brintos

brintos / linux-shallow public Read only

0
0
Text · 57.4 KiB · 6b119a7 Raw
2294 lines · c
1// SPDX-License-Identifier: GPL-2.02/*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#include "xfs_health.h"28 29#include <linux/iversion.h>30 31/* Radix tree tags for incore inode tree. */32 33/* inode is to be reclaimed */34#define XFS_ICI_RECLAIM_TAG	035/* Inode has speculative preallocations (posteof or cow) to clean. */36#define XFS_ICI_BLOCKGC_TAG	137 38/*39 * The goal for walking incore inodes.  These can correspond with incore inode40 * radix tree tags when convenient.  Avoid existing XFS_IWALK namespace.41 */42enum xfs_icwalk_goal {43	/* Goals directly associated with tagged inodes. */44	XFS_ICWALK_BLOCKGC	= XFS_ICI_BLOCKGC_TAG,45	XFS_ICWALK_RECLAIM	= XFS_ICI_RECLAIM_TAG,46};47 48static int xfs_icwalk(struct xfs_mount *mp,49		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);50static int xfs_icwalk_ag(struct xfs_perag *pag,51		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);52 53/*54 * Private inode cache walk flags for struct xfs_icwalk.  Must not55 * coincide with XFS_ICWALK_FLAGS_VALID.56 */57 58/* Stop scanning after icw_scan_limit inodes. */59#define XFS_ICWALK_FLAG_SCAN_LIMIT	(1U << 28)60 61#define XFS_ICWALK_FLAG_RECLAIM_SICK	(1U << 27)62#define XFS_ICWALK_FLAG_UNION		(1U << 26) /* union filter algorithm */63 64#define XFS_ICWALK_PRIVATE_FLAGS	(XFS_ICWALK_FLAG_SCAN_LIMIT | \65					 XFS_ICWALK_FLAG_RECLAIM_SICK | \66					 XFS_ICWALK_FLAG_UNION)67 68/* Marks for the perag xarray */69#define XFS_PERAG_RECLAIM_MARK	XA_MARK_070#define XFS_PERAG_BLOCKGC_MARK	XA_MARK_171 72static inline xa_mark_t ici_tag_to_mark(unsigned int tag)73{74	if (tag == XFS_ICI_RECLAIM_TAG)75		return XFS_PERAG_RECLAIM_MARK;76	ASSERT(tag == XFS_ICI_BLOCKGC_TAG);77	return XFS_PERAG_BLOCKGC_MARK;78}79 80/*81 * Allocate and initialise an xfs_inode.82 */83struct xfs_inode *84xfs_inode_alloc(85	struct xfs_mount	*mp,86	xfs_ino_t		ino)87{88	struct xfs_inode	*ip;89 90	/*91	 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL92	 * and return NULL here on ENOMEM.93	 */94	ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);95 96	if (inode_init_always(mp->m_super, VFS_I(ip))) {97		kmem_cache_free(xfs_inode_cache, ip);98		return NULL;99	}100 101	/* VFS doesn't initialise i_mode! */102	VFS_I(ip)->i_mode = 0;103	mapping_set_folio_min_order(VFS_I(ip)->i_mapping,104				    M_IGEO(mp)->min_folio_order);105 106	XFS_STATS_INC(mp, vn_active);107	ASSERT(atomic_read(&ip->i_pincount) == 0);108	ASSERT(ip->i_ino == 0);109 110	/* initialise the xfs inode */111	ip->i_ino = ino;112	ip->i_mount = mp;113	memset(&ip->i_imap, 0, sizeof(struct xfs_imap));114	ip->i_cowfp = NULL;115	memset(&ip->i_af, 0, sizeof(ip->i_af));116	ip->i_af.if_format = XFS_DINODE_FMT_EXTENTS;117	memset(&ip->i_df, 0, sizeof(ip->i_df));118	ip->i_flags = 0;119	ip->i_delayed_blks = 0;120	ip->i_diflags2 = mp->m_ino_geo.new_diflags2;121	ip->i_nblocks = 0;122	ip->i_forkoff = 0;123	ip->i_sick = 0;124	ip->i_checked = 0;125	INIT_WORK(&ip->i_ioend_work, xfs_end_io);126	INIT_LIST_HEAD(&ip->i_ioend_list);127	spin_lock_init(&ip->i_ioend_lock);128	ip->i_next_unlinked = NULLAGINO;129	ip->i_prev_unlinked = 0;130 131	return ip;132}133 134STATIC void135xfs_inode_free_callback(136	struct rcu_head		*head)137{138	struct inode		*inode = container_of(head, struct inode, i_rcu);139	struct xfs_inode	*ip = XFS_I(inode);140 141	switch (VFS_I(ip)->i_mode & S_IFMT) {142	case S_IFREG:143	case S_IFDIR:144	case S_IFLNK:145		xfs_idestroy_fork(&ip->i_df);146		break;147	}148 149	xfs_ifork_zap_attr(ip);150 151	if (ip->i_cowfp) {152		xfs_idestroy_fork(ip->i_cowfp);153		kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);154	}155	if (ip->i_itemp) {156		ASSERT(!test_bit(XFS_LI_IN_AIL,157				 &ip->i_itemp->ili_item.li_flags));158		xfs_inode_item_destroy(ip);159		ip->i_itemp = NULL;160	}161 162	kmem_cache_free(xfs_inode_cache, ip);163}164 165static void166__xfs_inode_free(167	struct xfs_inode	*ip)168{169	/* asserts to verify all state is correct here */170	ASSERT(atomic_read(&ip->i_pincount) == 0);171	ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));172	XFS_STATS_DEC(ip->i_mount, vn_active);173 174	call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);175}176 177void178xfs_inode_free(179	struct xfs_inode	*ip)180{181	ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));182 183	/*184	 * Because we use RCU freeing we need to ensure the inode always185	 * appears to be reclaimed with an invalid inode number when in the186	 * free state. The ip->i_flags_lock provides the barrier against lookup187	 * races.188	 */189	spin_lock(&ip->i_flags_lock);190	ip->i_flags = XFS_IRECLAIM;191	ip->i_ino = 0;192	spin_unlock(&ip->i_flags_lock);193 194	__xfs_inode_free(ip);195}196 197/*198 * Queue background inode reclaim work if there are reclaimable inodes and there199 * isn't reclaim work already scheduled or in progress.200 */201static void202xfs_reclaim_work_queue(203	struct xfs_mount        *mp)204{205 206	rcu_read_lock();207	if (xa_marked(&mp->m_perags, XFS_PERAG_RECLAIM_MARK)) {208		queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,209			msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));210	}211	rcu_read_unlock();212}213 214/*215 * Background scanning to trim preallocated space. This is queued based on the216 * 'speculative_prealloc_lifetime' tunable (5m by default).217 */218static inline void219xfs_blockgc_queue(220	struct xfs_perag	*pag)221{222	struct xfs_mount	*mp = pag->pag_mount;223 224	if (!xfs_is_blockgc_enabled(mp))225		return;226 227	rcu_read_lock();228	if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))229		queue_delayed_work(pag->pag_mount->m_blockgc_wq,230				   &pag->pag_blockgc_work,231				   msecs_to_jiffies(xfs_blockgc_secs * 1000));232	rcu_read_unlock();233}234 235/* Set a tag on both the AG incore inode tree and the AG radix tree. */236static void237xfs_perag_set_inode_tag(238	struct xfs_perag	*pag,239	xfs_agino_t		agino,240	unsigned int		tag)241{242	struct xfs_mount	*mp = pag->pag_mount;243	bool			was_tagged;244 245	lockdep_assert_held(&pag->pag_ici_lock);246 247	was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);248	radix_tree_tag_set(&pag->pag_ici_root, agino, tag);249 250	if (tag == XFS_ICI_RECLAIM_TAG)251		pag->pag_ici_reclaimable++;252 253	if (was_tagged)254		return;255 256	/* propagate the tag up into the perag radix tree */257	xa_set_mark(&mp->m_perags, pag->pag_agno, ici_tag_to_mark(tag));258 259	/* start background work */260	switch (tag) {261	case XFS_ICI_RECLAIM_TAG:262		xfs_reclaim_work_queue(mp);263		break;264	case XFS_ICI_BLOCKGC_TAG:265		xfs_blockgc_queue(pag);266		break;267	}268 269	trace_xfs_perag_set_inode_tag(pag, _RET_IP_);270}271 272/* Clear a tag on both the AG incore inode tree and the AG radix tree. */273static void274xfs_perag_clear_inode_tag(275	struct xfs_perag	*pag,276	xfs_agino_t		agino,277	unsigned int		tag)278{279	struct xfs_mount	*mp = pag->pag_mount;280 281	lockdep_assert_held(&pag->pag_ici_lock);282 283	/*284	 * Reclaim can signal (with a null agino) that it cleared its own tag285	 * by removing the inode from the radix tree.286	 */287	if (agino != NULLAGINO)288		radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);289	else290		ASSERT(tag == XFS_ICI_RECLAIM_TAG);291 292	if (tag == XFS_ICI_RECLAIM_TAG)293		pag->pag_ici_reclaimable--;294 295	if (radix_tree_tagged(&pag->pag_ici_root, tag))296		return;297 298	/* clear the tag from the perag radix tree */299	xa_clear_mark(&mp->m_perags, pag->pag_agno, ici_tag_to_mark(tag));300 301	trace_xfs_perag_clear_inode_tag(pag, _RET_IP_);302}303 304/*305 * Find the next AG after @pag, or the first AG if @pag is NULL.306 */307static struct xfs_perag *308xfs_perag_grab_next_tag(309	struct xfs_mount	*mp,310	struct xfs_perag	*pag,311	int			tag)312{313	unsigned long		index = 0;314 315	if (pag) {316		index = pag->pag_agno + 1;317		xfs_perag_rele(pag);318	}319 320	rcu_read_lock();321	pag = xa_find(&mp->m_perags, &index, ULONG_MAX, ici_tag_to_mark(tag));322	if (pag) {323		trace_xfs_perag_grab_next_tag(pag, _RET_IP_);324		if (!atomic_inc_not_zero(&pag->pag_active_ref))325			pag = NULL;326	}327	rcu_read_unlock();328	return pag;329}330 331/*332 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode333 * part of the structure. This is made more complex by the fact we store334 * information about the on-disk values in the VFS inode and so we can't just335 * overwrite the values unconditionally. Hence we save the parameters we336 * need to retain across reinitialisation, and rewrite them into the VFS inode337 * after reinitialisation even if it fails.338 */339static int340xfs_reinit_inode(341	struct xfs_mount	*mp,342	struct inode		*inode)343{344	int			error;345	uint32_t		nlink = inode->i_nlink;346	uint32_t		generation = inode->i_generation;347	uint64_t		version = inode_peek_iversion(inode);348	umode_t			mode = inode->i_mode;349	dev_t			dev = inode->i_rdev;350	kuid_t			uid = inode->i_uid;351	kgid_t			gid = inode->i_gid;352	unsigned long		state = inode->i_state;353 354	error = inode_init_always(mp->m_super, inode);355 356	set_nlink(inode, nlink);357	inode->i_generation = generation;358	inode_set_iversion_queried(inode, version);359	inode->i_mode = mode;360	inode->i_rdev = dev;361	inode->i_uid = uid;362	inode->i_gid = gid;363	inode->i_state = state;364	mapping_set_folio_min_order(inode->i_mapping,365				    M_IGEO(mp)->min_folio_order);366	return error;367}368 369/*370 * Carefully nudge an inode whose VFS state has been torn down back into a371 * usable state.  Drops the i_flags_lock and the rcu read lock.372 */373static int374xfs_iget_recycle(375	struct xfs_perag	*pag,376	struct xfs_inode	*ip) __releases(&ip->i_flags_lock)377{378	struct xfs_mount	*mp = ip->i_mount;379	struct inode		*inode = VFS_I(ip);380	int			error;381 382	trace_xfs_iget_recycle(ip);383 384	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))385		return -EAGAIN;386 387	/*388	 * We need to make it look like the inode is being reclaimed to prevent389	 * the actual reclaim workers from stomping over us while we recycle390	 * the inode.  We can't clear the radix tree tag yet as it requires391	 * pag_ici_lock to be held exclusive.392	 */393	ip->i_flags |= XFS_IRECLAIM;394 395	spin_unlock(&ip->i_flags_lock);396	rcu_read_unlock();397 398	ASSERT(!rwsem_is_locked(&inode->i_rwsem));399	error = xfs_reinit_inode(mp, inode);400	xfs_iunlock(ip, XFS_ILOCK_EXCL);401	if (error) {402		/*403		 * Re-initializing the inode failed, and we are in deep404		 * trouble.  Try to re-add it to the reclaim list.405		 */406		rcu_read_lock();407		spin_lock(&ip->i_flags_lock);408		ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);409		ASSERT(ip->i_flags & XFS_IRECLAIMABLE);410		spin_unlock(&ip->i_flags_lock);411		rcu_read_unlock();412 413		trace_xfs_iget_recycle_fail(ip);414		return error;415	}416 417	spin_lock(&pag->pag_ici_lock);418	spin_lock(&ip->i_flags_lock);419 420	/*421	 * Clear the per-lifetime state in the inode as we are now effectively422	 * a new inode and need to return to the initial state before reuse423	 * occurs.424	 */425	ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;426	ip->i_flags |= XFS_INEW;427	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),428			XFS_ICI_RECLAIM_TAG);429	inode->i_state = I_NEW;430	spin_unlock(&ip->i_flags_lock);431	spin_unlock(&pag->pag_ici_lock);432 433	return 0;434}435 436/*437 * If we are allocating a new inode, then check what was returned is438 * actually a free, empty inode. If we are not allocating an inode,439 * then check we didn't find a free inode.440 *441 * Returns:442 *	0		if the inode free state matches the lookup context443 *	-ENOENT		if the inode is free and we are not allocating444 *	-EFSCORRUPTED	if there is any state mismatch at all445 */446static int447xfs_iget_check_free_state(448	struct xfs_inode	*ip,449	int			flags)450{451	if (flags & XFS_IGET_CREATE) {452		/* should be a free inode */453		if (VFS_I(ip)->i_mode != 0) {454			xfs_warn(ip->i_mount,455"Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",456				ip->i_ino, VFS_I(ip)->i_mode);457			xfs_agno_mark_sick(ip->i_mount,458					XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),459					XFS_SICK_AG_INOBT);460			return -EFSCORRUPTED;461		}462 463		if (ip->i_nblocks != 0) {464			xfs_warn(ip->i_mount,465"Corruption detected! Free inode 0x%llx has blocks allocated!",466				ip->i_ino);467			xfs_agno_mark_sick(ip->i_mount,468					XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),469					XFS_SICK_AG_INOBT);470			return -EFSCORRUPTED;471		}472		return 0;473	}474 475	/* should be an allocated inode */476	if (VFS_I(ip)->i_mode == 0)477		return -ENOENT;478 479	return 0;480}481 482/* Make all pending inactivation work start immediately. */483static bool484xfs_inodegc_queue_all(485	struct xfs_mount	*mp)486{487	struct xfs_inodegc	*gc;488	int			cpu;489	bool			ret = false;490 491	for_each_cpu(cpu, &mp->m_inodegc_cpumask) {492		gc = per_cpu_ptr(mp->m_inodegc, cpu);493		if (!llist_empty(&gc->list)) {494			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);495			ret = true;496		}497	}498 499	return ret;500}501 502/* Wait for all queued work and collect errors */503static int504xfs_inodegc_wait_all(505	struct xfs_mount	*mp)506{507	int			cpu;508	int			error = 0;509 510	flush_workqueue(mp->m_inodegc_wq);511	for_each_cpu(cpu, &mp->m_inodegc_cpumask) {512		struct xfs_inodegc	*gc;513 514		gc = per_cpu_ptr(mp->m_inodegc, cpu);515		if (gc->error && !error)516			error = gc->error;517		gc->error = 0;518	}519 520	return error;521}522 523/*524 * Check the validity of the inode we just found it the cache525 */526static int527xfs_iget_cache_hit(528	struct xfs_perag	*pag,529	struct xfs_inode	*ip,530	xfs_ino_t		ino,531	int			flags,532	int			lock_flags) __releases(RCU)533{534	struct inode		*inode = VFS_I(ip);535	struct xfs_mount	*mp = ip->i_mount;536	int			error;537 538	/*539	 * check for re-use of an inode within an RCU grace period due to the540	 * radix tree nodes not being updated yet. We monitor for this by541	 * setting the inode number to zero before freeing the inode structure.542	 * If the inode has been reallocated and set up, then the inode number543	 * will not match, so check for that, too.544	 */545	spin_lock(&ip->i_flags_lock);546	if (ip->i_ino != ino)547		goto out_skip;548 549	/*550	 * If we are racing with another cache hit that is currently551	 * instantiating this inode or currently recycling it out of552	 * reclaimable state, wait for the initialisation to complete553	 * before continuing.554	 *555	 * If we're racing with the inactivation worker we also want to wait.556	 * If we're creating a new file, it's possible that the worker557	 * previously marked the inode as free on disk but hasn't finished558	 * updating the incore state yet.  The AGI buffer will be dirty and559	 * locked to the icreate transaction, so a synchronous push of the560	 * inodegc workers would result in deadlock.  For a regular iget, the561	 * worker is running already, so we might as well wait.562	 *563	 * XXX(hch): eventually we should do something equivalent to564	 *	     wait_on_inode to wait for these flags to be cleared565	 *	     instead of polling for it.566	 */567	if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))568		goto out_skip;569 570	if (ip->i_flags & XFS_NEED_INACTIVE) {571		/* Unlinked inodes cannot be re-grabbed. */572		if (VFS_I(ip)->i_nlink == 0) {573			error = -ENOENT;574			goto out_error;575		}576		goto out_inodegc_flush;577	}578 579	/*580	 * Check the inode free state is valid. This also detects lookup581	 * racing with unlinks.582	 */583	error = xfs_iget_check_free_state(ip, flags);584	if (error)585		goto out_error;586 587	/* Skip inodes that have no vfs state. */588	if ((flags & XFS_IGET_INCORE) &&589	    (ip->i_flags & XFS_IRECLAIMABLE))590		goto out_skip;591 592	/* The inode fits the selection criteria; process it. */593	if (ip->i_flags & XFS_IRECLAIMABLE) {594		/* Drops i_flags_lock and RCU read lock. */595		error = xfs_iget_recycle(pag, ip);596		if (error == -EAGAIN)597			goto out_skip;598		if (error)599			return error;600	} else {601		/* If the VFS inode is being torn down, pause and try again. */602		if (!igrab(inode))603			goto out_skip;604 605		/* We've got a live one. */606		spin_unlock(&ip->i_flags_lock);607		rcu_read_unlock();608		trace_xfs_iget_hit(ip);609	}610 611	if (lock_flags != 0)612		xfs_ilock(ip, lock_flags);613 614	if (!(flags & XFS_IGET_INCORE))615		xfs_iflags_clear(ip, XFS_ISTALE);616	XFS_STATS_INC(mp, xs_ig_found);617 618	return 0;619 620out_skip:621	trace_xfs_iget_skip(ip);622	XFS_STATS_INC(mp, xs_ig_frecycle);623	error = -EAGAIN;624out_error:625	spin_unlock(&ip->i_flags_lock);626	rcu_read_unlock();627	return error;628 629out_inodegc_flush:630	spin_unlock(&ip->i_flags_lock);631	rcu_read_unlock();632	/*633	 * Do not wait for the workers, because the caller could hold an AGI634	 * buffer lock.  We're just going to sleep in a loop anyway.635	 */636	if (xfs_is_inodegc_enabled(mp))637		xfs_inodegc_queue_all(mp);638	return -EAGAIN;639}640 641static int642xfs_iget_cache_miss(643	struct xfs_mount	*mp,644	struct xfs_perag	*pag,645	xfs_trans_t		*tp,646	xfs_ino_t		ino,647	struct xfs_inode	**ipp,648	int			flags,649	int			lock_flags)650{651	struct xfs_inode	*ip;652	int			error;653	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);654 655	ip = xfs_inode_alloc(mp, ino);656	if (!ip)657		return -ENOMEM;658 659	error = xfs_imap(pag, tp, ip->i_ino, &ip->i_imap, flags);660	if (error)661		goto out_destroy;662 663	/*664	 * For version 5 superblocks, if we are initialising a new inode and we665	 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can666	 * simply build the new inode core with a random generation number.667	 *668	 * For version 4 (and older) superblocks, log recovery is dependent on669	 * the i_flushiter field being initialised from the current on-disk670	 * value and hence we must also read the inode off disk even when671	 * initializing new inodes.672	 */673	if (xfs_has_v3inodes(mp) &&674	    (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {675		VFS_I(ip)->i_generation = get_random_u32();676	} else {677		struct xfs_buf		*bp;678 679		error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);680		if (error)681			goto out_destroy;682 683		error = xfs_inode_from_disk(ip,684				xfs_buf_offset(bp, ip->i_imap.im_boffset));685		if (!error)686			xfs_buf_set_ref(bp, XFS_INO_REF);687		else688			xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);689		xfs_trans_brelse(tp, bp);690 691		if (error)692			goto out_destroy;693	}694 695	trace_xfs_iget_miss(ip);696 697	/*698	 * Check the inode free state is valid. This also detects lookup699	 * racing with unlinks.700	 */701	error = xfs_iget_check_free_state(ip, flags);702	if (error)703		goto out_destroy;704 705	/*706	 * Preload the radix tree so we can insert safely under the707	 * write spinlock. Note that we cannot sleep inside the preload708	 * region.709	 */710	if (radix_tree_preload(GFP_KERNEL | __GFP_NOLOCKDEP)) {711		error = -EAGAIN;712		goto out_destroy;713	}714 715	/*716	 * Because the inode hasn't been added to the radix-tree yet it can't717	 * be found by another thread, so we can do the non-sleeping lock here.718	 */719	if (lock_flags) {720		if (!xfs_ilock_nowait(ip, lock_flags))721			BUG();722	}723 724	/*725	 * These values must be set before inserting the inode into the radix726	 * tree as the moment it is inserted a concurrent lookup (allowed by the727	 * RCU locking mechanism) can find it and that lookup must see that this728	 * is an inode currently under construction (i.e. that XFS_INEW is set).729	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the730	 * memory barrier that ensures this detection works correctly at lookup731	 * time.732	 */733	if (flags & XFS_IGET_DONTCACHE)734		d_mark_dontcache(VFS_I(ip));735	ip->i_udquot = NULL;736	ip->i_gdquot = NULL;737	ip->i_pdquot = NULL;738	xfs_iflags_set(ip, XFS_INEW);739 740	/* insert the new inode */741	spin_lock(&pag->pag_ici_lock);742	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);743	if (unlikely(error)) {744		WARN_ON(error != -EEXIST);745		XFS_STATS_INC(mp, xs_ig_dup);746		error = -EAGAIN;747		goto out_preload_end;748	}749	spin_unlock(&pag->pag_ici_lock);750	radix_tree_preload_end();751 752	*ipp = ip;753	return 0;754 755out_preload_end:756	spin_unlock(&pag->pag_ici_lock);757	radix_tree_preload_end();758	if (lock_flags)759		xfs_iunlock(ip, lock_flags);760out_destroy:761	__destroy_inode(VFS_I(ip));762	xfs_inode_free(ip);763	return error;764}765 766/*767 * Look up an inode by number in the given file system.  The inode is looked up768 * in the cache held in each AG.  If the inode is found in the cache, initialise769 * the vfs inode if necessary.770 *771 * If it is not in core, read it in from the file system's device, add it to the772 * cache and initialise the vfs inode.773 *774 * The inode is locked according to the value of the lock_flags parameter.775 * Inode lookup is only done during metadata operations and not as part of the776 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.777 */778int779xfs_iget(780	struct xfs_mount	*mp,781	struct xfs_trans	*tp,782	xfs_ino_t		ino,783	uint			flags,784	uint			lock_flags,785	struct xfs_inode	**ipp)786{787	struct xfs_inode	*ip;788	struct xfs_perag	*pag;789	xfs_agino_t		agino;790	int			error;791 792	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);793 794	/* reject inode numbers outside existing AGs */795	if (!xfs_verify_ino(mp, ino))796		return -EINVAL;797 798	XFS_STATS_INC(mp, xs_ig_attempts);799 800	/* get the perag structure and ensure that it's inode capable */801	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));802	agino = XFS_INO_TO_AGINO(mp, ino);803 804again:805	error = 0;806	rcu_read_lock();807	ip = radix_tree_lookup(&pag->pag_ici_root, agino);808 809	if (ip) {810		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);811		if (error)812			goto out_error_or_again;813	} else {814		rcu_read_unlock();815		if (flags & XFS_IGET_INCORE) {816			error = -ENODATA;817			goto out_error_or_again;818		}819		XFS_STATS_INC(mp, xs_ig_missed);820 821		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,822							flags, lock_flags);823		if (error)824			goto out_error_or_again;825	}826	xfs_perag_put(pag);827 828	*ipp = ip;829 830	/*831	 * If we have a real type for an on-disk inode, we can setup the inode832	 * now.	 If it's a new inode being created, xfs_init_new_inode will833	 * handle it.834	 */835	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)836		xfs_setup_existing_inode(ip);837	return 0;838 839out_error_or_again:840	if (!(flags & (XFS_IGET_INCORE | XFS_IGET_NORETRY)) &&841	    error == -EAGAIN) {842		delay(1);843		goto again;844	}845	xfs_perag_put(pag);846	return error;847}848 849/*850 * Grab the inode for reclaim exclusively.851 *852 * We have found this inode via a lookup under RCU, so the inode may have853 * already been freed, or it may be in the process of being recycled by854 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode855 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE856 * will not be set. Hence we need to check for both these flag conditions to857 * avoid inodes that are no longer reclaim candidates.858 *859 * Note: checking for other state flags here, under the i_flags_lock or not, is860 * racy and should be avoided. Those races should be resolved only after we have861 * ensured that we are able to reclaim this inode and the world can see that we862 * are going to reclaim it.863 *864 * Return true if we grabbed it, false otherwise.865 */866static bool867xfs_reclaim_igrab(868	struct xfs_inode	*ip,869	struct xfs_icwalk	*icw)870{871	ASSERT(rcu_read_lock_held());872 873	spin_lock(&ip->i_flags_lock);874	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||875	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {876		/* not a reclaim candidate. */877		spin_unlock(&ip->i_flags_lock);878		return false;879	}880 881	/* Don't reclaim a sick inode unless the caller asked for it. */882	if (ip->i_sick &&883	    (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {884		spin_unlock(&ip->i_flags_lock);885		return false;886	}887 888	__xfs_iflags_set(ip, XFS_IRECLAIM);889	spin_unlock(&ip->i_flags_lock);890	return true;891}892 893/*894 * Inode reclaim is non-blocking, so the default action if progress cannot be895 * made is to "requeue" the inode for reclaim by unlocking it and clearing the896 * XFS_IRECLAIM flag.  If we are in a shutdown state, we don't care about897 * blocking anymore and hence we can wait for the inode to be able to reclaim898 * it.899 *900 * We do no IO here - if callers require inodes to be cleaned they must push the901 * AIL first to trigger writeback of dirty inodes.  This enables writeback to be902 * done in the background in a non-blocking manner, and enables memory reclaim903 * to make progress without blocking.904 */905static void906xfs_reclaim_inode(907	struct xfs_inode	*ip,908	struct xfs_perag	*pag)909{910	xfs_ino_t		ino = ip->i_ino; /* for radix_tree_delete */911 912	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))913		goto out;914	if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))915		goto out_iunlock;916 917	/*918	 * Check for log shutdown because aborting the inode can move the log919	 * tail and corrupt in memory state. This is fine if the log is shut920	 * down, but if the log is still active and only the mount is shut down921	 * then the in-memory log tail movement caused by the abort can be922	 * incorrectly propagated to disk.923	 */924	if (xlog_is_shutdown(ip->i_mount->m_log)) {925		xfs_iunpin_wait(ip);926		xfs_iflush_shutdown_abort(ip);927		goto reclaim;928	}929	if (xfs_ipincount(ip))930		goto out_clear_flush;931	if (!xfs_inode_clean(ip))932		goto out_clear_flush;933 934	xfs_iflags_clear(ip, XFS_IFLUSHING);935reclaim:936	trace_xfs_inode_reclaiming(ip);937 938	/*939	 * Because we use RCU freeing we need to ensure the inode always appears940	 * to be reclaimed with an invalid inode number when in the free state.941	 * We do this as early as possible under the ILOCK so that942	 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to943	 * detect races with us here. By doing this, we guarantee that once944	 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that945	 * it will see either a valid inode that will serialise correctly, or it946	 * will see an invalid inode that it can skip.947	 */948	spin_lock(&ip->i_flags_lock);949	ip->i_flags = XFS_IRECLAIM;950	ip->i_ino = 0;951	ip->i_sick = 0;952	ip->i_checked = 0;953	spin_unlock(&ip->i_flags_lock);954 955	ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);956	xfs_iunlock(ip, XFS_ILOCK_EXCL);957 958	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);959	/*960	 * Remove the inode from the per-AG radix tree.961	 *962	 * Because radix_tree_delete won't complain even if the item was never963	 * added to the tree assert that it's been there before to catch964	 * problems with the inode life time early on.965	 */966	spin_lock(&pag->pag_ici_lock);967	if (!radix_tree_delete(&pag->pag_ici_root,968				XFS_INO_TO_AGINO(ip->i_mount, ino)))969		ASSERT(0);970	xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);971	spin_unlock(&pag->pag_ici_lock);972 973	/*974	 * Here we do an (almost) spurious inode lock in order to coordinate975	 * with inode cache radix tree lookups.  This is because the lookup976	 * can reference the inodes in the cache without taking references.977	 *978	 * We make that OK here by ensuring that we wait until the inode is979	 * unlocked after the lookup before we go ahead and free it.980	 */981	xfs_ilock(ip, XFS_ILOCK_EXCL);982	ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);983	xfs_iunlock(ip, XFS_ILOCK_EXCL);984	ASSERT(xfs_inode_clean(ip));985 986	__xfs_inode_free(ip);987	return;988 989out_clear_flush:990	xfs_iflags_clear(ip, XFS_IFLUSHING);991out_iunlock:992	xfs_iunlock(ip, XFS_ILOCK_EXCL);993out:994	xfs_iflags_clear(ip, XFS_IRECLAIM);995}996 997/* Reclaim sick inodes if we're unmounting or the fs went down. */998static inline bool999xfs_want_reclaim_sick(1000	struct xfs_mount	*mp)1001{1002	return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||1003	       xfs_is_shutdown(mp);1004}1005 1006void1007xfs_reclaim_inodes(1008	struct xfs_mount	*mp)1009{1010	struct xfs_icwalk	icw = {1011		.icw_flags	= 0,1012	};1013 1014	if (xfs_want_reclaim_sick(mp))1015		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;1016 1017	while (xa_marked(&mp->m_perags, XFS_PERAG_RECLAIM_MARK)) {1018		xfs_ail_push_all_sync(mp->m_ail);1019		xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);1020	}1021}1022 1023/*1024 * The shrinker infrastructure determines how many inodes we should scan for1025 * reclaim. We want as many clean inodes ready to reclaim as possible, so we1026 * push the AIL here. We also want to proactively free up memory if we can to1027 * minimise the amount of work memory reclaim has to do so we kick the1028 * background reclaim if it isn't already scheduled.1029 */1030long1031xfs_reclaim_inodes_nr(1032	struct xfs_mount	*mp,1033	unsigned long		nr_to_scan)1034{1035	struct xfs_icwalk	icw = {1036		.icw_flags	= XFS_ICWALK_FLAG_SCAN_LIMIT,1037		.icw_scan_limit	= min_t(unsigned long, LONG_MAX, nr_to_scan),1038	};1039 1040	if (xfs_want_reclaim_sick(mp))1041		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;1042 1043	/* kick background reclaimer and push the AIL */1044	xfs_reclaim_work_queue(mp);1045	xfs_ail_push_all(mp->m_ail);1046 1047	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);1048	return 0;1049}1050 1051/*1052 * Return the number of reclaimable inodes in the filesystem for1053 * the shrinker to determine how much to reclaim.1054 */1055long1056xfs_reclaim_inodes_count(1057	struct xfs_mount	*mp)1058{1059	XA_STATE		(xas, &mp->m_perags, 0);1060	long			reclaimable = 0;1061	struct xfs_perag	*pag;1062 1063	rcu_read_lock();1064	xas_for_each_marked(&xas, pag, ULONG_MAX, XFS_PERAG_RECLAIM_MARK) {1065		trace_xfs_reclaim_inodes_count(pag, _THIS_IP_);1066		reclaimable += pag->pag_ici_reclaimable;1067	}1068	rcu_read_unlock();1069 1070	return reclaimable;1071}1072 1073STATIC bool1074xfs_icwalk_match_id(1075	struct xfs_inode	*ip,1076	struct xfs_icwalk	*icw)1077{1078	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&1079	    !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))1080		return false;1081 1082	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&1083	    !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))1084		return false;1085 1086	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&1087	    ip->i_projid != icw->icw_prid)1088		return false;1089 1090	return true;1091}1092 1093/*1094 * A union-based inode filtering algorithm. Process the inode if any of the1095 * criteria match. This is for global/internal scans only.1096 */1097STATIC bool1098xfs_icwalk_match_id_union(1099	struct xfs_inode	*ip,1100	struct xfs_icwalk	*icw)1101{1102	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&1103	    uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))1104		return true;1105 1106	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&1107	    gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))1108		return true;1109 1110	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&1111	    ip->i_projid == icw->icw_prid)1112		return true;1113 1114	return false;1115}1116 1117/*1118 * Is this inode @ip eligible for eof/cow block reclamation, given some1119 * filtering parameters @icw?  The inode is eligible if @icw is null or1120 * if the predicate functions match.1121 */1122static bool1123xfs_icwalk_match(1124	struct xfs_inode	*ip,1125	struct xfs_icwalk	*icw)1126{1127	bool			match;1128 1129	if (!icw)1130		return true;1131 1132	if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)1133		match = xfs_icwalk_match_id_union(ip, icw);1134	else1135		match = xfs_icwalk_match_id(ip, icw);1136	if (!match)1137		return false;1138 1139	/* skip the inode if the file size is too small */1140	if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&1141	    XFS_ISIZE(ip) < icw->icw_min_file_size)1142		return false;1143 1144	return true;1145}1146 1147/*1148 * This is a fast pass over the inode cache to try to get reclaim moving on as1149 * many inodes as possible in a short period of time. It kicks itself every few1150 * seconds, as well as being kicked by the inode cache shrinker when memory1151 * goes low.1152 */1153void1154xfs_reclaim_worker(1155	struct work_struct *work)1156{1157	struct xfs_mount *mp = container_of(to_delayed_work(work),1158					struct xfs_mount, m_reclaim_work);1159 1160	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);1161	xfs_reclaim_work_queue(mp);1162}1163 1164STATIC int1165xfs_inode_free_eofblocks(1166	struct xfs_inode	*ip,1167	struct xfs_icwalk	*icw,1168	unsigned int		*lockflags)1169{1170	bool			wait;1171 1172	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);1173 1174	if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))1175		return 0;1176 1177	/*1178	 * If the mapping is dirty the operation can block and wait for some1179	 * time. Unless we are waiting, skip it.1180	 */1181	if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))1182		return 0;1183 1184	if (!xfs_icwalk_match(ip, icw))1185		return 0;1186 1187	/*1188	 * If the caller is waiting, return -EAGAIN to keep the background1189	 * scanner moving and revisit the inode in a subsequent pass.1190	 */1191	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {1192		if (wait)1193			return -EAGAIN;1194		return 0;1195	}1196	*lockflags |= XFS_IOLOCK_EXCL;1197 1198	if (xfs_can_free_eofblocks(ip))1199		return xfs_free_eofblocks(ip);1200 1201	/* inode could be preallocated */1202	trace_xfs_inode_free_eofblocks_invalid(ip);1203	xfs_inode_clear_eofblocks_tag(ip);1204	return 0;1205}1206 1207static void1208xfs_blockgc_set_iflag(1209	struct xfs_inode	*ip,1210	unsigned long		iflag)1211{1212	struct xfs_mount	*mp = ip->i_mount;1213	struct xfs_perag	*pag;1214 1215	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);1216 1217	/*1218	 * Don't bother locking the AG and looking up in the radix trees1219	 * if we already know that we have the tag set.1220	 */1221	if (ip->i_flags & iflag)1222		return;1223	spin_lock(&ip->i_flags_lock);1224	ip->i_flags |= iflag;1225	spin_unlock(&ip->i_flags_lock);1226 1227	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));1228	spin_lock(&pag->pag_ici_lock);1229 1230	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),1231			XFS_ICI_BLOCKGC_TAG);1232 1233	spin_unlock(&pag->pag_ici_lock);1234	xfs_perag_put(pag);1235}1236 1237void1238xfs_inode_set_eofblocks_tag(1239	xfs_inode_t	*ip)1240{1241	trace_xfs_inode_set_eofblocks_tag(ip);1242	return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);1243}1244 1245static void1246xfs_blockgc_clear_iflag(1247	struct xfs_inode	*ip,1248	unsigned long		iflag)1249{1250	struct xfs_mount	*mp = ip->i_mount;1251	struct xfs_perag	*pag;1252	bool			clear_tag;1253 1254	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);1255 1256	spin_lock(&ip->i_flags_lock);1257	ip->i_flags &= ~iflag;1258	clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;1259	spin_unlock(&ip->i_flags_lock);1260 1261	if (!clear_tag)1262		return;1263 1264	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));1265	spin_lock(&pag->pag_ici_lock);1266 1267	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),1268			XFS_ICI_BLOCKGC_TAG);1269 1270	spin_unlock(&pag->pag_ici_lock);1271	xfs_perag_put(pag);1272}1273 1274void1275xfs_inode_clear_eofblocks_tag(1276	xfs_inode_t	*ip)1277{1278	trace_xfs_inode_clear_eofblocks_tag(ip);1279	return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);1280}1281 1282/*1283 * Prepare to free COW fork blocks from an inode.1284 */1285static bool1286xfs_prep_free_cowblocks(1287	struct xfs_inode	*ip,1288	struct xfs_icwalk	*icw)1289{1290	bool			sync;1291 1292	sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);1293 1294	/*1295	 * Just clear the tag if we have an empty cow fork or none at all. It's1296	 * possible the inode was fully unshared since it was originally tagged.1297	 */1298	if (!xfs_inode_has_cow_data(ip)) {1299		trace_xfs_inode_free_cowblocks_invalid(ip);1300		xfs_inode_clear_cowblocks_tag(ip);1301		return false;1302	}1303 1304	/*1305	 * A cowblocks trim of an inode can have a significant effect on1306	 * fragmentation even when a reasonable COW extent size hint is set.1307	 * Therefore, we prefer to not process cowblocks unless they are clean1308	 * and idle. We can never process a cowblocks inode that is dirty or has1309	 * in-flight I/O under any circumstances, because outstanding writeback1310	 * or dio expects targeted COW fork blocks exist through write1311	 * completion where they can be remapped into the data fork.1312	 *1313	 * Therefore, the heuristic used here is to never process inodes1314	 * currently opened for write from background (i.e. non-sync) scans. For1315	 * sync scans, use the pagecache/dio state of the inode to ensure we1316	 * never free COW fork blocks out from under pending I/O.1317	 */1318	if (!sync && inode_is_open_for_write(VFS_I(ip)))1319		return false;1320	return xfs_can_free_cowblocks(ip);1321}1322 1323/*1324 * Automatic CoW Reservation Freeing1325 *1326 * These functions automatically garbage collect leftover CoW reservations1327 * that were made on behalf of a cowextsize hint when we start to run out1328 * of quota or when the reservations sit around for too long.  If the file1329 * has dirty pages or is undergoing writeback, its CoW reservations will1330 * be retained.1331 *1332 * The actual garbage collection piggybacks off the same code that runs1333 * the speculative EOF preallocation garbage collector.1334 */1335STATIC int1336xfs_inode_free_cowblocks(1337	struct xfs_inode	*ip,1338	struct xfs_icwalk	*icw,1339	unsigned int		*lockflags)1340{1341	bool			wait;1342	int			ret = 0;1343 1344	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);1345 1346	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))1347		return 0;1348 1349	if (!xfs_prep_free_cowblocks(ip, icw))1350		return 0;1351 1352	if (!xfs_icwalk_match(ip, icw))1353		return 0;1354 1355	/*1356	 * If the caller is waiting, return -EAGAIN to keep the background1357	 * scanner moving and revisit the inode in a subsequent pass.1358	 */1359	if (!(*lockflags & XFS_IOLOCK_EXCL) &&1360	    !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {1361		if (wait)1362			return -EAGAIN;1363		return 0;1364	}1365	*lockflags |= XFS_IOLOCK_EXCL;1366 1367	if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {1368		if (wait)1369			return -EAGAIN;1370		return 0;1371	}1372	*lockflags |= XFS_MMAPLOCK_EXCL;1373 1374	/*1375	 * Check again, nobody else should be able to dirty blocks or change1376	 * the reflink iflag now that we have the first two locks held.1377	 */1378	if (xfs_prep_free_cowblocks(ip, icw))1379		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);1380	return ret;1381}1382 1383void1384xfs_inode_set_cowblocks_tag(1385	xfs_inode_t	*ip)1386{1387	trace_xfs_inode_set_cowblocks_tag(ip);1388	return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);1389}1390 1391void1392xfs_inode_clear_cowblocks_tag(1393	xfs_inode_t	*ip)1394{1395	trace_xfs_inode_clear_cowblocks_tag(ip);1396	return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);1397}1398 1399/* Disable post-EOF and CoW block auto-reclamation. */1400void1401xfs_blockgc_stop(1402	struct xfs_mount	*mp)1403{1404	struct xfs_perag	*pag;1405	xfs_agnumber_t		agno;1406 1407	if (!xfs_clear_blockgc_enabled(mp))1408		return;1409 1410	for_each_perag(mp, agno, pag)1411		cancel_delayed_work_sync(&pag->pag_blockgc_work);1412	trace_xfs_blockgc_stop(mp, __return_address);1413}1414 1415/* Enable post-EOF and CoW block auto-reclamation. */1416void1417xfs_blockgc_start(1418	struct xfs_mount	*mp)1419{1420	struct xfs_perag	*pag = NULL;1421 1422	if (xfs_set_blockgc_enabled(mp))1423		return;1424 1425	trace_xfs_blockgc_start(mp, __return_address);1426	while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))1427		xfs_blockgc_queue(pag);1428}1429 1430/* Don't try to run block gc on an inode that's in any of these states. */1431#define XFS_BLOCKGC_NOGRAB_IFLAGS	(XFS_INEW | \1432					 XFS_NEED_INACTIVE | \1433					 XFS_INACTIVATING | \1434					 XFS_IRECLAIMABLE | \1435					 XFS_IRECLAIM)1436/*1437 * Decide if the given @ip is eligible for garbage collection of speculative1438 * preallocations, and grab it if so.  Returns true if it's ready to go or1439 * false if we should just ignore it.1440 */1441static bool1442xfs_blockgc_igrab(1443	struct xfs_inode	*ip)1444{1445	struct inode		*inode = VFS_I(ip);1446 1447	ASSERT(rcu_read_lock_held());1448 1449	/* Check for stale RCU freed inode */1450	spin_lock(&ip->i_flags_lock);1451	if (!ip->i_ino)1452		goto out_unlock_noent;1453 1454	if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)1455		goto out_unlock_noent;1456	spin_unlock(&ip->i_flags_lock);1457 1458	/* nothing to sync during shutdown */1459	if (xfs_is_shutdown(ip->i_mount))1460		return false;1461 1462	/* If we can't grab the inode, it must on it's way to reclaim. */1463	if (!igrab(inode))1464		return false;1465 1466	/* inode is valid */1467	return true;1468 1469out_unlock_noent:1470	spin_unlock(&ip->i_flags_lock);1471	return false;1472}1473 1474/* Scan one incore inode for block preallocations that we can remove. */1475static int1476xfs_blockgc_scan_inode(1477	struct xfs_inode	*ip,1478	struct xfs_icwalk	*icw)1479{1480	unsigned int		lockflags = 0;1481	int			error;1482 1483	error = xfs_inode_free_eofblocks(ip, icw, &lockflags);1484	if (error)1485		goto unlock;1486 1487	error = xfs_inode_free_cowblocks(ip, icw, &lockflags);1488unlock:1489	if (lockflags)1490		xfs_iunlock(ip, lockflags);1491	xfs_irele(ip);1492	return error;1493}1494 1495/* Background worker that trims preallocated space. */1496void1497xfs_blockgc_worker(1498	struct work_struct	*work)1499{1500	struct xfs_perag	*pag = container_of(to_delayed_work(work),1501					struct xfs_perag, pag_blockgc_work);1502	struct xfs_mount	*mp = pag->pag_mount;1503	int			error;1504 1505	trace_xfs_blockgc_worker(mp, __return_address);1506 1507	error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);1508	if (error)1509		xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",1510				pag->pag_agno, error);1511	xfs_blockgc_queue(pag);1512}1513 1514/*1515 * Try to free space in the filesystem by purging inactive inodes, eofblocks1516 * and cowblocks.1517 */1518int1519xfs_blockgc_free_space(1520	struct xfs_mount	*mp,1521	struct xfs_icwalk	*icw)1522{1523	int			error;1524 1525	trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);1526 1527	error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);1528	if (error)1529		return error;1530 1531	return xfs_inodegc_flush(mp);1532}1533 1534/*1535 * Reclaim all the free space that we can by scheduling the background blockgc1536 * and inodegc workers immediately and waiting for them all to clear.1537 */1538int1539xfs_blockgc_flush_all(1540	struct xfs_mount	*mp)1541{1542	struct xfs_perag	*pag = NULL;1543 1544	trace_xfs_blockgc_flush_all(mp, __return_address);1545 1546	/*1547	 * For each blockgc worker, move its queue time up to now.  If it wasn't1548	 * queued, it will not be requeued.  Then flush whatever is left.1549	 */1550	while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))1551		mod_delayed_work(pag->pag_mount->m_blockgc_wq,1552				&pag->pag_blockgc_work, 0);1553 1554	while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))1555		flush_delayed_work(&pag->pag_blockgc_work);1556 1557	return xfs_inodegc_flush(mp);1558}1559 1560/*1561 * Run cow/eofblocks scans on the supplied dquots.  We don't know exactly which1562 * quota caused an allocation failure, so we make a best effort by including1563 * each quota under low free space conditions (less than 1% free space) in the1564 * scan.1565 *1566 * Callers must not hold any inode's ILOCK.  If requesting a synchronous scan1567 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or1568 * MMAPLOCK.1569 */1570int1571xfs_blockgc_free_dquots(1572	struct xfs_mount	*mp,1573	struct xfs_dquot	*udqp,1574	struct xfs_dquot	*gdqp,1575	struct xfs_dquot	*pdqp,1576	unsigned int		iwalk_flags)1577{1578	struct xfs_icwalk	icw = {0};1579	bool			do_work = false;1580 1581	if (!udqp && !gdqp && !pdqp)1582		return 0;1583 1584	/*1585	 * Run a scan to free blocks using the union filter to cover all1586	 * applicable quotas in a single scan.1587	 */1588	icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;1589 1590	if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {1591		icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);1592		icw.icw_flags |= XFS_ICWALK_FLAG_UID;1593		do_work = true;1594	}1595 1596	if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {1597		icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);1598		icw.icw_flags |= XFS_ICWALK_FLAG_GID;1599		do_work = true;1600	}1601 1602	if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {1603		icw.icw_prid = pdqp->q_id;1604		icw.icw_flags |= XFS_ICWALK_FLAG_PRID;1605		do_work = true;1606	}1607 1608	if (!do_work)1609		return 0;1610 1611	return xfs_blockgc_free_space(mp, &icw);1612}1613 1614/* Run cow/eofblocks scans on the quotas attached to the inode. */1615int1616xfs_blockgc_free_quota(1617	struct xfs_inode	*ip,1618	unsigned int		iwalk_flags)1619{1620	return xfs_blockgc_free_dquots(ip->i_mount,1621			xfs_inode_dquot(ip, XFS_DQTYPE_USER),1622			xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),1623			xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);1624}1625 1626/* XFS Inode Cache Walking Code */1627 1628/*1629 * The inode lookup is done in batches to keep the amount of lock traffic and1630 * radix tree lookups to a minimum. The batch size is a trade off between1631 * lookup reduction and stack usage. This is in the reclaim path, so we can't1632 * be too greedy.1633 */1634#define XFS_LOOKUP_BATCH	321635 1636 1637/*1638 * Decide if we want to grab this inode in anticipation of doing work towards1639 * the goal.1640 */1641static inline bool1642xfs_icwalk_igrab(1643	enum xfs_icwalk_goal	goal,1644	struct xfs_inode	*ip,1645	struct xfs_icwalk	*icw)1646{1647	switch (goal) {1648	case XFS_ICWALK_BLOCKGC:1649		return xfs_blockgc_igrab(ip);1650	case XFS_ICWALK_RECLAIM:1651		return xfs_reclaim_igrab(ip, icw);1652	default:1653		return false;1654	}1655}1656 1657/*1658 * Process an inode.  Each processing function must handle any state changes1659 * made by the icwalk igrab function.  Return -EAGAIN to skip an inode.1660 */1661static inline int1662xfs_icwalk_process_inode(1663	enum xfs_icwalk_goal	goal,1664	struct xfs_inode	*ip,1665	struct xfs_perag	*pag,1666	struct xfs_icwalk	*icw)1667{1668	int			error = 0;1669 1670	switch (goal) {1671	case XFS_ICWALK_BLOCKGC:1672		error = xfs_blockgc_scan_inode(ip, icw);1673		break;1674	case XFS_ICWALK_RECLAIM:1675		xfs_reclaim_inode(ip, pag);1676		break;1677	}1678	return error;1679}1680 1681/*1682 * For a given per-AG structure @pag and a goal, grab qualifying inodes and1683 * process them in some manner.1684 */1685static int1686xfs_icwalk_ag(1687	struct xfs_perag	*pag,1688	enum xfs_icwalk_goal	goal,1689	struct xfs_icwalk	*icw)1690{1691	struct xfs_mount	*mp = pag->pag_mount;1692	uint32_t		first_index;1693	int			last_error = 0;1694	int			skipped;1695	bool			done;1696	int			nr_found;1697 1698restart:1699	done = false;1700	skipped = 0;1701	if (goal == XFS_ICWALK_RECLAIM)1702		first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);1703	else1704		first_index = 0;1705	nr_found = 0;1706	do {1707		struct xfs_inode *batch[XFS_LOOKUP_BATCH];1708		int		error = 0;1709		int		i;1710 1711		rcu_read_lock();1712 1713		nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,1714				(void **) batch, first_index,1715				XFS_LOOKUP_BATCH, goal);1716		if (!nr_found) {1717			done = true;1718			rcu_read_unlock();1719			break;1720		}1721 1722		/*1723		 * Grab the inodes before we drop the lock. if we found1724		 * nothing, nr == 0 and the loop will be skipped.1725		 */1726		for (i = 0; i < nr_found; i++) {1727			struct xfs_inode *ip = batch[i];1728 1729			if (done || !xfs_icwalk_igrab(goal, ip, icw))1730				batch[i] = NULL;1731 1732			/*1733			 * Update the index for the next lookup. Catch1734			 * overflows into the next AG range which can occur if1735			 * we have inodes in the last block of the AG and we1736			 * are currently pointing to the last inode.1737			 *1738			 * Because we may see inodes that are from the wrong AG1739			 * due to RCU freeing and reallocation, only update the1740			 * index if it lies in this AG. It was a race that lead1741			 * us to see this inode, so another lookup from the1742			 * same index will not find it again.1743			 */1744			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)1745				continue;1746			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);1747			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))1748				done = true;1749		}1750 1751		/* unlock now we've grabbed the inodes. */1752		rcu_read_unlock();1753 1754		for (i = 0; i < nr_found; i++) {1755			if (!batch[i])1756				continue;1757			error = xfs_icwalk_process_inode(goal, batch[i], pag,1758					icw);1759			if (error == -EAGAIN) {1760				skipped++;1761				continue;1762			}1763			if (error && last_error != -EFSCORRUPTED)1764				last_error = error;1765		}1766 1767		/* bail out if the filesystem is corrupted.  */1768		if (error == -EFSCORRUPTED)1769			break;1770 1771		cond_resched();1772 1773		if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {1774			icw->icw_scan_limit -= XFS_LOOKUP_BATCH;1775			if (icw->icw_scan_limit <= 0)1776				break;1777		}1778	} while (nr_found && !done);1779 1780	if (goal == XFS_ICWALK_RECLAIM) {1781		if (done)1782			first_index = 0;1783		WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);1784	}1785 1786	if (skipped) {1787		delay(1);1788		goto restart;1789	}1790	return last_error;1791}1792 1793/* Walk all incore inodes to achieve a given goal. */1794static int1795xfs_icwalk(1796	struct xfs_mount	*mp,1797	enum xfs_icwalk_goal	goal,1798	struct xfs_icwalk	*icw)1799{1800	struct xfs_perag	*pag = NULL;1801	int			error = 0;1802	int			last_error = 0;1803 1804	while ((pag = xfs_perag_grab_next_tag(mp, pag, goal))) {1805		error = xfs_icwalk_ag(pag, goal, icw);1806		if (error) {1807			last_error = error;1808			if (error == -EFSCORRUPTED) {1809				xfs_perag_rele(pag);1810				break;1811			}1812		}1813	}1814	return last_error;1815	BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);1816}1817 1818#ifdef DEBUG1819static void1820xfs_check_delalloc(1821	struct xfs_inode	*ip,1822	int			whichfork)1823{1824	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, whichfork);1825	struct xfs_bmbt_irec	got;1826	struct xfs_iext_cursor	icur;1827 1828	if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))1829		return;1830	do {1831		if (isnullstartblock(got.br_startblock)) {1832			xfs_warn(ip->i_mount,1833	"ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",1834				ip->i_ino,1835				whichfork == XFS_DATA_FORK ? "data" : "cow",1836				got.br_startoff, got.br_blockcount);1837		}1838	} while (xfs_iext_next_extent(ifp, &icur, &got));1839}1840#else1841#define xfs_check_delalloc(ip, whichfork)	do { } while (0)1842#endif1843 1844/* Schedule the inode for reclaim. */1845static void1846xfs_inodegc_set_reclaimable(1847	struct xfs_inode	*ip)1848{1849	struct xfs_mount	*mp = ip->i_mount;1850	struct xfs_perag	*pag;1851 1852	if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {1853		xfs_check_delalloc(ip, XFS_DATA_FORK);1854		xfs_check_delalloc(ip, XFS_COW_FORK);1855		ASSERT(0);1856	}1857 1858	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));1859	spin_lock(&pag->pag_ici_lock);1860	spin_lock(&ip->i_flags_lock);1861 1862	trace_xfs_inode_set_reclaimable(ip);1863	ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);1864	ip->i_flags |= XFS_IRECLAIMABLE;1865	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),1866			XFS_ICI_RECLAIM_TAG);1867 1868	spin_unlock(&ip->i_flags_lock);1869	spin_unlock(&pag->pag_ici_lock);1870	xfs_perag_put(pag);1871}1872 1873/*1874 * Free all speculative preallocations and possibly even the inode itself.1875 * This is the last chance to make changes to an otherwise unreferenced file1876 * before incore reclamation happens.1877 */1878static int1879xfs_inodegc_inactivate(1880	struct xfs_inode	*ip)1881{1882	int			error;1883 1884	trace_xfs_inode_inactivating(ip);1885	error = xfs_inactive(ip);1886	xfs_inodegc_set_reclaimable(ip);1887	return error;1888 1889}1890 1891void1892xfs_inodegc_worker(1893	struct work_struct	*work)1894{1895	struct xfs_inodegc	*gc = container_of(to_delayed_work(work),1896						struct xfs_inodegc, work);1897	struct llist_node	*node = llist_del_all(&gc->list);1898	struct xfs_inode	*ip, *n;1899	struct xfs_mount	*mp = gc->mp;1900	unsigned int		nofs_flag;1901 1902	/*1903	 * Clear the cpu mask bit and ensure that we have seen the latest1904	 * update of the gc structure associated with this CPU. This matches1905	 * with the release semantics used when setting the cpumask bit in1906	 * xfs_inodegc_queue.1907	 */1908	cpumask_clear_cpu(gc->cpu, &mp->m_inodegc_cpumask);1909	smp_mb__after_atomic();1910 1911	WRITE_ONCE(gc->items, 0);1912 1913	if (!node)1914		return;1915 1916	/*1917	 * We can allocate memory here while doing writeback on behalf of1918	 * memory reclaim.  To avoid memory allocation deadlocks set the1919	 * task-wide nofs context for the following operations.1920	 */1921	nofs_flag = memalloc_nofs_save();1922 1923	ip = llist_entry(node, struct xfs_inode, i_gclist);1924	trace_xfs_inodegc_worker(mp, READ_ONCE(gc->shrinker_hits));1925 1926	WRITE_ONCE(gc->shrinker_hits, 0);1927	llist_for_each_entry_safe(ip, n, node, i_gclist) {1928		int	error;1929 1930		xfs_iflags_set(ip, XFS_INACTIVATING);1931		error = xfs_inodegc_inactivate(ip);1932		if (error && !gc->error)1933			gc->error = error;1934	}1935 1936	memalloc_nofs_restore(nofs_flag);1937}1938 1939/*1940 * Expedite all pending inodegc work to run immediately. This does not wait for1941 * completion of the work.1942 */1943void1944xfs_inodegc_push(1945	struct xfs_mount	*mp)1946{1947	if (!xfs_is_inodegc_enabled(mp))1948		return;1949	trace_xfs_inodegc_push(mp, __return_address);1950	xfs_inodegc_queue_all(mp);1951}1952 1953/*1954 * Force all currently queued inode inactivation work to run immediately and1955 * wait for the work to finish.1956 */1957int1958xfs_inodegc_flush(1959	struct xfs_mount	*mp)1960{1961	xfs_inodegc_push(mp);1962	trace_xfs_inodegc_flush(mp, __return_address);1963	return xfs_inodegc_wait_all(mp);1964}1965 1966/*1967 * Flush all the pending work and then disable the inode inactivation background1968 * workers and wait for them to stop.  Caller must hold sb->s_umount to1969 * coordinate changes in the inodegc_enabled state.1970 */1971void1972xfs_inodegc_stop(1973	struct xfs_mount	*mp)1974{1975	bool			rerun;1976 1977	if (!xfs_clear_inodegc_enabled(mp))1978		return;1979 1980	/*1981	 * Drain all pending inodegc work, including inodes that could be1982	 * queued by racing xfs_inodegc_queue or xfs_inodegc_shrinker_scan1983	 * threads that sample the inodegc state just prior to us clearing it.1984	 * The inodegc flag state prevents new threads from queuing more1985	 * inodes, so we queue pending work items and flush the workqueue until1986	 * all inodegc lists are empty.  IOWs, we cannot use drain_workqueue1987	 * here because it does not allow other unserialized mechanisms to1988	 * reschedule inodegc work while this draining is in progress.1989	 */1990	xfs_inodegc_queue_all(mp);1991	do {1992		flush_workqueue(mp->m_inodegc_wq);1993		rerun = xfs_inodegc_queue_all(mp);1994	} while (rerun);1995 1996	trace_xfs_inodegc_stop(mp, __return_address);1997}1998 1999/*2000 * Enable the inode inactivation background workers and schedule deferred inode2001 * inactivation work if there is any.  Caller must hold sb->s_umount to2002 * coordinate changes in the inodegc_enabled state.2003 */2004void2005xfs_inodegc_start(2006	struct xfs_mount	*mp)2007{2008	if (xfs_set_inodegc_enabled(mp))2009		return;2010 2011	trace_xfs_inodegc_start(mp, __return_address);2012	xfs_inodegc_queue_all(mp);2013}2014 2015#ifdef CONFIG_XFS_RT2016static inline bool2017xfs_inodegc_want_queue_rt_file(2018	struct xfs_inode	*ip)2019{2020	struct xfs_mount	*mp = ip->i_mount;2021 2022	if (!XFS_IS_REALTIME_INODE(ip))2023		return false;2024 2025	if (__percpu_counter_compare(&mp->m_frextents,2026				mp->m_low_rtexts[XFS_LOWSP_5_PCNT],2027				XFS_FDBLOCKS_BATCH) < 0)2028		return true;2029 2030	return false;2031}2032#else2033# define xfs_inodegc_want_queue_rt_file(ip)	(false)2034#endif /* CONFIG_XFS_RT */2035 2036/*2037 * Schedule the inactivation worker when:2038 *2039 *  - We've accumulated more than one inode cluster buffer's worth of inodes.2040 *  - There is less than 5% free space left.2041 *  - Any of the quotas for this inode are near an enforcement limit.2042 */2043static inline bool2044xfs_inodegc_want_queue_work(2045	struct xfs_inode	*ip,2046	unsigned int		items)2047{2048	struct xfs_mount	*mp = ip->i_mount;2049 2050	if (items > mp->m_ino_geo.inodes_per_cluster)2051		return true;2052 2053	if (__percpu_counter_compare(&mp->m_fdblocks,2054				mp->m_low_space[XFS_LOWSP_5_PCNT],2055				XFS_FDBLOCKS_BATCH) < 0)2056		return true;2057 2058	if (xfs_inodegc_want_queue_rt_file(ip))2059		return true;2060 2061	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))2062		return true;2063 2064	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))2065		return true;2066 2067	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))2068		return true;2069 2070	return false;2071}2072 2073/*2074 * Upper bound on the number of inodes in each AG that can be queued for2075 * inactivation at any given time, to avoid monopolizing the workqueue.2076 */2077#define XFS_INODEGC_MAX_BACKLOG		(4 * XFS_INODES_PER_CHUNK)2078 2079/*2080 * Make the frontend wait for inactivations when:2081 *2082 *  - Memory shrinkers queued the inactivation worker and it hasn't finished.2083 *  - The queue depth exceeds the maximum allowable percpu backlog.2084 *2085 * Note: If we are in a NOFS context here (e.g. current thread is running a2086 * transaction) the we don't want to block here as inodegc progress may require2087 * filesystem resources we hold to make progress and that could result in a2088 * deadlock. Hence we skip out of here if we are in a scoped NOFS context.2089 */2090static inline bool2091xfs_inodegc_want_flush_work(2092	struct xfs_inode	*ip,2093	unsigned int		items,2094	unsigned int		shrinker_hits)2095{2096	if (current->flags & PF_MEMALLOC_NOFS)2097		return false;2098 2099	if (shrinker_hits > 0)2100		return true;2101 2102	if (items > XFS_INODEGC_MAX_BACKLOG)2103		return true;2104 2105	return false;2106}2107 2108/*2109 * Queue a background inactivation worker if there are inodes that need to be2110 * inactivated and higher level xfs code hasn't disabled the background2111 * workers.2112 */2113static void2114xfs_inodegc_queue(2115	struct xfs_inode	*ip)2116{2117	struct xfs_mount	*mp = ip->i_mount;2118	struct xfs_inodegc	*gc;2119	int			items;2120	unsigned int		shrinker_hits;2121	unsigned int		cpu_nr;2122	unsigned long		queue_delay = 1;2123 2124	trace_xfs_inode_set_need_inactive(ip);2125	spin_lock(&ip->i_flags_lock);2126	ip->i_flags |= XFS_NEED_INACTIVE;2127	spin_unlock(&ip->i_flags_lock);2128 2129	cpu_nr = get_cpu();2130	gc = this_cpu_ptr(mp->m_inodegc);2131	llist_add(&ip->i_gclist, &gc->list);2132	items = READ_ONCE(gc->items);2133	WRITE_ONCE(gc->items, items + 1);2134	shrinker_hits = READ_ONCE(gc->shrinker_hits);2135 2136	/*2137	 * Ensure the list add is always seen by anyone who finds the cpumask2138	 * bit set. This effectively gives the cpumask bit set operation2139	 * release ordering semantics.2140	 */2141	smp_mb__before_atomic();2142	if (!cpumask_test_cpu(cpu_nr, &mp->m_inodegc_cpumask))2143		cpumask_test_and_set_cpu(cpu_nr, &mp->m_inodegc_cpumask);2144 2145	/*2146	 * We queue the work while holding the current CPU so that the work2147	 * is scheduled to run on this CPU.2148	 */2149	if (!xfs_is_inodegc_enabled(mp)) {2150		put_cpu();2151		return;2152	}2153 2154	if (xfs_inodegc_want_queue_work(ip, items))2155		queue_delay = 0;2156 2157	trace_xfs_inodegc_queue(mp, __return_address);2158	mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,2159			queue_delay);2160	put_cpu();2161 2162	if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {2163		trace_xfs_inodegc_throttle(mp, __return_address);2164		flush_delayed_work(&gc->work);2165	}2166}2167 2168/*2169 * We set the inode flag atomically with the radix tree tag.  Once we get tag2170 * lookups on the radix tree, this inode flag can go away.2171 *2172 * We always use background reclaim here because even if the inode is clean, it2173 * still may be under IO and hence we have wait for IO completion to occur2174 * before we can reclaim the inode. The background reclaim path handles this2175 * more efficiently than we can here, so simply let background reclaim tear down2176 * all inodes.2177 */2178void2179xfs_inode_mark_reclaimable(2180	struct xfs_inode	*ip)2181{2182	struct xfs_mount	*mp = ip->i_mount;2183	bool			need_inactive;2184 2185	XFS_STATS_INC(mp, vn_reclaim);2186 2187	/*2188	 * We should never get here with any of the reclaim flags already set.2189	 */2190	ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));2191 2192	need_inactive = xfs_inode_needs_inactive(ip);2193	if (need_inactive) {2194		xfs_inodegc_queue(ip);2195		return;2196	}2197 2198	/* Going straight to reclaim, so drop the dquots. */2199	xfs_qm_dqdetach(ip);2200	xfs_inodegc_set_reclaimable(ip);2201}2202 2203/*2204 * Register a phony shrinker so that we can run background inodegc sooner when2205 * there's memory pressure.  Inactivation does not itself free any memory but2206 * it does make inodes reclaimable, which eventually frees memory.2207 *2208 * The count function, seek value, and batch value are crafted to trigger the2209 * scan function during the second round of scanning.  Hopefully this means2210 * that we reclaimed enough memory that initiating metadata transactions won't2211 * make things worse.2212 */2213#define XFS_INODEGC_SHRINKER_COUNT	(1UL << DEF_PRIORITY)2214#define XFS_INODEGC_SHRINKER_BATCH	((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)2215 2216static unsigned long2217xfs_inodegc_shrinker_count(2218	struct shrinker		*shrink,2219	struct shrink_control	*sc)2220{2221	struct xfs_mount	*mp = shrink->private_data;2222	struct xfs_inodegc	*gc;2223	int			cpu;2224 2225	if (!xfs_is_inodegc_enabled(mp))2226		return 0;2227 2228	for_each_cpu(cpu, &mp->m_inodegc_cpumask) {2229		gc = per_cpu_ptr(mp->m_inodegc, cpu);2230		if (!llist_empty(&gc->list))2231			return XFS_INODEGC_SHRINKER_COUNT;2232	}2233 2234	return 0;2235}2236 2237static unsigned long2238xfs_inodegc_shrinker_scan(2239	struct shrinker		*shrink,2240	struct shrink_control	*sc)2241{2242	struct xfs_mount	*mp = shrink->private_data;2243	struct xfs_inodegc	*gc;2244	int			cpu;2245	bool			no_items = true;2246 2247	if (!xfs_is_inodegc_enabled(mp))2248		return SHRINK_STOP;2249 2250	trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);2251 2252	for_each_cpu(cpu, &mp->m_inodegc_cpumask) {2253		gc = per_cpu_ptr(mp->m_inodegc, cpu);2254		if (!llist_empty(&gc->list)) {2255			unsigned int	h = READ_ONCE(gc->shrinker_hits);2256 2257			WRITE_ONCE(gc->shrinker_hits, h + 1);2258			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);2259			no_items = false;2260		}2261	}2262 2263	/*2264	 * If there are no inodes to inactivate, we don't want the shrinker2265	 * to think there's deferred work to call us back about.2266	 */2267	if (no_items)2268		return LONG_MAX;2269 2270	return SHRINK_STOP;2271}2272 2273/* Register a shrinker so we can accelerate inodegc and throttle queuing. */2274int2275xfs_inodegc_register_shrinker(2276	struct xfs_mount	*mp)2277{2278	mp->m_inodegc_shrinker = shrinker_alloc(SHRINKER_NONSLAB,2279						"xfs-inodegc:%s",2280						mp->m_super->s_id);2281	if (!mp->m_inodegc_shrinker)2282		return -ENOMEM;2283 2284	mp->m_inodegc_shrinker->count_objects = xfs_inodegc_shrinker_count;2285	mp->m_inodegc_shrinker->scan_objects = xfs_inodegc_shrinker_scan;2286	mp->m_inodegc_shrinker->seeks = 0;2287	mp->m_inodegc_shrinker->batch = XFS_INODEGC_SHRINKER_BATCH;2288	mp->m_inodegc_shrinker->private_data = mp;2289 2290	shrinker_register(mp->m_inodegc_shrinker);2291 2292	return 0;2293}2294