1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_mount.h"
  13#include "xfs_inode.h"
  14#include "xfs_trans.h"
  15#include "xfs_trans_priv.h"
  16#include "xfs_inode_item.h"
  17#include "xfs_quota.h"
  18#include "xfs_trace.h"
  19#include "xfs_icache.h"
  20#include "xfs_bmap_util.h"
  21#include "xfs_dquot_item.h"
  22#include "xfs_dquot.h"
  23#include "xfs_reflink.h"
  24#include "xfs_ialloc.h"
  25#include "xfs_ag.h"
  26#include "xfs_log_priv.h"
 
  27
  28#include <linux/iversion.h>
  29
  30/* Radix tree tags for incore inode tree. */
  31
  32/* inode is to be reclaimed */
  33#define XFS_ICI_RECLAIM_TAG	0
  34/* Inode has speculative preallocations (posteof or cow) to clean. */
  35#define XFS_ICI_BLOCKGC_TAG	1
  36
  37/*
  38 * The goal for walking incore inodes.  These can correspond with incore inode
  39 * radix tree tags when convenient.  Avoid existing XFS_IWALK namespace.
  40 */
  41enum xfs_icwalk_goal {
  42	/* Goals directly associated with tagged inodes. */
  43	XFS_ICWALK_BLOCKGC	= XFS_ICI_BLOCKGC_TAG,
  44	XFS_ICWALK_RECLAIM	= XFS_ICI_RECLAIM_TAG,
  45};
  46
  47static int xfs_icwalk(struct xfs_mount *mp,
  48		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
  49static int xfs_icwalk_ag(struct xfs_perag *pag,
  50		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
  51
  52/*
  53 * Private inode cache walk flags for struct xfs_icwalk.  Must not
  54 * coincide with XFS_ICWALK_FLAGS_VALID.
  55 */
  56
  57/* Stop scanning after icw_scan_limit inodes. */
  58#define XFS_ICWALK_FLAG_SCAN_LIMIT	(1U << 28)
  59
  60#define XFS_ICWALK_FLAG_RECLAIM_SICK	(1U << 27)
  61#define XFS_ICWALK_FLAG_UNION		(1U << 26) /* union filter algorithm */
  62
  63#define XFS_ICWALK_PRIVATE_FLAGS	(XFS_ICWALK_FLAG_SCAN_LIMIT | \
  64					 XFS_ICWALK_FLAG_RECLAIM_SICK | \
  65					 XFS_ICWALK_FLAG_UNION)
  66
  67/*
  68 * Allocate and initialise an xfs_inode.
  69 */
  70struct xfs_inode *
  71xfs_inode_alloc(
  72	struct xfs_mount	*mp,
  73	xfs_ino_t		ino)
  74{
  75	struct xfs_inode	*ip;
  76
  77	/*
  78	 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
  79	 * and return NULL here on ENOMEM.
  80	 */
  81	ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
  82
  83	if (inode_init_always(mp->m_super, VFS_I(ip))) {
  84		kmem_cache_free(xfs_inode_cache, ip);
  85		return NULL;
  86	}
  87
  88	/* VFS doesn't initialise i_mode or i_state! */
  89	VFS_I(ip)->i_mode = 0;
  90	VFS_I(ip)->i_state = 0;
  91	mapping_set_large_folios(VFS_I(ip)->i_mapping);
  92
  93	XFS_STATS_INC(mp, vn_active);
  94	ASSERT(atomic_read(&ip->i_pincount) == 0);
  95	ASSERT(ip->i_ino == 0);
  96
  97	/* initialise the xfs inode */
  98	ip->i_ino = ino;
  99	ip->i_mount = mp;
 100	memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
 101	ip->i_cowfp = NULL;
 102	memset(&ip->i_af, 0, sizeof(ip->i_af));
 103	ip->i_af.if_format = XFS_DINODE_FMT_EXTENTS;
 104	memset(&ip->i_df, 0, sizeof(ip->i_df));
 105	ip->i_flags = 0;
 106	ip->i_delayed_blks = 0;
 107	ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
 108	ip->i_nblocks = 0;
 109	ip->i_forkoff = 0;
 110	ip->i_sick = 0;
 111	ip->i_checked = 0;
 112	INIT_WORK(&ip->i_ioend_work, xfs_end_io);
 113	INIT_LIST_HEAD(&ip->i_ioend_list);
 114	spin_lock_init(&ip->i_ioend_lock);
 115	ip->i_next_unlinked = NULLAGINO;
 116	ip->i_prev_unlinked = NULLAGINO;
 117
 118	return ip;
 119}
 120
 121STATIC void
 122xfs_inode_free_callback(
 123	struct rcu_head		*head)
 124{
 125	struct inode		*inode = container_of(head, struct inode, i_rcu);
 126	struct xfs_inode	*ip = XFS_I(inode);
 127
 128	switch (VFS_I(ip)->i_mode & S_IFMT) {
 129	case S_IFREG:
 130	case S_IFDIR:
 131	case S_IFLNK:
 132		xfs_idestroy_fork(&ip->i_df);
 133		break;
 134	}
 135
 136	xfs_ifork_zap_attr(ip);
 137
 138	if (ip->i_cowfp) {
 139		xfs_idestroy_fork(ip->i_cowfp);
 140		kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
 141	}
 142	if (ip->i_itemp) {
 143		ASSERT(!test_bit(XFS_LI_IN_AIL,
 144				 &ip->i_itemp->ili_item.li_flags));
 145		xfs_inode_item_destroy(ip);
 146		ip->i_itemp = NULL;
 147	}
 148
 149	kmem_cache_free(xfs_inode_cache, ip);
 150}
 151
 152static void
 153__xfs_inode_free(
 154	struct xfs_inode	*ip)
 155{
 156	/* asserts to verify all state is correct here */
 157	ASSERT(atomic_read(&ip->i_pincount) == 0);
 158	ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
 159	XFS_STATS_DEC(ip->i_mount, vn_active);
 160
 161	call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
 162}
 163
 164void
 165xfs_inode_free(
 166	struct xfs_inode	*ip)
 167{
 168	ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
 169
 170	/*
 171	 * Because we use RCU freeing we need to ensure the inode always
 172	 * appears to be reclaimed with an invalid inode number when in the
 173	 * free state. The ip->i_flags_lock provides the barrier against lookup
 174	 * races.
 175	 */
 176	spin_lock(&ip->i_flags_lock);
 177	ip->i_flags = XFS_IRECLAIM;
 178	ip->i_ino = 0;
 179	spin_unlock(&ip->i_flags_lock);
 180
 181	__xfs_inode_free(ip);
 182}
 183
 184/*
 185 * Queue background inode reclaim work if there are reclaimable inodes and there
 186 * isn't reclaim work already scheduled or in progress.
 187 */
 188static void
 189xfs_reclaim_work_queue(
 190	struct xfs_mount        *mp)
 191{
 192
 193	rcu_read_lock();
 194	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
 195		queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
 196			msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
 197	}
 198	rcu_read_unlock();
 199}
 200
 201/*
 202 * Background scanning to trim preallocated space. This is queued based on the
 203 * 'speculative_prealloc_lifetime' tunable (5m by default).
 204 */
 205static inline void
 206xfs_blockgc_queue(
 207	struct xfs_perag	*pag)
 208{
 209	struct xfs_mount	*mp = pag->pag_mount;
 210
 211	if (!xfs_is_blockgc_enabled(mp))
 212		return;
 213
 214	rcu_read_lock();
 215	if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
 216		queue_delayed_work(pag->pag_mount->m_blockgc_wq,
 217				   &pag->pag_blockgc_work,
 218				   msecs_to_jiffies(xfs_blockgc_secs * 1000));
 219	rcu_read_unlock();
 220}
 221
 222/* Set a tag on both the AG incore inode tree and the AG radix tree. */
 223static void
 224xfs_perag_set_inode_tag(
 225	struct xfs_perag	*pag,
 226	xfs_agino_t		agino,
 227	unsigned int		tag)
 228{
 229	struct xfs_mount	*mp = pag->pag_mount;
 230	bool			was_tagged;
 231
 232	lockdep_assert_held(&pag->pag_ici_lock);
 233
 234	was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
 235	radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
 236
 237	if (tag == XFS_ICI_RECLAIM_TAG)
 238		pag->pag_ici_reclaimable++;
 239
 240	if (was_tagged)
 241		return;
 242
 243	/* propagate the tag up into the perag radix tree */
 244	spin_lock(&mp->m_perag_lock);
 245	radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
 246	spin_unlock(&mp->m_perag_lock);
 247
 248	/* start background work */
 249	switch (tag) {
 250	case XFS_ICI_RECLAIM_TAG:
 251		xfs_reclaim_work_queue(mp);
 252		break;
 253	case XFS_ICI_BLOCKGC_TAG:
 254		xfs_blockgc_queue(pag);
 255		break;
 256	}
 257
 258	trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
 259}
 260
 261/* Clear a tag on both the AG incore inode tree and the AG radix tree. */
 262static void
 263xfs_perag_clear_inode_tag(
 264	struct xfs_perag	*pag,
 265	xfs_agino_t		agino,
 266	unsigned int		tag)
 267{
 268	struct xfs_mount	*mp = pag->pag_mount;
 269
 270	lockdep_assert_held(&pag->pag_ici_lock);
 271
 272	/*
 273	 * Reclaim can signal (with a null agino) that it cleared its own tag
 274	 * by removing the inode from the radix tree.
 275	 */
 276	if (agino != NULLAGINO)
 277		radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
 278	else
 279		ASSERT(tag == XFS_ICI_RECLAIM_TAG);
 280
 281	if (tag == XFS_ICI_RECLAIM_TAG)
 282		pag->pag_ici_reclaimable--;
 283
 284	if (radix_tree_tagged(&pag->pag_ici_root, tag))
 285		return;
 286
 287	/* clear the tag from the perag radix tree */
 288	spin_lock(&mp->m_perag_lock);
 289	radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
 290	spin_unlock(&mp->m_perag_lock);
 291
 292	trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
 293}
 294
 295/*
 296 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
 297 * part of the structure. This is made more complex by the fact we store
 298 * information about the on-disk values in the VFS inode and so we can't just
 299 * overwrite the values unconditionally. Hence we save the parameters we
 300 * need to retain across reinitialisation, and rewrite them into the VFS inode
 301 * after reinitialisation even if it fails.
 302 */
 303static int
 304xfs_reinit_inode(
 305	struct xfs_mount	*mp,
 306	struct inode		*inode)
 307{
 308	int			error;
 309	uint32_t		nlink = inode->i_nlink;
 310	uint32_t		generation = inode->i_generation;
 311	uint64_t		version = inode_peek_iversion(inode);
 312	umode_t			mode = inode->i_mode;
 313	dev_t			dev = inode->i_rdev;
 314	kuid_t			uid = inode->i_uid;
 315	kgid_t			gid = inode->i_gid;
 316
 317	error = inode_init_always(mp->m_super, inode);
 318
 319	set_nlink(inode, nlink);
 320	inode->i_generation = generation;
 321	inode_set_iversion_queried(inode, version);
 322	inode->i_mode = mode;
 323	inode->i_rdev = dev;
 324	inode->i_uid = uid;
 325	inode->i_gid = gid;
 326	mapping_set_large_folios(inode->i_mapping);
 327	return error;
 328}
 329
 330/*
 331 * Carefully nudge an inode whose VFS state has been torn down back into a
 332 * usable state.  Drops the i_flags_lock and the rcu read lock.
 333 */
 334static int
 335xfs_iget_recycle(
 336	struct xfs_perag	*pag,
 337	struct xfs_inode	*ip) __releases(&ip->i_flags_lock)
 338{
 339	struct xfs_mount	*mp = ip->i_mount;
 340	struct inode		*inode = VFS_I(ip);
 341	int			error;
 342
 343	trace_xfs_iget_recycle(ip);
 344
 345	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
 346		return -EAGAIN;
 347
 348	/*
 349	 * We need to make it look like the inode is being reclaimed to prevent
 350	 * the actual reclaim workers from stomping over us while we recycle
 351	 * the inode.  We can't clear the radix tree tag yet as it requires
 352	 * pag_ici_lock to be held exclusive.
 353	 */
 354	ip->i_flags |= XFS_IRECLAIM;
 355
 356	spin_unlock(&ip->i_flags_lock);
 357	rcu_read_unlock();
 358
 359	ASSERT(!rwsem_is_locked(&inode->i_rwsem));
 360	error = xfs_reinit_inode(mp, inode);
 361	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 362	if (error) {
 363		/*
 364		 * Re-initializing the inode failed, and we are in deep
 365		 * trouble.  Try to re-add it to the reclaim list.
 366		 */
 367		rcu_read_lock();
 368		spin_lock(&ip->i_flags_lock);
 369		ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
 370		ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
 371		spin_unlock(&ip->i_flags_lock);
 372		rcu_read_unlock();
 373
 374		trace_xfs_iget_recycle_fail(ip);
 375		return error;
 376	}
 377
 378	spin_lock(&pag->pag_ici_lock);
 379	spin_lock(&ip->i_flags_lock);
 380
 381	/*
 382	 * Clear the per-lifetime state in the inode as we are now effectively
 383	 * a new inode and need to return to the initial state before reuse
 384	 * occurs.
 385	 */
 386	ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
 387	ip->i_flags |= XFS_INEW;
 388	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
 389			XFS_ICI_RECLAIM_TAG);
 390	inode->i_state = I_NEW;
 391	spin_unlock(&ip->i_flags_lock);
 392	spin_unlock(&pag->pag_ici_lock);
 393
 394	return 0;
 395}
 396
 397/*
 398 * If we are allocating a new inode, then check what was returned is
 399 * actually a free, empty inode. If we are not allocating an inode,
 400 * then check we didn't find a free inode.
 401 *
 402 * Returns:
 403 *	0		if the inode free state matches the lookup context
 404 *	-ENOENT		if the inode is free and we are not allocating
 405 *	-EFSCORRUPTED	if there is any state mismatch at all
 406 */
 407static int
 408xfs_iget_check_free_state(
 409	struct xfs_inode	*ip,
 410	int			flags)
 411{
 412	if (flags & XFS_IGET_CREATE) {
 413		/* should be a free inode */
 414		if (VFS_I(ip)->i_mode != 0) {
 415			xfs_warn(ip->i_mount,
 416"Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
 417				ip->i_ino, VFS_I(ip)->i_mode);
 
 
 
 418			return -EFSCORRUPTED;
 419		}
 420
 421		if (ip->i_nblocks != 0) {
 422			xfs_warn(ip->i_mount,
 423"Corruption detected! Free inode 0x%llx has blocks allocated!",
 424				ip->i_ino);
 
 
 
 425			return -EFSCORRUPTED;
 426		}
 427		return 0;
 428	}
 429
 430	/* should be an allocated inode */
 431	if (VFS_I(ip)->i_mode == 0)
 432		return -ENOENT;
 433
 434	return 0;
 435}
 436
 437/* Make all pending inactivation work start immediately. */
 438static void
 439xfs_inodegc_queue_all(
 440	struct xfs_mount	*mp)
 441{
 442	struct xfs_inodegc	*gc;
 443	int			cpu;
 
 444
 445	for_each_online_cpu(cpu) {
 446		gc = per_cpu_ptr(mp->m_inodegc, cpu);
 447		if (!llist_empty(&gc->list))
 448			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 
 
 449	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 450}
 451
 452/*
 453 * Check the validity of the inode we just found it the cache
 454 */
 455static int
 456xfs_iget_cache_hit(
 457	struct xfs_perag	*pag,
 458	struct xfs_inode	*ip,
 459	xfs_ino_t		ino,
 460	int			flags,
 461	int			lock_flags) __releases(RCU)
 462{
 463	struct inode		*inode = VFS_I(ip);
 464	struct xfs_mount	*mp = ip->i_mount;
 465	int			error;
 466
 467	/*
 468	 * check for re-use of an inode within an RCU grace period due to the
 469	 * radix tree nodes not being updated yet. We monitor for this by
 470	 * setting the inode number to zero before freeing the inode structure.
 471	 * If the inode has been reallocated and set up, then the inode number
 472	 * will not match, so check for that, too.
 473	 */
 474	spin_lock(&ip->i_flags_lock);
 475	if (ip->i_ino != ino)
 476		goto out_skip;
 477
 478	/*
 479	 * If we are racing with another cache hit that is currently
 480	 * instantiating this inode or currently recycling it out of
 481	 * reclaimable state, wait for the initialisation to complete
 482	 * before continuing.
 483	 *
 484	 * If we're racing with the inactivation worker we also want to wait.
 485	 * If we're creating a new file, it's possible that the worker
 486	 * previously marked the inode as free on disk but hasn't finished
 487	 * updating the incore state yet.  The AGI buffer will be dirty and
 488	 * locked to the icreate transaction, so a synchronous push of the
 489	 * inodegc workers would result in deadlock.  For a regular iget, the
 490	 * worker is running already, so we might as well wait.
 491	 *
 492	 * XXX(hch): eventually we should do something equivalent to
 493	 *	     wait_on_inode to wait for these flags to be cleared
 494	 *	     instead of polling for it.
 495	 */
 496	if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
 497		goto out_skip;
 498
 499	if (ip->i_flags & XFS_NEED_INACTIVE) {
 500		/* Unlinked inodes cannot be re-grabbed. */
 501		if (VFS_I(ip)->i_nlink == 0) {
 502			error = -ENOENT;
 503			goto out_error;
 504		}
 505		goto out_inodegc_flush;
 506	}
 507
 508	/*
 509	 * Check the inode free state is valid. This also detects lookup
 510	 * racing with unlinks.
 511	 */
 512	error = xfs_iget_check_free_state(ip, flags);
 513	if (error)
 514		goto out_error;
 515
 516	/* Skip inodes that have no vfs state. */
 517	if ((flags & XFS_IGET_INCORE) &&
 518	    (ip->i_flags & XFS_IRECLAIMABLE))
 519		goto out_skip;
 520
 521	/* The inode fits the selection criteria; process it. */
 522	if (ip->i_flags & XFS_IRECLAIMABLE) {
 523		/* Drops i_flags_lock and RCU read lock. */
 524		error = xfs_iget_recycle(pag, ip);
 525		if (error == -EAGAIN)
 526			goto out_skip;
 527		if (error)
 528			return error;
 529	} else {
 530		/* If the VFS inode is being torn down, pause and try again. */
 531		if (!igrab(inode))
 532			goto out_skip;
 533
 534		/* We've got a live one. */
 535		spin_unlock(&ip->i_flags_lock);
 536		rcu_read_unlock();
 537		trace_xfs_iget_hit(ip);
 538	}
 539
 540	if (lock_flags != 0)
 541		xfs_ilock(ip, lock_flags);
 542
 543	if (!(flags & XFS_IGET_INCORE))
 544		xfs_iflags_clear(ip, XFS_ISTALE);
 545	XFS_STATS_INC(mp, xs_ig_found);
 546
 547	return 0;
 548
 549out_skip:
 550	trace_xfs_iget_skip(ip);
 551	XFS_STATS_INC(mp, xs_ig_frecycle);
 552	error = -EAGAIN;
 553out_error:
 554	spin_unlock(&ip->i_flags_lock);
 555	rcu_read_unlock();
 556	return error;
 557
 558out_inodegc_flush:
 559	spin_unlock(&ip->i_flags_lock);
 560	rcu_read_unlock();
 561	/*
 562	 * Do not wait for the workers, because the caller could hold an AGI
 563	 * buffer lock.  We're just going to sleep in a loop anyway.
 564	 */
 565	if (xfs_is_inodegc_enabled(mp))
 566		xfs_inodegc_queue_all(mp);
 567	return -EAGAIN;
 568}
 569
 570static int
 571xfs_iget_cache_miss(
 572	struct xfs_mount	*mp,
 573	struct xfs_perag	*pag,
 574	xfs_trans_t		*tp,
 575	xfs_ino_t		ino,
 576	struct xfs_inode	**ipp,
 577	int			flags,
 578	int			lock_flags)
 579{
 580	struct xfs_inode	*ip;
 581	int			error;
 582	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);
 583	int			iflags;
 584
 585	ip = xfs_inode_alloc(mp, ino);
 586	if (!ip)
 587		return -ENOMEM;
 588
 589	error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
 590	if (error)
 591		goto out_destroy;
 592
 593	/*
 594	 * For version 5 superblocks, if we are initialising a new inode and we
 595	 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
 596	 * simply build the new inode core with a random generation number.
 597	 *
 598	 * For version 4 (and older) superblocks, log recovery is dependent on
 599	 * the i_flushiter field being initialised from the current on-disk
 600	 * value and hence we must also read the inode off disk even when
 601	 * initializing new inodes.
 602	 */
 603	if (xfs_has_v3inodes(mp) &&
 604	    (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
 605		VFS_I(ip)->i_generation = get_random_u32();
 606	} else {
 607		struct xfs_buf		*bp;
 608
 609		error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
 610		if (error)
 611			goto out_destroy;
 612
 613		error = xfs_inode_from_disk(ip,
 614				xfs_buf_offset(bp, ip->i_imap.im_boffset));
 615		if (!error)
 616			xfs_buf_set_ref(bp, XFS_INO_REF);
 
 
 617		xfs_trans_brelse(tp, bp);
 618
 619		if (error)
 620			goto out_destroy;
 621	}
 622
 623	trace_xfs_iget_miss(ip);
 624
 625	/*
 626	 * Check the inode free state is valid. This also detects lookup
 627	 * racing with unlinks.
 628	 */
 629	error = xfs_iget_check_free_state(ip, flags);
 630	if (error)
 631		goto out_destroy;
 632
 633	/*
 634	 * Preload the radix tree so we can insert safely under the
 635	 * write spinlock. Note that we cannot sleep inside the preload
 636	 * region. Since we can be called from transaction context, don't
 637	 * recurse into the file system.
 638	 */
 639	if (radix_tree_preload(GFP_NOFS)) {
 640		error = -EAGAIN;
 641		goto out_destroy;
 642	}
 643
 644	/*
 645	 * Because the inode hasn't been added to the radix-tree yet it can't
 646	 * be found by another thread, so we can do the non-sleeping lock here.
 647	 */
 648	if (lock_flags) {
 649		if (!xfs_ilock_nowait(ip, lock_flags))
 650			BUG();
 651	}
 652
 653	/*
 654	 * These values must be set before inserting the inode into the radix
 655	 * tree as the moment it is inserted a concurrent lookup (allowed by the
 656	 * RCU locking mechanism) can find it and that lookup must see that this
 657	 * is an inode currently under construction (i.e. that XFS_INEW is set).
 658	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
 659	 * memory barrier that ensures this detection works correctly at lookup
 660	 * time.
 661	 */
 662	iflags = XFS_INEW;
 663	if (flags & XFS_IGET_DONTCACHE)
 664		d_mark_dontcache(VFS_I(ip));
 665	ip->i_udquot = NULL;
 666	ip->i_gdquot = NULL;
 667	ip->i_pdquot = NULL;
 668	xfs_iflags_set(ip, iflags);
 669
 670	/* insert the new inode */
 671	spin_lock(&pag->pag_ici_lock);
 672	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
 673	if (unlikely(error)) {
 674		WARN_ON(error != -EEXIST);
 675		XFS_STATS_INC(mp, xs_ig_dup);
 676		error = -EAGAIN;
 677		goto out_preload_end;
 678	}
 679	spin_unlock(&pag->pag_ici_lock);
 680	radix_tree_preload_end();
 681
 682	*ipp = ip;
 683	return 0;
 684
 685out_preload_end:
 686	spin_unlock(&pag->pag_ici_lock);
 687	radix_tree_preload_end();
 688	if (lock_flags)
 689		xfs_iunlock(ip, lock_flags);
 690out_destroy:
 691	__destroy_inode(VFS_I(ip));
 692	xfs_inode_free(ip);
 693	return error;
 694}
 695
 696/*
 697 * Look up an inode by number in the given file system.  The inode is looked up
 698 * in the cache held in each AG.  If the inode is found in the cache, initialise
 699 * the vfs inode if necessary.
 700 *
 701 * If it is not in core, read it in from the file system's device, add it to the
 702 * cache and initialise the vfs inode.
 703 *
 704 * The inode is locked according to the value of the lock_flags parameter.
 705 * Inode lookup is only done during metadata operations and not as part of the
 706 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
 707 */
 708int
 709xfs_iget(
 710	struct xfs_mount	*mp,
 711	struct xfs_trans	*tp,
 712	xfs_ino_t		ino,
 713	uint			flags,
 714	uint			lock_flags,
 715	struct xfs_inode	**ipp)
 716{
 717	struct xfs_inode	*ip;
 718	struct xfs_perag	*pag;
 719	xfs_agino_t		agino;
 720	int			error;
 721
 722	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
 723
 724	/* reject inode numbers outside existing AGs */
 725	if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
 726		return -EINVAL;
 727
 728	XFS_STATS_INC(mp, xs_ig_attempts);
 729
 730	/* get the perag structure and ensure that it's inode capable */
 731	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
 732	agino = XFS_INO_TO_AGINO(mp, ino);
 733
 734again:
 735	error = 0;
 736	rcu_read_lock();
 737	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
 738
 739	if (ip) {
 740		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
 741		if (error)
 742			goto out_error_or_again;
 743	} else {
 744		rcu_read_unlock();
 745		if (flags & XFS_IGET_INCORE) {
 746			error = -ENODATA;
 747			goto out_error_or_again;
 748		}
 749		XFS_STATS_INC(mp, xs_ig_missed);
 750
 751		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
 752							flags, lock_flags);
 753		if (error)
 754			goto out_error_or_again;
 755	}
 756	xfs_perag_put(pag);
 757
 758	*ipp = ip;
 759
 760	/*
 761	 * If we have a real type for an on-disk inode, we can setup the inode
 762	 * now.	 If it's a new inode being created, xfs_init_new_inode will
 763	 * handle it.
 764	 */
 765	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
 766		xfs_setup_existing_inode(ip);
 767	return 0;
 768
 769out_error_or_again:
 770	if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
 
 771		delay(1);
 772		goto again;
 773	}
 774	xfs_perag_put(pag);
 775	return error;
 776}
 777
 778/*
 779 * "Is this a cached inode that's also allocated?"
 780 *
 781 * Look up an inode by number in the given file system.  If the inode is
 782 * in cache and isn't in purgatory, return 1 if the inode is allocated
 783 * and 0 if it is not.  For all other cases (not in cache, being torn
 784 * down, etc.), return a negative error code.
 785 *
 786 * The caller has to prevent inode allocation and freeing activity,
 787 * presumably by locking the AGI buffer.   This is to ensure that an
 788 * inode cannot transition from allocated to freed until the caller is
 789 * ready to allow that.  If the inode is in an intermediate state (new,
 790 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
 791 * inode is not in the cache, -ENOENT will be returned.  The caller must
 792 * deal with these scenarios appropriately.
 793 *
 794 * This is a specialized use case for the online scrubber; if you're
 795 * reading this, you probably want xfs_iget.
 796 */
 797int
 798xfs_icache_inode_is_allocated(
 799	struct xfs_mount	*mp,
 800	struct xfs_trans	*tp,
 801	xfs_ino_t		ino,
 802	bool			*inuse)
 803{
 804	struct xfs_inode	*ip;
 805	int			error;
 806
 807	error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
 808	if (error)
 809		return error;
 810
 811	*inuse = !!(VFS_I(ip)->i_mode);
 812	xfs_irele(ip);
 813	return 0;
 814}
 815
 816/*
 817 * Grab the inode for reclaim exclusively.
 818 *
 819 * We have found this inode via a lookup under RCU, so the inode may have
 820 * already been freed, or it may be in the process of being recycled by
 821 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
 822 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
 823 * will not be set. Hence we need to check for both these flag conditions to
 824 * avoid inodes that are no longer reclaim candidates.
 825 *
 826 * Note: checking for other state flags here, under the i_flags_lock or not, is
 827 * racy and should be avoided. Those races should be resolved only after we have
 828 * ensured that we are able to reclaim this inode and the world can see that we
 829 * are going to reclaim it.
 830 *
 831 * Return true if we grabbed it, false otherwise.
 832 */
 833static bool
 834xfs_reclaim_igrab(
 835	struct xfs_inode	*ip,
 836	struct xfs_icwalk	*icw)
 837{
 838	ASSERT(rcu_read_lock_held());
 839
 840	spin_lock(&ip->i_flags_lock);
 841	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
 842	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {
 843		/* not a reclaim candidate. */
 844		spin_unlock(&ip->i_flags_lock);
 845		return false;
 846	}
 847
 848	/* Don't reclaim a sick inode unless the caller asked for it. */
 849	if (ip->i_sick &&
 850	    (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
 851		spin_unlock(&ip->i_flags_lock);
 852		return false;
 853	}
 854
 855	__xfs_iflags_set(ip, XFS_IRECLAIM);
 856	spin_unlock(&ip->i_flags_lock);
 857	return true;
 858}
 859
 860/*
 861 * Inode reclaim is non-blocking, so the default action if progress cannot be
 862 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
 863 * XFS_IRECLAIM flag.  If we are in a shutdown state, we don't care about
 864 * blocking anymore and hence we can wait for the inode to be able to reclaim
 865 * it.
 866 *
 867 * We do no IO here - if callers require inodes to be cleaned they must push the
 868 * AIL first to trigger writeback of dirty inodes.  This enables writeback to be
 869 * done in the background in a non-blocking manner, and enables memory reclaim
 870 * to make progress without blocking.
 871 */
 872static void
 873xfs_reclaim_inode(
 874	struct xfs_inode	*ip,
 875	struct xfs_perag	*pag)
 876{
 877	xfs_ino_t		ino = ip->i_ino; /* for radix_tree_delete */
 878
 879	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
 880		goto out;
 881	if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
 882		goto out_iunlock;
 883
 884	/*
 885	 * Check for log shutdown because aborting the inode can move the log
 886	 * tail and corrupt in memory state. This is fine if the log is shut
 887	 * down, but if the log is still active and only the mount is shut down
 888	 * then the in-memory log tail movement caused by the abort can be
 889	 * incorrectly propagated to disk.
 890	 */
 891	if (xlog_is_shutdown(ip->i_mount->m_log)) {
 892		xfs_iunpin_wait(ip);
 893		xfs_iflush_shutdown_abort(ip);
 894		goto reclaim;
 895	}
 896	if (xfs_ipincount(ip))
 897		goto out_clear_flush;
 898	if (!xfs_inode_clean(ip))
 899		goto out_clear_flush;
 900
 901	xfs_iflags_clear(ip, XFS_IFLUSHING);
 902reclaim:
 903	trace_xfs_inode_reclaiming(ip);
 904
 905	/*
 906	 * Because we use RCU freeing we need to ensure the inode always appears
 907	 * to be reclaimed with an invalid inode number when in the free state.
 908	 * We do this as early as possible under the ILOCK so that
 909	 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
 910	 * detect races with us here. By doing this, we guarantee that once
 911	 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
 912	 * it will see either a valid inode that will serialise correctly, or it
 913	 * will see an invalid inode that it can skip.
 914	 */
 915	spin_lock(&ip->i_flags_lock);
 916	ip->i_flags = XFS_IRECLAIM;
 917	ip->i_ino = 0;
 918	ip->i_sick = 0;
 919	ip->i_checked = 0;
 920	spin_unlock(&ip->i_flags_lock);
 921
 922	ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);
 923	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 924
 925	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
 926	/*
 927	 * Remove the inode from the per-AG radix tree.
 928	 *
 929	 * Because radix_tree_delete won't complain even if the item was never
 930	 * added to the tree assert that it's been there before to catch
 931	 * problems with the inode life time early on.
 932	 */
 933	spin_lock(&pag->pag_ici_lock);
 934	if (!radix_tree_delete(&pag->pag_ici_root,
 935				XFS_INO_TO_AGINO(ip->i_mount, ino)))
 936		ASSERT(0);
 937	xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
 938	spin_unlock(&pag->pag_ici_lock);
 939
 940	/*
 941	 * Here we do an (almost) spurious inode lock in order to coordinate
 942	 * with inode cache radix tree lookups.  This is because the lookup
 943	 * can reference the inodes in the cache without taking references.
 944	 *
 945	 * We make that OK here by ensuring that we wait until the inode is
 946	 * unlocked after the lookup before we go ahead and free it.
 947	 */
 948	xfs_ilock(ip, XFS_ILOCK_EXCL);
 949	ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
 950	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 951	ASSERT(xfs_inode_clean(ip));
 952
 953	__xfs_inode_free(ip);
 954	return;
 955
 956out_clear_flush:
 957	xfs_iflags_clear(ip, XFS_IFLUSHING);
 958out_iunlock:
 959	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 960out:
 961	xfs_iflags_clear(ip, XFS_IRECLAIM);
 962}
 963
 964/* Reclaim sick inodes if we're unmounting or the fs went down. */
 965static inline bool
 966xfs_want_reclaim_sick(
 967	struct xfs_mount	*mp)
 968{
 969	return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
 970	       xfs_is_shutdown(mp);
 971}
 972
 973void
 974xfs_reclaim_inodes(
 975	struct xfs_mount	*mp)
 976{
 977	struct xfs_icwalk	icw = {
 978		.icw_flags	= 0,
 979	};
 980
 981	if (xfs_want_reclaim_sick(mp))
 982		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
 983
 984	while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
 985		xfs_ail_push_all_sync(mp->m_ail);
 986		xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
 987	}
 988}
 989
 990/*
 991 * The shrinker infrastructure determines how many inodes we should scan for
 992 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
 993 * push the AIL here. We also want to proactively free up memory if we can to
 994 * minimise the amount of work memory reclaim has to do so we kick the
 995 * background reclaim if it isn't already scheduled.
 996 */
 997long
 998xfs_reclaim_inodes_nr(
 999	struct xfs_mount	*mp,
1000	unsigned long		nr_to_scan)
1001{
1002	struct xfs_icwalk	icw = {
1003		.icw_flags	= XFS_ICWALK_FLAG_SCAN_LIMIT,
1004		.icw_scan_limit	= min_t(unsigned long, LONG_MAX, nr_to_scan),
1005	};
1006
1007	if (xfs_want_reclaim_sick(mp))
1008		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1009
1010	/* kick background reclaimer and push the AIL */
1011	xfs_reclaim_work_queue(mp);
1012	xfs_ail_push_all(mp->m_ail);
1013
1014	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1015	return 0;
1016}
1017
1018/*
1019 * Return the number of reclaimable inodes in the filesystem for
1020 * the shrinker to determine how much to reclaim.
1021 */
1022long
1023xfs_reclaim_inodes_count(
1024	struct xfs_mount	*mp)
1025{
1026	struct xfs_perag	*pag;
1027	xfs_agnumber_t		ag = 0;
1028	long			reclaimable = 0;
1029
1030	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1031		ag = pag->pag_agno + 1;
1032		reclaimable += pag->pag_ici_reclaimable;
1033		xfs_perag_put(pag);
1034	}
1035	return reclaimable;
1036}
1037
1038STATIC bool
1039xfs_icwalk_match_id(
1040	struct xfs_inode	*ip,
1041	struct xfs_icwalk	*icw)
1042{
1043	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1044	    !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1045		return false;
1046
1047	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1048	    !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1049		return false;
1050
1051	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1052	    ip->i_projid != icw->icw_prid)
1053		return false;
1054
1055	return true;
1056}
1057
1058/*
1059 * A union-based inode filtering algorithm. Process the inode if any of the
1060 * criteria match. This is for global/internal scans only.
1061 */
1062STATIC bool
1063xfs_icwalk_match_id_union(
1064	struct xfs_inode	*ip,
1065	struct xfs_icwalk	*icw)
1066{
1067	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1068	    uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1069		return true;
1070
1071	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1072	    gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1073		return true;
1074
1075	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1076	    ip->i_projid == icw->icw_prid)
1077		return true;
1078
1079	return false;
1080}
1081
1082/*
1083 * Is this inode @ip eligible for eof/cow block reclamation, given some
1084 * filtering parameters @icw?  The inode is eligible if @icw is null or
1085 * if the predicate functions match.
1086 */
1087static bool
1088xfs_icwalk_match(
1089	struct xfs_inode	*ip,
1090	struct xfs_icwalk	*icw)
1091{
1092	bool			match;
1093
1094	if (!icw)
1095		return true;
1096
1097	if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1098		match = xfs_icwalk_match_id_union(ip, icw);
1099	else
1100		match = xfs_icwalk_match_id(ip, icw);
1101	if (!match)
1102		return false;
1103
1104	/* skip the inode if the file size is too small */
1105	if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1106	    XFS_ISIZE(ip) < icw->icw_min_file_size)
1107		return false;
1108
1109	return true;
1110}
1111
1112/*
1113 * This is a fast pass over the inode cache to try to get reclaim moving on as
1114 * many inodes as possible in a short period of time. It kicks itself every few
1115 * seconds, as well as being kicked by the inode cache shrinker when memory
1116 * goes low.
1117 */
1118void
1119xfs_reclaim_worker(
1120	struct work_struct *work)
1121{
1122	struct xfs_mount *mp = container_of(to_delayed_work(work),
1123					struct xfs_mount, m_reclaim_work);
1124
1125	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1126	xfs_reclaim_work_queue(mp);
1127}
1128
1129STATIC int
1130xfs_inode_free_eofblocks(
1131	struct xfs_inode	*ip,
1132	struct xfs_icwalk	*icw,
1133	unsigned int		*lockflags)
1134{
1135	bool			wait;
1136
1137	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1138
1139	if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1140		return 0;
1141
1142	/*
1143	 * If the mapping is dirty the operation can block and wait for some
1144	 * time. Unless we are waiting, skip it.
1145	 */
1146	if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1147		return 0;
1148
1149	if (!xfs_icwalk_match(ip, icw))
1150		return 0;
1151
1152	/*
1153	 * If the caller is waiting, return -EAGAIN to keep the background
1154	 * scanner moving and revisit the inode in a subsequent pass.
1155	 */
1156	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1157		if (wait)
1158			return -EAGAIN;
1159		return 0;
1160	}
1161	*lockflags |= XFS_IOLOCK_EXCL;
1162
1163	if (xfs_can_free_eofblocks(ip, false))
1164		return xfs_free_eofblocks(ip);
1165
1166	/* inode could be preallocated or append-only */
1167	trace_xfs_inode_free_eofblocks_invalid(ip);
1168	xfs_inode_clear_eofblocks_tag(ip);
1169	return 0;
1170}
1171
1172static void
1173xfs_blockgc_set_iflag(
1174	struct xfs_inode	*ip,
1175	unsigned long		iflag)
1176{
1177	struct xfs_mount	*mp = ip->i_mount;
1178	struct xfs_perag	*pag;
1179
1180	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1181
1182	/*
1183	 * Don't bother locking the AG and looking up in the radix trees
1184	 * if we already know that we have the tag set.
1185	 */
1186	if (ip->i_flags & iflag)
1187		return;
1188	spin_lock(&ip->i_flags_lock);
1189	ip->i_flags |= iflag;
1190	spin_unlock(&ip->i_flags_lock);
1191
1192	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1193	spin_lock(&pag->pag_ici_lock);
1194
1195	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1196			XFS_ICI_BLOCKGC_TAG);
1197
1198	spin_unlock(&pag->pag_ici_lock);
1199	xfs_perag_put(pag);
1200}
1201
1202void
1203xfs_inode_set_eofblocks_tag(
1204	xfs_inode_t	*ip)
1205{
1206	trace_xfs_inode_set_eofblocks_tag(ip);
1207	return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1208}
1209
1210static void
1211xfs_blockgc_clear_iflag(
1212	struct xfs_inode	*ip,
1213	unsigned long		iflag)
1214{
1215	struct xfs_mount	*mp = ip->i_mount;
1216	struct xfs_perag	*pag;
1217	bool			clear_tag;
1218
1219	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1220
1221	spin_lock(&ip->i_flags_lock);
1222	ip->i_flags &= ~iflag;
1223	clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1224	spin_unlock(&ip->i_flags_lock);
1225
1226	if (!clear_tag)
1227		return;
1228
1229	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1230	spin_lock(&pag->pag_ici_lock);
1231
1232	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1233			XFS_ICI_BLOCKGC_TAG);
1234
1235	spin_unlock(&pag->pag_ici_lock);
1236	xfs_perag_put(pag);
1237}
1238
1239void
1240xfs_inode_clear_eofblocks_tag(
1241	xfs_inode_t	*ip)
1242{
1243	trace_xfs_inode_clear_eofblocks_tag(ip);
1244	return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1245}
1246
1247/*
1248 * Set ourselves up to free CoW blocks from this file.  If it's already clean
1249 * then we can bail out quickly, but otherwise we must back off if the file
1250 * is undergoing some kind of write.
1251 */
1252static bool
1253xfs_prep_free_cowblocks(
1254	struct xfs_inode	*ip)
1255{
1256	/*
1257	 * Just clear the tag if we have an empty cow fork or none at all. It's
1258	 * possible the inode was fully unshared since it was originally tagged.
1259	 */
1260	if (!xfs_inode_has_cow_data(ip)) {
1261		trace_xfs_inode_free_cowblocks_invalid(ip);
1262		xfs_inode_clear_cowblocks_tag(ip);
1263		return false;
1264	}
1265
1266	/*
1267	 * If the mapping is dirty or under writeback we cannot touch the
1268	 * CoW fork.  Leave it alone if we're in the midst of a directio.
1269	 */
1270	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1271	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1272	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1273	    atomic_read(&VFS_I(ip)->i_dio_count))
1274		return false;
1275
1276	return true;
1277}
1278
1279/*
1280 * Automatic CoW Reservation Freeing
1281 *
1282 * These functions automatically garbage collect leftover CoW reservations
1283 * that were made on behalf of a cowextsize hint when we start to run out
1284 * of quota or when the reservations sit around for too long.  If the file
1285 * has dirty pages or is undergoing writeback, its CoW reservations will
1286 * be retained.
1287 *
1288 * The actual garbage collection piggybacks off the same code that runs
1289 * the speculative EOF preallocation garbage collector.
1290 */
1291STATIC int
1292xfs_inode_free_cowblocks(
1293	struct xfs_inode	*ip,
1294	struct xfs_icwalk	*icw,
1295	unsigned int		*lockflags)
1296{
1297	bool			wait;
1298	int			ret = 0;
1299
1300	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1301
1302	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1303		return 0;
1304
1305	if (!xfs_prep_free_cowblocks(ip))
1306		return 0;
1307
1308	if (!xfs_icwalk_match(ip, icw))
1309		return 0;
1310
1311	/*
1312	 * If the caller is waiting, return -EAGAIN to keep the background
1313	 * scanner moving and revisit the inode in a subsequent pass.
1314	 */
1315	if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1316	    !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1317		if (wait)
1318			return -EAGAIN;
1319		return 0;
1320	}
1321	*lockflags |= XFS_IOLOCK_EXCL;
1322
1323	if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1324		if (wait)
1325			return -EAGAIN;
1326		return 0;
1327	}
1328	*lockflags |= XFS_MMAPLOCK_EXCL;
1329
1330	/*
1331	 * Check again, nobody else should be able to dirty blocks or change
1332	 * the reflink iflag now that we have the first two locks held.
1333	 */
1334	if (xfs_prep_free_cowblocks(ip))
1335		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1336	return ret;
1337}
1338
1339void
1340xfs_inode_set_cowblocks_tag(
1341	xfs_inode_t	*ip)
1342{
1343	trace_xfs_inode_set_cowblocks_tag(ip);
1344	return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1345}
1346
1347void
1348xfs_inode_clear_cowblocks_tag(
1349	xfs_inode_t	*ip)
1350{
1351	trace_xfs_inode_clear_cowblocks_tag(ip);
1352	return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1353}
1354
1355/* Disable post-EOF and CoW block auto-reclamation. */
1356void
1357xfs_blockgc_stop(
1358	struct xfs_mount	*mp)
1359{
1360	struct xfs_perag	*pag;
1361	xfs_agnumber_t		agno;
1362
1363	if (!xfs_clear_blockgc_enabled(mp))
1364		return;
1365
1366	for_each_perag(mp, agno, pag)
1367		cancel_delayed_work_sync(&pag->pag_blockgc_work);
1368	trace_xfs_blockgc_stop(mp, __return_address);
1369}
1370
1371/* Enable post-EOF and CoW block auto-reclamation. */
1372void
1373xfs_blockgc_start(
1374	struct xfs_mount	*mp)
1375{
1376	struct xfs_perag	*pag;
1377	xfs_agnumber_t		agno;
1378
1379	if (xfs_set_blockgc_enabled(mp))
1380		return;
1381
1382	trace_xfs_blockgc_start(mp, __return_address);
1383	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1384		xfs_blockgc_queue(pag);
1385}
1386
1387/* Don't try to run block gc on an inode that's in any of these states. */
1388#define XFS_BLOCKGC_NOGRAB_IFLAGS	(XFS_INEW | \
1389					 XFS_NEED_INACTIVE | \
1390					 XFS_INACTIVATING | \
1391					 XFS_IRECLAIMABLE | \
1392					 XFS_IRECLAIM)
1393/*
1394 * Decide if the given @ip is eligible for garbage collection of speculative
1395 * preallocations, and grab it if so.  Returns true if it's ready to go or
1396 * false if we should just ignore it.
1397 */
1398static bool
1399xfs_blockgc_igrab(
1400	struct xfs_inode	*ip)
1401{
1402	struct inode		*inode = VFS_I(ip);
1403
1404	ASSERT(rcu_read_lock_held());
1405
1406	/* Check for stale RCU freed inode */
1407	spin_lock(&ip->i_flags_lock);
1408	if (!ip->i_ino)
1409		goto out_unlock_noent;
1410
1411	if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1412		goto out_unlock_noent;
1413	spin_unlock(&ip->i_flags_lock);
1414
1415	/* nothing to sync during shutdown */
1416	if (xfs_is_shutdown(ip->i_mount))
1417		return false;
1418
1419	/* If we can't grab the inode, it must on it's way to reclaim. */
1420	if (!igrab(inode))
1421		return false;
1422
1423	/* inode is valid */
1424	return true;
1425
1426out_unlock_noent:
1427	spin_unlock(&ip->i_flags_lock);
1428	return false;
1429}
1430
1431/* Scan one incore inode for block preallocations that we can remove. */
1432static int
1433xfs_blockgc_scan_inode(
1434	struct xfs_inode	*ip,
1435	struct xfs_icwalk	*icw)
1436{
1437	unsigned int		lockflags = 0;
1438	int			error;
1439
1440	error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1441	if (error)
1442		goto unlock;
1443
1444	error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1445unlock:
1446	if (lockflags)
1447		xfs_iunlock(ip, lockflags);
1448	xfs_irele(ip);
1449	return error;
1450}
1451
1452/* Background worker that trims preallocated space. */
1453void
1454xfs_blockgc_worker(
1455	struct work_struct	*work)
1456{
1457	struct xfs_perag	*pag = container_of(to_delayed_work(work),
1458					struct xfs_perag, pag_blockgc_work);
1459	struct xfs_mount	*mp = pag->pag_mount;
1460	int			error;
1461
1462	trace_xfs_blockgc_worker(mp, __return_address);
1463
1464	error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1465	if (error)
1466		xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1467				pag->pag_agno, error);
1468	xfs_blockgc_queue(pag);
1469}
1470
1471/*
1472 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1473 * and cowblocks.
1474 */
1475int
1476xfs_blockgc_free_space(
1477	struct xfs_mount	*mp,
1478	struct xfs_icwalk	*icw)
1479{
1480	int			error;
1481
1482	trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1483
1484	error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1485	if (error)
1486		return error;
1487
1488	xfs_inodegc_flush(mp);
1489	return 0;
1490}
1491
1492/*
1493 * Reclaim all the free space that we can by scheduling the background blockgc
1494 * and inodegc workers immediately and waiting for them all to clear.
1495 */
1496void
1497xfs_blockgc_flush_all(
1498	struct xfs_mount	*mp)
1499{
1500	struct xfs_perag	*pag;
1501	xfs_agnumber_t		agno;
1502
1503	trace_xfs_blockgc_flush_all(mp, __return_address);
1504
1505	/*
1506	 * For each blockgc worker, move its queue time up to now.  If it
1507	 * wasn't queued, it will not be requeued.  Then flush whatever's
1508	 * left.
1509	 */
1510	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1511		mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1512				&pag->pag_blockgc_work, 0);
1513
1514	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1515		flush_delayed_work(&pag->pag_blockgc_work);
1516
1517	xfs_inodegc_flush(mp);
1518}
1519
1520/*
1521 * Run cow/eofblocks scans on the supplied dquots.  We don't know exactly which
1522 * quota caused an allocation failure, so we make a best effort by including
1523 * each quota under low free space conditions (less than 1% free space) in the
1524 * scan.
1525 *
1526 * Callers must not hold any inode's ILOCK.  If requesting a synchronous scan
1527 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1528 * MMAPLOCK.
1529 */
1530int
1531xfs_blockgc_free_dquots(
1532	struct xfs_mount	*mp,
1533	struct xfs_dquot	*udqp,
1534	struct xfs_dquot	*gdqp,
1535	struct xfs_dquot	*pdqp,
1536	unsigned int		iwalk_flags)
1537{
1538	struct xfs_icwalk	icw = {0};
1539	bool			do_work = false;
1540
1541	if (!udqp && !gdqp && !pdqp)
1542		return 0;
1543
1544	/*
1545	 * Run a scan to free blocks using the union filter to cover all
1546	 * applicable quotas in a single scan.
1547	 */
1548	icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1549
1550	if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1551		icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1552		icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1553		do_work = true;
1554	}
1555
1556	if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1557		icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1558		icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1559		do_work = true;
1560	}
1561
1562	if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1563		icw.icw_prid = pdqp->q_id;
1564		icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1565		do_work = true;
1566	}
1567
1568	if (!do_work)
1569		return 0;
1570
1571	return xfs_blockgc_free_space(mp, &icw);
1572}
1573
1574/* Run cow/eofblocks scans on the quotas attached to the inode. */
1575int
1576xfs_blockgc_free_quota(
1577	struct xfs_inode	*ip,
1578	unsigned int		iwalk_flags)
1579{
1580	return xfs_blockgc_free_dquots(ip->i_mount,
1581			xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1582			xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1583			xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1584}
1585
1586/* XFS Inode Cache Walking Code */
1587
1588/*
1589 * The inode lookup is done in batches to keep the amount of lock traffic and
1590 * radix tree lookups to a minimum. The batch size is a trade off between
1591 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1592 * be too greedy.
1593 */
1594#define XFS_LOOKUP_BATCH	32
1595
1596
1597/*
1598 * Decide if we want to grab this inode in anticipation of doing work towards
1599 * the goal.
1600 */
1601static inline bool
1602xfs_icwalk_igrab(
1603	enum xfs_icwalk_goal	goal,
1604	struct xfs_inode	*ip,
1605	struct xfs_icwalk	*icw)
1606{
1607	switch (goal) {
1608	case XFS_ICWALK_BLOCKGC:
1609		return xfs_blockgc_igrab(ip);
1610	case XFS_ICWALK_RECLAIM:
1611		return xfs_reclaim_igrab(ip, icw);
1612	default:
1613		return false;
1614	}
1615}
1616
1617/*
1618 * Process an inode.  Each processing function must handle any state changes
1619 * made by the icwalk igrab function.  Return -EAGAIN to skip an inode.
1620 */
1621static inline int
1622xfs_icwalk_process_inode(
1623	enum xfs_icwalk_goal	goal,
1624	struct xfs_inode	*ip,
1625	struct xfs_perag	*pag,
1626	struct xfs_icwalk	*icw)
1627{
1628	int			error = 0;
1629
1630	switch (goal) {
1631	case XFS_ICWALK_BLOCKGC:
1632		error = xfs_blockgc_scan_inode(ip, icw);
1633		break;
1634	case XFS_ICWALK_RECLAIM:
1635		xfs_reclaim_inode(ip, pag);
1636		break;
1637	}
1638	return error;
1639}
1640
1641/*
1642 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1643 * process them in some manner.
1644 */
1645static int
1646xfs_icwalk_ag(
1647	struct xfs_perag	*pag,
1648	enum xfs_icwalk_goal	goal,
1649	struct xfs_icwalk	*icw)
1650{
1651	struct xfs_mount	*mp = pag->pag_mount;
1652	uint32_t		first_index;
1653	int			last_error = 0;
1654	int			skipped;
1655	bool			done;
1656	int			nr_found;
1657
1658restart:
1659	done = false;
1660	skipped = 0;
1661	if (goal == XFS_ICWALK_RECLAIM)
1662		first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1663	else
1664		first_index = 0;
1665	nr_found = 0;
1666	do {
1667		struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1668		int		error = 0;
1669		int		i;
1670
1671		rcu_read_lock();
1672
1673		nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1674				(void **) batch, first_index,
1675				XFS_LOOKUP_BATCH, goal);
1676		if (!nr_found) {
1677			done = true;
1678			rcu_read_unlock();
1679			break;
1680		}
1681
1682		/*
1683		 * Grab the inodes before we drop the lock. if we found
1684		 * nothing, nr == 0 and the loop will be skipped.
1685		 */
1686		for (i = 0; i < nr_found; i++) {
1687			struct xfs_inode *ip = batch[i];
1688
1689			if (done || !xfs_icwalk_igrab(goal, ip, icw))
1690				batch[i] = NULL;
1691
1692			/*
1693			 * Update the index for the next lookup. Catch
1694			 * overflows into the next AG range which can occur if
1695			 * we have inodes in the last block of the AG and we
1696			 * are currently pointing to the last inode.
1697			 *
1698			 * Because we may see inodes that are from the wrong AG
1699			 * due to RCU freeing and reallocation, only update the
1700			 * index if it lies in this AG. It was a race that lead
1701			 * us to see this inode, so another lookup from the
1702			 * same index will not find it again.
1703			 */
1704			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1705				continue;
1706			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1707			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1708				done = true;
1709		}
1710
1711		/* unlock now we've grabbed the inodes. */
1712		rcu_read_unlock();
1713
1714		for (i = 0; i < nr_found; i++) {
1715			if (!batch[i])
1716				continue;
1717			error = xfs_icwalk_process_inode(goal, batch[i], pag,
1718					icw);
1719			if (error == -EAGAIN) {
1720				skipped++;
1721				continue;
1722			}
1723			if (error && last_error != -EFSCORRUPTED)
1724				last_error = error;
1725		}
1726
1727		/* bail out if the filesystem is corrupted.  */
1728		if (error == -EFSCORRUPTED)
1729			break;
1730
1731		cond_resched();
1732
1733		if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1734			icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1735			if (icw->icw_scan_limit <= 0)
1736				break;
1737		}
1738	} while (nr_found && !done);
1739
1740	if (goal == XFS_ICWALK_RECLAIM) {
1741		if (done)
1742			first_index = 0;
1743		WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1744	}
1745
1746	if (skipped) {
1747		delay(1);
1748		goto restart;
1749	}
1750	return last_error;
1751}
1752
1753/* Walk all incore inodes to achieve a given goal. */
1754static int
1755xfs_icwalk(
1756	struct xfs_mount	*mp,
1757	enum xfs_icwalk_goal	goal,
1758	struct xfs_icwalk	*icw)
1759{
1760	struct xfs_perag	*pag;
1761	int			error = 0;
1762	int			last_error = 0;
1763	xfs_agnumber_t		agno;
1764
1765	for_each_perag_tag(mp, agno, pag, goal) {
1766		error = xfs_icwalk_ag(pag, goal, icw);
1767		if (error) {
1768			last_error = error;
1769			if (error == -EFSCORRUPTED) {
1770				xfs_perag_put(pag);
1771				break;
1772			}
1773		}
1774	}
1775	return last_error;
1776	BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1777}
1778
1779#ifdef DEBUG
1780static void
1781xfs_check_delalloc(
1782	struct xfs_inode	*ip,
1783	int			whichfork)
1784{
1785	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, whichfork);
1786	struct xfs_bmbt_irec	got;
1787	struct xfs_iext_cursor	icur;
1788
1789	if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1790		return;
1791	do {
1792		if (isnullstartblock(got.br_startblock)) {
1793			xfs_warn(ip->i_mount,
1794	"ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1795				ip->i_ino,
1796				whichfork == XFS_DATA_FORK ? "data" : "cow",
1797				got.br_startoff, got.br_blockcount);
1798		}
1799	} while (xfs_iext_next_extent(ifp, &icur, &got));
1800}
1801#else
1802#define xfs_check_delalloc(ip, whichfork)	do { } while (0)
1803#endif
1804
1805/* Schedule the inode for reclaim. */
1806static void
1807xfs_inodegc_set_reclaimable(
1808	struct xfs_inode	*ip)
1809{
1810	struct xfs_mount	*mp = ip->i_mount;
1811	struct xfs_perag	*pag;
1812
1813	if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1814		xfs_check_delalloc(ip, XFS_DATA_FORK);
1815		xfs_check_delalloc(ip, XFS_COW_FORK);
1816		ASSERT(0);
1817	}
1818
1819	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1820	spin_lock(&pag->pag_ici_lock);
1821	spin_lock(&ip->i_flags_lock);
1822
1823	trace_xfs_inode_set_reclaimable(ip);
1824	ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1825	ip->i_flags |= XFS_IRECLAIMABLE;
1826	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1827			XFS_ICI_RECLAIM_TAG);
1828
1829	spin_unlock(&ip->i_flags_lock);
1830	spin_unlock(&pag->pag_ici_lock);
1831	xfs_perag_put(pag);
1832}
1833
1834/*
1835 * Free all speculative preallocations and possibly even the inode itself.
1836 * This is the last chance to make changes to an otherwise unreferenced file
1837 * before incore reclamation happens.
1838 */
1839static void
1840xfs_inodegc_inactivate(
1841	struct xfs_inode	*ip)
1842{
 
 
1843	trace_xfs_inode_inactivating(ip);
1844	xfs_inactive(ip);
1845	xfs_inodegc_set_reclaimable(ip);
 
 
1846}
1847
1848void
1849xfs_inodegc_worker(
1850	struct work_struct	*work)
1851{
1852	struct xfs_inodegc	*gc = container_of(to_delayed_work(work),
1853						struct xfs_inodegc, work);
1854	struct llist_node	*node = llist_del_all(&gc->list);
1855	struct xfs_inode	*ip, *n;
 
1856	unsigned int		nofs_flag;
1857
 
 
 
 
 
 
 
 
 
1858	WRITE_ONCE(gc->items, 0);
1859
1860	if (!node)
1861		return;
1862
1863	/*
1864	 * We can allocate memory here while doing writeback on behalf of
1865	 * memory reclaim.  To avoid memory allocation deadlocks set the
1866	 * task-wide nofs context for the following operations.
1867	 */
1868	nofs_flag = memalloc_nofs_save();
1869
1870	ip = llist_entry(node, struct xfs_inode, i_gclist);
1871	trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1872
1873	WRITE_ONCE(gc->shrinker_hits, 0);
1874	llist_for_each_entry_safe(ip, n, node, i_gclist) {
 
 
1875		xfs_iflags_set(ip, XFS_INACTIVATING);
1876		xfs_inodegc_inactivate(ip);
 
 
1877	}
1878
1879	memalloc_nofs_restore(nofs_flag);
1880}
1881
1882/*
1883 * Expedite all pending inodegc work to run immediately. This does not wait for
1884 * completion of the work.
1885 */
1886void
1887xfs_inodegc_push(
1888	struct xfs_mount	*mp)
1889{
1890	if (!xfs_is_inodegc_enabled(mp))
1891		return;
1892	trace_xfs_inodegc_push(mp, __return_address);
1893	xfs_inodegc_queue_all(mp);
1894}
1895
1896/*
1897 * Force all currently queued inode inactivation work to run immediately and
1898 * wait for the work to finish.
1899 */
1900void
1901xfs_inodegc_flush(
1902	struct xfs_mount	*mp)
1903{
1904	xfs_inodegc_push(mp);
1905	trace_xfs_inodegc_flush(mp, __return_address);
1906	flush_workqueue(mp->m_inodegc_wq);
1907}
1908
1909/*
1910 * Flush all the pending work and then disable the inode inactivation background
1911 * workers and wait for them to stop.
 
1912 */
1913void
1914xfs_inodegc_stop(
1915	struct xfs_mount	*mp)
1916{
 
 
1917	if (!xfs_clear_inodegc_enabled(mp))
1918		return;
1919
 
 
 
 
 
 
 
 
 
 
1920	xfs_inodegc_queue_all(mp);
1921	drain_workqueue(mp->m_inodegc_wq);
 
 
 
1922
1923	trace_xfs_inodegc_stop(mp, __return_address);
1924}
1925
1926/*
1927 * Enable the inode inactivation background workers and schedule deferred inode
1928 * inactivation work if there is any.
 
1929 */
1930void
1931xfs_inodegc_start(
1932	struct xfs_mount	*mp)
1933{
1934	if (xfs_set_inodegc_enabled(mp))
1935		return;
1936
1937	trace_xfs_inodegc_start(mp, __return_address);
1938	xfs_inodegc_queue_all(mp);
1939}
1940
1941#ifdef CONFIG_XFS_RT
1942static inline bool
1943xfs_inodegc_want_queue_rt_file(
1944	struct xfs_inode	*ip)
1945{
1946	struct xfs_mount	*mp = ip->i_mount;
1947
1948	if (!XFS_IS_REALTIME_INODE(ip))
1949		return false;
1950
1951	if (__percpu_counter_compare(&mp->m_frextents,
1952				mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
1953				XFS_FDBLOCKS_BATCH) < 0)
1954		return true;
1955
1956	return false;
1957}
1958#else
1959# define xfs_inodegc_want_queue_rt_file(ip)	(false)
1960#endif /* CONFIG_XFS_RT */
1961
1962/*
1963 * Schedule the inactivation worker when:
1964 *
1965 *  - We've accumulated more than one inode cluster buffer's worth of inodes.
1966 *  - There is less than 5% free space left.
1967 *  - Any of the quotas for this inode are near an enforcement limit.
1968 */
1969static inline bool
1970xfs_inodegc_want_queue_work(
1971	struct xfs_inode	*ip,
1972	unsigned int		items)
1973{
1974	struct xfs_mount	*mp = ip->i_mount;
1975
1976	if (items > mp->m_ino_geo.inodes_per_cluster)
1977		return true;
1978
1979	if (__percpu_counter_compare(&mp->m_fdblocks,
1980				mp->m_low_space[XFS_LOWSP_5_PCNT],
1981				XFS_FDBLOCKS_BATCH) < 0)
1982		return true;
1983
1984	if (xfs_inodegc_want_queue_rt_file(ip))
1985		return true;
1986
1987	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
1988		return true;
1989
1990	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
1991		return true;
1992
1993	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
1994		return true;
1995
1996	return false;
1997}
1998
1999/*
2000 * Upper bound on the number of inodes in each AG that can be queued for
2001 * inactivation at any given time, to avoid monopolizing the workqueue.
2002 */
2003#define XFS_INODEGC_MAX_BACKLOG		(4 * XFS_INODES_PER_CHUNK)
2004
2005/*
2006 * Make the frontend wait for inactivations when:
2007 *
2008 *  - Memory shrinkers queued the inactivation worker and it hasn't finished.
2009 *  - The queue depth exceeds the maximum allowable percpu backlog.
2010 *
2011 * Note: If the current thread is running a transaction, we don't ever want to
2012 * wait for other transactions because that could introduce a deadlock.
 
 
2013 */
2014static inline bool
2015xfs_inodegc_want_flush_work(
2016	struct xfs_inode	*ip,
2017	unsigned int		items,
2018	unsigned int		shrinker_hits)
2019{
2020	if (current->journal_info)
2021		return false;
2022
2023	if (shrinker_hits > 0)
2024		return true;
2025
2026	if (items > XFS_INODEGC_MAX_BACKLOG)
2027		return true;
2028
2029	return false;
2030}
2031
2032/*
2033 * Queue a background inactivation worker if there are inodes that need to be
2034 * inactivated and higher level xfs code hasn't disabled the background
2035 * workers.
2036 */
2037static void
2038xfs_inodegc_queue(
2039	struct xfs_inode	*ip)
2040{
2041	struct xfs_mount	*mp = ip->i_mount;
2042	struct xfs_inodegc	*gc;
2043	int			items;
2044	unsigned int		shrinker_hits;
 
2045	unsigned long		queue_delay = 1;
2046
2047	trace_xfs_inode_set_need_inactive(ip);
2048	spin_lock(&ip->i_flags_lock);
2049	ip->i_flags |= XFS_NEED_INACTIVE;
2050	spin_unlock(&ip->i_flags_lock);
2051
2052	gc = get_cpu_ptr(mp->m_inodegc);
 
2053	llist_add(&ip->i_gclist, &gc->list);
2054	items = READ_ONCE(gc->items);
2055	WRITE_ONCE(gc->items, items + 1);
2056	shrinker_hits = READ_ONCE(gc->shrinker_hits);
2057
2058	/*
 
 
 
 
 
 
 
 
 
2059	 * We queue the work while holding the current CPU so that the work
2060	 * is scheduled to run on this CPU.
2061	 */
2062	if (!xfs_is_inodegc_enabled(mp)) {
2063		put_cpu_ptr(gc);
2064		return;
2065	}
2066
2067	if (xfs_inodegc_want_queue_work(ip, items))
2068		queue_delay = 0;
2069
2070	trace_xfs_inodegc_queue(mp, __return_address);
2071	mod_delayed_work(mp->m_inodegc_wq, &gc->work, queue_delay);
2072	put_cpu_ptr(gc);
 
2073
2074	if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2075		trace_xfs_inodegc_throttle(mp, __return_address);
2076		flush_delayed_work(&gc->work);
2077	}
2078}
2079
2080/*
2081 * Fold the dead CPU inodegc queue into the current CPUs queue.
2082 */
2083void
2084xfs_inodegc_cpu_dead(
2085	struct xfs_mount	*mp,
2086	unsigned int		dead_cpu)
2087{
2088	struct xfs_inodegc	*dead_gc, *gc;
2089	struct llist_node	*first, *last;
2090	unsigned int		count = 0;
2091
2092	dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2093	cancel_delayed_work_sync(&dead_gc->work);
2094
2095	if (llist_empty(&dead_gc->list))
2096		return;
2097
2098	first = dead_gc->list.first;
2099	last = first;
2100	while (last->next) {
2101		last = last->next;
2102		count++;
2103	}
2104	dead_gc->list.first = NULL;
2105	dead_gc->items = 0;
2106
2107	/* Add pending work to current CPU */
2108	gc = get_cpu_ptr(mp->m_inodegc);
2109	llist_add_batch(first, last, &gc->list);
2110	count += READ_ONCE(gc->items);
2111	WRITE_ONCE(gc->items, count);
2112
2113	if (xfs_is_inodegc_enabled(mp)) {
2114		trace_xfs_inodegc_queue(mp, __return_address);
2115		mod_delayed_work(mp->m_inodegc_wq, &gc->work, 0);
2116	}
2117	put_cpu_ptr(gc);
2118}
2119
2120/*
2121 * We set the inode flag atomically with the radix tree tag.  Once we get tag
2122 * lookups on the radix tree, this inode flag can go away.
2123 *
2124 * We always use background reclaim here because even if the inode is clean, it
2125 * still may be under IO and hence we have wait for IO completion to occur
2126 * before we can reclaim the inode. The background reclaim path handles this
2127 * more efficiently than we can here, so simply let background reclaim tear down
2128 * all inodes.
2129 */
2130void
2131xfs_inode_mark_reclaimable(
2132	struct xfs_inode	*ip)
2133{
2134	struct xfs_mount	*mp = ip->i_mount;
2135	bool			need_inactive;
2136
2137	XFS_STATS_INC(mp, vn_reclaim);
2138
2139	/*
2140	 * We should never get here with any of the reclaim flags already set.
2141	 */
2142	ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2143
2144	need_inactive = xfs_inode_needs_inactive(ip);
2145	if (need_inactive) {
2146		xfs_inodegc_queue(ip);
2147		return;
2148	}
2149
2150	/* Going straight to reclaim, so drop the dquots. */
2151	xfs_qm_dqdetach(ip);
2152	xfs_inodegc_set_reclaimable(ip);
2153}
2154
2155/*
2156 * Register a phony shrinker so that we can run background inodegc sooner when
2157 * there's memory pressure.  Inactivation does not itself free any memory but
2158 * it does make inodes reclaimable, which eventually frees memory.
2159 *
2160 * The count function, seek value, and batch value are crafted to trigger the
2161 * scan function during the second round of scanning.  Hopefully this means
2162 * that we reclaimed enough memory that initiating metadata transactions won't
2163 * make things worse.
2164 */
2165#define XFS_INODEGC_SHRINKER_COUNT	(1UL << DEF_PRIORITY)
2166#define XFS_INODEGC_SHRINKER_BATCH	((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2167
2168static unsigned long
2169xfs_inodegc_shrinker_count(
2170	struct shrinker		*shrink,
2171	struct shrink_control	*sc)
2172{
2173	struct xfs_mount	*mp = container_of(shrink, struct xfs_mount,
2174						   m_inodegc_shrinker);
2175	struct xfs_inodegc	*gc;
2176	int			cpu;
2177
2178	if (!xfs_is_inodegc_enabled(mp))
2179		return 0;
2180
2181	for_each_online_cpu(cpu) {
2182		gc = per_cpu_ptr(mp->m_inodegc, cpu);
2183		if (!llist_empty(&gc->list))
2184			return XFS_INODEGC_SHRINKER_COUNT;
2185	}
2186
2187	return 0;
2188}
2189
2190static unsigned long
2191xfs_inodegc_shrinker_scan(
2192	struct shrinker		*shrink,
2193	struct shrink_control	*sc)
2194{
2195	struct xfs_mount	*mp = container_of(shrink, struct xfs_mount,
2196						   m_inodegc_shrinker);
2197	struct xfs_inodegc	*gc;
2198	int			cpu;
2199	bool			no_items = true;
2200
2201	if (!xfs_is_inodegc_enabled(mp))
2202		return SHRINK_STOP;
2203
2204	trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2205
2206	for_each_online_cpu(cpu) {
2207		gc = per_cpu_ptr(mp->m_inodegc, cpu);
2208		if (!llist_empty(&gc->list)) {
2209			unsigned int	h = READ_ONCE(gc->shrinker_hits);
2210
2211			WRITE_ONCE(gc->shrinker_hits, h + 1);
2212			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2213			no_items = false;
2214		}
2215	}
2216
2217	/*
2218	 * If there are no inodes to inactivate, we don't want the shrinker
2219	 * to think there's deferred work to call us back about.
2220	 */
2221	if (no_items)
2222		return LONG_MAX;
2223
2224	return SHRINK_STOP;
2225}
2226
2227/* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2228int
2229xfs_inodegc_register_shrinker(
2230	struct xfs_mount	*mp)
2231{
2232	struct shrinker		*shrink = &mp->m_inodegc_shrinker;
 
 
 
 
2233
2234	shrink->count_objects = xfs_inodegc_shrinker_count;
2235	shrink->scan_objects = xfs_inodegc_shrinker_scan;
2236	shrink->seeks = 0;
2237	shrink->flags = SHRINKER_NONSLAB;
2238	shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2239
2240	return register_shrinker(shrink, "xfs-inodegc:%s", mp->m_super->s_id);
 
 
2241}