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
  37#include <linux/kthread.h>
  38#include <linux/freezer.h>
  39
  40STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp,
  41				struct xfs_perag *pag, struct xfs_inode *ip);
  42
  43/*
  44 * Allocate and initialise an xfs_inode.
  45 */
  46struct xfs_inode *
  47xfs_inode_alloc(
  48	struct xfs_mount	*mp,
  49	xfs_ino_t		ino)
  50{
  51	struct xfs_inode	*ip;
  52
  53	/*
  54	 * if this didn't occur in transactions, we could use
  55	 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
  56	 * code up to do this anyway.
  57	 */
  58	ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
  59	if (!ip)
  60		return NULL;
  61	if (inode_init_always(mp->m_super, VFS_I(ip))) {
  62		kmem_zone_free(xfs_inode_zone, ip);
  63		return NULL;
  64	}
  65
  66	/* VFS doesn't initialise i_mode! */
  67	VFS_I(ip)->i_mode = 0;
  68
  69	XFS_STATS_INC(mp, vn_active);
  70	ASSERT(atomic_read(&ip->i_pincount) == 0);
  71	ASSERT(!spin_is_locked(&ip->i_flags_lock));
  72	ASSERT(!xfs_isiflocked(ip));
  73	ASSERT(ip->i_ino == 0);
  74
  75	mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
  76
  77	/* initialise the xfs inode */
  78	ip->i_ino = ino;
  79	ip->i_mount = mp;
  80	memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
  81	ip->i_afp = NULL;
  82	memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
  83	ip->i_flags = 0;
  84	ip->i_delayed_blks = 0;
  85	memset(&ip->i_d, 0, sizeof(ip->i_d));
  86
  87	return ip;
  88}
  89
  90STATIC void
  91xfs_inode_free_callback(
  92	struct rcu_head		*head)
  93{
  94	struct inode		*inode = container_of(head, struct inode, i_rcu);
  95	struct xfs_inode	*ip = XFS_I(inode);
  96
  97	kmem_zone_free(xfs_inode_zone, ip);
  98}
  99
 100void
 101xfs_inode_free(
 102	struct xfs_inode	*ip)
 103{
 104	switch (VFS_I(ip)->i_mode & S_IFMT) {
 105	case S_IFREG:
 106	case S_IFDIR:
 107	case S_IFLNK:
 108		xfs_idestroy_fork(ip, XFS_DATA_FORK);
 109		break;
 110	}
 111
 112	if (ip->i_afp)
 113		xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 114
 115	if (ip->i_itemp) {
 116		ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
 117		xfs_inode_item_destroy(ip);
 118		ip->i_itemp = NULL;
 119	}
 120
 121	/*
 122	 * Because we use RCU freeing we need to ensure the inode always
 123	 * appears to be reclaimed with an invalid inode number when in the
 124	 * free state. The ip->i_flags_lock provides the barrier against lookup
 125	 * races.
 126	 */
 127	spin_lock(&ip->i_flags_lock);
 128	ip->i_flags = XFS_IRECLAIM;
 129	ip->i_ino = 0;
 130	spin_unlock(&ip->i_flags_lock);
 131
 132	/* asserts to verify all state is correct here */
 133	ASSERT(atomic_read(&ip->i_pincount) == 0);
 134	ASSERT(!xfs_isiflocked(ip));
 135	XFS_STATS_DEC(ip->i_mount, vn_active);
 136
 137	call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
 138}
 139
 140/*
 141 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
 142 * part of the structure. This is made more complex by the fact we store
 143 * information about the on-disk values in the VFS inode and so we can't just
 144 * overwrite the values unconditionally. Hence we save the parameters we
 145 * need to retain across reinitialisation, and rewrite them into the VFS inode
 146 * after reinitialisation even if it fails.
 147 */
 148static int
 149xfs_reinit_inode(
 150	struct xfs_mount	*mp,
 151	struct inode		*inode)
 152{
 153	int		error;
 154	uint32_t	nlink = inode->i_nlink;
 155	uint32_t	generation = inode->i_generation;
 156	uint64_t	version = inode->i_version;
 157	umode_t		mode = inode->i_mode;
 158
 159	error = inode_init_always(mp->m_super, inode);
 160
 161	set_nlink(inode, nlink);
 162	inode->i_generation = generation;
 163	inode->i_version = version;
 164	inode->i_mode = mode;
 165	return error;
 166}
 167
 168/*
 169 * Check the validity of the inode we just found it the cache
 170 */
 171static int
 172xfs_iget_cache_hit(
 173	struct xfs_perag	*pag,
 174	struct xfs_inode	*ip,
 175	xfs_ino_t		ino,
 176	int			flags,
 177	int			lock_flags) __releases(RCU)
 178{
 179	struct inode		*inode = VFS_I(ip);
 180	struct xfs_mount	*mp = ip->i_mount;
 181	int			error;
 182
 183	/*
 184	 * check for re-use of an inode within an RCU grace period due to the
 185	 * radix tree nodes not being updated yet. We monitor for this by
 186	 * setting the inode number to zero before freeing the inode structure.
 187	 * If the inode has been reallocated and set up, then the inode number
 188	 * will not match, so check for that, too.
 189	 */
 190	spin_lock(&ip->i_flags_lock);
 191	if (ip->i_ino != ino) {
 192		trace_xfs_iget_skip(ip);
 193		XFS_STATS_INC(mp, xs_ig_frecycle);
 194		error = -EAGAIN;
 195		goto out_error;
 196	}
 197
 198
 199	/*
 200	 * If we are racing with another cache hit that is currently
 201	 * instantiating this inode or currently recycling it out of
 202	 * reclaimabe state, wait for the initialisation to complete
 203	 * before continuing.
 204	 *
 205	 * XXX(hch): eventually we should do something equivalent to
 206	 *	     wait_on_inode to wait for these flags to be cleared
 207	 *	     instead of polling for it.
 208	 */
 209	if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
 210		trace_xfs_iget_skip(ip);
 211		XFS_STATS_INC(mp, xs_ig_frecycle);
 212		error = -EAGAIN;
 213		goto out_error;
 214	}
 215
 216	/*
 217	 * If lookup is racing with unlink return an error immediately.
 218	 */
 219	if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
 220		error = -ENOENT;
 221		goto out_error;
 222	}
 223
 224	/*
 225	 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
 226	 * Need to carefully get it back into useable state.
 227	 */
 228	if (ip->i_flags & XFS_IRECLAIMABLE) {
 229		trace_xfs_iget_reclaim(ip);
 230
 231		/*
 232		 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
 233		 * from stomping over us while we recycle the inode.  We can't
 234		 * clear the radix tree reclaimable tag yet as it requires
 235		 * pag_ici_lock to be held exclusive.
 236		 */
 237		ip->i_flags |= XFS_IRECLAIM;
 238
 239		spin_unlock(&ip->i_flags_lock);
 240		rcu_read_unlock();
 241
 242		error = xfs_reinit_inode(mp, inode);
 243		if (error) {
 244			/*
 245			 * Re-initializing the inode failed, and we are in deep
 246			 * trouble.  Try to re-add it to the reclaim list.
 247			 */
 248			rcu_read_lock();
 249			spin_lock(&ip->i_flags_lock);
 250
 251			ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
 252			ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
 253			trace_xfs_iget_reclaim_fail(ip);
 254			goto out_error;
 255		}
 256
 257		spin_lock(&pag->pag_ici_lock);
 258		spin_lock(&ip->i_flags_lock);
 259
 260		/*
 261		 * Clear the per-lifetime state in the inode as we are now
 262		 * effectively a new inode and need to return to the initial
 263		 * state before reuse occurs.
 264		 */
 265		ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
 266		ip->i_flags |= XFS_INEW;
 267		__xfs_inode_clear_reclaim_tag(mp, pag, ip);
 268		inode->i_state = I_NEW;
 269
 270		ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
 271		mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
 272
 273		spin_unlock(&ip->i_flags_lock);
 274		spin_unlock(&pag->pag_ici_lock);
 275	} else {
 276		/* If the VFS inode is being torn down, pause and try again. */
 277		if (!igrab(inode)) {
 278			trace_xfs_iget_skip(ip);
 279			error = -EAGAIN;
 280			goto out_error;
 281		}
 282
 283		/* We've got a live one. */
 284		spin_unlock(&ip->i_flags_lock);
 285		rcu_read_unlock();
 286		trace_xfs_iget_hit(ip);
 287	}
 288
 289	if (lock_flags != 0)
 290		xfs_ilock(ip, lock_flags);
 291
 292	xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
 293	XFS_STATS_INC(mp, xs_ig_found);
 294
 295	return 0;
 296
 297out_error:
 298	spin_unlock(&ip->i_flags_lock);
 299	rcu_read_unlock();
 300	return error;
 301}
 302
 303
 304static int
 305xfs_iget_cache_miss(
 306	struct xfs_mount	*mp,
 307	struct xfs_perag	*pag,
 308	xfs_trans_t		*tp,
 309	xfs_ino_t		ino,
 310	struct xfs_inode	**ipp,
 311	int			flags,
 312	int			lock_flags)
 313{
 314	struct xfs_inode	*ip;
 315	int			error;
 316	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);
 317	int			iflags;
 318
 319	ip = xfs_inode_alloc(mp, ino);
 320	if (!ip)
 321		return -ENOMEM;
 322
 323	error = xfs_iread(mp, tp, ip, flags);
 324	if (error)
 325		goto out_destroy;
 326
 327	trace_xfs_iget_miss(ip);
 328
 329	if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
 330		error = -ENOENT;
 331		goto out_destroy;
 332	}
 333
 334	/*
 335	 * Preload the radix tree so we can insert safely under the
 336	 * write spinlock. Note that we cannot sleep inside the preload
 337	 * region. Since we can be called from transaction context, don't
 338	 * recurse into the file system.
 339	 */
 340	if (radix_tree_preload(GFP_NOFS)) {
 341		error = -EAGAIN;
 342		goto out_destroy;
 343	}
 344
 345	/*
 346	 * Because the inode hasn't been added to the radix-tree yet it can't
 347	 * be found by another thread, so we can do the non-sleeping lock here.
 348	 */
 349	if (lock_flags) {
 350		if (!xfs_ilock_nowait(ip, lock_flags))
 351			BUG();
 352	}
 353
 354	/*
 355	 * These values must be set before inserting the inode into the radix
 356	 * tree as the moment it is inserted a concurrent lookup (allowed by the
 357	 * RCU locking mechanism) can find it and that lookup must see that this
 358	 * is an inode currently under construction (i.e. that XFS_INEW is set).
 359	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
 360	 * memory barrier that ensures this detection works correctly at lookup
 361	 * time.
 362	 */
 363	iflags = XFS_INEW;
 364	if (flags & XFS_IGET_DONTCACHE)
 365		iflags |= XFS_IDONTCACHE;
 366	ip->i_udquot = NULL;
 367	ip->i_gdquot = NULL;
 368	ip->i_pdquot = NULL;
 369	xfs_iflags_set(ip, iflags);
 370
 371	/* insert the new inode */
 372	spin_lock(&pag->pag_ici_lock);
 373	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
 374	if (unlikely(error)) {
 375		WARN_ON(error != -EEXIST);
 376		XFS_STATS_INC(mp, xs_ig_dup);
 377		error = -EAGAIN;
 378		goto out_preload_end;
 379	}
 380	spin_unlock(&pag->pag_ici_lock);
 381	radix_tree_preload_end();
 382
 383	*ipp = ip;
 384	return 0;
 385
 386out_preload_end:
 387	spin_unlock(&pag->pag_ici_lock);
 388	radix_tree_preload_end();
 389	if (lock_flags)
 390		xfs_iunlock(ip, lock_flags);
 391out_destroy:
 392	__destroy_inode(VFS_I(ip));
 393	xfs_inode_free(ip);
 394	return error;
 395}
 396
 397/*
 398 * Look up an inode by number in the given file system.
 399 * The inode is looked up in the cache held in each AG.
 400 * If the inode is found in the cache, initialise the vfs inode
 401 * if necessary.
 402 *
 403 * If it is not in core, read it in from the file system's device,
 404 * add it to the cache and initialise the vfs inode.
 405 *
 406 * The inode is locked according to the value of the lock_flags parameter.
 407 * This flag parameter indicates how and if the inode's IO lock and inode lock
 408 * should be taken.
 409 *
 410 * mp -- the mount point structure for the current file system.  It points
 411 *       to the inode hash table.
 412 * tp -- a pointer to the current transaction if there is one.  This is
 413 *       simply passed through to the xfs_iread() call.
 414 * ino -- the number of the inode desired.  This is the unique identifier
 415 *        within the file system for the inode being requested.
 416 * lock_flags -- flags indicating how to lock the inode.  See the comment
 417 *		 for xfs_ilock() for a list of valid values.
 418 */
 419int
 420xfs_iget(
 421	xfs_mount_t	*mp,
 422	xfs_trans_t	*tp,
 423	xfs_ino_t	ino,
 424	uint		flags,
 425	uint		lock_flags,
 426	xfs_inode_t	**ipp)
 427{
 428	xfs_inode_t	*ip;
 429	int		error;
 430	xfs_perag_t	*pag;
 431	xfs_agino_t	agino;
 432
 433	/*
 434	 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
 435	 * doesn't get freed while it's being referenced during a
 436	 * radix tree traversal here.  It assumes this function
 437	 * aqcuires only the ILOCK (and therefore it has no need to
 438	 * involve the IOLOCK in this synchronization).
 439	 */
 440	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
 441
 442	/* reject inode numbers outside existing AGs */
 443	if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
 444		return -EINVAL;
 445
 446	XFS_STATS_INC(mp, xs_ig_attempts);
 447
 448	/* get the perag structure and ensure that it's inode capable */
 449	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
 450	agino = XFS_INO_TO_AGINO(mp, ino);
 451
 452again:
 453	error = 0;
 454	rcu_read_lock();
 455	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
 456
 457	if (ip) {
 458		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
 459		if (error)
 460			goto out_error_or_again;
 461	} else {
 462		rcu_read_unlock();
 463		XFS_STATS_INC(mp, xs_ig_missed);
 464
 465		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
 466							flags, lock_flags);
 467		if (error)
 468			goto out_error_or_again;
 469	}
 470	xfs_perag_put(pag);
 471
 472	*ipp = ip;
 473
 474	/*
 475	 * If we have a real type for an on-disk inode, we can setup the inode
 476	 * now.	 If it's a new inode being created, xfs_ialloc will handle it.
 477	 */
 478	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
 479		xfs_setup_existing_inode(ip);
 480	return 0;
 481
 482out_error_or_again:
 483	if (error == -EAGAIN) {
 484		delay(1);
 485		goto again;
 486	}
 487	xfs_perag_put(pag);
 488	return error;
 489}
 490
 491/*
 492 * The inode lookup is done in batches to keep the amount of lock traffic and
 493 * radix tree lookups to a minimum. The batch size is a trade off between
 494 * lookup reduction and stack usage. This is in the reclaim path, so we can't
 495 * be too greedy.
 496 */
 497#define XFS_LOOKUP_BATCH	32
 498
 499STATIC int
 500xfs_inode_ag_walk_grab(
 501	struct xfs_inode	*ip)
 502{
 503	struct inode		*inode = VFS_I(ip);
 504
 505	ASSERT(rcu_read_lock_held());
 506
 507	/*
 508	 * check for stale RCU freed inode
 509	 *
 510	 * If the inode has been reallocated, it doesn't matter if it's not in
 511	 * the AG we are walking - we are walking for writeback, so if it
 512	 * passes all the "valid inode" checks and is dirty, then we'll write
 513	 * it back anyway.  If it has been reallocated and still being
 514	 * initialised, the XFS_INEW check below will catch it.
 515	 */
 516	spin_lock(&ip->i_flags_lock);
 517	if (!ip->i_ino)
 518		goto out_unlock_noent;
 519
 520	/* avoid new or reclaimable inodes. Leave for reclaim code to flush */
 521	if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
 522		goto out_unlock_noent;
 523	spin_unlock(&ip->i_flags_lock);
 524
 525	/* nothing to sync during shutdown */
 526	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 527		return -EFSCORRUPTED;
 528
 529	/* If we can't grab the inode, it must on it's way to reclaim. */
 530	if (!igrab(inode))
 531		return -ENOENT;
 532
 533	/* inode is valid */
 534	return 0;
 535
 536out_unlock_noent:
 537	spin_unlock(&ip->i_flags_lock);
 538	return -ENOENT;
 539}
 540
 541STATIC int
 542xfs_inode_ag_walk(
 543	struct xfs_mount	*mp,
 544	struct xfs_perag	*pag,
 545	int			(*execute)(struct xfs_inode *ip, int flags,
 
 546					   void *args),
 547	int			flags,
 548	void			*args,
 549	int			tag)
 550{
 551	uint32_t		first_index;
 552	int			last_error = 0;
 553	int			skipped;
 554	int			done;
 555	int			nr_found;
 556
 557restart:
 558	done = 0;
 559	skipped = 0;
 560	first_index = 0;
 561	nr_found = 0;
 562	do {
 563		struct xfs_inode *batch[XFS_LOOKUP_BATCH];
 564		int		error = 0;
 565		int		i;
 566
 567		rcu_read_lock();
 568
 569		if (tag == -1)
 570			nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
 571					(void **)batch, first_index,
 572					XFS_LOOKUP_BATCH);
 573		else
 574			nr_found = radix_tree_gang_lookup_tag(
 575					&pag->pag_ici_root,
 576					(void **) batch, first_index,
 577					XFS_LOOKUP_BATCH, tag);
 578
 579		if (!nr_found) {
 580			rcu_read_unlock();
 581			break;
 582		}
 583
 584		/*
 585		 * Grab the inodes before we drop the lock. if we found
 586		 * nothing, nr == 0 and the loop will be skipped.
 587		 */
 588		for (i = 0; i < nr_found; i++) {
 589			struct xfs_inode *ip = batch[i];
 590
 591			if (done || xfs_inode_ag_walk_grab(ip))
 592				batch[i] = NULL;
 593
 594			/*
 595			 * Update the index for the next lookup. Catch
 596			 * overflows into the next AG range which can occur if
 597			 * we have inodes in the last block of the AG and we
 598			 * are currently pointing to the last inode.
 599			 *
 600			 * Because we may see inodes that are from the wrong AG
 601			 * due to RCU freeing and reallocation, only update the
 602			 * index if it lies in this AG. It was a race that lead
 603			 * us to see this inode, so another lookup from the
 604			 * same index will not find it again.
 605			 */
 606			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
 607				continue;
 608			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
 609			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
 610				done = 1;
 611		}
 612
 613		/* unlock now we've grabbed the inodes. */
 614		rcu_read_unlock();
 615
 616		for (i = 0; i < nr_found; i++) {
 617			if (!batch[i])
 618				continue;
 619			error = execute(batch[i], flags, args);
 620			IRELE(batch[i]);
 621			if (error == -EAGAIN) {
 622				skipped++;
 623				continue;
 624			}
 625			if (error && last_error != -EFSCORRUPTED)
 626				last_error = error;
 627		}
 628
 629		/* bail out if the filesystem is corrupted.  */
 630		if (error == -EFSCORRUPTED)
 631			break;
 632
 633		cond_resched();
 634
 635	} while (nr_found && !done);
 636
 637	if (skipped) {
 638		delay(1);
 639		goto restart;
 640	}
 641	return last_error;
 642}
 643
 644/*
 645 * Background scanning to trim post-EOF preallocated space. This is queued
 646 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
 647 */
 648STATIC void
 649xfs_queue_eofblocks(
 650	struct xfs_mount *mp)
 651{
 652	rcu_read_lock();
 653	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
 654		queue_delayed_work(mp->m_eofblocks_workqueue,
 655				   &mp->m_eofblocks_work,
 656				   msecs_to_jiffies(xfs_eofb_secs * 1000));
 657	rcu_read_unlock();
 658}
 659
 660void
 661xfs_eofblocks_worker(
 662	struct work_struct *work)
 663{
 664	struct xfs_mount *mp = container_of(to_delayed_work(work),
 665				struct xfs_mount, m_eofblocks_work);
 666	xfs_icache_free_eofblocks(mp, NULL);
 667	xfs_queue_eofblocks(mp);
 668}
 669
 670int
 671xfs_inode_ag_iterator(
 672	struct xfs_mount	*mp,
 673	int			(*execute)(struct xfs_inode *ip, int flags,
 
 674					   void *args),
 675	int			flags,
 676	void			*args)
 677{
 678	struct xfs_perag	*pag;
 679	int			error = 0;
 680	int			last_error = 0;
 681	xfs_agnumber_t		ag;
 682
 683	ag = 0;
 684	while ((pag = xfs_perag_get(mp, ag))) {
 685		ag = pag->pag_agno + 1;
 686		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1);
 687		xfs_perag_put(pag);
 688		if (error) {
 689			last_error = error;
 690			if (error == -EFSCORRUPTED)
 691				break;
 692		}
 693	}
 694	return last_error;
 695}
 696
 697int
 698xfs_inode_ag_iterator_tag(
 699	struct xfs_mount	*mp,
 700	int			(*execute)(struct xfs_inode *ip, int flags,
 
 701					   void *args),
 702	int			flags,
 703	void			*args,
 704	int			tag)
 705{
 706	struct xfs_perag	*pag;
 707	int			error = 0;
 708	int			last_error = 0;
 709	xfs_agnumber_t		ag;
 710
 711	ag = 0;
 712	while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
 713		ag = pag->pag_agno + 1;
 714		error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag);
 715		xfs_perag_put(pag);
 716		if (error) {
 717			last_error = error;
 718			if (error == -EFSCORRUPTED)
 719				break;
 720		}
 721	}
 722	return last_error;
 723}
 724
 725/*
 726 * Queue a new inode reclaim pass if there are reclaimable inodes and there
 727 * isn't a reclaim pass already in progress. By default it runs every 5s based
 728 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
 729 * tunable, but that can be done if this method proves to be ineffective or too
 730 * aggressive.
 731 */
 732static void
 733xfs_reclaim_work_queue(
 734	struct xfs_mount        *mp)
 735{
 736
 737	rcu_read_lock();
 738	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
 739		queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
 740			msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
 741	}
 742	rcu_read_unlock();
 743}
 744
 745/*
 746 * This is a fast pass over the inode cache to try to get reclaim moving on as
 747 * many inodes as possible in a short period of time. It kicks itself every few
 748 * seconds, as well as being kicked by the inode cache shrinker when memory
 749 * goes low. It scans as quickly as possible avoiding locked inodes or those
 750 * already being flushed, and once done schedules a future pass.
 751 */
 752void
 753xfs_reclaim_worker(
 754	struct work_struct *work)
 755{
 756	struct xfs_mount *mp = container_of(to_delayed_work(work),
 757					struct xfs_mount, m_reclaim_work);
 758
 759	xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
 760	xfs_reclaim_work_queue(mp);
 761}
 762
 763static void
 764__xfs_inode_set_reclaim_tag(
 765	struct xfs_perag	*pag,
 766	struct xfs_inode	*ip)
 767{
 768	radix_tree_tag_set(&pag->pag_ici_root,
 769			   XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
 770			   XFS_ICI_RECLAIM_TAG);
 771
 772	if (!pag->pag_ici_reclaimable) {
 773		/* propagate the reclaim tag up into the perag radix tree */
 774		spin_lock(&ip->i_mount->m_perag_lock);
 775		radix_tree_tag_set(&ip->i_mount->m_perag_tree,
 776				XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
 777				XFS_ICI_RECLAIM_TAG);
 778		spin_unlock(&ip->i_mount->m_perag_lock);
 779
 780		/* schedule periodic background inode reclaim */
 781		xfs_reclaim_work_queue(ip->i_mount);
 782
 783		trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
 784							-1, _RET_IP_);
 785	}
 786	pag->pag_ici_reclaimable++;
 787}
 788
 789/*
 790 * We set the inode flag atomically with the radix tree tag.
 791 * Once we get tag lookups on the radix tree, this inode flag
 792 * can go away.
 793 */
 794void
 795xfs_inode_set_reclaim_tag(
 796	xfs_inode_t	*ip)
 797{
 798	struct xfs_mount *mp = ip->i_mount;
 799	struct xfs_perag *pag;
 800
 801	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
 802	spin_lock(&pag->pag_ici_lock);
 803	spin_lock(&ip->i_flags_lock);
 804	__xfs_inode_set_reclaim_tag(pag, ip);
 805	__xfs_iflags_set(ip, XFS_IRECLAIMABLE);
 806	spin_unlock(&ip->i_flags_lock);
 807	spin_unlock(&pag->pag_ici_lock);
 808	xfs_perag_put(pag);
 809}
 810
 811STATIC void
 812__xfs_inode_clear_reclaim(
 813	xfs_perag_t	*pag,
 814	xfs_inode_t	*ip)
 815{
 816	pag->pag_ici_reclaimable--;
 817	if (!pag->pag_ici_reclaimable) {
 818		/* clear the reclaim tag from the perag radix tree */
 819		spin_lock(&ip->i_mount->m_perag_lock);
 820		radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
 821				XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
 822				XFS_ICI_RECLAIM_TAG);
 823		spin_unlock(&ip->i_mount->m_perag_lock);
 824		trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
 825							-1, _RET_IP_);
 826	}
 827}
 828
 829STATIC void
 830__xfs_inode_clear_reclaim_tag(
 831	xfs_mount_t	*mp,
 832	xfs_perag_t	*pag,
 833	xfs_inode_t	*ip)
 834{
 835	radix_tree_tag_clear(&pag->pag_ici_root,
 836			XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
 837	__xfs_inode_clear_reclaim(pag, ip);
 838}
 839
 840/*
 841 * Grab the inode for reclaim exclusively.
 842 * Return 0 if we grabbed it, non-zero otherwise.
 843 */
 844STATIC int
 845xfs_reclaim_inode_grab(
 846	struct xfs_inode	*ip,
 847	int			flags)
 848{
 849	ASSERT(rcu_read_lock_held());
 850
 851	/* quick check for stale RCU freed inode */
 852	if (!ip->i_ino)
 853		return 1;
 854
 855	/*
 856	 * If we are asked for non-blocking operation, do unlocked checks to
 857	 * see if the inode already is being flushed or in reclaim to avoid
 858	 * lock traffic.
 859	 */
 860	if ((flags & SYNC_TRYLOCK) &&
 861	    __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
 862		return 1;
 863
 864	/*
 865	 * The radix tree lock here protects a thread in xfs_iget from racing
 866	 * with us starting reclaim on the inode.  Once we have the
 867	 * XFS_IRECLAIM flag set it will not touch us.
 868	 *
 869	 * Due to RCU lookup, we may find inodes that have been freed and only
 870	 * have XFS_IRECLAIM set.  Indeed, we may see reallocated inodes that
 871	 * aren't candidates for reclaim at all, so we must check the
 872	 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
 873	 */
 874	spin_lock(&ip->i_flags_lock);
 875	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
 876	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {
 877		/* not a reclaim candidate. */
 878		spin_unlock(&ip->i_flags_lock);
 879		return 1;
 880	}
 881	__xfs_iflags_set(ip, XFS_IRECLAIM);
 882	spin_unlock(&ip->i_flags_lock);
 883	return 0;
 884}
 885
 886/*
 887 * Inodes in different states need to be treated differently. The following
 888 * table lists the inode states and the reclaim actions necessary:
 889 *
 890 *	inode state	     iflush ret		required action
 891 *      ---------------      ----------         ---------------
 892 *	bad			-		reclaim
 893 *	shutdown		EIO		unpin and reclaim
 894 *	clean, unpinned		0		reclaim
 895 *	stale, unpinned		0		reclaim
 896 *	clean, pinned(*)	0		requeue
 897 *	stale, pinned		EAGAIN		requeue
 898 *	dirty, async		-		requeue
 899 *	dirty, sync		0		reclaim
 900 *
 901 * (*) dgc: I don't think the clean, pinned state is possible but it gets
 902 * handled anyway given the order of checks implemented.
 903 *
 904 * Also, because we get the flush lock first, we know that any inode that has
 905 * been flushed delwri has had the flush completed by the time we check that
 906 * the inode is clean.
 907 *
 908 * Note that because the inode is flushed delayed write by AIL pushing, the
 909 * flush lock may already be held here and waiting on it can result in very
 910 * long latencies.  Hence for sync reclaims, where we wait on the flush lock,
 911 * the caller should push the AIL first before trying to reclaim inodes to
 912 * minimise the amount of time spent waiting.  For background relaim, we only
 913 * bother to reclaim clean inodes anyway.
 914 *
 915 * Hence the order of actions after gaining the locks should be:
 916 *	bad		=> reclaim
 917 *	shutdown	=> unpin and reclaim
 918 *	pinned, async	=> requeue
 919 *	pinned, sync	=> unpin
 920 *	stale		=> reclaim
 921 *	clean		=> reclaim
 922 *	dirty, async	=> requeue
 923 *	dirty, sync	=> flush, wait and reclaim
 924 */
 925STATIC int
 926xfs_reclaim_inode(
 927	struct xfs_inode	*ip,
 928	struct xfs_perag	*pag,
 929	int			sync_mode)
 930{
 931	struct xfs_buf		*bp = NULL;
 932	int			error;
 933
 934restart:
 935	error = 0;
 936	xfs_ilock(ip, XFS_ILOCK_EXCL);
 937	if (!xfs_iflock_nowait(ip)) {
 938		if (!(sync_mode & SYNC_WAIT))
 939			goto out;
 940		xfs_iflock(ip);
 941	}
 942
 943	if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 944		xfs_iunpin_wait(ip);
 945		xfs_iflush_abort(ip, false);
 946		goto reclaim;
 947	}
 948	if (xfs_ipincount(ip)) {
 949		if (!(sync_mode & SYNC_WAIT))
 950			goto out_ifunlock;
 951		xfs_iunpin_wait(ip);
 952	}
 953	if (xfs_iflags_test(ip, XFS_ISTALE))
 954		goto reclaim;
 955	if (xfs_inode_clean(ip))
 956		goto reclaim;
 957
 958	/*
 959	 * Never flush out dirty data during non-blocking reclaim, as it would
 960	 * just contend with AIL pushing trying to do the same job.
 961	 */
 962	if (!(sync_mode & SYNC_WAIT))
 963		goto out_ifunlock;
 964
 965	/*
 966	 * Now we have an inode that needs flushing.
 967	 *
 968	 * Note that xfs_iflush will never block on the inode buffer lock, as
 969	 * xfs_ifree_cluster() can lock the inode buffer before it locks the
 970	 * ip->i_lock, and we are doing the exact opposite here.  As a result,
 971	 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
 972	 * result in an ABBA deadlock with xfs_ifree_cluster().
 973	 *
 974	 * As xfs_ifree_cluser() must gather all inodes that are active in the
 975	 * cache to mark them stale, if we hit this case we don't actually want
 976	 * to do IO here - we want the inode marked stale so we can simply
 977	 * reclaim it.  Hence if we get an EAGAIN error here,  just unlock the
 978	 * inode, back off and try again.  Hopefully the next pass through will
 979	 * see the stale flag set on the inode.
 980	 */
 981	error = xfs_iflush(ip, &bp);
 982	if (error == -EAGAIN) {
 983		xfs_iunlock(ip, XFS_ILOCK_EXCL);
 984		/* backoff longer than in xfs_ifree_cluster */
 985		delay(2);
 986		goto restart;
 987	}
 988
 989	if (!error) {
 990		error = xfs_bwrite(bp);
 991		xfs_buf_relse(bp);
 992	}
 993
 994	xfs_iflock(ip);
 995reclaim:
 996	xfs_ifunlock(ip);
 997	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 998
 999	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1000	/*
1001	 * Remove the inode from the per-AG radix tree.
1002	 *
1003	 * Because radix_tree_delete won't complain even if the item was never
1004	 * added to the tree assert that it's been there before to catch
1005	 * problems with the inode life time early on.
1006	 */
1007	spin_lock(&pag->pag_ici_lock);
1008	if (!radix_tree_delete(&pag->pag_ici_root,
1009				XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
1010		ASSERT(0);
1011	__xfs_inode_clear_reclaim(pag, ip);
1012	spin_unlock(&pag->pag_ici_lock);
1013
1014	/*
1015	 * Here we do an (almost) spurious inode lock in order to coordinate
1016	 * with inode cache radix tree lookups.  This is because the lookup
1017	 * can reference the inodes in the cache without taking references.
1018	 *
1019	 * We make that OK here by ensuring that we wait until the inode is
1020	 * unlocked after the lookup before we go ahead and free it.
1021	 */
1022	xfs_ilock(ip, XFS_ILOCK_EXCL);
1023	xfs_qm_dqdetach(ip);
1024	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1025
1026	xfs_inode_free(ip);
1027	return error;
1028
1029out_ifunlock:
1030	xfs_ifunlock(ip);
1031out:
1032	xfs_iflags_clear(ip, XFS_IRECLAIM);
1033	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1034	/*
1035	 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1036	 * a short while. However, this just burns CPU time scanning the tree
1037	 * waiting for IO to complete and the reclaim work never goes back to
1038	 * the idle state. Instead, return 0 to let the next scheduled
1039	 * background reclaim attempt to reclaim the inode again.
1040	 */
1041	return 0;
1042}
1043
1044/*
1045 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1046 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1047 * then a shut down during filesystem unmount reclaim walk leak all the
1048 * unreclaimed inodes.
1049 */
1050STATIC int
1051xfs_reclaim_inodes_ag(
1052	struct xfs_mount	*mp,
1053	int			flags,
1054	int			*nr_to_scan)
1055{
1056	struct xfs_perag	*pag;
1057	int			error = 0;
1058	int			last_error = 0;
1059	xfs_agnumber_t		ag;
1060	int			trylock = flags & SYNC_TRYLOCK;
1061	int			skipped;
1062
1063restart:
1064	ag = 0;
1065	skipped = 0;
1066	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1067		unsigned long	first_index = 0;
1068		int		done = 0;
1069		int		nr_found = 0;
1070
1071		ag = pag->pag_agno + 1;
1072
1073		if (trylock) {
1074			if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1075				skipped++;
1076				xfs_perag_put(pag);
1077				continue;
1078			}
1079			first_index = pag->pag_ici_reclaim_cursor;
1080		} else
1081			mutex_lock(&pag->pag_ici_reclaim_lock);
1082
1083		do {
1084			struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1085			int	i;
1086
1087			rcu_read_lock();
1088			nr_found = radix_tree_gang_lookup_tag(
1089					&pag->pag_ici_root,
1090					(void **)batch, first_index,
1091					XFS_LOOKUP_BATCH,
1092					XFS_ICI_RECLAIM_TAG);
1093			if (!nr_found) {
1094				done = 1;
1095				rcu_read_unlock();
1096				break;
1097			}
1098
1099			/*
1100			 * Grab the inodes before we drop the lock. if we found
1101			 * nothing, nr == 0 and the loop will be skipped.
1102			 */
1103			for (i = 0; i < nr_found; i++) {
1104				struct xfs_inode *ip = batch[i];
1105
1106				if (done || xfs_reclaim_inode_grab(ip, flags))
1107					batch[i] = NULL;
1108
1109				/*
1110				 * Update the index for the next lookup. Catch
1111				 * overflows into the next AG range which can
1112				 * occur if we have inodes in the last block of
1113				 * the AG and we are currently pointing to the
1114				 * last inode.
1115				 *
1116				 * Because we may see inodes that are from the
1117				 * wrong AG due to RCU freeing and
1118				 * reallocation, only update the index if it
1119				 * lies in this AG. It was a race that lead us
1120				 * to see this inode, so another lookup from
1121				 * the same index will not find it again.
1122				 */
1123				if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1124								pag->pag_agno)
1125					continue;
1126				first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1127				if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1128					done = 1;
1129			}
1130
1131			/* unlock now we've grabbed the inodes. */
1132			rcu_read_unlock();
1133
1134			for (i = 0; i < nr_found; i++) {
1135				if (!batch[i])
1136					continue;
1137				error = xfs_reclaim_inode(batch[i], pag, flags);
1138				if (error && last_error != -EFSCORRUPTED)
1139					last_error = error;
1140			}
1141
1142			*nr_to_scan -= XFS_LOOKUP_BATCH;
1143
1144			cond_resched();
1145
1146		} while (nr_found && !done && *nr_to_scan > 0);
1147
1148		if (trylock && !done)
1149			pag->pag_ici_reclaim_cursor = first_index;
1150		else
1151			pag->pag_ici_reclaim_cursor = 0;
1152		mutex_unlock(&pag->pag_ici_reclaim_lock);
1153		xfs_perag_put(pag);
1154	}
1155
1156	/*
1157	 * if we skipped any AG, and we still have scan count remaining, do
1158	 * another pass this time using blocking reclaim semantics (i.e
1159	 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1160	 * ensure that when we get more reclaimers than AGs we block rather
1161	 * than spin trying to execute reclaim.
1162	 */
1163	if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1164		trylock = 0;
1165		goto restart;
1166	}
1167	return last_error;
1168}
1169
1170int
1171xfs_reclaim_inodes(
1172	xfs_mount_t	*mp,
1173	int		mode)
1174{
1175	int		nr_to_scan = INT_MAX;
1176
1177	return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1178}
1179
1180/*
1181 * Scan a certain number of inodes for reclaim.
1182 *
1183 * When called we make sure that there is a background (fast) inode reclaim in
1184 * progress, while we will throttle the speed of reclaim via doing synchronous
1185 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1186 * them to be cleaned, which we hope will not be very long due to the
1187 * background walker having already kicked the IO off on those dirty inodes.
1188 */
1189long
1190xfs_reclaim_inodes_nr(
1191	struct xfs_mount	*mp,
1192	int			nr_to_scan)
1193{
1194	/* kick background reclaimer and push the AIL */
1195	xfs_reclaim_work_queue(mp);
1196	xfs_ail_push_all(mp->m_ail);
1197
1198	return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1199}
1200
1201/*
1202 * Return the number of reclaimable inodes in the filesystem for
1203 * the shrinker to determine how much to reclaim.
1204 */
1205int
1206xfs_reclaim_inodes_count(
1207	struct xfs_mount	*mp)
1208{
1209	struct xfs_perag	*pag;
1210	xfs_agnumber_t		ag = 0;
1211	int			reclaimable = 0;
1212
1213	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1214		ag = pag->pag_agno + 1;
1215		reclaimable += pag->pag_ici_reclaimable;
1216		xfs_perag_put(pag);
1217	}
1218	return reclaimable;
1219}
1220
1221STATIC int
1222xfs_inode_match_id(
1223	struct xfs_inode	*ip,
1224	struct xfs_eofblocks	*eofb)
1225{
1226	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1227	    !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1228		return 0;
1229
1230	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1231	    !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1232		return 0;
1233
1234	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1235	    xfs_get_projid(ip) != eofb->eof_prid)
1236		return 0;
1237
1238	return 1;
1239}
1240
1241/*
1242 * A union-based inode filtering algorithm. Process the inode if any of the
1243 * criteria match. This is for global/internal scans only.
1244 */
1245STATIC int
1246xfs_inode_match_id_union(
1247	struct xfs_inode	*ip,
1248	struct xfs_eofblocks	*eofb)
1249{
1250	if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1251	    uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1252		return 1;
1253
1254	if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1255	    gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1256		return 1;
1257
1258	if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1259	    xfs_get_projid(ip) == eofb->eof_prid)
1260		return 1;
1261
1262	return 0;
1263}
1264
1265STATIC int
1266xfs_inode_free_eofblocks(
1267	struct xfs_inode	*ip,
 
1268	int			flags,
1269	void			*args)
1270{
1271	int ret;
1272	struct xfs_eofblocks *eofb = args;
1273	bool need_iolock = true;
1274	int match;
1275
1276	ASSERT(!eofb || (eofb && eofb->eof_scan_owner != 0));
1277
1278	if (!xfs_can_free_eofblocks(ip, false)) {
1279		/* inode could be preallocated or append-only */
1280		trace_xfs_inode_free_eofblocks_invalid(ip);
1281		xfs_inode_clear_eofblocks_tag(ip);
1282		return 0;
1283	}
1284
1285	/*
1286	 * If the mapping is dirty the operation can block and wait for some
1287	 * time. Unless we are waiting, skip it.
1288	 */
1289	if (!(flags & SYNC_WAIT) &&
1290	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1291		return 0;
1292
1293	if (eofb) {
1294		if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1295			match = xfs_inode_match_id_union(ip, eofb);
1296		else
1297			match = xfs_inode_match_id(ip, eofb);
1298		if (!match)
1299			return 0;
1300
1301		/* skip the inode if the file size is too small */
1302		if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1303		    XFS_ISIZE(ip) < eofb->eof_min_file_size)
1304			return 0;
1305
1306		/*
1307		 * A scan owner implies we already hold the iolock. Skip it in
1308		 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1309		 * the possibility of EAGAIN being returned.
1310		 */
1311		if (eofb->eof_scan_owner == ip->i_ino)
1312			need_iolock = false;
1313	}
1314
1315	ret = xfs_free_eofblocks(ip->i_mount, ip, need_iolock);
1316
1317	/* don't revisit the inode if we're not waiting */
1318	if (ret == -EAGAIN && !(flags & SYNC_WAIT))
1319		ret = 0;
1320
1321	return ret;
1322}
1323
1324int
1325xfs_icache_free_eofblocks(
1326	struct xfs_mount	*mp,
1327	struct xfs_eofblocks	*eofb)
1328{
1329	int flags = SYNC_TRYLOCK;
1330
1331	if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1332		flags = SYNC_WAIT;
1333
1334	return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags,
1335					 eofb, XFS_ICI_EOFBLOCKS_TAG);
1336}
1337
1338/*
1339 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1340 * multiple quotas, we don't know exactly which quota caused an allocation
1341 * failure. We make a best effort by including each quota under low free space
1342 * conditions (less than 1% free space) in the scan.
1343 */
1344int
1345xfs_inode_free_quota_eofblocks(
1346	struct xfs_inode *ip)
1347{
1348	int scan = 0;
1349	struct xfs_eofblocks eofb = {0};
1350	struct xfs_dquot *dq;
1351
1352	ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1353
1354	/*
1355	 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1356	 * can repeatedly trylock on the inode we're currently processing. We
1357	 * run a sync scan to increase effectiveness and use the union filter to
1358	 * cover all applicable quotas in a single scan.
1359	 */
1360	eofb.eof_scan_owner = ip->i_ino;
1361	eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1362
1363	if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1364		dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1365		if (dq && xfs_dquot_lowsp(dq)) {
1366			eofb.eof_uid = VFS_I(ip)->i_uid;
1367			eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1368			scan = 1;
1369		}
1370	}
1371
1372	if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1373		dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1374		if (dq && xfs_dquot_lowsp(dq)) {
1375			eofb.eof_gid = VFS_I(ip)->i_gid;
1376			eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1377			scan = 1;
1378		}
1379	}
1380
1381	if (scan)
1382		xfs_icache_free_eofblocks(ip->i_mount, &eofb);
1383
1384	return scan;
1385}
1386
1387void
1388xfs_inode_set_eofblocks_tag(
1389	xfs_inode_t	*ip)
1390{
1391	struct xfs_mount *mp = ip->i_mount;
1392	struct xfs_perag *pag;
1393	int tagged;
1394
1395	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1396	spin_lock(&pag->pag_ici_lock);
1397	trace_xfs_inode_set_eofblocks_tag(ip);
1398
1399	tagged = radix_tree_tagged(&pag->pag_ici_root,
1400				   XFS_ICI_EOFBLOCKS_TAG);
1401	radix_tree_tag_set(&pag->pag_ici_root,
1402			   XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1403			   XFS_ICI_EOFBLOCKS_TAG);
1404	if (!tagged) {
1405		/* propagate the eofblocks tag up into the perag radix tree */
1406		spin_lock(&ip->i_mount->m_perag_lock);
1407		radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1408				   XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1409				   XFS_ICI_EOFBLOCKS_TAG);
1410		spin_unlock(&ip->i_mount->m_perag_lock);
1411
1412		/* kick off background trimming */
1413		xfs_queue_eofblocks(ip->i_mount);
1414
1415		trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno,
1416					      -1, _RET_IP_);
1417	}
1418
1419	spin_unlock(&pag->pag_ici_lock);
1420	xfs_perag_put(pag);
1421}
1422
1423void
1424xfs_inode_clear_eofblocks_tag(
1425	xfs_inode_t	*ip)
1426{
1427	struct xfs_mount *mp = ip->i_mount;
1428	struct xfs_perag *pag;
1429
1430	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1431	spin_lock(&pag->pag_ici_lock);
1432	trace_xfs_inode_clear_eofblocks_tag(ip);
1433
1434	radix_tree_tag_clear(&pag->pag_ici_root,
1435			     XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1436			     XFS_ICI_EOFBLOCKS_TAG);
1437	if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) {
1438		/* clear the eofblocks tag from the perag radix tree */
1439		spin_lock(&ip->i_mount->m_perag_lock);
1440		radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1441				     XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1442				     XFS_ICI_EOFBLOCKS_TAG);
1443		spin_unlock(&ip->i_mount->m_perag_lock);
1444		trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno,
1445					       -1, _RET_IP_);
1446	}
1447
1448	spin_unlock(&pag->pag_ici_lock);
1449	xfs_perag_put(pag);
1450}
1451