xref: /openbmc/linux/fs/xfs/xfs_icache.c (revision 901181b7)
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 
27 #include <linux/iversion.h>
28 
29 /* Radix tree tags for incore inode tree. */
30 
31 /* inode is to be reclaimed */
32 #define XFS_ICI_RECLAIM_TAG	0
33 /* Inode has speculative preallocations (posteof or cow) to clean. */
34 #define XFS_ICI_BLOCKGC_TAG	1
35 
36 /*
37  * The goal for walking incore inodes.  These can correspond with incore inode
38  * radix tree tags when convenient.  Avoid existing XFS_IWALK namespace.
39  */
40 enum xfs_icwalk_goal {
41 	/* Goals directly associated with tagged inodes. */
42 	XFS_ICWALK_BLOCKGC	= XFS_ICI_BLOCKGC_TAG,
43 	XFS_ICWALK_RECLAIM	= XFS_ICI_RECLAIM_TAG,
44 };
45 
46 static int xfs_icwalk(struct xfs_mount *mp,
47 		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
48 static int xfs_icwalk_ag(struct xfs_perag *pag,
49 		enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
50 
51 /*
52  * Private inode cache walk flags for struct xfs_icwalk.  Must not
53  * coincide with XFS_ICWALK_FLAGS_VALID.
54  */
55 
56 /* Stop scanning after icw_scan_limit inodes. */
57 #define XFS_ICWALK_FLAG_SCAN_LIMIT	(1U << 28)
58 
59 #define XFS_ICWALK_FLAG_RECLAIM_SICK	(1U << 27)
60 #define XFS_ICWALK_FLAG_UNION		(1U << 26) /* union filter algorithm */
61 
62 #define XFS_ICWALK_PRIVATE_FLAGS	(XFS_ICWALK_FLAG_SCAN_LIMIT | \
63 					 XFS_ICWALK_FLAG_RECLAIM_SICK | \
64 					 XFS_ICWALK_FLAG_UNION)
65 
66 /*
67  * Allocate and initialise an xfs_inode.
68  */
69 struct xfs_inode *
70 xfs_inode_alloc(
71 	struct xfs_mount	*mp,
72 	xfs_ino_t		ino)
73 {
74 	struct xfs_inode	*ip;
75 
76 	/*
77 	 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
78 	 * and return NULL here on ENOMEM.
79 	 */
80 	ip = kmem_cache_alloc(xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
81 
82 	if (inode_init_always(mp->m_super, VFS_I(ip))) {
83 		kmem_cache_free(xfs_inode_cache, ip);
84 		return NULL;
85 	}
86 
87 	/* VFS doesn't initialise i_mode or i_state! */
88 	VFS_I(ip)->i_mode = 0;
89 	VFS_I(ip)->i_state = 0;
90 
91 	XFS_STATS_INC(mp, vn_active);
92 	ASSERT(atomic_read(&ip->i_pincount) == 0);
93 	ASSERT(ip->i_ino == 0);
94 
95 	/* initialise the xfs inode */
96 	ip->i_ino = ino;
97 	ip->i_mount = mp;
98 	memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
99 	ip->i_afp = NULL;
100 	ip->i_cowfp = NULL;
101 	memset(&ip->i_df, 0, sizeof(ip->i_df));
102 	ip->i_flags = 0;
103 	ip->i_delayed_blks = 0;
104 	ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
105 	ip->i_nblocks = 0;
106 	ip->i_forkoff = 0;
107 	ip->i_sick = 0;
108 	ip->i_checked = 0;
109 	INIT_WORK(&ip->i_ioend_work, xfs_end_io);
110 	INIT_LIST_HEAD(&ip->i_ioend_list);
111 	spin_lock_init(&ip->i_ioend_lock);
112 
113 	return ip;
114 }
115 
116 STATIC void
117 xfs_inode_free_callback(
118 	struct rcu_head		*head)
119 {
120 	struct inode		*inode = container_of(head, struct inode, i_rcu);
121 	struct xfs_inode	*ip = XFS_I(inode);
122 
123 	switch (VFS_I(ip)->i_mode & S_IFMT) {
124 	case S_IFREG:
125 	case S_IFDIR:
126 	case S_IFLNK:
127 		xfs_idestroy_fork(&ip->i_df);
128 		break;
129 	}
130 
131 	if (ip->i_afp) {
132 		xfs_idestroy_fork(ip->i_afp);
133 		kmem_cache_free(xfs_ifork_cache, ip->i_afp);
134 	}
135 	if (ip->i_cowfp) {
136 		xfs_idestroy_fork(ip->i_cowfp);
137 		kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
138 	}
139 	if (ip->i_itemp) {
140 		ASSERT(!test_bit(XFS_LI_IN_AIL,
141 				 &ip->i_itemp->ili_item.li_flags));
142 		xfs_inode_item_destroy(ip);
143 		ip->i_itemp = NULL;
144 	}
145 
146 	kmem_cache_free(xfs_inode_cache, ip);
147 }
148 
149 static void
150 __xfs_inode_free(
151 	struct xfs_inode	*ip)
152 {
153 	/* asserts to verify all state is correct here */
154 	ASSERT(atomic_read(&ip->i_pincount) == 0);
155 	ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
156 	XFS_STATS_DEC(ip->i_mount, vn_active);
157 
158 	call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
159 }
160 
161 void
162 xfs_inode_free(
163 	struct xfs_inode	*ip)
164 {
165 	ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
166 
167 	/*
168 	 * Because we use RCU freeing we need to ensure the inode always
169 	 * appears to be reclaimed with an invalid inode number when in the
170 	 * free state. The ip->i_flags_lock provides the barrier against lookup
171 	 * races.
172 	 */
173 	spin_lock(&ip->i_flags_lock);
174 	ip->i_flags = XFS_IRECLAIM;
175 	ip->i_ino = 0;
176 	spin_unlock(&ip->i_flags_lock);
177 
178 	__xfs_inode_free(ip);
179 }
180 
181 /*
182  * Queue background inode reclaim work if there are reclaimable inodes and there
183  * isn't reclaim work already scheduled or in progress.
184  */
185 static void
186 xfs_reclaim_work_queue(
187 	struct xfs_mount        *mp)
188 {
189 
190 	rcu_read_lock();
191 	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
192 		queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
193 			msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
194 	}
195 	rcu_read_unlock();
196 }
197 
198 /*
199  * Background scanning to trim preallocated space. This is queued based on the
200  * 'speculative_prealloc_lifetime' tunable (5m by default).
201  */
202 static inline void
203 xfs_blockgc_queue(
204 	struct xfs_perag	*pag)
205 {
206 	struct xfs_mount	*mp = pag->pag_mount;
207 
208 	if (!xfs_is_blockgc_enabled(mp))
209 		return;
210 
211 	rcu_read_lock();
212 	if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
213 		queue_delayed_work(pag->pag_mount->m_blockgc_wq,
214 				   &pag->pag_blockgc_work,
215 				   msecs_to_jiffies(xfs_blockgc_secs * 1000));
216 	rcu_read_unlock();
217 }
218 
219 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
220 static void
221 xfs_perag_set_inode_tag(
222 	struct xfs_perag	*pag,
223 	xfs_agino_t		agino,
224 	unsigned int		tag)
225 {
226 	struct xfs_mount	*mp = pag->pag_mount;
227 	bool			was_tagged;
228 
229 	lockdep_assert_held(&pag->pag_ici_lock);
230 
231 	was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
232 	radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
233 
234 	if (tag == XFS_ICI_RECLAIM_TAG)
235 		pag->pag_ici_reclaimable++;
236 
237 	if (was_tagged)
238 		return;
239 
240 	/* propagate the tag up into the perag radix tree */
241 	spin_lock(&mp->m_perag_lock);
242 	radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
243 	spin_unlock(&mp->m_perag_lock);
244 
245 	/* start background work */
246 	switch (tag) {
247 	case XFS_ICI_RECLAIM_TAG:
248 		xfs_reclaim_work_queue(mp);
249 		break;
250 	case XFS_ICI_BLOCKGC_TAG:
251 		xfs_blockgc_queue(pag);
252 		break;
253 	}
254 
255 	trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
256 }
257 
258 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
259 static void
260 xfs_perag_clear_inode_tag(
261 	struct xfs_perag	*pag,
262 	xfs_agino_t		agino,
263 	unsigned int		tag)
264 {
265 	struct xfs_mount	*mp = pag->pag_mount;
266 
267 	lockdep_assert_held(&pag->pag_ici_lock);
268 
269 	/*
270 	 * Reclaim can signal (with a null agino) that it cleared its own tag
271 	 * by removing the inode from the radix tree.
272 	 */
273 	if (agino != NULLAGINO)
274 		radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
275 	else
276 		ASSERT(tag == XFS_ICI_RECLAIM_TAG);
277 
278 	if (tag == XFS_ICI_RECLAIM_TAG)
279 		pag->pag_ici_reclaimable--;
280 
281 	if (radix_tree_tagged(&pag->pag_ici_root, tag))
282 		return;
283 
284 	/* clear the tag from the perag radix tree */
285 	spin_lock(&mp->m_perag_lock);
286 	radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
287 	spin_unlock(&mp->m_perag_lock);
288 
289 	trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
290 }
291 
292 static inline void
293 xfs_inew_wait(
294 	struct xfs_inode	*ip)
295 {
296 	wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
297 	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
298 
299 	do {
300 		prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
301 		if (!xfs_iflags_test(ip, XFS_INEW))
302 			break;
303 		schedule();
304 	} while (true);
305 	finish_wait(wq, &wait.wq_entry);
306 }
307 
308 /*
309  * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
310  * part of the structure. This is made more complex by the fact we store
311  * information about the on-disk values in the VFS inode and so we can't just
312  * overwrite the values unconditionally. Hence we save the parameters we
313  * need to retain across reinitialisation, and rewrite them into the VFS inode
314  * after reinitialisation even if it fails.
315  */
316 static int
317 xfs_reinit_inode(
318 	struct xfs_mount	*mp,
319 	struct inode		*inode)
320 {
321 	int			error;
322 	uint32_t		nlink = inode->i_nlink;
323 	uint32_t		generation = inode->i_generation;
324 	uint64_t		version = inode_peek_iversion(inode);
325 	umode_t			mode = inode->i_mode;
326 	dev_t			dev = inode->i_rdev;
327 	kuid_t			uid = inode->i_uid;
328 	kgid_t			gid = inode->i_gid;
329 
330 	error = inode_init_always(mp->m_super, inode);
331 
332 	set_nlink(inode, nlink);
333 	inode->i_generation = generation;
334 	inode_set_iversion_queried(inode, version);
335 	inode->i_mode = mode;
336 	inode->i_rdev = dev;
337 	inode->i_uid = uid;
338 	inode->i_gid = gid;
339 	return error;
340 }
341 
342 /*
343  * Carefully nudge an inode whose VFS state has been torn down back into a
344  * usable state.  Drops the i_flags_lock and the rcu read lock.
345  */
346 static int
347 xfs_iget_recycle(
348 	struct xfs_perag	*pag,
349 	struct xfs_inode	*ip) __releases(&ip->i_flags_lock)
350 {
351 	struct xfs_mount	*mp = ip->i_mount;
352 	struct inode		*inode = VFS_I(ip);
353 	int			error;
354 
355 	trace_xfs_iget_recycle(ip);
356 
357 	/*
358 	 * We need to make it look like the inode is being reclaimed to prevent
359 	 * the actual reclaim workers from stomping over us while we recycle
360 	 * the inode.  We can't clear the radix tree tag yet as it requires
361 	 * pag_ici_lock to be held exclusive.
362 	 */
363 	ip->i_flags |= XFS_IRECLAIM;
364 
365 	spin_unlock(&ip->i_flags_lock);
366 	rcu_read_unlock();
367 
368 	ASSERT(!rwsem_is_locked(&inode->i_rwsem));
369 	error = xfs_reinit_inode(mp, inode);
370 	if (error) {
371 		bool	wake;
372 
373 		/*
374 		 * Re-initializing the inode failed, and we are in deep
375 		 * trouble.  Try to re-add it to the reclaim list.
376 		 */
377 		rcu_read_lock();
378 		spin_lock(&ip->i_flags_lock);
379 		wake = !!__xfs_iflags_test(ip, XFS_INEW);
380 		ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
381 		if (wake)
382 			wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
383 		ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
384 		spin_unlock(&ip->i_flags_lock);
385 		rcu_read_unlock();
386 
387 		trace_xfs_iget_recycle_fail(ip);
388 		return error;
389 	}
390 
391 	spin_lock(&pag->pag_ici_lock);
392 	spin_lock(&ip->i_flags_lock);
393 
394 	/*
395 	 * Clear the per-lifetime state in the inode as we are now effectively
396 	 * a new inode and need to return to the initial state before reuse
397 	 * occurs.
398 	 */
399 	ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
400 	ip->i_flags |= XFS_INEW;
401 	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
402 			XFS_ICI_RECLAIM_TAG);
403 	inode->i_state = I_NEW;
404 	spin_unlock(&ip->i_flags_lock);
405 	spin_unlock(&pag->pag_ici_lock);
406 
407 	return 0;
408 }
409 
410 /*
411  * If we are allocating a new inode, then check what was returned is
412  * actually a free, empty inode. If we are not allocating an inode,
413  * then check we didn't find a free inode.
414  *
415  * Returns:
416  *	0		if the inode free state matches the lookup context
417  *	-ENOENT		if the inode is free and we are not allocating
418  *	-EFSCORRUPTED	if there is any state mismatch at all
419  */
420 static int
421 xfs_iget_check_free_state(
422 	struct xfs_inode	*ip,
423 	int			flags)
424 {
425 	if (flags & XFS_IGET_CREATE) {
426 		/* should be a free inode */
427 		if (VFS_I(ip)->i_mode != 0) {
428 			xfs_warn(ip->i_mount,
429 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
430 				ip->i_ino, VFS_I(ip)->i_mode);
431 			return -EFSCORRUPTED;
432 		}
433 
434 		if (ip->i_nblocks != 0) {
435 			xfs_warn(ip->i_mount,
436 "Corruption detected! Free inode 0x%llx has blocks allocated!",
437 				ip->i_ino);
438 			return -EFSCORRUPTED;
439 		}
440 		return 0;
441 	}
442 
443 	/* should be an allocated inode */
444 	if (VFS_I(ip)->i_mode == 0)
445 		return -ENOENT;
446 
447 	return 0;
448 }
449 
450 /* Make all pending inactivation work start immediately. */
451 static void
452 xfs_inodegc_queue_all(
453 	struct xfs_mount	*mp)
454 {
455 	struct xfs_inodegc	*gc;
456 	int			cpu;
457 
458 	for_each_online_cpu(cpu) {
459 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
460 		if (!llist_empty(&gc->list))
461 			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
462 	}
463 }
464 
465 /*
466  * Check the validity of the inode we just found it the cache
467  */
468 static int
469 xfs_iget_cache_hit(
470 	struct xfs_perag	*pag,
471 	struct xfs_inode	*ip,
472 	xfs_ino_t		ino,
473 	int			flags,
474 	int			lock_flags) __releases(RCU)
475 {
476 	struct inode		*inode = VFS_I(ip);
477 	struct xfs_mount	*mp = ip->i_mount;
478 	int			error;
479 
480 	/*
481 	 * check for re-use of an inode within an RCU grace period due to the
482 	 * radix tree nodes not being updated yet. We monitor for this by
483 	 * setting the inode number to zero before freeing the inode structure.
484 	 * If the inode has been reallocated and set up, then the inode number
485 	 * will not match, so check for that, too.
486 	 */
487 	spin_lock(&ip->i_flags_lock);
488 	if (ip->i_ino != ino)
489 		goto out_skip;
490 
491 	/*
492 	 * If we are racing with another cache hit that is currently
493 	 * instantiating this inode or currently recycling it out of
494 	 * reclaimable state, wait for the initialisation to complete
495 	 * before continuing.
496 	 *
497 	 * If we're racing with the inactivation worker we also want to wait.
498 	 * If we're creating a new file, it's possible that the worker
499 	 * previously marked the inode as free on disk but hasn't finished
500 	 * updating the incore state yet.  The AGI buffer will be dirty and
501 	 * locked to the icreate transaction, so a synchronous push of the
502 	 * inodegc workers would result in deadlock.  For a regular iget, the
503 	 * worker is running already, so we might as well wait.
504 	 *
505 	 * XXX(hch): eventually we should do something equivalent to
506 	 *	     wait_on_inode to wait for these flags to be cleared
507 	 *	     instead of polling for it.
508 	 */
509 	if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
510 		goto out_skip;
511 
512 	if (ip->i_flags & XFS_NEED_INACTIVE) {
513 		/* Unlinked inodes cannot be re-grabbed. */
514 		if (VFS_I(ip)->i_nlink == 0) {
515 			error = -ENOENT;
516 			goto out_error;
517 		}
518 		goto out_inodegc_flush;
519 	}
520 
521 	/*
522 	 * Check the inode free state is valid. This also detects lookup
523 	 * racing with unlinks.
524 	 */
525 	error = xfs_iget_check_free_state(ip, flags);
526 	if (error)
527 		goto out_error;
528 
529 	/* Skip inodes that have no vfs state. */
530 	if ((flags & XFS_IGET_INCORE) &&
531 	    (ip->i_flags & XFS_IRECLAIMABLE))
532 		goto out_skip;
533 
534 	/* The inode fits the selection criteria; process it. */
535 	if (ip->i_flags & XFS_IRECLAIMABLE) {
536 		/* Drops i_flags_lock and RCU read lock. */
537 		error = xfs_iget_recycle(pag, ip);
538 		if (error)
539 			return error;
540 	} else {
541 		/* If the VFS inode is being torn down, pause and try again. */
542 		if (!igrab(inode))
543 			goto out_skip;
544 
545 		/* We've got a live one. */
546 		spin_unlock(&ip->i_flags_lock);
547 		rcu_read_unlock();
548 		trace_xfs_iget_hit(ip);
549 	}
550 
551 	if (lock_flags != 0)
552 		xfs_ilock(ip, lock_flags);
553 
554 	if (!(flags & XFS_IGET_INCORE))
555 		xfs_iflags_clear(ip, XFS_ISTALE);
556 	XFS_STATS_INC(mp, xs_ig_found);
557 
558 	return 0;
559 
560 out_skip:
561 	trace_xfs_iget_skip(ip);
562 	XFS_STATS_INC(mp, xs_ig_frecycle);
563 	error = -EAGAIN;
564 out_error:
565 	spin_unlock(&ip->i_flags_lock);
566 	rcu_read_unlock();
567 	return error;
568 
569 out_inodegc_flush:
570 	spin_unlock(&ip->i_flags_lock);
571 	rcu_read_unlock();
572 	/*
573 	 * Do not wait for the workers, because the caller could hold an AGI
574 	 * buffer lock.  We're just going to sleep in a loop anyway.
575 	 */
576 	if (xfs_is_inodegc_enabled(mp))
577 		xfs_inodegc_queue_all(mp);
578 	return -EAGAIN;
579 }
580 
581 static int
582 xfs_iget_cache_miss(
583 	struct xfs_mount	*mp,
584 	struct xfs_perag	*pag,
585 	xfs_trans_t		*tp,
586 	xfs_ino_t		ino,
587 	struct xfs_inode	**ipp,
588 	int			flags,
589 	int			lock_flags)
590 {
591 	struct xfs_inode	*ip;
592 	int			error;
593 	xfs_agino_t		agino = XFS_INO_TO_AGINO(mp, ino);
594 	int			iflags;
595 
596 	ip = xfs_inode_alloc(mp, ino);
597 	if (!ip)
598 		return -ENOMEM;
599 
600 	error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
601 	if (error)
602 		goto out_destroy;
603 
604 	/*
605 	 * For version 5 superblocks, if we are initialising a new inode and we
606 	 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
607 	 * simply build the new inode core with a random generation number.
608 	 *
609 	 * For version 4 (and older) superblocks, log recovery is dependent on
610 	 * the i_flushiter field being initialised from the current on-disk
611 	 * value and hence we must also read the inode off disk even when
612 	 * initializing new inodes.
613 	 */
614 	if (xfs_has_v3inodes(mp) &&
615 	    (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
616 		VFS_I(ip)->i_generation = prandom_u32();
617 	} else {
618 		struct xfs_buf		*bp;
619 
620 		error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
621 		if (error)
622 			goto out_destroy;
623 
624 		error = xfs_inode_from_disk(ip,
625 				xfs_buf_offset(bp, ip->i_imap.im_boffset));
626 		if (!error)
627 			xfs_buf_set_ref(bp, XFS_INO_REF);
628 		xfs_trans_brelse(tp, bp);
629 
630 		if (error)
631 			goto out_destroy;
632 	}
633 
634 	trace_xfs_iget_miss(ip);
635 
636 	/*
637 	 * Check the inode free state is valid. This also detects lookup
638 	 * racing with unlinks.
639 	 */
640 	error = xfs_iget_check_free_state(ip, flags);
641 	if (error)
642 		goto out_destroy;
643 
644 	/*
645 	 * Preload the radix tree so we can insert safely under the
646 	 * write spinlock. Note that we cannot sleep inside the preload
647 	 * region. Since we can be called from transaction context, don't
648 	 * recurse into the file system.
649 	 */
650 	if (radix_tree_preload(GFP_NOFS)) {
651 		error = -EAGAIN;
652 		goto out_destroy;
653 	}
654 
655 	/*
656 	 * Because the inode hasn't been added to the radix-tree yet it can't
657 	 * be found by another thread, so we can do the non-sleeping lock here.
658 	 */
659 	if (lock_flags) {
660 		if (!xfs_ilock_nowait(ip, lock_flags))
661 			BUG();
662 	}
663 
664 	/*
665 	 * These values must be set before inserting the inode into the radix
666 	 * tree as the moment it is inserted a concurrent lookup (allowed by the
667 	 * RCU locking mechanism) can find it and that lookup must see that this
668 	 * is an inode currently under construction (i.e. that XFS_INEW is set).
669 	 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
670 	 * memory barrier that ensures this detection works correctly at lookup
671 	 * time.
672 	 */
673 	iflags = XFS_INEW;
674 	if (flags & XFS_IGET_DONTCACHE)
675 		d_mark_dontcache(VFS_I(ip));
676 	ip->i_udquot = NULL;
677 	ip->i_gdquot = NULL;
678 	ip->i_pdquot = NULL;
679 	xfs_iflags_set(ip, iflags);
680 
681 	/* insert the new inode */
682 	spin_lock(&pag->pag_ici_lock);
683 	error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
684 	if (unlikely(error)) {
685 		WARN_ON(error != -EEXIST);
686 		XFS_STATS_INC(mp, xs_ig_dup);
687 		error = -EAGAIN;
688 		goto out_preload_end;
689 	}
690 	spin_unlock(&pag->pag_ici_lock);
691 	radix_tree_preload_end();
692 
693 	*ipp = ip;
694 	return 0;
695 
696 out_preload_end:
697 	spin_unlock(&pag->pag_ici_lock);
698 	radix_tree_preload_end();
699 	if (lock_flags)
700 		xfs_iunlock(ip, lock_flags);
701 out_destroy:
702 	__destroy_inode(VFS_I(ip));
703 	xfs_inode_free(ip);
704 	return error;
705 }
706 
707 /*
708  * Look up an inode by number in the given file system.  The inode is looked up
709  * in the cache held in each AG.  If the inode is found in the cache, initialise
710  * the vfs inode if necessary.
711  *
712  * If it is not in core, read it in from the file system's device, add it to the
713  * cache and initialise the vfs inode.
714  *
715  * The inode is locked according to the value of the lock_flags parameter.
716  * Inode lookup is only done during metadata operations and not as part of the
717  * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
718  */
719 int
720 xfs_iget(
721 	struct xfs_mount	*mp,
722 	struct xfs_trans	*tp,
723 	xfs_ino_t		ino,
724 	uint			flags,
725 	uint			lock_flags,
726 	struct xfs_inode	**ipp)
727 {
728 	struct xfs_inode	*ip;
729 	struct xfs_perag	*pag;
730 	xfs_agino_t		agino;
731 	int			error;
732 
733 	ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
734 
735 	/* reject inode numbers outside existing AGs */
736 	if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
737 		return -EINVAL;
738 
739 	XFS_STATS_INC(mp, xs_ig_attempts);
740 
741 	/* get the perag structure and ensure that it's inode capable */
742 	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
743 	agino = XFS_INO_TO_AGINO(mp, ino);
744 
745 again:
746 	error = 0;
747 	rcu_read_lock();
748 	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
749 
750 	if (ip) {
751 		error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
752 		if (error)
753 			goto out_error_or_again;
754 	} else {
755 		rcu_read_unlock();
756 		if (flags & XFS_IGET_INCORE) {
757 			error = -ENODATA;
758 			goto out_error_or_again;
759 		}
760 		XFS_STATS_INC(mp, xs_ig_missed);
761 
762 		error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
763 							flags, lock_flags);
764 		if (error)
765 			goto out_error_or_again;
766 	}
767 	xfs_perag_put(pag);
768 
769 	*ipp = ip;
770 
771 	/*
772 	 * If we have a real type for an on-disk inode, we can setup the inode
773 	 * now.	 If it's a new inode being created, xfs_ialloc will handle it.
774 	 */
775 	if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
776 		xfs_setup_existing_inode(ip);
777 	return 0;
778 
779 out_error_or_again:
780 	if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
781 		delay(1);
782 		goto again;
783 	}
784 	xfs_perag_put(pag);
785 	return error;
786 }
787 
788 /*
789  * "Is this a cached inode that's also allocated?"
790  *
791  * Look up an inode by number in the given file system.  If the inode is
792  * in cache and isn't in purgatory, return 1 if the inode is allocated
793  * and 0 if it is not.  For all other cases (not in cache, being torn
794  * down, etc.), return a negative error code.
795  *
796  * The caller has to prevent inode allocation and freeing activity,
797  * presumably by locking the AGI buffer.   This is to ensure that an
798  * inode cannot transition from allocated to freed until the caller is
799  * ready to allow that.  If the inode is in an intermediate state (new,
800  * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
801  * inode is not in the cache, -ENOENT will be returned.  The caller must
802  * deal with these scenarios appropriately.
803  *
804  * This is a specialized use case for the online scrubber; if you're
805  * reading this, you probably want xfs_iget.
806  */
807 int
808 xfs_icache_inode_is_allocated(
809 	struct xfs_mount	*mp,
810 	struct xfs_trans	*tp,
811 	xfs_ino_t		ino,
812 	bool			*inuse)
813 {
814 	struct xfs_inode	*ip;
815 	int			error;
816 
817 	error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
818 	if (error)
819 		return error;
820 
821 	*inuse = !!(VFS_I(ip)->i_mode);
822 	xfs_irele(ip);
823 	return 0;
824 }
825 
826 /*
827  * Grab the inode for reclaim exclusively.
828  *
829  * We have found this inode via a lookup under RCU, so the inode may have
830  * already been freed, or it may be in the process of being recycled by
831  * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
832  * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
833  * will not be set. Hence we need to check for both these flag conditions to
834  * avoid inodes that are no longer reclaim candidates.
835  *
836  * Note: checking for other state flags here, under the i_flags_lock or not, is
837  * racy and should be avoided. Those races should be resolved only after we have
838  * ensured that we are able to reclaim this inode and the world can see that we
839  * are going to reclaim it.
840  *
841  * Return true if we grabbed it, false otherwise.
842  */
843 static bool
844 xfs_reclaim_igrab(
845 	struct xfs_inode	*ip,
846 	struct xfs_icwalk	*icw)
847 {
848 	ASSERT(rcu_read_lock_held());
849 
850 	spin_lock(&ip->i_flags_lock);
851 	if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
852 	    __xfs_iflags_test(ip, XFS_IRECLAIM)) {
853 		/* not a reclaim candidate. */
854 		spin_unlock(&ip->i_flags_lock);
855 		return false;
856 	}
857 
858 	/* Don't reclaim a sick inode unless the caller asked for it. */
859 	if (ip->i_sick &&
860 	    (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
861 		spin_unlock(&ip->i_flags_lock);
862 		return false;
863 	}
864 
865 	__xfs_iflags_set(ip, XFS_IRECLAIM);
866 	spin_unlock(&ip->i_flags_lock);
867 	return true;
868 }
869 
870 /*
871  * Inode reclaim is non-blocking, so the default action if progress cannot be
872  * made is to "requeue" the inode for reclaim by unlocking it and clearing the
873  * XFS_IRECLAIM flag.  If we are in a shutdown state, we don't care about
874  * blocking anymore and hence we can wait for the inode to be able to reclaim
875  * it.
876  *
877  * We do no IO here - if callers require inodes to be cleaned they must push the
878  * AIL first to trigger writeback of dirty inodes.  This enables writeback to be
879  * done in the background in a non-blocking manner, and enables memory reclaim
880  * to make progress without blocking.
881  */
882 static void
883 xfs_reclaim_inode(
884 	struct xfs_inode	*ip,
885 	struct xfs_perag	*pag)
886 {
887 	xfs_ino_t		ino = ip->i_ino; /* for radix_tree_delete */
888 
889 	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
890 		goto out;
891 	if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
892 		goto out_iunlock;
893 
894 	if (xfs_is_shutdown(ip->i_mount)) {
895 		xfs_iunpin_wait(ip);
896 		xfs_iflush_abort(ip);
897 		goto reclaim;
898 	}
899 	if (xfs_ipincount(ip))
900 		goto out_clear_flush;
901 	if (!xfs_inode_clean(ip))
902 		goto out_clear_flush;
903 
904 	xfs_iflags_clear(ip, XFS_IFLUSHING);
905 reclaim:
906 	trace_xfs_inode_reclaiming(ip);
907 
908 	/*
909 	 * Because we use RCU freeing we need to ensure the inode always appears
910 	 * to be reclaimed with an invalid inode number when in the free state.
911 	 * We do this as early as possible under the ILOCK so that
912 	 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
913 	 * detect races with us here. By doing this, we guarantee that once
914 	 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
915 	 * it will see either a valid inode that will serialise correctly, or it
916 	 * will see an invalid inode that it can skip.
917 	 */
918 	spin_lock(&ip->i_flags_lock);
919 	ip->i_flags = XFS_IRECLAIM;
920 	ip->i_ino = 0;
921 	ip->i_sick = 0;
922 	ip->i_checked = 0;
923 	spin_unlock(&ip->i_flags_lock);
924 
925 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
926 
927 	XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
928 	/*
929 	 * Remove the inode from the per-AG radix tree.
930 	 *
931 	 * Because radix_tree_delete won't complain even if the item was never
932 	 * added to the tree assert that it's been there before to catch
933 	 * problems with the inode life time early on.
934 	 */
935 	spin_lock(&pag->pag_ici_lock);
936 	if (!radix_tree_delete(&pag->pag_ici_root,
937 				XFS_INO_TO_AGINO(ip->i_mount, ino)))
938 		ASSERT(0);
939 	xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
940 	spin_unlock(&pag->pag_ici_lock);
941 
942 	/*
943 	 * Here we do an (almost) spurious inode lock in order to coordinate
944 	 * with inode cache radix tree lookups.  This is because the lookup
945 	 * can reference the inodes in the cache without taking references.
946 	 *
947 	 * We make that OK here by ensuring that we wait until the inode is
948 	 * unlocked after the lookup before we go ahead and free it.
949 	 */
950 	xfs_ilock(ip, XFS_ILOCK_EXCL);
951 	ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
952 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
953 	ASSERT(xfs_inode_clean(ip));
954 
955 	__xfs_inode_free(ip);
956 	return;
957 
958 out_clear_flush:
959 	xfs_iflags_clear(ip, XFS_IFLUSHING);
960 out_iunlock:
961 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
962 out:
963 	xfs_iflags_clear(ip, XFS_IRECLAIM);
964 }
965 
966 /* Reclaim sick inodes if we're unmounting or the fs went down. */
967 static inline bool
968 xfs_want_reclaim_sick(
969 	struct xfs_mount	*mp)
970 {
971 	return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
972 	       xfs_is_shutdown(mp);
973 }
974 
975 void
976 xfs_reclaim_inodes(
977 	struct xfs_mount	*mp)
978 {
979 	struct xfs_icwalk	icw = {
980 		.icw_flags	= 0,
981 	};
982 
983 	if (xfs_want_reclaim_sick(mp))
984 		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
985 
986 	while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
987 		xfs_ail_push_all_sync(mp->m_ail);
988 		xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
989 	}
990 }
991 
992 /*
993  * The shrinker infrastructure determines how many inodes we should scan for
994  * reclaim. We want as many clean inodes ready to reclaim as possible, so we
995  * push the AIL here. We also want to proactively free up memory if we can to
996  * minimise the amount of work memory reclaim has to do so we kick the
997  * background reclaim if it isn't already scheduled.
998  */
999 long
1000 xfs_reclaim_inodes_nr(
1001 	struct xfs_mount	*mp,
1002 	unsigned long		nr_to_scan)
1003 {
1004 	struct xfs_icwalk	icw = {
1005 		.icw_flags	= XFS_ICWALK_FLAG_SCAN_LIMIT,
1006 		.icw_scan_limit	= min_t(unsigned long, LONG_MAX, nr_to_scan),
1007 	};
1008 
1009 	if (xfs_want_reclaim_sick(mp))
1010 		icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1011 
1012 	/* kick background reclaimer and push the AIL */
1013 	xfs_reclaim_work_queue(mp);
1014 	xfs_ail_push_all(mp->m_ail);
1015 
1016 	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1017 	return 0;
1018 }
1019 
1020 /*
1021  * Return the number of reclaimable inodes in the filesystem for
1022  * the shrinker to determine how much to reclaim.
1023  */
1024 long
1025 xfs_reclaim_inodes_count(
1026 	struct xfs_mount	*mp)
1027 {
1028 	struct xfs_perag	*pag;
1029 	xfs_agnumber_t		ag = 0;
1030 	long			reclaimable = 0;
1031 
1032 	while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1033 		ag = pag->pag_agno + 1;
1034 		reclaimable += pag->pag_ici_reclaimable;
1035 		xfs_perag_put(pag);
1036 	}
1037 	return reclaimable;
1038 }
1039 
1040 STATIC bool
1041 xfs_icwalk_match_id(
1042 	struct xfs_inode	*ip,
1043 	struct xfs_icwalk	*icw)
1044 {
1045 	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1046 	    !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1047 		return false;
1048 
1049 	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1050 	    !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1051 		return false;
1052 
1053 	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1054 	    ip->i_projid != icw->icw_prid)
1055 		return false;
1056 
1057 	return true;
1058 }
1059 
1060 /*
1061  * A union-based inode filtering algorithm. Process the inode if any of the
1062  * criteria match. This is for global/internal scans only.
1063  */
1064 STATIC bool
1065 xfs_icwalk_match_id_union(
1066 	struct xfs_inode	*ip,
1067 	struct xfs_icwalk	*icw)
1068 {
1069 	if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1070 	    uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1071 		return true;
1072 
1073 	if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1074 	    gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1075 		return true;
1076 
1077 	if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1078 	    ip->i_projid == icw->icw_prid)
1079 		return true;
1080 
1081 	return false;
1082 }
1083 
1084 /*
1085  * Is this inode @ip eligible for eof/cow block reclamation, given some
1086  * filtering parameters @icw?  The inode is eligible if @icw is null or
1087  * if the predicate functions match.
1088  */
1089 static bool
1090 xfs_icwalk_match(
1091 	struct xfs_inode	*ip,
1092 	struct xfs_icwalk	*icw)
1093 {
1094 	bool			match;
1095 
1096 	if (!icw)
1097 		return true;
1098 
1099 	if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1100 		match = xfs_icwalk_match_id_union(ip, icw);
1101 	else
1102 		match = xfs_icwalk_match_id(ip, icw);
1103 	if (!match)
1104 		return false;
1105 
1106 	/* skip the inode if the file size is too small */
1107 	if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1108 	    XFS_ISIZE(ip) < icw->icw_min_file_size)
1109 		return false;
1110 
1111 	return true;
1112 }
1113 
1114 /*
1115  * This is a fast pass over the inode cache to try to get reclaim moving on as
1116  * many inodes as possible in a short period of time. It kicks itself every few
1117  * seconds, as well as being kicked by the inode cache shrinker when memory
1118  * goes low.
1119  */
1120 void
1121 xfs_reclaim_worker(
1122 	struct work_struct *work)
1123 {
1124 	struct xfs_mount *mp = container_of(to_delayed_work(work),
1125 					struct xfs_mount, m_reclaim_work);
1126 
1127 	xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1128 	xfs_reclaim_work_queue(mp);
1129 }
1130 
1131 STATIC int
1132 xfs_inode_free_eofblocks(
1133 	struct xfs_inode	*ip,
1134 	struct xfs_icwalk	*icw,
1135 	unsigned int		*lockflags)
1136 {
1137 	bool			wait;
1138 
1139 	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1140 
1141 	if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1142 		return 0;
1143 
1144 	/*
1145 	 * If the mapping is dirty the operation can block and wait for some
1146 	 * time. Unless we are waiting, skip it.
1147 	 */
1148 	if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1149 		return 0;
1150 
1151 	if (!xfs_icwalk_match(ip, icw))
1152 		return 0;
1153 
1154 	/*
1155 	 * If the caller is waiting, return -EAGAIN to keep the background
1156 	 * scanner moving and revisit the inode in a subsequent pass.
1157 	 */
1158 	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1159 		if (wait)
1160 			return -EAGAIN;
1161 		return 0;
1162 	}
1163 	*lockflags |= XFS_IOLOCK_EXCL;
1164 
1165 	if (xfs_can_free_eofblocks(ip, false))
1166 		return xfs_free_eofblocks(ip);
1167 
1168 	/* inode could be preallocated or append-only */
1169 	trace_xfs_inode_free_eofblocks_invalid(ip);
1170 	xfs_inode_clear_eofblocks_tag(ip);
1171 	return 0;
1172 }
1173 
1174 static void
1175 xfs_blockgc_set_iflag(
1176 	struct xfs_inode	*ip,
1177 	unsigned long		iflag)
1178 {
1179 	struct xfs_mount	*mp = ip->i_mount;
1180 	struct xfs_perag	*pag;
1181 
1182 	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1183 
1184 	/*
1185 	 * Don't bother locking the AG and looking up in the radix trees
1186 	 * if we already know that we have the tag set.
1187 	 */
1188 	if (ip->i_flags & iflag)
1189 		return;
1190 	spin_lock(&ip->i_flags_lock);
1191 	ip->i_flags |= iflag;
1192 	spin_unlock(&ip->i_flags_lock);
1193 
1194 	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1195 	spin_lock(&pag->pag_ici_lock);
1196 
1197 	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1198 			XFS_ICI_BLOCKGC_TAG);
1199 
1200 	spin_unlock(&pag->pag_ici_lock);
1201 	xfs_perag_put(pag);
1202 }
1203 
1204 void
1205 xfs_inode_set_eofblocks_tag(
1206 	xfs_inode_t	*ip)
1207 {
1208 	trace_xfs_inode_set_eofblocks_tag(ip);
1209 	return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1210 }
1211 
1212 static void
1213 xfs_blockgc_clear_iflag(
1214 	struct xfs_inode	*ip,
1215 	unsigned long		iflag)
1216 {
1217 	struct xfs_mount	*mp = ip->i_mount;
1218 	struct xfs_perag	*pag;
1219 	bool			clear_tag;
1220 
1221 	ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1222 
1223 	spin_lock(&ip->i_flags_lock);
1224 	ip->i_flags &= ~iflag;
1225 	clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1226 	spin_unlock(&ip->i_flags_lock);
1227 
1228 	if (!clear_tag)
1229 		return;
1230 
1231 	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1232 	spin_lock(&pag->pag_ici_lock);
1233 
1234 	xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1235 			XFS_ICI_BLOCKGC_TAG);
1236 
1237 	spin_unlock(&pag->pag_ici_lock);
1238 	xfs_perag_put(pag);
1239 }
1240 
1241 void
1242 xfs_inode_clear_eofblocks_tag(
1243 	xfs_inode_t	*ip)
1244 {
1245 	trace_xfs_inode_clear_eofblocks_tag(ip);
1246 	return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1247 }
1248 
1249 /*
1250  * Set ourselves up to free CoW blocks from this file.  If it's already clean
1251  * then we can bail out quickly, but otherwise we must back off if the file
1252  * is undergoing some kind of write.
1253  */
1254 static bool
1255 xfs_prep_free_cowblocks(
1256 	struct xfs_inode	*ip)
1257 {
1258 	/*
1259 	 * Just clear the tag if we have an empty cow fork or none at all. It's
1260 	 * possible the inode was fully unshared since it was originally tagged.
1261 	 */
1262 	if (!xfs_inode_has_cow_data(ip)) {
1263 		trace_xfs_inode_free_cowblocks_invalid(ip);
1264 		xfs_inode_clear_cowblocks_tag(ip);
1265 		return false;
1266 	}
1267 
1268 	/*
1269 	 * If the mapping is dirty or under writeback we cannot touch the
1270 	 * CoW fork.  Leave it alone if we're in the midst of a directio.
1271 	 */
1272 	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1273 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1274 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1275 	    atomic_read(&VFS_I(ip)->i_dio_count))
1276 		return false;
1277 
1278 	return true;
1279 }
1280 
1281 /*
1282  * Automatic CoW Reservation Freeing
1283  *
1284  * These functions automatically garbage collect leftover CoW reservations
1285  * that were made on behalf of a cowextsize hint when we start to run out
1286  * of quota or when the reservations sit around for too long.  If the file
1287  * has dirty pages or is undergoing writeback, its CoW reservations will
1288  * be retained.
1289  *
1290  * The actual garbage collection piggybacks off the same code that runs
1291  * the speculative EOF preallocation garbage collector.
1292  */
1293 STATIC int
1294 xfs_inode_free_cowblocks(
1295 	struct xfs_inode	*ip,
1296 	struct xfs_icwalk	*icw,
1297 	unsigned int		*lockflags)
1298 {
1299 	bool			wait;
1300 	int			ret = 0;
1301 
1302 	wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1303 
1304 	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1305 		return 0;
1306 
1307 	if (!xfs_prep_free_cowblocks(ip))
1308 		return 0;
1309 
1310 	if (!xfs_icwalk_match(ip, icw))
1311 		return 0;
1312 
1313 	/*
1314 	 * If the caller is waiting, return -EAGAIN to keep the background
1315 	 * scanner moving and revisit the inode in a subsequent pass.
1316 	 */
1317 	if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1318 	    !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1319 		if (wait)
1320 			return -EAGAIN;
1321 		return 0;
1322 	}
1323 	*lockflags |= XFS_IOLOCK_EXCL;
1324 
1325 	if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1326 		if (wait)
1327 			return -EAGAIN;
1328 		return 0;
1329 	}
1330 	*lockflags |= XFS_MMAPLOCK_EXCL;
1331 
1332 	/*
1333 	 * Check again, nobody else should be able to dirty blocks or change
1334 	 * the reflink iflag now that we have the first two locks held.
1335 	 */
1336 	if (xfs_prep_free_cowblocks(ip))
1337 		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1338 	return ret;
1339 }
1340 
1341 void
1342 xfs_inode_set_cowblocks_tag(
1343 	xfs_inode_t	*ip)
1344 {
1345 	trace_xfs_inode_set_cowblocks_tag(ip);
1346 	return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1347 }
1348 
1349 void
1350 xfs_inode_clear_cowblocks_tag(
1351 	xfs_inode_t	*ip)
1352 {
1353 	trace_xfs_inode_clear_cowblocks_tag(ip);
1354 	return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1355 }
1356 
1357 /* Disable post-EOF and CoW block auto-reclamation. */
1358 void
1359 xfs_blockgc_stop(
1360 	struct xfs_mount	*mp)
1361 {
1362 	struct xfs_perag	*pag;
1363 	xfs_agnumber_t		agno;
1364 
1365 	if (!xfs_clear_blockgc_enabled(mp))
1366 		return;
1367 
1368 	for_each_perag(mp, agno, pag)
1369 		cancel_delayed_work_sync(&pag->pag_blockgc_work);
1370 	trace_xfs_blockgc_stop(mp, __return_address);
1371 }
1372 
1373 /* Enable post-EOF and CoW block auto-reclamation. */
1374 void
1375 xfs_blockgc_start(
1376 	struct xfs_mount	*mp)
1377 {
1378 	struct xfs_perag	*pag;
1379 	xfs_agnumber_t		agno;
1380 
1381 	if (xfs_set_blockgc_enabled(mp))
1382 		return;
1383 
1384 	trace_xfs_blockgc_start(mp, __return_address);
1385 	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1386 		xfs_blockgc_queue(pag);
1387 }
1388 
1389 /* Don't try to run block gc on an inode that's in any of these states. */
1390 #define XFS_BLOCKGC_NOGRAB_IFLAGS	(XFS_INEW | \
1391 					 XFS_NEED_INACTIVE | \
1392 					 XFS_INACTIVATING | \
1393 					 XFS_IRECLAIMABLE | \
1394 					 XFS_IRECLAIM)
1395 /*
1396  * Decide if the given @ip is eligible for garbage collection of speculative
1397  * preallocations, and grab it if so.  Returns true if it's ready to go or
1398  * false if we should just ignore it.
1399  */
1400 static bool
1401 xfs_blockgc_igrab(
1402 	struct xfs_inode	*ip)
1403 {
1404 	struct inode		*inode = VFS_I(ip);
1405 
1406 	ASSERT(rcu_read_lock_held());
1407 
1408 	/* Check for stale RCU freed inode */
1409 	spin_lock(&ip->i_flags_lock);
1410 	if (!ip->i_ino)
1411 		goto out_unlock_noent;
1412 
1413 	if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1414 		goto out_unlock_noent;
1415 	spin_unlock(&ip->i_flags_lock);
1416 
1417 	/* nothing to sync during shutdown */
1418 	if (xfs_is_shutdown(ip->i_mount))
1419 		return false;
1420 
1421 	/* If we can't grab the inode, it must on it's way to reclaim. */
1422 	if (!igrab(inode))
1423 		return false;
1424 
1425 	/* inode is valid */
1426 	return true;
1427 
1428 out_unlock_noent:
1429 	spin_unlock(&ip->i_flags_lock);
1430 	return false;
1431 }
1432 
1433 /* Scan one incore inode for block preallocations that we can remove. */
1434 static int
1435 xfs_blockgc_scan_inode(
1436 	struct xfs_inode	*ip,
1437 	struct xfs_icwalk	*icw)
1438 {
1439 	unsigned int		lockflags = 0;
1440 	int			error;
1441 
1442 	error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1443 	if (error)
1444 		goto unlock;
1445 
1446 	error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1447 unlock:
1448 	if (lockflags)
1449 		xfs_iunlock(ip, lockflags);
1450 	xfs_irele(ip);
1451 	return error;
1452 }
1453 
1454 /* Background worker that trims preallocated space. */
1455 void
1456 xfs_blockgc_worker(
1457 	struct work_struct	*work)
1458 {
1459 	struct xfs_perag	*pag = container_of(to_delayed_work(work),
1460 					struct xfs_perag, pag_blockgc_work);
1461 	struct xfs_mount	*mp = pag->pag_mount;
1462 	int			error;
1463 
1464 	trace_xfs_blockgc_worker(mp, __return_address);
1465 
1466 	error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1467 	if (error)
1468 		xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1469 				pag->pag_agno, error);
1470 	xfs_blockgc_queue(pag);
1471 }
1472 
1473 /*
1474  * Try to free space in the filesystem by purging inactive inodes, eofblocks
1475  * and cowblocks.
1476  */
1477 int
1478 xfs_blockgc_free_space(
1479 	struct xfs_mount	*mp,
1480 	struct xfs_icwalk	*icw)
1481 {
1482 	int			error;
1483 
1484 	trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1485 
1486 	error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1487 	if (error)
1488 		return error;
1489 
1490 	xfs_inodegc_flush(mp);
1491 	return 0;
1492 }
1493 
1494 /*
1495  * Reclaim all the free space that we can by scheduling the background blockgc
1496  * and inodegc workers immediately and waiting for them all to clear.
1497  */
1498 void
1499 xfs_blockgc_flush_all(
1500 	struct xfs_mount	*mp)
1501 {
1502 	struct xfs_perag	*pag;
1503 	xfs_agnumber_t		agno;
1504 
1505 	trace_xfs_blockgc_flush_all(mp, __return_address);
1506 
1507 	/*
1508 	 * For each blockgc worker, move its queue time up to now.  If it
1509 	 * wasn't queued, it will not be requeued.  Then flush whatever's
1510 	 * left.
1511 	 */
1512 	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1513 		mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1514 				&pag->pag_blockgc_work, 0);
1515 
1516 	for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1517 		flush_delayed_work(&pag->pag_blockgc_work);
1518 
1519 	xfs_inodegc_flush(mp);
1520 }
1521 
1522 /*
1523  * Run cow/eofblocks scans on the supplied dquots.  We don't know exactly which
1524  * quota caused an allocation failure, so we make a best effort by including
1525  * each quota under low free space conditions (less than 1% free space) in the
1526  * scan.
1527  *
1528  * Callers must not hold any inode's ILOCK.  If requesting a synchronous scan
1529  * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1530  * MMAPLOCK.
1531  */
1532 int
1533 xfs_blockgc_free_dquots(
1534 	struct xfs_mount	*mp,
1535 	struct xfs_dquot	*udqp,
1536 	struct xfs_dquot	*gdqp,
1537 	struct xfs_dquot	*pdqp,
1538 	unsigned int		iwalk_flags)
1539 {
1540 	struct xfs_icwalk	icw = {0};
1541 	bool			do_work = false;
1542 
1543 	if (!udqp && !gdqp && !pdqp)
1544 		return 0;
1545 
1546 	/*
1547 	 * Run a scan to free blocks using the union filter to cover all
1548 	 * applicable quotas in a single scan.
1549 	 */
1550 	icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1551 
1552 	if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1553 		icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1554 		icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1555 		do_work = true;
1556 	}
1557 
1558 	if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1559 		icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1560 		icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1561 		do_work = true;
1562 	}
1563 
1564 	if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1565 		icw.icw_prid = pdqp->q_id;
1566 		icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1567 		do_work = true;
1568 	}
1569 
1570 	if (!do_work)
1571 		return 0;
1572 
1573 	return xfs_blockgc_free_space(mp, &icw);
1574 }
1575 
1576 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1577 int
1578 xfs_blockgc_free_quota(
1579 	struct xfs_inode	*ip,
1580 	unsigned int		iwalk_flags)
1581 {
1582 	return xfs_blockgc_free_dquots(ip->i_mount,
1583 			xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1584 			xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1585 			xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1586 }
1587 
1588 /* XFS Inode Cache Walking Code */
1589 
1590 /*
1591  * The inode lookup is done in batches to keep the amount of lock traffic and
1592  * radix tree lookups to a minimum. The batch size is a trade off between
1593  * lookup reduction and stack usage. This is in the reclaim path, so we can't
1594  * be too greedy.
1595  */
1596 #define XFS_LOOKUP_BATCH	32
1597 
1598 
1599 /*
1600  * Decide if we want to grab this inode in anticipation of doing work towards
1601  * the goal.
1602  */
1603 static inline bool
1604 xfs_icwalk_igrab(
1605 	enum xfs_icwalk_goal	goal,
1606 	struct xfs_inode	*ip,
1607 	struct xfs_icwalk	*icw)
1608 {
1609 	switch (goal) {
1610 	case XFS_ICWALK_BLOCKGC:
1611 		return xfs_blockgc_igrab(ip);
1612 	case XFS_ICWALK_RECLAIM:
1613 		return xfs_reclaim_igrab(ip, icw);
1614 	default:
1615 		return false;
1616 	}
1617 }
1618 
1619 /*
1620  * Process an inode.  Each processing function must handle any state changes
1621  * made by the icwalk igrab function.  Return -EAGAIN to skip an inode.
1622  */
1623 static inline int
1624 xfs_icwalk_process_inode(
1625 	enum xfs_icwalk_goal	goal,
1626 	struct xfs_inode	*ip,
1627 	struct xfs_perag	*pag,
1628 	struct xfs_icwalk	*icw)
1629 {
1630 	int			error = 0;
1631 
1632 	switch (goal) {
1633 	case XFS_ICWALK_BLOCKGC:
1634 		error = xfs_blockgc_scan_inode(ip, icw);
1635 		break;
1636 	case XFS_ICWALK_RECLAIM:
1637 		xfs_reclaim_inode(ip, pag);
1638 		break;
1639 	}
1640 	return error;
1641 }
1642 
1643 /*
1644  * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1645  * process them in some manner.
1646  */
1647 static int
1648 xfs_icwalk_ag(
1649 	struct xfs_perag	*pag,
1650 	enum xfs_icwalk_goal	goal,
1651 	struct xfs_icwalk	*icw)
1652 {
1653 	struct xfs_mount	*mp = pag->pag_mount;
1654 	uint32_t		first_index;
1655 	int			last_error = 0;
1656 	int			skipped;
1657 	bool			done;
1658 	int			nr_found;
1659 
1660 restart:
1661 	done = false;
1662 	skipped = 0;
1663 	if (goal == XFS_ICWALK_RECLAIM)
1664 		first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1665 	else
1666 		first_index = 0;
1667 	nr_found = 0;
1668 	do {
1669 		struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1670 		int		error = 0;
1671 		int		i;
1672 
1673 		rcu_read_lock();
1674 
1675 		nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1676 				(void **) batch, first_index,
1677 				XFS_LOOKUP_BATCH, goal);
1678 		if (!nr_found) {
1679 			done = true;
1680 			rcu_read_unlock();
1681 			break;
1682 		}
1683 
1684 		/*
1685 		 * Grab the inodes before we drop the lock. if we found
1686 		 * nothing, nr == 0 and the loop will be skipped.
1687 		 */
1688 		for (i = 0; i < nr_found; i++) {
1689 			struct xfs_inode *ip = batch[i];
1690 
1691 			if (done || !xfs_icwalk_igrab(goal, ip, icw))
1692 				batch[i] = NULL;
1693 
1694 			/*
1695 			 * Update the index for the next lookup. Catch
1696 			 * overflows into the next AG range which can occur if
1697 			 * we have inodes in the last block of the AG and we
1698 			 * are currently pointing to the last inode.
1699 			 *
1700 			 * Because we may see inodes that are from the wrong AG
1701 			 * due to RCU freeing and reallocation, only update the
1702 			 * index if it lies in this AG. It was a race that lead
1703 			 * us to see this inode, so another lookup from the
1704 			 * same index will not find it again.
1705 			 */
1706 			if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1707 				continue;
1708 			first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1709 			if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1710 				done = true;
1711 		}
1712 
1713 		/* unlock now we've grabbed the inodes. */
1714 		rcu_read_unlock();
1715 
1716 		for (i = 0; i < nr_found; i++) {
1717 			if (!batch[i])
1718 				continue;
1719 			error = xfs_icwalk_process_inode(goal, batch[i], pag,
1720 					icw);
1721 			if (error == -EAGAIN) {
1722 				skipped++;
1723 				continue;
1724 			}
1725 			if (error && last_error != -EFSCORRUPTED)
1726 				last_error = error;
1727 		}
1728 
1729 		/* bail out if the filesystem is corrupted.  */
1730 		if (error == -EFSCORRUPTED)
1731 			break;
1732 
1733 		cond_resched();
1734 
1735 		if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1736 			icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1737 			if (icw->icw_scan_limit <= 0)
1738 				break;
1739 		}
1740 	} while (nr_found && !done);
1741 
1742 	if (goal == XFS_ICWALK_RECLAIM) {
1743 		if (done)
1744 			first_index = 0;
1745 		WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1746 	}
1747 
1748 	if (skipped) {
1749 		delay(1);
1750 		goto restart;
1751 	}
1752 	return last_error;
1753 }
1754 
1755 /* Walk all incore inodes to achieve a given goal. */
1756 static int
1757 xfs_icwalk(
1758 	struct xfs_mount	*mp,
1759 	enum xfs_icwalk_goal	goal,
1760 	struct xfs_icwalk	*icw)
1761 {
1762 	struct xfs_perag	*pag;
1763 	int			error = 0;
1764 	int			last_error = 0;
1765 	xfs_agnumber_t		agno;
1766 
1767 	for_each_perag_tag(mp, agno, pag, goal) {
1768 		error = xfs_icwalk_ag(pag, goal, icw);
1769 		if (error) {
1770 			last_error = error;
1771 			if (error == -EFSCORRUPTED) {
1772 				xfs_perag_put(pag);
1773 				break;
1774 			}
1775 		}
1776 	}
1777 	return last_error;
1778 	BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1779 }
1780 
1781 #ifdef DEBUG
1782 static void
1783 xfs_check_delalloc(
1784 	struct xfs_inode	*ip,
1785 	int			whichfork)
1786 {
1787 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, whichfork);
1788 	struct xfs_bmbt_irec	got;
1789 	struct xfs_iext_cursor	icur;
1790 
1791 	if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1792 		return;
1793 	do {
1794 		if (isnullstartblock(got.br_startblock)) {
1795 			xfs_warn(ip->i_mount,
1796 	"ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1797 				ip->i_ino,
1798 				whichfork == XFS_DATA_FORK ? "data" : "cow",
1799 				got.br_startoff, got.br_blockcount);
1800 		}
1801 	} while (xfs_iext_next_extent(ifp, &icur, &got));
1802 }
1803 #else
1804 #define xfs_check_delalloc(ip, whichfork)	do { } while (0)
1805 #endif
1806 
1807 /* Schedule the inode for reclaim. */
1808 static void
1809 xfs_inodegc_set_reclaimable(
1810 	struct xfs_inode	*ip)
1811 {
1812 	struct xfs_mount	*mp = ip->i_mount;
1813 	struct xfs_perag	*pag;
1814 
1815 	if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1816 		xfs_check_delalloc(ip, XFS_DATA_FORK);
1817 		xfs_check_delalloc(ip, XFS_COW_FORK);
1818 		ASSERT(0);
1819 	}
1820 
1821 	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1822 	spin_lock(&pag->pag_ici_lock);
1823 	spin_lock(&ip->i_flags_lock);
1824 
1825 	trace_xfs_inode_set_reclaimable(ip);
1826 	ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1827 	ip->i_flags |= XFS_IRECLAIMABLE;
1828 	xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1829 			XFS_ICI_RECLAIM_TAG);
1830 
1831 	spin_unlock(&ip->i_flags_lock);
1832 	spin_unlock(&pag->pag_ici_lock);
1833 	xfs_perag_put(pag);
1834 }
1835 
1836 /*
1837  * Free all speculative preallocations and possibly even the inode itself.
1838  * This is the last chance to make changes to an otherwise unreferenced file
1839  * before incore reclamation happens.
1840  */
1841 static void
1842 xfs_inodegc_inactivate(
1843 	struct xfs_inode	*ip)
1844 {
1845 	trace_xfs_inode_inactivating(ip);
1846 	xfs_inactive(ip);
1847 	xfs_inodegc_set_reclaimable(ip);
1848 }
1849 
1850 void
1851 xfs_inodegc_worker(
1852 	struct work_struct	*work)
1853 {
1854 	struct xfs_inodegc	*gc = container_of(work, struct xfs_inodegc,
1855 							work);
1856 	struct llist_node	*node = llist_del_all(&gc->list);
1857 	struct xfs_inode	*ip, *n;
1858 
1859 	WRITE_ONCE(gc->items, 0);
1860 
1861 	if (!node)
1862 		return;
1863 
1864 	ip = llist_entry(node, struct xfs_inode, i_gclist);
1865 	trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1866 
1867 	WRITE_ONCE(gc->shrinker_hits, 0);
1868 	llist_for_each_entry_safe(ip, n, node, i_gclist) {
1869 		xfs_iflags_set(ip, XFS_INACTIVATING);
1870 		xfs_inodegc_inactivate(ip);
1871 	}
1872 }
1873 
1874 /*
1875  * Force all currently queued inode inactivation work to run immediately, and
1876  * wait for the work to finish. Two pass - queue all the work first pass, wait
1877  * for it in a second pass.
1878  */
1879 void
1880 xfs_inodegc_flush(
1881 	struct xfs_mount	*mp)
1882 {
1883 	struct xfs_inodegc	*gc;
1884 	int			cpu;
1885 
1886 	if (!xfs_is_inodegc_enabled(mp))
1887 		return;
1888 
1889 	trace_xfs_inodegc_flush(mp, __return_address);
1890 
1891 	xfs_inodegc_queue_all(mp);
1892 
1893 	for_each_online_cpu(cpu) {
1894 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
1895 		flush_work(&gc->work);
1896 	}
1897 }
1898 
1899 /*
1900  * Flush all the pending work and then disable the inode inactivation background
1901  * workers and wait for them to stop.
1902  */
1903 void
1904 xfs_inodegc_stop(
1905 	struct xfs_mount	*mp)
1906 {
1907 	struct xfs_inodegc	*gc;
1908 	int			cpu;
1909 
1910 	if (!xfs_clear_inodegc_enabled(mp))
1911 		return;
1912 
1913 	xfs_inodegc_queue_all(mp);
1914 
1915 	for_each_online_cpu(cpu) {
1916 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
1917 		cancel_work_sync(&gc->work);
1918 	}
1919 	trace_xfs_inodegc_stop(mp, __return_address);
1920 }
1921 
1922 /*
1923  * Enable the inode inactivation background workers and schedule deferred inode
1924  * inactivation work if there is any.
1925  */
1926 void
1927 xfs_inodegc_start(
1928 	struct xfs_mount	*mp)
1929 {
1930 	if (xfs_set_inodegc_enabled(mp))
1931 		return;
1932 
1933 	trace_xfs_inodegc_start(mp, __return_address);
1934 	xfs_inodegc_queue_all(mp);
1935 }
1936 
1937 #ifdef CONFIG_XFS_RT
1938 static inline bool
1939 xfs_inodegc_want_queue_rt_file(
1940 	struct xfs_inode	*ip)
1941 {
1942 	struct xfs_mount	*mp = ip->i_mount;
1943 	uint64_t		freertx;
1944 
1945 	if (!XFS_IS_REALTIME_INODE(ip))
1946 		return false;
1947 
1948 	freertx = READ_ONCE(mp->m_sb.sb_frextents);
1949 	return freertx < mp->m_low_rtexts[XFS_LOWSP_5_PCNT];
1950 }
1951 #else
1952 # define xfs_inodegc_want_queue_rt_file(ip)	(false)
1953 #endif /* CONFIG_XFS_RT */
1954 
1955 /*
1956  * Schedule the inactivation worker when:
1957  *
1958  *  - We've accumulated more than one inode cluster buffer's worth of inodes.
1959  *  - There is less than 5% free space left.
1960  *  - Any of the quotas for this inode are near an enforcement limit.
1961  */
1962 static inline bool
1963 xfs_inodegc_want_queue_work(
1964 	struct xfs_inode	*ip,
1965 	unsigned int		items)
1966 {
1967 	struct xfs_mount	*mp = ip->i_mount;
1968 
1969 	if (items > mp->m_ino_geo.inodes_per_cluster)
1970 		return true;
1971 
1972 	if (__percpu_counter_compare(&mp->m_fdblocks,
1973 				mp->m_low_space[XFS_LOWSP_5_PCNT],
1974 				XFS_FDBLOCKS_BATCH) < 0)
1975 		return true;
1976 
1977 	if (xfs_inodegc_want_queue_rt_file(ip))
1978 		return true;
1979 
1980 	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
1981 		return true;
1982 
1983 	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
1984 		return true;
1985 
1986 	if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
1987 		return true;
1988 
1989 	return false;
1990 }
1991 
1992 /*
1993  * Upper bound on the number of inodes in each AG that can be queued for
1994  * inactivation at any given time, to avoid monopolizing the workqueue.
1995  */
1996 #define XFS_INODEGC_MAX_BACKLOG		(4 * XFS_INODES_PER_CHUNK)
1997 
1998 /*
1999  * Make the frontend wait for inactivations when:
2000  *
2001  *  - Memory shrinkers queued the inactivation worker and it hasn't finished.
2002  *  - The queue depth exceeds the maximum allowable percpu backlog.
2003  *
2004  * Note: If the current thread is running a transaction, we don't ever want to
2005  * wait for other transactions because that could introduce a deadlock.
2006  */
2007 static inline bool
2008 xfs_inodegc_want_flush_work(
2009 	struct xfs_inode	*ip,
2010 	unsigned int		items,
2011 	unsigned int		shrinker_hits)
2012 {
2013 	if (current->journal_info)
2014 		return false;
2015 
2016 	if (shrinker_hits > 0)
2017 		return true;
2018 
2019 	if (items > XFS_INODEGC_MAX_BACKLOG)
2020 		return true;
2021 
2022 	return false;
2023 }
2024 
2025 /*
2026  * Queue a background inactivation worker if there are inodes that need to be
2027  * inactivated and higher level xfs code hasn't disabled the background
2028  * workers.
2029  */
2030 static void
2031 xfs_inodegc_queue(
2032 	struct xfs_inode	*ip)
2033 {
2034 	struct xfs_mount	*mp = ip->i_mount;
2035 	struct xfs_inodegc	*gc;
2036 	int			items;
2037 	unsigned int		shrinker_hits;
2038 
2039 	trace_xfs_inode_set_need_inactive(ip);
2040 	spin_lock(&ip->i_flags_lock);
2041 	ip->i_flags |= XFS_NEED_INACTIVE;
2042 	spin_unlock(&ip->i_flags_lock);
2043 
2044 	gc = get_cpu_ptr(mp->m_inodegc);
2045 	llist_add(&ip->i_gclist, &gc->list);
2046 	items = READ_ONCE(gc->items);
2047 	WRITE_ONCE(gc->items, items + 1);
2048 	shrinker_hits = READ_ONCE(gc->shrinker_hits);
2049 	put_cpu_ptr(gc);
2050 
2051 	if (!xfs_is_inodegc_enabled(mp))
2052 		return;
2053 
2054 	if (xfs_inodegc_want_queue_work(ip, items)) {
2055 		trace_xfs_inodegc_queue(mp, __return_address);
2056 		queue_work(mp->m_inodegc_wq, &gc->work);
2057 	}
2058 
2059 	if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2060 		trace_xfs_inodegc_throttle(mp, __return_address);
2061 		flush_work(&gc->work);
2062 	}
2063 }
2064 
2065 /*
2066  * Fold the dead CPU inodegc queue into the current CPUs queue.
2067  */
2068 void
2069 xfs_inodegc_cpu_dead(
2070 	struct xfs_mount	*mp,
2071 	unsigned int		dead_cpu)
2072 {
2073 	struct xfs_inodegc	*dead_gc, *gc;
2074 	struct llist_node	*first, *last;
2075 	unsigned int		count = 0;
2076 
2077 	dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2078 	cancel_work_sync(&dead_gc->work);
2079 
2080 	if (llist_empty(&dead_gc->list))
2081 		return;
2082 
2083 	first = dead_gc->list.first;
2084 	last = first;
2085 	while (last->next) {
2086 		last = last->next;
2087 		count++;
2088 	}
2089 	dead_gc->list.first = NULL;
2090 	dead_gc->items = 0;
2091 
2092 	/* Add pending work to current CPU */
2093 	gc = get_cpu_ptr(mp->m_inodegc);
2094 	llist_add_batch(first, last, &gc->list);
2095 	count += READ_ONCE(gc->items);
2096 	WRITE_ONCE(gc->items, count);
2097 	put_cpu_ptr(gc);
2098 
2099 	if (xfs_is_inodegc_enabled(mp)) {
2100 		trace_xfs_inodegc_queue(mp, __return_address);
2101 		queue_work(mp->m_inodegc_wq, &gc->work);
2102 	}
2103 }
2104 
2105 /*
2106  * We set the inode flag atomically with the radix tree tag.  Once we get tag
2107  * lookups on the radix tree, this inode flag can go away.
2108  *
2109  * We always use background reclaim here because even if the inode is clean, it
2110  * still may be under IO and hence we have wait for IO completion to occur
2111  * before we can reclaim the inode. The background reclaim path handles this
2112  * more efficiently than we can here, so simply let background reclaim tear down
2113  * all inodes.
2114  */
2115 void
2116 xfs_inode_mark_reclaimable(
2117 	struct xfs_inode	*ip)
2118 {
2119 	struct xfs_mount	*mp = ip->i_mount;
2120 	bool			need_inactive;
2121 
2122 	XFS_STATS_INC(mp, vn_reclaim);
2123 
2124 	/*
2125 	 * We should never get here with any of the reclaim flags already set.
2126 	 */
2127 	ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2128 
2129 	need_inactive = xfs_inode_needs_inactive(ip);
2130 	if (need_inactive) {
2131 		xfs_inodegc_queue(ip);
2132 		return;
2133 	}
2134 
2135 	/* Going straight to reclaim, so drop the dquots. */
2136 	xfs_qm_dqdetach(ip);
2137 	xfs_inodegc_set_reclaimable(ip);
2138 }
2139 
2140 /*
2141  * Register a phony shrinker so that we can run background inodegc sooner when
2142  * there's memory pressure.  Inactivation does not itself free any memory but
2143  * it does make inodes reclaimable, which eventually frees memory.
2144  *
2145  * The count function, seek value, and batch value are crafted to trigger the
2146  * scan function during the second round of scanning.  Hopefully this means
2147  * that we reclaimed enough memory that initiating metadata transactions won't
2148  * make things worse.
2149  */
2150 #define XFS_INODEGC_SHRINKER_COUNT	(1UL << DEF_PRIORITY)
2151 #define XFS_INODEGC_SHRINKER_BATCH	((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2152 
2153 static unsigned long
2154 xfs_inodegc_shrinker_count(
2155 	struct shrinker		*shrink,
2156 	struct shrink_control	*sc)
2157 {
2158 	struct xfs_mount	*mp = container_of(shrink, struct xfs_mount,
2159 						   m_inodegc_shrinker);
2160 	struct xfs_inodegc	*gc;
2161 	int			cpu;
2162 
2163 	if (!xfs_is_inodegc_enabled(mp))
2164 		return 0;
2165 
2166 	for_each_online_cpu(cpu) {
2167 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
2168 		if (!llist_empty(&gc->list))
2169 			return XFS_INODEGC_SHRINKER_COUNT;
2170 	}
2171 
2172 	return 0;
2173 }
2174 
2175 static unsigned long
2176 xfs_inodegc_shrinker_scan(
2177 	struct shrinker		*shrink,
2178 	struct shrink_control	*sc)
2179 {
2180 	struct xfs_mount	*mp = container_of(shrink, struct xfs_mount,
2181 						   m_inodegc_shrinker);
2182 	struct xfs_inodegc	*gc;
2183 	int			cpu;
2184 	bool			no_items = true;
2185 
2186 	if (!xfs_is_inodegc_enabled(mp))
2187 		return SHRINK_STOP;
2188 
2189 	trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2190 
2191 	for_each_online_cpu(cpu) {
2192 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
2193 		if (!llist_empty(&gc->list)) {
2194 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
2195 
2196 			WRITE_ONCE(gc->shrinker_hits, h + 1);
2197 			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
2198 			no_items = false;
2199 		}
2200 	}
2201 
2202 	/*
2203 	 * If there are no inodes to inactivate, we don't want the shrinker
2204 	 * to think there's deferred work to call us back about.
2205 	 */
2206 	if (no_items)
2207 		return LONG_MAX;
2208 
2209 	return SHRINK_STOP;
2210 }
2211 
2212 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2213 int
2214 xfs_inodegc_register_shrinker(
2215 	struct xfs_mount	*mp)
2216 {
2217 	struct shrinker		*shrink = &mp->m_inodegc_shrinker;
2218 
2219 	shrink->count_objects = xfs_inodegc_shrinker_count;
2220 	shrink->scan_objects = xfs_inodegc_shrinker_scan;
2221 	shrink->seeks = 0;
2222 	shrink->flags = SHRINKER_NONSLAB;
2223 	shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2224 
2225 	return register_shrinker(shrink);
2226 }
2227