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
 
 
 
  41/*
  42 * Allocate and initialise an xfs_inode.
  43 */
  44struct xfs_inode *
  45xfs_inode_alloc(
  46	struct xfs_mount	*mp,
  47	xfs_ino_t		ino)
  48{
  49	struct xfs_inode	*ip;
  50
  51	/*
  52	 * if this didn't occur in transactions, we could use
  53	 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
  54	 * code up to do this anyway.
  55	 */
  56	ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
  57	if (!ip)
  58		return NULL;
  59	if (inode_init_always(mp->m_super, VFS_I(ip))) {
  60		kmem_zone_free(xfs_inode_zone, ip);
  61		return NULL;
  62	}
  63
  64	/* VFS doesn't initialise i_mode! */
  65	VFS_I(ip)->i_mode = 0;
  66
  67	XFS_STATS_INC(mp, vn_active);
  68	ASSERT(atomic_read(&ip->i_pincount) == 0);
  69	ASSERT(!spin_is_locked(&ip->i_flags_lock));
  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	ASSERT(spin_is_locked(&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	ASSERT(spin_is_locked(&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
 265/*
 266 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
 267 * part of the structure. This is made more complex by the fact we store
 268 * information about the on-disk values in the VFS inode and so we can't just
 269 * overwrite the values unconditionally. Hence we save the parameters we
 270 * need to retain across reinitialisation, and rewrite them into the VFS inode
 271 * after reinitialisation even if it fails.
 272 */
 273static int
 274xfs_reinit_inode(
 275	struct xfs_mount	*mp,
 276	struct inode		*inode)
 277{
 278	int		error;
 279	uint32_t	nlink = inode->i_nlink;
 280	uint32_t	generation = inode->i_generation;
 281	uint64_t	version = inode->i_version;
 282	umode_t		mode = inode->i_mode;
 283
 284	error = inode_init_always(mp->m_super, inode);
 285
 286	set_nlink(inode, nlink);
 287	inode->i_generation = generation;
 288	inode->i_version = version;
 289	inode->i_mode = mode;
 290	return error;
 291}
 292
 293/*
 294 * Check the validity of the inode we just found it the cache
 295 */
 296static int
 297xfs_iget_cache_hit(
 298	struct xfs_perag	*pag,
 299	struct xfs_inode	*ip,
 300	xfs_ino_t		ino,
 301	int			flags,
 302	int			lock_flags) __releases(RCU)
 303{
 304	struct inode		*inode = VFS_I(ip);
 305	struct xfs_mount	*mp = ip->i_mount;
 306	int			error;
 307
 308	/*
 309	 * check for re-use of an inode within an RCU grace period due to the
 310	 * radix tree nodes not being updated yet. We monitor for this by
 311	 * setting the inode number to zero before freeing the inode structure.
 312	 * If the inode has been reallocated and set up, then the inode number
 313	 * will not match, so check for that, too.
 314	 */
 315	spin_lock(&ip->i_flags_lock);
 316	if (ip->i_ino != ino) {
 317		trace_xfs_iget_skip(ip);
 318		XFS_STATS_INC(mp, xs_ig_frecycle);
 319		error = -EAGAIN;
 320		goto out_error;
 321	}
 322
 323
 324	/*
 325	 * If we are racing with another cache hit that is currently
 326	 * instantiating this inode or currently recycling it out of
 327	 * reclaimabe state, wait for the initialisation to complete
 328	 * before continuing.
 329	 *
 330	 * XXX(hch): eventually we should do something equivalent to
 331	 *	     wait_on_inode to wait for these flags to be cleared
 332	 *	     instead of polling for it.
 333	 */
 334	if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
 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	 * If lookup is racing with unlink return an error immediately.
 343	 */
 344	if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
 345		error = -ENOENT;
 346		goto out_error;
 347	}
 348
 349	/*
 350	 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
 351	 * Need to carefully get it back into useable state.
 352	 */
 353	if (ip->i_flags & XFS_IRECLAIMABLE) {
 354		trace_xfs_iget_reclaim(ip);
 355
 356		/*
 357		 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
 358		 * from stomping over us while we recycle the inode.  We can't
 359		 * clear the radix tree reclaimable tag yet as it requires
 360		 * pag_ici_lock to be held exclusive.
 361		 */
 362		ip->i_flags |= XFS_IRECLAIM;
 363
 364		spin_unlock(&ip->i_flags_lock);
 365		rcu_read_unlock();
 366
 367		error = xfs_reinit_inode(mp, inode);
 368		if (error) {
 369			/*
 370			 * Re-initializing the inode failed, and we are in deep
 371			 * trouble.  Try to re-add it to the reclaim list.
 372			 */
 373			rcu_read_lock();
 374			spin_lock(&ip->i_flags_lock);
 375
 376			ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
 377			ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
 378			trace_xfs_iget_reclaim_fail(ip);
 379			goto out_error;
 380		}
 381
 382		spin_lock(&pag->pag_ici_lock);
 383		spin_lock(&ip->i_flags_lock);
 384
 385		/*
 386		 * Clear the per-lifetime state in the inode as we are now
 387		 * effectively a new inode and need to return to the initial
 388		 * state before reuse occurs.
 389		 */
 390		ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
 391		ip->i_flags |= XFS_INEW;
 392		xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
 393		inode->i_state = I_NEW;
 394
 395		ASSERT(!rwsem_is_locked(&inode->i_rwsem));
 396		init_rwsem(&inode->i_rwsem);
 397
 398		spin_unlock(&ip->i_flags_lock);
 399		spin_unlock(&pag->pag_ici_lock);
 400	} else {
 401		/* If the VFS inode is being torn down, pause and try again. */
 402		if (!igrab(inode)) {
 403			trace_xfs_iget_skip(ip);
 404			error = -EAGAIN;
 405			goto out_error;
 406		}
 407
 408		/* We've got a live one. */
 409		spin_unlock(&ip->i_flags_lock);
 410		rcu_read_unlock();
 411		trace_xfs_iget_hit(ip);
 412	}
 413
 414	if (lock_flags != 0)
 415		xfs_ilock(ip, lock_flags);
 416
 417	xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
 418	XFS_STATS_INC(mp, xs_ig_found);
 419
 420	return 0;
 421
 422out_error:
 423	spin_unlock(&ip->i_flags_lock);
 424	rcu_read_unlock();
 425	return error;
 426}
 427
 428
 429static int
 430xfs_iget_cache_miss(
 431	struct xfs_mount	*mp,
 432	struct xfs_perag	*pag,
 433	xfs_trans_t		*tp,
 434	xfs_ino_t		ino,
 435	struct xfs_inode	**ipp,
 436	int			flags,
 437	int			lock_flags)
 438{
 439	struct xfs_inode	*ip;
 440	int			error;
 441	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);
 442	int			iflags;
 443
 444	ip = xfs_inode_alloc(mp, ino);
 445	if (!ip)
 446		return -ENOMEM;
 447
 448	error = xfs_iread(mp, tp, ip, flags);
 449	if (error)
 450		goto out_destroy;
 451
 452	trace_xfs_iget_miss(ip);
 453
 454	if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
 455		error = -ENOENT;
 456		goto out_destroy;
 457	}
 458
 459	/*
 460	 * Preload the radix tree so we can insert safely under the
 461	 * write spinlock. Note that we cannot sleep inside the preload
 462	 * region. Since we can be called from transaction context, don't
 463	 * recurse into the file system.
 464	 */
 465	if (radix_tree_preload(GFP_NOFS)) {
 466		error = -EAGAIN;
 467		goto out_destroy;
 468	}
 469
 470	/*
 471	 * Because the inode hasn't been added to the radix-tree yet it can't
 472	 * be found by another thread, so we can do the non-sleeping lock here.
 473	 */
 474	if (lock_flags) {
 475		if (!xfs_ilock_nowait(ip, lock_flags))
 476			BUG();
 477	}
 478
 479	/*
 480	 * These values must be set before inserting the inode into the radix
 481	 * tree as the moment it is inserted a concurrent lookup (allowed by the
 482	 * RCU locking mechanism) can find it and that lookup must see that this
 483	 * is an inode currently under construction (i.e. that XFS_INEW is set).
 484	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
 485	 * memory barrier that ensures this detection works correctly at lookup
 486	 * time.
 487	 */
 488	iflags = XFS_INEW;
 489	if (flags & XFS_IGET_DONTCACHE)
 490		iflags |= XFS_IDONTCACHE;
 491	ip->i_udquot = NULL;
 492	ip->i_gdquot = NULL;
 493	ip->i_pdquot = NULL;
 494	xfs_iflags_set(ip, iflags);
 495
 496	/* insert the new inode */
 497	spin_lock(&pag->pag_ici_lock);
 498	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
 499	if (unlikely(error)) {
 500		WARN_ON(error != -EEXIST);
 501		XFS_STATS_INC(mp, xs_ig_dup);
 502		error = -EAGAIN;
 503		goto out_preload_end;
 504	}
 505	spin_unlock(&pag->pag_ici_lock);
 506	radix_tree_preload_end();
 507
 508	*ipp = ip;
 509	return 0;
 510
 511out_preload_end:
 512	spin_unlock(&pag->pag_ici_lock);
 513	radix_tree_preload_end();
 514	if (lock_flags)
 515		xfs_iunlock(ip, lock_flags);
 516out_destroy:
 517	__destroy_inode(VFS_I(ip));
 518	xfs_inode_free(ip);
 519	return error;
 520}
 521
 522/*
 523 * Look up an inode by number in the given file system.
 524 * The inode is looked up in the cache held in each AG.
 525 * If the inode is found in the cache, initialise the vfs inode
 526 * if necessary.
 527 *
 528 * If it is not in core, read it in from the file system's device,
 529 * add it to the cache and initialise the vfs inode.
 530 *
 531 * The inode is locked according to the value of the lock_flags parameter.
 532 * This flag parameter indicates how and if the inode's IO lock and inode lock
 533 * should be taken.
 534 *
 535 * mp -- the mount point structure for the current file system.  It points
 536 *       to the inode hash table.
 537 * tp -- a pointer to the current transaction if there is one.  This is
 538 *       simply passed through to the xfs_iread() call.
 539 * ino -- the number of the inode desired.  This is the unique identifier
 540 *        within the file system for the inode being requested.
 541 * lock_flags -- flags indicating how to lock the inode.  See the comment
 542 *		 for xfs_ilock() for a list of valid values.
 543 */
 544int
 545xfs_iget(
 546	xfs_mount_t	*mp,
 547	xfs_trans_t	*tp,
 548	xfs_ino_t	ino,
 549	uint		flags,
 550	uint		lock_flags,
 551	xfs_inode_t	**ipp)
 552{
 553	xfs_inode_t	*ip;
 554	int		error;
 555	xfs_perag_t	*pag;
 556	xfs_agino_t	agino;
 557
 558	/*
 559	 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
 560	 * doesn't get freed while it's being referenced during a
 561	 * radix tree traversal here.  It assumes this function
 562	 * aqcuires only the ILOCK (and therefore it has no need to
 563	 * involve the IOLOCK in this synchronization).
 564	 */
 565	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
 566
 567	/* reject inode numbers outside existing AGs */
 568	if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
 569		return -EINVAL;
 570
 571	XFS_STATS_INC(mp, xs_ig_attempts);
 572
 573	/* get the perag structure and ensure that it's inode capable */
 574	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
 575	agino = XFS_INO_TO_AGINO(mp, ino);
 576
 577again:
 578	error = 0;
 579	rcu_read_lock();
 580	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
 581
 582	if (ip) {
 583		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
 584		if (error)
 585			goto out_error_or_again;
 586	} else {
 587		rcu_read_unlock();
 588		XFS_STATS_INC(mp, xs_ig_missed);
 589
 590		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
 591							flags, lock_flags);
 592		if (error)
 593			goto out_error_or_again;
 594	}
 595	xfs_perag_put(pag);
 596
 597	*ipp = ip;
 598
 599	/*
 600	 * If we have a real type for an on-disk inode, we can setup the inode
 601	 * now.	 If it's a new inode being created, xfs_ialloc will handle it.
 602	 */
 603	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
 604		xfs_setup_existing_inode(ip);
 605	return 0;
 606
 607out_error_or_again:
 608	if (error == -EAGAIN) {
 609		delay(1);
 610		goto again;
 611	}
 612	xfs_perag_put(pag);
 613	return error;
 614}
 615
 616/*
 617 * The inode lookup is done in batches to keep the amount of lock traffic and
 618 * radix tree lookups to a minimum. The batch size is a trade off between
 619 * lookup reduction and stack usage. This is in the reclaim path, so we can't
 620 * be too greedy.
 621 */
 622#define XFS_LOOKUP_BATCH	32
 623
 624STATIC int
 625xfs_inode_ag_walk_grab(
 626	struct xfs_inode	*ip)
 627{
 628	struct inode		*inode = VFS_I(ip);
 629
 630	ASSERT(rcu_read_lock_held());
 631
 632	/*
 633	 * check for stale RCU freed inode
 634	 *
 635	 * If the inode has been reallocated, it doesn't matter if it's not in
 636	 * the AG we are walking - we are walking for writeback, so if it
 637	 * passes all the "valid inode" checks and is dirty, then we'll write
 638	 * it back anyway.  If it has been reallocated and still being
 639	 * initialised, the XFS_INEW check below will catch it.
 640	 */
 641	spin_lock(&ip->i_flags_lock);
 642	if (!ip->i_ino)
 643		goto out_unlock_noent;
 644
 645	/* avoid new or reclaimable inodes. Leave for reclaim code to flush */
 646	if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
 647		goto out_unlock_noent;
 648	spin_unlock(&ip->i_flags_lock);
 649
 650	/* nothing to sync during shutdown */
 651	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 652		return -EFSCORRUPTED;
 653
 654	/* If we can't grab the inode, it must on it's way to reclaim. */
 655	if (!igrab(inode))
 656		return -ENOENT;
 657
 658	/* inode is valid */
 659	return 0;
 660
 661out_unlock_noent:
 662	spin_unlock(&ip->i_flags_lock);
 663	return -ENOENT;
 664}
 665
 666STATIC int
 667xfs_inode_ag_walk(
 668	struct xfs_mount	*mp,
 669	struct xfs_perag	*pag,
 670	int			(*execute)(struct xfs_inode *ip, int flags,
 
 671					   void *args),
 672	int			flags,
 673	void			*args,
 674	int			tag)
 675{
 676	uint32_t		first_index;
 677	int			last_error = 0;
 678	int			skipped;
 679	int			done;
 680	int			nr_found;
 681
 682restart:
 683	done = 0;
 684	skipped = 0;
 685	first_index = 0;
 686	nr_found = 0;
 687	do {
 688		struct xfs_inode *batch[XFS_LOOKUP_BATCH];
 689		int		error = 0;
 690		int		i;
 691
 692		rcu_read_lock();
 693
 694		if (tag == -1)
 695			nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
 696					(void **)batch, first_index,
 697					XFS_LOOKUP_BATCH);
 698		else
 699			nr_found = radix_tree_gang_lookup_tag(
 700					&pag->pag_ici_root,
 701					(void **) batch, first_index,
 702					XFS_LOOKUP_BATCH, tag);
 703
 704		if (!nr_found) {
 705			rcu_read_unlock();
 706			break;
 707		}
 708
 709		/*
 710		 * Grab the inodes before we drop the lock. if we found
 711		 * nothing, nr == 0 and the loop will be skipped.
 712		 */
 713		for (i = 0; i < nr_found; i++) {
 714			struct xfs_inode *ip = batch[i];
 715
 716			if (done || xfs_inode_ag_walk_grab(ip))
 717				batch[i] = NULL;
 718
 719			/*
 720			 * Update the index for the next lookup. Catch
 721			 * overflows into the next AG range which can occur if
 722			 * we have inodes in the last block of the AG and we
 723			 * are currently pointing to the last inode.
 724			 *
 725			 * Because we may see inodes that are from the wrong AG
 726			 * due to RCU freeing and reallocation, only update the
 727			 * index if it lies in this AG. It was a race that lead
 728			 * us to see this inode, so another lookup from the
 729			 * same index will not find it again.
 730			 */
 731			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
 732				continue;
 733			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
 734			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
 735				done = 1;
 736		}
 737
 738		/* unlock now we've grabbed the inodes. */
 739		rcu_read_unlock();
 740
 741		for (i = 0; i < nr_found; i++) {
 742			if (!batch[i])
 743				continue;
 744			error = execute(batch[i], flags, args);
 745			IRELE(batch[i]);
 746			if (error == -EAGAIN) {
 747				skipped++;
 748				continue;
 749			}
 750			if (error && last_error != -EFSCORRUPTED)
 751				last_error = error;
 752		}
 753
 754		/* bail out if the filesystem is corrupted.  */
 755		if (error == -EFSCORRUPTED)
 756			break;
 757
 758		cond_resched();
 759
 760	} while (nr_found && !done);
 761
 762	if (skipped) {
 763		delay(1);
 764		goto restart;
 765	}
 766	return last_error;
 767}
 768
 769/*
 770 * Background scanning to trim post-EOF preallocated space. This is queued
 771 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
 772 */
 773void
 774xfs_queue_eofblocks(
 775	struct xfs_mount *mp)
 776{
 777	rcu_read_lock();
 778	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
 779		queue_delayed_work(mp->m_eofblocks_workqueue,
 780				   &mp->m_eofblocks_work,
 781				   msecs_to_jiffies(xfs_eofb_secs * 1000));
 782	rcu_read_unlock();
 783}
 784
 785void
 786xfs_eofblocks_worker(
 787	struct work_struct *work)
 788{
 789	struct xfs_mount *mp = container_of(to_delayed_work(work),
 790				struct xfs_mount, m_eofblocks_work);
 791	xfs_icache_free_eofblocks(mp, NULL);
 792	xfs_queue_eofblocks(mp);
 793}
 794
 795/*
 796 * Background scanning to trim preallocated CoW space. This is queued
 797 * based on the 'speculative_cow_prealloc_lifetime' tunable (5m by default).
 798 * (We'll just piggyback on the post-EOF prealloc space workqueue.)
 799 */
 800STATIC void
 801xfs_queue_cowblocks(
 802	struct xfs_mount *mp)
 803{
 804	rcu_read_lock();
 805	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_COWBLOCKS_TAG))
 806		queue_delayed_work(mp->m_eofblocks_workqueue,
 807				   &mp->m_cowblocks_work,
 808				   msecs_to_jiffies(xfs_cowb_secs * 1000));
 809	rcu_read_unlock();
 810}
 811
 812void
 813xfs_cowblocks_worker(
 814	struct work_struct *work)
 815{
 816	struct xfs_mount *mp = container_of(to_delayed_work(work),
 817				struct xfs_mount, m_cowblocks_work);
 818	xfs_icache_free_cowblocks(mp, NULL);
 819	xfs_queue_cowblocks(mp);
 820}
 821
 822int
 823xfs_inode_ag_iterator(
 824	struct xfs_mount	*mp,
 825	int			(*execute)(struct xfs_inode *ip, int flags,
 
 826					   void *args),
 827	int			flags,
 828	void			*args)
 829{
 830	struct xfs_perag	*pag;
 831	int			error = 0;
 832	int			last_error = 0;
 833	xfs_agnumber_t		ag;
 834
 835	ag = 0;
 836	while ((pag = xfs_perag_get(mp, ag))) {
 837		ag = pag->pag_agno + 1;
 838		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1);
 839		xfs_perag_put(pag);
 840		if (error) {
 841			last_error = error;
 842			if (error == -EFSCORRUPTED)
 843				break;
 844		}
 845	}
 846	return last_error;
 847}
 848
 849int
 850xfs_inode_ag_iterator_tag(
 851	struct xfs_mount	*mp,
 852	int			(*execute)(struct xfs_inode *ip, int flags,
 
 853					   void *args),
 854	int			flags,
 855	void			*args,
 856	int			tag)
 857{
 858	struct xfs_perag	*pag;
 859	int			error = 0;
 860	int			last_error = 0;
 861	xfs_agnumber_t		ag;
 862
 863	ag = 0;
 864	while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
 865		ag = pag->pag_agno + 1;
 866		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag);
 867		xfs_perag_put(pag);
 868		if (error) {
 869			last_error = error;
 870			if (error == -EFSCORRUPTED)
 871				break;
 872		}
 873	}
 874	return last_error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 875}
 876
 877/*
 878 * Grab the inode for reclaim exclusively.
 879 * Return 0 if we grabbed it, non-zero otherwise.
 880 */
 881STATIC int
 882xfs_reclaim_inode_grab(
 883	struct xfs_inode	*ip,
 884	int			flags)
 885{
 886	ASSERT(rcu_read_lock_held());
 887
 888	/* quick check for stale RCU freed inode */
 889	if (!ip->i_ino)
 890		return 1;
 891
 892	/*
 893	 * If we are asked for non-blocking operation, do unlocked checks to
 894	 * see if the inode already is being flushed or in reclaim to avoid
 895	 * lock traffic.
 896	 */
 897	if ((flags & SYNC_TRYLOCK) &&
 898	    __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
 899		return 1;
 900
 901	/*
 902	 * The radix tree lock here protects a thread in xfs_iget from racing
 903	 * with us starting reclaim on the inode.  Once we have the
 904	 * XFS_IRECLAIM flag set it will not touch us.
 905	 *
 906	 * Due to RCU lookup, we may find inodes that have been freed and only
 907	 * have XFS_IRECLAIM set.  Indeed, we may see reallocated inodes that
 908	 * aren't candidates for reclaim at all, so we must check the
 909	 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
 910	 */
 911	spin_lock(&ip->i_flags_lock);
 912	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
 913	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {
 914		/* not a reclaim candidate. */
 915		spin_unlock(&ip->i_flags_lock);
 916		return 1;
 917	}
 918	__xfs_iflags_set(ip, XFS_IRECLAIM);
 919	spin_unlock(&ip->i_flags_lock);
 920	return 0;
 921}
 922
 923/*
 924 * Inodes in different states need to be treated differently. The following
 925 * table lists the inode states and the reclaim actions necessary:
 926 *
 927 *	inode state	     iflush ret		required action
 928 *      ---------------      ----------         ---------------
 929 *	bad			-		reclaim
 930 *	shutdown		EIO		unpin and reclaim
 931 *	clean, unpinned		0		reclaim
 932 *	stale, unpinned		0		reclaim
 933 *	clean, pinned(*)	0		requeue
 934 *	stale, pinned		EAGAIN		requeue
 935 *	dirty, async		-		requeue
 936 *	dirty, sync		0		reclaim
 937 *
 938 * (*) dgc: I don't think the clean, pinned state is possible but it gets
 939 * handled anyway given the order of checks implemented.
 940 *
 941 * Also, because we get the flush lock first, we know that any inode that has
 942 * been flushed delwri has had the flush completed by the time we check that
 943 * the inode is clean.
 944 *
 945 * Note that because the inode is flushed delayed write by AIL pushing, the
 946 * flush lock may already be held here and waiting on it can result in very
 947 * long latencies.  Hence for sync reclaims, where we wait on the flush lock,
 948 * the caller should push the AIL first before trying to reclaim inodes to
 949 * minimise the amount of time spent waiting.  For background relaim, we only
 950 * bother to reclaim clean inodes anyway.
 951 *
 952 * Hence the order of actions after gaining the locks should be:
 953 *	bad		=> reclaim
 954 *	shutdown	=> unpin and reclaim
 955 *	pinned, async	=> requeue
 956 *	pinned, sync	=> unpin
 957 *	stale		=> reclaim
 958 *	clean		=> reclaim
 959 *	dirty, async	=> requeue
 960 *	dirty, sync	=> flush, wait and reclaim
 961 */
 962STATIC int
 963xfs_reclaim_inode(
 964	struct xfs_inode	*ip,
 965	struct xfs_perag	*pag,
 966	int			sync_mode)
 967{
 968	struct xfs_buf		*bp = NULL;
 969	xfs_ino_t		ino = ip->i_ino; /* for radix_tree_delete */
 970	int			error;
 971
 972restart:
 973	error = 0;
 974	xfs_ilock(ip, XFS_ILOCK_EXCL);
 975	if (!xfs_iflock_nowait(ip)) {
 976		if (!(sync_mode & SYNC_WAIT))
 977			goto out;
 978		xfs_iflock(ip);
 979	}
 980
 981	if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 982		xfs_iunpin_wait(ip);
 983		/* xfs_iflush_abort() drops the flush lock */
 984		xfs_iflush_abort(ip, false);
 985		goto reclaim;
 986	}
 987	if (xfs_ipincount(ip)) {
 988		if (!(sync_mode & SYNC_WAIT))
 989			goto out_ifunlock;
 990		xfs_iunpin_wait(ip);
 991	}
 992	if (xfs_iflags_test(ip, XFS_ISTALE) || xfs_inode_clean(ip)) {
 993		xfs_ifunlock(ip);
 
 994		goto reclaim;
 995	}
 996
 997	/*
 998	 * Never flush out dirty data during non-blocking reclaim, as it would
 999	 * just contend with AIL pushing trying to do the same job.
1000	 */
1001	if (!(sync_mode & SYNC_WAIT))
1002		goto out_ifunlock;
1003
1004	/*
1005	 * Now we have an inode that needs flushing.
1006	 *
1007	 * Note that xfs_iflush will never block on the inode buffer lock, as
1008	 * xfs_ifree_cluster() can lock the inode buffer before it locks the
1009	 * ip->i_lock, and we are doing the exact opposite here.  As a result,
1010	 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
1011	 * result in an ABBA deadlock with xfs_ifree_cluster().
1012	 *
1013	 * As xfs_ifree_cluser() must gather all inodes that are active in the
1014	 * cache to mark them stale, if we hit this case we don't actually want
1015	 * to do IO here - we want the inode marked stale so we can simply
1016	 * reclaim it.  Hence if we get an EAGAIN error here,  just unlock the
1017	 * inode, back off and try again.  Hopefully the next pass through will
1018	 * see the stale flag set on the inode.
1019	 */
1020	error = xfs_iflush(ip, &bp);
1021	if (error == -EAGAIN) {
1022		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1023		/* backoff longer than in xfs_ifree_cluster */
1024		delay(2);
1025		goto restart;
1026	}
1027
1028	if (!error) {
1029		error = xfs_bwrite(bp);
1030		xfs_buf_relse(bp);
1031	}
1032
 
1033reclaim:
1034	ASSERT(!xfs_isiflocked(ip));
1035
1036	/*
1037	 * Because we use RCU freeing we need to ensure the inode always appears
1038	 * to be reclaimed with an invalid inode number when in the free state.
1039	 * We do this as early as possible under the ILOCK so that
1040	 * xfs_iflush_cluster() can be guaranteed to detect races with us here.
1041	 * By doing this, we guarantee that once xfs_iflush_cluster has locked
1042	 * XFS_ILOCK that it will see either a valid, flushable inode that will
1043	 * serialise correctly, or it will see a clean (and invalid) inode that
1044	 * it can skip.
1045	 */
1046	spin_lock(&ip->i_flags_lock);
1047	ip->i_flags = XFS_IRECLAIM;
1048	ip->i_ino = 0;
1049	spin_unlock(&ip->i_flags_lock);
1050
1051	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1052
1053	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1054	/*
1055	 * Remove the inode from the per-AG radix tree.
1056	 *
1057	 * Because radix_tree_delete won't complain even if the item was never
1058	 * added to the tree assert that it's been there before to catch
1059	 * problems with the inode life time early on.
1060	 */
1061	spin_lock(&pag->pag_ici_lock);
1062	if (!radix_tree_delete(&pag->pag_ici_root,
1063				XFS_INO_TO_AGINO(ip->i_mount, ino)))
1064		ASSERT(0);
1065	xfs_perag_clear_reclaim_tag(pag);
1066	spin_unlock(&pag->pag_ici_lock);
1067
1068	/*
1069	 * Here we do an (almost) spurious inode lock in order to coordinate
1070	 * with inode cache radix tree lookups.  This is because the lookup
1071	 * can reference the inodes in the cache without taking references.
1072	 *
1073	 * We make that OK here by ensuring that we wait until the inode is
1074	 * unlocked after the lookup before we go ahead and free it.
1075	 */
1076	xfs_ilock(ip, XFS_ILOCK_EXCL);
1077	xfs_qm_dqdetach(ip);
1078	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1079
1080	__xfs_inode_free(ip);
1081	return error;
1082
1083out_ifunlock:
1084	xfs_ifunlock(ip);
1085out:
1086	xfs_iflags_clear(ip, XFS_IRECLAIM);
1087	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1088	/*
1089	 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1090	 * a short while. However, this just burns CPU time scanning the tree
1091	 * waiting for IO to complete and the reclaim work never goes back to
1092	 * the idle state. Instead, return 0 to let the next scheduled
1093	 * background reclaim attempt to reclaim the inode again.
1094	 */
1095	return 0;
1096}
1097
1098/*
1099 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1100 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1101 * then a shut down during filesystem unmount reclaim walk leak all the
1102 * unreclaimed inodes.
1103 */
1104STATIC int
1105xfs_reclaim_inodes_ag(
1106	struct xfs_mount	*mp,
1107	int			flags,
1108	int			*nr_to_scan)
1109{
1110	struct xfs_perag	*pag;
1111	int			error = 0;
1112	int			last_error = 0;
1113	xfs_agnumber_t		ag;
1114	int			trylock = flags & SYNC_TRYLOCK;
1115	int			skipped;
1116
1117restart:
1118	ag = 0;
1119	skipped = 0;
1120	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1121		unsigned long	first_index = 0;
1122		int		done = 0;
1123		int		nr_found = 0;
1124
1125		ag = pag->pag_agno + 1;
1126
1127		if (trylock) {
1128			if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1129				skipped++;
1130				xfs_perag_put(pag);
1131				continue;
1132			}
1133			first_index = pag->pag_ici_reclaim_cursor;
1134		} else
1135			mutex_lock(&pag->pag_ici_reclaim_lock);
1136
1137		do {
1138			struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1139			int	i;
1140
1141			rcu_read_lock();
1142			nr_found = radix_tree_gang_lookup_tag(
1143					&pag->pag_ici_root,
1144					(void **)batch, first_index,
1145					XFS_LOOKUP_BATCH,
1146					XFS_ICI_RECLAIM_TAG);
1147			if (!nr_found) {
1148				done = 1;
1149				rcu_read_unlock();
1150				break;
1151			}
1152
1153			/*
1154			 * Grab the inodes before we drop the lock. if we found
1155			 * nothing, nr == 0 and the loop will be skipped.
1156			 */
1157			for (i = 0; i < nr_found; i++) {
1158				struct xfs_inode *ip = batch[i];
1159
1160				if (done || xfs_reclaim_inode_grab(ip, flags))
1161					batch[i] = NULL;
1162
1163				/*
1164				 * Update the index for the next lookup. Catch
1165				 * overflows into the next AG range which can
1166				 * occur if we have inodes in the last block of
1167				 * the AG and we are currently pointing to the
1168				 * last inode.
1169				 *
1170				 * Because we may see inodes that are from the
1171				 * wrong AG due to RCU freeing and
1172				 * reallocation, only update the index if it
1173				 * lies in this AG. It was a race that lead us
1174				 * to see this inode, so another lookup from
1175				 * the same index will not find it again.
1176				 */
1177				if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1178								pag->pag_agno)
1179					continue;
1180				first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1181				if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1182					done = 1;
1183			}
1184
1185			/* unlock now we've grabbed the inodes. */
1186			rcu_read_unlock();
1187
1188			for (i = 0; i < nr_found; i++) {
1189				if (!batch[i])
1190					continue;
1191				error = xfs_reclaim_inode(batch[i], pag, flags);
1192				if (error && last_error != -EFSCORRUPTED)
1193					last_error = error;
1194			}
1195
1196			*nr_to_scan -= XFS_LOOKUP_BATCH;
1197
1198			cond_resched();
1199
1200		} while (nr_found && !done && *nr_to_scan > 0);
1201
1202		if (trylock && !done)
1203			pag->pag_ici_reclaim_cursor = first_index;
1204		else
1205			pag->pag_ici_reclaim_cursor = 0;
1206		mutex_unlock(&pag->pag_ici_reclaim_lock);
1207		xfs_perag_put(pag);
1208	}
1209
1210	/*
1211	 * if we skipped any AG, and we still have scan count remaining, do
1212	 * another pass this time using blocking reclaim semantics (i.e
1213	 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1214	 * ensure that when we get more reclaimers than AGs we block rather
1215	 * than spin trying to execute reclaim.
1216	 */
1217	if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1218		trylock = 0;
1219		goto restart;
1220	}
1221	return last_error;
1222}
1223
1224int
1225xfs_reclaim_inodes(
1226	xfs_mount_t	*mp,
1227	int		mode)
1228{
1229	int		nr_to_scan = INT_MAX;
1230
1231	return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1232}
1233
1234/*
1235 * Scan a certain number of inodes for reclaim.
1236 *
1237 * When called we make sure that there is a background (fast) inode reclaim in
1238 * progress, while we will throttle the speed of reclaim via doing synchronous
1239 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1240 * them to be cleaned, which we hope will not be very long due to the
1241 * background walker having already kicked the IO off on those dirty inodes.
1242 */
1243long
1244xfs_reclaim_inodes_nr(
1245	struct xfs_mount	*mp,
1246	int			nr_to_scan)
1247{
1248	/* kick background reclaimer and push the AIL */
1249	xfs_reclaim_work_queue(mp);
1250	xfs_ail_push_all(mp->m_ail);
1251
1252	return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1253}
1254
1255/*
1256 * Return the number of reclaimable inodes in the filesystem for
1257 * the shrinker to determine how much to reclaim.
1258 */
1259int
1260xfs_reclaim_inodes_count(
1261	struct xfs_mount	*mp)
1262{
1263	struct xfs_perag	*pag;
1264	xfs_agnumber_t		ag = 0;
1265	int			reclaimable = 0;
1266
1267	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1268		ag = pag->pag_agno + 1;
1269		reclaimable += pag->pag_ici_reclaimable;
1270		xfs_perag_put(pag);
1271	}
1272	return reclaimable;
1273}
1274
1275STATIC int
1276xfs_inode_match_id(
1277	struct xfs_inode	*ip,
1278	struct xfs_eofblocks	*eofb)
1279{
1280	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1281	    !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1282		return 0;
1283
1284	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1285	    !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1286		return 0;
1287
1288	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1289	    xfs_get_projid(ip) != eofb->eof_prid)
1290		return 0;
1291
1292	return 1;
1293}
1294
1295/*
1296 * A union-based inode filtering algorithm. Process the inode if any of the
1297 * criteria match. This is for global/internal scans only.
1298 */
1299STATIC int
1300xfs_inode_match_id_union(
1301	struct xfs_inode	*ip,
1302	struct xfs_eofblocks	*eofb)
1303{
1304	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1305	    uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1306		return 1;
1307
1308	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1309	    gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1310		return 1;
1311
1312	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1313	    xfs_get_projid(ip) == eofb->eof_prid)
1314		return 1;
1315
1316	return 0;
1317}
1318
1319STATIC int
1320xfs_inode_free_eofblocks(
1321	struct xfs_inode	*ip,
 
1322	int			flags,
1323	void			*args)
1324{
1325	int ret = 0;
1326	struct xfs_eofblocks *eofb = args;
1327	int match;
1328
1329	if (!xfs_can_free_eofblocks(ip, false)) {
1330		/* inode could be preallocated or append-only */
1331		trace_xfs_inode_free_eofblocks_invalid(ip);
1332		xfs_inode_clear_eofblocks_tag(ip);
1333		return 0;
1334	}
1335
1336	/*
1337	 * If the mapping is dirty the operation can block and wait for some
1338	 * time. Unless we are waiting, skip it.
1339	 */
1340	if (!(flags & SYNC_WAIT) &&
1341	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1342		return 0;
1343
1344	if (eofb) {
1345		if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1346			match = xfs_inode_match_id_union(ip, eofb);
1347		else
1348			match = xfs_inode_match_id(ip, eofb);
1349		if (!match)
1350			return 0;
1351
1352		/* skip the inode if the file size is too small */
1353		if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1354		    XFS_ISIZE(ip) < eofb->eof_min_file_size)
1355			return 0;
1356	}
1357
1358	/*
1359	 * If the caller is waiting, return -EAGAIN to keep the background
1360	 * scanner moving and revisit the inode in a subsequent pass.
1361	 */
1362	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1363		if (flags & SYNC_WAIT)
1364			ret = -EAGAIN;
1365		return ret;
1366	}
1367	ret = xfs_free_eofblocks(ip);
1368	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1369
1370	return ret;
1371}
1372
1373static int
1374__xfs_icache_free_eofblocks(
1375	struct xfs_mount	*mp,
1376	struct xfs_eofblocks	*eofb,
1377	int			(*execute)(struct xfs_inode *ip, int flags,
1378					   void *args),
1379	int			tag)
1380{
1381	int flags = SYNC_TRYLOCK;
1382
1383	if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1384		flags = SYNC_WAIT;
 
1385
1386	return xfs_inode_ag_iterator_tag(mp, execute, flags,
1387					 eofb, tag);
1388}
1389
1390int
1391xfs_icache_free_eofblocks(
1392	struct xfs_mount	*mp,
1393	struct xfs_eofblocks	*eofb)
1394{
1395	return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_eofblocks,
1396			XFS_ICI_EOFBLOCKS_TAG);
1397}
1398
1399/*
1400 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1401 * multiple quotas, we don't know exactly which quota caused an allocation
1402 * failure. We make a best effort by including each quota under low free space
1403 * conditions (less than 1% free space) in the scan.
1404 */
1405static int
1406__xfs_inode_free_quota_eofblocks(
1407	struct xfs_inode	*ip,
1408	int			(*execute)(struct xfs_mount *mp,
1409					   struct xfs_eofblocks	*eofb))
1410{
1411	int scan = 0;
1412	struct xfs_eofblocks eofb = {0};
1413	struct xfs_dquot *dq;
1414
1415	/*
1416	 * Run a sync scan to increase effectiveness and use the union filter to
1417	 * cover all applicable quotas in a single scan.
1418	 */
1419	eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1420
1421	if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1422		dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1423		if (dq && xfs_dquot_lowsp(dq)) {
1424			eofb.eof_uid = VFS_I(ip)->i_uid;
1425			eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1426			scan = 1;
1427		}
1428	}
1429
1430	if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1431		dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1432		if (dq && xfs_dquot_lowsp(dq)) {
1433			eofb.eof_gid = VFS_I(ip)->i_gid;
1434			eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1435			scan = 1;
1436		}
1437	}
1438
1439	if (scan)
1440		execute(ip->i_mount, &eofb);
1441
1442	return scan;
 
1443}
1444
1445int
1446xfs_inode_free_quota_eofblocks(
1447	struct xfs_inode *ip)
1448{
1449	return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_eofblocks);
1450}
1451
1452static void
1453__xfs_inode_set_eofblocks_tag(
1454	xfs_inode_t	*ip,
1455	void		(*execute)(struct xfs_mount *mp),
1456	void		(*set_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1457				  int error, unsigned long caller_ip),
1458	int		tag)
1459{
1460	struct xfs_mount *mp = ip->i_mount;
1461	struct xfs_perag *pag;
1462	int tagged;
1463
1464	/*
1465	 * Don't bother locking the AG and looking up in the radix trees
1466	 * if we already know that we have the tag set.
1467	 */
1468	if (ip->i_flags & XFS_IEOFBLOCKS)
1469		return;
1470	spin_lock(&ip->i_flags_lock);
1471	ip->i_flags |= XFS_IEOFBLOCKS;
1472	spin_unlock(&ip->i_flags_lock);
1473
1474	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1475	spin_lock(&pag->pag_ici_lock);
 
1476
1477	tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
 
1478	radix_tree_tag_set(&pag->pag_ici_root,
1479			   XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
 
1480	if (!tagged) {
1481		/* propagate the eofblocks tag up into the perag radix tree */
1482		spin_lock(&ip->i_mount->m_perag_lock);
1483		radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1484				   XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1485				   tag);
1486		spin_unlock(&ip->i_mount->m_perag_lock);
1487
1488		/* kick off background trimming */
1489		execute(ip->i_mount);
1490
1491		set_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
 
1492	}
1493
1494	spin_unlock(&pag->pag_ici_lock);
1495	xfs_perag_put(pag);
1496}
1497
1498void
1499xfs_inode_set_eofblocks_tag(
1500	xfs_inode_t	*ip)
1501{
1502	trace_xfs_inode_set_eofblocks_tag(ip);
1503	return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_eofblocks,
1504			trace_xfs_perag_set_eofblocks,
1505			XFS_ICI_EOFBLOCKS_TAG);
1506}
1507
1508static void
1509__xfs_inode_clear_eofblocks_tag(
1510	xfs_inode_t	*ip,
1511	void		(*clear_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1512				    int error, unsigned long caller_ip),
1513	int		tag)
1514{
1515	struct xfs_mount *mp = ip->i_mount;
1516	struct xfs_perag *pag;
1517
1518	spin_lock(&ip->i_flags_lock);
1519	ip->i_flags &= ~XFS_IEOFBLOCKS;
1520	spin_unlock(&ip->i_flags_lock);
1521
1522	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1523	spin_lock(&pag->pag_ici_lock);
 
1524
1525	radix_tree_tag_clear(&pag->pag_ici_root,
1526			     XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1527	if (!radix_tree_tagged(&pag->pag_ici_root, tag)) {
 
1528		/* clear the eofblocks tag from the perag radix tree */
1529		spin_lock(&ip->i_mount->m_perag_lock);
1530		radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1531				     XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1532				     tag);
1533		spin_unlock(&ip->i_mount->m_perag_lock);
1534		clear_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
 
1535	}
1536
1537	spin_unlock(&pag->pag_ici_lock);
1538	xfs_perag_put(pag);
1539}
1540
1541void
1542xfs_inode_clear_eofblocks_tag(
1543	xfs_inode_t	*ip)
1544{
1545	trace_xfs_inode_clear_eofblocks_tag(ip);
1546	return __xfs_inode_clear_eofblocks_tag(ip,
1547			trace_xfs_perag_clear_eofblocks, XFS_ICI_EOFBLOCKS_TAG);
1548}
1549
1550/*
1551 * Automatic CoW Reservation Freeing
1552 *
1553 * These functions automatically garbage collect leftover CoW reservations
1554 * that were made on behalf of a cowextsize hint when we start to run out
1555 * of quota or when the reservations sit around for too long.  If the file
1556 * has dirty pages or is undergoing writeback, its CoW reservations will
1557 * be retained.
1558 *
1559 * The actual garbage collection piggybacks off the same code that runs
1560 * the speculative EOF preallocation garbage collector.
1561 */
1562STATIC int
1563xfs_inode_free_cowblocks(
1564	struct xfs_inode	*ip,
1565	int			flags,
1566	void			*args)
1567{
1568	int ret;
1569	struct xfs_eofblocks *eofb = args;
1570	int match;
1571	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1572
1573	/*
1574	 * Just clear the tag if we have an empty cow fork or none at all. It's
1575	 * possible the inode was fully unshared since it was originally tagged.
1576	 */
1577	if (!xfs_is_reflink_inode(ip) || !ifp->if_bytes) {
1578		trace_xfs_inode_free_cowblocks_invalid(ip);
1579		xfs_inode_clear_cowblocks_tag(ip);
1580		return 0;
1581	}
1582
1583	/*
1584	 * If the mapping is dirty or under writeback we cannot touch the
1585	 * CoW fork.  Leave it alone if we're in the midst of a directio.
1586	 */
1587	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1588	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1589	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1590	    atomic_read(&VFS_I(ip)->i_dio_count))
1591		return 0;
1592
1593	if (eofb) {
1594		if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1595			match = xfs_inode_match_id_union(ip, eofb);
1596		else
1597			match = xfs_inode_match_id(ip, eofb);
1598		if (!match)
1599			return 0;
1600
1601		/* skip the inode if the file size is too small */
1602		if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1603		    XFS_ISIZE(ip) < eofb->eof_min_file_size)
1604			return 0;
1605	}
1606
1607	/* Free the CoW blocks */
1608	xfs_ilock(ip, XFS_IOLOCK_EXCL);
1609	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1610
1611	ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1612
1613	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1614	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1615
1616	return ret;
1617}
1618
1619int
1620xfs_icache_free_cowblocks(
1621	struct xfs_mount	*mp,
1622	struct xfs_eofblocks	*eofb)
1623{
1624	return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_cowblocks,
1625			XFS_ICI_COWBLOCKS_TAG);
1626}
1627
1628int
1629xfs_inode_free_quota_cowblocks(
1630	struct xfs_inode *ip)
1631{
1632	return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_cowblocks);
1633}
1634
1635void
1636xfs_inode_set_cowblocks_tag(
1637	xfs_inode_t	*ip)
1638{
1639	trace_xfs_inode_set_cowblocks_tag(ip);
1640	return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_cowblocks,
1641			trace_xfs_perag_set_cowblocks,
1642			XFS_ICI_COWBLOCKS_TAG);
1643}
1644
1645void
1646xfs_inode_clear_cowblocks_tag(
1647	xfs_inode_t	*ip)
1648{
1649	trace_xfs_inode_clear_cowblocks_tag(ip);
1650	return __xfs_inode_clear_eofblocks_tag(ip,
1651			trace_xfs_perag_clear_cowblocks, XFS_ICI_COWBLOCKS_TAG);
1652}