1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Copyright (c) 2000-2006 Silicon Graphics, Inc.
4	* All Rights Reserved.
5	*/
6	#include <linux/iversion.h>
7
8	#include "xfs.h"
9	#include "xfs_fs.h"
10	#include "xfs_shared.h"
11	#include "xfs_format.h"
12	#include "xfs_log_format.h"
13	#include "xfs_trans_resv.h"
14	#include "xfs_mount.h"
15	#include "xfs_defer.h"
16	#include "xfs_inode.h"
17	#include "xfs_dir2.h"
18	#include "xfs_attr.h"
19	#include "xfs_bit.h"
20	#include "xfs_trans_space.h"
21	#include "xfs_trans.h"
22	#include "xfs_buf_item.h"
23	#include "xfs_inode_item.h"
24	#include "xfs_iunlink_item.h"
25	#include "xfs_ialloc.h"
26	#include "xfs_bmap.h"
27	#include "xfs_bmap_util.h"
28	#include "xfs_errortag.h"
29	#include "xfs_error.h"
30	#include "xfs_quota.h"
31	#include "xfs_filestream.h"
32	#include "xfs_trace.h"
33	#include "xfs_icache.h"
34	#include "xfs_symlink.h"
35	#include "xfs_trans_priv.h"
36	#include "xfs_log.h"
37	#include "xfs_bmap_btree.h"
38	#include "xfs_reflink.h"
39	#include "xfs_ag.h"
40	#include "xfs_log_priv.h"
41	#include "xfs_health.h"
42	#include "xfs_pnfs.h"
43	#include "xfs_parent.h"
44	#include "xfs_xattr.h"
45	#include "xfs_inode_util.h"
46	#include "xfs_metafile.h"
47
48	struct kmem_cache *xfs_inode_cache;
49
50	/*
51	* These two are wrapper routines around the xfs_ilock() routine used to
52	* centralize some grungy code. They are used in places that wish to lock the
53	* inode solely for reading the extents. The reason these places can't just
54	* call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
55	* bringing in of the extents from disk for a file in b-tree format. If the
56	* inode is in b-tree format, then we need to lock the inode exclusively until
57	* the extents are read in. Locking it exclusively all the time would limit
58	* our parallelism unnecessarily, though. What we do instead is check to see
59	* if the extents have been read in yet, and only lock the inode exclusively
60	* if they have not.
61	*
62	* The functions return a value which should be given to the corresponding
63	* xfs_iunlock() call.
64	*/
65	uint
66	xfs_ilock_data_map_shared(
67	struct xfs_inode *ip)
68	{
69	uint lock_mode = XFS_ILOCK_SHARED;
70
71	if (xfs_need_iread_extents(&ip->i_df))
72	lock_mode = XFS_ILOCK_EXCL;
73	xfs_ilock(ip, lock_mode);
74	return lock_mode;
75	}
76
77	uint
78	xfs_ilock_attr_map_shared(
79	struct xfs_inode *ip)
80	{
81	uint lock_mode = XFS_ILOCK_SHARED;
82
83	if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
84	lock_mode = XFS_ILOCK_EXCL;
85	xfs_ilock(ip, lock_mode);
86	return lock_mode;
87	}
88
89	/*
90	* You can't set both SHARED and EXCL for the same lock,
91	* and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
92	* XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
93	* to set in lock_flags.
94	*/
95	static inline void
96	xfs_lock_flags_assert(
97	uint lock_flags)
98	{
99	ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
100	(XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
101	ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
102	(XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
103	ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
104	(XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
105	ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
106	ASSERT(lock_flags != 0);
107	}
108
109	/*
110	* In addition to i_rwsem in the VFS inode, the xfs inode contains 2
111	* multi-reader locks: invalidate_lock and the i_lock. This routine allows
112	* various combinations of the locks to be obtained.
113	*
114	* The 3 locks should always be ordered so that the IO lock is obtained first,
115	* the mmap lock second and the ilock last in order to prevent deadlock.
116	*
117	* Basic locking order:
118	*
119	* i_rwsem -> invalidate_lock -> page_lock -> i_ilock
120	*
121	* mmap_lock locking order:
122	*
123	* i_rwsem -> page lock -> mmap_lock
124	* mmap_lock -> invalidate_lock -> page_lock
125	*
126	* The difference in mmap_lock locking order mean that we cannot hold the
127	* invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
128	* can fault in pages during copy in/out (for buffered IO) or require the
129	* mmap_lock in get_user_pages() to map the user pages into the kernel address
130	* space for direct IO. Similarly the i_rwsem cannot be taken inside a page
131	* fault because page faults already hold the mmap_lock.
132	*
133	* Hence to serialise fully against both syscall and mmap based IO, we need to
134	* take both the i_rwsem and the invalidate_lock. These locks should *only* be
135	* both taken in places where we need to invalidate the page cache in a race
136	* free manner (e.g. truncate, hole punch and other extent manipulation
137	* functions).
138	*/
139	void
140	xfs_ilock(
141	xfs_inode_t *ip,
142	uint lock_flags)
143	{
144	trace_xfs_ilock(ip, lock_flags, _RET_IP_);
145
146	xfs_lock_flags_assert(lock_flags);
147
148	if (lock_flags & XFS_IOLOCK_EXCL) {
149	down_write_nested(sem: &VFS_I(ip)->i_rwsem,
150	XFS_IOLOCK_DEP(lock_flags));
151	} else if (lock_flags & XFS_IOLOCK_SHARED) {
152	down_read_nested(sem: &VFS_I(ip)->i_rwsem,
153	XFS_IOLOCK_DEP(lock_flags));
154	}
155
156	if (lock_flags & XFS_MMAPLOCK_EXCL) {
157	down_write_nested(sem: &VFS_I(ip)->i_mapping->invalidate_lock,
158	XFS_MMAPLOCK_DEP(lock_flags));
159	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
160	down_read_nested(sem: &VFS_I(ip)->i_mapping->invalidate_lock,
161	XFS_MMAPLOCK_DEP(lock_flags));
162	}
163
164	if (lock_flags & XFS_ILOCK_EXCL)
165	down_write_nested(sem: &ip->i_lock, XFS_ILOCK_DEP(lock_flags));
166	else if (lock_flags & XFS_ILOCK_SHARED)
167	down_read_nested(sem: &ip->i_lock, XFS_ILOCK_DEP(lock_flags));
168	}
169
170	/*
171	* This is just like xfs_ilock(), except that the caller
172	* is guaranteed not to sleep. It returns 1 if it gets
173	* the requested locks and 0 otherwise. If the IO lock is
174	* obtained but the inode lock cannot be, then the IO lock
175	* is dropped before returning.
176	*
177	* ip -- the inode being locked
178	* lock_flags -- this parameter indicates the inode's locks to be
179	* to be locked. See the comment for xfs_ilock() for a list
180	* of valid values.
181	*/
182	int
183	xfs_ilock_nowait(
184	xfs_inode_t *ip,
185	uint lock_flags)
186	{
187	trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
188
189	xfs_lock_flags_assert(lock_flags);
190
191	if (lock_flags & XFS_IOLOCK_EXCL) {
192	if (!down_write_trylock(sem: &VFS_I(ip)->i_rwsem))
193	goto out;
194	} else if (lock_flags & XFS_IOLOCK_SHARED) {
195	if (!down_read_trylock(sem: &VFS_I(ip)->i_rwsem))
196	goto out;
197	}
198
199	if (lock_flags & XFS_MMAPLOCK_EXCL) {
200	if (!down_write_trylock(sem: &VFS_I(ip)->i_mapping->invalidate_lock))
201	goto out_undo_iolock;
202	} else if (lock_flags & XFS_MMAPLOCK_SHARED) {
203	if (!down_read_trylock(sem: &VFS_I(ip)->i_mapping->invalidate_lock))
204	goto out_undo_iolock;
205	}
206
207	if (lock_flags & XFS_ILOCK_EXCL) {
208	if (!down_write_trylock(sem: &ip->i_lock))
209	goto out_undo_mmaplock;
210	} else if (lock_flags & XFS_ILOCK_SHARED) {
211	if (!down_read_trylock(sem: &ip->i_lock))
212	goto out_undo_mmaplock;
213	}
214	return 1;
215
216	out_undo_mmaplock:
217	if (lock_flags & XFS_MMAPLOCK_EXCL)
218	up_write(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
219	else if (lock_flags & XFS_MMAPLOCK_SHARED)
220	up_read(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
221	out_undo_iolock:
222	if (lock_flags & XFS_IOLOCK_EXCL)
223	up_write(sem: &VFS_I(ip)->i_rwsem);
224	else if (lock_flags & XFS_IOLOCK_SHARED)
225	up_read(sem: &VFS_I(ip)->i_rwsem);
226	out:
227	return 0;
228	}
229
230	/*
231	* xfs_iunlock() is used to drop the inode locks acquired with
232	* xfs_ilock() and xfs_ilock_nowait(). The caller must pass
233	* in the flags given to xfs_ilock() or xfs_ilock_nowait() so
234	* that we know which locks to drop.
235	*
236	* ip -- the inode being unlocked
237	* lock_flags -- this parameter indicates the inode's locks to be
238	* to be unlocked. See the comment for xfs_ilock() for a list
239	* of valid values for this parameter.
240	*
241	*/
242	void
243	xfs_iunlock(
244	xfs_inode_t *ip,
245	uint lock_flags)
246	{
247	xfs_lock_flags_assert(lock_flags);
248
249	if (lock_flags & XFS_IOLOCK_EXCL)
250	up_write(sem: &VFS_I(ip)->i_rwsem);
251	else if (lock_flags & XFS_IOLOCK_SHARED)
252	up_read(sem: &VFS_I(ip)->i_rwsem);
253
254	if (lock_flags & XFS_MMAPLOCK_EXCL)
255	up_write(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
256	else if (lock_flags & XFS_MMAPLOCK_SHARED)
257	up_read(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
258
259	if (lock_flags & XFS_ILOCK_EXCL)
260	up_write(sem: &ip->i_lock);
261	else if (lock_flags & XFS_ILOCK_SHARED)
262	up_read(sem: &ip->i_lock);
263
264	trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
265	}
266
267	/*
268	* give up write locks. the i/o lock cannot be held nested
269	* if it is being demoted.
270	*/
271	void
272	xfs_ilock_demote(
273	xfs_inode_t *ip,
274	uint lock_flags)
275	{
276	ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
277	ASSERT((lock_flags &
278	~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
279
280	if (lock_flags & XFS_ILOCK_EXCL)
281	downgrade_write(sem: &ip->i_lock);
282	if (lock_flags & XFS_MMAPLOCK_EXCL)
283	downgrade_write(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
284	if (lock_flags & XFS_IOLOCK_EXCL)
285	downgrade_write(sem: &VFS_I(ip)->i_rwsem);
286
287	trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
288	}
289
290	void
291	xfs_assert_ilocked(
292	struct xfs_inode *ip,
293	uint lock_flags)
294	{
295	/*
296	* Sometimes we assert the ILOCK is held exclusively, but we're in
297	* a workqueue, so lockdep doesn't know we're the owner.
298	*/
299	if (lock_flags & XFS_ILOCK_SHARED)
300	rwsem_assert_held(sem: &ip->i_lock);
301	else if (lock_flags & XFS_ILOCK_EXCL)
302	rwsem_assert_held_write_nolockdep(sem: &ip->i_lock);
303
304	if (lock_flags & XFS_MMAPLOCK_SHARED)
305	rwsem_assert_held(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
306	else if (lock_flags & XFS_MMAPLOCK_EXCL)
307	rwsem_assert_held_write(sem: &VFS_I(ip)->i_mapping->invalidate_lock);
308
309	if (lock_flags & XFS_IOLOCK_SHARED)
310	rwsem_assert_held(sem: &VFS_I(ip)->i_rwsem);
311	else if (lock_flags & XFS_IOLOCK_EXCL)
312	rwsem_assert_held_write(sem: &VFS_I(ip)->i_rwsem);
313	}
314
315	/*
316	* xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
317	* DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
318	* when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
319	* errors and warnings.
320	*/
321	#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
322	static bool
323	xfs_lockdep_subclass_ok(
324	int subclass)
325	{
326	return subclass < MAX_LOCKDEP_SUBCLASSES;
327	}
328	#else
329	#define xfs_lockdep_subclass_ok(subclass) (true)
330	#endif
331
332	/*
333	* Bump the subclass so xfs_lock_inodes() acquires each lock with a different
334	* value. This can be called for any type of inode lock combination, including
335	* parent locking. Care must be taken to ensure we don't overrun the subclass
336	* storage fields in the class mask we build.
337	*/
338	static inline uint
339	xfs_lock_inumorder(
340	uint lock_mode,
341	uint subclass)
342	{
343	uint class = 0;
344
345	ASSERT(!(lock_mode & XFS_ILOCK_PARENT));
346	ASSERT(xfs_lockdep_subclass_ok(subclass));
347
348	if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
349	ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
350	class += subclass << XFS_IOLOCK_SHIFT;
351	}
352
353	if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
354	ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
355	class += subclass << XFS_MMAPLOCK_SHIFT;
356	}
357
358	if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
359	ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
360	class += subclass << XFS_ILOCK_SHIFT;
361	}
362
363	return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
364	}
365
366	/*
367	* The following routine will lock n inodes in exclusive mode. We assume the
368	* caller calls us with the inodes in i_ino order.
369	*
370	* We need to detect deadlock where an inode that we lock is in the AIL and we
371	* start waiting for another inode that is locked by a thread in a long running
372	* transaction (such as truncate). This can result in deadlock since the long
373	* running trans might need to wait for the inode we just locked in order to
374	* push the tail and free space in the log.
375	*
376	* xfs_lock_inodes() can only be used to lock one type of lock at a time -
377	* the iolock, the mmaplock or the ilock, but not more than one at a time. If we
378	* lock more than one at a time, lockdep will report false positives saying we
379	* have violated locking orders.
380	*/
381	void
382	xfs_lock_inodes(
383	struct xfs_inode **ips,
384	int inodes,
385	uint lock_mode)
386	{
387	int attempts = 0;
388	uint i;
389	int j;
390	bool try_lock;
391	struct xfs_log_item *lp;
392
393	/*
394	* Currently supports between 2 and 5 inodes with exclusive locking. We
395	* support an arbitrary depth of locking here, but absolute limits on
396	* inodes depend on the type of locking and the limits placed by
397	* lockdep annotations in xfs_lock_inumorder. These are all checked by
398	* the asserts.
399	*/
400	ASSERT(ips && inodes >= 2 && inodes <= 5);
401	ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
402	XFS_ILOCK_EXCL));
403	ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
404	XFS_ILOCK_SHARED)));
405	ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
406	inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
407	ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
408	inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
409
410	if (lock_mode & XFS_IOLOCK_EXCL) {
411	ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
412	} else if (lock_mode & XFS_MMAPLOCK_EXCL)
413	ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
414
415	again:
416	try_lock = false;
417	i = 0;
418	for (; i < inodes; i++) {
419	ASSERT(ips[i]);
420
421	if (i && (ips[i] == ips[i - 1])) /* Already locked */
422	continue;
423
424	/*
425	* If try_lock is not set yet, make sure all locked inodes are
426	* not in the AIL. If any are, set try_lock to be used later.
427	*/
428	if (!try_lock) {
429	for (j = (i - 1); j >= 0 && !try_lock; j--) {
430	lp = &ips[j]->i_itemp->ili_item;
431	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
432	try_lock = true;
433	}
434	}
435
436	/*
437	* If any of the previous locks we have locked is in the AIL,
438	* we must TRY to get the second and subsequent locks. If
439	* we can't get any, we must release all we have
440	* and try again.
441	*/
442	if (!try_lock) {
443	xfs_ilock(ip: ips[i], lock_flags: xfs_lock_inumorder(lock_mode, subclass: i));
444	continue;
445	}
446
447	/* try_lock means we have an inode locked that is in the AIL. */
448	ASSERT(i != 0);
449	if (xfs_ilock_nowait(ip: ips[i], lock_flags: xfs_lock_inumorder(lock_mode, subclass: i)))
450	continue;
451
452	/*
453	* Unlock all previous guys and try again. xfs_iunlock will try
454	* to push the tail if the inode is in the AIL.
455	*/
456	attempts++;
457	for (j = i - 1; j >= 0; j--) {
458	/*
459	* Check to see if we've already unlocked this one. Not
460	* the first one going back, and the inode ptr is the
461	* same.
462	*/
463	if (j != (i - 1) && ips[j] == ips[j + 1])
464	continue;
465
466	xfs_iunlock(ip: ips[j], lock_flags: lock_mode);
467	}
468
469	if ((attempts % 5) == 0) {
470	delay(ticks: 1); /* Don't just spin the CPU */
471	}
472	goto again;
473	}
474	}
475
476	/*
477	* xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
478	* mmaplock must be double-locked separately since we use i_rwsem and
479	* invalidate_lock for that. We now support taking one lock EXCL and the
480	* other SHARED.
481	*/
482	void
483	xfs_lock_two_inodes(
484	struct xfs_inode *ip0,
485	uint ip0_mode,
486	struct xfs_inode *ip1,
487	uint ip1_mode)
488	{
489	int attempts = 0;
490	struct xfs_log_item *lp;
491
492	ASSERT(hweight32(ip0_mode) == 1);
493	ASSERT(hweight32(ip1_mode) == 1);
494	ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
495	ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
496	ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
497	ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
498	ASSERT(ip0->i_ino != ip1->i_ino);
499
500	if (ip0->i_ino > ip1->i_ino) {
501	swap(ip0, ip1);
502	swap(ip0_mode, ip1_mode);
503	}
504
505	again:
506	xfs_ilock(ip: ip0, lock_flags: xfs_lock_inumorder(lock_mode: ip0_mode, subclass: 0));
507
508	/*
509	* If the first lock we have locked is in the AIL, we must TRY to get
510	* the second lock. If we can't get it, we must release the first one
511	* and try again.
512	*/
513	lp = &ip0->i_itemp->ili_item;
514	if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
515	if (!xfs_ilock_nowait(ip: ip1, lock_flags: xfs_lock_inumorder(lock_mode: ip1_mode, subclass: 1))) {
516	xfs_iunlock(ip: ip0, lock_flags: ip0_mode);
517	if ((++attempts % 5) == 0)
518	delay(ticks: 1); /* Don't just spin the CPU */
519	goto again;
520	}
521	} else {
522	xfs_ilock(ip: ip1, lock_flags: xfs_lock_inumorder(lock_mode: ip1_mode, subclass: 1));
523	}
524	}
525
526	/*
527	* Lookups up an inode from "name". If ci_name is not NULL, then a CI match
528	* is allowed, otherwise it has to be an exact match. If a CI match is found,
529	* ci_name->name will point to a the actual name (caller must free) or
530	* will be set to NULL if an exact match is found.
531	*/
532	int
533	xfs_lookup(
534	struct xfs_inode *dp,
535	const struct xfs_name *name,
536	struct xfs_inode **ipp,
537	struct xfs_name *ci_name)
538	{
539	xfs_ino_t inum;
540	int error;
541
542	trace_xfs_lookup(dp, xfs_lookup: name);
543
544	if (xfs_is_shutdown(mp: dp->i_mount))
545	return -EIO;
546	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
547	return -EIO;
548
549	error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
550	if (error)
551	goto out_unlock;
552
553	error = xfs_iget(mp: dp->i_mount, NULL, ino: inum, flags: 0, lock_flags: 0, ipp);
554	if (error)
555	goto out_free_name;
556
557	/*
558	* Fail if a directory entry in the regular directory tree points to
559	* a metadata file.
560	*/
561	if (XFS_IS_CORRUPT(dp->i_mount, xfs_is_metadir_inode(*ipp))) {
562	xfs_fs_mark_sick(dp->i_mount, XFS_SICK_FS_METADIR);
563	error = -EFSCORRUPTED;
564	goto out_irele;
565	}
566
567	return 0;
568
569	out_irele:
570	xfs_irele(ip: *ipp);
571	out_free_name:
572	if (ci_name)
573	kfree(objp: ci_name->name);
574	out_unlock:
575	*ipp = NULL;
576	return error;
577	}
578
579	/*
580	* Initialise a newly allocated inode and return the in-core inode to the
581	* caller locked exclusively.
582	*
583	* Caller is responsible for unlocking the inode manually upon return
584	*/
585	int
586	xfs_icreate(
587	struct xfs_trans *tp,
588	xfs_ino_t ino,
589	const struct xfs_icreate_args *args,
590	struct xfs_inode **ipp)
591	{
592	struct xfs_mount *mp = tp->t_mountp;
593	struct xfs_inode *ip = NULL;
594	int error;
595
596	/*
597	* Get the in-core inode with the lock held exclusively to prevent
598	* others from looking at until we're done.
599	*/
600	error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, ipp: &ip);
601	if (error)
602	return error;
603
604	ASSERT(ip != NULL);
605	xfs_trans_ijoin(tp, ip, 0);
606	xfs_inode_init(tp, args, ip);
607
608	/* now that we have an i_mode we can setup the inode structure */
609	xfs_setup_inode(ip);
610
611	*ipp = ip;
612	return 0;
613	}
614
615	/* Return dquots for the ids that will be assigned to a new file. */
616	int
617	xfs_icreate_dqalloc(
618	const struct xfs_icreate_args *args,
619	struct xfs_dquot **udqpp,
620	struct xfs_dquot **gdqpp,
621	struct xfs_dquot **pdqpp)
622	{
623	struct inode *dir = VFS_I(ip: args->pip);
624	kuid_t uid = GLOBAL_ROOT_UID;
625	kgid_t gid = GLOBAL_ROOT_GID;
626	prid_t prid = 0;
627	unsigned int flags = XFS_QMOPT_QUOTALL;
628
629	if (args->idmap) {
630	/*
631	* The uid/gid computation code must match what the VFS uses to
632	* assign i_[ug]id. INHERIT adjusts the gid computation for
633	* setgid/grpid systems.
634	*/
635	uid = mapped_fsuid(idmap: args->idmap, fs_userns: i_user_ns(inode: dir));
636	gid = mapped_fsgid(idmap: args->idmap, fs_userns: i_user_ns(inode: dir));
637	prid = xfs_get_initial_prid(args->pip);
638	flags |= XFS_QMOPT_INHERIT;
639	}
640
641	*udqpp = *gdqpp = *pdqpp = NULL;
642
643	return xfs_qm_vop_dqalloc(args->pip, uid, gid, prid, flags, udqpp,
644	gdqpp, pdqpp);
645	}
646
647	int
648	xfs_create(
649	const struct xfs_icreate_args *args,
650	struct xfs_name *name,
651	struct xfs_inode **ipp)
652	{
653	struct xfs_inode *dp = args->pip;
654	struct xfs_dir_update du = {
655	.dp = dp,
656	.name = name,
657	};
658	struct xfs_mount *mp = dp->i_mount;
659	struct xfs_trans *tp = NULL;
660	struct xfs_dquot *udqp;
661	struct xfs_dquot *gdqp;
662	struct xfs_dquot *pdqp;
663	struct xfs_trans_res *tres;
664	xfs_ino_t ino;
665	bool unlock_dp_on_error = false;
666	bool is_dir = S_ISDIR(args->mode);
667	uint resblks;
668	int error;
669
670	trace_xfs_create(dp, xfs_create: name);
671
672	if (xfs_is_shutdown(mp))
673	return -EIO;
674	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
675	return -EIO;
676
677	/* Make sure that we have allocated dquot(s) on disk. */
678	error = xfs_icreate_dqalloc(args, udqpp: &udqp, gdqpp: &gdqp, pdqpp: &pdqp);
679	if (error)
680	return error;
681
682	if (is_dir) {
683	resblks = xfs_mkdir_space_res(mp, name->len);
684	tres = &M_RES(mp)->tr_mkdir;
685	} else {
686	resblks = xfs_create_space_res(mp, name->len);
687	tres = &M_RES(mp)->tr_create;
688	}
689
690	error = xfs_parent_start(mp, &du.ppargs);
691	if (error)
692	goto out_release_dquots;
693
694	/*
695	* Initially assume that the file does not exist and
696	* reserve the resources for that case. If that is not
697	* the case we'll drop the one we have and get a more
698	* appropriate transaction later.
699	*/
700	error = xfs_trans_alloc_icreate(mp, resv: tres, udqp, gdqp, pdqp, dblocks: resblks,
701	tpp: &tp);
702	if (error == -ENOSPC) {
703	/* flush outstanding delalloc blocks and retry */
704	xfs_flush_inodes(mp);
705	error = xfs_trans_alloc_icreate(mp, resv: tres, udqp, gdqp, pdqp,
706	dblocks: resblks, tpp: &tp);
707	}
708	if (error)
709	goto out_parent;
710
711	xfs_ilock(ip: dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
712	unlock_dp_on_error = true;
713
714	/*
715	* A newly created regular or special file just has one directory
716	* entry pointing to them, but a directory also the "." entry
717	* pointing to itself.
718	*/
719	error = xfs_dialloc(&tp, args, &ino);
720	if (!error)
721	error = xfs_icreate(tp, ino, args, ipp: &du.ip);
722	if (error)
723	goto out_trans_cancel;
724
725	/*
726	* Now we join the directory inode to the transaction. We do not do it
727	* earlier because xfs_dialloc might commit the previous transaction
728	* (and release all the locks). An error from here on will result in
729	* the transaction cancel unlocking dp so don't do it explicitly in the
730	* error path.
731	*/
732	xfs_trans_ijoin(tp, dp, 0);
733
734	error = xfs_dir_create_child(tp, resblks, &du);
735	if (error)
736	goto out_trans_cancel;
737
738	/*
739	* If this is a synchronous mount, make sure that the
740	* create transaction goes to disk before returning to
741	* the user.
742	*/
743	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
744	xfs_trans_set_sync(tp);
745
746	/*
747	* Attach the dquot(s) to the inodes and modify them incore.
748	* These ids of the inode couldn't have changed since the new
749	* inode has been locked ever since it was created.
750	*/
751	xfs_qm_vop_create_dqattach(tp, du.ip, udqp, gdqp, pdqp);
752
753	error = xfs_trans_commit(tp);
754	if (error)
755	goto out_release_inode;
756
757	xfs_qm_dqrele(udqp);
758	xfs_qm_dqrele(gdqp);
759	xfs_qm_dqrele(pdqp);
760
761	*ipp = du.ip;
762	xfs_iunlock(ip: du.ip, XFS_ILOCK_EXCL);
763	xfs_iunlock(ip: dp, XFS_ILOCK_EXCL);
764	xfs_parent_finish(mp, du.ppargs);
765	return 0;
766
767	out_trans_cancel:
768	xfs_trans_cancel(tp);
769	out_release_inode:
770	/*
771	* Wait until after the current transaction is aborted to finish the
772	* setup of the inode and release the inode. This prevents recursive
773	* transactions and deadlocks from xfs_inactive.
774	*/
775	if (du.ip) {
776	xfs_iunlock(ip: du.ip, XFS_ILOCK_EXCL);
777	xfs_finish_inode_setup(ip: du.ip);
778	xfs_irele(ip: du.ip);
779	}
780	out_parent:
781	xfs_parent_finish(mp, du.ppargs);
782	out_release_dquots:
783	xfs_qm_dqrele(udqp);
784	xfs_qm_dqrele(gdqp);
785	xfs_qm_dqrele(pdqp);
786
787	if (unlock_dp_on_error)
788	xfs_iunlock(ip: dp, XFS_ILOCK_EXCL);
789	return error;
790	}
791
792	int
793	xfs_create_tmpfile(
794	const struct xfs_icreate_args *args,
795	struct xfs_inode **ipp)
796	{
797	struct xfs_inode *dp = args->pip;
798	struct xfs_mount *mp = dp->i_mount;
799	struct xfs_inode *ip = NULL;
800	struct xfs_trans *tp = NULL;
801	struct xfs_dquot *udqp;
802	struct xfs_dquot *gdqp;
803	struct xfs_dquot *pdqp;
804	struct xfs_trans_res *tres;
805	xfs_ino_t ino;
806	uint resblks;
807	int error;
808
809	ASSERT(args->flags & XFS_ICREATE_TMPFILE);
810
811	if (xfs_is_shutdown(mp))
812	return -EIO;
813
814	/* Make sure that we have allocated dquot(s) on disk. */
815	error = xfs_icreate_dqalloc(args, udqpp: &udqp, gdqpp: &gdqp, pdqpp: &pdqp);
816	if (error)
817	return error;
818
819	resblks = XFS_IALLOC_SPACE_RES(mp);
820	tres = &M_RES(mp)->tr_create_tmpfile;
821
822	error = xfs_trans_alloc_icreate(mp, resv: tres, udqp, gdqp, pdqp, dblocks: resblks,
823	tpp: &tp);
824	if (error)
825	goto out_release_dquots;
826
827	error = xfs_dialloc(&tp, args, &ino);
828	if (!error)
829	error = xfs_icreate(tp, ino, args, ipp: &ip);
830	if (error)
831	goto out_trans_cancel;
832
833	if (xfs_has_wsync(mp))
834	xfs_trans_set_sync(tp);
835
836	/*
837	* Attach the dquot(s) to the inodes and modify them incore.
838	* These ids of the inode couldn't have changed since the new
839	* inode has been locked ever since it was created.
840	*/
841	xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
842
843	error = xfs_iunlink(tp, ip);
844	if (error)
845	goto out_trans_cancel;
846
847	error = xfs_trans_commit(tp);
848	if (error)
849	goto out_release_inode;
850
851	xfs_qm_dqrele(udqp);
852	xfs_qm_dqrele(gdqp);
853	xfs_qm_dqrele(pdqp);
854
855	*ipp = ip;
856	xfs_iunlock(ip, XFS_ILOCK_EXCL);
857	return 0;
858
859	out_trans_cancel:
860	xfs_trans_cancel(tp);
861	out_release_inode:
862	/*
863	* Wait until after the current transaction is aborted to finish the
864	* setup of the inode and release the inode. This prevents recursive
865	* transactions and deadlocks from xfs_inactive.
866	*/
867	if (ip) {
868	xfs_iunlock(ip, XFS_ILOCK_EXCL);
869	xfs_finish_inode_setup(ip);
870	xfs_irele(ip);
871	}
872	out_release_dquots:
873	xfs_qm_dqrele(udqp);
874	xfs_qm_dqrele(gdqp);
875	xfs_qm_dqrele(pdqp);
876
877	return error;
878	}
879
880	static inline int
881	xfs_projid_differ(
882	struct xfs_inode *tdp,
883	struct xfs_inode *sip)
884	{
885	/*
886	* If we are using project inheritance, we only allow hard link/renames
887	* creation in our tree when the project IDs are the same; else
888	* the tree quota mechanism could be circumvented.
889	*/
890	if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
891	tdp->i_projid != sip->i_projid)) {
892	/*
893	* Project quota setup skips special files which can
894	* leave inodes in a PROJINHERIT directory without a
895	* project ID set. We need to allow links to be made
896	* to these "project-less" inodes because userspace
897	* expects them to succeed after project ID setup,
898	* but everything else should be rejected.
899	*/
900	if (!special_file(VFS_I(sip)->i_mode) ||
901	sip->i_projid != 0) {
902	return -EXDEV;
903	}
904	}
905
906	return 0;
907	}
908
909	int
910	xfs_link(
911	struct xfs_inode *tdp,
912	struct xfs_inode *sip,
913	struct xfs_name *target_name)
914	{
915	struct xfs_dir_update du = {
916	.dp = tdp,
917	.name = target_name,
918	.ip = sip,
919	};
920	struct xfs_mount *mp = tdp->i_mount;
921	struct xfs_trans *tp;
922	int error, nospace_error = 0;
923	int resblks;
924
925	trace_xfs_link(dp: tdp, xfs_link: target_name);
926
927	ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
928
929	if (xfs_is_shutdown(mp))
930	return -EIO;
931	if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
932	return -EIO;
933
934	error = xfs_qm_dqattach(sip);
935	if (error)
936	goto std_return;
937
938	error = xfs_qm_dqattach(tdp);
939	if (error)
940	goto std_return;
941
942	error = xfs_parent_start(mp, &du.ppargs);
943	if (error)
944	goto std_return;
945
946	resblks = xfs_link_space_res(mp, target_name->len);
947	error = xfs_trans_alloc_dir(dp: tdp, resv: &M_RES(mp)->tr_link, ip: sip, dblocks: &resblks,
948	tpp: &tp, nospace_error: &nospace_error);
949	if (error)
950	goto out_parent;
951
952	/*
953	* We don't allow reservationless or quotaless hardlinking when parent
954	* pointers are enabled because we can't back out if the xattrs must
955	* grow.
956	*/
957	if (du.ppargs && nospace_error) {
958	error = nospace_error;
959	goto error_return;
960	}
961
962	error = xfs_projid_differ(tdp, sip);
963	if (error)
964	goto error_return;
965
966	error = xfs_dir_add_child(tp, resblks, &du);
967	if (error)
968	goto error_return;
969
970	/*
971	* If this is a synchronous mount, make sure that the
972	* link transaction goes to disk before returning to
973	* the user.
974	*/
975	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
976	xfs_trans_set_sync(tp);
977
978	error = xfs_trans_commit(tp);
979	xfs_iunlock(ip: tdp, XFS_ILOCK_EXCL);
980	xfs_iunlock(ip: sip, XFS_ILOCK_EXCL);
981	xfs_parent_finish(mp, du.ppargs);
982	return error;
983
984	error_return:
985	xfs_trans_cancel(tp);
986	xfs_iunlock(ip: tdp, XFS_ILOCK_EXCL);
987	xfs_iunlock(ip: sip, XFS_ILOCK_EXCL);
988	out_parent:
989	xfs_parent_finish(mp, du.ppargs);
990	std_return:
991	if (error == -ENOSPC && nospace_error)
992	error = nospace_error;
993	return error;
994	}
995
996	/* Clear the reflink flag and the cowblocks tag if possible. */
997	static void
998	xfs_itruncate_clear_reflink_flags(
999	struct xfs_inode *ip)
1000	{
1001	struct xfs_ifork *dfork;
1002	struct xfs_ifork *cfork;
1003
1004	if (!xfs_is_reflink_inode(ip))
1005	return;
1006	dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1007	cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
1008	if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1009	ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1010	if (cfork->if_bytes == 0)
1011	xfs_inode_clear_cowblocks_tag(ip);
1012	}
1013
1014	/*
1015	* Free up the underlying blocks past new_size. The new size must be smaller
1016	* than the current size. This routine can be used both for the attribute and
1017	* data fork, and does not modify the inode size, which is left to the caller.
1018	*
1019	* The transaction passed to this routine must have made a permanent log
1020	* reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1021	* given transaction and start new ones, so make sure everything involved in
1022	* the transaction is tidy before calling here. Some transaction will be
1023	* returned to the caller to be committed. The incoming transaction must
1024	* already include the inode, and both inode locks must be held exclusively.
1025	* The inode must also be "held" within the transaction. On return the inode
1026	* will be "held" within the returned transaction. This routine does NOT
1027	* require any disk space to be reserved for it within the transaction.
1028	*
1029	* If we get an error, we must return with the inode locked and linked into the
1030	* current transaction. This keeps things simple for the higher level code,
1031	* because it always knows that the inode is locked and held in the transaction
1032	* that returns to it whether errors occur or not. We don't mark the inode
1033	* dirty on error so that transactions can be easily aborted if possible.
1034	*/
1035	int
1036	xfs_itruncate_extents_flags(
1037	struct xfs_trans **tpp,
1038	struct xfs_inode *ip,
1039	int whichfork,
1040	xfs_fsize_t new_size,
1041	int flags)
1042	{
1043	struct xfs_mount *mp = ip->i_mount;
1044	struct xfs_trans *tp = *tpp;
1045	xfs_fileoff_t first_unmap_block;
1046	int error = 0;
1047
1048	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1049	if (icount_read(inode: VFS_I(ip)))
1050	xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1051	ASSERT(new_size <= XFS_ISIZE(ip));
1052	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1053	ASSERT(ip->i_itemp != NULL);
1054	ASSERT(ip->i_itemp->ili_lock_flags == 0);
1055	ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1056
1057	trace_xfs_itruncate_extents_start(ip, new_size);
1058
1059	flags |= xfs_bmapi_aflag(whichfork);
1060
1061	/*
1062	* Since it is possible for space to become allocated beyond
1063	* the end of the file (in a crash where the space is allocated
1064	* but the inode size is not yet updated), simply remove any
1065	* blocks which show up between the new EOF and the maximum
1066	* possible file size.
1067	*
1068	* We have to free all the blocks to the bmbt maximum offset, even if
1069	* the page cache can't scale that far.
1070	*/
1071	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1072	if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1073	WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1074	return 0;
1075	}
1076
1077	error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1078	XFS_MAX_FILEOFF);
1079	if (error)
1080	goto out;
1081
1082	if (whichfork == XFS_DATA_FORK) {
1083	/* Remove all pending CoW reservations. */
1084	error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1085	first_unmap_block, XFS_MAX_FILEOFF, true);
1086	if (error)
1087	goto out;
1088
1089	xfs_itruncate_clear_reflink_flags(ip);
1090	}
1091
1092	/*
1093	* Always re-log the inode so that our permanent transaction can keep
1094	* on rolling it forward in the log.
1095	*/
1096	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1097
1098	trace_xfs_itruncate_extents_end(ip, new_size);
1099
1100	out:
1101	*tpp = tp;
1102	return error;
1103	}
1104
1105	/*
1106	* Mark all the buffers attached to this directory stale. In theory we should
1107	* never be freeing a directory with any blocks at all, but this covers the
1108	* case where we've recovered a directory swap with a "temporary" directory
1109	* created by online repair and now need to dump it.
1110	*/
1111	STATIC void
1112	xfs_inactive_dir(
1113	struct xfs_inode *dp)
1114	{
1115	struct xfs_iext_cursor icur;
1116	struct xfs_bmbt_irec got;
1117	struct xfs_mount *mp = dp->i_mount;
1118	struct xfs_da_geometry *geo = mp->m_dir_geo;
1119	struct xfs_ifork *ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1120	xfs_fileoff_t off;
1121
1122	/*
1123	* Invalidate each directory block. All directory blocks are of
1124	* fsbcount length and alignment, so we only need to walk those same
1125	* offsets. We hold the only reference to this inode, so we must wait
1126	* for the buffer locks.
1127	*/
1128	for_each_xfs_iext(ifp, &icur, &got) {
1129	for (off = round_up(got.br_startoff, geo->fsbcount);
1130	off < got.br_startoff + got.br_blockcount;
1131	off += geo->fsbcount) {
1132	struct xfs_buf *bp = NULL;
1133	xfs_fsblock_t fsbno;
1134	int error;
1135
1136	fsbno = (off - got.br_startoff) + got.br_startblock;
1137	error = xfs_buf_incore(mp->m_ddev_targp,
1138	XFS_FSB_TO_DADDR(mp, fsbno),
1139	XFS_FSB_TO_BB(mp, geo->fsbcount),
1140	XBF_LIVESCAN, &bp);
1141	if (error)
1142	continue;
1143
1144	xfs_buf_stale(bp);
1145	xfs_buf_relse(bp);
1146	}
1147	}
1148	}
1149
1150	/*
1151	* xfs_inactive_truncate
1152	*
1153	* Called to perform a truncate when an inode becomes unlinked.
1154	*/
1155	STATIC int
1156	xfs_inactive_truncate(
1157	struct xfs_inode *ip)
1158	{
1159	struct xfs_mount *mp = ip->i_mount;
1160	struct xfs_trans *tp;
1161	int error;
1162
1163	error = xfs_trans_alloc(mp, resp: &M_RES(mp)->tr_itruncate, blocks: 0, rtextents: 0, flags: 0, tpp: &tp);
1164	if (error) {
1165	ASSERT(xfs_is_shutdown(mp));
1166	return error;
1167	}
1168	xfs_ilock(ip, XFS_ILOCK_EXCL);
1169	xfs_trans_ijoin(tp, ip, 0);
1170
1171	/*
1172	* Log the inode size first to prevent stale data exposure in the event
1173	* of a system crash before the truncate completes. See the related
1174	* comment in xfs_vn_setattr_size() for details.
1175	*/
1176	ip->i_disk_size = 0;
1177	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1178
1179	error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1180	if (error)
1181	goto error_trans_cancel;
1182
1183	ASSERT(ip->i_df.if_nextents == 0);
1184
1185	error = xfs_trans_commit(tp);
1186	if (error)
1187	goto error_unlock;
1188
1189	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1190	return 0;
1191
1192	error_trans_cancel:
1193	xfs_trans_cancel(tp);
1194	error_unlock:
1195	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1196	return error;
1197	}
1198
1199	/*
1200	* xfs_inactive_ifree()
1201	*
1202	* Perform the inode free when an inode is unlinked.
1203	*/
1204	STATIC int
1205	xfs_inactive_ifree(
1206	struct xfs_inode *ip)
1207	{
1208	struct xfs_mount *mp = ip->i_mount;
1209	struct xfs_trans *tp;
1210	int error;
1211
1212	/*
1213	* We try to use a per-AG reservation for any block needed by the finobt
1214	* tree, but as the finobt feature predates the per-AG reservation
1215	* support a degraded file system might not have enough space for the
1216	* reservation at mount time. In that case try to dip into the reserved
1217	* pool and pray.
1218	*
1219	* Send a warning if the reservation does happen to fail, as the inode
1220	* now remains allocated and sits on the unlinked list until the fs is
1221	* repaired.
1222	*/
1223	if (unlikely(mp->m_finobt_nores)) {
1224	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1225	XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1226	&tp);
1227	} else {
1228	error = xfs_trans_alloc(mp, resp: &M_RES(mp)->tr_ifree, blocks: 0, rtextents: 0, flags: 0, tpp: &tp);
1229	}
1230	if (error) {
1231	if (error == -ENOSPC) {
1232	xfs_warn_ratelimited(mp,
1233	"Failed to remove inode(s) from unlinked list. "
1234	"Please free space, unmount and run xfs_repair.");
1235	} else {
1236	ASSERT(xfs_is_shutdown(mp));
1237	}
1238	return error;
1239	}
1240
1241	/*
1242	* We do not hold the inode locked across the entire rolling transaction
1243	* here. We only need to hold it for the first transaction that
1244	* xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1245	* underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1246	* here breaks the relationship between cluster buffer invalidation and
1247	* stale inode invalidation on cluster buffer item journal commit
1248	* completion, and can result in leaving dirty stale inodes hanging
1249	* around in memory.
1250	*
1251	* We have no need for serialising this inode operation against other
1252	* operations - we freed the inode and hence reallocation is required
1253	* and that will serialise on reallocating the space the deferops need
1254	* to free. Hence we can unlock the inode on the first commit of
1255	* the transaction rather than roll it right through the deferops. This
1256	* avoids relogging the XFS_ISTALE inode.
1257	*
1258	* We check that xfs_ifree() hasn't grown an internal transaction roll
1259	* by asserting that the inode is still locked when it returns.
1260	*/
1261	xfs_ilock(ip, XFS_ILOCK_EXCL);
1262	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1263
1264	error = xfs_ifree(tp, ip);
1265	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1266	if (error) {
1267	/*
1268	* If we fail to free the inode, shut down. The cancel
1269	* might do that, we need to make sure. Otherwise the
1270	* inode might be lost for a long time or forever.
1271	*/
1272	if (!xfs_is_shutdown(mp)) {
1273	xfs_notice(mp, "%s: xfs_ifree returned error %d",
1274	__func__, error);
1275	xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1276	}
1277	xfs_trans_cancel(tp);
1278	return error;
1279	}
1280
1281	/*
1282	* Credit the quota account(s). The inode is gone.
1283	*/
1284	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1285
1286	return xfs_trans_commit(tp);
1287	}
1288
1289	/*
1290	* Returns true if we need to update the on-disk metadata before we can free
1291	* the memory used by this inode. Updates include freeing post-eof
1292	* preallocations; freeing COW staging extents; and marking the inode free in
1293	* the inobt if it is on the unlinked list.
1294	*/
1295	bool
1296	xfs_inode_needs_inactive(
1297	struct xfs_inode *ip)
1298	{
1299	struct xfs_mount *mp = ip->i_mount;
1300	struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1301
1302	/*
1303	* If the inode is already free, then there can be nothing
1304	* to clean up here.
1305	*/
1306	if (VFS_I(ip)->i_mode == 0)
1307	return false;
1308
1309	/*
1310	* If this is a read-only mount, don't do this (would generate I/O)
1311	* unless we're in log recovery and cleaning the iunlinked list.
1312	*/
1313	if (xfs_is_readonly(mp) && !xlog_recovery_needed(log: mp->m_log))
1314	return false;
1315
1316	/* If the log isn't running, push inodes straight to reclaim. */
1317	if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1318	return false;
1319
1320	/* Metadata inodes require explicit resource cleanup. */
1321	if (xfs_is_internal_inode(ip))
1322	return false;
1323
1324	/* Want to clean out the cow blocks if there are any. */
1325	if (cow_ifp && cow_ifp->if_bytes > 0)
1326	return true;
1327
1328	/* Unlinked files must be freed. */
1329	if (VFS_I(ip)->i_nlink == 0)
1330	return true;
1331
1332	/*
1333	* This file isn't being freed, so check if there are post-eof blocks
1334	* to free.
1335	*
1336	* Note: don't bother with iolock here since lockdep complains about
1337	* acquiring it in reclaim context. We have the only reference to the
1338	* inode at this point anyways.
1339	*/
1340	return xfs_can_free_eofblocks(ip);
1341	}
1342
1343	/*
1344	* Save health status somewhere, if we're dumping an inode with uncorrected
1345	* errors and online repair isn't running.
1346	*/
1347	static inline void
1348	xfs_inactive_health(
1349	struct xfs_inode *ip)
1350	{
1351	struct xfs_mount *mp = ip->i_mount;
1352	struct xfs_perag *pag;
1353	unsigned int sick;
1354	unsigned int checked;
1355
1356	xfs_inode_measure_sickness(ip, &sick, &checked);
1357	if (!sick)
1358	return;
1359
1360	trace_xfs_inode_unfixed_corruption(ip, flags: sick);
1361
1362	if (sick & XFS_SICK_INO_FORGET)
1363	return;
1364
1365	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1366	if (!pag) {
1367	/* There had better still be a perag structure! */
1368	ASSERT(0);
1369	return;
1370	}
1371
1372	xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1373	xfs_perag_put(pag);
1374	}
1375
1376	/*
1377	* xfs_inactive
1378	*
1379	* This is called when the vnode reference count for the vnode
1380	* goes to zero. If the file has been unlinked, then it must
1381	* now be truncated. Also, we clear all of the read-ahead state
1382	* kept for the inode here since the file is now closed.
1383	*/
1384	int
1385	xfs_inactive(
1386	xfs_inode_t *ip)
1387	{
1388	struct xfs_mount *mp;
1389	int error = 0;
1390	int truncate = 0;
1391
1392	/*
1393	* If the inode is already free, then there can be nothing
1394	* to clean up here.
1395	*/
1396	if (VFS_I(ip)->i_mode == 0) {
1397	ASSERT(ip->i_df.if_broot_bytes == 0);
1398	goto out;
1399	}
1400
1401	mp = ip->i_mount;
1402	ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1403
1404	xfs_inactive_health(ip);
1405
1406	/*
1407	* If this is a read-only mount, don't do this (would generate I/O)
1408	* unless we're in log recovery and cleaning the iunlinked list.
1409	*/
1410	if (xfs_is_readonly(mp) && !xlog_recovery_needed(log: mp->m_log))
1411	goto out;
1412
1413	/* Metadata inodes require explicit resource cleanup. */
1414	if (xfs_is_internal_inode(ip))
1415	goto out;
1416
1417	/* Try to clean out the cow blocks if there are any. */
1418	if (xfs_inode_has_cow_data(ip)) {
1419	error = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1420	if (error)
1421	goto out;
1422	}
1423
1424	if (VFS_I(ip)->i_nlink != 0) {
1425	/*
1426	* Note: don't bother with iolock here since lockdep complains
1427	* about acquiring it in reclaim context. We have the only
1428	* reference to the inode at this point anyways.
1429	*/
1430	if (xfs_can_free_eofblocks(ip))
1431	error = xfs_free_eofblocks(ip);
1432
1433	goto out;
1434	}
1435
1436	if (S_ISREG(VFS_I(ip)->i_mode) &&
1437	(ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1438	xfs_inode_has_filedata(ip)))
1439	truncate = 1;
1440
1441	if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1442	/*
1443	* If this inode is being inactivated during a quotacheck and
1444	* has not yet been scanned by quotacheck, we /must/ remove
1445	* the dquots from the inode before inactivation changes the
1446	* block and inode counts. Most probably this is a result of
1447	* reloading the incore iunlinked list to purge unrecovered
1448	* unlinked inodes.
1449	*/
1450	xfs_qm_dqdetach(ip);
1451	} else {
1452	error = xfs_qm_dqattach(ip);
1453	if (error)
1454	goto out;
1455	}
1456
1457	if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1458	xfs_inactive_dir(dp: ip);
1459	truncate = 1;
1460	}
1461
1462	if (S_ISLNK(VFS_I(ip)->i_mode))
1463	error = xfs_inactive_symlink(ip);
1464	else if (truncate)
1465	error = xfs_inactive_truncate(ip);
1466	if (error)
1467	goto out;
1468
1469	/*
1470	* If there are attributes associated with the file then blow them away
1471	* now. The code calls a routine that recursively deconstructs the
1472	* attribute fork. If also blows away the in-core attribute fork.
1473	*/
1474	if (xfs_inode_has_attr_fork(ip)) {
1475	error = xfs_attr_inactive(ip);
1476	if (error)
1477	goto out;
1478	}
1479
1480	ASSERT(ip->i_forkoff == 0);
1481
1482	/*
1483	* Free the inode.
1484	*/
1485	error = xfs_inactive_ifree(ip);
1486
1487	out:
1488	/*
1489	* We're done making metadata updates for this inode, so we can release
1490	* the attached dquots.
1491	*/
1492	xfs_qm_dqdetach(ip);
1493	return error;
1494	}
1495
1496	/*
1497	* Find an inode on the unlinked list. This does not take references to the
1498	* inode as we have existence guarantees by holding the AGI buffer lock and that
1499	* only unlinked, referenced inodes can be on the unlinked inode list. If we
1500	* don't find the inode in cache, then let the caller handle the situation.
1501	*/
1502	struct xfs_inode *
1503	xfs_iunlink_lookup(
1504	struct xfs_perag *pag,
1505	xfs_agino_t agino)
1506	{
1507	struct xfs_inode *ip;
1508
1509	rcu_read_lock();
1510	ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1511	if (!ip) {
1512	/* Caller can handle inode not being in memory. */
1513	rcu_read_unlock();
1514	return NULL;
1515	}
1516
1517	/*
1518	* Inode in RCU freeing limbo should not happen. Warn about this and
1519	* let the caller handle the failure.
1520	*/
1521	if (WARN_ON_ONCE(!ip->i_ino)) {
1522	rcu_read_unlock();
1523	return NULL;
1524	}
1525	ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1526	rcu_read_unlock();
1527	return ip;
1528	}
1529
1530	/*
1531	* Load the inode @next_agino into the cache and set its prev_unlinked pointer
1532	* to @prev_agino. Caller must hold the AGI to synchronize with other changes
1533	* to the unlinked list.
1534	*/
1535	int
1536	xfs_iunlink_reload_next(
1537	struct xfs_trans *tp,
1538	struct xfs_buf *agibp,
1539	xfs_agino_t prev_agino,
1540	xfs_agino_t next_agino)
1541	{
1542	struct xfs_perag *pag = agibp->b_pag;
1543	struct xfs_mount *mp = pag_mount(pag);
1544	struct xfs_inode *next_ip = NULL;
1545	int error;
1546
1547	ASSERT(next_agino != NULLAGINO);
1548
1549	#ifdef DEBUG
1550	rcu_read_lock();
1551	next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1552	ASSERT(next_ip == NULL);
1553	rcu_read_unlock();
1554	#endif
1555
1556	xfs_info_ratelimited(mp,
1557	"Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
1558	next_agino, pag_agno(pag));
1559
1560	/*
1561	* Use an untrusted lookup just to be cautious in case the AGI has been
1562	* corrupted and now points at a free inode. That shouldn't happen,
1563	* but we'd rather shut down now since we're already running in a weird
1564	* situation.
1565	*/
1566	error = xfs_iget(mp, tp, ino: xfs_agino_to_ino(pag, next_agino),
1567	XFS_IGET_UNTRUSTED, lock_flags: 0, ipp: &next_ip);
1568	if (error) {
1569	xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1570	return error;
1571	}
1572
1573	/* If this is not an unlinked inode, something is very wrong. */
1574	if (VFS_I(ip: next_ip)->i_nlink != 0) {
1575	xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1576	error = -EFSCORRUPTED;
1577	goto rele;
1578	}
1579
1580	next_ip->i_prev_unlinked = prev_agino;
1581	trace_xfs_iunlink_reload_next(ip: next_ip);
1582	rele:
1583	ASSERT(!(inode_state_read_once(VFS_I(next_ip)) & I_DONTCACHE));
1584	if (xfs_is_quotacheck_running(mp) && next_ip)
1585	xfs_iflags_set(ip: next_ip, XFS_IQUOTAUNCHECKED);
1586	xfs_irele(ip: next_ip);
1587	return error;
1588	}
1589
1590	/*
1591	* Look up the inode number specified and if it is not already marked XFS_ISTALE
1592	* mark it stale. We should only find clean inodes in this lookup that aren't
1593	* already stale.
1594	*/
1595	static void
1596	xfs_ifree_mark_inode_stale(
1597	struct xfs_perag *pag,
1598	struct xfs_inode *free_ip,
1599	xfs_ino_t inum)
1600	{
1601	struct xfs_mount *mp = pag_mount(pag);
1602	struct xfs_inode_log_item *iip;
1603	struct xfs_inode *ip;
1604
1605	retry:
1606	rcu_read_lock();
1607	ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1608
1609	/* Inode not in memory, nothing to do */
1610	if (!ip) {
1611	rcu_read_unlock();
1612	return;
1613	}
1614
1615	/*
1616	* because this is an RCU protected lookup, we could find a recently
1617	* freed or even reallocated inode during the lookup. We need to check
1618	* under the i_flags_lock for a valid inode here. Skip it if it is not
1619	* valid, the wrong inode or stale.
1620	*/
1621	spin_lock(lock: &ip->i_flags_lock);
1622	if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1623	goto out_iflags_unlock;
1624
1625	/*
1626	* Don't try to lock/unlock the current inode, but we _cannot_ skip the
1627	* other inodes that we did not find in the list attached to the buffer
1628	* and are not already marked stale. If we can't lock it, back off and
1629	* retry.
1630	*/
1631	if (ip != free_ip) {
1632	if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1633	spin_unlock(lock: &ip->i_flags_lock);
1634	rcu_read_unlock();
1635	delay(ticks: 1);
1636	goto retry;
1637	}
1638	}
1639	ip->i_flags |= XFS_ISTALE;
1640
1641	/*
1642	* If the inode is flushing, it is already attached to the buffer. All
1643	* we needed to do here is mark the inode stale so buffer IO completion
1644	* will remove it from the AIL.
1645	*/
1646	iip = ip->i_itemp;
1647	if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1648	ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1649	ASSERT(iip->ili_last_fields || xlog_is_shutdown(mp->m_log));
1650	goto out_iunlock;
1651	}
1652
1653	/*
1654	* Inodes not attached to the buffer can be released immediately.
1655	* Everything else has to go through xfs_iflush_abort() on journal
1656	* commit as the flock synchronises removal of the inode from the
1657	* cluster buffer against inode reclaim.
1658	*/
1659	if (!iip || list_empty(head: &iip->ili_item.li_bio_list))
1660	goto out_iunlock;
1661
1662	__xfs_iflags_set(ip, XFS_IFLUSHING);
1663	spin_unlock(lock: &ip->i_flags_lock);
1664	rcu_read_unlock();
1665
1666	/* we have a dirty inode in memory that has not yet been flushed. */
1667	spin_lock(lock: &iip->ili_lock);
1668	iip->ili_last_fields = iip->ili_fields;
1669	iip->ili_fields = 0;
1670	spin_unlock(lock: &iip->ili_lock);
1671	ASSERT(iip->ili_last_fields);
1672
1673	if (ip != free_ip)
1674	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1675	return;
1676
1677	out_iunlock:
1678	if (ip != free_ip)
1679	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1680	out_iflags_unlock:
1681	spin_unlock(lock: &ip->i_flags_lock);
1682	rcu_read_unlock();
1683	}
1684
1685	/*
1686	* A big issue when freeing the inode cluster is that we _cannot_ skip any
1687	* inodes that are in memory - they all must be marked stale and attached to
1688	* the cluster buffer.
1689	*/
1690	static int
1691	xfs_ifree_cluster(
1692	struct xfs_trans *tp,
1693	struct xfs_perag *pag,
1694	struct xfs_inode *free_ip,
1695	struct xfs_icluster *xic)
1696	{
1697	struct xfs_mount *mp = free_ip->i_mount;
1698	struct xfs_ino_geometry *igeo = M_IGEO(mp);
1699	struct xfs_buf *bp;
1700	xfs_daddr_t blkno;
1701	xfs_ino_t inum = xic->first_ino;
1702	int nbufs;
1703	int i, j;
1704	int ioffset;
1705	int error;
1706
1707	nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1708
1709	for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1710	/*
1711	* The allocation bitmap tells us which inodes of the chunk were
1712	* physically allocated. Skip the cluster if an inode falls into
1713	* a sparse region.
1714	*/
1715	ioffset = inum - xic->first_ino;
1716	if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1717	ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1718	continue;
1719	}
1720
1721	blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1722	XFS_INO_TO_AGBNO(mp, inum));
1723
1724	/*
1725	* We obtain and lock the backing buffer first in the process
1726	* here to ensure dirty inodes attached to the buffer remain in
1727	* the flushing state while we mark them stale.
1728	*
1729	* If we scan the in-memory inodes first, then buffer IO can
1730	* complete before we get a lock on it, and hence we may fail
1731	* to mark all the active inodes on the buffer stale.
1732	*/
1733	error = xfs_trans_get_buf(tp, target: mp->m_ddev_targp, blkno,
1734	numblks: mp->m_bsize * igeo->blocks_per_cluster, flags: 0, bpp: &bp);
1735	if (error)
1736	return error;
1737
1738	/*
1739	* This buffer may not have been correctly initialised as we
1740	* didn't read it from disk. That's not important because we are
1741	* only using to mark the buffer as stale in the log, and to
1742	* attach stale cached inodes on it.
1743	*
1744	* For the inode that triggered the cluster freeing, this
1745	* attachment may occur in xfs_inode_item_precommit() after we
1746	* have marked this buffer stale. If this buffer was not in
1747	* memory before xfs_ifree_cluster() started, it will not be
1748	* marked XBF_DONE and this will cause problems later in
1749	* xfs_inode_item_precommit() when we trip over a (stale, !done)
1750	* buffer to attached to the transaction.
1751	*
1752	* Hence we have to mark the buffer as XFS_DONE here. This is
1753	* safe because we are also marking the buffer as XBF_STALE and
1754	* XFS_BLI_STALE. That means it will never be dispatched for
1755	* IO and it won't be unlocked until the cluster freeing has
1756	* been committed to the journal and the buffer unpinned. If it
1757	* is written, we want to know about it, and we want it to
1758	* fail. We can acheive this by adding a write verifier to the
1759	* buffer.
1760	*/
1761	bp->b_flags |= XBF_DONE;
1762	bp->b_ops = &xfs_inode_buf_ops;
1763
1764	/*
1765	* Now we need to set all the cached clean inodes as XFS_ISTALE,
1766	* too. This requires lookups, and will skip inodes that we've
1767	* already marked XFS_ISTALE.
1768	*/
1769	for (i = 0; i < igeo->inodes_per_cluster; i++)
1770	xfs_ifree_mark_inode_stale(pag, free_ip, inum: inum + i);
1771
1772	xfs_trans_stale_inode_buf(tp, bp);
1773	xfs_trans_binval(tp, bp);
1774	}
1775	return 0;
1776	}
1777
1778	/*
1779	* This is called to return an inode to the inode free list. The inode should
1780	* already be truncated to 0 length and have no pages associated with it. This
1781	* routine also assumes that the inode is already a part of the transaction.
1782	*
1783	* The on-disk copy of the inode will have been added to the list of unlinked
1784	* inodes in the AGI. We need to remove the inode from that list atomically with
1785	* respect to freeing it here.
1786	*/
1787	int
1788	xfs_ifree(
1789	struct xfs_trans *tp,
1790	struct xfs_inode *ip)
1791	{
1792	struct xfs_mount *mp = ip->i_mount;
1793	struct xfs_perag *pag;
1794	struct xfs_icluster xic = { 0 };
1795	struct xfs_inode_log_item *iip = ip->i_itemp;
1796	int error;
1797
1798	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1799	ASSERT(VFS_I(ip)->i_nlink == 0);
1800	ASSERT(ip->i_df.if_nextents == 0);
1801	ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1802	ASSERT(ip->i_nblocks == 0);
1803
1804	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1805
1806	error = xfs_inode_uninit(tp, pag, ip, &xic);
1807	if (error)
1808	goto out;
1809
1810	if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1811	xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1812
1813	/* Don't attempt to replay owner changes for a deleted inode */
1814	spin_lock(lock: &iip->ili_lock);
1815	iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1816	spin_unlock(lock: &iip->ili_lock);
1817
1818	if (xic.deleted)
1819	error = xfs_ifree_cluster(tp, pag, free_ip: ip, xic: &xic);
1820	out:
1821	xfs_perag_put(pag);
1822	return error;
1823	}
1824
1825	/*
1826	* This is called to unpin an inode. The caller must have the inode locked
1827	* in at least shared mode so that the buffer cannot be subsequently pinned
1828	* once someone is waiting for it to be unpinned.
1829	*/
1830	static void
1831	xfs_iunpin(
1832	struct xfs_inode *ip)
1833	{
1834	struct xfs_inode_log_item *iip = ip->i_itemp;
1835	xfs_csn_t seq = 0;
1836
1837	trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1838	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1839
1840	spin_lock(lock: &iip->ili_lock);
1841	seq = iip->ili_commit_seq;
1842	spin_unlock(lock: &iip->ili_lock);
1843	if (!seq)
1844	return;
1845
1846	/* Give the log a push to start the unpinning I/O */
1847	xfs_log_force_seq(ip->i_mount, seq, 0, NULL);
1848
1849	}
1850
1851	static void
1852	__xfs_iunpin_wait(
1853	struct xfs_inode *ip)
1854	{
1855	wait_queue_head_t *wq = bit_waitqueue(word: &ip->i_flags, __XFS_IPINNED_BIT);
1856	DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1857
1858	xfs_iunpin(ip);
1859
1860	do {
1861	prepare_to_wait(wq_head: wq, wq_entry: &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1862	if (xfs_ipincount(ip))
1863	io_schedule();
1864	} while (xfs_ipincount(ip));
1865	finish_wait(wq_head: wq, wq_entry: &wait.wq_entry);
1866	}
1867
1868	void
1869	xfs_iunpin_wait(
1870	struct xfs_inode *ip)
1871	{
1872	if (xfs_ipincount(ip))
1873	__xfs_iunpin_wait(ip);
1874	}
1875
1876	/*
1877	* Removing an inode from the namespace involves removing the directory entry
1878	* and dropping the link count on the inode. Removing the directory entry can
1879	* result in locking an AGF (directory blocks were freed) and removing a link
1880	* count can result in placing the inode on an unlinked list which results in
1881	* locking an AGI.
1882	*
1883	* The big problem here is that we have an ordering constraint on AGF and AGI
1884	* locking - inode allocation locks the AGI, then can allocate a new extent for
1885	* new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1886	* removes the inode from the unlinked list, requiring that we lock the AGI
1887	* first, and then freeing the inode can result in an inode chunk being freed
1888	* and hence freeing disk space requiring that we lock an AGF.
1889	*
1890	* Hence the ordering that is imposed by other parts of the code is AGI before
1891	* AGF. This means we cannot remove the directory entry before we drop the inode
1892	* reference count and put it on the unlinked list as this results in a lock
1893	* order of AGF then AGI, and this can deadlock against inode allocation and
1894	* freeing. Therefore we must drop the link counts before we remove the
1895	* directory entry.
1896	*
1897	* This is still safe from a transactional point of view - it is not until we
1898	* get to xfs_defer_finish() that we have the possibility of multiple
1899	* transactions in this operation. Hence as long as we remove the directory
1900	* entry and drop the link count in the first transaction of the remove
1901	* operation, there are no transactional constraints on the ordering here.
1902	*/
1903	int
1904	xfs_remove(
1905	struct xfs_inode *dp,
1906	struct xfs_name *name,
1907	struct xfs_inode *ip)
1908	{
1909	struct xfs_dir_update du = {
1910	.dp = dp,
1911	.name = name,
1912	.ip = ip,
1913	};
1914	struct xfs_mount *mp = dp->i_mount;
1915	struct xfs_trans *tp = NULL;
1916	int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1917	int dontcare;
1918	int error = 0;
1919	uint resblks;
1920
1921	trace_xfs_remove(dp, xfs_remove: name);
1922
1923	if (xfs_is_shutdown(mp))
1924	return -EIO;
1925	if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1926	return -EIO;
1927
1928	error = xfs_qm_dqattach(dp);
1929	if (error)
1930	goto std_return;
1931
1932	error = xfs_qm_dqattach(ip);
1933	if (error)
1934	goto std_return;
1935
1936	error = xfs_parent_start(mp, &du.ppargs);
1937	if (error)
1938	goto std_return;
1939
1940	/*
1941	* We try to get the real space reservation first, allowing for
1942	* directory btree deletion(s) implying possible bmap insert(s). If we
1943	* can't get the space reservation then we use 0 instead, and avoid the
1944	* bmap btree insert(s) in the directory code by, if the bmap insert
1945	* tries to happen, instead trimming the LAST block from the directory.
1946	*
1947	* Ignore EDQUOT and ENOSPC being returned via nospace_error because
1948	* the directory code can handle a reservationless update and we don't
1949	* want to prevent a user from trying to free space by deleting things.
1950	*/
1951	resblks = xfs_remove_space_res(mp, name->len);
1952	error = xfs_trans_alloc_dir(dp, resv: &M_RES(mp)->tr_remove, ip, dblocks: &resblks,
1953	tpp: &tp, nospace_error: &dontcare);
1954	if (error) {
1955	ASSERT(error != -ENOSPC);
1956	goto out_parent;
1957	}
1958
1959	error = xfs_dir_remove_child(tp, resblks, &du);
1960	if (error)
1961	goto out_trans_cancel;
1962
1963	/*
1964	* If this is a synchronous mount, make sure that the
1965	* remove transaction goes to disk before returning to
1966	* the user.
1967	*/
1968	if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1969	xfs_trans_set_sync(tp);
1970
1971	error = xfs_trans_commit(tp);
1972	if (error)
1973	goto out_unlock;
1974
1975	if (is_dir && xfs_inode_is_filestream(ip))
1976	xfs_filestream_deassociate(ip);
1977
1978	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1979	xfs_iunlock(ip: dp, XFS_ILOCK_EXCL);
1980	xfs_parent_finish(mp, du.ppargs);
1981	return 0;
1982
1983	out_trans_cancel:
1984	xfs_trans_cancel(tp);
1985	out_unlock:
1986	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1987	xfs_iunlock(ip: dp, XFS_ILOCK_EXCL);
1988	out_parent:
1989	xfs_parent_finish(mp, du.ppargs);
1990	std_return:
1991	return error;
1992	}
1993
1994	static inline void
1995	xfs_iunlock_rename(
1996	struct xfs_inode **i_tab,
1997	int num_inodes)
1998	{
1999	int i;
2000
2001	for (i = num_inodes - 1; i >= 0; i--) {
2002	/* Skip duplicate inodes if src and target dps are the same */
2003	if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
2004	continue;
2005	xfs_iunlock(ip: i_tab[i], XFS_ILOCK_EXCL);
2006	}
2007	}
2008
2009	/*
2010	* Enter all inodes for a rename transaction into a sorted array.
2011	*/
2012	#define __XFS_SORT_INODES 5
2013	STATIC void
2014	xfs_sort_for_rename(
2015	struct xfs_inode *dp1, /* in: old (source) directory inode */
2016	struct xfs_inode *dp2, /* in: new (target) directory inode */
2017	struct xfs_inode *ip1, /* in: inode of old entry */
2018	struct xfs_inode *ip2, /* in: inode of new entry */
2019	struct xfs_inode *wip, /* in: whiteout inode */
2020	struct xfs_inode **i_tab,/* out: sorted array of inodes */
2021	int *num_inodes) /* in/out: inodes in array */
2022	{
2023	int i;
2024
2025	ASSERT(*num_inodes == __XFS_SORT_INODES);
2026	memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2027
2028	/*
2029	* i_tab contains a list of pointers to inodes. We initialize
2030	* the table here & we'll sort it. We will then use it to
2031	* order the acquisition of the inode locks.
2032	*
2033	* Note that the table may contain duplicates. e.g., dp1 == dp2.
2034	*/
2035	i = 0;
2036	i_tab[i++] = dp1;
2037	i_tab[i++] = dp2;
2038	i_tab[i++] = ip1;
2039	if (ip2)
2040	i_tab[i++] = ip2;
2041	if (wip)
2042	i_tab[i++] = wip;
2043	*num_inodes = i;
2044
2045	xfs_sort_inodes(i_tab, num_inodes: *num_inodes);
2046	}
2047
2048	void
2049	xfs_sort_inodes(
2050	struct xfs_inode **i_tab,
2051	unsigned int num_inodes)
2052	{
2053	int i, j;
2054
2055	ASSERT(num_inodes <= __XFS_SORT_INODES);
2056
2057	/*
2058	* Sort the elements via bubble sort. (Remember, there are at
2059	* most 5 elements to sort, so this is adequate.)
2060	*/
2061	for (i = 0; i < num_inodes; i++) {
2062	for (j = 1; j < num_inodes; j++) {
2063	if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2064	swap(i_tab[j], i_tab[j - 1]);
2065	}
2066	}
2067	}
2068
2069	/*
2070	* xfs_rename_alloc_whiteout()
2071	*
2072	* Return a referenced, unlinked, unlocked inode that can be used as a
2073	* whiteout in a rename transaction. We use a tmpfile inode here so that if we
2074	* crash between allocating the inode and linking it into the rename transaction
2075	* recovery will free the inode and we won't leak it.
2076	*/
2077	static int
2078	xfs_rename_alloc_whiteout(
2079	struct mnt_idmap *idmap,
2080	struct xfs_name *src_name,
2081	struct xfs_inode *dp,
2082	struct xfs_inode **wip)
2083	{
2084	struct xfs_icreate_args args = {
2085	.idmap = idmap,
2086	.pip = dp,
2087	.mode = S_IFCHR | WHITEOUT_MODE,
2088	.flags = XFS_ICREATE_TMPFILE,
2089	};
2090	struct xfs_inode *tmpfile;
2091	struct qstr name;
2092	int error;
2093
2094	error = xfs_create_tmpfile(iargs: &args, ipp: &tmpfile);
2095	if (error)
2096	return error;
2097
2098	name.name = src_name->name;
2099	name.len = src_name->len;
2100	error = xfs_inode_init_security(inode: VFS_I(ip: tmpfile), dir: VFS_I(ip: dp), qstr: &name);
2101	if (error) {
2102	xfs_finish_inode_setup(ip: tmpfile);
2103	xfs_irele(ip: tmpfile);
2104	return error;
2105	}
2106
2107	/*
2108	* Prepare the tmpfile inode as if it were created through the VFS.
2109	* Complete the inode setup and flag it as linkable. nlink is already
2110	* zero, so we can skip the drop_nlink.
2111	*/
2112	xfs_setup_iops(ip: tmpfile);
2113	xfs_finish_inode_setup(ip: tmpfile);
2114	inode_state_set_raw(inode: VFS_I(ip: tmpfile), flags: I_LINKABLE);
2115
2116	*wip = tmpfile;
2117	return 0;
2118	}
2119
2120	/*
2121	* xfs_rename
2122	*/
2123	int
2124	xfs_rename(
2125	struct mnt_idmap *idmap,
2126	struct xfs_inode *src_dp,
2127	struct xfs_name *src_name,
2128	struct xfs_inode *src_ip,
2129	struct xfs_inode *target_dp,
2130	struct xfs_name *target_name,
2131	struct xfs_inode *target_ip,
2132	unsigned int flags)
2133	{
2134	struct xfs_dir_update du_src = {
2135	.dp = src_dp,
2136	.name = src_name,
2137	.ip = src_ip,
2138	};
2139	struct xfs_dir_update du_tgt = {
2140	.dp = target_dp,
2141	.name = target_name,
2142	.ip = target_ip,
2143	};
2144	struct xfs_dir_update du_wip = { };
2145	struct xfs_mount *mp = src_dp->i_mount;
2146	struct xfs_trans *tp;
2147	struct xfs_inode *inodes[__XFS_SORT_INODES];
2148	int i;
2149	int num_inodes = __XFS_SORT_INODES;
2150	bool new_parent = (src_dp != target_dp);
2151	bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2152	int spaceres;
2153	bool retried = false;
2154	int error, nospace_error = 0;
2155
2156	trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2157
2158	if ((flags & RENAME_EXCHANGE) && !target_ip)
2159	return -EINVAL;
2160
2161	/*
2162	* If we are doing a whiteout operation, allocate the whiteout inode
2163	* we will be placing at the target and ensure the type is set
2164	* appropriately.
2165	*/
2166	if (flags & RENAME_WHITEOUT) {
2167	error = xfs_rename_alloc_whiteout(idmap, src_name, dp: target_dp,
2168	wip: &du_wip.ip);
2169	if (error)
2170	return error;
2171
2172	/* setup target dirent info as whiteout */
2173	src_name->type = XFS_DIR3_FT_CHRDEV;
2174	}
2175
2176	xfs_sort_for_rename(dp1: src_dp, dp2: target_dp, ip1: src_ip, ip2: target_ip, wip: du_wip.ip,
2177	i_tab: inodes, num_inodes: &num_inodes);
2178
2179	error = xfs_parent_start(mp, &du_src.ppargs);
2180	if (error)
2181	goto out_release_wip;
2182
2183	if (du_wip.ip) {
2184	error = xfs_parent_start(mp, &du_wip.ppargs);
2185	if (error)
2186	goto out_src_ppargs;
2187	}
2188
2189	if (target_ip) {
2190	error = xfs_parent_start(mp, &du_tgt.ppargs);
2191	if (error)
2192	goto out_wip_ppargs;
2193	}
2194
2195	retry:
2196	nospace_error = 0;
2197	spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2198	target_name->len, du_wip.ip != NULL);
2199	error = xfs_trans_alloc(mp, resp: &M_RES(mp)->tr_rename, blocks: spaceres, rtextents: 0, flags: 0, tpp: &tp);
2200	if (error == -ENOSPC) {
2201	nospace_error = error;
2202	spaceres = 0;
2203	error = xfs_trans_alloc(mp, resp: &M_RES(mp)->tr_rename, blocks: 0, rtextents: 0, flags: 0,
2204	tpp: &tp);
2205	}
2206	if (error)
2207	goto out_tgt_ppargs;
2208
2209	/*
2210	* We don't allow reservationless renaming when parent pointers are
2211	* enabled because we can't back out if the xattrs must grow.
2212	*/
2213	if (du_src.ppargs && nospace_error) {
2214	error = nospace_error;
2215	xfs_trans_cancel(tp);
2216	goto out_tgt_ppargs;
2217	}
2218
2219	/*
2220	* Attach the dquots to the inodes
2221	*/
2222	error = xfs_qm_vop_rename_dqattach(inodes);
2223	if (error) {
2224	xfs_trans_cancel(tp);
2225	goto out_tgt_ppargs;
2226	}
2227
2228	/*
2229	* Lock all the participating inodes. Depending upon whether
2230	* the target_name exists in the target directory, and
2231	* whether the target directory is the same as the source
2232	* directory, we can lock from 2 to 5 inodes.
2233	*/
2234	xfs_lock_inodes(ips: inodes, inodes: num_inodes, XFS_ILOCK_EXCL);
2235
2236	/*
2237	* Join all the inodes to the transaction.
2238	*/
2239	xfs_trans_ijoin(tp, src_dp, 0);
2240	if (new_parent)
2241	xfs_trans_ijoin(tp, target_dp, 0);
2242	xfs_trans_ijoin(tp, src_ip, 0);
2243	if (target_ip)
2244	xfs_trans_ijoin(tp, target_ip, 0);
2245	if (du_wip.ip)
2246	xfs_trans_ijoin(tp, du_wip.ip, 0);
2247
2248	error = xfs_projid_differ(tdp: target_dp, sip: src_ip);
2249	if (error)
2250	goto out_trans_cancel;
2251
2252	/* RENAME_EXCHANGE is unique from here on. */
2253	if (flags & RENAME_EXCHANGE) {
2254	error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2255	spaceres);
2256	if (error)
2257	goto out_trans_cancel;
2258	goto out_commit;
2259	}
2260
2261	/*
2262	* Try to reserve quota to handle an expansion of the target directory.
2263	* We'll allow the rename to continue in reservationless mode if we hit
2264	* a space usage constraint. If we trigger reservationless mode, save
2265	* the errno if there isn't any free space in the target directory.
2266	*/
2267	if (spaceres != 0) {
2268	error = xfs_trans_reserve_quota_nblks(tp, ip: target_dp, dblocks: spaceres,
2269	rblocks: 0, force: false);
2270	if (error == -EDQUOT || error == -ENOSPC) {
2271	if (!retried) {
2272	xfs_trans_cancel(tp);
2273	xfs_iunlock_rename(i_tab: inodes, num_inodes);
2274	xfs_blockgc_free_quota(ip: target_dp, iwalk_flags: 0);
2275	retried = true;
2276	goto retry;
2277	}
2278
2279	nospace_error = error;
2280	spaceres = 0;
2281	error = 0;
2282	}
2283	if (error)
2284	goto out_trans_cancel;
2285	}
2286
2287	/*
2288	* We don't allow quotaless renaming when parent pointers are enabled
2289	* because we can't back out if the xattrs must grow.
2290	*/
2291	if (du_src.ppargs && nospace_error) {
2292	error = nospace_error;
2293	goto out_trans_cancel;
2294	}
2295
2296	/*
2297	* Lock the AGI buffers we need to handle bumping the nlink of the
2298	* whiteout inode off the unlinked list and to handle dropping the
2299	* nlink of the target inode. Per locking order rules, do this in
2300	* increasing AG order and before directory block allocation tries to
2301	* grab AGFs because we grab AGIs before AGFs.
2302	*
2303	* The (vfs) caller must ensure that if src is a directory then
2304	* target_ip is either null or an empty directory.
2305	*/
2306	for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2307	if (inodes[i] == du_wip.ip ||
2308	(inodes[i] == target_ip &&
2309	(VFS_I(ip: target_ip)->i_nlink == 1 || src_is_directory))) {
2310	struct xfs_perag *pag;
2311	struct xfs_buf *bp;
2312
2313	pag = xfs_perag_get(mp,
2314	XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2315	error = xfs_read_agi(pag, tp, 0, &bp);
2316	xfs_perag_put(pag);
2317	if (error)
2318	goto out_trans_cancel;
2319	}
2320	}
2321
2322	error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2323	&du_wip);
2324	if (error)
2325	goto out_trans_cancel;
2326
2327	if (du_wip.ip) {
2328	/*
2329	* Now we have a real link, clear the "I'm a tmpfile" state
2330	* flag from the inode so it doesn't accidentally get misused in
2331	* future.
2332	*/
2333	inode_state_clear_raw(inode: VFS_I(ip: du_wip.ip), flags: I_LINKABLE);
2334	}
2335
2336	out_commit:
2337	/*
2338	* If this is a synchronous mount, make sure that the rename
2339	* transaction goes to disk before returning to the user.
2340	*/
2341	if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2342	xfs_trans_set_sync(tp);
2343
2344	error = xfs_trans_commit(tp);
2345	nospace_error = 0;
2346	goto out_unlock;
2347
2348	out_trans_cancel:
2349	xfs_trans_cancel(tp);
2350	out_unlock:
2351	xfs_iunlock_rename(i_tab: inodes, num_inodes);
2352	out_tgt_ppargs:
2353	xfs_parent_finish(mp, du_tgt.ppargs);
2354	out_wip_ppargs:
2355	xfs_parent_finish(mp, du_wip.ppargs);
2356	out_src_ppargs:
2357	xfs_parent_finish(mp, du_src.ppargs);
2358	out_release_wip:
2359	if (du_wip.ip)
2360	xfs_irele(ip: du_wip.ip);
2361	if (error == -ENOSPC && nospace_error)
2362	error = nospace_error;
2363	return error;
2364	}
2365
2366	static int
2367	xfs_iflush(
2368	struct xfs_inode *ip,
2369	struct xfs_buf *bp)
2370	{
2371	struct xfs_inode_log_item *iip = ip->i_itemp;
2372	struct xfs_dinode *dip;
2373	struct xfs_mount *mp = ip->i_mount;
2374	int error;
2375
2376	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2377	ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2378	ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2379	ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2380	ASSERT(iip->ili_item.li_buf == bp);
2381
2382	dip = xfs_buf_offset(bp, offset: ip->i_imap.im_boffset);
2383
2384	/*
2385	* We don't flush the inode if any of the following checks fail, but we
2386	* do still update the log item and attach to the backing buffer as if
2387	* the flush happened. This is a formality to facilitate predictable
2388	* error handling as the caller will shutdown and fail the buffer.
2389	*/
2390	error = -EFSCORRUPTED;
2391	if (dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC) ||
2392	XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_1)) {
2393	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2394	"%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2395	__func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2396	goto flush_out;
2397	}
2398	if (ip->i_df.if_format == XFS_DINODE_FMT_META_BTREE) {
2399	if (!S_ISREG(VFS_I(ip)->i_mode) ||
2400	!(ip->i_diflags2 & XFS_DIFLAG2_METADATA)) {
2401	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2402	"%s: Bad %s meta btree inode %Lu, ptr "PTR_FMT,
2403	__func__, xfs_metafile_type_str(ip->i_metatype),
2404	ip->i_ino, ip);
2405	goto flush_out;
2406	}
2407	} else if (S_ISREG(VFS_I(ip)->i_mode)) {
2408	if ((ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2409	ip->i_df.if_format != XFS_DINODE_FMT_BTREE) ||
2410	XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_3)) {
2411	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2412	"%s: Bad regular inode %llu, ptr "PTR_FMT,
2413	__func__, ip->i_ino, ip);
2414	goto flush_out;
2415	}
2416	} else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2417	if ((ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2418	ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2419	ip->i_df.if_format != XFS_DINODE_FMT_LOCAL) ||
2420	XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_4)) {
2421	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2422	"%s: Bad directory inode %llu, ptr "PTR_FMT,
2423	__func__, ip->i_ino, ip);
2424	goto flush_out;
2425	}
2426	}
2427	if (ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2428	ip->i_nblocks || XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_5)) {
2429	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2430	"%s: detected corrupt incore inode %llu, "
2431	"total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2432	__func__, ip->i_ino,
2433	ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2434	ip->i_nblocks, ip);
2435	goto flush_out;
2436	}
2437	if (ip->i_forkoff > mp->m_sb.sb_inodesize ||
2438	XFS_TEST_ERROR(mp, XFS_ERRTAG_IFLUSH_6)) {
2439	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2440	"%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2441	__func__, ip->i_ino, ip->i_forkoff, ip);
2442	goto flush_out;
2443	}
2444
2445	if (xfs_inode_has_attr_fork(ip) &&
2446	ip->i_af.if_format == XFS_DINODE_FMT_META_BTREE) {
2447	xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2448	"%s: meta btree in inode %Lu attr fork, ptr "PTR_FMT,
2449	__func__, ip->i_ino, ip);
2450	goto flush_out;
2451	}
2452
2453	/*
2454	* Inode item log recovery for v2 inodes are dependent on the flushiter
2455	* count for correct sequencing. We bump the flush iteration count so
2456	* we can detect flushes which postdate a log record during recovery.
2457	* This is redundant as we now log every change and hence this can't
2458	* happen but we need to still do it to ensure backwards compatibility
2459	* with old kernels that predate logging all inode changes.
2460	*/
2461	if (!xfs_has_v3inodes(mp))
2462	ip->i_flushiter++;
2463
2464	/*
2465	* If there are inline format data / attr forks attached to this inode,
2466	* make sure they are not corrupt.
2467	*/
2468	if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2469	xfs_ifork_verify_local_data(ip))
2470	goto flush_out;
2471	if (xfs_inode_has_attr_fork(ip) &&
2472	ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2473	xfs_ifork_verify_local_attr(ip))
2474	goto flush_out;
2475
2476	/*
2477	* Copy the dirty parts of the inode into the on-disk inode. We always
2478	* copy out the core of the inode, because if the inode is dirty at all
2479	* the core must be.
2480	*/
2481	xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2482
2483	/* Wrap, we never let the log put out DI_MAX_FLUSH */
2484	if (!xfs_has_v3inodes(mp)) {
2485	if (ip->i_flushiter == DI_MAX_FLUSH)
2486	ip->i_flushiter = 0;
2487	}
2488
2489	xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2490	if (xfs_inode_has_attr_fork(ip))
2491	xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2492
2493	/*
2494	* We've recorded everything logged in the inode, so we'd like to clear
2495	* the ili_fields bits so we don't log and flush things unnecessarily.
2496	* However, we can't stop logging all this information until the data
2497	* we've copied into the disk buffer is written to disk. If we did we
2498	* might overwrite the copy of the inode in the log with all the data
2499	* after re-logging only part of it, and in the face of a crash we
2500	* wouldn't have all the data we need to recover.
2501	*
2502	* What we do is move the bits to the ili_last_fields field. When
2503	* logging the inode, these bits are moved back to the ili_fields field.
2504	* In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2505	* we know that the information those bits represent is permanently on
2506	* disk. As long as the flush completes before the inode is logged
2507	* again, then both ili_fields and ili_last_fields will be cleared.
2508	*/
2509	error = 0;
2510	flush_out:
2511	spin_lock(lock: &iip->ili_lock);
2512	iip->ili_last_fields = iip->ili_fields;
2513	iip->ili_fields = 0;
2514	set_bit(XFS_LI_FLUSHING, addr: &iip->ili_item.li_flags);
2515	spin_unlock(lock: &iip->ili_lock);
2516
2517	/*
2518	* Store the current LSN of the inode so that we can tell whether the
2519	* item has moved in the AIL from xfs_buf_inode_iodone().
2520	*/
2521	xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2522	&iip->ili_item.li_lsn);
2523
2524	/* generate the checksum. */
2525	xfs_dinode_calc_crc(mp, dip);
2526	if (error)
2527	xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2528	return error;
2529	}
2530
2531	/*
2532	* Non-blocking flush of dirty inode metadata into the backing buffer.
2533	*
2534	* The caller must have a reference to the inode and hold the cluster buffer
2535	* locked. The function will walk across all the inodes on the cluster buffer it
2536	* can find and lock without blocking, and flush them to the cluster buffer.
2537	*
2538	* On successful flushing of at least one inode, the caller must write out the
2539	* buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2540	* the caller needs to release the buffer. On failure, the filesystem will be
2541	* shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2542	* will be returned.
2543	*/
2544	int
2545	xfs_iflush_cluster(
2546	struct xfs_buf *bp)
2547	{
2548	struct xfs_mount *mp = bp->b_mount;
2549	struct xfs_log_item *lip, *n;
2550	struct xfs_inode *ip;
2551	struct xfs_inode_log_item *iip;
2552	int clcount = 0;
2553	int error = 0;
2554
2555	/*
2556	* We must use the safe variant here as on shutdown xfs_iflush_abort()
2557	* will remove itself from the list.
2558	*/
2559	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2560	iip = (struct xfs_inode_log_item *)lip;
2561	ip = iip->ili_inode;
2562
2563	/*
2564	* Quick and dirty check to avoid locks if possible.
2565	*/
2566	if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2567	continue;
2568	if (xfs_ipincount(ip))
2569	continue;
2570
2571	/*
2572	* The inode is still attached to the buffer, which means it is
2573	* dirty but reclaim might try to grab it. Check carefully for
2574	* that, and grab the ilock while still holding the i_flags_lock
2575	* to guarantee reclaim will not be able to reclaim this inode
2576	* once we drop the i_flags_lock.
2577	*/
2578	spin_lock(lock: &ip->i_flags_lock);
2579	ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2580	if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2581	spin_unlock(lock: &ip->i_flags_lock);
2582	continue;
2583	}
2584
2585	/*
2586	* ILOCK will pin the inode against reclaim and prevent
2587	* concurrent transactions modifying the inode while we are
2588	* flushing the inode. If we get the lock, set the flushing
2589	* state before we drop the i_flags_lock.
2590	*/
2591	if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2592	spin_unlock(lock: &ip->i_flags_lock);
2593	continue;
2594	}
2595	__xfs_iflags_set(ip, XFS_IFLUSHING);
2596	spin_unlock(lock: &ip->i_flags_lock);
2597
2598	/*
2599	* Abort flushing this inode if we are shut down because the
2600	* inode may not currently be in the AIL. This can occur when
2601	* log I/O failure unpins the inode without inserting into the
2602	* AIL, leaving a dirty/unpinned inode attached to the buffer
2603	* that otherwise looks like it should be flushed.
2604	*/
2605	if (xlog_is_shutdown(log: mp->m_log)) {
2606	xfs_iunpin_wait(ip);
2607	xfs_iflush_abort(ip);
2608	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2609	error = -EIO;
2610	continue;
2611	}
2612
2613	/* don't block waiting on a log force to unpin dirty inodes */
2614	if (xfs_ipincount(ip)) {
2615	xfs_iflags_clear(ip, XFS_IFLUSHING);
2616	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2617	continue;
2618	}
2619
2620	if (!xfs_inode_clean(ip))
2621	error = xfs_iflush(ip, bp);
2622	else
2623	xfs_iflags_clear(ip, XFS_IFLUSHING);
2624	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2625	if (error)
2626	break;
2627	clcount++;
2628	}
2629
2630	if (error) {
2631	/*
2632	* Shutdown first so we kill the log before we release this
2633	* buffer. If it is an INODE_ALLOC buffer and pins the tail
2634	* of the log, failing it before the _log_ is shut down can
2635	* result in the log tail being moved forward in the journal
2636	* on disk because log writes can still be taking place. Hence
2637	* unpinning the tail will allow the ICREATE intent to be
2638	* removed from the log an recovery will fail with uninitialised
2639	* inode cluster buffers.
2640	*/
2641	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2642	bp->b_flags |= XBF_ASYNC;
2643	xfs_buf_ioend_fail(bp);
2644	return error;
2645	}
2646
2647	if (!clcount)
2648	return -EAGAIN;
2649
2650	XFS_STATS_INC(mp, xs_icluster_flushcnt);
2651	XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2652	return 0;
2653
2654	}
2655
2656	/* Release an inode. */
2657	void
2658	xfs_irele(
2659	struct xfs_inode *ip)
2660	{
2661	trace_xfs_irele(ip, _RET_IP_);
2662	iput(VFS_I(ip));
2663	}
2664
2665	/*
2666	* Ensure all commited transactions touching the inode are written to the log.
2667	*/
2668	int
2669	xfs_log_force_inode(
2670	struct xfs_inode *ip)
2671	{
2672	struct xfs_inode_log_item *iip = ip->i_itemp;
2673	xfs_csn_t seq = 0;
2674
2675	if (!iip)
2676	return 0;
2677
2678	spin_lock(lock: &iip->ili_lock);
2679	seq = iip->ili_commit_seq;
2680	spin_unlock(lock: &iip->ili_lock);
2681
2682	if (!seq)
2683	return 0;
2684	return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2685	}
2686
2687	/*
2688	* Grab the exclusive iolock for a data copy from src to dest, making sure to
2689	* abide vfs locking order (lowest pointer value goes first) and breaking the
2690	* layout leases before proceeding. The loop is needed because we cannot call
2691	* the blocking break_layout() with the iolocks held, and therefore have to
2692	* back out both locks.
2693	*/
2694	static int
2695	xfs_iolock_two_inodes_and_break_layout(
2696	struct inode *src,
2697	struct inode *dest)
2698	{
2699	int error;
2700
2701	if (src > dest)
2702	swap(src, dest);
2703
2704	retry:
2705	/* Wait to break both inodes' layouts before we start locking. */
2706	error = break_layout(inode: src, wait: true);
2707	if (error)
2708	return error;
2709	if (src != dest) {
2710	error = break_layout(inode: dest, wait: true);
2711	if (error)
2712	return error;
2713	}
2714
2715	/* Lock one inode and make sure nobody got in and leased it. */
2716	inode_lock(inode: src);
2717	error = break_layout(inode: src, wait: false);
2718	if (error) {
2719	inode_unlock(inode: src);
2720	if (error == -EWOULDBLOCK)
2721	goto retry;
2722	return error;
2723	}
2724
2725	if (src == dest)
2726	return 0;
2727
2728	/* Lock the other inode and make sure nobody got in and leased it. */
2729	inode_lock_nested(inode: dest, subclass: I_MUTEX_NONDIR2);
2730	error = break_layout(inode: dest, wait: false);
2731	if (error) {
2732	inode_unlock(inode: src);
2733	inode_unlock(inode: dest);
2734	if (error == -EWOULDBLOCK)
2735	goto retry;
2736	return error;
2737	}
2738
2739	return 0;
2740	}
2741
2742	static int
2743	xfs_mmaplock_two_inodes_and_break_dax_layout(
2744	struct xfs_inode *ip1,
2745	struct xfs_inode *ip2)
2746	{
2747	int error;
2748
2749	if (ip1->i_ino > ip2->i_ino)
2750	swap(ip1, ip2);
2751
2752	again:
2753	/* Lock the first inode */
2754	xfs_ilock(ip: ip1, XFS_MMAPLOCK_EXCL);
2755	error = xfs_break_dax_layouts(inode: VFS_I(ip: ip1));
2756	if (error) {
2757	xfs_iunlock(ip: ip1, XFS_MMAPLOCK_EXCL);
2758	return error;
2759	}
2760
2761	if (ip1 == ip2)
2762	return 0;
2763
2764	/* Nested lock the second inode */
2765	xfs_ilock(ip: ip2, lock_flags: xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, subclass: 1));
2766	/*
2767	* We cannot use xfs_break_dax_layouts() directly here because it may
2768	* need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2769	* for this nested lock case.
2770	*/
2771	error = dax_break_layout(inode: VFS_I(ip: ip2), start: 0, end: -1, NULL);
2772	if (error) {
2773	xfs_iunlock(ip: ip2, XFS_MMAPLOCK_EXCL);
2774	xfs_iunlock(ip: ip1, XFS_MMAPLOCK_EXCL);
2775	goto again;
2776	}
2777
2778	return 0;
2779	}
2780
2781	/*
2782	* Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2783	* mmap activity.
2784	*/
2785	int
2786	xfs_ilock2_io_mmap(
2787	struct xfs_inode *ip1,
2788	struct xfs_inode *ip2)
2789	{
2790	int ret;
2791
2792	ret = xfs_iolock_two_inodes_and_break_layout(src: VFS_I(ip: ip1), dest: VFS_I(ip: ip2));
2793	if (ret)
2794	return ret;
2795
2796	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2797	ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2798	if (ret) {
2799	inode_unlock(inode: VFS_I(ip: ip2));
2800	if (ip1 != ip2)
2801	inode_unlock(inode: VFS_I(ip: ip1));
2802	return ret;
2803	}
2804	} else
2805	filemap_invalidate_lock_two(mapping1: VFS_I(ip: ip1)->i_mapping,
2806	mapping2: VFS_I(ip: ip2)->i_mapping);
2807
2808	return 0;
2809	}
2810
2811	/* Unlock both inodes to allow IO and mmap activity. */
2812	void
2813	xfs_iunlock2_io_mmap(
2814	struct xfs_inode *ip1,
2815	struct xfs_inode *ip2)
2816	{
2817	if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2818	xfs_iunlock(ip: ip2, XFS_MMAPLOCK_EXCL);
2819	if (ip1 != ip2)
2820	xfs_iunlock(ip: ip1, XFS_MMAPLOCK_EXCL);
2821	} else
2822	filemap_invalidate_unlock_two(mapping1: VFS_I(ip: ip1)->i_mapping,
2823	mapping2: VFS_I(ip: ip2)->i_mapping);
2824
2825	inode_unlock(inode: VFS_I(ip: ip2));
2826	if (ip1 != ip2)
2827	inode_unlock(inode: VFS_I(ip: ip1));
2828	}
2829
2830	/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2831	void
2832	xfs_iunlock2_remapping(
2833	struct xfs_inode *ip1,
2834	struct xfs_inode *ip2)
2835	{
2836	xfs_iflags_clear(ip: ip1, XFS_IREMAPPING);
2837
2838	if (ip1 != ip2)
2839	xfs_iunlock(ip: ip1, XFS_MMAPLOCK_SHARED);
2840	xfs_iunlock(ip: ip2, XFS_MMAPLOCK_EXCL);
2841
2842	if (ip1 != ip2)
2843	inode_unlock_shared(inode: VFS_I(ip: ip1));
2844	inode_unlock(inode: VFS_I(ip: ip2));
2845	}
2846
2847	/*
2848	* Reload the incore inode list for this inode. Caller should ensure that
2849	* the link count cannot change, either by taking ILOCK_SHARED or otherwise
2850	* preventing other threads from executing.
2851	*/
2852	int
2853	xfs_inode_reload_unlinked_bucket(
2854	struct xfs_trans *tp,
2855	struct xfs_inode *ip)
2856	{
2857	struct xfs_mount *mp = tp->t_mountp;
2858	struct xfs_buf *agibp;
2859	struct xfs_agi *agi;
2860	struct xfs_perag *pag;
2861	xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2862	xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2863	xfs_agino_t prev_agino, next_agino;
2864	unsigned int bucket;
2865	bool foundit = false;
2866	int error;
2867
2868	/* Grab the first inode in the list */
2869	pag = xfs_perag_get(mp, agno);
2870	error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2871	xfs_perag_put(pag);
2872	if (error)
2873	return error;
2874
2875	/*
2876	* We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2877	* incore unlinked list pointers for this inode. Check once more to
2878	* see if we raced with anyone else to reload the unlinked list.
2879	*/
2880	if (!xfs_inode_unlinked_incomplete(ip)) {
2881	foundit = true;
2882	goto out_agibp;
2883	}
2884
2885	bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2886	agi = agibp->b_addr;
2887
2888	trace_xfs_inode_reload_unlinked_bucket(ip);
2889
2890	xfs_info_ratelimited(mp,
2891	"Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
2892	agino, agno);
2893
2894	prev_agino = NULLAGINO;
2895	next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2896	while (next_agino != NULLAGINO) {
2897	struct xfs_inode *next_ip = NULL;
2898
2899	/* Found this caller's inode, set its backlink. */
2900	if (next_agino == agino) {
2901	next_ip = ip;
2902	next_ip->i_prev_unlinked = prev_agino;
2903	foundit = true;
2904	goto next_inode;
2905	}
2906
2907	/* Try in-memory lookup first. */
2908	next_ip = xfs_iunlink_lookup(pag, next_agino);
2909	if (next_ip)
2910	goto next_inode;
2911
2912	/* Inode not in memory, try reloading it. */
2913	error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2914	next_agino);
2915	if (error)
2916	break;
2917
2918	/* Grab the reloaded inode. */
2919	next_ip = xfs_iunlink_lookup(pag, next_agino);
2920	if (!next_ip) {
2921	/* No incore inode at all? We reloaded it... */
2922	ASSERT(next_ip != NULL);
2923	error = -EFSCORRUPTED;
2924	break;
2925	}
2926
2927	next_inode:
2928	prev_agino = next_agino;
2929	next_agino = next_ip->i_next_unlinked;
2930	}
2931
2932	out_agibp:
2933	xfs_trans_brelse(tp, agibp);
2934	/* Should have found this inode somewhere in the iunlinked bucket. */
2935	if (!error && !foundit)
2936	error = -EFSCORRUPTED;
2937	return error;
2938	}
2939
2940	/* Decide if this inode is missing its unlinked list and reload it. */
2941	int
2942	xfs_inode_reload_unlinked(
2943	struct xfs_inode *ip)
2944	{
2945	struct xfs_trans *tp;
2946	int error = 0;
2947
2948	tp = xfs_trans_alloc_empty(mp: ip->i_mount);
2949	xfs_ilock(ip, XFS_ILOCK_SHARED);
2950	if (xfs_inode_unlinked_incomplete(ip))
2951	error = xfs_inode_reload_unlinked_bucket(tp, ip);
2952	xfs_iunlock(ip, XFS_ILOCK_SHARED);
2953	xfs_trans_cancel(tp);
2954
2955	return error;
2956	}
2957
2958	/* Has this inode fork been zapped by repair? */
2959	bool
2960	xfs_ifork_zapped(
2961	const struct xfs_inode *ip,
2962	int whichfork)
2963	{
2964	unsigned int datamask = 0;
2965
2966	switch (whichfork) {
2967	case XFS_DATA_FORK:
2968	switch (ip->i_vnode.i_mode & S_IFMT) {
2969	case S_IFDIR:
2970	datamask = XFS_SICK_INO_DIR_ZAPPED;
2971	break;
2972	case S_IFLNK:
2973	datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
2974	break;
2975	}
2976	return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
2977	case XFS_ATTR_FORK:
2978	return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
2979	default:
2980	return false;
2981	}
2982	}
2983
2984	/* Compute the number of data and realtime blocks used by a file. */
2985	void
2986	xfs_inode_count_blocks(
2987	struct xfs_trans *tp,
2988	struct xfs_inode *ip,
2989	xfs_filblks_t *dblocks,
2990	xfs_filblks_t *rblocks)
2991	{
2992	struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
2993
2994	*rblocks = 0;
2995	if (XFS_IS_REALTIME_INODE(ip))
2996	xfs_bmap_count_leaves(ifp, rblocks);
2997	*dblocks = ip->i_nblocks - *rblocks;
2998	}
2999
3000	static void
3001	xfs_wait_dax_page(
3002	struct inode *inode)
3003	{
3004	struct xfs_inode *ip = XFS_I(inode);
3005
3006	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
3007	schedule();
3008	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
3009	}
3010
3011	int
3012	xfs_break_dax_layouts(
3013	struct inode *inode)
3014	{
3015	xfs_assert_ilocked(ip: XFS_I(inode), XFS_MMAPLOCK_EXCL);
3016
3017	return dax_break_layout_inode(inode, cb: xfs_wait_dax_page);
3018	}
3019
3020	int
3021	xfs_break_layouts(
3022	struct inode *inode,
3023	uint *iolock,
3024	enum layout_break_reason reason)
3025	{
3026	bool retry;
3027	int error;
3028
3029	xfs_assert_ilocked(ip: XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3030
3031	do {
3032	retry = false;
3033	switch (reason) {
3034	case BREAK_UNMAP:
3035	error = xfs_break_dax_layouts(inode);
3036	if (error)
3037	break;
3038	fallthrough;
3039	case BREAK_WRITE:
3040	error = xfs_break_leased_layouts(inode, iolock, did_unlock: &retry);
3041	break;
3042	default:
3043	WARN_ON_ONCE(1);
3044	error = -EINVAL;
3045	}
3046	} while (error == 0 && retry);
3047
3048	return error;
3049	}
3050
3051	/* Returns the size of fundamental allocation unit for a file, in bytes. */
3052	unsigned int
3053	xfs_inode_alloc_unitsize(
3054	struct xfs_inode *ip)
3055	{
3056	unsigned int blocks = 1;
3057
3058	if (XFS_IS_REALTIME_INODE(ip))
3059	blocks = ip->i_mount->m_sb.sb_rextsize;
3060
3061	return XFS_FSB_TO_B(ip->i_mount, blocks);
3062	}
3063
3064	/* Should we always be using copy on write for file writes? */
3065	bool
3066	xfs_is_always_cow_inode(
3067	const struct xfs_inode *ip)
3068	{
3069	return xfs_is_zoned_inode(ip) ||
3070	(ip->i_mount->m_always_cow && xfs_has_reflink(mp: ip->i_mount));
3071	}
3072