1/*
   2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_format.h"
  21#include "xfs_log_format.h"
  22#include "xfs_trans_resv.h"
 
  23#include "xfs_sb.h"
 
  24#include "xfs_mount.h"
  25#include "xfs_inode.h"
  26#include "xfs_error.h"
  27#include "xfs_trans.h"
  28#include "xfs_trans_priv.h"
  29#include "xfs_inode_item.h"
  30#include "xfs_quota.h"
  31#include "xfs_trace.h"
  32#include "xfs_icache.h"
  33#include "xfs_bmap_util.h"
  34#include "xfs_dquot_item.h"
  35#include "xfs_dquot.h"
  36#include "xfs_reflink.h"
  37
  38#include <linux/kthread.h>
  39#include <linux/freezer.h>
  40#include <linux/iversion.h>
 
 
  41
  42/*
  43 * Allocate and initialise an xfs_inode.
  44 */
  45struct xfs_inode *
  46xfs_inode_alloc(
  47	struct xfs_mount	*mp,
  48	xfs_ino_t		ino)
  49{
  50	struct xfs_inode	*ip;
  51
  52	/*
  53	 * if this didn't occur in transactions, we could use
  54	 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
  55	 * code up to do this anyway.
  56	 */
  57	ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
  58	if (!ip)
  59		return NULL;
  60	if (inode_init_always(mp->m_super, VFS_I(ip))) {
  61		kmem_zone_free(xfs_inode_zone, ip);
  62		return NULL;
  63	}
  64
  65	/* VFS doesn't initialise i_mode! */
  66	VFS_I(ip)->i_mode = 0;
  67
  68	XFS_STATS_INC(mp, vn_active);
  69	ASSERT(atomic_read(&ip->i_pincount) == 0);
 
  70	ASSERT(!xfs_isiflocked(ip));
  71	ASSERT(ip->i_ino == 0);
  72
 
 
  73	/* initialise the xfs inode */
  74	ip->i_ino = ino;
  75	ip->i_mount = mp;
  76	memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
  77	ip->i_afp = NULL;
  78	ip->i_cowfp = NULL;
  79	ip->i_cnextents = 0;
  80	ip->i_cformat = XFS_DINODE_FMT_EXTENTS;
  81	memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
  82	ip->i_flags = 0;
  83	ip->i_delayed_blks = 0;
  84	memset(&ip->i_d, 0, sizeof(ip->i_d));
  85
  86	return ip;
  87}
  88
  89STATIC void
  90xfs_inode_free_callback(
  91	struct rcu_head		*head)
  92{
  93	struct inode		*inode = container_of(head, struct inode, i_rcu);
  94	struct xfs_inode	*ip = XFS_I(inode);
  95
  96	switch (VFS_I(ip)->i_mode & S_IFMT) {
 
 
 
 
 
 
 
  97	case S_IFREG:
  98	case S_IFDIR:
  99	case S_IFLNK:
 100		xfs_idestroy_fork(ip, XFS_DATA_FORK);
 101		break;
 102	}
 103
 104	if (ip->i_afp)
 105		xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 106	if (ip->i_cowfp)
 107		xfs_idestroy_fork(ip, XFS_COW_FORK);
 108
 109	if (ip->i_itemp) {
 110		ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
 111		xfs_inode_item_destroy(ip);
 112		ip->i_itemp = NULL;
 113	}
 114
 115	kmem_zone_free(xfs_inode_zone, ip);
 116}
 117
 118static void
 119__xfs_inode_free(
 120	struct xfs_inode	*ip)
 121{
 122	/* asserts to verify all state is correct here */
 123	ASSERT(atomic_read(&ip->i_pincount) == 0);
 124	XFS_STATS_DEC(ip->i_mount, vn_active);
 125
 126	call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
 127}
 128
 129void
 130xfs_inode_free(
 131	struct xfs_inode	*ip)
 132{
 133	ASSERT(!xfs_isiflocked(ip));
 134
 135	/*
 136	 * Because we use RCU freeing we need to ensure the inode always
 137	 * appears to be reclaimed with an invalid inode number when in the
 138	 * free state. The ip->i_flags_lock provides the barrier against lookup
 139	 * races.
 140	 */
 141	spin_lock(&ip->i_flags_lock);
 142	ip->i_flags = XFS_IRECLAIM;
 143	ip->i_ino = 0;
 144	spin_unlock(&ip->i_flags_lock);
 145
 146	__xfs_inode_free(ip);
 147}
 148
 149/*
 150 * Queue a new inode reclaim pass if there are reclaimable inodes and there
 151 * isn't a reclaim pass already in progress. By default it runs every 5s based
 152 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
 153 * tunable, but that can be done if this method proves to be ineffective or too
 154 * aggressive.
 155 */
 156static void
 157xfs_reclaim_work_queue(
 158	struct xfs_mount        *mp)
 159{
 160
 161	rcu_read_lock();
 162	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
 163		queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
 164			msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
 165	}
 166	rcu_read_unlock();
 167}
 168
 169/*
 170 * This is a fast pass over the inode cache to try to get reclaim moving on as
 171 * many inodes as possible in a short period of time. It kicks itself every few
 172 * seconds, as well as being kicked by the inode cache shrinker when memory
 173 * goes low. It scans as quickly as possible avoiding locked inodes or those
 174 * already being flushed, and once done schedules a future pass.
 175 */
 176void
 177xfs_reclaim_worker(
 178	struct work_struct *work)
 179{
 180	struct xfs_mount *mp = container_of(to_delayed_work(work),
 181					struct xfs_mount, m_reclaim_work);
 182
 183	xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
 184	xfs_reclaim_work_queue(mp);
 185}
 186
 187static void
 188xfs_perag_set_reclaim_tag(
 189	struct xfs_perag	*pag)
 190{
 191	struct xfs_mount	*mp = pag->pag_mount;
 192
 193	lockdep_assert_held(&pag->pag_ici_lock);
 194	if (pag->pag_ici_reclaimable++)
 195		return;
 196
 197	/* propagate the reclaim tag up into the perag radix tree */
 198	spin_lock(&mp->m_perag_lock);
 199	radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno,
 200			   XFS_ICI_RECLAIM_TAG);
 201	spin_unlock(&mp->m_perag_lock);
 202
 203	/* schedule periodic background inode reclaim */
 204	xfs_reclaim_work_queue(mp);
 205
 206	trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
 207}
 208
 209static void
 210xfs_perag_clear_reclaim_tag(
 211	struct xfs_perag	*pag)
 212{
 213	struct xfs_mount	*mp = pag->pag_mount;
 214
 215	lockdep_assert_held(&pag->pag_ici_lock);
 216	if (--pag->pag_ici_reclaimable)
 217		return;
 218
 219	/* clear the reclaim tag from the perag radix tree */
 220	spin_lock(&mp->m_perag_lock);
 221	radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno,
 222			     XFS_ICI_RECLAIM_TAG);
 223	spin_unlock(&mp->m_perag_lock);
 224	trace_xfs_perag_clear_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
 225}
 226
 227
 228/*
 229 * We set the inode flag atomically with the radix tree tag.
 230 * Once we get tag lookups on the radix tree, this inode flag
 231 * can go away.
 232 */
 233void
 234xfs_inode_set_reclaim_tag(
 235	struct xfs_inode	*ip)
 236{
 237	struct xfs_mount	*mp = ip->i_mount;
 238	struct xfs_perag	*pag;
 239
 240	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
 241	spin_lock(&pag->pag_ici_lock);
 242	spin_lock(&ip->i_flags_lock);
 243
 244	radix_tree_tag_set(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino),
 245			   XFS_ICI_RECLAIM_TAG);
 246	xfs_perag_set_reclaim_tag(pag);
 247	__xfs_iflags_set(ip, XFS_IRECLAIMABLE);
 248
 249	spin_unlock(&ip->i_flags_lock);
 250	spin_unlock(&pag->pag_ici_lock);
 251	xfs_perag_put(pag);
 252}
 253
 254STATIC void
 255xfs_inode_clear_reclaim_tag(
 256	struct xfs_perag	*pag,
 257	xfs_ino_t		ino)
 258{
 259	radix_tree_tag_clear(&pag->pag_ici_root,
 260			     XFS_INO_TO_AGINO(pag->pag_mount, ino),
 261			     XFS_ICI_RECLAIM_TAG);
 262	xfs_perag_clear_reclaim_tag(pag);
 263}
 264
 265static void
 266xfs_inew_wait(
 267	struct xfs_inode	*ip)
 268{
 269	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
 270	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
 271
 272	do {
 273		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
 274		if (!xfs_iflags_test(ip, XFS_INEW))
 275			break;
 276		schedule();
 277	} while (true);
 278	finish_wait(wq, &wait.wq_entry);
 279}
 280
 281/*
 282 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
 283 * part of the structure. This is made more complex by the fact we store
 284 * information about the on-disk values in the VFS inode and so we can't just
 285 * overwrite the values unconditionally. Hence we save the parameters we
 286 * need to retain across reinitialisation, and rewrite them into the VFS inode
 287 * after reinitialisation even if it fails.
 288 */
 289static int
 290xfs_reinit_inode(
 291	struct xfs_mount	*mp,
 292	struct inode		*inode)
 293{
 294	int		error;
 295	uint32_t	nlink = inode->i_nlink;
 296	uint32_t	generation = inode->i_generation;
 297	uint64_t	version = inode_peek_iversion(inode);
 298	umode_t		mode = inode->i_mode;
 299	dev_t		dev = inode->i_rdev;
 300
 301	error = inode_init_always(mp->m_super, inode);
 302
 303	set_nlink(inode, nlink);
 304	inode->i_generation = generation;
 305	inode_set_iversion_queried(inode, version);
 306	inode->i_mode = mode;
 307	inode->i_rdev = dev;
 308	return error;
 309}
 310
 311/*
 312 * Check the validity of the inode we just found it the cache
 313 */
 314static int
 315xfs_iget_cache_hit(
 316	struct xfs_perag	*pag,
 317	struct xfs_inode	*ip,
 318	xfs_ino_t		ino,
 319	int			flags,
 320	int			lock_flags) __releases(RCU)
 321{
 322	struct inode		*inode = VFS_I(ip);
 323	struct xfs_mount	*mp = ip->i_mount;
 324	int			error;
 325
 326	/*
 327	 * check for re-use of an inode within an RCU grace period due to the
 328	 * radix tree nodes not being updated yet. We monitor for this by
 329	 * setting the inode number to zero before freeing the inode structure.
 330	 * If the inode has been reallocated and set up, then the inode number
 331	 * will not match, so check for that, too.
 332	 */
 333	spin_lock(&ip->i_flags_lock);
 334	if (ip->i_ino != ino) {
 335		trace_xfs_iget_skip(ip);
 336		XFS_STATS_INC(mp, xs_ig_frecycle);
 337		error = -EAGAIN;
 338		goto out_error;
 339	}
 340
 341
 342	/*
 343	 * If we are racing with another cache hit that is currently
 344	 * instantiating this inode or currently recycling it out of
 345	 * reclaimabe state, wait for the initialisation to complete
 346	 * before continuing.
 347	 *
 348	 * XXX(hch): eventually we should do something equivalent to
 349	 *	     wait_on_inode to wait for these flags to be cleared
 350	 *	     instead of polling for it.
 351	 */
 352	if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
 353		trace_xfs_iget_skip(ip);
 354		XFS_STATS_INC(mp, xs_ig_frecycle);
 355		error = -EAGAIN;
 356		goto out_error;
 357	}
 358
 359	/*
 360	 * If lookup is racing with unlink return an error immediately.
 361	 */
 362	if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
 363		error = -ENOENT;
 364		goto out_error;
 365	}
 366
 367	/*
 368	 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
 369	 * Need to carefully get it back into useable state.
 370	 */
 371	if (ip->i_flags & XFS_IRECLAIMABLE) {
 372		trace_xfs_iget_reclaim(ip);
 373
 374		if (flags & XFS_IGET_INCORE) {
 375			error = -EAGAIN;
 376			goto out_error;
 377		}
 378
 379		/*
 380		 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
 381		 * from stomping over us while we recycle the inode.  We can't
 382		 * clear the radix tree reclaimable tag yet as it requires
 383		 * pag_ici_lock to be held exclusive.
 384		 */
 385		ip->i_flags |= XFS_IRECLAIM;
 386
 387		spin_unlock(&ip->i_flags_lock);
 388		rcu_read_unlock();
 389
 390		error = xfs_reinit_inode(mp, inode);
 391		if (error) {
 392			bool wake;
 393			/*
 394			 * Re-initializing the inode failed, and we are in deep
 395			 * trouble.  Try to re-add it to the reclaim list.
 396			 */
 397			rcu_read_lock();
 398			spin_lock(&ip->i_flags_lock);
 399			wake = !!__xfs_iflags_test(ip, XFS_INEW);
 400			ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
 401			if (wake)
 402				wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
 403			ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
 404			trace_xfs_iget_reclaim_fail(ip);
 405			goto out_error;
 406		}
 407
 408		spin_lock(&pag->pag_ici_lock);
 409		spin_lock(&ip->i_flags_lock);
 410
 411		/*
 412		 * Clear the per-lifetime state in the inode as we are now
 413		 * effectively a new inode and need to return to the initial
 414		 * state before reuse occurs.
 415		 */
 416		ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
 417		ip->i_flags |= XFS_INEW;
 418		xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
 419		inode->i_state = I_NEW;
 420
 421		ASSERT(!rwsem_is_locked(&inode->i_rwsem));
 422		init_rwsem(&inode->i_rwsem);
 423
 424		spin_unlock(&ip->i_flags_lock);
 425		spin_unlock(&pag->pag_ici_lock);
 426	} else {
 427		/* If the VFS inode is being torn down, pause and try again. */
 428		if (!igrab(inode)) {
 429			trace_xfs_iget_skip(ip);
 430			error = -EAGAIN;
 431			goto out_error;
 432		}
 433
 434		/* We've got a live one. */
 435		spin_unlock(&ip->i_flags_lock);
 436		rcu_read_unlock();
 437		trace_xfs_iget_hit(ip);
 438	}
 439
 440	if (lock_flags != 0)
 441		xfs_ilock(ip, lock_flags);
 442
 443	if (!(flags & XFS_IGET_INCORE))
 444		xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
 445	XFS_STATS_INC(mp, xs_ig_found);
 446
 447	return 0;
 448
 449out_error:
 450	spin_unlock(&ip->i_flags_lock);
 451	rcu_read_unlock();
 452	return error;
 453}
 454
 455
 456static int
 457xfs_iget_cache_miss(
 458	struct xfs_mount	*mp,
 459	struct xfs_perag	*pag,
 460	xfs_trans_t		*tp,
 461	xfs_ino_t		ino,
 462	struct xfs_inode	**ipp,
 463	int			flags,
 464	int			lock_flags)
 465{
 466	struct xfs_inode	*ip;
 467	int			error;
 468	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);
 469	int			iflags;
 470
 471	ip = xfs_inode_alloc(mp, ino);
 472	if (!ip)
 473		return -ENOMEM;
 474
 475	error = xfs_iread(mp, tp, ip, flags);
 476	if (error)
 477		goto out_destroy;
 478
 479	if (!xfs_inode_verify_forks(ip)) {
 480		error = -EFSCORRUPTED;
 481		goto out_destroy;
 482	}
 483
 484	trace_xfs_iget_miss(ip);
 485
 486
 487	/*
 488	 * If we are allocating a new inode, then check what was returned is
 489	 * actually a free, empty inode. If we are not allocating an inode,
 490	 * the check we didn't find a free inode.
 491	 */
 492	if (flags & XFS_IGET_CREATE) {
 493		if (VFS_I(ip)->i_mode != 0) {
 494			xfs_warn(mp,
 495"Corruption detected! Free inode 0x%llx not marked free on disk",
 496				ino);
 497			error = -EFSCORRUPTED;
 498			goto out_destroy;
 499		}
 500		if (ip->i_d.di_nblocks != 0) {
 501			xfs_warn(mp,
 502"Corruption detected! Free inode 0x%llx has blocks allocated!",
 503				ino);
 504			error = -EFSCORRUPTED;
 505			goto out_destroy;
 506		}
 507	} else if (VFS_I(ip)->i_mode == 0) {
 508		error = -ENOENT;
 509		goto out_destroy;
 510	}
 511
 512	/*
 513	 * Preload the radix tree so we can insert safely under the
 514	 * write spinlock. Note that we cannot sleep inside the preload
 515	 * region. Since we can be called from transaction context, don't
 516	 * recurse into the file system.
 517	 */
 518	if (radix_tree_preload(GFP_NOFS)) {
 519		error = -EAGAIN;
 520		goto out_destroy;
 521	}
 522
 523	/*
 524	 * Because the inode hasn't been added to the radix-tree yet it can't
 525	 * be found by another thread, so we can do the non-sleeping lock here.
 526	 */
 527	if (lock_flags) {
 528		if (!xfs_ilock_nowait(ip, lock_flags))
 529			BUG();
 530	}
 531
 532	/*
 533	 * These values must be set before inserting the inode into the radix
 534	 * tree as the moment it is inserted a concurrent lookup (allowed by the
 535	 * RCU locking mechanism) can find it and that lookup must see that this
 536	 * is an inode currently under construction (i.e. that XFS_INEW is set).
 537	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
 538	 * memory barrier that ensures this detection works correctly at lookup
 539	 * time.
 540	 */
 541	iflags = XFS_INEW;
 542	if (flags & XFS_IGET_DONTCACHE)
 543		iflags |= XFS_IDONTCACHE;
 544	ip->i_udquot = NULL;
 545	ip->i_gdquot = NULL;
 546	ip->i_pdquot = NULL;
 547	xfs_iflags_set(ip, iflags);
 548
 549	/* insert the new inode */
 550	spin_lock(&pag->pag_ici_lock);
 551	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
 552	if (unlikely(error)) {
 553		WARN_ON(error != -EEXIST);
 554		XFS_STATS_INC(mp, xs_ig_dup);
 555		error = -EAGAIN;
 556		goto out_preload_end;
 557	}
 558	spin_unlock(&pag->pag_ici_lock);
 559	radix_tree_preload_end();
 560
 561	*ipp = ip;
 562	return 0;
 563
 564out_preload_end:
 565	spin_unlock(&pag->pag_ici_lock);
 566	radix_tree_preload_end();
 567	if (lock_flags)
 568		xfs_iunlock(ip, lock_flags);
 569out_destroy:
 570	__destroy_inode(VFS_I(ip));
 571	xfs_inode_free(ip);
 572	return error;
 573}
 574
 575/*
 576 * Look up an inode by number in the given file system.
 577 * The inode is looked up in the cache held in each AG.
 578 * If the inode is found in the cache, initialise the vfs inode
 579 * if necessary.
 580 *
 581 * If it is not in core, read it in from the file system's device,
 582 * add it to the cache and initialise the vfs inode.
 583 *
 584 * The inode is locked according to the value of the lock_flags parameter.
 585 * This flag parameter indicates how and if the inode's IO lock and inode lock
 586 * should be taken.
 587 *
 588 * mp -- the mount point structure for the current file system.  It points
 589 *       to the inode hash table.
 590 * tp -- a pointer to the current transaction if there is one.  This is
 591 *       simply passed through to the xfs_iread() call.
 592 * ino -- the number of the inode desired.  This is the unique identifier
 593 *        within the file system for the inode being requested.
 594 * lock_flags -- flags indicating how to lock the inode.  See the comment
 595 *		 for xfs_ilock() for a list of valid values.
 596 */
 597int
 598xfs_iget(
 599	xfs_mount_t	*mp,
 600	xfs_trans_t	*tp,
 601	xfs_ino_t	ino,
 602	uint		flags,
 603	uint		lock_flags,
 604	xfs_inode_t	**ipp)
 605{
 606	xfs_inode_t	*ip;
 607	int		error;
 608	xfs_perag_t	*pag;
 609	xfs_agino_t	agino;
 610
 611	/*
 612	 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
 613	 * doesn't get freed while it's being referenced during a
 614	 * radix tree traversal here.  It assumes this function
 615	 * aqcuires only the ILOCK (and therefore it has no need to
 616	 * involve the IOLOCK in this synchronization).
 617	 */
 618	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
 619
 620	/* reject inode numbers outside existing AGs */
 621	if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
 622		return -EINVAL;
 623
 624	XFS_STATS_INC(mp, xs_ig_attempts);
 625
 626	/* get the perag structure and ensure that it's inode capable */
 627	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
 628	agino = XFS_INO_TO_AGINO(mp, ino);
 629
 630again:
 631	error = 0;
 632	rcu_read_lock();
 633	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
 634
 635	if (ip) {
 636		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
 637		if (error)
 638			goto out_error_or_again;
 639	} else {
 640		rcu_read_unlock();
 641		if (flags & XFS_IGET_INCORE) {
 642			error = -ENODATA;
 643			goto out_error_or_again;
 644		}
 645		XFS_STATS_INC(mp, xs_ig_missed);
 646
 647		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
 648							flags, lock_flags);
 649		if (error)
 650			goto out_error_or_again;
 651	}
 652	xfs_perag_put(pag);
 653
 654	*ipp = ip;
 655
 656	/*
 657	 * If we have a real type for an on-disk inode, we can setup the inode
 658	 * now.	 If it's a new inode being created, xfs_ialloc will handle it.
 659	 */
 660	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
 661		xfs_setup_existing_inode(ip);
 662	return 0;
 663
 664out_error_or_again:
 665	if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
 666		delay(1);
 667		goto again;
 668	}
 669	xfs_perag_put(pag);
 670	return error;
 671}
 672
 673/*
 674 * "Is this a cached inode that's also allocated?"
 675 *
 676 * Look up an inode by number in the given file system.  If the inode is
 677 * in cache and isn't in purgatory, return 1 if the inode is allocated
 678 * and 0 if it is not.  For all other cases (not in cache, being torn
 679 * down, etc.), return a negative error code.
 680 *
 681 * The caller has to prevent inode allocation and freeing activity,
 682 * presumably by locking the AGI buffer.   This is to ensure that an
 683 * inode cannot transition from allocated to freed until the caller is
 684 * ready to allow that.  If the inode is in an intermediate state (new,
 685 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
 686 * inode is not in the cache, -ENOENT will be returned.  The caller must
 687 * deal with these scenarios appropriately.
 688 *
 689 * This is a specialized use case for the online scrubber; if you're
 690 * reading this, you probably want xfs_iget.
 691 */
 692int
 693xfs_icache_inode_is_allocated(
 694	struct xfs_mount	*mp,
 695	struct xfs_trans	*tp,
 696	xfs_ino_t		ino,
 697	bool			*inuse)
 698{
 699	struct xfs_inode	*ip;
 700	int			error;
 701
 702	error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
 703	if (error)
 704		return error;
 705
 706	*inuse = !!(VFS_I(ip)->i_mode);
 707	IRELE(ip);
 708	return 0;
 709}
 710
 711/*
 712 * The inode lookup is done in batches to keep the amount of lock traffic and
 713 * radix tree lookups to a minimum. The batch size is a trade off between
 714 * lookup reduction and stack usage. This is in the reclaim path, so we can't
 715 * be too greedy.
 716 */
 717#define XFS_LOOKUP_BATCH	32
 718
 719STATIC int
 720xfs_inode_ag_walk_grab(
 721	struct xfs_inode	*ip,
 722	int			flags)
 723{
 724	struct inode		*inode = VFS_I(ip);
 725	bool			newinos = !!(flags & XFS_AGITER_INEW_WAIT);
 726
 727	ASSERT(rcu_read_lock_held());
 728
 729	/*
 730	 * check for stale RCU freed inode
 731	 *
 732	 * If the inode has been reallocated, it doesn't matter if it's not in
 733	 * the AG we are walking - we are walking for writeback, so if it
 734	 * passes all the "valid inode" checks and is dirty, then we'll write
 735	 * it back anyway.  If it has been reallocated and still being
 736	 * initialised, the XFS_INEW check below will catch it.
 737	 */
 738	spin_lock(&ip->i_flags_lock);
 739	if (!ip->i_ino)
 740		goto out_unlock_noent;
 741
 742	/* avoid new or reclaimable inodes. Leave for reclaim code to flush */
 743	if ((!newinos && __xfs_iflags_test(ip, XFS_INEW)) ||
 744	    __xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM))
 745		goto out_unlock_noent;
 746	spin_unlock(&ip->i_flags_lock);
 747
 748	/* nothing to sync during shutdown */
 749	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 750		return -EFSCORRUPTED;
 751
 752	/* If we can't grab the inode, it must on it's way to reclaim. */
 753	if (!igrab(inode))
 754		return -ENOENT;
 755
 756	/* inode is valid */
 757	return 0;
 758
 759out_unlock_noent:
 760	spin_unlock(&ip->i_flags_lock);
 761	return -ENOENT;
 762}
 763
 764STATIC int
 765xfs_inode_ag_walk(
 766	struct xfs_mount	*mp,
 767	struct xfs_perag	*pag,
 768	int			(*execute)(struct xfs_inode *ip, int flags,
 
 769					   void *args),
 770	int			flags,
 771	void			*args,
 772	int			tag,
 773	int			iter_flags)
 774{
 775	uint32_t		first_index;
 776	int			last_error = 0;
 777	int			skipped;
 778	int			done;
 779	int			nr_found;
 780
 781restart:
 782	done = 0;
 783	skipped = 0;
 784	first_index = 0;
 785	nr_found = 0;
 786	do {
 787		struct xfs_inode *batch[XFS_LOOKUP_BATCH];
 788		int		error = 0;
 789		int		i;
 790
 791		rcu_read_lock();
 792
 793		if (tag == -1)
 794			nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
 795					(void **)batch, first_index,
 796					XFS_LOOKUP_BATCH);
 797		else
 798			nr_found = radix_tree_gang_lookup_tag(
 799					&pag->pag_ici_root,
 800					(void **) batch, first_index,
 801					XFS_LOOKUP_BATCH, tag);
 802
 803		if (!nr_found) {
 804			rcu_read_unlock();
 805			break;
 806		}
 807
 808		/*
 809		 * Grab the inodes before we drop the lock. if we found
 810		 * nothing, nr == 0 and the loop will be skipped.
 811		 */
 812		for (i = 0; i < nr_found; i++) {
 813			struct xfs_inode *ip = batch[i];
 814
 815			if (done || xfs_inode_ag_walk_grab(ip, iter_flags))
 816				batch[i] = NULL;
 817
 818			/*
 819			 * Update the index for the next lookup. Catch
 820			 * overflows into the next AG range which can occur if
 821			 * we have inodes in the last block of the AG and we
 822			 * are currently pointing to the last inode.
 823			 *
 824			 * Because we may see inodes that are from the wrong AG
 825			 * due to RCU freeing and reallocation, only update the
 826			 * index if it lies in this AG. It was a race that lead
 827			 * us to see this inode, so another lookup from the
 828			 * same index will not find it again.
 829			 */
 830			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
 831				continue;
 832			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
 833			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
 834				done = 1;
 835		}
 836
 837		/* unlock now we've grabbed the inodes. */
 838		rcu_read_unlock();
 839
 840		for (i = 0; i < nr_found; i++) {
 841			if (!batch[i])
 842				continue;
 843			if ((iter_flags & XFS_AGITER_INEW_WAIT) &&
 844			    xfs_iflags_test(batch[i], XFS_INEW))
 845				xfs_inew_wait(batch[i]);
 846			error = execute(batch[i], flags, args);
 847			IRELE(batch[i]);
 848			if (error == -EAGAIN) {
 849				skipped++;
 850				continue;
 851			}
 852			if (error && last_error != -EFSCORRUPTED)
 853				last_error = error;
 854		}
 855
 856		/* bail out if the filesystem is corrupted.  */
 857		if (error == -EFSCORRUPTED)
 858			break;
 859
 860		cond_resched();
 861
 862	} while (nr_found && !done);
 863
 864	if (skipped) {
 865		delay(1);
 866		goto restart;
 867	}
 868	return last_error;
 869}
 870
 871/*
 872 * Background scanning to trim post-EOF preallocated space. This is queued
 873 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
 874 */
 875void
 876xfs_queue_eofblocks(
 877	struct xfs_mount *mp)
 878{
 879	rcu_read_lock();
 880	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
 881		queue_delayed_work(mp->m_eofblocks_workqueue,
 882				   &mp->m_eofblocks_work,
 883				   msecs_to_jiffies(xfs_eofb_secs * 1000));
 884	rcu_read_unlock();
 885}
 886
 887void
 888xfs_eofblocks_worker(
 889	struct work_struct *work)
 890{
 891	struct xfs_mount *mp = container_of(to_delayed_work(work),
 892				struct xfs_mount, m_eofblocks_work);
 893	xfs_icache_free_eofblocks(mp, NULL);
 894	xfs_queue_eofblocks(mp);
 895}
 896
 897/*
 898 * Background scanning to trim preallocated CoW space. This is queued
 899 * based on the 'speculative_cow_prealloc_lifetime' tunable (5m by default).
 900 * (We'll just piggyback on the post-EOF prealloc space workqueue.)
 901 */
 902void
 903xfs_queue_cowblocks(
 904	struct xfs_mount *mp)
 905{
 906	rcu_read_lock();
 907	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_COWBLOCKS_TAG))
 908		queue_delayed_work(mp->m_eofblocks_workqueue,
 909				   &mp->m_cowblocks_work,
 910				   msecs_to_jiffies(xfs_cowb_secs * 1000));
 911	rcu_read_unlock();
 912}
 913
 914void
 915xfs_cowblocks_worker(
 916	struct work_struct *work)
 917{
 918	struct xfs_mount *mp = container_of(to_delayed_work(work),
 919				struct xfs_mount, m_cowblocks_work);
 920	xfs_icache_free_cowblocks(mp, NULL);
 921	xfs_queue_cowblocks(mp);
 922}
 923
 924int
 925xfs_inode_ag_iterator_flags(
 926	struct xfs_mount	*mp,
 927	int			(*execute)(struct xfs_inode *ip, int flags,
 
 928					   void *args),
 929	int			flags,
 930	void			*args,
 931	int			iter_flags)
 932{
 933	struct xfs_perag	*pag;
 934	int			error = 0;
 935	int			last_error = 0;
 936	xfs_agnumber_t		ag;
 937
 938	ag = 0;
 939	while ((pag = xfs_perag_get(mp, ag))) {
 940		ag = pag->pag_agno + 1;
 941		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1,
 942					  iter_flags);
 943		xfs_perag_put(pag);
 944		if (error) {
 945			last_error = error;
 946			if (error == -EFSCORRUPTED)
 947				break;
 948		}
 949	}
 950	return last_error;
 951}
 952
 953int
 954xfs_inode_ag_iterator(
 955	struct xfs_mount	*mp,
 956	int			(*execute)(struct xfs_inode *ip, int flags,
 957					   void *args),
 958	int			flags,
 959	void			*args)
 960{
 961	return xfs_inode_ag_iterator_flags(mp, execute, flags, args, 0);
 962}
 963
 964int
 965xfs_inode_ag_iterator_tag(
 966	struct xfs_mount	*mp,
 967	int			(*execute)(struct xfs_inode *ip, int flags,
 
 968					   void *args),
 969	int			flags,
 970	void			*args,
 971	int			tag)
 972{
 973	struct xfs_perag	*pag;
 974	int			error = 0;
 975	int			last_error = 0;
 976	xfs_agnumber_t		ag;
 977
 978	ag = 0;
 979	while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
 980		ag = pag->pag_agno + 1;
 981		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag,
 982					  0);
 983		xfs_perag_put(pag);
 984		if (error) {
 985			last_error = error;
 986			if (error == -EFSCORRUPTED)
 987				break;
 988		}
 989	}
 990	return last_error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 991}
 992
 993/*
 994 * Grab the inode for reclaim exclusively.
 995 * Return 0 if we grabbed it, non-zero otherwise.
 996 */
 997STATIC int
 998xfs_reclaim_inode_grab(
 999	struct xfs_inode	*ip,
1000	int			flags)
1001{
1002	ASSERT(rcu_read_lock_held());
1003
1004	/* quick check for stale RCU freed inode */
1005	if (!ip->i_ino)
1006		return 1;
1007
1008	/*
1009	 * If we are asked for non-blocking operation, do unlocked checks to
1010	 * see if the inode already is being flushed or in reclaim to avoid
1011	 * lock traffic.
1012	 */
1013	if ((flags & SYNC_TRYLOCK) &&
1014	    __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
1015		return 1;
1016
1017	/*
1018	 * The radix tree lock here protects a thread in xfs_iget from racing
1019	 * with us starting reclaim on the inode.  Once we have the
1020	 * XFS_IRECLAIM flag set it will not touch us.
1021	 *
1022	 * Due to RCU lookup, we may find inodes that have been freed and only
1023	 * have XFS_IRECLAIM set.  Indeed, we may see reallocated inodes that
1024	 * aren't candidates for reclaim at all, so we must check the
1025	 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
1026	 */
1027	spin_lock(&ip->i_flags_lock);
1028	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
1029	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {
1030		/* not a reclaim candidate. */
1031		spin_unlock(&ip->i_flags_lock);
1032		return 1;
1033	}
1034	__xfs_iflags_set(ip, XFS_IRECLAIM);
1035	spin_unlock(&ip->i_flags_lock);
1036	return 0;
1037}
1038
1039/*
1040 * Inodes in different states need to be treated differently. The following
1041 * table lists the inode states and the reclaim actions necessary:
1042 *
1043 *	inode state	     iflush ret		required action
1044 *      ---------------      ----------         ---------------
1045 *	bad			-		reclaim
1046 *	shutdown		EIO		unpin and reclaim
1047 *	clean, unpinned		0		reclaim
1048 *	stale, unpinned		0		reclaim
1049 *	clean, pinned(*)	0		requeue
1050 *	stale, pinned		EAGAIN		requeue
1051 *	dirty, async		-		requeue
1052 *	dirty, sync		0		reclaim
1053 *
1054 * (*) dgc: I don't think the clean, pinned state is possible but it gets
1055 * handled anyway given the order of checks implemented.
1056 *
1057 * Also, because we get the flush lock first, we know that any inode that has
1058 * been flushed delwri has had the flush completed by the time we check that
1059 * the inode is clean.
1060 *
1061 * Note that because the inode is flushed delayed write by AIL pushing, the
1062 * flush lock may already be held here and waiting on it can result in very
1063 * long latencies.  Hence for sync reclaims, where we wait on the flush lock,
1064 * the caller should push the AIL first before trying to reclaim inodes to
1065 * minimise the amount of time spent waiting.  For background relaim, we only
1066 * bother to reclaim clean inodes anyway.
1067 *
1068 * Hence the order of actions after gaining the locks should be:
1069 *	bad		=> reclaim
1070 *	shutdown	=> unpin and reclaim
1071 *	pinned, async	=> requeue
1072 *	pinned, sync	=> unpin
1073 *	stale		=> reclaim
1074 *	clean		=> reclaim
1075 *	dirty, async	=> requeue
1076 *	dirty, sync	=> flush, wait and reclaim
1077 */
1078STATIC int
1079xfs_reclaim_inode(
1080	struct xfs_inode	*ip,
1081	struct xfs_perag	*pag,
1082	int			sync_mode)
1083{
1084	struct xfs_buf		*bp = NULL;
1085	xfs_ino_t		ino = ip->i_ino; /* for radix_tree_delete */
1086	int			error;
1087
1088restart:
1089	error = 0;
1090	xfs_ilock(ip, XFS_ILOCK_EXCL);
1091	if (!xfs_iflock_nowait(ip)) {
1092		if (!(sync_mode & SYNC_WAIT))
1093			goto out;
1094		xfs_iflock(ip);
1095	}
1096
1097	if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1098		xfs_iunpin_wait(ip);
1099		/* xfs_iflush_abort() drops the flush lock */
1100		xfs_iflush_abort(ip, false);
1101		goto reclaim;
1102	}
1103	if (xfs_ipincount(ip)) {
1104		if (!(sync_mode & SYNC_WAIT))
1105			goto out_ifunlock;
1106		xfs_iunpin_wait(ip);
1107	}
1108	if (xfs_iflags_test(ip, XFS_ISTALE) || xfs_inode_clean(ip)) {
1109		xfs_ifunlock(ip);
 
1110		goto reclaim;
1111	}
1112
1113	/*
1114	 * Never flush out dirty data during non-blocking reclaim, as it would
1115	 * just contend with AIL pushing trying to do the same job.
1116	 */
1117	if (!(sync_mode & SYNC_WAIT))
1118		goto out_ifunlock;
1119
1120	/*
1121	 * Now we have an inode that needs flushing.
1122	 *
1123	 * Note that xfs_iflush will never block on the inode buffer lock, as
1124	 * xfs_ifree_cluster() can lock the inode buffer before it locks the
1125	 * ip->i_lock, and we are doing the exact opposite here.  As a result,
1126	 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
1127	 * result in an ABBA deadlock with xfs_ifree_cluster().
1128	 *
1129	 * As xfs_ifree_cluser() must gather all inodes that are active in the
1130	 * cache to mark them stale, if we hit this case we don't actually want
1131	 * to do IO here - we want the inode marked stale so we can simply
1132	 * reclaim it.  Hence if we get an EAGAIN error here,  just unlock the
1133	 * inode, back off and try again.  Hopefully the next pass through will
1134	 * see the stale flag set on the inode.
1135	 */
1136	error = xfs_iflush(ip, &bp);
1137	if (error == -EAGAIN) {
1138		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1139		/* backoff longer than in xfs_ifree_cluster */
1140		delay(2);
1141		goto restart;
1142	}
1143
1144	if (!error) {
1145		error = xfs_bwrite(bp);
1146		xfs_buf_relse(bp);
1147	}
1148
 
1149reclaim:
1150	ASSERT(!xfs_isiflocked(ip));
1151
1152	/*
1153	 * Because we use RCU freeing we need to ensure the inode always appears
1154	 * to be reclaimed with an invalid inode number when in the free state.
1155	 * We do this as early as possible under the ILOCK so that
1156	 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
1157	 * detect races with us here. By doing this, we guarantee that once
1158	 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
1159	 * it will see either a valid inode that will serialise correctly, or it
1160	 * will see an invalid inode that it can skip.
1161	 */
1162	spin_lock(&ip->i_flags_lock);
1163	ip->i_flags = XFS_IRECLAIM;
1164	ip->i_ino = 0;
1165	spin_unlock(&ip->i_flags_lock);
1166
1167	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1168
1169	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1170	/*
1171	 * Remove the inode from the per-AG radix tree.
1172	 *
1173	 * Because radix_tree_delete won't complain even if the item was never
1174	 * added to the tree assert that it's been there before to catch
1175	 * problems with the inode life time early on.
1176	 */
1177	spin_lock(&pag->pag_ici_lock);
1178	if (!radix_tree_delete(&pag->pag_ici_root,
1179				XFS_INO_TO_AGINO(ip->i_mount, ino)))
1180		ASSERT(0);
1181	xfs_perag_clear_reclaim_tag(pag);
1182	spin_unlock(&pag->pag_ici_lock);
1183
1184	/*
1185	 * Here we do an (almost) spurious inode lock in order to coordinate
1186	 * with inode cache radix tree lookups.  This is because the lookup
1187	 * can reference the inodes in the cache without taking references.
1188	 *
1189	 * We make that OK here by ensuring that we wait until the inode is
1190	 * unlocked after the lookup before we go ahead and free it.
1191	 */
1192	xfs_ilock(ip, XFS_ILOCK_EXCL);
1193	xfs_qm_dqdetach(ip);
1194	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1195
1196	__xfs_inode_free(ip);
1197	return error;
1198
1199out_ifunlock:
1200	xfs_ifunlock(ip);
1201out:
1202	xfs_iflags_clear(ip, XFS_IRECLAIM);
1203	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1204	/*
1205	 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1206	 * a short while. However, this just burns CPU time scanning the tree
1207	 * waiting for IO to complete and the reclaim work never goes back to
1208	 * the idle state. Instead, return 0 to let the next scheduled
1209	 * background reclaim attempt to reclaim the inode again.
1210	 */
1211	return 0;
1212}
1213
1214/*
1215 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1216 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1217 * then a shut down during filesystem unmount reclaim walk leak all the
1218 * unreclaimed inodes.
1219 */
1220STATIC int
1221xfs_reclaim_inodes_ag(
1222	struct xfs_mount	*mp,
1223	int			flags,
1224	int			*nr_to_scan)
1225{
1226	struct xfs_perag	*pag;
1227	int			error = 0;
1228	int			last_error = 0;
1229	xfs_agnumber_t		ag;
1230	int			trylock = flags & SYNC_TRYLOCK;
1231	int			skipped;
1232
1233restart:
1234	ag = 0;
1235	skipped = 0;
1236	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1237		unsigned long	first_index = 0;
1238		int		done = 0;
1239		int		nr_found = 0;
1240
1241		ag = pag->pag_agno + 1;
1242
1243		if (trylock) {
1244			if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1245				skipped++;
1246				xfs_perag_put(pag);
1247				continue;
1248			}
1249			first_index = pag->pag_ici_reclaim_cursor;
1250		} else
1251			mutex_lock(&pag->pag_ici_reclaim_lock);
1252
1253		do {
1254			struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1255			int	i;
1256
1257			rcu_read_lock();
1258			nr_found = radix_tree_gang_lookup_tag(
1259					&pag->pag_ici_root,
1260					(void **)batch, first_index,
1261					XFS_LOOKUP_BATCH,
1262					XFS_ICI_RECLAIM_TAG);
1263			if (!nr_found) {
1264				done = 1;
1265				rcu_read_unlock();
1266				break;
1267			}
1268
1269			/*
1270			 * Grab the inodes before we drop the lock. if we found
1271			 * nothing, nr == 0 and the loop will be skipped.
1272			 */
1273			for (i = 0; i < nr_found; i++) {
1274				struct xfs_inode *ip = batch[i];
1275
1276				if (done || xfs_reclaim_inode_grab(ip, flags))
1277					batch[i] = NULL;
1278
1279				/*
1280				 * Update the index for the next lookup. Catch
1281				 * overflows into the next AG range which can
1282				 * occur if we have inodes in the last block of
1283				 * the AG and we are currently pointing to the
1284				 * last inode.
1285				 *
1286				 * Because we may see inodes that are from the
1287				 * wrong AG due to RCU freeing and
1288				 * reallocation, only update the index if it
1289				 * lies in this AG. It was a race that lead us
1290				 * to see this inode, so another lookup from
1291				 * the same index will not find it again.
1292				 */
1293				if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1294								pag->pag_agno)
1295					continue;
1296				first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1297				if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1298					done = 1;
1299			}
1300
1301			/* unlock now we've grabbed the inodes. */
1302			rcu_read_unlock();
1303
1304			for (i = 0; i < nr_found; i++) {
1305				if (!batch[i])
1306					continue;
1307				error = xfs_reclaim_inode(batch[i], pag, flags);
1308				if (error && last_error != -EFSCORRUPTED)
1309					last_error = error;
1310			}
1311
1312			*nr_to_scan -= XFS_LOOKUP_BATCH;
1313
1314			cond_resched();
1315
1316		} while (nr_found && !done && *nr_to_scan > 0);
1317
1318		if (trylock && !done)
1319			pag->pag_ici_reclaim_cursor = first_index;
1320		else
1321			pag->pag_ici_reclaim_cursor = 0;
1322		mutex_unlock(&pag->pag_ici_reclaim_lock);
1323		xfs_perag_put(pag);
1324	}
1325
1326	/*
1327	 * if we skipped any AG, and we still have scan count remaining, do
1328	 * another pass this time using blocking reclaim semantics (i.e
1329	 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1330	 * ensure that when we get more reclaimers than AGs we block rather
1331	 * than spin trying to execute reclaim.
1332	 */
1333	if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1334		trylock = 0;
1335		goto restart;
1336	}
1337	return last_error;
1338}
1339
1340int
1341xfs_reclaim_inodes(
1342	xfs_mount_t	*mp,
1343	int		mode)
1344{
1345	int		nr_to_scan = INT_MAX;
1346
1347	return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1348}
1349
1350/*
1351 * Scan a certain number of inodes for reclaim.
1352 *
1353 * When called we make sure that there is a background (fast) inode reclaim in
1354 * progress, while we will throttle the speed of reclaim via doing synchronous
1355 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1356 * them to be cleaned, which we hope will not be very long due to the
1357 * background walker having already kicked the IO off on those dirty inodes.
1358 */
1359long
1360xfs_reclaim_inodes_nr(
1361	struct xfs_mount	*mp,
1362	int			nr_to_scan)
1363{
1364	/* kick background reclaimer and push the AIL */
1365	xfs_reclaim_work_queue(mp);
1366	xfs_ail_push_all(mp->m_ail);
1367
1368	return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1369}
1370
1371/*
1372 * Return the number of reclaimable inodes in the filesystem for
1373 * the shrinker to determine how much to reclaim.
1374 */
1375int
1376xfs_reclaim_inodes_count(
1377	struct xfs_mount	*mp)
1378{
1379	struct xfs_perag	*pag;
1380	xfs_agnumber_t		ag = 0;
1381	int			reclaimable = 0;
1382
1383	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1384		ag = pag->pag_agno + 1;
1385		reclaimable += pag->pag_ici_reclaimable;
1386		xfs_perag_put(pag);
1387	}
1388	return reclaimable;
1389}
1390
1391STATIC int
1392xfs_inode_match_id(
1393	struct xfs_inode	*ip,
1394	struct xfs_eofblocks	*eofb)
1395{
1396	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1397	    !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1398		return 0;
1399
1400	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1401	    !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1402		return 0;
1403
1404	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1405	    xfs_get_projid(ip) != eofb->eof_prid)
1406		return 0;
1407
1408	return 1;
1409}
1410
1411/*
1412 * A union-based inode filtering algorithm. Process the inode if any of the
1413 * criteria match. This is for global/internal scans only.
1414 */
1415STATIC int
1416xfs_inode_match_id_union(
1417	struct xfs_inode	*ip,
1418	struct xfs_eofblocks	*eofb)
1419{
1420	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1421	    uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1422		return 1;
1423
1424	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1425	    gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1426		return 1;
1427
1428	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1429	    xfs_get_projid(ip) == eofb->eof_prid)
1430		return 1;
1431
1432	return 0;
1433}
1434
1435STATIC int
1436xfs_inode_free_eofblocks(
1437	struct xfs_inode	*ip,
 
1438	int			flags,
1439	void			*args)
1440{
1441	int ret = 0;
1442	struct xfs_eofblocks *eofb = args;
1443	int match;
1444
1445	if (!xfs_can_free_eofblocks(ip, false)) {
1446		/* inode could be preallocated or append-only */
1447		trace_xfs_inode_free_eofblocks_invalid(ip);
1448		xfs_inode_clear_eofblocks_tag(ip);
1449		return 0;
1450	}
1451
1452	/*
1453	 * If the mapping is dirty the operation can block and wait for some
1454	 * time. Unless we are waiting, skip it.
1455	 */
1456	if (!(flags & SYNC_WAIT) &&
1457	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1458		return 0;
1459
1460	if (eofb) {
1461		if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1462			match = xfs_inode_match_id_union(ip, eofb);
1463		else
1464			match = xfs_inode_match_id(ip, eofb);
1465		if (!match)
1466			return 0;
1467
1468		/* skip the inode if the file size is too small */
1469		if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1470		    XFS_ISIZE(ip) < eofb->eof_min_file_size)
1471			return 0;
1472	}
1473
1474	/*
1475	 * If the caller is waiting, return -EAGAIN to keep the background
1476	 * scanner moving and revisit the inode in a subsequent pass.
1477	 */
1478	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1479		if (flags & SYNC_WAIT)
1480			ret = -EAGAIN;
1481		return ret;
1482	}
1483	ret = xfs_free_eofblocks(ip);
1484	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1485
1486	return ret;
1487}
1488
1489static int
1490__xfs_icache_free_eofblocks(
1491	struct xfs_mount	*mp,
1492	struct xfs_eofblocks	*eofb,
1493	int			(*execute)(struct xfs_inode *ip, int flags,
1494					   void *args),
1495	int			tag)
1496{
1497	int flags = SYNC_TRYLOCK;
1498
1499	if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1500		flags = SYNC_WAIT;
 
1501
1502	return xfs_inode_ag_iterator_tag(mp, execute, flags,
1503					 eofb, tag);
1504}
1505
1506int
1507xfs_icache_free_eofblocks(
1508	struct xfs_mount	*mp,
1509	struct xfs_eofblocks	*eofb)
1510{
1511	return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_eofblocks,
1512			XFS_ICI_EOFBLOCKS_TAG);
1513}
1514
1515/*
1516 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1517 * multiple quotas, we don't know exactly which quota caused an allocation
1518 * failure. We make a best effort by including each quota under low free space
1519 * conditions (less than 1% free space) in the scan.
1520 */
1521static int
1522__xfs_inode_free_quota_eofblocks(
1523	struct xfs_inode	*ip,
1524	int			(*execute)(struct xfs_mount *mp,
1525					   struct xfs_eofblocks	*eofb))
1526{
1527	int scan = 0;
1528	struct xfs_eofblocks eofb = {0};
1529	struct xfs_dquot *dq;
1530
1531	/*
1532	 * Run a sync scan to increase effectiveness and use the union filter to
1533	 * cover all applicable quotas in a single scan.
1534	 */
1535	eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1536
1537	if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1538		dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1539		if (dq && xfs_dquot_lowsp(dq)) {
1540			eofb.eof_uid = VFS_I(ip)->i_uid;
1541			eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1542			scan = 1;
1543		}
1544	}
1545
1546	if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1547		dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1548		if (dq && xfs_dquot_lowsp(dq)) {
1549			eofb.eof_gid = VFS_I(ip)->i_gid;
1550			eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1551			scan = 1;
1552		}
1553	}
1554
1555	if (scan)
1556		execute(ip->i_mount, &eofb);
1557
1558	return scan;
1559}
1560
1561int
1562xfs_inode_free_quota_eofblocks(
1563	struct xfs_inode *ip)
1564{
1565	return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_eofblocks);
1566}
1567
1568static inline unsigned long
1569xfs_iflag_for_tag(
1570	int		tag)
1571{
1572	switch (tag) {
1573	case XFS_ICI_EOFBLOCKS_TAG:
1574		return XFS_IEOFBLOCKS;
1575	case XFS_ICI_COWBLOCKS_TAG:
1576		return XFS_ICOWBLOCKS;
1577	default:
1578		ASSERT(0);
1579		return 0;
1580	}
1581}
1582
1583static void
1584__xfs_inode_set_blocks_tag(
1585	xfs_inode_t	*ip,
1586	void		(*execute)(struct xfs_mount *mp),
1587	void		(*set_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1588				  int error, unsigned long caller_ip),
1589	int		tag)
1590{
1591	struct xfs_mount *mp = ip->i_mount;
1592	struct xfs_perag *pag;
1593	int tagged;
1594
1595	/*
1596	 * Don't bother locking the AG and looking up in the radix trees
1597	 * if we already know that we have the tag set.
1598	 */
1599	if (ip->i_flags & xfs_iflag_for_tag(tag))
1600		return;
1601	spin_lock(&ip->i_flags_lock);
1602	ip->i_flags |= xfs_iflag_for_tag(tag);
1603	spin_unlock(&ip->i_flags_lock);
1604
1605	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1606	spin_lock(&pag->pag_ici_lock);
 
1607
1608	tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
 
1609	radix_tree_tag_set(&pag->pag_ici_root,
1610			   XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
 
1611	if (!tagged) {
1612		/* propagate the eofblocks tag up into the perag radix tree */
1613		spin_lock(&ip->i_mount->m_perag_lock);
1614		radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1615				   XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1616				   tag);
1617		spin_unlock(&ip->i_mount->m_perag_lock);
1618
1619		/* kick off background trimming */
1620		execute(ip->i_mount);
1621
1622		set_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
 
1623	}
1624
1625	spin_unlock(&pag->pag_ici_lock);
1626	xfs_perag_put(pag);
1627}
1628
1629void
1630xfs_inode_set_eofblocks_tag(
1631	xfs_inode_t	*ip)
1632{
1633	trace_xfs_inode_set_eofblocks_tag(ip);
1634	return __xfs_inode_set_blocks_tag(ip, xfs_queue_eofblocks,
1635			trace_xfs_perag_set_eofblocks,
1636			XFS_ICI_EOFBLOCKS_TAG);
1637}
1638
1639static void
1640__xfs_inode_clear_blocks_tag(
1641	xfs_inode_t	*ip,
1642	void		(*clear_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1643				    int error, unsigned long caller_ip),
1644	int		tag)
1645{
1646	struct xfs_mount *mp = ip->i_mount;
1647	struct xfs_perag *pag;
1648
1649	spin_lock(&ip->i_flags_lock);
1650	ip->i_flags &= ~xfs_iflag_for_tag(tag);
1651	spin_unlock(&ip->i_flags_lock);
1652
1653	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1654	spin_lock(&pag->pag_ici_lock);
 
1655
1656	radix_tree_tag_clear(&pag->pag_ici_root,
1657			     XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1658	if (!radix_tree_tagged(&pag->pag_ici_root, tag)) {
 
1659		/* clear the eofblocks tag from the perag radix tree */
1660		spin_lock(&ip->i_mount->m_perag_lock);
1661		radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1662				     XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1663				     tag);
1664		spin_unlock(&ip->i_mount->m_perag_lock);
1665		clear_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
 
1666	}
1667
1668	spin_unlock(&pag->pag_ici_lock);
1669	xfs_perag_put(pag);
1670}
1671
1672void
1673xfs_inode_clear_eofblocks_tag(
1674	xfs_inode_t	*ip)
1675{
1676	trace_xfs_inode_clear_eofblocks_tag(ip);
1677	return __xfs_inode_clear_blocks_tag(ip,
1678			trace_xfs_perag_clear_eofblocks, XFS_ICI_EOFBLOCKS_TAG);
1679}
1680
1681/*
1682 * Set ourselves up to free CoW blocks from this file.  If it's already clean
1683 * then we can bail out quickly, but otherwise we must back off if the file
1684 * is undergoing some kind of write.
1685 */
1686static bool
1687xfs_prep_free_cowblocks(
1688	struct xfs_inode	*ip,
1689	struct xfs_ifork	*ifp)
1690{
1691	/*
1692	 * Just clear the tag if we have an empty cow fork or none at all. It's
1693	 * possible the inode was fully unshared since it was originally tagged.
1694	 */
1695	if (!xfs_is_reflink_inode(ip) || !ifp->if_bytes) {
1696		trace_xfs_inode_free_cowblocks_invalid(ip);
1697		xfs_inode_clear_cowblocks_tag(ip);
1698		return false;
1699	}
1700
1701	/*
1702	 * If the mapping is dirty or under writeback we cannot touch the
1703	 * CoW fork.  Leave it alone if we're in the midst of a directio.
1704	 */
1705	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1706	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1707	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1708	    atomic_read(&VFS_I(ip)->i_dio_count))
1709		return false;
1710
1711	return true;
1712}
1713
1714/*
1715 * Automatic CoW Reservation Freeing
1716 *
1717 * These functions automatically garbage collect leftover CoW reservations
1718 * that were made on behalf of a cowextsize hint when we start to run out
1719 * of quota or when the reservations sit around for too long.  If the file
1720 * has dirty pages or is undergoing writeback, its CoW reservations will
1721 * be retained.
1722 *
1723 * The actual garbage collection piggybacks off the same code that runs
1724 * the speculative EOF preallocation garbage collector.
1725 */
1726STATIC int
1727xfs_inode_free_cowblocks(
1728	struct xfs_inode	*ip,
1729	int			flags,
1730	void			*args)
1731{
1732	struct xfs_eofblocks	*eofb = args;
1733	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1734	int			match;
1735	int			ret = 0;
1736
1737	if (!xfs_prep_free_cowblocks(ip, ifp))
1738		return 0;
1739
1740	if (eofb) {
1741		if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1742			match = xfs_inode_match_id_union(ip, eofb);
1743		else
1744			match = xfs_inode_match_id(ip, eofb);
1745		if (!match)
1746			return 0;
1747
1748		/* skip the inode if the file size is too small */
1749		if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1750		    XFS_ISIZE(ip) < eofb->eof_min_file_size)
1751			return 0;
1752	}
1753
1754	/* Free the CoW blocks */
1755	xfs_ilock(ip, XFS_IOLOCK_EXCL);
1756	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1757
1758	/*
1759	 * Check again, nobody else should be able to dirty blocks or change
1760	 * the reflink iflag now that we have the first two locks held.
1761	 */
1762	if (xfs_prep_free_cowblocks(ip, ifp))
1763		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1764
1765	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1766	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1767
1768	return ret;
1769}
1770
1771int
1772xfs_icache_free_cowblocks(
1773	struct xfs_mount	*mp,
1774	struct xfs_eofblocks	*eofb)
1775{
1776	return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_cowblocks,
1777			XFS_ICI_COWBLOCKS_TAG);
1778}
1779
1780int
1781xfs_inode_free_quota_cowblocks(
1782	struct xfs_inode *ip)
1783{
1784	return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_cowblocks);
1785}
1786
1787void
1788xfs_inode_set_cowblocks_tag(
1789	xfs_inode_t	*ip)
1790{
1791	trace_xfs_inode_set_cowblocks_tag(ip);
1792	return __xfs_inode_set_blocks_tag(ip, xfs_queue_cowblocks,
1793			trace_xfs_perag_set_cowblocks,
1794			XFS_ICI_COWBLOCKS_TAG);
1795}
1796
1797void
1798xfs_inode_clear_cowblocks_tag(
1799	xfs_inode_t	*ip)
1800{
1801	trace_xfs_inode_clear_cowblocks_tag(ip);
1802	return __xfs_inode_clear_blocks_tag(ip,
1803			trace_xfs_perag_clear_cowblocks, XFS_ICI_COWBLOCKS_TAG);
1804}