208		if (next_agino != NULLAGINO)
2209			error = -EFSCORRUPTED;
2210		goto out;
2211	}
2212
2213	/* Ok, update the new pointer. */
2214	xfs_iunlink_update_dinode(tp, agno, XFS_INO_TO_AGINO(mp, ip->i_ino),
2215			ibp, dip, &ip->i_imap, next_agino);
2216	return 0;
2217out:
2218	xfs_trans_brelse(tp, ibp);
 
2219	return error;
2220}
2221
2222/*
2223 * This is called when the inode's link count has gone to 0 or we are creating
2224 * a tmpfile via O_TMPFILE.  The inode @ip must have nlink == 0.
2225 *
2226 * We place the on-disk inode on a list in the AGI.  It will be pulled from this
2227 * list when the inode is freed.
2228 */
2229STATIC int
2230xfs_iunlink(
2231	struct xfs_trans	*tp,
 
 
2232	struct xfs_inode	*ip)
2233{
2234	struct xfs_mount	*mp = tp->t_mountp;
2235	struct xfs_agi		*agi;
2236	struct xfs_buf		*agibp;
2237	xfs_agino_t		next_agino;
2238	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2239	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2240	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2241	int			error;
2242
2243	ASSERT(VFS_I(ip)->i_nlink == 0);
2244	ASSERT(VFS_I(ip)->i_mode != 0);
2245	trace_xfs_iunlink(ip);
2246
2247	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2248	error = xfs_read_agi(mp, tp, agno, &agibp);
2249	if (error)
2250		return error;
2251	agi = XFS_BUF_TO_AGI(agibp);
2252
2253	/*
2254	 * Get the index into the agi hash table for the list this inode will
2255	 * go on.  Make sure the pointer isn't garbage and that this inode
2256	 * isn't already on the list.
2257	 */
2258	next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2259	if (next_agino == agino ||
2260	    !xfs_verify_agino_or_null(mp, agno, next_agino))
 
 
2261		return -EFSCORRUPTED;
 
2262
2263	if (next_agino != NULLAGINO) {
2264		struct xfs_perag	*pag;
2265		xfs_agino_t		old_agino;
 
 
 
 
 
 
2266
 
2267		/*
2268		 * There is already another inode in the bucket, so point this
2269		 * inode to the current head of the list.
2270		 */
2271		error = xfs_iunlink_update_inode(tp, ip, agno, next_agino,
2272				&old_agino);
2273		if (error)
2274			return error;
2275		ASSERT(old_agino == NULLAGINO);
2276
2277		/*
2278		 * agino has been unlinked, add a backref from the next inode
2279		 * back to agino.
2280		 */
2281		pag = xfs_perag_get(mp, agno);
2282		error = xfs_iunlink_add_backref(pag, agino, next_agino);
2283		xfs_perag_put(pag);
2284		if (error)
2285			return error;
 
2286	}
2287
2288	/* Point the head of the list to point to this inode. */
2289	return xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index, agino);
 
2290}
2291
2292/* Return the imap, dinode pointer, and buffer for an inode. */
 
 
 
 
 
 
2293STATIC int
2294xfs_iunlink_map_ino(
2295	struct xfs_trans	*tp,
2296	xfs_agnumber_t		agno,
2297	xfs_agino_t		agino,
2298	struct xfs_imap		*imap,
2299	struct xfs_dinode	**dipp,
2300	struct xfs_buf		**bpp)
2301{
2302	struct xfs_mount	*mp = tp->t_mountp;
 
 
2303	int			error;
2304
2305	imap->im_blkno = 0;
2306	error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2307	if (error) {
2308		xfs_warn(mp, "%s: xfs_imap returned error %d.",
2309				__func__, error);
2310		return error;
2311	}
2312
2313	error = xfs_imap_to_bp(mp, tp, imap, dipp, bpp, 0, 0);
2314	if (error) {
2315		xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2316				__func__, error);
2317		return error;
2318	}
2319
2320	return 0;
 
 
 
 
 
 
 
 
2321}
2322
2323/*
2324 * Walk the unlinked chain from @head_agino until we find the inode that
2325 * points to @target_agino.  Return the inode number, map, dinode pointer,
2326 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2327 *
2328 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2329 * @agino, @imap, @dipp, and @bpp are all output parameters.
2330 *
2331 * Do not call this function if @target_agino is the head of the list.
2332 */
2333STATIC int
2334xfs_iunlink_map_prev(
2335	struct xfs_trans	*tp,
2336	xfs_agnumber_t		agno,
2337	xfs_agino_t		head_agino,
2338	xfs_agino_t		target_agino,
2339	xfs_agino_t		*agino,
2340	struct xfs_imap		*imap,
2341	struct xfs_dinode	**dipp,
2342	struct xfs_buf		**bpp,
2343	struct xfs_perag	*pag)
2344{
2345	struct xfs_mount	*mp = tp->t_mountp;
2346	xfs_agino_t		next_agino;
 
 
 
2347	int			error;
2348
2349	ASSERT(head_agino != target_agino);
2350	*bpp = NULL;
2351
2352	/* See if our backref cache can find it faster. */
2353	*agino = xfs_iunlink_lookup_backref(pag, target_agino);
2354	if (*agino != NULLAGINO) {
2355		error = xfs_iunlink_map_ino(tp, agno, *agino, imap, dipp, bpp);
2356		if (error)
2357			return error;
2358
2359		if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2360			return 0;
2361
2362		/*
2363		 * If we get here the cache contents were corrupt, so drop the
2364		 * buffer and fall back to walking the bucket list.
2365		 */
2366		xfs_trans_brelse(tp, *bpp);
2367		*bpp = NULL;
2368		WARN_ON_ONCE(1);
 
 
 
2369	}
2370
2371	trace_xfs_iunlink_map_prev_fallback(mp, agno);
2372
2373	/* Otherwise, walk the entire bucket until we find it. */
2374	next_agino = head_agino;
2375	while (next_agino != target_agino) {
2376		xfs_agino_t	unlinked_agino;
 
 
2377
2378		if (*bpp)
2379			xfs_trans_brelse(tp, *bpp);
 
 
 
 
 
 
 
 
 
2380
2381		*agino = next_agino;
2382		error = xfs_iunlink_map_ino(tp, agno, next_agino, imap, dipp,
2383				bpp);
2384		if (error)
2385			return error;
2386
2387		unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2388		/*
2389		 * Make sure this pointer is valid and isn't an obvious
2390		 * infinite loop.
2391		 */
2392		if (!xfs_verify_agino(mp, agno, unlinked_agino) ||
2393		    next_agino == unlinked_agino) {
2394			XFS_CORRUPTION_ERROR(__func__,
2395					XFS_ERRLEVEL_LOW, mp,
2396					*dipp, sizeof(**dipp));
2397			error = -EFSCORRUPTED;
2398			return error;
2399		}
2400		next_agino = unlinked_agino;
 
 
 
 
 
 
 
2401	}
2402
2403	return 0;
 
 
2404}
2405
2406/*
2407 * Pull the on-disk inode from the AGI unlinked list.
2408 */
2409STATIC int
2410xfs_iunlink_remove(
2411	struct xfs_trans	*tp,
 
2412	struct xfs_inode	*ip)
2413{
2414	struct xfs_mount	*mp = tp->t_mountp;
2415	struct xfs_agi		*agi;
2416	struct xfs_buf		*agibp;
2417	struct xfs_buf		*last_ibp;
2418	struct xfs_dinode	*last_dip = NULL;
2419	struct xfs_perag	*pag = NULL;
2420	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2421	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2422	xfs_agino_t		next_agino;
2423	xfs_agino_t		head_agino;
2424	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2425	int			error;
2426
2427	trace_xfs_iunlink_remove(ip);
2428
2429	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2430	error = xfs_read_agi(mp, tp, agno, &agibp);
2431	if (error)
2432		return error;
2433	agi = XFS_BUF_TO_AGI(agibp);
2434
2435	/*
2436	 * Get the index into the agi hash table for the list this inode will
2437	 * go on.  Make sure the head pointer isn't garbage.
2438	 */
2439	head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2440	if (!xfs_verify_agino(mp, agno, head_agino)) {
2441		XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2442				agi, sizeof(*agi));
2443		return -EFSCORRUPTED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2444	}
2445
2446	/*
2447	 * Set our inode's next_unlinked pointer to NULL and then return
2448	 * the old pointer value so that we can update whatever was previous
2449	 * to us in the list to point to whatever was next in the list.
 
 
 
 
 
 
 
 
 
 
 
2450	 */
2451	error = xfs_iunlink_update_inode(tp, ip, agno, NULLAGINO, &next_agino);
2452	if (error)
2453		return error;
 
 
 
 
 
 
2454
2455	/*
2456	 * If there was a backref pointing from the next inode back to this
2457	 * one, remove it because we've removed this inode from the list.
2458	 *
2459	 * Later, if this inode was in the middle of the list we'll update
2460	 * this inode's backref to point from the next inode.
2461	 */
2462	if (next_agino != NULLAGINO) {
2463		pag = xfs_perag_get(mp, agno);
2464		error = xfs_iunlink_change_backref(pag, next_agino,
2465				NULLAGINO);
2466		if (error)
2467			goto out;
2468	}
2469
2470	if (head_agino == agino) {
2471		/* Point the head of the list to the next unlinked inode. */
2472		error = xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index,
2473				next_agino);
2474		if (error)
2475			goto out;
2476	} else {
2477		struct xfs_imap	imap;
2478		xfs_agino_t	prev_agino;
2479
2480		if (!pag)
2481			pag = xfs_perag_get(mp, agno);
2482
2483		/* We need to search the list for the inode being freed. */
2484		error = xfs_iunlink_map_prev(tp, agno, head_agino, agino,
2485				&prev_agino, &imap, &last_dip, &last_ibp,
2486				pag);
2487		if (error)
2488			goto out;
2489
2490		/* Point the previous inode on the list to the next inode. */
2491		xfs_iunlink_update_dinode(tp, agno, prev_agino, last_ibp,
2492				last_dip, &imap, next_agino);
 
 
 
 
2493
2494		/*
2495		 * Now we deal with the backref for this inode.  If this inode
2496		 * pointed at a real inode, change the backref that pointed to
2497		 * us to point to our old next.  If this inode was the end of
2498		 * the list, delete the backref that pointed to us.  Note that
2499		 * change_backref takes care of deleting the backref if
2500		 * next_agino is NULLAGINO.
2501		 */
2502		error = xfs_iunlink_change_backref(pag, agino, next_agino);
2503		if (error)
2504			goto out;
2505	}
2506
2507out:
2508	if (pag)
2509		xfs_perag_put(pag);
2510	return error;
 
 
2511}
2512
2513/*
2514 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2515 * inodes that are in memory - they all must be marked stale and attached to
2516 * the cluster buffer.
2517 */
2518STATIC int
2519xfs_ifree_cluster(
2520	xfs_inode_t		*free_ip,
2521	xfs_trans_t		*tp,
 
2522	struct xfs_icluster	*xic)
2523{
2524	xfs_mount_t		*mp = free_ip->i_mount;
 
 
 
 
2525	int			nbufs;
2526	int			i, j;
2527	int			ioffset;
2528	xfs_daddr_t		blkno;
2529	xfs_buf_t		*bp;
2530	xfs_inode_t		*ip;
2531	xfs_inode_log_item_t	*iip;
2532	struct xfs_log_item	*lip;
2533	struct xfs_perag	*pag;
2534	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2535	xfs_ino_t		inum;
2536
2537	inum = xic->first_ino;
2538	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2539	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2540
2541	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2542		/*
2543		 * The allocation bitmap tells us which inodes of the chunk were
2544		 * physically allocated. Skip the cluster if an inode falls into
2545		 * a sparse region.
2546		 */
2547		ioffset = inum - xic->first_ino;
2548		if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2549			ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2550			continue;
2551		}
2552
2553		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2554					 XFS_INO_TO_AGBNO(mp, inum));
2555
2556		/*
2557		 * We obtain and lock the backing buffer first in the process
2558		 * here, as we have to ensure that any dirty inode that we
2559		 * can't get the flush lock on is attached to the buffer.
 
2560		 * If we scan the in-memory inodes first, then buffer IO can
2561		 * complete before we get a lock on it, and hence we may fail
2562		 * to mark all the active inodes on the buffer stale.
2563		 */
2564		bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2565					mp->m_bsize * igeo->blocks_per_cluster,
2566					XBF_UNMAPPED);
2567
2568		if (!bp)
2569			return -ENOMEM;
2570
2571		/*
2572		 * This buffer may not have been correctly initialised as we
2573		 * didn't read it from disk. That's not important because we are
2574		 * only using to mark the buffer as stale in the log, and to
2575		 * attach stale cached inodes on it. That means it will never be
2576		 * dispatched for IO. If it is, we want to know about it, and we
2577		 * want it to fail. We can acheive this by adding a write
2578		 * verifier to the buffer.
2579		 */
2580		bp->b_ops = &xfs_inode_buf_ops;
2581
2582		/*
2583		 * Walk the inodes already attached to the buffer and mark them
2584		 * stale. These will all have the flush locks held, so an
2585		 * in-memory inode walk can't lock them. By marking them all
2586		 * stale first, we will not attempt to lock them in the loop
2587		 * below as the XFS_ISTALE flag will be set.
2588		 */
2589		list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
2590			if (lip->li_type == XFS_LI_INODE) {
2591				iip = (xfs_inode_log_item_t *)lip;
2592				ASSERT(iip->ili_logged == 1);
2593				lip->li_cb = xfs_istale_done;
2594				xfs_trans_ail_copy_lsn(mp->m_ail,
2595							&iip->ili_flush_lsn,
2596							&iip->ili_item.li_lsn);
2597				xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2598			}
2599		}
2600
2601
2602		/*
2603		 * For each inode in memory attempt to add it to the inode
2604		 * buffer and set it up for being staled on buffer IO
2605		 * completion.  This is safe as we've locked out tail pushing
2606		 * and flushing by locking the buffer.
2607		 *
2608		 * We have already marked every inode that was part of a
2609		 * transaction stale above, which means there is no point in
2610		 * even trying to lock them.
2611		 */
2612		for (i = 0; i < igeo->inodes_per_cluster; i++) {
2613retry:
2614			rcu_read_lock();
2615			ip = radix_tree_lookup(&pag->pag_ici_root,
2616					XFS_INO_TO_AGINO(mp, (inum + i)));
2617
2618			/* Inode not in memory, nothing to do */
2619			if (!ip) {
2620				rcu_read_unlock();
2621				continue;
2622			}
2623
2624			/*
2625			 * because this is an RCU protected lookup, we could
2626			 * find a recently freed or even reallocated inode
2627			 * during the lookup. We need to check under the
2628			 * i_flags_lock for a valid inode here. Skip it if it
2629			 * is not valid, the wrong inode or stale.
2630			 */
2631			spin_lock(&ip->i_flags_lock);
2632			if (ip->i_ino != inum + i ||
2633			    __xfs_iflags_test(ip, XFS_ISTALE)) {
2634				spin_unlock(&ip->i_flags_lock);
2635				rcu_read_unlock();
2636				continue;
2637			}
2638			spin_unlock(&ip->i_flags_lock);
2639
2640			/*
2641			 * Don't try to lock/unlock the current inode, but we
2642			 * _cannot_ skip the other inodes that we did not find
2643			 * in the list attached to the buffer and are not
2644			 * already marked stale. If we can't lock it, back off
2645			 * and retry.
2646			 */
2647			if (ip != free_ip) {
2648				if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2649					rcu_read_unlock();
2650					delay(1);
2651					goto retry;
2652				}
2653
2654				/*
2655				 * Check the inode number again in case we're
2656				 * racing with freeing in xfs_reclaim_inode().
2657				 * See the comments in that function for more
2658				 * information as to why the initial check is
2659				 * not sufficient.
2660				 */
2661				if (ip->i_ino != inum + i) {
2662					xfs_iunlock(ip, XFS_ILOCK_EXCL);
2663					rcu_read_unlock();
2664					continue;
2665				}
2666			}
2667			rcu_read_unlock();
2668
2669			xfs_iflock(ip);
2670			xfs_iflags_set(ip, XFS_ISTALE);
2671
2672			/*
2673			 * we don't need to attach clean inodes or those only
2674			 * with unlogged changes (which we throw away, anyway).
2675			 */
2676			iip = ip->i_itemp;
2677			if (!iip || xfs_inode_clean(ip)) {
2678				ASSERT(ip != free_ip);
2679				xfs_ifunlock(ip);
2680				xfs_iunlock(ip, XFS_ILOCK_EXCL);
2681				continue;
2682			}
2683
2684			iip->ili_last_fields = iip->ili_fields;
2685			iip->ili_fields = 0;
2686			iip->ili_fsync_fields = 0;
2687			iip->ili_logged = 1;
2688			xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2689						&iip->ili_item.li_lsn);
2690
2691			xfs_buf_attach_iodone(bp, xfs_istale_done,
2692						  &iip->ili_item);
2693
2694			if (ip != free_ip)
2695				xfs_iunlock(ip, XFS_ILOCK_EXCL);
2696		}
2697
2698		xfs_trans_stale_inode_buf(tp, bp);
2699		xfs_trans_binval(tp, bp);
2700	}
2701
2702	xfs_perag_put(pag);
2703	return 0;
2704}
2705
2706/*
2707 * Free any local-format buffers sitting around before we reset to
2708 * extents format.
2709 */
2710static inline void
2711xfs_ifree_local_data(
2712	struct xfs_inode	*ip,
2713	int			whichfork)
2714{
2715	struct xfs_ifork	*ifp;
2716
2717	if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
2718		return;
2719
2720	ifp = XFS_IFORK_PTR(ip, whichfork);
2721	xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
2722}
2723
2724/*
2725 * This is called to return an inode to the inode free list.
2726 * The inode should already be truncated to 0 length and have
2727 * no pages associated with it.  This routine also assumes that
2728 * the inode is already a part of the transaction.
2729 *
2730 * The on-disk copy of the inode will have been added to the list
2731 * of unlinked inodes in the AGI. We need to remove the inode from
2732 * that list atomically with respect to freeing it here.
2733 */
2734int
2735xfs_ifree(
2736	struct xfs_trans	*tp,
2737	struct xfs_inode	*ip)
2738{
2739	int			error;
 
2740	struct xfs_icluster	xic = { 0 };
 
 
2741
2742	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2743	ASSERT(VFS_I(ip)->i_nlink == 0);
2744	ASSERT(ip->i_d.di_nextents == 0);
2745	ASSERT(ip->i_d.di_anextents == 0);
2746	ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2747	ASSERT(ip->i_d.di_nblocks == 0);
 
2748
2749	/*
2750	 * Pull the on-disk inode from the AGI unlinked list.
 
 
 
2751	 */
2752	error = xfs_iunlink_remove(tp, ip);
2753	if (error)
2754		return error;
2755
2756	error = xfs_difree(tp, ip->i_ino, &xic);
2757	if (error)
2758		return error;
2759
2760	xfs_ifree_local_data(ip, XFS_DATA_FORK);
2761	xfs_ifree_local_data(ip, XFS_ATTR_FORK);
 
 
 
 
 
 
 
 
2762
2763	VFS_I(ip)->i_mode = 0;		/* mark incore inode as free */
2764	ip->i_d.di_flags = 0;
2765	ip->i_d.di_flags2 = 0;
2766	ip->i_d.di_dmevmask = 0;
2767	ip->i_d.di_forkoff = 0;		/* mark the attr fork not in use */
2768	ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2769	ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2770
2771	/* Don't attempt to replay owner changes for a deleted inode */
2772	ip->i_itemp->ili_fields &= ~(XFS_ILOG_AOWNER|XFS_ILOG_DOWNER);
 
 
2773
2774	/*
2775	 * Bump the generation count so no one will be confused
2776	 * by reincarnations of this inode.
2777	 */
2778	VFS_I(ip)->i_generation++;
2779	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2780
2781	if (xic.deleted)
2782		error = xfs_ifree_cluster(ip, tp, &xic);
2783
 
2784	return error;
2785}
2786
2787/*
2788 * This is called to unpin an inode.  The caller must have the inode locked
2789 * in at least shared mode so that the buffer cannot be subsequently pinned
2790 * once someone is waiting for it to be unpinned.
2791 */
2792static void
2793xfs_iunpin(
2794	struct xfs_inode	*ip)
2795{
2796	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2797
2798	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2799
2800	/* Give the log a push to start the unpinning I/O */
2801	xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0, NULL);
2802
2803}
2804
2805static void
2806__xfs_iunpin_wait(
2807	struct xfs_inode	*ip)
2808{
2809	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2810	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2811
2812	xfs_iunpin(ip);
2813
2814	do {
2815		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2816		if (xfs_ipincount(ip))
2817			io_schedule();
2818	} while (xfs_ipincount(ip));
2819	finish_wait(wq, &wait.wq_entry);
2820}
2821
2822void
2823xfs_iunpin_wait(
2824	struct xfs_inode	*ip)
2825{
2826	if (xfs_ipincount(ip))
2827		__xfs_iunpin_wait(ip);
2828}
2829
2830/*
2831 * Removing an inode from the namespace involves removing the directory entry
2832 * and dropping the link count on the inode. Removing the directory entry can
2833 * result in locking an AGF (directory blocks were freed) and removing a link
2834 * count can result in placing the inode on an unlinked list which results in
2835 * locking an AGI.
2836 *
2837 * The big problem here is that we have an ordering constraint on AGF and AGI
2838 * locking - inode allocation locks the AGI, then can allocate a new extent for
2839 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2840 * removes the inode from the unlinked list, requiring that we lock the AGI
2841 * first, and then freeing the inode can result in an inode chunk being freed
2842 * and hence freeing disk space requiring that we lock an AGF.
2843 *
2844 * Hence the ordering that is imposed by other parts of the code is AGI before
2845 * AGF. This means we cannot remove the directory entry before we drop the inode
2846 * reference count and put it on the unlinked list as this results in a lock
2847 * order of AGF then AGI, and this can deadlock against inode allocation and
2848 * freeing. Therefore we must drop the link counts before we remove the
2849 * directory entry.
2850 *
2851 * This is still safe from a transactional point of view - it is not until we
2852 * get to xfs_defer_finish() that we have the possibility of multiple
2853 * transactions in this operation. Hence as long as we remove the directory
2854 * entry and drop the link count in the first transaction of the remove
2855 * operation, there are no transactional constraints on the ordering here.
2856 */
2857int
2858xfs_remove(
2859	xfs_inode_t             *dp,
2860	struct xfs_name		*name,
2861	xfs_inode_t		*ip)
2862{
2863	xfs_mount_t		*mp = dp->i_mount;
2864	xfs_trans_t             *tp = NULL;
2865	int			is_dir = S_ISDIR(VFS_I(ip)->i_mode);
 
2866	int                     error = 0;
2867	uint			resblks;
2868
2869	trace_xfs_remove(dp, name);
2870
2871	if (XFS_FORCED_SHUTDOWN(mp))
 
 
2872		return -EIO;
2873
2874	error = xfs_qm_dqattach(dp);
2875	if (error)
2876		goto std_return;
2877
2878	error = xfs_qm_dqattach(ip);
2879	if (error)
2880		goto std_return;
2881
2882	/*
2883	 * We try to get the real space reservation first,
2884	 * allowing for directory btree deletion(s) implying
2885	 * possible bmap insert(s).  If we can't get the space
2886	 * reservation then we use 0 instead, and avoid the bmap
2887	 * btree insert(s) in the directory code by, if the bmap
2888	 * insert tries to happen, instead trimming the LAST
2889	 * block from the directory.
 
 
2890	 */
2891	resblks = XFS_REMOVE_SPACE_RES(mp);
2892	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2893	if (error == -ENOSPC) {
2894		resblks = 0;
2895		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2896				&tp);
2897	}
2898	if (error) {
2899		ASSERT(error != -ENOSPC);
2900		goto std_return;
2901	}
2902
2903	xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2904
2905	xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2906	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2907
2908	/*
2909	 * If we're removing a directory perform some additional validation.
2910	 */
2911	if (is_dir) {
2912		ASSERT(VFS_I(ip)->i_nlink >= 2);
2913		if (VFS_I(ip)->i_nlink != 2) {
2914			error = -ENOTEMPTY;
2915			goto out_trans_cancel;
2916		}
2917		if (!xfs_dir_isempty(ip)) {
2918			error = -ENOTEMPTY;
2919			goto out_trans_cancel;
2920		}
2921
2922		/* Drop the link from ip's "..".  */
2923		error = xfs_droplink(tp, dp);
2924		if (error)
2925			goto out_trans_cancel;
2926
2927		/* Drop the "." link from ip to self.  */
2928		error = xfs_droplink(tp, ip);
2929		if (error)
2930			goto out_trans_cancel;
 
 
 
 
 
 
 
 
 
 
 
 
 
2931	} else {
2932		/*
2933		 * When removing a non-directory we need to log the parent
2934		 * inode here.  For a directory this is done implicitly
2935		 * by the xfs_droplink call for the ".." entry.
2936		 */
2937		xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2938	}
2939	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2940
2941	/* Drop the link from dp to ip. */
2942	error = xfs_droplink(tp, ip);
2943	if (error)
2944		goto out_trans_cancel;
2945
2946	error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2947	if (error) {
2948		ASSERT(error != -ENOENT);
2949		goto out_trans_cancel;
2950	}
2951
2952	/*
 
 
 
 
 
 
2953	 * If this is a synchronous mount, make sure that the
2954	 * remove transaction goes to disk before returning to
2955	 * the user.
2956	 */
2957	if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2958		xfs_trans_set_sync(tp);
2959
2960	error = xfs_trans_commit(tp);
2961	if (error)
2962		goto std_return;
2963
2964	if (is_dir && xfs_inode_is_filestream(ip))
2965		xfs_filestream_deassociate(ip);
2966
2967	return 0;
2968
2969 out_trans_cancel:
2970	xfs_trans_cancel(tp);
2971 std_return:
2972	return error;
2973}
2974
2975/*
2976 * Enter all inodes for a rename transaction into a sorted array.
2977 */
2978#define __XFS_SORT_INODES	5
2979STATIC void
2980xfs_sort_for_rename(
2981	struct xfs_inode	*dp1,	/* in: old (source) directory inode */
2982	struct xfs_inode	*dp2,	/* in: new (target) directory inode */
2983	struct xfs_inode	*ip1,	/* in: inode of old entry */
2984	struct xfs_inode	*ip2,	/* in: inode of new entry */
2985	struct xfs_inode	*wip,	/* in: whiteout inode */
2986	struct xfs_inode	**i_tab,/* out: sorted array of inodes */
2987	int			*num_inodes)  /* in/out: inodes in array */
2988{
2989	int			i, j;
2990
2991	ASSERT(*num_inodes == __XFS_SORT_INODES);
2992	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2993
2994	/*
2995	 * i_tab contains a list of pointers to inodes.  We initialize
2996	 * the table here & we'll sort it.  We will then use it to
2997	 * order the acquisition of the inode locks.
2998	 *
2999	 * Note that the table may contain duplicates.  e.g., dp1 == dp2.
3000	 */
3001	i = 0;
3002	i_tab[i++] = dp1;
3003	i_tab[i++] = dp2;
3004	i_tab[i++] = ip1;
3005	if (ip2)
3006		i_tab[i++] = ip2;
3007	if (wip)
3008		i_tab[i++] = wip;
3009	*num_inodes = i;
3010
3011	/*
3012	 * Sort the elements via bubble sort.  (Remember, there are at
3013	 * most 5 elements to sort, so this is adequate.)
3014	 */
3015	for (i = 0; i < *num_inodes; i++) {
3016		for (j = 1; j < *num_inodes; j++) {
3017			if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
3018				struct xfs_inode *temp = i_tab[j];
3019				i_tab[j] = i_tab[j-1];
3020				i_tab[j-1] = temp;
3021			}
3022		}
3023	}
3024}
3025
3026static int
3027xfs_finish_rename(
3028	struct xfs_trans	*tp)
3029{
3030	/*
3031	 * If this is a synchronous mount, make sure that the rename transaction
3032	 * goes to disk before returning to the user.
3033	 */
3034	if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
3035		xfs_trans_set_sync(tp);
3036
3037	return xfs_trans_commit(tp);
3038}
3039
3040/*
3041 * xfs_cross_rename()
3042 *
3043 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
3044 */
3045STATIC int
3046xfs_cross_rename(
3047	struct xfs_trans	*tp,
3048	struct xfs_inode	*dp1,
3049	struct xfs_name		*name1,
3050	struct xfs_inode	*ip1,
3051	struct xfs_inode	*dp2,
3052	struct xfs_name		*name2,
3053	struct xfs_inode	*ip2,
3054	int			spaceres)
3055{
3056	int		error = 0;
3057	int		ip1_flags = 0;
3058	int		ip2_flags = 0;
3059	int		dp2_flags = 0;
3060
3061	/* Swap inode number for dirent in first parent */
3062	error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
3063	if (error)
3064		goto out_trans_abort;
3065
3066	/* Swap inode number for dirent in second parent */
3067	error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
3068	if (error)
3069		goto out_trans_abort;
3070
3071	/*
3072	 * If we're renaming one or more directories across different parents,
3073	 * update the respective ".." entries (and link counts) to match the new
3074	 * parents.
3075	 */
3076	if (dp1 != dp2) {
3077		dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3078
3079		if (S_ISDIR(VFS_I(ip2)->i_mode)) {
3080			error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
3081						dp1->i_ino, spaceres);
3082			if (error)
3083				goto out_trans_abort;
3084
3085			/* transfer ip2 ".." reference to dp1 */
3086			if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
3087				error = xfs_droplink(tp, dp2);
3088				if (error)
3089					goto out_trans_abort;
3090				xfs_bumplink(tp, dp1);
3091			}
3092
3093			/*
3094			 * Although ip1 isn't changed here, userspace needs
3095			 * to be warned about the change, so that applications
3096			 * relying on it (like backup ones), will properly
3097			 * notify the change
3098			 */
3099			ip1_flags |= XFS_ICHGTIME_CHG;
3100			ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3101		}
3102
3103		if (S_ISDIR(VFS_I(ip1)->i_mode)) {
3104			error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
3105						dp2->i_ino, spaceres);
3106			if (error)
3107				goto out_trans_abort;
3108
3109			/* transfer ip1 ".." reference to dp2 */
3110			if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
3111				error = xfs_droplink(tp, dp1);
3112				if (error)
3113					goto out_trans_abort;
3114				xfs_bumplink(tp, dp2);
3115			}
3116
3117			/*
3118			 * Although ip2 isn't changed here, userspace needs
3119			 * to be warned about the change, so that applications
3120			 * relying on it (like backup ones), will properly
3121			 * notify the change
3122			 */
3123			ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3124			ip2_flags |= XFS_ICHGTIME_CHG;
3125		}
3126	}
3127
3128	if (ip1_flags) {
3129		xfs_trans_ichgtime(tp, ip1, ip1_flags);
3130		xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
3131	}
3132	if (ip2_flags) {
3133		xfs_trans_ichgtime(tp, ip2, ip2_flags);
3134		xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
3135	}
3136	if (dp2_flags) {
3137		xfs_trans_ichgtime(tp, dp2, dp2_flags);
3138		xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
3139	}
3140	xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3141	xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3142	return xfs_finish_rename(tp);
3143
3144out_trans_abort:
3145	xfs_trans_cancel(tp);
3146	return error;
3147}
3148
3149/*
3150 * xfs_rename_alloc_whiteout()
3151 *
3152 * Return a referenced, unlinked, unlocked inode that that can be used as a
3153 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3154 * crash between allocating the inode and linking it into the rename transaction
3155 * recovery will free the inode and we won't leak it.
3156 */
3157static int
3158xfs_rename_alloc_whiteout(
 
 
3159	struct xfs_inode	*dp,
3160	struct xfs_inode	**wip)
3161{
3162	struct xfs_inode	*tmpfile;
 
3163	int			error;
3164
3165	error = xfs_create_tmpfile(dp, S_IFCHR | WHITEOUT_MODE, &tmpfile);
 
3166	if (error)
3167		return error;
3168
 
 
 
 
 
 
 
 
 
3169	/*
3170	 * Prepare the tmpfile inode as if it were created through the VFS.
3171	 * Complete the inode setup and flag it as linkable.  nlink is already
3172	 * zero, so we can skip the drop_nlink.
3173	 */
3174	xfs_setup_iops(tmpfile);
3175	xfs_finish_inode_setup(tmpfile);
3176	VFS_I(tmpfile)->i_state |= I_LINKABLE;
3177
3178	*wip = tmpfile;
3179	return 0;
3180}
3181
3182/*
3183 * xfs_rename
3184 */
3185int
3186xfs_rename(
 
3187	struct xfs_inode	*src_dp,
3188	struct xfs_name		*src_name,
3189	struct xfs_inode	*src_ip,
3190	struct xfs_inode	*target_dp,
3191	struct xfs_name		*target_name,
3192	struct xfs_inode	*target_ip,
3193	unsigned int		flags)
3194{
3195	struct xfs_mount	*mp = src_dp->i_mount;
3196	struct xfs_trans	*tp;
3197	struct xfs_inode	*wip = NULL;		/* whiteout inode */
3198	struct xfs_inode	*inodes[__XFS_SORT_INODES];
 
3199	int			num_inodes = __XFS_SORT_INODES;
3200	bool			new_parent = (src_dp != target_dp);
3201	bool			src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3202	int			spaceres;
3203	int			error;
 
3204
3205	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3206
3207	if ((flags & RENAME_EXCHANGE) && !target_ip)
3208		return -EINVAL;
3209
3210	/*
3211	 * If we are doing a whiteout operation, allocate the whiteout inode
3212	 * we will be placing at the target and ensure the type is set
3213	 * appropriately.
3214	 */
3215	if (flags & RENAME_WHITEOUT) {
3216		ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3217		error = xfs_rename_alloc_whiteout(target_dp, &wip);
3218		if (error)
3219			return error;
3220
3221		/* setup target dirent info as whiteout */
3222		src_name->type = XFS_DIR3_FT_CHRDEV;
3223	}
3224
3225	xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3226				inodes, &num_inodes);
3227
 
 
3228	spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3229	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3230	if (error == -ENOSPC) {
 
3231		spaceres = 0;
3232		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3233				&tp);
3234	}
3235	if (error)
3236		goto out_release_wip;
3237
3238	/*
3239	 * Attach the dquots to the inodes
3240	 */
3241	error = xfs_qm_vop_rename_dqattach(inodes);
3242	if (error)
3243		goto out_trans_cancel;
3244
3245	/*
3246	 * Lock all the participating inodes. Depending upon whether
3247	 * the target_name exists in the target directory, and
3248	 * whether the target directory is the same as the source
3249	 * directory, we can lock from 2 to 4 inodes.
3250	 */
3251	xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3252
3253	/*
3254	 * Join all the inodes to the transaction. From this point on,
3255	 * we can rely on either trans_commit or trans_cancel to unlock
3256	 * them.
3257	 */
3258	xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3259	if (new_parent)
3260		xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3261	xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3262	if (target_ip)
3263		xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3264	if (wip)
3265		xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3266
3267	/*
3268	 * If we are using project inheritance, we only allow renames
3269	 * into our tree when the project IDs are the same; else the
3270	 * tree quota mechanism would be circumvented.
3271	 */
3272	if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
3273		     (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
3274		error = -EXDEV;
3275		goto out_trans_cancel;
3276	}
3277
3278	/* RENAME_EXCHANGE is unique from here on. */
3279	if (flags & RENAME_EXCHANGE)
3280		return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3281					target_dp, target_name, target_ip,
3282					spaceres);
3283
3284	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3285	 * Check for expected errors before we dirty the transaction
3286	 * so we can return an error without a transaction abort.
3287	 */
3288	if (target_ip == NULL) {
3289		/*
3290		 * If there's no space reservation, check the entry will
3291		 * fit before actually inserting it.
3292		 */
3293		if (!spaceres) {
3294			error = xfs_dir_canenter(tp, target_dp, target_name);
3295			if (error)
3296				goto out_trans_cancel;
3297		}
3298	} else {
3299		/*
3300		 * If target exists and it's a directory, check that whether
3301		 * it can be destroyed.
3302		 */
3303		if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3304		    (!xfs_dir_isempty(target_ip) ||
3305		     (VFS_I(target_ip)->i_nlink > 2))) {
3306			error = -EEXIST;
3307			goto out_trans_cancel;
3308		}
3309	}
3310
3311	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3312	 * Directory entry creation below may acquire the AGF. Remove
3313	 * the whiteout from the unlinked list first to preserve correct
3314	 * AGI/AGF locking order. This dirties the transaction so failures
3315	 * after this point will abort and log recovery will clean up the
3316	 * mess.
3317	 *
3318	 * For whiteouts, we need to bump the link count on the whiteout
3319	 * inode. After this point, we have a real link, clear the tmpfile
3320	 * state flag from the inode so it doesn't accidentally get misused
3321	 * in future.
3322	 */
3323	if (wip) {
 
 
3324		ASSERT(VFS_I(wip)->i_nlink == 0);
3325		error = xfs_iunlink_remove(tp, wip);
 
 
 
3326		if (error)
3327			goto out_trans_cancel;
3328
3329		xfs_bumplink(tp, wip);
3330		xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3331		VFS_I(wip)->i_state &= ~I_LINKABLE;
3332	}
3333
3334	/*
3335	 * Set up the target.
3336	 */
3337	if (target_ip == NULL) {
3338		/*
3339		 * If target does not exist and the rename crosses
3340		 * directories, adjust the target directory link count
3341		 * to account for the ".." reference from the new entry.
3342		 */
3343		error = xfs_dir_createname(tp, target_dp, target_name,
3344					   src_ip->i_ino, spaceres);
3345		if (error)
3346			goto out_trans_cancel;
3347
3348		xfs_trans_ichgtime(tp, target_dp,
3349					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3350
3351		if (new_parent && src_is_directory) {
3352			xfs_bumplink(tp, target_dp);
3353		}
3354	} else { /* target_ip != NULL */
3355		/*
3356		 * Link the source inode under the target name.
3357		 * If the source inode is a directory and we are moving
3358		 * it across directories, its ".." entry will be
3359		 * inconsistent until we replace that down below.
3360		 *
3361		 * In case there is already an entry with the same
3362		 * name at the destination directory, remove it first.
3363		 */
3364		error = xfs_dir_replace(tp, target_dp, target_name,
3365					src_ip->i_ino, spaceres);
3366		if (error)
3367			goto out_trans_cancel;
3368
3369		xfs_trans_ichgtime(tp, target_dp,
3370					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3371
3372		/*
3373		 * Decrement the link count on the target since the target
3374		 * dir no longer points to it.
3375		 */
3376		error = xfs_droplink(tp, target_ip);
3377		if (error)
3378			goto out_trans_cancel;
3379
3380		if (src_is_directory) {
3381			/*
3382			 * Drop the link from the old "." entry.
3383			 */
3384			error = xfs_droplink(tp, target_ip);
3385			if (error)
3386				goto out_trans_cancel;
3387		}
3388	} /* target_ip != NULL */
3389
3390	/*
3391	 * Remove the source.
3392	 */
3393	if (new_parent && src_is_directory) {
3394		/*
3395		 * Rewrite the ".." entry to point to the new
3396		 * directory.
3397		 */
3398		error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3399					target_dp->i_ino, spaceres);
3400		ASSERT(error != -EEXIST);
3401		if (error)
3402			goto out_trans_cancel;
3403	}
3404
3405	/*
3406	 * We always want to hit the ctime on the source inode.
3407	 *
3408	 * This isn't strictly required by the standards since the source
3409	 * inode isn't really being changed, but old unix file systems did
3410	 * it and some incremental backup programs won't work without it.
3411	 */
3412	xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3413	xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3414
3415	/*
3416	 * Adjust the link count on src_dp.  This is necessary when
3417	 * renaming a directory, either within one parent when
3418	 * the target existed, or across two parent directories.
3419	 */
3420	if (src_is_directory && (new_parent || target_ip != NULL)) {
3421
3422		/*
3423		 * Decrement link count on src_directory since the
3424		 * entry that's moved no longer points to it.
3425		 */
3426		error = xfs_droplink(tp, src_dp);
3427		if (error)
3428			goto out_trans_cancel;
3429	}
3430
3431	/*
3432	 * For whiteouts, we only need to update the source dirent with the
3433	 * inode number of the whiteout inode rather than removing it
3434	 * altogether.
3435	 */
3436	if (wip) {
3437		error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3438					spaceres);
3439	} else
3440		error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3441					   spaceres);
 
3442	if (error)
3443		goto out_trans_cancel;
3444
3445	xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3446	xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3447	if (new_parent)
3448		xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3449
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3450	error = xfs_finish_rename(tp);
3451	if (wip)
3452		xfs_irele(wip);
3453	return error;
3454
3455out_trans_cancel:
3456	xfs_trans_cancel(tp);
3457out_release_wip:
3458	if (wip)
3459		xfs_irele(wip);
 
 
3460	return error;
3461}
3462
3463STATIC int
3464xfs_iflush_cluster(
3465	struct xfs_inode	*ip,
3466	struct xfs_buf		*bp)
3467{
3468	struct xfs_mount	*mp = ip->i_mount;
3469	struct xfs_perag	*pag;
3470	unsigned long		first_index, mask;
3471	int			cilist_size;
3472	struct xfs_inode	**cilist;
3473	struct xfs_inode	*cip;
3474	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
3475	int			nr_found;
3476	int			clcount = 0;
3477	int			i;
3478
3479	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3480
3481	cilist_size = igeo->inodes_per_cluster * sizeof(struct xfs_inode *);
3482	cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3483	if (!cilist)
3484		goto out_put;
3485
3486	mask = ~(igeo->inodes_per_cluster - 1);
3487	first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3488	rcu_read_lock();
3489	/* really need a gang lookup range call here */
3490	nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3491					first_index, igeo->inodes_per_cluster);
3492	if (nr_found == 0)
3493		goto out_free;
3494
3495	for (i = 0; i < nr_found; i++) {
3496		cip = cilist[i];
3497		if (cip == ip)
3498			continue;
3499
3500		/*
3501		 * because this is an RCU protected lookup, we could find a
3502		 * recently freed or even reallocated inode during the lookup.
3503		 * We need to check under the i_flags_lock for a valid inode
3504		 * here. Skip it if it is not valid or the wrong inode.
3505		 */
3506		spin_lock(&cip->i_flags_lock);
3507		if (!cip->i_ino ||
3508		    __xfs_iflags_test(cip, XFS_ISTALE)) {
3509			spin_unlock(&cip->i_flags_lock);
3510			continue;
3511		}
3512
3513		/*
3514		 * Once we fall off the end of the cluster, no point checking
3515		 * any more inodes in the list because they will also all be
3516		 * outside the cluster.
3517		 */
3518		if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3519			spin_unlock(&cip->i_flags_lock);
3520			break;
3521		}
3522		spin_unlock(&cip->i_flags_lock);
3523
3524		/*
3525		 * Do an un-protected check to see if the inode is dirty and
3526		 * is a candidate for flushing.  These checks will be repeated
3527		 * later after the appropriate locks are acquired.
3528		 */
3529		if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3530			continue;
3531
3532		/*
3533		 * Try to get locks.  If any are unavailable or it is pinned,
3534		 * then this inode cannot be flushed and is skipped.
3535		 */
3536
3537		if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3538			continue;
3539		if (!xfs_iflock_nowait(cip)) {
3540			xfs_iunlock(cip, XFS_ILOCK_SHARED);
3541			continue;
3542		}
3543		if (xfs_ipincount(cip)) {
3544			xfs_ifunlock(cip);
3545			xfs_iunlock(cip, XFS_ILOCK_SHARED);
3546			continue;
3547		}
3548
3549
3550		/*
3551		 * Check the inode number again, just to be certain we are not
3552		 * racing with freeing in xfs_reclaim_inode(). See the comments
3553		 * in that function for more information as to why the initial
3554		 * check is not sufficient.
3555		 */
3556		if (!cip->i_ino) {
3557			xfs_ifunlock(cip);
3558			xfs_iunlock(cip, XFS_ILOCK_SHARED);
3559			continue;
3560		}
3561
3562		/*
3563		 * arriving here means that this inode can be flushed.  First
3564		 * re-check that it's dirty before flushing.
3565		 */
3566		if (!xfs_inode_clean(cip)) {
3567			int	error;
3568			error = xfs_iflush_int(cip, bp);
3569			if (error) {
3570				xfs_iunlock(cip, XFS_ILOCK_SHARED);
3571				goto cluster_corrupt_out;
3572			}
3573			clcount++;
3574		} else {
3575			xfs_ifunlock(cip);
3576		}
3577		xfs_iunlock(cip, XFS_ILOCK_SHARED);
3578	}
3579
3580	if (clcount) {
3581		XFS_STATS_INC(mp, xs_icluster_flushcnt);
3582		XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3583	}
3584
3585out_free:
3586	rcu_read_unlock();
3587	kmem_free(cilist);
3588out_put:
3589	xfs_perag_put(pag);
3590	return 0;
3591
3592
3593cluster_corrupt_out:
3594	/*
3595	 * Corruption detected in the clustering loop.  Invalidate the
3596	 * inode buffer and shut down the filesystem.
3597	 */
3598	rcu_read_unlock();
3599
3600	/*
3601	 * We'll always have an inode attached to the buffer for completion
3602	 * process by the time we are called from xfs_iflush(). Hence we have
3603	 * always need to do IO completion processing to abort the inodes
3604	 * attached to the buffer.  handle them just like the shutdown case in
3605	 * xfs_buf_submit().
3606	 */
3607	ASSERT(bp->b_iodone);
3608	bp->b_flags |= XBF_ASYNC;
3609	bp->b_flags &= ~XBF_DONE;
3610	xfs_buf_stale(bp);
3611	xfs_buf_ioerror(bp, -EIO);
3612	xfs_buf_ioend(bp);
3613
3614	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3615
3616	/* abort the corrupt inode, as it was not attached to the buffer */
3617	xfs_iflush_abort(cip, false);
3618	kmem_free(cilist);
3619	xfs_perag_put(pag);
3620	return -EFSCORRUPTED;
3621}
3622
3623/*
3624 * Flush dirty inode metadata into the backing buffer.
3625 *
3626 * The caller must have the inode lock and the inode flush lock held.  The
3627 * inode lock will still be held upon return to the caller, and the inode
3628 * flush lock will be released after the inode has reached the disk.
3629 *
3630 * The caller must write out the buffer returned in *bpp and release it.
3631 */
3632int
3633xfs_iflush(
3634	struct xfs_inode	*ip,
3635	struct xfs_buf		**bpp)
3636{
3637	struct xfs_mount	*mp = ip->i_mount;
3638	struct xfs_buf		*bp = NULL;
3639	struct xfs_dinode	*dip;
3640	int			error;
3641
3642	XFS_STATS_INC(mp, xs_iflush_count);
3643
3644	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3645	ASSERT(xfs_isiflocked(ip));
3646	ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3647	       ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3648
3649	*bpp = NULL;
3650
3651	xfs_iunpin_wait(ip);
3652
3653	/*
3654	 * For stale inodes we cannot rely on the backing buffer remaining
3655	 * stale in cache for the remaining life of the stale inode and so
3656	 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3657	 * inodes below. We have to check this after ensuring the inode is
3658	 * unpinned so that it is safe to reclaim the stale inode after the
3659	 * flush call.
3660	 */
3661	if (xfs_iflags_test(ip, XFS_ISTALE)) {
3662		xfs_ifunlock(ip);
3663		return 0;
3664	}
3665
3666	/*
3667	 * This may have been unpinned because the filesystem is shutting
3668	 * down forcibly. If that's the case we must not write this inode
3669	 * to disk, because the log record didn't make it to disk.
3670	 *
3671	 * We also have to remove the log item from the AIL in this case,
3672	 * as we wait for an empty AIL as part of the unmount process.
3673	 */
3674	if (XFS_FORCED_SHUTDOWN(mp)) {
3675		error = -EIO;
3676		goto abort_out;
3677	}
3678
3679	/*
3680	 * Get the buffer containing the on-disk inode. We are doing a try-lock
3681	 * operation here, so we may get  an EAGAIN error. In that case, we
3682	 * simply want to return with the inode still dirty.
3683	 *
3684	 * If we get any other error, we effectively have a corruption situation
3685	 * and we cannot flush the inode, so we treat it the same as failing
3686	 * xfs_iflush_int().
3687	 */
3688	error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3689			       0);
3690	if (error == -EAGAIN) {
3691		xfs_ifunlock(ip);
3692		return error;
3693	}
3694	if (error)
3695		goto corrupt_out;
3696
3697	/*
3698	 * First flush out the inode that xfs_iflush was called with.
3699	 */
3700	error = xfs_iflush_int(ip, bp);
3701	if (error)
3702		goto corrupt_out;
3703
3704	/*
3705	 * If the buffer is pinned then push on the log now so we won't
3706	 * get stuck waiting in the write for too long.
3707	 */
3708	if (xfs_buf_ispinned(bp))
3709		xfs_log_force(mp, 0);
3710
3711	/*
3712	 * inode clustering: try to gather other inodes into this write
3713	 *
3714	 * Note: Any error during clustering will result in the filesystem
3715	 * being shut down and completion callbacks run on the cluster buffer.
3716	 * As we have already flushed and attached this inode to the buffer,
3717	 * it has already been aborted and released by xfs_iflush_cluster() and
3718	 * so we have no further error handling to do here.
3719	 */
3720	error = xfs_iflush_cluster(ip, bp);
3721	if (error)
3722		return error;
3723
3724	*bpp = bp;
3725	return 0;
3726
3727corrupt_out:
3728	if (bp)
3729		xfs_buf_relse(bp);
3730	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3731abort_out:
3732	/* abort the corrupt inode, as it was not attached to the buffer */
3733	xfs_iflush_abort(ip, false);
3734	return error;
3735}
3736
3737/*
3738 * If there are inline format data / attr forks attached to this inode,
3739 * make sure they're not corrupt.
3740 */
3741bool
3742xfs_inode_verify_forks(
3743	struct xfs_inode	*ip)
3744{
3745	struct xfs_ifork	*ifp;
3746	xfs_failaddr_t		fa;
3747
3748	fa = xfs_ifork_verify_data(ip, &xfs_default_ifork_ops);
3749	if (fa) {
3750		ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
3751		xfs_inode_verifier_error(ip, -EFSCORRUPTED, "data fork",
3752				ifp->if_u1.if_data, ifp->if_bytes, fa);
3753		return false;
3754	}
3755
3756	fa = xfs_ifork_verify_attr(ip, &xfs_default_ifork_ops);
3757	if (fa) {
3758		ifp = XFS_IFORK_PTR(ip, XFS_ATTR_FORK);
3759		xfs_inode_verifier_error(ip, -EFSCORRUPTED, "attr fork",
3760				ifp ? ifp->if_u1.if_data : NULL,
3761				ifp ? ifp->if_bytes : 0, fa);
3762		return false;
3763	}
3764	return true;
3765}
3766
3767STATIC int
3768xfs_iflush_int(
3769	struct xfs_inode	*ip,
3770	struct xfs_buf		*bp)
3771{
3772	struct xfs_inode_log_item *iip = ip->i_itemp;
3773	struct xfs_dinode	*dip;
3774	struct xfs_mount	*mp = ip->i_mount;
 
3775
3776	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3777	ASSERT(xfs_isiflocked(ip));
3778	ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3779	       ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3780	ASSERT(iip != NULL && iip->ili_fields != 0);
3781	ASSERT(ip->i_d.di_version > 1);
3782
3783	/* set *dip = inode's place in the buffer */
3784	dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3785
 
 
 
 
 
 
 
3786	if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3787			       mp, XFS_ERRTAG_IFLUSH_1)) {
3788		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3789			"%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3790			__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3791		goto corrupt_out;
3792	}
3793	if (S_ISREG(VFS_I(ip)->i_mode)) {
3794		if (XFS_TEST_ERROR(
3795		    (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3796		    (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3797		    mp, XFS_ERRTAG_IFLUSH_3)) {
3798			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3799				"%s: Bad regular inode %Lu, ptr "PTR_FMT,
3800				__func__, ip->i_ino, ip);
3801			goto corrupt_out;
3802		}
3803	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3804		if (XFS_TEST_ERROR(
3805		    (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3806		    (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3807		    (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3808		    mp, XFS_ERRTAG_IFLUSH_4)) {
3809			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3810				"%s: Bad directory inode %Lu, ptr "PTR_FMT,
3811				__func__, ip->i_ino, ip);
3812			goto corrupt_out;
3813		}
3814	}
3815	if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3816				ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3817		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3818			"%s: detected corrupt incore inode %Lu, "
3819			"total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3820			__func__, ip->i_ino,
3821			ip->i_d.di_nextents + ip->i_d.di_anextents,
3822			ip->i_d.di_nblocks, ip);
3823		goto corrupt_out;
3824	}
3825	if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3826				mp, XFS_ERRTAG_IFLUSH_6)) {
3827		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3828			"%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3829			__func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3830		goto corrupt_out;
3831	}
3832
3833	/*
3834	 * Inode item log recovery for v2 inodes are dependent on the
3835	 * di_flushiter count for correct sequencing. We bump the flush
3836	 * iteration count so we can detect flushes which postdate a log record
3837	 * during recovery. This is redundant as we now log every change and
3838	 * hence this can't happen but we need to still do it to ensure
3839	 * backwards compatibility with old kernels that predate logging all
3840	 * inode changes.
3841	 */
3842	if (ip->i_d.di_version < 3)
3843		ip->i_d.di_flushiter++;
3844
3845	/* Check the inline fork data before we write out. */
3846	if (!xfs_inode_verify_forks(ip))
3847		goto corrupt_out;
 
 
 
 
 
 
 
3848
3849	/*
3850	 * Copy the dirty parts of the inode into the on-disk inode.  We always
3851	 * copy out the core of the inode, because if the inode is dirty at all
3852	 * the core must be.
3853	 */
3854	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3855
3856	/* Wrap, we never let the log put out DI_MAX_FLUSH */
3857	if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3858		ip->i_d.di_flushiter = 0;
 
 
3859
3860	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3861	if (XFS_IFORK_Q(ip))
3862		xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3863	xfs_inobp_check(mp, bp);
3864
3865	/*
3866	 * We've recorded everything logged in the inode, so we'd like to clear
3867	 * the ili_fields bits so we don't log and flush things unnecessarily.
3868	 * However, we can't stop logging all this information until the data
3869	 * we've copied into the disk buffer is written to disk.  If we did we
3870	 * might overwrite the copy of the inode in the log with all the data
3871	 * after re-logging only part of it, and in the face of a crash we
3872	 * wouldn't have all the data we need to recover.
3873	 *
3874	 * What we do is move the bits to the ili_last_fields field.  When
3875	 * logging the inode, these bits are moved back to the ili_fields field.
3876	 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3877	 * know that the information those bits represent is permanently on
3878	 * disk.  As long as the flush completes before the inode is logged
3879	 * again, then both ili_fields and ili_last_fields will be cleared.
3880	 *
3881	 * We can play with the ili_fields bits here, because the inode lock
3882	 * must be held exclusively in order to set bits there and the flush
3883	 * lock protects the ili_last_fields bits.  Set ili_logged so the flush
3884	 * done routine can tell whether or not to look in the AIL.  Also, store
3885	 * the current LSN of the inode so that we can tell whether the item has
3886	 * moved in the AIL from xfs_iflush_done().  In order to read the lsn we
3887	 * need the AIL lock, because it is a 64 bit value that cannot be read
3888	 * atomically.
3889	 */
 
 
 
3890	iip->ili_last_fields = iip->ili_fields;
3891	iip->ili_fields = 0;
3892	iip->ili_fsync_fields = 0;
3893	iip->ili_logged = 1;
3894
 
 
 
 
3895	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3896				&iip->ili_item.li_lsn);
3897
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3898	/*
3899	 * Attach the function xfs_iflush_done to the inode's
3900	 * buffer.  This will remove the inode from the AIL
3901	 * and unlock the inode's flush lock when the inode is
3902	 * completely written to disk.
3903	 */
3904	xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
 
 
3905
3906	/* generate the checksum. */
3907	xfs_dinode_calc_crc(mp, dip);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3908
3909	ASSERT(!list_empty(&bp->b_li_list));
3910	ASSERT(bp->b_iodone != NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3911	return 0;
3912
3913corrupt_out:
3914	return -EFSCORRUPTED;
3915}
3916
3917/* Release an inode. */
3918void
3919xfs_irele(
3920	struct xfs_inode	*ip)
3921{
3922	trace_xfs_irele(ip, _RET_IP_);
3923	iput(VFS_I(ip));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3924}
v6.9.4
   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_iunlink_item.h"
  24#include "xfs_ialloc.h"
  25#include "xfs_bmap.h"
  26#include "xfs_bmap_util.h"
  27#include "xfs_errortag.h"
  28#include "xfs_error.h"
  29#include "xfs_quota.h"
  30#include "xfs_filestream.h"
  31#include "xfs_trace.h"
  32#include "xfs_icache.h"
  33#include "xfs_symlink.h"
  34#include "xfs_trans_priv.h"
  35#include "xfs_log.h"
  36#include "xfs_bmap_btree.h"
  37#include "xfs_reflink.h"
  38#include "xfs_ag.h"
  39#include "xfs_log_priv.h"
  40#include "xfs_health.h"
  41
  42struct kmem_cache *xfs_inode_cache;
  43
 
 
 
 
 
 
 
  44STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
  45STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
  46	struct xfs_inode *);
  47
  48/*
  49 * helper function to extract extent size hint from inode
  50 */
  51xfs_extlen_t
  52xfs_get_extsz_hint(
  53	struct xfs_inode	*ip)
  54{
  55	/*
  56	 * No point in aligning allocations if we need to COW to actually
  57	 * write to them.
  58	 */
  59	if (xfs_is_always_cow_inode(ip))
  60		return 0;
  61	if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
  62		return ip->i_extsize;
  63	if (XFS_IS_REALTIME_INODE(ip))
  64		return ip->i_mount->m_sb.sb_rextsize;
  65	return 0;
  66}
  67
  68/*
  69 * Helper function to extract CoW extent size hint from inode.
  70 * Between the extent size hint and the CoW extent size hint, we
  71 * return the greater of the two.  If the value is zero (automatic),
  72 * use the default size.
  73 */
  74xfs_extlen_t
  75xfs_get_cowextsz_hint(
  76	struct xfs_inode	*ip)
  77{
  78	xfs_extlen_t		a, b;
  79
  80	a = 0;
  81	if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
  82		a = ip->i_cowextsize;
  83	b = xfs_get_extsz_hint(ip);
  84
  85	a = max(a, b);
  86	if (a == 0)
  87		return XFS_DEFAULT_COWEXTSZ_HINT;
  88	return a;
  89}
  90
  91/*
  92 * These two are wrapper routines around the xfs_ilock() routine used to
  93 * centralize some grungy code.  They are used in places that wish to lock the
  94 * inode solely for reading the extents.  The reason these places can't just
  95 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
  96 * bringing in of the extents from disk for a file in b-tree format.  If the
  97 * inode is in b-tree format, then we need to lock the inode exclusively until
  98 * the extents are read in.  Locking it exclusively all the time would limit
  99 * our parallelism unnecessarily, though.  What we do instead is check to see
 100 * if the extents have been read in yet, and only lock the inode exclusively
 101 * if they have not.
 102 *
 103 * The functions return a value which should be given to the corresponding
 104 * xfs_iunlock() call.
 105 */
 106uint
 107xfs_ilock_data_map_shared(
 108	struct xfs_inode	*ip)
 109{
 110	uint			lock_mode = XFS_ILOCK_SHARED;
 111
 112	if (xfs_need_iread_extents(&ip->i_df))
 
 113		lock_mode = XFS_ILOCK_EXCL;
 114	xfs_ilock(ip, lock_mode);
 115	return lock_mode;
 116}
 117
 118uint
 119xfs_ilock_attr_map_shared(
 120	struct xfs_inode	*ip)
 121{
 122	uint			lock_mode = XFS_ILOCK_SHARED;
 123
 124	if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
 
 125		lock_mode = XFS_ILOCK_EXCL;
 126	xfs_ilock(ip, lock_mode);
 127	return lock_mode;
 128}
 129
 130/*
 131 * You can't set both SHARED and EXCL for the same lock,
 132 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
 133 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
 134 * to set in lock_flags.
 135 */
 136static inline void
 137xfs_lock_flags_assert(
 138	uint		lock_flags)
 139{
 140	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
 141		(XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
 142	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
 143		(XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
 144	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
 145		(XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
 146	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
 147	ASSERT(lock_flags != 0);
 148}
 149
 150/*
 151 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
 152 * multi-reader locks: invalidate_lock and the i_lock.  This routine allows
 153 * various combinations of the locks to be obtained.
 154 *
 155 * The 3 locks should always be ordered so that the IO lock is obtained first,
 156 * the mmap lock second and the ilock last in order to prevent deadlock.
 157 *
 158 * Basic locking order:
 159 *
 160 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
 161 *
 162 * mmap_lock locking order:
 163 *
 164 * i_rwsem -> page lock -> mmap_lock
 165 * mmap_lock -> invalidate_lock -> page_lock
 166 *
 167 * The difference in mmap_lock locking order mean that we cannot hold the
 168 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
 169 * can fault in pages during copy in/out (for buffered IO) or require the
 170 * mmap_lock in get_user_pages() to map the user pages into the kernel address
 171 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
 172 * fault because page faults already hold the mmap_lock.
 173 *
 174 * Hence to serialise fully against both syscall and mmap based IO, we need to
 175 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
 176 * both taken in places where we need to invalidate the page cache in a race
 177 * free manner (e.g. truncate, hole punch and other extent manipulation
 178 * functions).
 179 */
 180void
 181xfs_ilock(
 182	xfs_inode_t		*ip,
 183	uint			lock_flags)
 184{
 185	trace_xfs_ilock(ip, lock_flags, _RET_IP_);
 186
 187	xfs_lock_flags_assert(lock_flags);
 
 
 
 
 
 
 
 
 
 
 
 188
 189	if (lock_flags & XFS_IOLOCK_EXCL) {
 190		down_write_nested(&VFS_I(ip)->i_rwsem,
 191				  XFS_IOLOCK_DEP(lock_flags));
 192	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 193		down_read_nested(&VFS_I(ip)->i_rwsem,
 194				 XFS_IOLOCK_DEP(lock_flags));
 195	}
 196
 197	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 198		down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
 199				  XFS_MMAPLOCK_DEP(lock_flags));
 200	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 201		down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
 202				 XFS_MMAPLOCK_DEP(lock_flags));
 203	}
 204
 205	if (lock_flags & XFS_ILOCK_EXCL)
 206		down_write_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 207	else if (lock_flags & XFS_ILOCK_SHARED)
 208		down_read_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 209}
 210
 211/*
 212 * This is just like xfs_ilock(), except that the caller
 213 * is guaranteed not to sleep.  It returns 1 if it gets
 214 * the requested locks and 0 otherwise.  If the IO lock is
 215 * obtained but the inode lock cannot be, then the IO lock
 216 * is dropped before returning.
 217 *
 218 * ip -- the inode being locked
 219 * lock_flags -- this parameter indicates the inode's locks to be
 220 *       to be locked.  See the comment for xfs_ilock() for a list
 221 *	 of valid values.
 222 */
 223int
 224xfs_ilock_nowait(
 225	xfs_inode_t		*ip,
 226	uint			lock_flags)
 227{
 228	trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
 229
 230	xfs_lock_flags_assert(lock_flags);
 
 
 
 
 
 
 
 
 
 
 
 231
 232	if (lock_flags & XFS_IOLOCK_EXCL) {
 233		if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
 234			goto out;
 235	} else if (lock_flags & XFS_IOLOCK_SHARED) {
 236		if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
 237			goto out;
 238	}
 239
 240	if (lock_flags & XFS_MMAPLOCK_EXCL) {
 241		if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
 242			goto out_undo_iolock;
 243	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
 244		if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
 245			goto out_undo_iolock;
 246	}
 247
 248	if (lock_flags & XFS_ILOCK_EXCL) {
 249		if (!down_write_trylock(&ip->i_lock))
 250			goto out_undo_mmaplock;
 251	} else if (lock_flags & XFS_ILOCK_SHARED) {
 252		if (!down_read_trylock(&ip->i_lock))
 253			goto out_undo_mmaplock;
 254	}
 255	return 1;
 256
 257out_undo_mmaplock:
 258	if (lock_flags & XFS_MMAPLOCK_EXCL)
 259		up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 260	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 261		up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
 262out_undo_iolock:
 263	if (lock_flags & XFS_IOLOCK_EXCL)
 264		up_write(&VFS_I(ip)->i_rwsem);
 265	else if (lock_flags & XFS_IOLOCK_SHARED)
 266		up_read(&VFS_I(ip)->i_rwsem);
 267out:
 268	return 0;
 269}
 270
 271/*
 272 * xfs_iunlock() is used to drop the inode locks acquired with
 273 * xfs_ilock() and xfs_ilock_nowait().  The caller must pass
 274 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
 275 * that we know which locks to drop.
 276 *
 277 * ip -- the inode being unlocked
 278 * lock_flags -- this parameter indicates the inode's locks to be
 279 *       to be unlocked.  See the comment for xfs_ilock() for a list
 280 *	 of valid values for this parameter.
 281 *
 282 */
 283void
 284xfs_iunlock(
 285	xfs_inode_t		*ip,
 286	uint			lock_flags)
 287{
 288	xfs_lock_flags_assert(lock_flags);
 
 
 
 
 
 
 
 
 
 
 
 
 289
 290	if (lock_flags & XFS_IOLOCK_EXCL)
 291		up_write(&VFS_I(ip)->i_rwsem);
 292	else if (lock_flags & XFS_IOLOCK_SHARED)
 293		up_read(&VFS_I(ip)->i_rwsem);
 294
 295	if (lock_flags & XFS_MMAPLOCK_EXCL)
 296		up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 297	else if (lock_flags & XFS_MMAPLOCK_SHARED)
 298		up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
 299
 300	if (lock_flags & XFS_ILOCK_EXCL)
 301		up_write(&ip->i_lock);
 302	else if (lock_flags & XFS_ILOCK_SHARED)
 303		up_read(&ip->i_lock);
 304
 305	trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
 306}
 307
 308/*
 309 * give up write locks.  the i/o lock cannot be held nested
 310 * if it is being demoted.
 311 */
 312void
 313xfs_ilock_demote(
 314	xfs_inode_t		*ip,
 315	uint			lock_flags)
 316{
 317	ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
 318	ASSERT((lock_flags &
 319		~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
 320
 321	if (lock_flags & XFS_ILOCK_EXCL)
 322		downgrade_write(&ip->i_lock);
 323	if (lock_flags & XFS_MMAPLOCK_EXCL)
 324		downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 325	if (lock_flags & XFS_IOLOCK_EXCL)
 326		downgrade_write(&VFS_I(ip)->i_rwsem);
 327
 328	trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
 329}
 330
 331void
 332xfs_assert_ilocked(
 333	struct xfs_inode	*ip,
 
 334	uint			lock_flags)
 335{
 336	/*
 337	 * Sometimes we assert the ILOCK is held exclusively, but we're in
 338	 * a workqueue, so lockdep doesn't know we're the owner.
 339	 */
 340	if (lock_flags & XFS_ILOCK_SHARED)
 341		rwsem_assert_held(&ip->i_lock);
 342	else if (lock_flags & XFS_ILOCK_EXCL)
 343		rwsem_assert_held_write_nolockdep(&ip->i_lock);
 344
 345	if (lock_flags & XFS_MMAPLOCK_SHARED)
 346		rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
 347	else if (lock_flags & XFS_MMAPLOCK_EXCL)
 348		rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
 349
 350	if (lock_flags & XFS_IOLOCK_SHARED)
 351		rwsem_assert_held(&VFS_I(ip)->i_rwsem);
 352	else if (lock_flags & XFS_IOLOCK_EXCL)
 353		rwsem_assert_held_write(&VFS_I(ip)->i_rwsem);
 
 
 
 354}
 
 355
 356/*
 357 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
 358 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
 359 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
 360 * errors and warnings.
 361 */
 362#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
 363static bool
 364xfs_lockdep_subclass_ok(
 365	int subclass)
 366{
 367	return subclass < MAX_LOCKDEP_SUBCLASSES;
 368}
 369#else
 370#define xfs_lockdep_subclass_ok(subclass)	(true)
 371#endif
 372
 373/*
 374 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
 375 * value. This can be called for any type of inode lock combination, including
 376 * parent locking. Care must be taken to ensure we don't overrun the subclass
 377 * storage fields in the class mask we build.
 378 */
 379static inline uint
 380xfs_lock_inumorder(
 381	uint	lock_mode,
 382	uint	subclass)
 383{
 384	uint	class = 0;
 385
 386	ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
 387			      XFS_ILOCK_RTSUM)));
 388	ASSERT(xfs_lockdep_subclass_ok(subclass));
 389
 390	if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
 391		ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
 392		class += subclass << XFS_IOLOCK_SHIFT;
 393	}
 394
 395	if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
 396		ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
 397		class += subclass << XFS_MMAPLOCK_SHIFT;
 398	}
 399
 400	if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
 401		ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
 402		class += subclass << XFS_ILOCK_SHIFT;
 403	}
 404
 405	return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
 406}
 407
 408/*
 409 * The following routine will lock n inodes in exclusive mode.  We assume the
 410 * caller calls us with the inodes in i_ino order.
 411 *
 412 * We need to detect deadlock where an inode that we lock is in the AIL and we
 413 * start waiting for another inode that is locked by a thread in a long running
 414 * transaction (such as truncate). This can result in deadlock since the long
 415 * running trans might need to wait for the inode we just locked in order to
 416 * push the tail and free space in the log.
 417 *
 418 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
 419 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
 420 * lock more than one at a time, lockdep will report false positives saying we
 421 * have violated locking orders.
 422 */
 423static void
 424xfs_lock_inodes(
 425	struct xfs_inode	**ips,
 426	int			inodes,
 427	uint			lock_mode)
 428{
 429	int			attempts = 0;
 430	uint			i;
 431	int			j;
 432	bool			try_lock;
 433	struct xfs_log_item	*lp;
 434
 435	/*
 436	 * Currently supports between 2 and 5 inodes with exclusive locking.  We
 437	 * support an arbitrary depth of locking here, but absolute limits on
 438	 * inodes depend on the type of locking and the limits placed by
 439	 * lockdep annotations in xfs_lock_inumorder.  These are all checked by
 440	 * the asserts.
 441	 */
 442	ASSERT(ips && inodes >= 2 && inodes <= 5);
 443	ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
 444			    XFS_ILOCK_EXCL));
 445	ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
 446			      XFS_ILOCK_SHARED)));
 447	ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
 448		inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
 449	ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
 450		inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
 451
 452	if (lock_mode & XFS_IOLOCK_EXCL) {
 453		ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
 454	} else if (lock_mode & XFS_MMAPLOCK_EXCL)
 455		ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
 456
 
 
 457again:
 458	try_lock = false;
 459	i = 0;
 460	for (; i < inodes; i++) {
 461		ASSERT(ips[i]);
 462
 463		if (i && (ips[i] == ips[i - 1]))	/* Already locked */
 464			continue;
 465
 466		/*
 467		 * If try_lock is not set yet, make sure all locked inodes are
 468		 * not in the AIL.  If any are, set try_lock to be used later.
 469		 */
 470		if (!try_lock) {
 471			for (j = (i - 1); j >= 0 && !try_lock; j--) {
 472				lp = &ips[j]->i_itemp->ili_item;
 473				if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
 474					try_lock = true;
 475			}
 476		}
 477
 478		/*
 479		 * If any of the previous locks we have locked is in the AIL,
 480		 * we must TRY to get the second and subsequent locks. If
 481		 * we can't get any, we must release all we have
 482		 * and try again.
 483		 */
 484		if (!try_lock) {
 485			xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
 486			continue;
 487		}
 488
 489		/* try_lock means we have an inode locked that is in the AIL. */
 490		ASSERT(i != 0);
 491		if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
 492			continue;
 493
 494		/*
 495		 * Unlock all previous guys and try again.  xfs_iunlock will try
 496		 * to push the tail if the inode is in the AIL.
 497		 */
 498		attempts++;
 499		for (j = i - 1; j >= 0; j--) {
 500			/*
 501			 * Check to see if we've already unlocked this one.  Not
 502			 * the first one going back, and the inode ptr is the
 503			 * same.
 504			 */
 505			if (j != (i - 1) && ips[j] == ips[j + 1])
 506				continue;
 507
 508			xfs_iunlock(ips[j], lock_mode);
 509		}
 510
 511		if ((attempts % 5) == 0) {
 512			delay(1); /* Don't just spin the CPU */
 513		}
 
 
 514		goto again;
 515	}
 516}
 517
 518/*
 519 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
 520 * mmaplock must be double-locked separately since we use i_rwsem and
 521 * invalidate_lock for that. We now support taking one lock EXCL and the
 522 * other SHARED.
 
 
 523 */
 524void
 525xfs_lock_two_inodes(
 526	struct xfs_inode	*ip0,
 527	uint			ip0_mode,
 528	struct xfs_inode	*ip1,
 529	uint			ip1_mode)
 530{
 
 
 531	int			attempts = 0;
 532	struct xfs_log_item	*lp;
 533
 534	ASSERT(hweight32(ip0_mode) == 1);
 535	ASSERT(hweight32(ip1_mode) == 1);
 536	ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 537	ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
 538	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
 539	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
 
 
 
 
 
 
 
 540	ASSERT(ip0->i_ino != ip1->i_ino);
 541
 542	if (ip0->i_ino > ip1->i_ino) {
 543		swap(ip0, ip1);
 544		swap(ip0_mode, ip1_mode);
 
 
 
 
 545	}
 546
 547 again:
 548	xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
 549
 550	/*
 551	 * If the first lock we have locked is in the AIL, we must TRY to get
 552	 * the second lock. If we can't get it, we must release the first one
 553	 * and try again.
 554	 */
 555	lp = &ip0->i_itemp->ili_item;
 556	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
 557		if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
 558			xfs_iunlock(ip0, ip0_mode);
 559			if ((++attempts % 5) == 0)
 560				delay(1); /* Don't just spin the CPU */
 561			goto again;
 562		}
 563	} else {
 564		xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
 565	}
 566}
 567
 568uint
 569xfs_ip2xflags(
 570	struct xfs_inode	*ip)
 571{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 572	uint			flags = 0;
 573
 574	if (ip->i_diflags & XFS_DIFLAG_ANY) {
 575		if (ip->i_diflags & XFS_DIFLAG_REALTIME)
 576			flags |= FS_XFLAG_REALTIME;
 577		if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
 578			flags |= FS_XFLAG_PREALLOC;
 579		if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
 580			flags |= FS_XFLAG_IMMUTABLE;
 581		if (ip->i_diflags & XFS_DIFLAG_APPEND)
 582			flags |= FS_XFLAG_APPEND;
 583		if (ip->i_diflags & XFS_DIFLAG_SYNC)
 584			flags |= FS_XFLAG_SYNC;
 585		if (ip->i_diflags & XFS_DIFLAG_NOATIME)
 586			flags |= FS_XFLAG_NOATIME;
 587		if (ip->i_diflags & XFS_DIFLAG_NODUMP)
 588			flags |= FS_XFLAG_NODUMP;
 589		if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
 590			flags |= FS_XFLAG_RTINHERIT;
 591		if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
 592			flags |= FS_XFLAG_PROJINHERIT;
 593		if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
 594			flags |= FS_XFLAG_NOSYMLINKS;
 595		if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
 596			flags |= FS_XFLAG_EXTSIZE;
 597		if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
 598			flags |= FS_XFLAG_EXTSZINHERIT;
 599		if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
 600			flags |= FS_XFLAG_NODEFRAG;
 601		if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
 602			flags |= FS_XFLAG_FILESTREAM;
 603	}
 604
 605	if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
 606		if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
 607			flags |= FS_XFLAG_DAX;
 608		if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
 609			flags |= FS_XFLAG_COWEXTSIZE;
 610	}
 611
 612	if (xfs_inode_has_attr_fork(ip))
 613		flags |= FS_XFLAG_HASATTR;
 
 614	return flags;
 615}
 616
 
 
 
 
 
 
 
 
 
 617/*
 618 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 619 * is allowed, otherwise it has to be an exact match. If a CI match is found,
 620 * ci_name->name will point to a the actual name (caller must free) or
 621 * will be set to NULL if an exact match is found.
 622 */
 623int
 624xfs_lookup(
 625	struct xfs_inode	*dp,
 626	const struct xfs_name	*name,
 627	struct xfs_inode	**ipp,
 628	struct xfs_name		*ci_name)
 629{
 630	xfs_ino_t		inum;
 631	int			error;
 632
 633	trace_xfs_lookup(dp, name);
 634
 635	if (xfs_is_shutdown(dp->i_mount))
 636		return -EIO;
 637	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
 638		return -EIO;
 639
 640	error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 641	if (error)
 642		goto out_unlock;
 643
 644	error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 645	if (error)
 646		goto out_free_name;
 647
 648	return 0;
 649
 650out_free_name:
 651	if (ci_name)
 652		kfree(ci_name->name);
 653out_unlock:
 654	*ipp = NULL;
 655	return error;
 656}
 657
 658/* Propagate di_flags from a parent inode to a child inode. */
 659static void
 660xfs_inode_inherit_flags(
 661	struct xfs_inode	*ip,
 662	const struct xfs_inode	*pip)
 663{
 664	unsigned int		di_flags = 0;
 665	xfs_failaddr_t		failaddr;
 666	umode_t			mode = VFS_I(ip)->i_mode;
 667
 668	if (S_ISDIR(mode)) {
 669		if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
 670			di_flags |= XFS_DIFLAG_RTINHERIT;
 671		if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
 672			di_flags |= XFS_DIFLAG_EXTSZINHERIT;
 673			ip->i_extsize = pip->i_extsize;
 674		}
 675		if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
 676			di_flags |= XFS_DIFLAG_PROJINHERIT;
 677	} else if (S_ISREG(mode)) {
 678		if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
 679		    xfs_has_realtime(ip->i_mount))
 680			di_flags |= XFS_DIFLAG_REALTIME;
 681		if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
 682			di_flags |= XFS_DIFLAG_EXTSIZE;
 683			ip->i_extsize = pip->i_extsize;
 684		}
 685	}
 686	if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
 687	    xfs_inherit_noatime)
 688		di_flags |= XFS_DIFLAG_NOATIME;
 689	if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
 690	    xfs_inherit_nodump)
 691		di_flags |= XFS_DIFLAG_NODUMP;
 692	if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
 693	    xfs_inherit_sync)
 694		di_flags |= XFS_DIFLAG_SYNC;
 695	if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
 696	    xfs_inherit_nosymlinks)
 697		di_flags |= XFS_DIFLAG_NOSYMLINKS;
 698	if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
 699	    xfs_inherit_nodefrag)
 700		di_flags |= XFS_DIFLAG_NODEFRAG;
 701	if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
 702		di_flags |= XFS_DIFLAG_FILESTREAM;
 703
 704	ip->i_diflags |= di_flags;
 705
 706	/*
 707	 * Inode verifiers on older kernels only check that the extent size
 708	 * hint is an integer multiple of the rt extent size on realtime files.
 709	 * They did not check the hint alignment on a directory with both
 710	 * rtinherit and extszinherit flags set.  If the misaligned hint is
 711	 * propagated from a directory into a new realtime file, new file
 712	 * allocations will fail due to math errors in the rt allocator and/or
 713	 * trip the verifiers.  Validate the hint settings in the new file so
 714	 * that we don't let broken hints propagate.
 715	 */
 716	failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
 717			VFS_I(ip)->i_mode, ip->i_diflags);
 718	if (failaddr) {
 719		ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
 720				   XFS_DIFLAG_EXTSZINHERIT);
 721		ip->i_extsize = 0;
 722	}
 723}
 724
 725/* Propagate di_flags2 from a parent inode to a child inode. */
 726static void
 727xfs_inode_inherit_flags2(
 728	struct xfs_inode	*ip,
 729	const struct xfs_inode	*pip)
 730{
 731	xfs_failaddr_t		failaddr;
 732
 733	if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
 734		ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
 735		ip->i_cowextsize = pip->i_cowextsize;
 736	}
 737	if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
 738		ip->i_diflags2 |= XFS_DIFLAG2_DAX;
 739
 740	/* Don't let invalid cowextsize hints propagate. */
 741	failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
 742			VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
 743	if (failaddr) {
 744		ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
 745		ip->i_cowextsize = 0;
 746	}
 747}
 748
 749/*
 750 * Initialise a newly allocated inode and return the in-core inode to the
 751 * caller locked exclusively.
 752 */
 753int
 754xfs_init_new_inode(
 755	struct mnt_idmap	*idmap,
 756	struct xfs_trans	*tp,
 757	struct xfs_inode	*pip,
 758	xfs_ino_t		ino,
 759	umode_t			mode,
 760	xfs_nlink_t		nlink,
 761	dev_t			rdev,
 762	prid_t			prid,
 763	bool			init_xattrs,
 764	struct xfs_inode	**ipp)
 765{
 766	struct inode		*dir = pip ? VFS_I(pip) : NULL;
 767	struct xfs_mount	*mp = tp->t_mountp;
 768	struct xfs_inode	*ip;
 769	unsigned int		flags;
 770	int			error;
 771	struct timespec64	tv;
 772	struct inode		*inode;
 773
 774	/*
 775	 * Protect against obviously corrupt allocation btree records. Later
 776	 * xfs_iget checks will catch re-allocation of other active in-memory
 777	 * and on-disk inodes. If we don't catch reallocating the parent inode
 778	 * here we will deadlock in xfs_iget() so we have to do these checks
 779	 * first.
 780	 */
 781	if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
 782		xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
 783		xfs_agno_mark_sick(mp, XFS_INO_TO_AGNO(mp, ino),
 784				XFS_SICK_AG_INOBT);
 785		return -EFSCORRUPTED;
 786	}
 787
 788	/*
 789	 * Get the in-core inode with the lock held exclusively to prevent
 790	 * others from looking at until we're done.
 
 791	 */
 792	error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
 
 793	if (error)
 794		return error;
 795
 796	ASSERT(ip != NULL);
 797	inode = VFS_I(ip);
 
 
 
 
 
 
 
 
 
 
 798	set_nlink(inode, nlink);
 
 
 799	inode->i_rdev = rdev;
 800	ip->i_projid = prid;
 801
 802	if (dir && !(dir->i_mode & S_ISGID) && xfs_has_grpid(mp)) {
 803		inode_fsuid_set(inode, idmap);
 804		inode->i_gid = dir->i_gid;
 805		inode->i_mode = mode;
 806	} else {
 807		inode_init_owner(idmap, inode, dir, mode);
 808	}
 809
 810	/*
 811	 * If the group ID of the new file does not match the effective group
 812	 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
 813	 * (and only if the irix_sgid_inherit compatibility variable is set).
 814	 */
 815	if (irix_sgid_inherit && (inode->i_mode & S_ISGID) &&
 816	    !vfsgid_in_group_p(i_gid_into_vfsgid(idmap, inode)))
 
 817		inode->i_mode &= ~S_ISGID;
 818
 819	ip->i_disk_size = 0;
 820	ip->i_df.if_nextents = 0;
 821	ASSERT(ip->i_nblocks == 0);
 822
 823	tv = inode_set_ctime_current(inode);
 824	inode_set_mtime_to_ts(inode, tv);
 825	inode_set_atime_to_ts(inode, tv);
 
 
 
 
 
 
 826
 827	ip->i_extsize = 0;
 828	ip->i_diflags = 0;
 829
 830	if (xfs_has_v3inodes(mp)) {
 831		inode_set_iversion(inode, 1);
 832		ip->i_cowextsize = 0;
 833		ip->i_crtime = tv;
 
 
 834	}
 835
 
 836	flags = XFS_ILOG_CORE;
 837	switch (mode & S_IFMT) {
 838	case S_IFIFO:
 839	case S_IFCHR:
 840	case S_IFBLK:
 841	case S_IFSOCK:
 842		ip->i_df.if_format = XFS_DINODE_FMT_DEV;
 
 843		flags |= XFS_ILOG_DEV;
 844		break;
 845	case S_IFREG:
 846	case S_IFDIR:
 847		if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
 848			xfs_inode_inherit_flags(ip, pip);
 849		if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
 850			xfs_inode_inherit_flags2(ip, pip);
 851		fallthrough;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 852	case S_IFLNK:
 853		ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
 
 854		ip->i_df.if_bytes = 0;
 855		ip->i_df.if_data = NULL;
 856		break;
 857	default:
 858		ASSERT(0);
 859	}
 860
 861	/*
 862	 * If we need to create attributes immediately after allocating the
 863	 * inode, initialise an empty attribute fork right now. We use the
 864	 * default fork offset for attributes here as we don't know exactly what
 865	 * size or how many attributes we might be adding. We can do this
 866	 * safely here because we know the data fork is completely empty and
 867	 * this saves us from needing to run a separate transaction to set the
 868	 * fork offset in the immediate future.
 869	 */
 870	if (init_xattrs && xfs_has_attr(mp)) {
 871		ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
 872		xfs_ifork_init_attr(ip, XFS_DINODE_FMT_EXTENTS, 0);
 873	}
 874
 875	/*
 876	 * Log the new values stuffed into the inode.
 877	 */
 878	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 879	xfs_trans_log_inode(tp, ip, flags);
 880
 881	/* now that we have an i_mode we can setup the inode structure */
 882	xfs_setup_inode(ip);
 883
 884	*ipp = ip;
 885	return 0;
 886}
 887
 888/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 889 * Decrement the link count on an inode & log the change.  If this causes the
 890 * link count to go to zero, move the inode to AGI unlinked list so that it can
 891 * be freed when the last active reference goes away via xfs_inactive().
 892 */
 893static int			/* error */
 894xfs_droplink(
 895	xfs_trans_t *tp,
 896	xfs_inode_t *ip)
 897{
 898	if (VFS_I(ip)->i_nlink == 0) {
 899		xfs_alert(ip->i_mount,
 900			  "%s: Attempt to drop inode (%llu) with nlink zero.",
 901			  __func__, ip->i_ino);
 902		return -EFSCORRUPTED;
 903	}
 904
 905	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
 906
 907	drop_nlink(VFS_I(ip));
 908	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 909
 910	if (VFS_I(ip)->i_nlink)
 911		return 0;
 912
 913	return xfs_iunlink(tp, ip);
 914}
 915
 916/*
 917 * Increment the link count on an inode & log the change.
 918 */
 919static void
 920xfs_bumplink(
 921	xfs_trans_t *tp,
 922	xfs_inode_t *ip)
 923{
 924	xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
 925
 
 926	inc_nlink(VFS_I(ip));
 927	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 928}
 929
 930#ifdef CONFIG_XFS_LIVE_HOOKS
 931/*
 932 * Use a static key here to reduce the overhead of directory live update hooks.
 933 * If the compiler supports jump labels, the static branch will be replaced by
 934 * a nop sled when there are no hook users.  Online fsck is currently the only
 935 * caller, so this is a reasonable tradeoff.
 936 *
 937 * Note: Patching the kernel code requires taking the cpu hotplug lock.  Other
 938 * parts of the kernel allocate memory with that lock held, which means that
 939 * XFS callers cannot hold any locks that might be used by memory reclaim or
 940 * writeback when calling the static_branch_{inc,dec} functions.
 941 */
 942DEFINE_STATIC_XFS_HOOK_SWITCH(xfs_dir_hooks_switch);
 943
 944void
 945xfs_dir_hook_disable(void)
 946{
 947	xfs_hooks_switch_off(&xfs_dir_hooks_switch);
 948}
 949
 950void
 951xfs_dir_hook_enable(void)
 952{
 953	xfs_hooks_switch_on(&xfs_dir_hooks_switch);
 954}
 955
 956/* Call hooks for a directory update relating to a child dirent update. */
 957inline void
 958xfs_dir_update_hook(
 959	struct xfs_inode		*dp,
 960	struct xfs_inode		*ip,
 961	int				delta,
 962	const struct xfs_name		*name)
 963{
 964	if (xfs_hooks_switched_on(&xfs_dir_hooks_switch)) {
 965		struct xfs_dir_update_params	p = {
 966			.dp		= dp,
 967			.ip		= ip,
 968			.delta		= delta,
 969			.name		= name,
 970		};
 971		struct xfs_mount	*mp = ip->i_mount;
 972
 973		xfs_hooks_call(&mp->m_dir_update_hooks, 0, &p);
 974	}
 975}
 976
 977/* Call the specified function during a directory update. */
 978int
 979xfs_dir_hook_add(
 980	struct xfs_mount	*mp,
 981	struct xfs_dir_hook	*hook)
 982{
 983	return xfs_hooks_add(&mp->m_dir_update_hooks, &hook->dirent_hook);
 984}
 985
 986/* Stop calling the specified function during a directory update. */
 987void
 988xfs_dir_hook_del(
 989	struct xfs_mount	*mp,
 990	struct xfs_dir_hook	*hook)
 991{
 992	xfs_hooks_del(&mp->m_dir_update_hooks, &hook->dirent_hook);
 993}
 994
 995/* Configure directory update hook functions. */
 996void
 997xfs_dir_hook_setup(
 998	struct xfs_dir_hook	*hook,
 999	notifier_fn_t		mod_fn)
1000{
1001	xfs_hook_setup(&hook->dirent_hook, mod_fn);
1002}
1003#endif /* CONFIG_XFS_LIVE_HOOKS */
1004
1005int
1006xfs_create(
1007	struct mnt_idmap	*idmap,
1008	xfs_inode_t		*dp,
1009	struct xfs_name		*name,
1010	umode_t			mode,
1011	dev_t			rdev,
1012	bool			init_xattrs,
1013	xfs_inode_t		**ipp)
1014{
1015	int			is_dir = S_ISDIR(mode);
1016	struct xfs_mount	*mp = dp->i_mount;
1017	struct xfs_inode	*ip = NULL;
1018	struct xfs_trans	*tp = NULL;
1019	int			error;
1020	bool                    unlock_dp_on_error = false;
1021	prid_t			prid;
1022	struct xfs_dquot	*udqp = NULL;
1023	struct xfs_dquot	*gdqp = NULL;
1024	struct xfs_dquot	*pdqp = NULL;
1025	struct xfs_trans_res	*tres;
1026	uint			resblks;
1027	xfs_ino_t		ino;
1028
1029	trace_xfs_create(dp, name);
1030
1031	if (xfs_is_shutdown(mp))
1032		return -EIO;
1033	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1034		return -EIO;
1035
1036	prid = xfs_get_initial_prid(dp);
1037
1038	/*
1039	 * Make sure that we have allocated dquot(s) on disk.
1040	 */
1041	error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
1042			mapped_fsgid(idmap, &init_user_ns), prid,
1043			XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1044			&udqp, &gdqp, &pdqp);
1045	if (error)
1046		return error;
1047
1048	if (is_dir) {
1049		resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1050		tres = &M_RES(mp)->tr_mkdir;
1051	} else {
1052		resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1053		tres = &M_RES(mp)->tr_create;
1054	}
1055
1056	/*
1057	 * Initially assume that the file does not exist and
1058	 * reserve the resources for that case.  If that is not
1059	 * the case we'll drop the one we have and get a more
1060	 * appropriate transaction later.
1061	 */
1062	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1063			&tp);
1064	if (error == -ENOSPC) {
1065		/* flush outstanding delalloc blocks and retry */
1066		xfs_flush_inodes(mp);
1067		error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1068				resblks, &tp);
1069	}
1070	if (error)
1071		goto out_release_dquots;
1072
1073	xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1074	unlock_dp_on_error = true;
1075
1076	/*
 
 
 
 
 
 
 
 
1077	 * A newly created regular or special file just has one directory
1078	 * entry pointing to them, but a directory also the "." entry
1079	 * pointing to itself.
1080	 */
1081	error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1082	if (!error)
1083		error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1084				is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1085	if (error)
1086		goto out_trans_cancel;
1087
1088	/*
1089	 * Now we join the directory inode to the transaction.  We do not do it
1090	 * earlier because xfs_dialloc might commit the previous transaction
1091	 * (and release all the locks).  An error from here on will result in
1092	 * the transaction cancel unlocking dp so don't do it explicitly in the
1093	 * error path.
1094	 */
1095	xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1096	unlock_dp_on_error = false;
1097
1098	error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1099					resblks - XFS_IALLOC_SPACE_RES(mp));
 
1100	if (error) {
1101		ASSERT(error != -ENOSPC);
1102		goto out_trans_cancel;
1103	}
1104	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1105	xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1106
1107	if (is_dir) {
1108		error = xfs_dir_init(tp, ip, dp);
1109		if (error)
1110			goto out_trans_cancel;
1111
1112		xfs_bumplink(tp, dp);
1113	}
1114
1115	/*
1116	 * Create ip with a reference from dp, and add '.' and '..' references
1117	 * if it's a directory.
1118	 */
1119	xfs_dir_update_hook(dp, ip, 1, name);
1120
1121	/*
1122	 * If this is a synchronous mount, make sure that the
1123	 * create transaction goes to disk before returning to
1124	 * the user.
1125	 */
1126	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1127		xfs_trans_set_sync(tp);
1128
1129	/*
1130	 * Attach the dquot(s) to the inodes and modify them incore.
1131	 * These ids of the inode couldn't have changed since the new
1132	 * inode has been locked ever since it was created.
1133	 */
1134	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1135
1136	error = xfs_trans_commit(tp);
1137	if (error)
1138		goto out_release_inode;
1139
1140	xfs_qm_dqrele(udqp);
1141	xfs_qm_dqrele(gdqp);
1142	xfs_qm_dqrele(pdqp);
1143
1144	*ipp = ip;
1145	return 0;
1146
1147 out_trans_cancel:
1148	xfs_trans_cancel(tp);
1149 out_release_inode:
1150	/*
1151	 * Wait until after the current transaction is aborted to finish the
1152	 * setup of the inode and release the inode.  This prevents recursive
1153	 * transactions and deadlocks from xfs_inactive.
1154	 */
1155	if (ip) {
1156		xfs_finish_inode_setup(ip);
1157		xfs_irele(ip);
1158	}
1159 out_release_dquots:
1160	xfs_qm_dqrele(udqp);
1161	xfs_qm_dqrele(gdqp);
1162	xfs_qm_dqrele(pdqp);
1163
1164	if (unlock_dp_on_error)
1165		xfs_iunlock(dp, XFS_ILOCK_EXCL);
1166	return error;
1167}
1168
1169int
1170xfs_create_tmpfile(
1171	struct mnt_idmap	*idmap,
1172	struct xfs_inode	*dp,
1173	umode_t			mode,
1174	struct xfs_inode	**ipp)
1175{
1176	struct xfs_mount	*mp = dp->i_mount;
1177	struct xfs_inode	*ip = NULL;
1178	struct xfs_trans	*tp = NULL;
1179	int			error;
1180	prid_t                  prid;
1181	struct xfs_dquot	*udqp = NULL;
1182	struct xfs_dquot	*gdqp = NULL;
1183	struct xfs_dquot	*pdqp = NULL;
1184	struct xfs_trans_res	*tres;
1185	uint			resblks;
1186	xfs_ino_t		ino;
1187
1188	if (xfs_is_shutdown(mp))
1189		return -EIO;
1190
1191	prid = xfs_get_initial_prid(dp);
1192
1193	/*
1194	 * Make sure that we have allocated dquot(s) on disk.
1195	 */
1196	error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
1197			mapped_fsgid(idmap, &init_user_ns), prid,
1198			XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1199			&udqp, &gdqp, &pdqp);
1200	if (error)
1201		return error;
1202
1203	resblks = XFS_IALLOC_SPACE_RES(mp);
1204	tres = &M_RES(mp)->tr_create_tmpfile;
1205
1206	error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1207			&tp);
 
 
 
 
1208	if (error)
1209		goto out_release_dquots;
1210
1211	error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1212	if (!error)
1213		error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1214				0, 0, prid, false, &ip);
1215	if (error)
1216		goto out_trans_cancel;
1217
1218	if (xfs_has_wsync(mp))
1219		xfs_trans_set_sync(tp);
1220
1221	/*
1222	 * Attach the dquot(s) to the inodes and modify them incore.
1223	 * These ids of the inode couldn't have changed since the new
1224	 * inode has been locked ever since it was created.
1225	 */
1226	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1227
1228	error = xfs_iunlink(tp, ip);
1229	if (error)
1230		goto out_trans_cancel;
1231
1232	error = xfs_trans_commit(tp);
1233	if (error)
1234		goto out_release_inode;
1235
1236	xfs_qm_dqrele(udqp);
1237	xfs_qm_dqrele(gdqp);
1238	xfs_qm_dqrele(pdqp);
1239
1240	*ipp = ip;
1241	return 0;
1242
1243 out_trans_cancel:
1244	xfs_trans_cancel(tp);
1245 out_release_inode:
1246	/*
1247	 * Wait until after the current transaction is aborted to finish the
1248	 * setup of the inode and release the inode.  This prevents recursive
1249	 * transactions and deadlocks from xfs_inactive.
1250	 */
1251	if (ip) {
1252		xfs_finish_inode_setup(ip);
1253		xfs_irele(ip);
1254	}
1255 out_release_dquots:
1256	xfs_qm_dqrele(udqp);
1257	xfs_qm_dqrele(gdqp);
1258	xfs_qm_dqrele(pdqp);
1259
1260	return error;
1261}
1262
1263int
1264xfs_link(
1265	xfs_inode_t		*tdp,
1266	xfs_inode_t		*sip,
1267	struct xfs_name		*target_name)
1268{
1269	xfs_mount_t		*mp = tdp->i_mount;
1270	xfs_trans_t		*tp;
1271	int			error, nospace_error = 0;
1272	int			resblks;
1273
1274	trace_xfs_link(tdp, target_name);
1275
1276	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1277
1278	if (xfs_is_shutdown(mp))
1279		return -EIO;
1280	if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
1281		return -EIO;
1282
1283	error = xfs_qm_dqattach(sip);
1284	if (error)
1285		goto std_return;
1286
1287	error = xfs_qm_dqattach(tdp);
1288	if (error)
1289		goto std_return;
1290
1291	resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1292	error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
1293			&tp, &nospace_error);
 
 
 
1294	if (error)
1295		goto std_return;
1296
 
 
 
 
 
1297	/*
1298	 * If we are using project inheritance, we only allow hard link
1299	 * creation in our tree when the project IDs are the same; else
1300	 * the tree quota mechanism could be circumvented.
1301	 */
1302	if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1303		     tdp->i_projid != sip->i_projid)) {
1304		/*
1305		 * Project quota setup skips special files which can
1306		 * leave inodes in a PROJINHERIT directory without a
1307		 * project ID set. We need to allow links to be made
1308		 * to these "project-less" inodes because userspace
1309		 * expects them to succeed after project ID setup,
1310		 * but everything else should be rejected.
1311		 */
1312		if (!special_file(VFS_I(sip)->i_mode) ||
1313		    sip->i_projid != 0) {
1314			error = -EXDEV;
1315			goto error_return;
1316		}
1317	}
1318
1319	if (!resblks) {
1320		error = xfs_dir_canenter(tp, tdp, target_name);
1321		if (error)
1322			goto error_return;
1323	}
1324
1325	/*
1326	 * Handle initial link state of O_TMPFILE inode
1327	 */
1328	if (VFS_I(sip)->i_nlink == 0) {
1329		struct xfs_perag	*pag;
1330
1331		pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1332		error = xfs_iunlink_remove(tp, pag, sip);
1333		xfs_perag_put(pag);
1334		if (error)
1335			goto error_return;
1336	}
1337
1338	error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1339				   resblks);
1340	if (error)
1341		goto error_return;
1342	xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1343	xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1344
1345	xfs_bumplink(tp, sip);
1346	xfs_dir_update_hook(tdp, sip, 1, target_name);
1347
1348	/*
1349	 * If this is a synchronous mount, make sure that the
1350	 * link transaction goes to disk before returning to
1351	 * the user.
1352	 */
1353	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1354		xfs_trans_set_sync(tp);
1355
1356	return xfs_trans_commit(tp);
1357
1358 error_return:
1359	xfs_trans_cancel(tp);
1360 std_return:
1361	if (error == -ENOSPC && nospace_error)
1362		error = nospace_error;
1363	return error;
1364}
1365
1366/* Clear the reflink flag and the cowblocks tag if possible. */
1367static void
1368xfs_itruncate_clear_reflink_flags(
1369	struct xfs_inode	*ip)
1370{
1371	struct xfs_ifork	*dfork;
1372	struct xfs_ifork	*cfork;
1373
1374	if (!xfs_is_reflink_inode(ip))
1375		return;
1376	dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1377	cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
1378	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1379		ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1380	if (cfork->if_bytes == 0)
1381		xfs_inode_clear_cowblocks_tag(ip);
1382}
1383
1384/*
1385 * Free up the underlying blocks past new_size.  The new size must be smaller
1386 * than the current size.  This routine can be used both for the attribute and
1387 * data fork, and does not modify the inode size, which is left to the caller.
1388 *
1389 * The transaction passed to this routine must have made a permanent log
1390 * reservation of at least XFS_ITRUNCATE_LOG_RES.  This routine may commit the
1391 * given transaction and start new ones, so make sure everything involved in
1392 * the transaction is tidy before calling here.  Some transaction will be
1393 * returned to the caller to be committed.  The incoming transaction must
1394 * already include the inode, and both inode locks must be held exclusively.
1395 * The inode must also be "held" within the transaction.  On return the inode
1396 * will be "held" within the returned transaction.  This routine does NOT
1397 * require any disk space to be reserved for it within the transaction.
1398 *
1399 * If we get an error, we must return with the inode locked and linked into the
1400 * current transaction. This keeps things simple for the higher level code,
1401 * because it always knows that the inode is locked and held in the transaction
1402 * that returns to it whether errors occur or not.  We don't mark the inode
1403 * dirty on error so that transactions can be easily aborted if possible.
1404 */
1405int
1406xfs_itruncate_extents_flags(
1407	struct xfs_trans	**tpp,
1408	struct xfs_inode	*ip,
1409	int			whichfork,
1410	xfs_fsize_t		new_size,
1411	int			flags)
1412{
1413	struct xfs_mount	*mp = ip->i_mount;
1414	struct xfs_trans	*tp = *tpp;
1415	xfs_fileoff_t		first_unmap_block;
 
 
1416	int			error = 0;
 
1417
1418	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1419	if (atomic_read(&VFS_I(ip)->i_count))
1420		xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1421	ASSERT(new_size <= XFS_ISIZE(ip));
1422	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1423	ASSERT(ip->i_itemp != NULL);
1424	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1425	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1426
1427	trace_xfs_itruncate_extents_start(ip, new_size);
1428
1429	flags |= xfs_bmapi_aflag(whichfork);
1430
1431	/*
1432	 * Since it is possible for space to become allocated beyond
1433	 * the end of the file (in a crash where the space is allocated
1434	 * but the inode size is not yet updated), simply remove any
1435	 * blocks which show up between the new EOF and the maximum
1436	 * possible file size.
1437	 *
1438	 * We have to free all the blocks to the bmbt maximum offset, even if
1439	 * the page cache can't scale that far.
1440	 */
1441	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1442	if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1443		WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1444		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1445	}
1446
1447	error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1448			XFS_MAX_FILEOFF);
1449	if (error)
1450		goto out;
1451
1452	if (whichfork == XFS_DATA_FORK) {
1453		/* Remove all pending CoW reservations. */
1454		error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1455				first_unmap_block, XFS_MAX_FILEOFF, true);
1456		if (error)
1457			goto out;
1458
1459		xfs_itruncate_clear_reflink_flags(ip);
1460	}
1461
1462	/*
1463	 * Always re-log the inode so that our permanent transaction can keep
1464	 * on rolling it forward in the log.
1465	 */
1466	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1467
1468	trace_xfs_itruncate_extents_end(ip, new_size);
1469
1470out:
1471	*tpp = tp;
1472	return error;
1473}
1474
1475int
1476xfs_release(
1477	xfs_inode_t	*ip)
1478{
1479	xfs_mount_t	*mp = ip->i_mount;
1480	int		error = 0;
1481
1482	if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1483		return 0;
1484
1485	/* If this is a read-only mount, don't do this (would generate I/O) */
1486	if (xfs_is_readonly(mp))
1487		return 0;
1488
1489	if (!xfs_is_shutdown(mp)) {
1490		int truncated;
1491
1492		/*
1493		 * If we previously truncated this file and removed old data
1494		 * in the process, we want to initiate "early" writeout on
1495		 * the last close.  This is an attempt to combat the notorious
1496		 * NULL files problem which is particularly noticeable from a
1497		 * truncate down, buffered (re-)write (delalloc), followed by
1498		 * a crash.  What we are effectively doing here is
1499		 * significantly reducing the time window where we'd otherwise
1500		 * be exposed to that problem.
1501		 */
1502		truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1503		if (truncated) {
1504			xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1505			if (ip->i_delayed_blks > 0) {
1506				error = filemap_flush(VFS_I(ip)->i_mapping);
1507				if (error)
1508					return error;
1509			}
1510		}
1511	}
1512
1513	if (VFS_I(ip)->i_nlink == 0)
1514		return 0;
1515
1516	/*
1517	 * If we can't get the iolock just skip truncating the blocks past EOF
1518	 * because we could deadlock with the mmap_lock otherwise. We'll get
1519	 * another chance to drop them once the last reference to the inode is
1520	 * dropped, so we'll never leak blocks permanently.
1521	 */
1522	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1523		return 0;
1524
1525	if (xfs_can_free_eofblocks(ip, false)) {
1526		/*
1527		 * Check if the inode is being opened, written and closed
1528		 * frequently and we have delayed allocation blocks outstanding
1529		 * (e.g. streaming writes from the NFS server), truncating the
1530		 * blocks past EOF will cause fragmentation to occur.
1531		 *
1532		 * In this case don't do the truncation, but we have to be
1533		 * careful how we detect this case. Blocks beyond EOF show up as
1534		 * i_delayed_blks even when the inode is clean, so we need to
1535		 * truncate them away first before checking for a dirty release.
1536		 * Hence on the first dirty close we will still remove the
1537		 * speculative allocation, but after that we will leave it in
1538		 * place.
1539		 */
1540		if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1541			goto out_unlock;
1542
1543		error = xfs_free_eofblocks(ip);
1544		if (error)
1545			goto out_unlock;
 
 
 
 
 
 
 
 
 
1546
1547		/* delalloc blocks after truncation means it really is dirty */
1548		if (ip->i_delayed_blks)
1549			xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1550	}
1551
1552out_unlock:
1553	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1554	return error;
1555}
1556
1557/*
1558 * xfs_inactive_truncate
1559 *
1560 * Called to perform a truncate when an inode becomes unlinked.
1561 */
1562STATIC int
1563xfs_inactive_truncate(
1564	struct xfs_inode *ip)
1565{
1566	struct xfs_mount	*mp = ip->i_mount;
1567	struct xfs_trans	*tp;
1568	int			error;
1569
1570	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1571	if (error) {
1572		ASSERT(xfs_is_shutdown(mp));
1573		return error;
1574	}
1575	xfs_ilock(ip, XFS_ILOCK_EXCL);
1576	xfs_trans_ijoin(tp, ip, 0);
1577
1578	/*
1579	 * Log the inode size first to prevent stale data exposure in the event
1580	 * of a system crash before the truncate completes. See the related
1581	 * comment in xfs_vn_setattr_size() for details.
1582	 */
1583	ip->i_disk_size = 0;
1584	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1585
1586	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1587	if (error)
1588		goto error_trans_cancel;
1589
1590	ASSERT(ip->i_df.if_nextents == 0);
1591
1592	error = xfs_trans_commit(tp);
1593	if (error)
1594		goto error_unlock;
1595
1596	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1597	return 0;
1598
1599error_trans_cancel:
1600	xfs_trans_cancel(tp);
1601error_unlock:
1602	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1603	return error;
1604}
1605
1606/*
1607 * xfs_inactive_ifree()
1608 *
1609 * Perform the inode free when an inode is unlinked.
1610 */
1611STATIC int
1612xfs_inactive_ifree(
1613	struct xfs_inode *ip)
1614{
1615	struct xfs_mount	*mp = ip->i_mount;
1616	struct xfs_trans	*tp;
1617	int			error;
1618
1619	/*
1620	 * We try to use a per-AG reservation for any block needed by the finobt
1621	 * tree, but as the finobt feature predates the per-AG reservation
1622	 * support a degraded file system might not have enough space for the
1623	 * reservation at mount time.  In that case try to dip into the reserved
1624	 * pool and pray.
1625	 *
1626	 * Send a warning if the reservation does happen to fail, as the inode
1627	 * now remains allocated and sits on the unlinked list until the fs is
1628	 * repaired.
1629	 */
1630	if (unlikely(mp->m_finobt_nores)) {
1631		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1632				XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1633				&tp);
1634	} else {
1635		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1636	}
1637	if (error) {
1638		if (error == -ENOSPC) {
1639			xfs_warn_ratelimited(mp,
1640			"Failed to remove inode(s) from unlinked list. "
1641			"Please free space, unmount and run xfs_repair.");
1642		} else {
1643			ASSERT(xfs_is_shutdown(mp));
1644		}
1645		return error;
1646	}
1647
1648	/*
1649	 * We do not hold the inode locked across the entire rolling transaction
1650	 * here. We only need to hold it for the first transaction that
1651	 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1652	 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1653	 * here breaks the relationship between cluster buffer invalidation and
1654	 * stale inode invalidation on cluster buffer item journal commit
1655	 * completion, and can result in leaving dirty stale inodes hanging
1656	 * around in memory.
1657	 *
1658	 * We have no need for serialising this inode operation against other
1659	 * operations - we freed the inode and hence reallocation is required
1660	 * and that will serialise on reallocating the space the deferops need
1661	 * to free. Hence we can unlock the inode on the first commit of
1662	 * the transaction rather than roll it right through the deferops. This
1663	 * avoids relogging the XFS_ISTALE inode.
1664	 *
1665	 * We check that xfs_ifree() hasn't grown an internal transaction roll
1666	 * by asserting that the inode is still locked when it returns.
1667	 */
1668	xfs_ilock(ip, XFS_ILOCK_EXCL);
1669	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1670
1671	error = xfs_ifree(tp, ip);
1672	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1673	if (error) {
1674		/*
1675		 * If we fail to free the inode, shut down.  The cancel
1676		 * might do that, we need to make sure.  Otherwise the
1677		 * inode might be lost for a long time or forever.
1678		 */
1679		if (!xfs_is_shutdown(mp)) {
1680			xfs_notice(mp, "%s: xfs_ifree returned error %d",
1681				__func__, error);
1682			xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1683		}
1684		xfs_trans_cancel(tp);
 
1685		return error;
1686	}
1687
1688	/*
1689	 * Credit the quota account(s). The inode is gone.
1690	 */
1691	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1692
1693	return xfs_trans_commit(tp);
1694}
1695
1696/*
1697 * Returns true if we need to update the on-disk metadata before we can free
1698 * the memory used by this inode.  Updates include freeing post-eof
1699 * preallocations; freeing COW staging extents; and marking the inode free in
1700 * the inobt if it is on the unlinked list.
1701 */
1702bool
1703xfs_inode_needs_inactive(
1704	struct xfs_inode	*ip)
1705{
1706	struct xfs_mount	*mp = ip->i_mount;
1707	struct xfs_ifork	*cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1708
1709	/*
1710	 * If the inode is already free, then there can be nothing
1711	 * to clean up here.
1712	 */
1713	if (VFS_I(ip)->i_mode == 0)
1714		return false;
 
 
1715
1716	/*
1717	 * If this is a read-only mount, don't do this (would generate I/O)
1718	 * unless we're in log recovery and cleaning the iunlinked list.
1719	 */
1720	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1721		return false;
1722
1723	/* If the log isn't running, push inodes straight to reclaim. */
1724	if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1725		return false;
1726
1727	/* Metadata inodes require explicit resource cleanup. */
1728	if (xfs_is_metadata_inode(ip))
1729		return false;
1730
1731	/* Want to clean out the cow blocks if there are any. */
1732	if (cow_ifp && cow_ifp->if_bytes > 0)
1733		return true;
1734
1735	/* Unlinked files must be freed. */
1736	if (VFS_I(ip)->i_nlink == 0)
1737		return true;
1738
1739	/*
1740	 * This file isn't being freed, so check if there are post-eof blocks
1741	 * to free.  @force is true because we are evicting an inode from the
1742	 * cache.  Post-eof blocks must be freed, lest we end up with broken
1743	 * free space accounting.
1744	 *
1745	 * Note: don't bother with iolock here since lockdep complains about
1746	 * acquiring it in reclaim context. We have the only reference to the
1747	 * inode at this point anyways.
1748	 */
1749	return xfs_can_free_eofblocks(ip, true);
1750}
1751
1752/*
1753 * Save health status somewhere, if we're dumping an inode with uncorrected
1754 * errors and online repair isn't running.
1755 */
1756static inline void
1757xfs_inactive_health(
1758	struct xfs_inode	*ip)
1759{
1760	struct xfs_mount	*mp = ip->i_mount;
1761	struct xfs_perag	*pag;
1762	unsigned int		sick;
1763	unsigned int		checked;
1764
1765	xfs_inode_measure_sickness(ip, &sick, &checked);
1766	if (!sick)
1767		return;
1768
1769	trace_xfs_inode_unfixed_corruption(ip, sick);
1770
1771	if (sick & XFS_SICK_INO_FORGET)
1772		return;
1773
1774	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1775	if (!pag) {
1776		/* There had better still be a perag structure! */
1777		ASSERT(0);
1778		return;
1779	}
1780
1781	xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1782	xfs_perag_put(pag);
1783}
1784
1785/*
1786 * xfs_inactive
1787 *
1788 * This is called when the vnode reference count for the vnode
1789 * goes to zero.  If the file has been unlinked, then it must
1790 * now be truncated.  Also, we clear all of the read-ahead state
1791 * kept for the inode here since the file is now closed.
1792 */
1793int
1794xfs_inactive(
1795	xfs_inode_t	*ip)
1796{
1797	struct xfs_mount	*mp;
1798	int			error = 0;
1799	int			truncate = 0;
1800
1801	/*
1802	 * If the inode is already free, then there can be nothing
1803	 * to clean up here.
1804	 */
1805	if (VFS_I(ip)->i_mode == 0) {
1806		ASSERT(ip->i_df.if_broot_bytes == 0);
1807		goto out;
1808	}
1809
1810	mp = ip->i_mount;
1811	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1812
1813	xfs_inactive_health(ip);
1814
1815	/*
1816	 * If this is a read-only mount, don't do this (would generate I/O)
1817	 * unless we're in log recovery and cleaning the iunlinked list.
1818	 */
1819	if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1820		goto out;
1821
1822	/* Metadata inodes require explicit resource cleanup. */
1823	if (xfs_is_metadata_inode(ip))
1824		goto out;
1825
1826	/* Try to clean out the cow blocks if there are any. */
1827	if (xfs_inode_has_cow_data(ip))
1828		xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1829
1830	if (VFS_I(ip)->i_nlink != 0) {
1831		/*
1832		 * force is true because we are evicting an inode from the
1833		 * cache. Post-eof blocks must be freed, lest we end up with
1834		 * broken free space accounting.
1835		 *
1836		 * Note: don't bother with iolock here since lockdep complains
1837		 * about acquiring it in reclaim context. We have the only
1838		 * reference to the inode at this point anyways.
1839		 */
1840		if (xfs_can_free_eofblocks(ip, true))
1841			error = xfs_free_eofblocks(ip);
1842
1843		goto out;
1844	}
1845
1846	if (S_ISREG(VFS_I(ip)->i_mode) &&
1847	    (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1848	     ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1849		truncate = 1;
1850
1851	if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1852		/*
1853		 * If this inode is being inactivated during a quotacheck and
1854		 * has not yet been scanned by quotacheck, we /must/ remove
1855		 * the dquots from the inode before inactivation changes the
1856		 * block and inode counts.  Most probably this is a result of
1857		 * reloading the incore iunlinked list to purge unrecovered
1858		 * unlinked inodes.
1859		 */
1860		xfs_qm_dqdetach(ip);
1861	} else {
1862		error = xfs_qm_dqattach(ip);
1863		if (error)
1864			goto out;
1865	}
1866
1867	if (S_ISLNK(VFS_I(ip)->i_mode))
1868		error = xfs_inactive_symlink(ip);
1869	else if (truncate)
1870		error = xfs_inactive_truncate(ip);
1871	if (error)
1872		goto out;
1873
1874	/*
1875	 * If there are attributes associated with the file then blow them away
1876	 * now.  The code calls a routine that recursively deconstructs the
1877	 * attribute fork. If also blows away the in-core attribute fork.
1878	 */
1879	if (xfs_inode_has_attr_fork(ip)) {
1880		error = xfs_attr_inactive(ip);
1881		if (error)
1882			goto out;
1883	}
1884
1885	ASSERT(ip->i_forkoff == 0);
 
 
1886
1887	/*
1888	 * Free the inode.
1889	 */
1890	error = xfs_inactive_ifree(ip);
 
 
1891
1892out:
1893	/*
1894	 * We're done making metadata updates for this inode, so we can release
1895	 * the attached dquots.
1896	 */
1897	xfs_qm_dqdetach(ip);
1898	return error;
1899}
1900
1901/*
1902 * In-Core Unlinked List Lookups
1903 * =============================
1904 *
1905 * Every inode is supposed to be reachable from some other piece of metadata
1906 * with the exception of the root directory.  Inodes with a connection to a
1907 * file descriptor but not linked from anywhere in the on-disk directory tree
1908 * are collectively known as unlinked inodes, though the filesystem itself
1909 * maintains links to these inodes so that on-disk metadata are consistent.
1910 *
1911 * XFS implements a per-AG on-disk hash table of unlinked inodes.  The AGI
1912 * header contains a number of buckets that point to an inode, and each inode
1913 * record has a pointer to the next inode in the hash chain.  This
1914 * singly-linked list causes scaling problems in the iunlink remove function
1915 * because we must walk that list to find the inode that points to the inode
1916 * being removed from the unlinked hash bucket list.
1917 *
1918 * Hence we keep an in-memory double linked list to link each inode on an
1919 * unlinked list. Because there are 64 unlinked lists per AGI, keeping pointer
1920 * based lists would require having 64 list heads in the perag, one for each
1921 * list. This is expensive in terms of memory (think millions of AGs) and cache
1922 * misses on lookups. Instead, use the fact that inodes on the unlinked list
1923 * must be referenced at the VFS level to keep them on the list and hence we
1924 * have an existence guarantee for inodes on the unlinked list.
1925 *
1926 * Given we have an existence guarantee, we can use lockless inode cache lookups
1927 * to resolve aginos to xfs inodes. This means we only need 8 bytes per inode
1928 * for the double linked unlinked list, and we don't need any extra locking to
1929 * keep the list safe as all manipulations are done under the AGI buffer lock.
1930 * Keeping the list up to date does not require memory allocation, just finding
1931 * the XFS inode and updating the next/prev unlinked list aginos.
1932 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1933
1934/*
1935 * Find an inode on the unlinked list. This does not take references to the
1936 * inode as we have existence guarantees by holding the AGI buffer lock and that
1937 * only unlinked, referenced inodes can be on the unlinked inode list.  If we
1938 * don't find the inode in cache, then let the caller handle the situation.
1939 */
1940static struct xfs_inode *
1941xfs_iunlink_lookup(
1942	struct xfs_perag	*pag,
1943	xfs_agino_t		agino)
1944{
1945	struct xfs_inode	*ip;
 
 
 
 
 
1946
1947	rcu_read_lock();
1948	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1949	if (!ip) {
1950		/* Caller can handle inode not being in memory. */
1951		rcu_read_unlock();
1952		return NULL;
1953	}
 
 
 
 
1954
 
 
1955	/*
1956	 * Inode in RCU freeing limbo should not happen.  Warn about this and
1957	 * let the caller handle the failure.
 
 
1958	 */
1959	if (WARN_ON_ONCE(!ip->i_ino)) {
1960		rcu_read_unlock();
1961		return NULL;
1962	}
1963	ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1964	rcu_read_unlock();
1965	return ip;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1966}
1967
1968/*
1969 * Update the prev pointer of the next agino.  Returns -ENOLINK if the inode
1970 * is not in cache.
 
1971 */
1972static int
1973xfs_iunlink_update_backref(
1974	struct xfs_perag	*pag,
1975	xfs_agino_t		prev_agino,
1976	xfs_agino_t		next_agino)
1977{
1978	struct xfs_inode	*ip;
 
 
 
 
 
 
 
1979
1980	/* No update necessary if we are at the end of the list. */
1981	if (next_agino == NULLAGINO)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1982		return 0;
 
1983
1984	ip = xfs_iunlink_lookup(pag, next_agino);
1985	if (!ip)
1986		return -ENOLINK;
 
1987
1988	ip->i_prev_unlinked = prev_agino;
1989	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1990}
1991
1992/*
1993 * Point the AGI unlinked bucket at an inode and log the results.  The caller
1994 * is responsible for validating the old value.
1995 */
1996STATIC int
1997xfs_iunlink_update_bucket(
1998	struct xfs_trans	*tp,
1999	struct xfs_perag	*pag,
2000	struct xfs_buf		*agibp,
2001	unsigned int		bucket_index,
2002	xfs_agino_t		new_agino)
2003{
2004	struct xfs_agi		*agi = agibp->b_addr;
2005	xfs_agino_t		old_value;
2006	int			offset;
2007
2008	ASSERT(xfs_verify_agino_or_null(pag, new_agino));
2009
2010	old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2011	trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
2012			old_value, new_agino);
2013
2014	/*
2015	 * We should never find the head of the list already set to the value
2016	 * passed in because either we're adding or removing ourselves from the
2017	 * head of the list.
2018	 */
2019	if (old_value == new_agino) {
2020		xfs_buf_mark_corrupt(agibp);
2021		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2022		return -EFSCORRUPTED;
2023	}
2024
2025	agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2026	offset = offsetof(struct xfs_agi, agi_unlinked) +
2027			(sizeof(xfs_agino_t) * bucket_index);
2028	xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2029	return 0;
2030}
2031
2032/*
2033 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
2034 * to @prev_agino.  Caller must hold the AGI to synchronize with other changes
2035 * to the unlinked list.
2036 */
2037STATIC int
2038xfs_iunlink_reload_next(
2039	struct xfs_trans	*tp,
2040	struct xfs_buf		*agibp,
2041	xfs_agino_t		prev_agino,
 
 
 
2042	xfs_agino_t		next_agino)
2043{
2044	struct xfs_perag	*pag = agibp->b_pag;
2045	struct xfs_mount	*mp = pag->pag_mount;
2046	struct xfs_inode	*next_ip = NULL;
2047	xfs_ino_t		ino;
2048	int			error;
 
 
 
 
 
 
2049
2050	ASSERT(next_agino != NULLAGINO);
 
 
 
 
 
2051
2052#ifdef DEBUG
2053	rcu_read_lock();
2054	next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
2055	ASSERT(next_ip == NULL);
2056	rcu_read_unlock();
2057#endif
 
 
 
 
 
 
 
 
2058
2059	xfs_info_ratelimited(mp,
2060 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating recovery.",
2061			next_agino, pag->pag_agno);
2062
2063	/*
2064	 * Use an untrusted lookup just to be cautious in case the AGI has been
2065	 * corrupted and now points at a free inode.  That shouldn't happen,
2066	 * but we'd rather shut down now since we're already running in a weird
2067	 * situation.
2068	 */
2069	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, next_agino);
2070	error = xfs_iget(mp, tp, ino, XFS_IGET_UNTRUSTED, 0, &next_ip);
2071	if (error) {
2072		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2073		return error;
 
 
 
 
 
 
2074	}
2075
2076	/* If this is not an unlinked inode, something is very wrong. */
2077	if (VFS_I(next_ip)->i_nlink != 0) {
2078		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2079		error = -EFSCORRUPTED;
2080		goto rele;
 
 
 
 
 
2081	}
2082
2083	next_ip->i_prev_unlinked = prev_agino;
2084	trace_xfs_iunlink_reload_next(next_ip);
2085rele:
2086	ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
2087	if (xfs_is_quotacheck_running(mp) && next_ip)
2088		xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
2089	xfs_irele(next_ip);
2090	return error;
2091}
2092
2093static int
2094xfs_iunlink_insert_inode(
 
 
 
 
 
 
 
2095	struct xfs_trans	*tp,
2096	struct xfs_perag	*pag,
2097	struct xfs_buf		*agibp,
2098	struct xfs_inode	*ip)
2099{
2100	struct xfs_mount	*mp = tp->t_mountp;
2101	struct xfs_agi		*agi = agibp->b_addr;
 
2102	xfs_agino_t		next_agino;
 
2103	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2104	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2105	int			error;
2106
 
 
 
 
 
 
 
 
 
 
2107	/*
2108	 * Get the index into the agi hash table for the list this inode will
2109	 * go on.  Make sure the pointer isn't garbage and that this inode
2110	 * isn't already on the list.
2111	 */
2112	next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2113	if (next_agino == agino ||
2114	    !xfs_verify_agino_or_null(pag, next_agino)) {
2115		xfs_buf_mark_corrupt(agibp);
2116		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2117		return -EFSCORRUPTED;
2118	}
2119
2120	/*
2121	 * Update the prev pointer in the next inode to point back to this
2122	 * inode.
2123	 */
2124	error = xfs_iunlink_update_backref(pag, agino, next_agino);
2125	if (error == -ENOLINK)
2126		error = xfs_iunlink_reload_next(tp, agibp, agino, next_agino);
2127	if (error)
2128		return error;
2129
2130	if (next_agino != NULLAGINO) {
2131		/*
2132		 * There is already another inode in the bucket, so point this
2133		 * inode to the current head of the list.
2134		 */
2135		error = xfs_iunlink_log_inode(tp, ip, pag, next_agino);
 
 
 
 
 
 
 
 
 
 
 
 
2136		if (error)
2137			return error;
2138		ip->i_next_unlinked = next_agino;
2139	}
2140
2141	/* Point the head of the list to point to this inode. */
2142	ip->i_prev_unlinked = NULLAGINO;
2143	return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2144}
2145
2146/*
2147 * This is called when the inode's link count has gone to 0 or we are creating
2148 * a tmpfile via O_TMPFILE.  The inode @ip must have nlink == 0.
2149 *
2150 * We place the on-disk inode on a list in the AGI.  It will be pulled from this
2151 * list when the inode is freed.
2152 */
2153STATIC int
2154xfs_iunlink(
2155	struct xfs_trans	*tp,
2156	struct xfs_inode	*ip)
 
 
 
 
2157{
2158	struct xfs_mount	*mp = tp->t_mountp;
2159	struct xfs_perag	*pag;
2160	struct xfs_buf		*agibp;
2161	int			error;
2162
2163	ASSERT(VFS_I(ip)->i_nlink == 0);
2164	ASSERT(VFS_I(ip)->i_mode != 0);
2165	trace_xfs_iunlink(ip);
 
 
 
 
2166
2167	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
 
 
 
 
 
2168
2169	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2170	error = xfs_read_agi(pag, tp, &agibp);
2171	if (error)
2172		goto out;
2173
2174	error = xfs_iunlink_insert_inode(tp, pag, agibp, ip);
2175out:
2176	xfs_perag_put(pag);
2177	return error;
2178}
2179
2180static int
2181xfs_iunlink_remove_inode(
 
 
 
 
 
 
 
 
 
 
2182	struct xfs_trans	*tp,
2183	struct xfs_perag	*pag,
2184	struct xfs_buf		*agibp,
2185	struct xfs_inode	*ip)
 
 
 
 
 
2186{
2187	struct xfs_mount	*mp = tp->t_mountp;
2188	struct xfs_agi		*agi = agibp->b_addr;
2189	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2190	xfs_agino_t		head_agino;
2191	short			bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2192	int			error;
2193
2194	trace_xfs_iunlink_remove(ip);
 
 
 
 
 
 
 
 
 
 
 
2195
2196	/*
2197	 * Get the index into the agi hash table for the list this inode will
2198	 * go on.  Make sure the head pointer isn't garbage.
2199	 */
2200	head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2201	if (!xfs_verify_agino(pag, head_agino)) {
2202		XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2203				agi, sizeof(*agi));
2204		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2205		return -EFSCORRUPTED;
2206	}
2207
2208	/*
2209	 * Set our inode's next_unlinked pointer to NULL and then return
2210	 * the old pointer value so that we can update whatever was previous
2211	 * to us in the list to point to whatever was next in the list.
2212	 */
2213	error = xfs_iunlink_log_inode(tp, ip, pag, NULLAGINO);
2214	if (error)
2215		return error;
2216
2217	/*
2218	 * Update the prev pointer in the next inode to point back to previous
2219	 * inode in the chain.
2220	 */
2221	error = xfs_iunlink_update_backref(pag, ip->i_prev_unlinked,
2222			ip->i_next_unlinked);
2223	if (error == -ENOLINK)
2224		error = xfs_iunlink_reload_next(tp, agibp, ip->i_prev_unlinked,
2225				ip->i_next_unlinked);
2226	if (error)
2227		return error;
2228
2229	if (head_agino != agino) {
2230		struct xfs_inode	*prev_ip;
 
 
 
2231
2232		prev_ip = xfs_iunlink_lookup(pag, ip->i_prev_unlinked);
2233		if (!prev_ip) {
2234			xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2235			return -EFSCORRUPTED;
 
 
 
 
 
 
 
 
2236		}
2237
2238		error = xfs_iunlink_log_inode(tp, prev_ip, pag,
2239				ip->i_next_unlinked);
2240		prev_ip->i_next_unlinked = ip->i_next_unlinked;
2241	} else {
2242		/* Point the head of the list to the next unlinked inode. */
2243		error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2244				ip->i_next_unlinked);
2245	}
2246
2247	ip->i_next_unlinked = NULLAGINO;
2248	ip->i_prev_unlinked = 0;
2249	return error;
2250}
2251
2252/*
2253 * Pull the on-disk inode from the AGI unlinked list.
2254 */
2255STATIC int
2256xfs_iunlink_remove(
2257	struct xfs_trans	*tp,
2258	struct xfs_perag	*pag,
2259	struct xfs_inode	*ip)
2260{
 
 
2261	struct xfs_buf		*agibp;
 
 
 
 
 
 
 
 
2262	int			error;
2263
2264	trace_xfs_iunlink_remove(ip);
2265
2266	/* Get the agi buffer first.  It ensures lock ordering on the list. */
2267	error = xfs_read_agi(pag, tp, &agibp);
2268	if (error)
2269		return error;
 
2270
2271	return xfs_iunlink_remove_inode(tp, pag, agibp, ip);
2272}
2273
2274/*
2275 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2276 * mark it stale. We should only find clean inodes in this lookup that aren't
2277 * already stale.
2278 */
2279static void
2280xfs_ifree_mark_inode_stale(
2281	struct xfs_perag	*pag,
2282	struct xfs_inode	*free_ip,
2283	xfs_ino_t		inum)
2284{
2285	struct xfs_mount	*mp = pag->pag_mount;
2286	struct xfs_inode_log_item *iip;
2287	struct xfs_inode	*ip;
2288
2289retry:
2290	rcu_read_lock();
2291	ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2292
2293	/* Inode not in memory, nothing to do */
2294	if (!ip) {
2295		rcu_read_unlock();
2296		return;
2297	}
2298
2299	/*
2300	 * because this is an RCU protected lookup, we could find a recently
2301	 * freed or even reallocated inode during the lookup. We need to check
2302	 * under the i_flags_lock for a valid inode here. Skip it if it is not
2303	 * valid, the wrong inode or stale.
2304	 */
2305	spin_lock(&ip->i_flags_lock);
2306	if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2307		goto out_iflags_unlock;
2308
2309	/*
2310	 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2311	 * other inodes that we did not find in the list attached to the buffer
2312	 * and are not already marked stale. If we can't lock it, back off and
2313	 * retry.
2314	 */
2315	if (ip != free_ip) {
2316		if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2317			spin_unlock(&ip->i_flags_lock);
2318			rcu_read_unlock();
2319			delay(1);
2320			goto retry;
2321		}
2322	}
2323	ip->i_flags |= XFS_ISTALE;
2324
2325	/*
2326	 * If the inode is flushing, it is already attached to the buffer.  All
2327	 * we needed to do here is mark the inode stale so buffer IO completion
2328	 * will remove it from the AIL.
 
 
2329	 */
2330	iip = ip->i_itemp;
2331	if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2332		ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2333		ASSERT(iip->ili_last_fields);
2334		goto out_iunlock;
 
2335	}
2336
2337	/*
2338	 * Inodes not attached to the buffer can be released immediately.
2339	 * Everything else has to go through xfs_iflush_abort() on journal
2340	 * commit as the flock synchronises removal of the inode from the
2341	 * cluster buffer against inode reclaim.
2342	 */
2343	if (!iip || list_empty(&iip->ili_item.li_bio_list))
2344		goto out_iunlock;
 
 
 
 
2345
2346	__xfs_iflags_set(ip, XFS_IFLUSHING);
2347	spin_unlock(&ip->i_flags_lock);
2348	rcu_read_unlock();
 
 
 
2349
2350	/* we have a dirty inode in memory that has not yet been flushed. */
2351	spin_lock(&iip->ili_lock);
2352	iip->ili_last_fields = iip->ili_fields;
2353	iip->ili_fields = 0;
2354	iip->ili_fsync_fields = 0;
2355	spin_unlock(&iip->ili_lock);
2356	ASSERT(iip->ili_last_fields);
2357
2358	if (ip != free_ip)
2359		xfs_iunlock(ip, XFS_ILOCK_EXCL);
2360	return;
 
 
 
 
 
 
 
 
 
2361
2362out_iunlock:
2363	if (ip != free_ip)
2364		xfs_iunlock(ip, XFS_ILOCK_EXCL);
2365out_iflags_unlock:
2366	spin_unlock(&ip->i_flags_lock);
2367	rcu_read_unlock();
2368}
2369
2370/*
2371 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2372 * inodes that are in memory - they all must be marked stale and attached to
2373 * the cluster buffer.
2374 */
2375static int
2376xfs_ifree_cluster(
2377	struct xfs_trans	*tp,
2378	struct xfs_perag	*pag,
2379	struct xfs_inode	*free_ip,
2380	struct xfs_icluster	*xic)
2381{
2382	struct xfs_mount	*mp = free_ip->i_mount;
2383	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2384	struct xfs_buf		*bp;
2385	xfs_daddr_t		blkno;
2386	xfs_ino_t		inum = xic->first_ino;
2387	int			nbufs;
2388	int			i, j;
2389	int			ioffset;
2390	int			error;
 
 
 
 
 
 
 
2391
 
 
2392	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2393
2394	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2395		/*
2396		 * The allocation bitmap tells us which inodes of the chunk were
2397		 * physically allocated. Skip the cluster if an inode falls into
2398		 * a sparse region.
2399		 */
2400		ioffset = inum - xic->first_ino;
2401		if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2402			ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2403			continue;
2404		}
2405
2406		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2407					 XFS_INO_TO_AGBNO(mp, inum));
2408
2409		/*
2410		 * We obtain and lock the backing buffer first in the process
2411		 * here to ensure dirty inodes attached to the buffer remain in
2412		 * the flushing state while we mark them stale.
2413		 *
2414		 * If we scan the in-memory inodes first, then buffer IO can
2415		 * complete before we get a lock on it, and hence we may fail
2416		 * to mark all the active inodes on the buffer stale.
2417		 */
2418		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2419				mp->m_bsize * igeo->blocks_per_cluster,
2420				XBF_UNMAPPED, &bp);
2421		if (error)
2422			return error;
 
2423
2424		/*
2425		 * This buffer may not have been correctly initialised as we
2426		 * didn't read it from disk. That's not important because we are
2427		 * only using to mark the buffer as stale in the log, and to
2428		 * attach stale cached inodes on it. That means it will never be
2429		 * dispatched for IO. If it is, we want to know about it, and we
2430		 * want it to fail. We can acheive this by adding a write
2431		 * verifier to the buffer.
2432		 */
2433		bp->b_ops = &xfs_inode_buf_ops;
2434
2435		/*
2436		 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2437		 * too. This requires lookups, and will skip inodes that we've
2438		 * already marked XFS_ISTALE.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2439		 */
2440		for (i = 0; i < igeo->inodes_per_cluster; i++)
2441			xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2442
2443		xfs_trans_stale_inode_buf(tp, bp);
2444		xfs_trans_binval(tp, bp);
2445	}
 
 
2446	return 0;
2447}
2448
2449/*
2450 * This is called to return an inode to the inode free list.  The inode should
2451 * already be truncated to 0 length and have no pages associated with it.  This
2452 * routine also assumes that the inode is already a part of the transaction.
2453 *
2454 * The on-disk copy of the inode will have been added to the list of unlinked
2455 * inodes in the AGI. We need to remove the inode from that list atomically with
2456 * respect to freeing it here.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2457 */
2458int
2459xfs_ifree(
2460	struct xfs_trans	*tp,
2461	struct xfs_inode	*ip)
2462{
2463	struct xfs_mount	*mp = ip->i_mount;
2464	struct xfs_perag	*pag;
2465	struct xfs_icluster	xic = { 0 };
2466	struct xfs_inode_log_item *iip = ip->i_itemp;
2467	int			error;
2468
2469	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
2470	ASSERT(VFS_I(ip)->i_nlink == 0);
2471	ASSERT(ip->i_df.if_nextents == 0);
2472	ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2473	ASSERT(ip->i_nblocks == 0);
2474
2475	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2476
2477	/*
2478	 * Free the inode first so that we guarantee that the AGI lock is going
2479	 * to be taken before we remove the inode from the unlinked list. This
2480	 * makes the AGI lock -> unlinked list modification order the same as
2481	 * used in O_TMPFILE creation.
2482	 */
2483	error = xfs_difree(tp, pag, ip->i_ino, &xic);
2484	if (error)
2485		goto out;
2486
2487	error = xfs_iunlink_remove(tp, pag, ip);
2488	if (error)
2489		goto out;
2490
2491	/*
2492	 * Free any local-format data sitting around before we reset the
2493	 * data fork to extents format.  Note that the attr fork data has
2494	 * already been freed by xfs_attr_inactive.
2495	 */
2496	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2497		kfree(ip->i_df.if_data);
2498		ip->i_df.if_data = NULL;
2499		ip->i_df.if_bytes = 0;
2500	}
2501
2502	VFS_I(ip)->i_mode = 0;		/* mark incore inode as free */
2503	ip->i_diflags = 0;
2504	ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2505	ip->i_forkoff = 0;		/* mark the attr fork not in use */
2506	ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2507	if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2508		xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2509
2510	/* Don't attempt to replay owner changes for a deleted inode */
2511	spin_lock(&iip->ili_lock);
2512	iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2513	spin_unlock(&iip->ili_lock);
2514
2515	/*
2516	 * Bump the generation count so no one will be confused
2517	 * by reincarnations of this inode.
2518	 */
2519	VFS_I(ip)->i_generation++;
2520	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2521
2522	if (xic.deleted)
2523		error = xfs_ifree_cluster(tp, pag, ip, &xic);
2524out:
2525	xfs_perag_put(pag);
2526	return error;
2527}
2528
2529/*
2530 * This is called to unpin an inode.  The caller must have the inode locked
2531 * in at least shared mode so that the buffer cannot be subsequently pinned
2532 * once someone is waiting for it to be unpinned.
2533 */
2534static void
2535xfs_iunpin(
2536	struct xfs_inode	*ip)
2537{
2538	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2539
2540	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2541
2542	/* Give the log a push to start the unpinning I/O */
2543	xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2544
2545}
2546
2547static void
2548__xfs_iunpin_wait(
2549	struct xfs_inode	*ip)
2550{
2551	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2552	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2553
2554	xfs_iunpin(ip);
2555
2556	do {
2557		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2558		if (xfs_ipincount(ip))
2559			io_schedule();
2560	} while (xfs_ipincount(ip));
2561	finish_wait(wq, &wait.wq_entry);
2562}
2563
2564void
2565xfs_iunpin_wait(
2566	struct xfs_inode	*ip)
2567{
2568	if (xfs_ipincount(ip))
2569		__xfs_iunpin_wait(ip);
2570}
2571
2572/*
2573 * Removing an inode from the namespace involves removing the directory entry
2574 * and dropping the link count on the inode. Removing the directory entry can
2575 * result in locking an AGF (directory blocks were freed) and removing a link
2576 * count can result in placing the inode on an unlinked list which results in
2577 * locking an AGI.
2578 *
2579 * The big problem here is that we have an ordering constraint on AGF and AGI
2580 * locking - inode allocation locks the AGI, then can allocate a new extent for
2581 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2582 * removes the inode from the unlinked list, requiring that we lock the AGI
2583 * first, and then freeing the inode can result in an inode chunk being freed
2584 * and hence freeing disk space requiring that we lock an AGF.
2585 *
2586 * Hence the ordering that is imposed by other parts of the code is AGI before
2587 * AGF. This means we cannot remove the directory entry before we drop the inode
2588 * reference count and put it on the unlinked list as this results in a lock
2589 * order of AGF then AGI, and this can deadlock against inode allocation and
2590 * freeing. Therefore we must drop the link counts before we remove the
2591 * directory entry.
2592 *
2593 * This is still safe from a transactional point of view - it is not until we
2594 * get to xfs_defer_finish() that we have the possibility of multiple
2595 * transactions in this operation. Hence as long as we remove the directory
2596 * entry and drop the link count in the first transaction of the remove
2597 * operation, there are no transactional constraints on the ordering here.
2598 */
2599int
2600xfs_remove(
2601	xfs_inode_t             *dp,
2602	struct xfs_name		*name,
2603	xfs_inode_t		*ip)
2604{
2605	xfs_mount_t		*mp = dp->i_mount;
2606	xfs_trans_t             *tp = NULL;
2607	int			is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2608	int			dontcare;
2609	int                     error = 0;
2610	uint			resblks;
2611
2612	trace_xfs_remove(dp, name);
2613
2614	if (xfs_is_shutdown(mp))
2615		return -EIO;
2616	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
2617		return -EIO;
2618
2619	error = xfs_qm_dqattach(dp);
2620	if (error)
2621		goto std_return;
2622
2623	error = xfs_qm_dqattach(ip);
2624	if (error)
2625		goto std_return;
2626
2627	/*
2628	 * We try to get the real space reservation first, allowing for
2629	 * directory btree deletion(s) implying possible bmap insert(s).  If we
2630	 * can't get the space reservation then we use 0 instead, and avoid the
2631	 * bmap btree insert(s) in the directory code by, if the bmap insert
2632	 * tries to happen, instead trimming the LAST block from the directory.
2633	 *
2634	 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
2635	 * the directory code can handle a reservationless update and we don't
2636	 * want to prevent a user from trying to free space by deleting things.
2637	 */
2638	resblks = XFS_REMOVE_SPACE_RES(mp);
2639	error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2640			&tp, &dontcare);
 
 
 
 
2641	if (error) {
2642		ASSERT(error != -ENOSPC);
2643		goto std_return;
2644	}
2645
 
 
 
 
 
2646	/*
2647	 * If we're removing a directory perform some additional validation.
2648	 */
2649	if (is_dir) {
2650		ASSERT(VFS_I(ip)->i_nlink >= 2);
2651		if (VFS_I(ip)->i_nlink != 2) {
2652			error = -ENOTEMPTY;
2653			goto out_trans_cancel;
2654		}
2655		if (!xfs_dir_isempty(ip)) {
2656			error = -ENOTEMPTY;
2657			goto out_trans_cancel;
2658		}
2659
2660		/* Drop the link from ip's "..".  */
2661		error = xfs_droplink(tp, dp);
2662		if (error)
2663			goto out_trans_cancel;
2664
2665		/* Drop the "." link from ip to self.  */
2666		error = xfs_droplink(tp, ip);
2667		if (error)
2668			goto out_trans_cancel;
2669
2670		/*
2671		 * Point the unlinked child directory's ".." entry to the root
2672		 * directory to eliminate back-references to inodes that may
2673		 * get freed before the child directory is closed.  If the fs
2674		 * gets shrunk, this can lead to dirent inode validation errors.
2675		 */
2676		if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2677			error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2678					tp->t_mountp->m_sb.sb_rootino, 0);
2679			if (error)
2680				goto out_trans_cancel;
2681		}
2682	} else {
2683		/*
2684		 * When removing a non-directory we need to log the parent
2685		 * inode here.  For a directory this is done implicitly
2686		 * by the xfs_droplink call for the ".." entry.
2687		 */
2688		xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2689	}
2690	xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2691
2692	/* Drop the link from dp to ip. */
2693	error = xfs_droplink(tp, ip);
2694	if (error)
2695		goto out_trans_cancel;
2696
2697	error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2698	if (error) {
2699		ASSERT(error != -ENOENT);
2700		goto out_trans_cancel;
2701	}
2702
2703	/*
2704	 * Drop the link from dp to ip, and if ip was a directory, remove the
2705	 * '.' and '..' references since we freed the directory.
2706	 */
2707	xfs_dir_update_hook(dp, ip, -1, name);
2708
2709	/*
2710	 * If this is a synchronous mount, make sure that the
2711	 * remove transaction goes to disk before returning to
2712	 * the user.
2713	 */
2714	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2715		xfs_trans_set_sync(tp);
2716
2717	error = xfs_trans_commit(tp);
2718	if (error)
2719		goto std_return;
2720
2721	if (is_dir && xfs_inode_is_filestream(ip))
2722		xfs_filestream_deassociate(ip);
2723
2724	return 0;
2725
2726 out_trans_cancel:
2727	xfs_trans_cancel(tp);
2728 std_return:
2729	return error;
2730}
2731
2732/*
2733 * Enter all inodes for a rename transaction into a sorted array.
2734 */
2735#define __XFS_SORT_INODES	5
2736STATIC void
2737xfs_sort_for_rename(
2738	struct xfs_inode	*dp1,	/* in: old (source) directory inode */
2739	struct xfs_inode	*dp2,	/* in: new (target) directory inode */
2740	struct xfs_inode	*ip1,	/* in: inode of old entry */
2741	struct xfs_inode	*ip2,	/* in: inode of new entry */
2742	struct xfs_inode	*wip,	/* in: whiteout inode */
2743	struct xfs_inode	**i_tab,/* out: sorted array of inodes */
2744	int			*num_inodes)  /* in/out: inodes in array */
2745{
2746	int			i, j;
2747
2748	ASSERT(*num_inodes == __XFS_SORT_INODES);
2749	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2750
2751	/*
2752	 * i_tab contains a list of pointers to inodes.  We initialize
2753	 * the table here & we'll sort it.  We will then use it to
2754	 * order the acquisition of the inode locks.
2755	 *
2756	 * Note that the table may contain duplicates.  e.g., dp1 == dp2.
2757	 */
2758	i = 0;
2759	i_tab[i++] = dp1;
2760	i_tab[i++] = dp2;
2761	i_tab[i++] = ip1;
2762	if (ip2)
2763		i_tab[i++] = ip2;
2764	if (wip)
2765		i_tab[i++] = wip;
2766	*num_inodes = i;
2767
2768	/*
2769	 * Sort the elements via bubble sort.  (Remember, there are at
2770	 * most 5 elements to sort, so this is adequate.)
2771	 */
2772	for (i = 0; i < *num_inodes; i++) {
2773		for (j = 1; j < *num_inodes; j++) {
2774			if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2775				struct xfs_inode *temp = i_tab[j];
2776				i_tab[j] = i_tab[j-1];
2777				i_tab[j-1] = temp;
2778			}
2779		}
2780	}
2781}
2782
2783static int
2784xfs_finish_rename(
2785	struct xfs_trans	*tp)
2786{
2787	/*
2788	 * If this is a synchronous mount, make sure that the rename transaction
2789	 * goes to disk before returning to the user.
2790	 */
2791	if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2792		xfs_trans_set_sync(tp);
2793
2794	return xfs_trans_commit(tp);
2795}
2796
2797/*
2798 * xfs_cross_rename()
2799 *
2800 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2801 */
2802STATIC int
2803xfs_cross_rename(
2804	struct xfs_trans	*tp,
2805	struct xfs_inode	*dp1,
2806	struct xfs_name		*name1,
2807	struct xfs_inode	*ip1,
2808	struct xfs_inode	*dp2,
2809	struct xfs_name		*name2,
2810	struct xfs_inode	*ip2,
2811	int			spaceres)
2812{
2813	int		error = 0;
2814	int		ip1_flags = 0;
2815	int		ip2_flags = 0;
2816	int		dp2_flags = 0;
2817
2818	/* Swap inode number for dirent in first parent */
2819	error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2820	if (error)
2821		goto out_trans_abort;
2822
2823	/* Swap inode number for dirent in second parent */
2824	error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2825	if (error)
2826		goto out_trans_abort;
2827
2828	/*
2829	 * If we're renaming one or more directories across different parents,
2830	 * update the respective ".." entries (and link counts) to match the new
2831	 * parents.
2832	 */
2833	if (dp1 != dp2) {
2834		dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2835
2836		if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2837			error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2838						dp1->i_ino, spaceres);
2839			if (error)
2840				goto out_trans_abort;
2841
2842			/* transfer ip2 ".." reference to dp1 */
2843			if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2844				error = xfs_droplink(tp, dp2);
2845				if (error)
2846					goto out_trans_abort;
2847				xfs_bumplink(tp, dp1);
2848			}
2849
2850			/*
2851			 * Although ip1 isn't changed here, userspace needs
2852			 * to be warned about the change, so that applications
2853			 * relying on it (like backup ones), will properly
2854			 * notify the change
2855			 */
2856			ip1_flags |= XFS_ICHGTIME_CHG;
2857			ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2858		}
2859
2860		if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2861			error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2862						dp2->i_ino, spaceres);
2863			if (error)
2864				goto out_trans_abort;
2865
2866			/* transfer ip1 ".." reference to dp2 */
2867			if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2868				error = xfs_droplink(tp, dp1);
2869				if (error)
2870					goto out_trans_abort;
2871				xfs_bumplink(tp, dp2);
2872			}
2873
2874			/*
2875			 * Although ip2 isn't changed here, userspace needs
2876			 * to be warned about the change, so that applications
2877			 * relying on it (like backup ones), will properly
2878			 * notify the change
2879			 */
2880			ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2881			ip2_flags |= XFS_ICHGTIME_CHG;
2882		}
2883	}
2884
2885	if (ip1_flags) {
2886		xfs_trans_ichgtime(tp, ip1, ip1_flags);
2887		xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2888	}
2889	if (ip2_flags) {
2890		xfs_trans_ichgtime(tp, ip2, ip2_flags);
2891		xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2892	}
2893	if (dp2_flags) {
2894		xfs_trans_ichgtime(tp, dp2, dp2_flags);
2895		xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2896	}
2897	xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2898	xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2899
2900	/*
2901	 * Inform our hook clients that we've finished an exchange operation as
2902	 * follows: removed the source and target files from their directories;
2903	 * added the target to the source directory; and added the source to
2904	 * the target directory.  All inodes are locked, so it's ok to model a
2905	 * rename this way so long as we say we deleted entries before we add
2906	 * new ones.
2907	 */
2908	xfs_dir_update_hook(dp1, ip1, -1, name1);
2909	xfs_dir_update_hook(dp2, ip2, -1, name2);
2910	xfs_dir_update_hook(dp1, ip2, 1, name1);
2911	xfs_dir_update_hook(dp2, ip1, 1, name2);
2912
2913	return xfs_finish_rename(tp);
2914
2915out_trans_abort:
2916	xfs_trans_cancel(tp);
2917	return error;
2918}
2919
2920/*
2921 * xfs_rename_alloc_whiteout()
2922 *
2923 * Return a referenced, unlinked, unlocked inode that can be used as a
2924 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2925 * crash between allocating the inode and linking it into the rename transaction
2926 * recovery will free the inode and we won't leak it.
2927 */
2928static int
2929xfs_rename_alloc_whiteout(
2930	struct mnt_idmap	*idmap,
2931	struct xfs_name		*src_name,
2932	struct xfs_inode	*dp,
2933	struct xfs_inode	**wip)
2934{
2935	struct xfs_inode	*tmpfile;
2936	struct qstr		name;
2937	int			error;
2938
2939	error = xfs_create_tmpfile(idmap, dp, S_IFCHR | WHITEOUT_MODE,
2940				   &tmpfile);
2941	if (error)
2942		return error;
2943
2944	name.name = src_name->name;
2945	name.len = src_name->len;
2946	error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2947	if (error) {
2948		xfs_finish_inode_setup(tmpfile);
2949		xfs_irele(tmpfile);
2950		return error;
2951	}
2952
2953	/*
2954	 * Prepare the tmpfile inode as if it were created through the VFS.
2955	 * Complete the inode setup and flag it as linkable.  nlink is already
2956	 * zero, so we can skip the drop_nlink.
2957	 */
2958	xfs_setup_iops(tmpfile);
2959	xfs_finish_inode_setup(tmpfile);
2960	VFS_I(tmpfile)->i_state |= I_LINKABLE;
2961
2962	*wip = tmpfile;
2963	return 0;
2964}
2965
2966/*
2967 * xfs_rename
2968 */
2969int
2970xfs_rename(
2971	struct mnt_idmap	*idmap,
2972	struct xfs_inode	*src_dp,
2973	struct xfs_name		*src_name,
2974	struct xfs_inode	*src_ip,
2975	struct xfs_inode	*target_dp,
2976	struct xfs_name		*target_name,
2977	struct xfs_inode	*target_ip,
2978	unsigned int		flags)
2979{
2980	struct xfs_mount	*mp = src_dp->i_mount;
2981	struct xfs_trans	*tp;
2982	struct xfs_inode	*wip = NULL;		/* whiteout inode */
2983	struct xfs_inode	*inodes[__XFS_SORT_INODES];
2984	int			i;
2985	int			num_inodes = __XFS_SORT_INODES;
2986	bool			new_parent = (src_dp != target_dp);
2987	bool			src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2988	int			spaceres;
2989	bool			retried = false;
2990	int			error, nospace_error = 0;
2991
2992	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2993
2994	if ((flags & RENAME_EXCHANGE) && !target_ip)
2995		return -EINVAL;
2996
2997	/*
2998	 * If we are doing a whiteout operation, allocate the whiteout inode
2999	 * we will be placing at the target and ensure the type is set
3000	 * appropriately.
3001	 */
3002	if (flags & RENAME_WHITEOUT) {
3003		error = xfs_rename_alloc_whiteout(idmap, src_name,
3004						  target_dp, &wip);
3005		if (error)
3006			return error;
3007
3008		/* setup target dirent info as whiteout */
3009		src_name->type = XFS_DIR3_FT_CHRDEV;
3010	}
3011
3012	xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3013				inodes, &num_inodes);
3014
3015retry:
3016	nospace_error = 0;
3017	spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3018	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3019	if (error == -ENOSPC) {
3020		nospace_error = error;
3021		spaceres = 0;
3022		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3023				&tp);
3024	}
3025	if (error)
3026		goto out_release_wip;
3027
3028	/*
3029	 * Attach the dquots to the inodes
3030	 */
3031	error = xfs_qm_vop_rename_dqattach(inodes);
3032	if (error)
3033		goto out_trans_cancel;
3034
3035	/*
3036	 * Lock all the participating inodes. Depending upon whether
3037	 * the target_name exists in the target directory, and
3038	 * whether the target directory is the same as the source
3039	 * directory, we can lock from 2 to 5 inodes.
3040	 */
3041	xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3042
3043	/*
3044	 * Join all the inodes to the transaction. From this point on,
3045	 * we can rely on either trans_commit or trans_cancel to unlock
3046	 * them.
3047	 */
3048	xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3049	if (new_parent)
3050		xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3051	xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3052	if (target_ip)
3053		xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3054	if (wip)
3055		xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3056
3057	/*
3058	 * If we are using project inheritance, we only allow renames
3059	 * into our tree when the project IDs are the same; else the
3060	 * tree quota mechanism would be circumvented.
3061	 */
3062	if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
3063		     target_dp->i_projid != src_ip->i_projid)) {
3064		error = -EXDEV;
3065		goto out_trans_cancel;
3066	}
3067
3068	/* RENAME_EXCHANGE is unique from here on. */
3069	if (flags & RENAME_EXCHANGE)
3070		return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3071					target_dp, target_name, target_ip,
3072					spaceres);
3073
3074	/*
3075	 * Try to reserve quota to handle an expansion of the target directory.
3076	 * We'll allow the rename to continue in reservationless mode if we hit
3077	 * a space usage constraint.  If we trigger reservationless mode, save
3078	 * the errno if there isn't any free space in the target directory.
3079	 */
3080	if (spaceres != 0) {
3081		error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
3082				0, false);
3083		if (error == -EDQUOT || error == -ENOSPC) {
3084			if (!retried) {
3085				xfs_trans_cancel(tp);
3086				xfs_blockgc_free_quota(target_dp, 0);
3087				retried = true;
3088				goto retry;
3089			}
3090
3091			nospace_error = error;
3092			spaceres = 0;
3093			error = 0;
3094		}
3095		if (error)
3096			goto out_trans_cancel;
3097	}
3098
3099	/*
3100	 * Check for expected errors before we dirty the transaction
3101	 * so we can return an error without a transaction abort.
3102	 */
3103	if (target_ip == NULL) {
3104		/*
3105		 * If there's no space reservation, check the entry will
3106		 * fit before actually inserting it.
3107		 */
3108		if (!spaceres) {
3109			error = xfs_dir_canenter(tp, target_dp, target_name);
3110			if (error)
3111				goto out_trans_cancel;
3112		}
3113	} else {
3114		/*
3115		 * If target exists and it's a directory, check that whether
3116		 * it can be destroyed.
3117		 */
3118		if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3119		    (!xfs_dir_isempty(target_ip) ||
3120		     (VFS_I(target_ip)->i_nlink > 2))) {
3121			error = -EEXIST;
3122			goto out_trans_cancel;
3123		}
3124	}
3125
3126	/*
3127	 * Lock the AGI buffers we need to handle bumping the nlink of the
3128	 * whiteout inode off the unlinked list and to handle dropping the
3129	 * nlink of the target inode.  Per locking order rules, do this in
3130	 * increasing AG order and before directory block allocation tries to
3131	 * grab AGFs because we grab AGIs before AGFs.
3132	 *
3133	 * The (vfs) caller must ensure that if src is a directory then
3134	 * target_ip is either null or an empty directory.
3135	 */
3136	for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3137		if (inodes[i] == wip ||
3138		    (inodes[i] == target_ip &&
3139		     (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3140			struct xfs_perag	*pag;
3141			struct xfs_buf		*bp;
3142
3143			pag = xfs_perag_get(mp,
3144					XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
3145			error = xfs_read_agi(pag, tp, &bp);
3146			xfs_perag_put(pag);
3147			if (error)
3148				goto out_trans_cancel;
3149		}
3150	}
3151
3152	/*
3153	 * Directory entry creation below may acquire the AGF. Remove
3154	 * the whiteout from the unlinked list first to preserve correct
3155	 * AGI/AGF locking order. This dirties the transaction so failures
3156	 * after this point will abort and log recovery will clean up the
3157	 * mess.
3158	 *
3159	 * For whiteouts, we need to bump the link count on the whiteout
3160	 * inode. After this point, we have a real link, clear the tmpfile
3161	 * state flag from the inode so it doesn't accidentally get misused
3162	 * in future.
3163	 */
3164	if (wip) {
3165		struct xfs_perag	*pag;
3166
3167		ASSERT(VFS_I(wip)->i_nlink == 0);
3168
3169		pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3170		error = xfs_iunlink_remove(tp, pag, wip);
3171		xfs_perag_put(pag);
3172		if (error)
3173			goto out_trans_cancel;
3174
3175		xfs_bumplink(tp, wip);
 
3176		VFS_I(wip)->i_state &= ~I_LINKABLE;
3177	}
3178
3179	/*
3180	 * Set up the target.
3181	 */
3182	if (target_ip == NULL) {
3183		/*
3184		 * If target does not exist and the rename crosses
3185		 * directories, adjust the target directory link count
3186		 * to account for the ".." reference from the new entry.
3187		 */
3188		error = xfs_dir_createname(tp, target_dp, target_name,
3189					   src_ip->i_ino, spaceres);
3190		if (error)
3191			goto out_trans_cancel;
3192
3193		xfs_trans_ichgtime(tp, target_dp,
3194					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3195
3196		if (new_parent && src_is_directory) {
3197			xfs_bumplink(tp, target_dp);
3198		}
3199	} else { /* target_ip != NULL */
3200		/*
3201		 * Link the source inode under the target name.
3202		 * If the source inode is a directory and we are moving
3203		 * it across directories, its ".." entry will be
3204		 * inconsistent until we replace that down below.
3205		 *
3206		 * In case there is already an entry with the same
3207		 * name at the destination directory, remove it first.
3208		 */
3209		error = xfs_dir_replace(tp, target_dp, target_name,
3210					src_ip->i_ino, spaceres);
3211		if (error)
3212			goto out_trans_cancel;
3213
3214		xfs_trans_ichgtime(tp, target_dp,
3215					XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3216
3217		/*
3218		 * Decrement the link count on the target since the target
3219		 * dir no longer points to it.
3220		 */
3221		error = xfs_droplink(tp, target_ip);
3222		if (error)
3223			goto out_trans_cancel;
3224
3225		if (src_is_directory) {
3226			/*
3227			 * Drop the link from the old "." entry.
3228			 */
3229			error = xfs_droplink(tp, target_ip);
3230			if (error)
3231				goto out_trans_cancel;
3232		}
3233	} /* target_ip != NULL */
3234
3235	/*
3236	 * Remove the source.
3237	 */
3238	if (new_parent && src_is_directory) {
3239		/*
3240		 * Rewrite the ".." entry to point to the new
3241		 * directory.
3242		 */
3243		error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3244					target_dp->i_ino, spaceres);
3245		ASSERT(error != -EEXIST);
3246		if (error)
3247			goto out_trans_cancel;
3248	}
3249
3250	/*
3251	 * We always want to hit the ctime on the source inode.
3252	 *
3253	 * This isn't strictly required by the standards since the source
3254	 * inode isn't really being changed, but old unix file systems did
3255	 * it and some incremental backup programs won't work without it.
3256	 */
3257	xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3258	xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3259
3260	/*
3261	 * Adjust the link count on src_dp.  This is necessary when
3262	 * renaming a directory, either within one parent when
3263	 * the target existed, or across two parent directories.
3264	 */
3265	if (src_is_directory && (new_parent || target_ip != NULL)) {
3266
3267		/*
3268		 * Decrement link count on src_directory since the
3269		 * entry that's moved no longer points to it.
3270		 */
3271		error = xfs_droplink(tp, src_dp);
3272		if (error)
3273			goto out_trans_cancel;
3274	}
3275
3276	/*
3277	 * For whiteouts, we only need to update the source dirent with the
3278	 * inode number of the whiteout inode rather than removing it
3279	 * altogether.
3280	 */
3281	if (wip)
3282		error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3283					spaceres);
3284	else
3285		error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3286					   spaceres);
3287
3288	if (error)
3289		goto out_trans_cancel;
3290
3291	xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3292	xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3293	if (new_parent)
3294		xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3295
3296	/*
3297	 * Inform our hook clients that we've finished a rename operation as
3298	 * follows: removed the source and target files from their directories;
3299	 * that we've added the source to the target directory; and finally
3300	 * that we've added the whiteout, if there was one.  All inodes are
3301	 * locked, so it's ok to model a rename this way so long as we say we
3302	 * deleted entries before we add new ones.
3303	 */
3304	if (target_ip)
3305		xfs_dir_update_hook(target_dp, target_ip, -1, target_name);
3306	xfs_dir_update_hook(src_dp, src_ip, -1, src_name);
3307	xfs_dir_update_hook(target_dp, src_ip, 1, target_name);
3308	if (wip)
3309		xfs_dir_update_hook(src_dp, wip, 1, src_name);
3310
3311	error = xfs_finish_rename(tp);
3312	if (wip)
3313		xfs_irele(wip);
3314	return error;
3315
3316out_trans_cancel:
3317	xfs_trans_cancel(tp);
3318out_release_wip:
3319	if (wip)
3320		xfs_irele(wip);
3321	if (error == -ENOSPC && nospace_error)
3322		error = nospace_error;
3323	return error;
3324}
3325
3326static int
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3327xfs_iflush(
3328	struct xfs_inode	*ip,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3329	struct xfs_buf		*bp)
3330{
3331	struct xfs_inode_log_item *iip = ip->i_itemp;
3332	struct xfs_dinode	*dip;
3333	struct xfs_mount	*mp = ip->i_mount;
3334	int			error;
3335
3336	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
3337	ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3338	ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3339	       ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3340	ASSERT(iip->ili_item.li_buf == bp);
 
3341
 
3342	dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3343
3344	/*
3345	 * We don't flush the inode if any of the following checks fail, but we
3346	 * do still update the log item and attach to the backing buffer as if
3347	 * the flush happened. This is a formality to facilitate predictable
3348	 * error handling as the caller will shutdown and fail the buffer.
3349	 */
3350	error = -EFSCORRUPTED;
3351	if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3352			       mp, XFS_ERRTAG_IFLUSH_1)) {
3353		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3354			"%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
3355			__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3356		goto flush_out;
3357	}
3358	if (S_ISREG(VFS_I(ip)->i_mode)) {
3359		if (XFS_TEST_ERROR(
3360		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3361		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3362		    mp, XFS_ERRTAG_IFLUSH_3)) {
3363			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3364				"%s: Bad regular inode %llu, ptr "PTR_FMT,
3365				__func__, ip->i_ino, ip);
3366			goto flush_out;
3367		}
3368	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3369		if (XFS_TEST_ERROR(
3370		    ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3371		    ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3372		    ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3373		    mp, XFS_ERRTAG_IFLUSH_4)) {
3374			xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3375				"%s: Bad directory inode %llu, ptr "PTR_FMT,
3376				__func__, ip->i_ino, ip);
3377			goto flush_out;
3378		}
3379	}
3380	if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
3381				ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3382		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3383			"%s: detected corrupt incore inode %llu, "
3384			"total extents = %llu nblocks = %lld, ptr "PTR_FMT,
3385			__func__, ip->i_ino,
3386			ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
3387			ip->i_nblocks, ip);
3388		goto flush_out;
3389	}
3390	if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3391				mp, XFS_ERRTAG_IFLUSH_6)) {
3392		xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3393			"%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
3394			__func__, ip->i_ino, ip->i_forkoff, ip);
3395		goto flush_out;
3396	}
3397
3398	/*
3399	 * Inode item log recovery for v2 inodes are dependent on the flushiter
3400	 * count for correct sequencing.  We bump the flush iteration count so
3401	 * we can detect flushes which postdate a log record during recovery.
3402	 * This is redundant as we now log every change and hence this can't
3403	 * happen but we need to still do it to ensure backwards compatibility
3404	 * with old kernels that predate logging all inode changes.
3405	 */
3406	if (!xfs_has_v3inodes(mp))
3407		ip->i_flushiter++;
3408
3409	/*
3410	 * If there are inline format data / attr forks attached to this inode,
3411	 * make sure they are not corrupt.
3412	 */
3413	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3414	    xfs_ifork_verify_local_data(ip))
3415		goto flush_out;
3416	if (xfs_inode_has_attr_fork(ip) &&
3417	    ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
3418	    xfs_ifork_verify_local_attr(ip))
3419		goto flush_out;
3420
3421	/*
3422	 * Copy the dirty parts of the inode into the on-disk inode.  We always
3423	 * copy out the core of the inode, because if the inode is dirty at all
3424	 * the core must be.
3425	 */
3426	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3427
3428	/* Wrap, we never let the log put out DI_MAX_FLUSH */
3429	if (!xfs_has_v3inodes(mp)) {
3430		if (ip->i_flushiter == DI_MAX_FLUSH)
3431			ip->i_flushiter = 0;
3432	}
3433
3434	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3435	if (xfs_inode_has_attr_fork(ip))
3436		xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
 
3437
3438	/*
3439	 * We've recorded everything logged in the inode, so we'd like to clear
3440	 * the ili_fields bits so we don't log and flush things unnecessarily.
3441	 * However, we can't stop logging all this information until the data
3442	 * we've copied into the disk buffer is written to disk.  If we did we
3443	 * might overwrite the copy of the inode in the log with all the data
3444	 * after re-logging only part of it, and in the face of a crash we
3445	 * wouldn't have all the data we need to recover.
3446	 *
3447	 * What we do is move the bits to the ili_last_fields field.  When
3448	 * logging the inode, these bits are moved back to the ili_fields field.
3449	 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3450	 * we know that the information those bits represent is permanently on
3451	 * disk.  As long as the flush completes before the inode is logged
3452	 * again, then both ili_fields and ili_last_fields will be cleared.
 
 
 
 
 
 
 
 
 
3453	 */
3454	error = 0;
3455flush_out:
3456	spin_lock(&iip->ili_lock);
3457	iip->ili_last_fields = iip->ili_fields;
3458	iip->ili_fields = 0;
3459	iip->ili_fsync_fields = 0;
3460	spin_unlock(&iip->ili_lock);
3461
3462	/*
3463	 * Store the current LSN of the inode so that we can tell whether the
3464	 * item has moved in the AIL from xfs_buf_inode_iodone().
3465	 */
3466	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3467				&iip->ili_item.li_lsn);
3468
3469	/* generate the checksum. */
3470	xfs_dinode_calc_crc(mp, dip);
3471	if (error)
3472		xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
3473	return error;
3474}
3475
3476/*
3477 * Non-blocking flush of dirty inode metadata into the backing buffer.
3478 *
3479 * The caller must have a reference to the inode and hold the cluster buffer
3480 * locked. The function will walk across all the inodes on the cluster buffer it
3481 * can find and lock without blocking, and flush them to the cluster buffer.
3482 *
3483 * On successful flushing of at least one inode, the caller must write out the
3484 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3485 * the caller needs to release the buffer. On failure, the filesystem will be
3486 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3487 * will be returned.
3488 */
3489int
3490xfs_iflush_cluster(
3491	struct xfs_buf		*bp)
3492{
3493	struct xfs_mount	*mp = bp->b_mount;
3494	struct xfs_log_item	*lip, *n;
3495	struct xfs_inode	*ip;
3496	struct xfs_inode_log_item *iip;
3497	int			clcount = 0;
3498	int			error = 0;
3499
3500	/*
3501	 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3502	 * will remove itself from the list.
 
 
3503	 */
3504	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3505		iip = (struct xfs_inode_log_item *)lip;
3506		ip = iip->ili_inode;
3507
3508		/*
3509		 * Quick and dirty check to avoid locks if possible.
3510		 */
3511		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3512			continue;
3513		if (xfs_ipincount(ip))
3514			continue;
3515
3516		/*
3517		 * The inode is still attached to the buffer, which means it is
3518		 * dirty but reclaim might try to grab it. Check carefully for
3519		 * that, and grab the ilock while still holding the i_flags_lock
3520		 * to guarantee reclaim will not be able to reclaim this inode
3521		 * once we drop the i_flags_lock.
3522		 */
3523		spin_lock(&ip->i_flags_lock);
3524		ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3525		if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3526			spin_unlock(&ip->i_flags_lock);
3527			continue;
3528		}
3529
3530		/*
3531		 * ILOCK will pin the inode against reclaim and prevent
3532		 * concurrent transactions modifying the inode while we are
3533		 * flushing the inode. If we get the lock, set the flushing
3534		 * state before we drop the i_flags_lock.
3535		 */
3536		if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3537			spin_unlock(&ip->i_flags_lock);
3538			continue;
3539		}
3540		__xfs_iflags_set(ip, XFS_IFLUSHING);
3541		spin_unlock(&ip->i_flags_lock);
3542
3543		/*
3544		 * Abort flushing this inode if we are shut down because the
3545		 * inode may not currently be in the AIL. This can occur when
3546		 * log I/O failure unpins the inode without inserting into the
3547		 * AIL, leaving a dirty/unpinned inode attached to the buffer
3548		 * that otherwise looks like it should be flushed.
3549		 */
3550		if (xlog_is_shutdown(mp->m_log)) {
3551			xfs_iunpin_wait(ip);
3552			xfs_iflush_abort(ip);
3553			xfs_iunlock(ip, XFS_ILOCK_SHARED);
3554			error = -EIO;
3555			continue;
3556		}
3557
3558		/* don't block waiting on a log force to unpin dirty inodes */
3559		if (xfs_ipincount(ip)) {
3560			xfs_iflags_clear(ip, XFS_IFLUSHING);
3561			xfs_iunlock(ip, XFS_ILOCK_SHARED);
3562			continue;
3563		}
3564
3565		if (!xfs_inode_clean(ip))
3566			error = xfs_iflush(ip, bp);
3567		else
3568			xfs_iflags_clear(ip, XFS_IFLUSHING);
3569		xfs_iunlock(ip, XFS_ILOCK_SHARED);
3570		if (error)
3571			break;
3572		clcount++;
3573	}
3574
3575	if (error) {
3576		/*
3577		 * Shutdown first so we kill the log before we release this
3578		 * buffer. If it is an INODE_ALLOC buffer and pins the tail
3579		 * of the log, failing it before the _log_ is shut down can
3580		 * result in the log tail being moved forward in the journal
3581		 * on disk because log writes can still be taking place. Hence
3582		 * unpinning the tail will allow the ICREATE intent to be
3583		 * removed from the log an recovery will fail with uninitialised
3584		 * inode cluster buffers.
3585		 */
3586		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3587		bp->b_flags |= XBF_ASYNC;
3588		xfs_buf_ioend_fail(bp);
3589		return error;
3590	}
3591
3592	if (!clcount)
3593		return -EAGAIN;
3594
3595	XFS_STATS_INC(mp, xs_icluster_flushcnt);
3596	XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3597	return 0;
3598
 
 
3599}
3600
3601/* Release an inode. */
3602void
3603xfs_irele(
3604	struct xfs_inode	*ip)
3605{
3606	trace_xfs_irele(ip, _RET_IP_);
3607	iput(VFS_I(ip));
3608}
3609
3610/*
3611 * Ensure all commited transactions touching the inode are written to the log.
3612 */
3613int
3614xfs_log_force_inode(
3615	struct xfs_inode	*ip)
3616{
3617	xfs_csn_t		seq = 0;
3618
3619	xfs_ilock(ip, XFS_ILOCK_SHARED);
3620	if (xfs_ipincount(ip))
3621		seq = ip->i_itemp->ili_commit_seq;
3622	xfs_iunlock(ip, XFS_ILOCK_SHARED);
3623
3624	if (!seq)
3625		return 0;
3626	return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3627}
3628
3629/*
3630 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3631 * abide vfs locking order (lowest pointer value goes first) and breaking the
3632 * layout leases before proceeding.  The loop is needed because we cannot call
3633 * the blocking break_layout() with the iolocks held, and therefore have to
3634 * back out both locks.
3635 */
3636static int
3637xfs_iolock_two_inodes_and_break_layout(
3638	struct inode		*src,
3639	struct inode		*dest)
3640{
3641	int			error;
3642
3643	if (src > dest)
3644		swap(src, dest);
3645
3646retry:
3647	/* Wait to break both inodes' layouts before we start locking. */
3648	error = break_layout(src, true);
3649	if (error)
3650		return error;
3651	if (src != dest) {
3652		error = break_layout(dest, true);
3653		if (error)
3654			return error;
3655	}
3656
3657	/* Lock one inode and make sure nobody got in and leased it. */
3658	inode_lock(src);
3659	error = break_layout(src, false);
3660	if (error) {
3661		inode_unlock(src);
3662		if (error == -EWOULDBLOCK)
3663			goto retry;
3664		return error;
3665	}
3666
3667	if (src == dest)
3668		return 0;
3669
3670	/* Lock the other inode and make sure nobody got in and leased it. */
3671	inode_lock_nested(dest, I_MUTEX_NONDIR2);
3672	error = break_layout(dest, false);
3673	if (error) {
3674		inode_unlock(src);
3675		inode_unlock(dest);
3676		if (error == -EWOULDBLOCK)
3677			goto retry;
3678		return error;
3679	}
3680
3681	return 0;
3682}
3683
3684static int
3685xfs_mmaplock_two_inodes_and_break_dax_layout(
3686	struct xfs_inode	*ip1,
3687	struct xfs_inode	*ip2)
3688{
3689	int			error;
3690	bool			retry;
3691	struct page		*page;
3692
3693	if (ip1->i_ino > ip2->i_ino)
3694		swap(ip1, ip2);
3695
3696again:
3697	retry = false;
3698	/* Lock the first inode */
3699	xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3700	error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
3701	if (error || retry) {
3702		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3703		if (error == 0 && retry)
3704			goto again;
3705		return error;
3706	}
3707
3708	if (ip1 == ip2)
3709		return 0;
3710
3711	/* Nested lock the second inode */
3712	xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
3713	/*
3714	 * We cannot use xfs_break_dax_layouts() directly here because it may
3715	 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
3716	 * for this nested lock case.
3717	 */
3718	page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
3719	if (page && page_ref_count(page) != 1) {
3720		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3721		xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3722		goto again;
3723	}
3724
3725	return 0;
3726}
3727
3728/*
3729 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3730 * mmap activity.
3731 */
3732int
3733xfs_ilock2_io_mmap(
3734	struct xfs_inode	*ip1,
3735	struct xfs_inode	*ip2)
3736{
3737	int			ret;
3738
3739	ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3740	if (ret)
3741		return ret;
3742
3743	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3744		ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
3745		if (ret) {
3746			inode_unlock(VFS_I(ip2));
3747			if (ip1 != ip2)
3748				inode_unlock(VFS_I(ip1));
3749			return ret;
3750		}
3751	} else
3752		filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
3753					    VFS_I(ip2)->i_mapping);
3754
3755	return 0;
3756}
3757
3758/* Unlock both inodes to allow IO and mmap activity. */
3759void
3760xfs_iunlock2_io_mmap(
3761	struct xfs_inode	*ip1,
3762	struct xfs_inode	*ip2)
3763{
3764	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3765		xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3766		if (ip1 != ip2)
3767			xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3768	} else
3769		filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
3770					      VFS_I(ip2)->i_mapping);
3771
3772	inode_unlock(VFS_I(ip2));
3773	if (ip1 != ip2)
3774		inode_unlock(VFS_I(ip1));
3775}
3776
3777/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
3778void
3779xfs_iunlock2_remapping(
3780	struct xfs_inode	*ip1,
3781	struct xfs_inode	*ip2)
3782{
3783	xfs_iflags_clear(ip1, XFS_IREMAPPING);
3784
3785	if (ip1 != ip2)
3786		xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
3787	xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3788
3789	if (ip1 != ip2)
3790		inode_unlock_shared(VFS_I(ip1));
3791	inode_unlock(VFS_I(ip2));
3792}
3793
3794/*
3795 * Reload the incore inode list for this inode.  Caller should ensure that
3796 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
3797 * preventing other threads from executing.
3798 */
3799int
3800xfs_inode_reload_unlinked_bucket(
3801	struct xfs_trans	*tp,
3802	struct xfs_inode	*ip)
3803{
3804	struct xfs_mount	*mp = tp->t_mountp;
3805	struct xfs_buf		*agibp;
3806	struct xfs_agi		*agi;
3807	struct xfs_perag	*pag;
3808	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
3809	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
3810	xfs_agino_t		prev_agino, next_agino;
3811	unsigned int		bucket;
3812	bool			foundit = false;
3813	int			error;
3814
3815	/* Grab the first inode in the list */
3816	pag = xfs_perag_get(mp, agno);
3817	error = xfs_ialloc_read_agi(pag, tp, &agibp);
3818	xfs_perag_put(pag);
3819	if (error)
3820		return error;
3821
3822	/*
3823	 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
3824	 * incore unlinked list pointers for this inode.  Check once more to
3825	 * see if we raced with anyone else to reload the unlinked list.
3826	 */
3827	if (!xfs_inode_unlinked_incomplete(ip)) {
3828		foundit = true;
3829		goto out_agibp;
3830	}
3831
3832	bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
3833	agi = agibp->b_addr;
3834
3835	trace_xfs_inode_reload_unlinked_bucket(ip);
3836
3837	xfs_info_ratelimited(mp,
3838 "Found unrecovered unlinked inode 0x%x in AG 0x%x.  Initiating list recovery.",
3839			agino, agno);
3840
3841	prev_agino = NULLAGINO;
3842	next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3843	while (next_agino != NULLAGINO) {
3844		struct xfs_inode	*next_ip = NULL;
3845
3846		/* Found this caller's inode, set its backlink. */
3847		if (next_agino == agino) {
3848			next_ip = ip;
3849			next_ip->i_prev_unlinked = prev_agino;
3850			foundit = true;
3851			goto next_inode;
3852		}
3853
3854		/* Try in-memory lookup first. */
3855		next_ip = xfs_iunlink_lookup(pag, next_agino);
3856		if (next_ip)
3857			goto next_inode;
3858
3859		/* Inode not in memory, try reloading it. */
3860		error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
3861				next_agino);
3862		if (error)
3863			break;
3864
3865		/* Grab the reloaded inode. */
3866		next_ip = xfs_iunlink_lookup(pag, next_agino);
3867		if (!next_ip) {
3868			/* No incore inode at all?  We reloaded it... */
3869			ASSERT(next_ip != NULL);
3870			error = -EFSCORRUPTED;
3871			break;
3872		}
3873
3874next_inode:
3875		prev_agino = next_agino;
3876		next_agino = next_ip->i_next_unlinked;
3877	}
3878
3879out_agibp:
3880	xfs_trans_brelse(tp, agibp);
3881	/* Should have found this inode somewhere in the iunlinked bucket. */
3882	if (!error && !foundit)
3883		error = -EFSCORRUPTED;
3884	return error;
3885}
3886
3887/* Decide if this inode is missing its unlinked list and reload it. */
3888int
3889xfs_inode_reload_unlinked(
3890	struct xfs_inode	*ip)
3891{
3892	struct xfs_trans	*tp;
3893	int			error;
3894
3895	error = xfs_trans_alloc_empty(ip->i_mount, &tp);
3896	if (error)
3897		return error;
3898
3899	xfs_ilock(ip, XFS_ILOCK_SHARED);
3900	if (xfs_inode_unlinked_incomplete(ip))
3901		error = xfs_inode_reload_unlinked_bucket(tp, ip);
3902	xfs_iunlock(ip, XFS_ILOCK_SHARED);
3903	xfs_trans_cancel(tp);
3904
3905	return error;
3906}
3907
3908/* Has this inode fork been zapped by repair? */
3909bool
3910xfs_ifork_zapped(
3911	const struct xfs_inode	*ip,
3912	int			whichfork)
3913{
3914	unsigned int		datamask = 0;
3915
3916	switch (whichfork) {
3917	case XFS_DATA_FORK:
3918		switch (ip->i_vnode.i_mode & S_IFMT) {
3919		case S_IFDIR:
3920			datamask = XFS_SICK_INO_DIR_ZAPPED;
3921			break;
3922		case S_IFLNK:
3923			datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
3924			break;
3925		}
3926		return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
3927	case XFS_ATTR_FORK:
3928		return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
3929	default:
3930		return false;
3931	}
3932}
3933
3934/* Compute the number of data and realtime blocks used by a file. */
3935void
3936xfs_inode_count_blocks(
3937	struct xfs_trans	*tp,
3938	struct xfs_inode	*ip,
3939	xfs_filblks_t		*dblocks,
3940	xfs_filblks_t		*rblocks)
3941{
3942	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
3943
3944	*rblocks = 0;
3945	if (XFS_IS_REALTIME_INODE(ip))
3946		xfs_bmap_count_leaves(ifp, rblocks);
3947	*dblocks = ip->i_nblocks - *rblocks;
3948}