Loading...
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
48struct 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 */
65uint
66xfs_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
77uint
78xfs_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 */
95static inline void
96xfs_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 */
139void
140xfs_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(&VFS_I(ip)->i_rwsem,
150 XFS_IOLOCK_DEP(lock_flags));
151 } else if (lock_flags & XFS_IOLOCK_SHARED) {
152 down_read_nested(&VFS_I(ip)->i_rwsem,
153 XFS_IOLOCK_DEP(lock_flags));
154 }
155
156 if (lock_flags & XFS_MMAPLOCK_EXCL) {
157 down_write_nested(&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(&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(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
166 else if (lock_flags & XFS_ILOCK_SHARED)
167 down_read_nested(&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 */
182int
183xfs_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(&VFS_I(ip)->i_rwsem))
193 goto out;
194 } else if (lock_flags & XFS_IOLOCK_SHARED) {
195 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
196 goto out;
197 }
198
199 if (lock_flags & XFS_MMAPLOCK_EXCL) {
200 if (!down_write_trylock(&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(&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(&ip->i_lock))
209 goto out_undo_mmaplock;
210 } else if (lock_flags & XFS_ILOCK_SHARED) {
211 if (!down_read_trylock(&ip->i_lock))
212 goto out_undo_mmaplock;
213 }
214 return 1;
215
216out_undo_mmaplock:
217 if (lock_flags & XFS_MMAPLOCK_EXCL)
218 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
219 else if (lock_flags & XFS_MMAPLOCK_SHARED)
220 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
221out_undo_iolock:
222 if (lock_flags & XFS_IOLOCK_EXCL)
223 up_write(&VFS_I(ip)->i_rwsem);
224 else if (lock_flags & XFS_IOLOCK_SHARED)
225 up_read(&VFS_I(ip)->i_rwsem);
226out:
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 */
242void
243xfs_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(&VFS_I(ip)->i_rwsem);
251 else if (lock_flags & XFS_IOLOCK_SHARED)
252 up_read(&VFS_I(ip)->i_rwsem);
253
254 if (lock_flags & XFS_MMAPLOCK_EXCL)
255 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
256 else if (lock_flags & XFS_MMAPLOCK_SHARED)
257 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
258
259 if (lock_flags & XFS_ILOCK_EXCL)
260 up_write(&ip->i_lock);
261 else if (lock_flags & XFS_ILOCK_SHARED)
262 up_read(&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 */
271void
272xfs_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(&ip->i_lock);
282 if (lock_flags & XFS_MMAPLOCK_EXCL)
283 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
284 if (lock_flags & XFS_IOLOCK_EXCL)
285 downgrade_write(&VFS_I(ip)->i_rwsem);
286
287 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
288}
289
290void
291xfs_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(&ip->i_lock);
301 else if (lock_flags & XFS_ILOCK_EXCL)
302 rwsem_assert_held_write_nolockdep(&ip->i_lock);
303
304 if (lock_flags & XFS_MMAPLOCK_SHARED)
305 rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
306 else if (lock_flags & XFS_MMAPLOCK_EXCL)
307 rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
308
309 if (lock_flags & XFS_IOLOCK_SHARED)
310 rwsem_assert_held(&VFS_I(ip)->i_rwsem);
311 else if (lock_flags & XFS_IOLOCK_EXCL)
312 rwsem_assert_held_write(&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)
322static bool
323xfs_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 */
338static inline uint
339xfs_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 */
381void
382xfs_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
415again:
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(ips[i], xfs_lock_inumorder(lock_mode, 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(ips[i], xfs_lock_inumorder(lock_mode, 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(ips[j], lock_mode);
467 }
468
469 if ((attempts % 5) == 0) {
470 delay(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 */
482void
483xfs_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(ip0, xfs_lock_inumorder(ip0_mode, 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(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
516 xfs_iunlock(ip0, ip0_mode);
517 if ((++attempts % 5) == 0)
518 delay(1); /* Don't just spin the CPU */
519 goto again;
520 }
521 } else {
522 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 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 */
532int
533xfs_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, name);
543
544 if (xfs_is_shutdown(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(dp->i_mount, NULL, inum, 0, 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
569out_irele:
570 xfs_irele(*ipp);
571out_free_name:
572 if (ci_name)
573 kfree(ci_name->name);
574out_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 */
585int
586xfs_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, &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. */
616int
617xfs_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(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(args->idmap, i_user_ns(dir));
636 gid = mapped_fsgid(args->idmap, i_user_ns(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
647int
648xfs_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, 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, &udqp, &gdqp, &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, tres, udqp, gdqp, pdqp, resblks,
701 &tp);
702 if (error == -ENOSPC) {
703 /* flush outstanding delalloc blocks and retry */
704 xfs_flush_inodes(mp);
705 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
706 resblks, &tp);
707 }
708 if (error)
709 goto out_parent;
710
711 xfs_ilock(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, &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(du.ip, XFS_ILOCK_EXCL);
763 xfs_iunlock(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(du.ip, XFS_ILOCK_EXCL);
777 xfs_finish_inode_setup(du.ip);
778 xfs_irele(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(dp, XFS_ILOCK_EXCL);
789 return error;
790}
791
792int
793xfs_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, &udqp, &gdqp, &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, tres, udqp, gdqp, pdqp, resblks,
823 &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, &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
880int
881xfs_link(
882 struct xfs_inode *tdp,
883 struct xfs_inode *sip,
884 struct xfs_name *target_name)
885{
886 struct xfs_dir_update du = {
887 .dp = tdp,
888 .name = target_name,
889 .ip = sip,
890 };
891 struct xfs_mount *mp = tdp->i_mount;
892 struct xfs_trans *tp;
893 int error, nospace_error = 0;
894 int resblks;
895
896 trace_xfs_link(tdp, target_name);
897
898 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
899
900 if (xfs_is_shutdown(mp))
901 return -EIO;
902 if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
903 return -EIO;
904
905 error = xfs_qm_dqattach(sip);
906 if (error)
907 goto std_return;
908
909 error = xfs_qm_dqattach(tdp);
910 if (error)
911 goto std_return;
912
913 error = xfs_parent_start(mp, &du.ppargs);
914 if (error)
915 goto std_return;
916
917 resblks = xfs_link_space_res(mp, target_name->len);
918 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
919 &tp, &nospace_error);
920 if (error)
921 goto out_parent;
922
923 /*
924 * We don't allow reservationless or quotaless hardlinking when parent
925 * pointers are enabled because we can't back out if the xattrs must
926 * grow.
927 */
928 if (du.ppargs && nospace_error) {
929 error = nospace_error;
930 goto error_return;
931 }
932
933 /*
934 * If we are using project inheritance, we only allow hard link
935 * creation in our tree when the project IDs are the same; else
936 * the tree quota mechanism could be circumvented.
937 */
938 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
939 tdp->i_projid != sip->i_projid)) {
940 /*
941 * Project quota setup skips special files which can
942 * leave inodes in a PROJINHERIT directory without a
943 * project ID set. We need to allow links to be made
944 * to these "project-less" inodes because userspace
945 * expects them to succeed after project ID setup,
946 * but everything else should be rejected.
947 */
948 if (!special_file(VFS_I(sip)->i_mode) ||
949 sip->i_projid != 0) {
950 error = -EXDEV;
951 goto error_return;
952 }
953 }
954
955 error = xfs_dir_add_child(tp, resblks, &du);
956 if (error)
957 goto error_return;
958
959 /*
960 * If this is a synchronous mount, make sure that the
961 * link transaction goes to disk before returning to
962 * the user.
963 */
964 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
965 xfs_trans_set_sync(tp);
966
967 error = xfs_trans_commit(tp);
968 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
969 xfs_iunlock(sip, XFS_ILOCK_EXCL);
970 xfs_parent_finish(mp, du.ppargs);
971 return error;
972
973 error_return:
974 xfs_trans_cancel(tp);
975 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
976 xfs_iunlock(sip, XFS_ILOCK_EXCL);
977 out_parent:
978 xfs_parent_finish(mp, du.ppargs);
979 std_return:
980 if (error == -ENOSPC && nospace_error)
981 error = nospace_error;
982 return error;
983}
984
985/* Clear the reflink flag and the cowblocks tag if possible. */
986static void
987xfs_itruncate_clear_reflink_flags(
988 struct xfs_inode *ip)
989{
990 struct xfs_ifork *dfork;
991 struct xfs_ifork *cfork;
992
993 if (!xfs_is_reflink_inode(ip))
994 return;
995 dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
996 cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
997 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
998 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
999 if (cfork->if_bytes == 0)
1000 xfs_inode_clear_cowblocks_tag(ip);
1001}
1002
1003/*
1004 * Free up the underlying blocks past new_size. The new size must be smaller
1005 * than the current size. This routine can be used both for the attribute and
1006 * data fork, and does not modify the inode size, which is left to the caller.
1007 *
1008 * The transaction passed to this routine must have made a permanent log
1009 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1010 * given transaction and start new ones, so make sure everything involved in
1011 * the transaction is tidy before calling here. Some transaction will be
1012 * returned to the caller to be committed. The incoming transaction must
1013 * already include the inode, and both inode locks must be held exclusively.
1014 * The inode must also be "held" within the transaction. On return the inode
1015 * will be "held" within the returned transaction. This routine does NOT
1016 * require any disk space to be reserved for it within the transaction.
1017 *
1018 * If we get an error, we must return with the inode locked and linked into the
1019 * current transaction. This keeps things simple for the higher level code,
1020 * because it always knows that the inode is locked and held in the transaction
1021 * that returns to it whether errors occur or not. We don't mark the inode
1022 * dirty on error so that transactions can be easily aborted if possible.
1023 */
1024int
1025xfs_itruncate_extents_flags(
1026 struct xfs_trans **tpp,
1027 struct xfs_inode *ip,
1028 int whichfork,
1029 xfs_fsize_t new_size,
1030 int flags)
1031{
1032 struct xfs_mount *mp = ip->i_mount;
1033 struct xfs_trans *tp = *tpp;
1034 xfs_fileoff_t first_unmap_block;
1035 int error = 0;
1036
1037 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1038 if (atomic_read(&VFS_I(ip)->i_count))
1039 xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1040 ASSERT(new_size <= XFS_ISIZE(ip));
1041 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1042 ASSERT(ip->i_itemp != NULL);
1043 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1044 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1045
1046 trace_xfs_itruncate_extents_start(ip, new_size);
1047
1048 flags |= xfs_bmapi_aflag(whichfork);
1049
1050 /*
1051 * Since it is possible for space to become allocated beyond
1052 * the end of the file (in a crash where the space is allocated
1053 * but the inode size is not yet updated), simply remove any
1054 * blocks which show up between the new EOF and the maximum
1055 * possible file size.
1056 *
1057 * We have to free all the blocks to the bmbt maximum offset, even if
1058 * the page cache can't scale that far.
1059 */
1060 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1061 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1062 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1063 return 0;
1064 }
1065
1066 error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1067 XFS_MAX_FILEOFF);
1068 if (error)
1069 goto out;
1070
1071 if (whichfork == XFS_DATA_FORK) {
1072 /* Remove all pending CoW reservations. */
1073 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1074 first_unmap_block, XFS_MAX_FILEOFF, true);
1075 if (error)
1076 goto out;
1077
1078 xfs_itruncate_clear_reflink_flags(ip);
1079 }
1080
1081 /*
1082 * Always re-log the inode so that our permanent transaction can keep
1083 * on rolling it forward in the log.
1084 */
1085 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1086
1087 trace_xfs_itruncate_extents_end(ip, new_size);
1088
1089out:
1090 *tpp = tp;
1091 return error;
1092}
1093
1094/*
1095 * Mark all the buffers attached to this directory stale. In theory we should
1096 * never be freeing a directory with any blocks at all, but this covers the
1097 * case where we've recovered a directory swap with a "temporary" directory
1098 * created by online repair and now need to dump it.
1099 */
1100STATIC void
1101xfs_inactive_dir(
1102 struct xfs_inode *dp)
1103{
1104 struct xfs_iext_cursor icur;
1105 struct xfs_bmbt_irec got;
1106 struct xfs_mount *mp = dp->i_mount;
1107 struct xfs_da_geometry *geo = mp->m_dir_geo;
1108 struct xfs_ifork *ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1109 xfs_fileoff_t off;
1110
1111 /*
1112 * Invalidate each directory block. All directory blocks are of
1113 * fsbcount length and alignment, so we only need to walk those same
1114 * offsets. We hold the only reference to this inode, so we must wait
1115 * for the buffer locks.
1116 */
1117 for_each_xfs_iext(ifp, &icur, &got) {
1118 for (off = round_up(got.br_startoff, geo->fsbcount);
1119 off < got.br_startoff + got.br_blockcount;
1120 off += geo->fsbcount) {
1121 struct xfs_buf *bp = NULL;
1122 xfs_fsblock_t fsbno;
1123 int error;
1124
1125 fsbno = (off - got.br_startoff) + got.br_startblock;
1126 error = xfs_buf_incore(mp->m_ddev_targp,
1127 XFS_FSB_TO_DADDR(mp, fsbno),
1128 XFS_FSB_TO_BB(mp, geo->fsbcount),
1129 XBF_LIVESCAN, &bp);
1130 if (error)
1131 continue;
1132
1133 xfs_buf_stale(bp);
1134 xfs_buf_relse(bp);
1135 }
1136 }
1137}
1138
1139/*
1140 * xfs_inactive_truncate
1141 *
1142 * Called to perform a truncate when an inode becomes unlinked.
1143 */
1144STATIC int
1145xfs_inactive_truncate(
1146 struct xfs_inode *ip)
1147{
1148 struct xfs_mount *mp = ip->i_mount;
1149 struct xfs_trans *tp;
1150 int error;
1151
1152 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1153 if (error) {
1154 ASSERT(xfs_is_shutdown(mp));
1155 return error;
1156 }
1157 xfs_ilock(ip, XFS_ILOCK_EXCL);
1158 xfs_trans_ijoin(tp, ip, 0);
1159
1160 /*
1161 * Log the inode size first to prevent stale data exposure in the event
1162 * of a system crash before the truncate completes. See the related
1163 * comment in xfs_vn_setattr_size() for details.
1164 */
1165 ip->i_disk_size = 0;
1166 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1167
1168 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1169 if (error)
1170 goto error_trans_cancel;
1171
1172 ASSERT(ip->i_df.if_nextents == 0);
1173
1174 error = xfs_trans_commit(tp);
1175 if (error)
1176 goto error_unlock;
1177
1178 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1179 return 0;
1180
1181error_trans_cancel:
1182 xfs_trans_cancel(tp);
1183error_unlock:
1184 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1185 return error;
1186}
1187
1188/*
1189 * xfs_inactive_ifree()
1190 *
1191 * Perform the inode free when an inode is unlinked.
1192 */
1193STATIC int
1194xfs_inactive_ifree(
1195 struct xfs_inode *ip)
1196{
1197 struct xfs_mount *mp = ip->i_mount;
1198 struct xfs_trans *tp;
1199 int error;
1200
1201 /*
1202 * We try to use a per-AG reservation for any block needed by the finobt
1203 * tree, but as the finobt feature predates the per-AG reservation
1204 * support a degraded file system might not have enough space for the
1205 * reservation at mount time. In that case try to dip into the reserved
1206 * pool and pray.
1207 *
1208 * Send a warning if the reservation does happen to fail, as the inode
1209 * now remains allocated and sits on the unlinked list until the fs is
1210 * repaired.
1211 */
1212 if (unlikely(mp->m_finobt_nores)) {
1213 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1214 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1215 &tp);
1216 } else {
1217 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1218 }
1219 if (error) {
1220 if (error == -ENOSPC) {
1221 xfs_warn_ratelimited(mp,
1222 "Failed to remove inode(s) from unlinked list. "
1223 "Please free space, unmount and run xfs_repair.");
1224 } else {
1225 ASSERT(xfs_is_shutdown(mp));
1226 }
1227 return error;
1228 }
1229
1230 /*
1231 * We do not hold the inode locked across the entire rolling transaction
1232 * here. We only need to hold it for the first transaction that
1233 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1234 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1235 * here breaks the relationship between cluster buffer invalidation and
1236 * stale inode invalidation on cluster buffer item journal commit
1237 * completion, and can result in leaving dirty stale inodes hanging
1238 * around in memory.
1239 *
1240 * We have no need for serialising this inode operation against other
1241 * operations - we freed the inode and hence reallocation is required
1242 * and that will serialise on reallocating the space the deferops need
1243 * to free. Hence we can unlock the inode on the first commit of
1244 * the transaction rather than roll it right through the deferops. This
1245 * avoids relogging the XFS_ISTALE inode.
1246 *
1247 * We check that xfs_ifree() hasn't grown an internal transaction roll
1248 * by asserting that the inode is still locked when it returns.
1249 */
1250 xfs_ilock(ip, XFS_ILOCK_EXCL);
1251 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1252
1253 error = xfs_ifree(tp, ip);
1254 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1255 if (error) {
1256 /*
1257 * If we fail to free the inode, shut down. The cancel
1258 * might do that, we need to make sure. Otherwise the
1259 * inode might be lost for a long time or forever.
1260 */
1261 if (!xfs_is_shutdown(mp)) {
1262 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1263 __func__, error);
1264 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1265 }
1266 xfs_trans_cancel(tp);
1267 return error;
1268 }
1269
1270 /*
1271 * Credit the quota account(s). The inode is gone.
1272 */
1273 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1274
1275 return xfs_trans_commit(tp);
1276}
1277
1278/*
1279 * Returns true if we need to update the on-disk metadata before we can free
1280 * the memory used by this inode. Updates include freeing post-eof
1281 * preallocations; freeing COW staging extents; and marking the inode free in
1282 * the inobt if it is on the unlinked list.
1283 */
1284bool
1285xfs_inode_needs_inactive(
1286 struct xfs_inode *ip)
1287{
1288 struct xfs_mount *mp = ip->i_mount;
1289 struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1290
1291 /*
1292 * If the inode is already free, then there can be nothing
1293 * to clean up here.
1294 */
1295 if (VFS_I(ip)->i_mode == 0)
1296 return false;
1297
1298 /*
1299 * If this is a read-only mount, don't do this (would generate I/O)
1300 * unless we're in log recovery and cleaning the iunlinked list.
1301 */
1302 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1303 return false;
1304
1305 /* If the log isn't running, push inodes straight to reclaim. */
1306 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1307 return false;
1308
1309 /* Metadata inodes require explicit resource cleanup. */
1310 if (xfs_is_internal_inode(ip))
1311 return false;
1312
1313 /* Want to clean out the cow blocks if there are any. */
1314 if (cow_ifp && cow_ifp->if_bytes > 0)
1315 return true;
1316
1317 /* Unlinked files must be freed. */
1318 if (VFS_I(ip)->i_nlink == 0)
1319 return true;
1320
1321 /*
1322 * This file isn't being freed, so check if there are post-eof blocks
1323 * to free.
1324 *
1325 * Note: don't bother with iolock here since lockdep complains about
1326 * acquiring it in reclaim context. We have the only reference to the
1327 * inode at this point anyways.
1328 */
1329 return xfs_can_free_eofblocks(ip);
1330}
1331
1332/*
1333 * Save health status somewhere, if we're dumping an inode with uncorrected
1334 * errors and online repair isn't running.
1335 */
1336static inline void
1337xfs_inactive_health(
1338 struct xfs_inode *ip)
1339{
1340 struct xfs_mount *mp = ip->i_mount;
1341 struct xfs_perag *pag;
1342 unsigned int sick;
1343 unsigned int checked;
1344
1345 xfs_inode_measure_sickness(ip, &sick, &checked);
1346 if (!sick)
1347 return;
1348
1349 trace_xfs_inode_unfixed_corruption(ip, sick);
1350
1351 if (sick & XFS_SICK_INO_FORGET)
1352 return;
1353
1354 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1355 if (!pag) {
1356 /* There had better still be a perag structure! */
1357 ASSERT(0);
1358 return;
1359 }
1360
1361 xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1362 xfs_perag_put(pag);
1363}
1364
1365/*
1366 * xfs_inactive
1367 *
1368 * This is called when the vnode reference count for the vnode
1369 * goes to zero. If the file has been unlinked, then it must
1370 * now be truncated. Also, we clear all of the read-ahead state
1371 * kept for the inode here since the file is now closed.
1372 */
1373int
1374xfs_inactive(
1375 xfs_inode_t *ip)
1376{
1377 struct xfs_mount *mp;
1378 int error = 0;
1379 int truncate = 0;
1380
1381 /*
1382 * If the inode is already free, then there can be nothing
1383 * to clean up here.
1384 */
1385 if (VFS_I(ip)->i_mode == 0) {
1386 ASSERT(ip->i_df.if_broot_bytes == 0);
1387 goto out;
1388 }
1389
1390 mp = ip->i_mount;
1391 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1392
1393 xfs_inactive_health(ip);
1394
1395 /*
1396 * If this is a read-only mount, don't do this (would generate I/O)
1397 * unless we're in log recovery and cleaning the iunlinked list.
1398 */
1399 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1400 goto out;
1401
1402 /* Metadata inodes require explicit resource cleanup. */
1403 if (xfs_is_internal_inode(ip))
1404 goto out;
1405
1406 /* Try to clean out the cow blocks if there are any. */
1407 if (xfs_inode_has_cow_data(ip)) {
1408 error = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1409 if (error)
1410 goto out;
1411 }
1412
1413 if (VFS_I(ip)->i_nlink != 0) {
1414 /*
1415 * Note: don't bother with iolock here since lockdep complains
1416 * about acquiring it in reclaim context. We have the only
1417 * reference to the inode at this point anyways.
1418 */
1419 if (xfs_can_free_eofblocks(ip))
1420 error = xfs_free_eofblocks(ip);
1421
1422 goto out;
1423 }
1424
1425 if (S_ISREG(VFS_I(ip)->i_mode) &&
1426 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1427 xfs_inode_has_filedata(ip)))
1428 truncate = 1;
1429
1430 if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1431 /*
1432 * If this inode is being inactivated during a quotacheck and
1433 * has not yet been scanned by quotacheck, we /must/ remove
1434 * the dquots from the inode before inactivation changes the
1435 * block and inode counts. Most probably this is a result of
1436 * reloading the incore iunlinked list to purge unrecovered
1437 * unlinked inodes.
1438 */
1439 xfs_qm_dqdetach(ip);
1440 } else {
1441 error = xfs_qm_dqattach(ip);
1442 if (error)
1443 goto out;
1444 }
1445
1446 if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1447 xfs_inactive_dir(ip);
1448 truncate = 1;
1449 }
1450
1451 if (S_ISLNK(VFS_I(ip)->i_mode))
1452 error = xfs_inactive_symlink(ip);
1453 else if (truncate)
1454 error = xfs_inactive_truncate(ip);
1455 if (error)
1456 goto out;
1457
1458 /*
1459 * If there are attributes associated with the file then blow them away
1460 * now. The code calls a routine that recursively deconstructs the
1461 * attribute fork. If also blows away the in-core attribute fork.
1462 */
1463 if (xfs_inode_has_attr_fork(ip)) {
1464 error = xfs_attr_inactive(ip);
1465 if (error)
1466 goto out;
1467 }
1468
1469 ASSERT(ip->i_forkoff == 0);
1470
1471 /*
1472 * Free the inode.
1473 */
1474 error = xfs_inactive_ifree(ip);
1475
1476out:
1477 /*
1478 * We're done making metadata updates for this inode, so we can release
1479 * the attached dquots.
1480 */
1481 xfs_qm_dqdetach(ip);
1482 return error;
1483}
1484
1485/*
1486 * Find an inode on the unlinked list. This does not take references to the
1487 * inode as we have existence guarantees by holding the AGI buffer lock and that
1488 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1489 * don't find the inode in cache, then let the caller handle the situation.
1490 */
1491struct xfs_inode *
1492xfs_iunlink_lookup(
1493 struct xfs_perag *pag,
1494 xfs_agino_t agino)
1495{
1496 struct xfs_inode *ip;
1497
1498 rcu_read_lock();
1499 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1500 if (!ip) {
1501 /* Caller can handle inode not being in memory. */
1502 rcu_read_unlock();
1503 return NULL;
1504 }
1505
1506 /*
1507 * Inode in RCU freeing limbo should not happen. Warn about this and
1508 * let the caller handle the failure.
1509 */
1510 if (WARN_ON_ONCE(!ip->i_ino)) {
1511 rcu_read_unlock();
1512 return NULL;
1513 }
1514 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1515 rcu_read_unlock();
1516 return ip;
1517}
1518
1519/*
1520 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1521 * to @prev_agino. Caller must hold the AGI to synchronize with other changes
1522 * to the unlinked list.
1523 */
1524int
1525xfs_iunlink_reload_next(
1526 struct xfs_trans *tp,
1527 struct xfs_buf *agibp,
1528 xfs_agino_t prev_agino,
1529 xfs_agino_t next_agino)
1530{
1531 struct xfs_perag *pag = agibp->b_pag;
1532 struct xfs_mount *mp = pag_mount(pag);
1533 struct xfs_inode *next_ip = NULL;
1534 int error;
1535
1536 ASSERT(next_agino != NULLAGINO);
1537
1538#ifdef DEBUG
1539 rcu_read_lock();
1540 next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1541 ASSERT(next_ip == NULL);
1542 rcu_read_unlock();
1543#endif
1544
1545 xfs_info_ratelimited(mp,
1546 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
1547 next_agino, pag_agno(pag));
1548
1549 /*
1550 * Use an untrusted lookup just to be cautious in case the AGI has been
1551 * corrupted and now points at a free inode. That shouldn't happen,
1552 * but we'd rather shut down now since we're already running in a weird
1553 * situation.
1554 */
1555 error = xfs_iget(mp, tp, xfs_agino_to_ino(pag, next_agino),
1556 XFS_IGET_UNTRUSTED, 0, &next_ip);
1557 if (error) {
1558 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1559 return error;
1560 }
1561
1562 /* If this is not an unlinked inode, something is very wrong. */
1563 if (VFS_I(next_ip)->i_nlink != 0) {
1564 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1565 error = -EFSCORRUPTED;
1566 goto rele;
1567 }
1568
1569 next_ip->i_prev_unlinked = prev_agino;
1570 trace_xfs_iunlink_reload_next(next_ip);
1571rele:
1572 ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
1573 if (xfs_is_quotacheck_running(mp) && next_ip)
1574 xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1575 xfs_irele(next_ip);
1576 return error;
1577}
1578
1579/*
1580 * Look up the inode number specified and if it is not already marked XFS_ISTALE
1581 * mark it stale. We should only find clean inodes in this lookup that aren't
1582 * already stale.
1583 */
1584static void
1585xfs_ifree_mark_inode_stale(
1586 struct xfs_perag *pag,
1587 struct xfs_inode *free_ip,
1588 xfs_ino_t inum)
1589{
1590 struct xfs_mount *mp = pag_mount(pag);
1591 struct xfs_inode_log_item *iip;
1592 struct xfs_inode *ip;
1593
1594retry:
1595 rcu_read_lock();
1596 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1597
1598 /* Inode not in memory, nothing to do */
1599 if (!ip) {
1600 rcu_read_unlock();
1601 return;
1602 }
1603
1604 /*
1605 * because this is an RCU protected lookup, we could find a recently
1606 * freed or even reallocated inode during the lookup. We need to check
1607 * under the i_flags_lock for a valid inode here. Skip it if it is not
1608 * valid, the wrong inode or stale.
1609 */
1610 spin_lock(&ip->i_flags_lock);
1611 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1612 goto out_iflags_unlock;
1613
1614 /*
1615 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
1616 * other inodes that we did not find in the list attached to the buffer
1617 * and are not already marked stale. If we can't lock it, back off and
1618 * retry.
1619 */
1620 if (ip != free_ip) {
1621 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1622 spin_unlock(&ip->i_flags_lock);
1623 rcu_read_unlock();
1624 delay(1);
1625 goto retry;
1626 }
1627 }
1628 ip->i_flags |= XFS_ISTALE;
1629
1630 /*
1631 * If the inode is flushing, it is already attached to the buffer. All
1632 * we needed to do here is mark the inode stale so buffer IO completion
1633 * will remove it from the AIL.
1634 */
1635 iip = ip->i_itemp;
1636 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1637 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1638 ASSERT(iip->ili_last_fields);
1639 goto out_iunlock;
1640 }
1641
1642 /*
1643 * Inodes not attached to the buffer can be released immediately.
1644 * Everything else has to go through xfs_iflush_abort() on journal
1645 * commit as the flock synchronises removal of the inode from the
1646 * cluster buffer against inode reclaim.
1647 */
1648 if (!iip || list_empty(&iip->ili_item.li_bio_list))
1649 goto out_iunlock;
1650
1651 __xfs_iflags_set(ip, XFS_IFLUSHING);
1652 spin_unlock(&ip->i_flags_lock);
1653 rcu_read_unlock();
1654
1655 /* we have a dirty inode in memory that has not yet been flushed. */
1656 spin_lock(&iip->ili_lock);
1657 iip->ili_last_fields = iip->ili_fields;
1658 iip->ili_fields = 0;
1659 iip->ili_fsync_fields = 0;
1660 spin_unlock(&iip->ili_lock);
1661 ASSERT(iip->ili_last_fields);
1662
1663 if (ip != free_ip)
1664 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1665 return;
1666
1667out_iunlock:
1668 if (ip != free_ip)
1669 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1670out_iflags_unlock:
1671 spin_unlock(&ip->i_flags_lock);
1672 rcu_read_unlock();
1673}
1674
1675/*
1676 * A big issue when freeing the inode cluster is that we _cannot_ skip any
1677 * inodes that are in memory - they all must be marked stale and attached to
1678 * the cluster buffer.
1679 */
1680static int
1681xfs_ifree_cluster(
1682 struct xfs_trans *tp,
1683 struct xfs_perag *pag,
1684 struct xfs_inode *free_ip,
1685 struct xfs_icluster *xic)
1686{
1687 struct xfs_mount *mp = free_ip->i_mount;
1688 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1689 struct xfs_buf *bp;
1690 xfs_daddr_t blkno;
1691 xfs_ino_t inum = xic->first_ino;
1692 int nbufs;
1693 int i, j;
1694 int ioffset;
1695 int error;
1696
1697 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1698
1699 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1700 /*
1701 * The allocation bitmap tells us which inodes of the chunk were
1702 * physically allocated. Skip the cluster if an inode falls into
1703 * a sparse region.
1704 */
1705 ioffset = inum - xic->first_ino;
1706 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1707 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1708 continue;
1709 }
1710
1711 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1712 XFS_INO_TO_AGBNO(mp, inum));
1713
1714 /*
1715 * We obtain and lock the backing buffer first in the process
1716 * here to ensure dirty inodes attached to the buffer remain in
1717 * the flushing state while we mark them stale.
1718 *
1719 * If we scan the in-memory inodes first, then buffer IO can
1720 * complete before we get a lock on it, and hence we may fail
1721 * to mark all the active inodes on the buffer stale.
1722 */
1723 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
1724 mp->m_bsize * igeo->blocks_per_cluster,
1725 XBF_UNMAPPED, &bp);
1726 if (error)
1727 return error;
1728
1729 /*
1730 * This buffer may not have been correctly initialised as we
1731 * didn't read it from disk. That's not important because we are
1732 * only using to mark the buffer as stale in the log, and to
1733 * attach stale cached inodes on it.
1734 *
1735 * For the inode that triggered the cluster freeing, this
1736 * attachment may occur in xfs_inode_item_precommit() after we
1737 * have marked this buffer stale. If this buffer was not in
1738 * memory before xfs_ifree_cluster() started, it will not be
1739 * marked XBF_DONE and this will cause problems later in
1740 * xfs_inode_item_precommit() when we trip over a (stale, !done)
1741 * buffer to attached to the transaction.
1742 *
1743 * Hence we have to mark the buffer as XFS_DONE here. This is
1744 * safe because we are also marking the buffer as XBF_STALE and
1745 * XFS_BLI_STALE. That means it will never be dispatched for
1746 * IO and it won't be unlocked until the cluster freeing has
1747 * been committed to the journal and the buffer unpinned. If it
1748 * is written, we want to know about it, and we want it to
1749 * fail. We can acheive this by adding a write verifier to the
1750 * buffer.
1751 */
1752 bp->b_flags |= XBF_DONE;
1753 bp->b_ops = &xfs_inode_buf_ops;
1754
1755 /*
1756 * Now we need to set all the cached clean inodes as XFS_ISTALE,
1757 * too. This requires lookups, and will skip inodes that we've
1758 * already marked XFS_ISTALE.
1759 */
1760 for (i = 0; i < igeo->inodes_per_cluster; i++)
1761 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
1762
1763 xfs_trans_stale_inode_buf(tp, bp);
1764 xfs_trans_binval(tp, bp);
1765 }
1766 return 0;
1767}
1768
1769/*
1770 * This is called to return an inode to the inode free list. The inode should
1771 * already be truncated to 0 length and have no pages associated with it. This
1772 * routine also assumes that the inode is already a part of the transaction.
1773 *
1774 * The on-disk copy of the inode will have been added to the list of unlinked
1775 * inodes in the AGI. We need to remove the inode from that list atomically with
1776 * respect to freeing it here.
1777 */
1778int
1779xfs_ifree(
1780 struct xfs_trans *tp,
1781 struct xfs_inode *ip)
1782{
1783 struct xfs_mount *mp = ip->i_mount;
1784 struct xfs_perag *pag;
1785 struct xfs_icluster xic = { 0 };
1786 struct xfs_inode_log_item *iip = ip->i_itemp;
1787 int error;
1788
1789 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1790 ASSERT(VFS_I(ip)->i_nlink == 0);
1791 ASSERT(ip->i_df.if_nextents == 0);
1792 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1793 ASSERT(ip->i_nblocks == 0);
1794
1795 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1796
1797 error = xfs_inode_uninit(tp, pag, ip, &xic);
1798 if (error)
1799 goto out;
1800
1801 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1802 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1803
1804 /* Don't attempt to replay owner changes for a deleted inode */
1805 spin_lock(&iip->ili_lock);
1806 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1807 spin_unlock(&iip->ili_lock);
1808
1809 if (xic.deleted)
1810 error = xfs_ifree_cluster(tp, pag, ip, &xic);
1811out:
1812 xfs_perag_put(pag);
1813 return error;
1814}
1815
1816/*
1817 * This is called to unpin an inode. The caller must have the inode locked
1818 * in at least shared mode so that the buffer cannot be subsequently pinned
1819 * once someone is waiting for it to be unpinned.
1820 */
1821static void
1822xfs_iunpin(
1823 struct xfs_inode *ip)
1824{
1825 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1826
1827 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1828
1829 /* Give the log a push to start the unpinning I/O */
1830 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
1831
1832}
1833
1834static void
1835__xfs_iunpin_wait(
1836 struct xfs_inode *ip)
1837{
1838 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
1839 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1840
1841 xfs_iunpin(ip);
1842
1843 do {
1844 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1845 if (xfs_ipincount(ip))
1846 io_schedule();
1847 } while (xfs_ipincount(ip));
1848 finish_wait(wq, &wait.wq_entry);
1849}
1850
1851void
1852xfs_iunpin_wait(
1853 struct xfs_inode *ip)
1854{
1855 if (xfs_ipincount(ip))
1856 __xfs_iunpin_wait(ip);
1857}
1858
1859/*
1860 * Removing an inode from the namespace involves removing the directory entry
1861 * and dropping the link count on the inode. Removing the directory entry can
1862 * result in locking an AGF (directory blocks were freed) and removing a link
1863 * count can result in placing the inode on an unlinked list which results in
1864 * locking an AGI.
1865 *
1866 * The big problem here is that we have an ordering constraint on AGF and AGI
1867 * locking - inode allocation locks the AGI, then can allocate a new extent for
1868 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1869 * removes the inode from the unlinked list, requiring that we lock the AGI
1870 * first, and then freeing the inode can result in an inode chunk being freed
1871 * and hence freeing disk space requiring that we lock an AGF.
1872 *
1873 * Hence the ordering that is imposed by other parts of the code is AGI before
1874 * AGF. This means we cannot remove the directory entry before we drop the inode
1875 * reference count and put it on the unlinked list as this results in a lock
1876 * order of AGF then AGI, and this can deadlock against inode allocation and
1877 * freeing. Therefore we must drop the link counts before we remove the
1878 * directory entry.
1879 *
1880 * This is still safe from a transactional point of view - it is not until we
1881 * get to xfs_defer_finish() that we have the possibility of multiple
1882 * transactions in this operation. Hence as long as we remove the directory
1883 * entry and drop the link count in the first transaction of the remove
1884 * operation, there are no transactional constraints on the ordering here.
1885 */
1886int
1887xfs_remove(
1888 struct xfs_inode *dp,
1889 struct xfs_name *name,
1890 struct xfs_inode *ip)
1891{
1892 struct xfs_dir_update du = {
1893 .dp = dp,
1894 .name = name,
1895 .ip = ip,
1896 };
1897 struct xfs_mount *mp = dp->i_mount;
1898 struct xfs_trans *tp = NULL;
1899 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1900 int dontcare;
1901 int error = 0;
1902 uint resblks;
1903
1904 trace_xfs_remove(dp, name);
1905
1906 if (xfs_is_shutdown(mp))
1907 return -EIO;
1908 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1909 return -EIO;
1910
1911 error = xfs_qm_dqattach(dp);
1912 if (error)
1913 goto std_return;
1914
1915 error = xfs_qm_dqattach(ip);
1916 if (error)
1917 goto std_return;
1918
1919 error = xfs_parent_start(mp, &du.ppargs);
1920 if (error)
1921 goto std_return;
1922
1923 /*
1924 * We try to get the real space reservation first, allowing for
1925 * directory btree deletion(s) implying possible bmap insert(s). If we
1926 * can't get the space reservation then we use 0 instead, and avoid the
1927 * bmap btree insert(s) in the directory code by, if the bmap insert
1928 * tries to happen, instead trimming the LAST block from the directory.
1929 *
1930 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
1931 * the directory code can handle a reservationless update and we don't
1932 * want to prevent a user from trying to free space by deleting things.
1933 */
1934 resblks = xfs_remove_space_res(mp, name->len);
1935 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
1936 &tp, &dontcare);
1937 if (error) {
1938 ASSERT(error != -ENOSPC);
1939 goto out_parent;
1940 }
1941
1942 error = xfs_dir_remove_child(tp, resblks, &du);
1943 if (error)
1944 goto out_trans_cancel;
1945
1946 /*
1947 * If this is a synchronous mount, make sure that the
1948 * remove transaction goes to disk before returning to
1949 * the user.
1950 */
1951 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1952 xfs_trans_set_sync(tp);
1953
1954 error = xfs_trans_commit(tp);
1955 if (error)
1956 goto out_unlock;
1957
1958 if (is_dir && xfs_inode_is_filestream(ip))
1959 xfs_filestream_deassociate(ip);
1960
1961 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1962 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1963 xfs_parent_finish(mp, du.ppargs);
1964 return 0;
1965
1966 out_trans_cancel:
1967 xfs_trans_cancel(tp);
1968 out_unlock:
1969 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1970 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1971 out_parent:
1972 xfs_parent_finish(mp, du.ppargs);
1973 std_return:
1974 return error;
1975}
1976
1977static inline void
1978xfs_iunlock_rename(
1979 struct xfs_inode **i_tab,
1980 int num_inodes)
1981{
1982 int i;
1983
1984 for (i = num_inodes - 1; i >= 0; i--) {
1985 /* Skip duplicate inodes if src and target dps are the same */
1986 if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
1987 continue;
1988 xfs_iunlock(i_tab[i], XFS_ILOCK_EXCL);
1989 }
1990}
1991
1992/*
1993 * Enter all inodes for a rename transaction into a sorted array.
1994 */
1995#define __XFS_SORT_INODES 5
1996STATIC void
1997xfs_sort_for_rename(
1998 struct xfs_inode *dp1, /* in: old (source) directory inode */
1999 struct xfs_inode *dp2, /* in: new (target) directory inode */
2000 struct xfs_inode *ip1, /* in: inode of old entry */
2001 struct xfs_inode *ip2, /* in: inode of new entry */
2002 struct xfs_inode *wip, /* in: whiteout inode */
2003 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2004 int *num_inodes) /* in/out: inodes in array */
2005{
2006 int i;
2007
2008 ASSERT(*num_inodes == __XFS_SORT_INODES);
2009 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2010
2011 /*
2012 * i_tab contains a list of pointers to inodes. We initialize
2013 * the table here & we'll sort it. We will then use it to
2014 * order the acquisition of the inode locks.
2015 *
2016 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2017 */
2018 i = 0;
2019 i_tab[i++] = dp1;
2020 i_tab[i++] = dp2;
2021 i_tab[i++] = ip1;
2022 if (ip2)
2023 i_tab[i++] = ip2;
2024 if (wip)
2025 i_tab[i++] = wip;
2026 *num_inodes = i;
2027
2028 xfs_sort_inodes(i_tab, *num_inodes);
2029}
2030
2031void
2032xfs_sort_inodes(
2033 struct xfs_inode **i_tab,
2034 unsigned int num_inodes)
2035{
2036 int i, j;
2037
2038 ASSERT(num_inodes <= __XFS_SORT_INODES);
2039
2040 /*
2041 * Sort the elements via bubble sort. (Remember, there are at
2042 * most 5 elements to sort, so this is adequate.)
2043 */
2044 for (i = 0; i < num_inodes; i++) {
2045 for (j = 1; j < num_inodes; j++) {
2046 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2047 swap(i_tab[j], i_tab[j - 1]);
2048 }
2049 }
2050}
2051
2052/*
2053 * xfs_rename_alloc_whiteout()
2054 *
2055 * Return a referenced, unlinked, unlocked inode that can be used as a
2056 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2057 * crash between allocating the inode and linking it into the rename transaction
2058 * recovery will free the inode and we won't leak it.
2059 */
2060static int
2061xfs_rename_alloc_whiteout(
2062 struct mnt_idmap *idmap,
2063 struct xfs_name *src_name,
2064 struct xfs_inode *dp,
2065 struct xfs_inode **wip)
2066{
2067 struct xfs_icreate_args args = {
2068 .idmap = idmap,
2069 .pip = dp,
2070 .mode = S_IFCHR | WHITEOUT_MODE,
2071 .flags = XFS_ICREATE_TMPFILE,
2072 };
2073 struct xfs_inode *tmpfile;
2074 struct qstr name;
2075 int error;
2076
2077 error = xfs_create_tmpfile(&args, &tmpfile);
2078 if (error)
2079 return error;
2080
2081 name.name = src_name->name;
2082 name.len = src_name->len;
2083 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2084 if (error) {
2085 xfs_finish_inode_setup(tmpfile);
2086 xfs_irele(tmpfile);
2087 return error;
2088 }
2089
2090 /*
2091 * Prepare the tmpfile inode as if it were created through the VFS.
2092 * Complete the inode setup and flag it as linkable. nlink is already
2093 * zero, so we can skip the drop_nlink.
2094 */
2095 xfs_setup_iops(tmpfile);
2096 xfs_finish_inode_setup(tmpfile);
2097 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2098
2099 *wip = tmpfile;
2100 return 0;
2101}
2102
2103/*
2104 * xfs_rename
2105 */
2106int
2107xfs_rename(
2108 struct mnt_idmap *idmap,
2109 struct xfs_inode *src_dp,
2110 struct xfs_name *src_name,
2111 struct xfs_inode *src_ip,
2112 struct xfs_inode *target_dp,
2113 struct xfs_name *target_name,
2114 struct xfs_inode *target_ip,
2115 unsigned int flags)
2116{
2117 struct xfs_dir_update du_src = {
2118 .dp = src_dp,
2119 .name = src_name,
2120 .ip = src_ip,
2121 };
2122 struct xfs_dir_update du_tgt = {
2123 .dp = target_dp,
2124 .name = target_name,
2125 .ip = target_ip,
2126 };
2127 struct xfs_dir_update du_wip = { };
2128 struct xfs_mount *mp = src_dp->i_mount;
2129 struct xfs_trans *tp;
2130 struct xfs_inode *inodes[__XFS_SORT_INODES];
2131 int i;
2132 int num_inodes = __XFS_SORT_INODES;
2133 bool new_parent = (src_dp != target_dp);
2134 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2135 int spaceres;
2136 bool retried = false;
2137 int error, nospace_error = 0;
2138
2139 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2140
2141 if ((flags & RENAME_EXCHANGE) && !target_ip)
2142 return -EINVAL;
2143
2144 /*
2145 * If we are doing a whiteout operation, allocate the whiteout inode
2146 * we will be placing at the target and ensure the type is set
2147 * appropriately.
2148 */
2149 if (flags & RENAME_WHITEOUT) {
2150 error = xfs_rename_alloc_whiteout(idmap, src_name, target_dp,
2151 &du_wip.ip);
2152 if (error)
2153 return error;
2154
2155 /* setup target dirent info as whiteout */
2156 src_name->type = XFS_DIR3_FT_CHRDEV;
2157 }
2158
2159 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, du_wip.ip,
2160 inodes, &num_inodes);
2161
2162 error = xfs_parent_start(mp, &du_src.ppargs);
2163 if (error)
2164 goto out_release_wip;
2165
2166 if (du_wip.ip) {
2167 error = xfs_parent_start(mp, &du_wip.ppargs);
2168 if (error)
2169 goto out_src_ppargs;
2170 }
2171
2172 if (target_ip) {
2173 error = xfs_parent_start(mp, &du_tgt.ppargs);
2174 if (error)
2175 goto out_wip_ppargs;
2176 }
2177
2178retry:
2179 nospace_error = 0;
2180 spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2181 target_name->len, du_wip.ip != NULL);
2182 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2183 if (error == -ENOSPC) {
2184 nospace_error = error;
2185 spaceres = 0;
2186 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2187 &tp);
2188 }
2189 if (error)
2190 goto out_tgt_ppargs;
2191
2192 /*
2193 * We don't allow reservationless renaming when parent pointers are
2194 * enabled because we can't back out if the xattrs must grow.
2195 */
2196 if (du_src.ppargs && nospace_error) {
2197 error = nospace_error;
2198 xfs_trans_cancel(tp);
2199 goto out_tgt_ppargs;
2200 }
2201
2202 /*
2203 * Attach the dquots to the inodes
2204 */
2205 error = xfs_qm_vop_rename_dqattach(inodes);
2206 if (error) {
2207 xfs_trans_cancel(tp);
2208 goto out_tgt_ppargs;
2209 }
2210
2211 /*
2212 * Lock all the participating inodes. Depending upon whether
2213 * the target_name exists in the target directory, and
2214 * whether the target directory is the same as the source
2215 * directory, we can lock from 2 to 5 inodes.
2216 */
2217 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2218
2219 /*
2220 * Join all the inodes to the transaction.
2221 */
2222 xfs_trans_ijoin(tp, src_dp, 0);
2223 if (new_parent)
2224 xfs_trans_ijoin(tp, target_dp, 0);
2225 xfs_trans_ijoin(tp, src_ip, 0);
2226 if (target_ip)
2227 xfs_trans_ijoin(tp, target_ip, 0);
2228 if (du_wip.ip)
2229 xfs_trans_ijoin(tp, du_wip.ip, 0);
2230
2231 /*
2232 * If we are using project inheritance, we only allow renames
2233 * into our tree when the project IDs are the same; else the
2234 * tree quota mechanism would be circumvented.
2235 */
2236 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2237 target_dp->i_projid != src_ip->i_projid)) {
2238 error = -EXDEV;
2239 goto out_trans_cancel;
2240 }
2241
2242 /* RENAME_EXCHANGE is unique from here on. */
2243 if (flags & RENAME_EXCHANGE) {
2244 error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2245 spaceres);
2246 if (error)
2247 goto out_trans_cancel;
2248 goto out_commit;
2249 }
2250
2251 /*
2252 * Try to reserve quota to handle an expansion of the target directory.
2253 * We'll allow the rename to continue in reservationless mode if we hit
2254 * a space usage constraint. If we trigger reservationless mode, save
2255 * the errno if there isn't any free space in the target directory.
2256 */
2257 if (spaceres != 0) {
2258 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2259 0, false);
2260 if (error == -EDQUOT || error == -ENOSPC) {
2261 if (!retried) {
2262 xfs_trans_cancel(tp);
2263 xfs_iunlock_rename(inodes, num_inodes);
2264 xfs_blockgc_free_quota(target_dp, 0);
2265 retried = true;
2266 goto retry;
2267 }
2268
2269 nospace_error = error;
2270 spaceres = 0;
2271 error = 0;
2272 }
2273 if (error)
2274 goto out_trans_cancel;
2275 }
2276
2277 /*
2278 * We don't allow quotaless renaming when parent pointers are enabled
2279 * because we can't back out if the xattrs must grow.
2280 */
2281 if (du_src.ppargs && nospace_error) {
2282 error = nospace_error;
2283 goto out_trans_cancel;
2284 }
2285
2286 /*
2287 * Lock the AGI buffers we need to handle bumping the nlink of the
2288 * whiteout inode off the unlinked list and to handle dropping the
2289 * nlink of the target inode. Per locking order rules, do this in
2290 * increasing AG order and before directory block allocation tries to
2291 * grab AGFs because we grab AGIs before AGFs.
2292 *
2293 * The (vfs) caller must ensure that if src is a directory then
2294 * target_ip is either null or an empty directory.
2295 */
2296 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2297 if (inodes[i] == du_wip.ip ||
2298 (inodes[i] == target_ip &&
2299 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2300 struct xfs_perag *pag;
2301 struct xfs_buf *bp;
2302
2303 pag = xfs_perag_get(mp,
2304 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2305 error = xfs_read_agi(pag, tp, 0, &bp);
2306 xfs_perag_put(pag);
2307 if (error)
2308 goto out_trans_cancel;
2309 }
2310 }
2311
2312 error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2313 &du_wip);
2314 if (error)
2315 goto out_trans_cancel;
2316
2317 if (du_wip.ip) {
2318 /*
2319 * Now we have a real link, clear the "I'm a tmpfile" state
2320 * flag from the inode so it doesn't accidentally get misused in
2321 * future.
2322 */
2323 VFS_I(du_wip.ip)->i_state &= ~I_LINKABLE;
2324 }
2325
2326out_commit:
2327 /*
2328 * If this is a synchronous mount, make sure that the rename
2329 * transaction goes to disk before returning to the user.
2330 */
2331 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2332 xfs_trans_set_sync(tp);
2333
2334 error = xfs_trans_commit(tp);
2335 nospace_error = 0;
2336 goto out_unlock;
2337
2338out_trans_cancel:
2339 xfs_trans_cancel(tp);
2340out_unlock:
2341 xfs_iunlock_rename(inodes, num_inodes);
2342out_tgt_ppargs:
2343 xfs_parent_finish(mp, du_tgt.ppargs);
2344out_wip_ppargs:
2345 xfs_parent_finish(mp, du_wip.ppargs);
2346out_src_ppargs:
2347 xfs_parent_finish(mp, du_src.ppargs);
2348out_release_wip:
2349 if (du_wip.ip)
2350 xfs_irele(du_wip.ip);
2351 if (error == -ENOSPC && nospace_error)
2352 error = nospace_error;
2353 return error;
2354}
2355
2356static int
2357xfs_iflush(
2358 struct xfs_inode *ip,
2359 struct xfs_buf *bp)
2360{
2361 struct xfs_inode_log_item *iip = ip->i_itemp;
2362 struct xfs_dinode *dip;
2363 struct xfs_mount *mp = ip->i_mount;
2364 int error;
2365
2366 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2367 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2368 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2369 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2370 ASSERT(iip->ili_item.li_buf == bp);
2371
2372 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
2373
2374 /*
2375 * We don't flush the inode if any of the following checks fail, but we
2376 * do still update the log item and attach to the backing buffer as if
2377 * the flush happened. This is a formality to facilitate predictable
2378 * error handling as the caller will shutdown and fail the buffer.
2379 */
2380 error = -EFSCORRUPTED;
2381 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
2382 mp, XFS_ERRTAG_IFLUSH_1)) {
2383 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2384 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2385 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2386 goto flush_out;
2387 }
2388 if (S_ISREG(VFS_I(ip)->i_mode)) {
2389 if (XFS_TEST_ERROR(
2390 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2391 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
2392 mp, XFS_ERRTAG_IFLUSH_3)) {
2393 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2394 "%s: Bad regular inode %llu, ptr "PTR_FMT,
2395 __func__, ip->i_ino, ip);
2396 goto flush_out;
2397 }
2398 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2399 if (XFS_TEST_ERROR(
2400 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2401 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2402 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
2403 mp, XFS_ERRTAG_IFLUSH_4)) {
2404 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2405 "%s: Bad directory inode %llu, ptr "PTR_FMT,
2406 __func__, ip->i_ino, ip);
2407 goto flush_out;
2408 }
2409 }
2410 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2411 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
2412 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2413 "%s: detected corrupt incore inode %llu, "
2414 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2415 __func__, ip->i_ino,
2416 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2417 ip->i_nblocks, ip);
2418 goto flush_out;
2419 }
2420 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
2421 mp, XFS_ERRTAG_IFLUSH_6)) {
2422 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2423 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2424 __func__, ip->i_ino, ip->i_forkoff, ip);
2425 goto flush_out;
2426 }
2427
2428 /*
2429 * Inode item log recovery for v2 inodes are dependent on the flushiter
2430 * count for correct sequencing. We bump the flush iteration count so
2431 * we can detect flushes which postdate a log record during recovery.
2432 * This is redundant as we now log every change and hence this can't
2433 * happen but we need to still do it to ensure backwards compatibility
2434 * with old kernels that predate logging all inode changes.
2435 */
2436 if (!xfs_has_v3inodes(mp))
2437 ip->i_flushiter++;
2438
2439 /*
2440 * If there are inline format data / attr forks attached to this inode,
2441 * make sure they are not corrupt.
2442 */
2443 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2444 xfs_ifork_verify_local_data(ip))
2445 goto flush_out;
2446 if (xfs_inode_has_attr_fork(ip) &&
2447 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2448 xfs_ifork_verify_local_attr(ip))
2449 goto flush_out;
2450
2451 /*
2452 * Copy the dirty parts of the inode into the on-disk inode. We always
2453 * copy out the core of the inode, because if the inode is dirty at all
2454 * the core must be.
2455 */
2456 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2457
2458 /* Wrap, we never let the log put out DI_MAX_FLUSH */
2459 if (!xfs_has_v3inodes(mp)) {
2460 if (ip->i_flushiter == DI_MAX_FLUSH)
2461 ip->i_flushiter = 0;
2462 }
2463
2464 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2465 if (xfs_inode_has_attr_fork(ip))
2466 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2467
2468 /*
2469 * We've recorded everything logged in the inode, so we'd like to clear
2470 * the ili_fields bits so we don't log and flush things unnecessarily.
2471 * However, we can't stop logging all this information until the data
2472 * we've copied into the disk buffer is written to disk. If we did we
2473 * might overwrite the copy of the inode in the log with all the data
2474 * after re-logging only part of it, and in the face of a crash we
2475 * wouldn't have all the data we need to recover.
2476 *
2477 * What we do is move the bits to the ili_last_fields field. When
2478 * logging the inode, these bits are moved back to the ili_fields field.
2479 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2480 * we know that the information those bits represent is permanently on
2481 * disk. As long as the flush completes before the inode is logged
2482 * again, then both ili_fields and ili_last_fields will be cleared.
2483 */
2484 error = 0;
2485flush_out:
2486 spin_lock(&iip->ili_lock);
2487 iip->ili_last_fields = iip->ili_fields;
2488 iip->ili_fields = 0;
2489 iip->ili_fsync_fields = 0;
2490 set_bit(XFS_LI_FLUSHING, &iip->ili_item.li_flags);
2491 spin_unlock(&iip->ili_lock);
2492
2493 /*
2494 * Store the current LSN of the inode so that we can tell whether the
2495 * item has moved in the AIL from xfs_buf_inode_iodone().
2496 */
2497 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2498 &iip->ili_item.li_lsn);
2499
2500 /* generate the checksum. */
2501 xfs_dinode_calc_crc(mp, dip);
2502 if (error)
2503 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2504 return error;
2505}
2506
2507/*
2508 * Non-blocking flush of dirty inode metadata into the backing buffer.
2509 *
2510 * The caller must have a reference to the inode and hold the cluster buffer
2511 * locked. The function will walk across all the inodes on the cluster buffer it
2512 * can find and lock without blocking, and flush them to the cluster buffer.
2513 *
2514 * On successful flushing of at least one inode, the caller must write out the
2515 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2516 * the caller needs to release the buffer. On failure, the filesystem will be
2517 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2518 * will be returned.
2519 */
2520int
2521xfs_iflush_cluster(
2522 struct xfs_buf *bp)
2523{
2524 struct xfs_mount *mp = bp->b_mount;
2525 struct xfs_log_item *lip, *n;
2526 struct xfs_inode *ip;
2527 struct xfs_inode_log_item *iip;
2528 int clcount = 0;
2529 int error = 0;
2530
2531 /*
2532 * We must use the safe variant here as on shutdown xfs_iflush_abort()
2533 * will remove itself from the list.
2534 */
2535 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2536 iip = (struct xfs_inode_log_item *)lip;
2537 ip = iip->ili_inode;
2538
2539 /*
2540 * Quick and dirty check to avoid locks if possible.
2541 */
2542 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2543 continue;
2544 if (xfs_ipincount(ip))
2545 continue;
2546
2547 /*
2548 * The inode is still attached to the buffer, which means it is
2549 * dirty but reclaim might try to grab it. Check carefully for
2550 * that, and grab the ilock while still holding the i_flags_lock
2551 * to guarantee reclaim will not be able to reclaim this inode
2552 * once we drop the i_flags_lock.
2553 */
2554 spin_lock(&ip->i_flags_lock);
2555 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2556 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2557 spin_unlock(&ip->i_flags_lock);
2558 continue;
2559 }
2560
2561 /*
2562 * ILOCK will pin the inode against reclaim and prevent
2563 * concurrent transactions modifying the inode while we are
2564 * flushing the inode. If we get the lock, set the flushing
2565 * state before we drop the i_flags_lock.
2566 */
2567 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2568 spin_unlock(&ip->i_flags_lock);
2569 continue;
2570 }
2571 __xfs_iflags_set(ip, XFS_IFLUSHING);
2572 spin_unlock(&ip->i_flags_lock);
2573
2574 /*
2575 * Abort flushing this inode if we are shut down because the
2576 * inode may not currently be in the AIL. This can occur when
2577 * log I/O failure unpins the inode without inserting into the
2578 * AIL, leaving a dirty/unpinned inode attached to the buffer
2579 * that otherwise looks like it should be flushed.
2580 */
2581 if (xlog_is_shutdown(mp->m_log)) {
2582 xfs_iunpin_wait(ip);
2583 xfs_iflush_abort(ip);
2584 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2585 error = -EIO;
2586 continue;
2587 }
2588
2589 /* don't block waiting on a log force to unpin dirty inodes */
2590 if (xfs_ipincount(ip)) {
2591 xfs_iflags_clear(ip, XFS_IFLUSHING);
2592 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2593 continue;
2594 }
2595
2596 if (!xfs_inode_clean(ip))
2597 error = xfs_iflush(ip, bp);
2598 else
2599 xfs_iflags_clear(ip, XFS_IFLUSHING);
2600 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2601 if (error)
2602 break;
2603 clcount++;
2604 }
2605
2606 if (error) {
2607 /*
2608 * Shutdown first so we kill the log before we release this
2609 * buffer. If it is an INODE_ALLOC buffer and pins the tail
2610 * of the log, failing it before the _log_ is shut down can
2611 * result in the log tail being moved forward in the journal
2612 * on disk because log writes can still be taking place. Hence
2613 * unpinning the tail will allow the ICREATE intent to be
2614 * removed from the log an recovery will fail with uninitialised
2615 * inode cluster buffers.
2616 */
2617 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2618 bp->b_flags |= XBF_ASYNC;
2619 xfs_buf_ioend_fail(bp);
2620 return error;
2621 }
2622
2623 if (!clcount)
2624 return -EAGAIN;
2625
2626 XFS_STATS_INC(mp, xs_icluster_flushcnt);
2627 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2628 return 0;
2629
2630}
2631
2632/* Release an inode. */
2633void
2634xfs_irele(
2635 struct xfs_inode *ip)
2636{
2637 trace_xfs_irele(ip, _RET_IP_);
2638 iput(VFS_I(ip));
2639}
2640
2641/*
2642 * Ensure all commited transactions touching the inode are written to the log.
2643 */
2644int
2645xfs_log_force_inode(
2646 struct xfs_inode *ip)
2647{
2648 xfs_csn_t seq = 0;
2649
2650 xfs_ilock(ip, XFS_ILOCK_SHARED);
2651 if (xfs_ipincount(ip))
2652 seq = ip->i_itemp->ili_commit_seq;
2653 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2654
2655 if (!seq)
2656 return 0;
2657 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2658}
2659
2660/*
2661 * Grab the exclusive iolock for a data copy from src to dest, making sure to
2662 * abide vfs locking order (lowest pointer value goes first) and breaking the
2663 * layout leases before proceeding. The loop is needed because we cannot call
2664 * the blocking break_layout() with the iolocks held, and therefore have to
2665 * back out both locks.
2666 */
2667static int
2668xfs_iolock_two_inodes_and_break_layout(
2669 struct inode *src,
2670 struct inode *dest)
2671{
2672 int error;
2673
2674 if (src > dest)
2675 swap(src, dest);
2676
2677retry:
2678 /* Wait to break both inodes' layouts before we start locking. */
2679 error = break_layout(src, true);
2680 if (error)
2681 return error;
2682 if (src != dest) {
2683 error = break_layout(dest, true);
2684 if (error)
2685 return error;
2686 }
2687
2688 /* Lock one inode and make sure nobody got in and leased it. */
2689 inode_lock(src);
2690 error = break_layout(src, false);
2691 if (error) {
2692 inode_unlock(src);
2693 if (error == -EWOULDBLOCK)
2694 goto retry;
2695 return error;
2696 }
2697
2698 if (src == dest)
2699 return 0;
2700
2701 /* Lock the other inode and make sure nobody got in and leased it. */
2702 inode_lock_nested(dest, I_MUTEX_NONDIR2);
2703 error = break_layout(dest, false);
2704 if (error) {
2705 inode_unlock(src);
2706 inode_unlock(dest);
2707 if (error == -EWOULDBLOCK)
2708 goto retry;
2709 return error;
2710 }
2711
2712 return 0;
2713}
2714
2715static int
2716xfs_mmaplock_two_inodes_and_break_dax_layout(
2717 struct xfs_inode *ip1,
2718 struct xfs_inode *ip2)
2719{
2720 int error;
2721 bool retry;
2722 struct page *page;
2723
2724 if (ip1->i_ino > ip2->i_ino)
2725 swap(ip1, ip2);
2726
2727again:
2728 retry = false;
2729 /* Lock the first inode */
2730 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
2731 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
2732 if (error || retry) {
2733 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2734 if (error == 0 && retry)
2735 goto again;
2736 return error;
2737 }
2738
2739 if (ip1 == ip2)
2740 return 0;
2741
2742 /* Nested lock the second inode */
2743 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
2744 /*
2745 * We cannot use xfs_break_dax_layouts() directly here because it may
2746 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2747 * for this nested lock case.
2748 */
2749 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
2750 if (page && page_ref_count(page) != 1) {
2751 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2752 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2753 goto again;
2754 }
2755
2756 return 0;
2757}
2758
2759/*
2760 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2761 * mmap activity.
2762 */
2763int
2764xfs_ilock2_io_mmap(
2765 struct xfs_inode *ip1,
2766 struct xfs_inode *ip2)
2767{
2768 int ret;
2769
2770 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
2771 if (ret)
2772 return ret;
2773
2774 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2775 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2776 if (ret) {
2777 inode_unlock(VFS_I(ip2));
2778 if (ip1 != ip2)
2779 inode_unlock(VFS_I(ip1));
2780 return ret;
2781 }
2782 } else
2783 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
2784 VFS_I(ip2)->i_mapping);
2785
2786 return 0;
2787}
2788
2789/* Unlock both inodes to allow IO and mmap activity. */
2790void
2791xfs_iunlock2_io_mmap(
2792 struct xfs_inode *ip1,
2793 struct xfs_inode *ip2)
2794{
2795 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2796 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2797 if (ip1 != ip2)
2798 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2799 } else
2800 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
2801 VFS_I(ip2)->i_mapping);
2802
2803 inode_unlock(VFS_I(ip2));
2804 if (ip1 != ip2)
2805 inode_unlock(VFS_I(ip1));
2806}
2807
2808/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2809void
2810xfs_iunlock2_remapping(
2811 struct xfs_inode *ip1,
2812 struct xfs_inode *ip2)
2813{
2814 xfs_iflags_clear(ip1, XFS_IREMAPPING);
2815
2816 if (ip1 != ip2)
2817 xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
2818 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2819
2820 if (ip1 != ip2)
2821 inode_unlock_shared(VFS_I(ip1));
2822 inode_unlock(VFS_I(ip2));
2823}
2824
2825/*
2826 * Reload the incore inode list for this inode. Caller should ensure that
2827 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
2828 * preventing other threads from executing.
2829 */
2830int
2831xfs_inode_reload_unlinked_bucket(
2832 struct xfs_trans *tp,
2833 struct xfs_inode *ip)
2834{
2835 struct xfs_mount *mp = tp->t_mountp;
2836 struct xfs_buf *agibp;
2837 struct xfs_agi *agi;
2838 struct xfs_perag *pag;
2839 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2840 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2841 xfs_agino_t prev_agino, next_agino;
2842 unsigned int bucket;
2843 bool foundit = false;
2844 int error;
2845
2846 /* Grab the first inode in the list */
2847 pag = xfs_perag_get(mp, agno);
2848 error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2849 xfs_perag_put(pag);
2850 if (error)
2851 return error;
2852
2853 /*
2854 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2855 * incore unlinked list pointers for this inode. Check once more to
2856 * see if we raced with anyone else to reload the unlinked list.
2857 */
2858 if (!xfs_inode_unlinked_incomplete(ip)) {
2859 foundit = true;
2860 goto out_agibp;
2861 }
2862
2863 bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2864 agi = agibp->b_addr;
2865
2866 trace_xfs_inode_reload_unlinked_bucket(ip);
2867
2868 xfs_info_ratelimited(mp,
2869 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
2870 agino, agno);
2871
2872 prev_agino = NULLAGINO;
2873 next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2874 while (next_agino != NULLAGINO) {
2875 struct xfs_inode *next_ip = NULL;
2876
2877 /* Found this caller's inode, set its backlink. */
2878 if (next_agino == agino) {
2879 next_ip = ip;
2880 next_ip->i_prev_unlinked = prev_agino;
2881 foundit = true;
2882 goto next_inode;
2883 }
2884
2885 /* Try in-memory lookup first. */
2886 next_ip = xfs_iunlink_lookup(pag, next_agino);
2887 if (next_ip)
2888 goto next_inode;
2889
2890 /* Inode not in memory, try reloading it. */
2891 error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2892 next_agino);
2893 if (error)
2894 break;
2895
2896 /* Grab the reloaded inode. */
2897 next_ip = xfs_iunlink_lookup(pag, next_agino);
2898 if (!next_ip) {
2899 /* No incore inode at all? We reloaded it... */
2900 ASSERT(next_ip != NULL);
2901 error = -EFSCORRUPTED;
2902 break;
2903 }
2904
2905next_inode:
2906 prev_agino = next_agino;
2907 next_agino = next_ip->i_next_unlinked;
2908 }
2909
2910out_agibp:
2911 xfs_trans_brelse(tp, agibp);
2912 /* Should have found this inode somewhere in the iunlinked bucket. */
2913 if (!error && !foundit)
2914 error = -EFSCORRUPTED;
2915 return error;
2916}
2917
2918/* Decide if this inode is missing its unlinked list and reload it. */
2919int
2920xfs_inode_reload_unlinked(
2921 struct xfs_inode *ip)
2922{
2923 struct xfs_trans *tp;
2924 int error;
2925
2926 error = xfs_trans_alloc_empty(ip->i_mount, &tp);
2927 if (error)
2928 return error;
2929
2930 xfs_ilock(ip, XFS_ILOCK_SHARED);
2931 if (xfs_inode_unlinked_incomplete(ip))
2932 error = xfs_inode_reload_unlinked_bucket(tp, ip);
2933 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2934 xfs_trans_cancel(tp);
2935
2936 return error;
2937}
2938
2939/* Has this inode fork been zapped by repair? */
2940bool
2941xfs_ifork_zapped(
2942 const struct xfs_inode *ip,
2943 int whichfork)
2944{
2945 unsigned int datamask = 0;
2946
2947 switch (whichfork) {
2948 case XFS_DATA_FORK:
2949 switch (ip->i_vnode.i_mode & S_IFMT) {
2950 case S_IFDIR:
2951 datamask = XFS_SICK_INO_DIR_ZAPPED;
2952 break;
2953 case S_IFLNK:
2954 datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
2955 break;
2956 }
2957 return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
2958 case XFS_ATTR_FORK:
2959 return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
2960 default:
2961 return false;
2962 }
2963}
2964
2965/* Compute the number of data and realtime blocks used by a file. */
2966void
2967xfs_inode_count_blocks(
2968 struct xfs_trans *tp,
2969 struct xfs_inode *ip,
2970 xfs_filblks_t *dblocks,
2971 xfs_filblks_t *rblocks)
2972{
2973 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
2974
2975 *rblocks = 0;
2976 if (XFS_IS_REALTIME_INODE(ip))
2977 xfs_bmap_count_leaves(ifp, rblocks);
2978 *dblocks = ip->i_nblocks - *rblocks;
2979}
2980
2981static void
2982xfs_wait_dax_page(
2983 struct inode *inode)
2984{
2985 struct xfs_inode *ip = XFS_I(inode);
2986
2987 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
2988 schedule();
2989 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
2990}
2991
2992int
2993xfs_break_dax_layouts(
2994 struct inode *inode,
2995 bool *retry)
2996{
2997 struct page *page;
2998
2999 xfs_assert_ilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL);
3000
3001 page = dax_layout_busy_page(inode->i_mapping);
3002 if (!page)
3003 return 0;
3004
3005 *retry = true;
3006 return ___wait_var_event(&page->_refcount,
3007 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
3008 0, 0, xfs_wait_dax_page(inode));
3009}
3010
3011int
3012xfs_break_layouts(
3013 struct inode *inode,
3014 uint *iolock,
3015 enum layout_break_reason reason)
3016{
3017 bool retry;
3018 int error;
3019
3020 xfs_assert_ilocked(XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3021
3022 do {
3023 retry = false;
3024 switch (reason) {
3025 case BREAK_UNMAP:
3026 error = xfs_break_dax_layouts(inode, &retry);
3027 if (error || retry)
3028 break;
3029 fallthrough;
3030 case BREAK_WRITE:
3031 error = xfs_break_leased_layouts(inode, iolock, &retry);
3032 break;
3033 default:
3034 WARN_ON_ONCE(1);
3035 error = -EINVAL;
3036 }
3037 } while (error == 0 && retry);
3038
3039 return error;
3040}
3041
3042/* Returns the size of fundamental allocation unit for a file, in bytes. */
3043unsigned int
3044xfs_inode_alloc_unitsize(
3045 struct xfs_inode *ip)
3046{
3047 unsigned int blocks = 1;
3048
3049 if (XFS_IS_REALTIME_INODE(ip))
3050 blocks = ip->i_mount->m_sb.sb_rextsize;
3051
3052 return XFS_FSB_TO_B(ip->i_mount, blocks);
3053}
3054
3055/* Should we always be using copy on write for file writes? */
3056bool
3057xfs_is_always_cow_inode(
3058 const struct xfs_inode *ip)
3059{
3060 return ip->i_mount->m_always_cow && xfs_has_reflink(ip->i_mount);
3061}
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_sb.h"
15#include "xfs_mount.h"
16#include "xfs_defer.h"
17#include "xfs_inode.h"
18#include "xfs_dir2.h"
19#include "xfs_attr.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_ialloc.h"
25#include "xfs_bmap.h"
26#include "xfs_bmap_util.h"
27#include "xfs_errortag.h"
28#include "xfs_error.h"
29#include "xfs_quota.h"
30#include "xfs_filestream.h"
31#include "xfs_trace.h"
32#include "xfs_icache.h"
33#include "xfs_symlink.h"
34#include "xfs_trans_priv.h"
35#include "xfs_log.h"
36#include "xfs_bmap_btree.h"
37#include "xfs_reflink.h"
38
39kmem_zone_t *xfs_inode_zone;
40
41/*
42 * Used in xfs_itruncate_extents(). This is the maximum number of extents
43 * freed from a file in a single transaction.
44 */
45#define XFS_ITRUNC_MAX_EXTENTS 2
46
47STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
48STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
49
50/*
51 * helper function to extract extent size hint from inode
52 */
53xfs_extlen_t
54xfs_get_extsz_hint(
55 struct xfs_inode *ip)
56{
57 /*
58 * No point in aligning allocations if we need to COW to actually
59 * write to them.
60 */
61 if (xfs_is_always_cow_inode(ip))
62 return 0;
63 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
64 return ip->i_d.di_extsize;
65 if (XFS_IS_REALTIME_INODE(ip))
66 return ip->i_mount->m_sb.sb_rextsize;
67 return 0;
68}
69
70/*
71 * Helper function to extract CoW extent size hint from inode.
72 * Between the extent size hint and the CoW extent size hint, we
73 * return the greater of the two. If the value is zero (automatic),
74 * use the default size.
75 */
76xfs_extlen_t
77xfs_get_cowextsz_hint(
78 struct xfs_inode *ip)
79{
80 xfs_extlen_t a, b;
81
82 a = 0;
83 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
84 a = ip->i_d.di_cowextsize;
85 b = xfs_get_extsz_hint(ip);
86
87 a = max(a, b);
88 if (a == 0)
89 return XFS_DEFAULT_COWEXTSZ_HINT;
90 return a;
91}
92
93/*
94 * These two are wrapper routines around the xfs_ilock() routine used to
95 * centralize some grungy code. They are used in places that wish to lock the
96 * inode solely for reading the extents. The reason these places can't just
97 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
98 * bringing in of the extents from disk for a file in b-tree format. If the
99 * inode is in b-tree format, then we need to lock the inode exclusively until
100 * the extents are read in. Locking it exclusively all the time would limit
101 * our parallelism unnecessarily, though. What we do instead is check to see
102 * if the extents have been read in yet, and only lock the inode exclusively
103 * if they have not.
104 *
105 * The functions return a value which should be given to the corresponding
106 * xfs_iunlock() call.
107 */
108uint
109xfs_ilock_data_map_shared(
110 struct xfs_inode *ip)
111{
112 uint lock_mode = XFS_ILOCK_SHARED;
113
114 if (ip->i_df.if_format == XFS_DINODE_FMT_BTREE &&
115 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
116 lock_mode = XFS_ILOCK_EXCL;
117 xfs_ilock(ip, lock_mode);
118 return lock_mode;
119}
120
121uint
122xfs_ilock_attr_map_shared(
123 struct xfs_inode *ip)
124{
125 uint lock_mode = XFS_ILOCK_SHARED;
126
127 if (ip->i_afp &&
128 ip->i_afp->if_format == XFS_DINODE_FMT_BTREE &&
129 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
130 lock_mode = XFS_ILOCK_EXCL;
131 xfs_ilock(ip, lock_mode);
132 return lock_mode;
133}
134
135/*
136 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
137 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
138 * various combinations of the locks to be obtained.
139 *
140 * The 3 locks should always be ordered so that the IO lock is obtained first,
141 * the mmap lock second and the ilock last in order to prevent deadlock.
142 *
143 * Basic locking order:
144 *
145 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
146 *
147 * mmap_lock locking order:
148 *
149 * i_rwsem -> page lock -> mmap_lock
150 * mmap_lock -> i_mmap_lock -> page_lock
151 *
152 * The difference in mmap_lock locking order mean that we cannot hold the
153 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
154 * fault in pages during copy in/out (for buffered IO) or require the mmap_lock
155 * in get_user_pages() to map the user pages into the kernel address space for
156 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
157 * page faults already hold the mmap_lock.
158 *
159 * Hence to serialise fully against both syscall and mmap based IO, we need to
160 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
161 * taken in places where we need to invalidate the page cache in a race
162 * free manner (e.g. truncate, hole punch and other extent manipulation
163 * functions).
164 */
165void
166xfs_ilock(
167 xfs_inode_t *ip,
168 uint lock_flags)
169{
170 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
171
172 /*
173 * You can't set both SHARED and EXCL for the same lock,
174 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
175 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
176 */
177 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
178 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
179 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
180 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
181 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
182 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
183 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
184
185 if (lock_flags & XFS_IOLOCK_EXCL) {
186 down_write_nested(&VFS_I(ip)->i_rwsem,
187 XFS_IOLOCK_DEP(lock_flags));
188 } else if (lock_flags & XFS_IOLOCK_SHARED) {
189 down_read_nested(&VFS_I(ip)->i_rwsem,
190 XFS_IOLOCK_DEP(lock_flags));
191 }
192
193 if (lock_flags & XFS_MMAPLOCK_EXCL)
194 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
195 else if (lock_flags & XFS_MMAPLOCK_SHARED)
196 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
197
198 if (lock_flags & XFS_ILOCK_EXCL)
199 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
200 else if (lock_flags & XFS_ILOCK_SHARED)
201 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
202}
203
204/*
205 * This is just like xfs_ilock(), except that the caller
206 * is guaranteed not to sleep. It returns 1 if it gets
207 * the requested locks and 0 otherwise. If the IO lock is
208 * obtained but the inode lock cannot be, then the IO lock
209 * is dropped before returning.
210 *
211 * ip -- the inode being locked
212 * lock_flags -- this parameter indicates the inode's locks to be
213 * to be locked. See the comment for xfs_ilock() for a list
214 * of valid values.
215 */
216int
217xfs_ilock_nowait(
218 xfs_inode_t *ip,
219 uint lock_flags)
220{
221 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
222
223 /*
224 * You can't set both SHARED and EXCL for the same lock,
225 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
226 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
227 */
228 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
229 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
230 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
231 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
232 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
233 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
234 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
235
236 if (lock_flags & XFS_IOLOCK_EXCL) {
237 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
238 goto out;
239 } else if (lock_flags & XFS_IOLOCK_SHARED) {
240 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
241 goto out;
242 }
243
244 if (lock_flags & XFS_MMAPLOCK_EXCL) {
245 if (!mrtryupdate(&ip->i_mmaplock))
246 goto out_undo_iolock;
247 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
248 if (!mrtryaccess(&ip->i_mmaplock))
249 goto out_undo_iolock;
250 }
251
252 if (lock_flags & XFS_ILOCK_EXCL) {
253 if (!mrtryupdate(&ip->i_lock))
254 goto out_undo_mmaplock;
255 } else if (lock_flags & XFS_ILOCK_SHARED) {
256 if (!mrtryaccess(&ip->i_lock))
257 goto out_undo_mmaplock;
258 }
259 return 1;
260
261out_undo_mmaplock:
262 if (lock_flags & XFS_MMAPLOCK_EXCL)
263 mrunlock_excl(&ip->i_mmaplock);
264 else if (lock_flags & XFS_MMAPLOCK_SHARED)
265 mrunlock_shared(&ip->i_mmaplock);
266out_undo_iolock:
267 if (lock_flags & XFS_IOLOCK_EXCL)
268 up_write(&VFS_I(ip)->i_rwsem);
269 else if (lock_flags & XFS_IOLOCK_SHARED)
270 up_read(&VFS_I(ip)->i_rwsem);
271out:
272 return 0;
273}
274
275/*
276 * xfs_iunlock() is used to drop the inode locks acquired with
277 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
278 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
279 * that we know which locks to drop.
280 *
281 * ip -- the inode being unlocked
282 * lock_flags -- this parameter indicates the inode's locks to be
283 * to be unlocked. See the comment for xfs_ilock() for a list
284 * of valid values for this parameter.
285 *
286 */
287void
288xfs_iunlock(
289 xfs_inode_t *ip,
290 uint lock_flags)
291{
292 /*
293 * You can't set both SHARED and EXCL for the same lock,
294 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
295 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
296 */
297 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
298 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
299 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
300 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
301 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
302 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
303 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
304 ASSERT(lock_flags != 0);
305
306 if (lock_flags & XFS_IOLOCK_EXCL)
307 up_write(&VFS_I(ip)->i_rwsem);
308 else if (lock_flags & XFS_IOLOCK_SHARED)
309 up_read(&VFS_I(ip)->i_rwsem);
310
311 if (lock_flags & XFS_MMAPLOCK_EXCL)
312 mrunlock_excl(&ip->i_mmaplock);
313 else if (lock_flags & XFS_MMAPLOCK_SHARED)
314 mrunlock_shared(&ip->i_mmaplock);
315
316 if (lock_flags & XFS_ILOCK_EXCL)
317 mrunlock_excl(&ip->i_lock);
318 else if (lock_flags & XFS_ILOCK_SHARED)
319 mrunlock_shared(&ip->i_lock);
320
321 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
322}
323
324/*
325 * give up write locks. the i/o lock cannot be held nested
326 * if it is being demoted.
327 */
328void
329xfs_ilock_demote(
330 xfs_inode_t *ip,
331 uint lock_flags)
332{
333 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
334 ASSERT((lock_flags &
335 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
336
337 if (lock_flags & XFS_ILOCK_EXCL)
338 mrdemote(&ip->i_lock);
339 if (lock_flags & XFS_MMAPLOCK_EXCL)
340 mrdemote(&ip->i_mmaplock);
341 if (lock_flags & XFS_IOLOCK_EXCL)
342 downgrade_write(&VFS_I(ip)->i_rwsem);
343
344 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
345}
346
347#if defined(DEBUG) || defined(XFS_WARN)
348int
349xfs_isilocked(
350 xfs_inode_t *ip,
351 uint lock_flags)
352{
353 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
354 if (!(lock_flags & XFS_ILOCK_SHARED))
355 return !!ip->i_lock.mr_writer;
356 return rwsem_is_locked(&ip->i_lock.mr_lock);
357 }
358
359 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
360 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
361 return !!ip->i_mmaplock.mr_writer;
362 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
363 }
364
365 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
366 if (!(lock_flags & XFS_IOLOCK_SHARED))
367 return !debug_locks ||
368 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
369 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
370 }
371
372 ASSERT(0);
373 return 0;
374}
375#endif
376
377/*
378 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
379 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
380 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
381 * errors and warnings.
382 */
383#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
384static bool
385xfs_lockdep_subclass_ok(
386 int subclass)
387{
388 return subclass < MAX_LOCKDEP_SUBCLASSES;
389}
390#else
391#define xfs_lockdep_subclass_ok(subclass) (true)
392#endif
393
394/*
395 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
396 * value. This can be called for any type of inode lock combination, including
397 * parent locking. Care must be taken to ensure we don't overrun the subclass
398 * storage fields in the class mask we build.
399 */
400static inline int
401xfs_lock_inumorder(int lock_mode, int subclass)
402{
403 int class = 0;
404
405 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
406 XFS_ILOCK_RTSUM)));
407 ASSERT(xfs_lockdep_subclass_ok(subclass));
408
409 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
410 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
411 class += subclass << XFS_IOLOCK_SHIFT;
412 }
413
414 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
415 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
416 class += subclass << XFS_MMAPLOCK_SHIFT;
417 }
418
419 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
420 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
421 class += subclass << XFS_ILOCK_SHIFT;
422 }
423
424 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
425}
426
427/*
428 * The following routine will lock n inodes in exclusive mode. We assume the
429 * caller calls us with the inodes in i_ino order.
430 *
431 * We need to detect deadlock where an inode that we lock is in the AIL and we
432 * start waiting for another inode that is locked by a thread in a long running
433 * transaction (such as truncate). This can result in deadlock since the long
434 * running trans might need to wait for the inode we just locked in order to
435 * push the tail and free space in the log.
436 *
437 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
438 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
439 * lock more than one at a time, lockdep will report false positives saying we
440 * have violated locking orders.
441 */
442static void
443xfs_lock_inodes(
444 struct xfs_inode **ips,
445 int inodes,
446 uint lock_mode)
447{
448 int attempts = 0, i, j, try_lock;
449 struct xfs_log_item *lp;
450
451 /*
452 * Currently supports between 2 and 5 inodes with exclusive locking. We
453 * support an arbitrary depth of locking here, but absolute limits on
454 * inodes depend on the type of locking and the limits placed by
455 * lockdep annotations in xfs_lock_inumorder. These are all checked by
456 * the asserts.
457 */
458 ASSERT(ips && inodes >= 2 && inodes <= 5);
459 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
460 XFS_ILOCK_EXCL));
461 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
462 XFS_ILOCK_SHARED)));
463 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
464 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
465 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
466 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
467
468 if (lock_mode & XFS_IOLOCK_EXCL) {
469 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
470 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
471 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
472
473 try_lock = 0;
474 i = 0;
475again:
476 for (; i < inodes; i++) {
477 ASSERT(ips[i]);
478
479 if (i && (ips[i] == ips[i - 1])) /* Already locked */
480 continue;
481
482 /*
483 * If try_lock is not set yet, make sure all locked inodes are
484 * not in the AIL. If any are, set try_lock to be used later.
485 */
486 if (!try_lock) {
487 for (j = (i - 1); j >= 0 && !try_lock; j--) {
488 lp = &ips[j]->i_itemp->ili_item;
489 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
490 try_lock++;
491 }
492 }
493
494 /*
495 * If any of the previous locks we have locked is in the AIL,
496 * we must TRY to get the second and subsequent locks. If
497 * we can't get any, we must release all we have
498 * and try again.
499 */
500 if (!try_lock) {
501 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
502 continue;
503 }
504
505 /* try_lock means we have an inode locked that is in the AIL. */
506 ASSERT(i != 0);
507 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
508 continue;
509
510 /*
511 * Unlock all previous guys and try again. xfs_iunlock will try
512 * to push the tail if the inode is in the AIL.
513 */
514 attempts++;
515 for (j = i - 1; j >= 0; j--) {
516 /*
517 * Check to see if we've already unlocked this one. Not
518 * the first one going back, and the inode ptr is the
519 * same.
520 */
521 if (j != (i - 1) && ips[j] == ips[j + 1])
522 continue;
523
524 xfs_iunlock(ips[j], lock_mode);
525 }
526
527 if ((attempts % 5) == 0) {
528 delay(1); /* Don't just spin the CPU */
529 }
530 i = 0;
531 try_lock = 0;
532 goto again;
533 }
534}
535
536/*
537 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
538 * the mmaplock or the ilock, but not more than one type at a time. If we lock
539 * more than one at a time, lockdep will report false positives saying we have
540 * violated locking orders. The iolock must be double-locked separately since
541 * we use i_rwsem for that. We now support taking one lock EXCL and the other
542 * SHARED.
543 */
544void
545xfs_lock_two_inodes(
546 struct xfs_inode *ip0,
547 uint ip0_mode,
548 struct xfs_inode *ip1,
549 uint ip1_mode)
550{
551 struct xfs_inode *temp;
552 uint mode_temp;
553 int attempts = 0;
554 struct xfs_log_item *lp;
555
556 ASSERT(hweight32(ip0_mode) == 1);
557 ASSERT(hweight32(ip1_mode) == 1);
558 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
559 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
560 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
561 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
562 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
563 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
564 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
565 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
566 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
567 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
568
569 ASSERT(ip0->i_ino != ip1->i_ino);
570
571 if (ip0->i_ino > ip1->i_ino) {
572 temp = ip0;
573 ip0 = ip1;
574 ip1 = temp;
575 mode_temp = ip0_mode;
576 ip0_mode = ip1_mode;
577 ip1_mode = mode_temp;
578 }
579
580 again:
581 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
582
583 /*
584 * If the first lock we have locked is in the AIL, we must TRY to get
585 * the second lock. If we can't get it, we must release the first one
586 * and try again.
587 */
588 lp = &ip0->i_itemp->ili_item;
589 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
590 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
591 xfs_iunlock(ip0, ip0_mode);
592 if ((++attempts % 5) == 0)
593 delay(1); /* Don't just spin the CPU */
594 goto again;
595 }
596 } else {
597 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
598 }
599}
600
601void
602__xfs_iflock(
603 struct xfs_inode *ip)
604{
605 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
606 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
607
608 do {
609 prepare_to_wait_exclusive(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
610 if (xfs_isiflocked(ip))
611 io_schedule();
612 } while (!xfs_iflock_nowait(ip));
613
614 finish_wait(wq, &wait.wq_entry);
615}
616
617STATIC uint
618_xfs_dic2xflags(
619 uint16_t di_flags,
620 uint64_t di_flags2,
621 bool has_attr)
622{
623 uint flags = 0;
624
625 if (di_flags & XFS_DIFLAG_ANY) {
626 if (di_flags & XFS_DIFLAG_REALTIME)
627 flags |= FS_XFLAG_REALTIME;
628 if (di_flags & XFS_DIFLAG_PREALLOC)
629 flags |= FS_XFLAG_PREALLOC;
630 if (di_flags & XFS_DIFLAG_IMMUTABLE)
631 flags |= FS_XFLAG_IMMUTABLE;
632 if (di_flags & XFS_DIFLAG_APPEND)
633 flags |= FS_XFLAG_APPEND;
634 if (di_flags & XFS_DIFLAG_SYNC)
635 flags |= FS_XFLAG_SYNC;
636 if (di_flags & XFS_DIFLAG_NOATIME)
637 flags |= FS_XFLAG_NOATIME;
638 if (di_flags & XFS_DIFLAG_NODUMP)
639 flags |= FS_XFLAG_NODUMP;
640 if (di_flags & XFS_DIFLAG_RTINHERIT)
641 flags |= FS_XFLAG_RTINHERIT;
642 if (di_flags & XFS_DIFLAG_PROJINHERIT)
643 flags |= FS_XFLAG_PROJINHERIT;
644 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
645 flags |= FS_XFLAG_NOSYMLINKS;
646 if (di_flags & XFS_DIFLAG_EXTSIZE)
647 flags |= FS_XFLAG_EXTSIZE;
648 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
649 flags |= FS_XFLAG_EXTSZINHERIT;
650 if (di_flags & XFS_DIFLAG_NODEFRAG)
651 flags |= FS_XFLAG_NODEFRAG;
652 if (di_flags & XFS_DIFLAG_FILESTREAM)
653 flags |= FS_XFLAG_FILESTREAM;
654 }
655
656 if (di_flags2 & XFS_DIFLAG2_ANY) {
657 if (di_flags2 & XFS_DIFLAG2_DAX)
658 flags |= FS_XFLAG_DAX;
659 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
660 flags |= FS_XFLAG_COWEXTSIZE;
661 }
662
663 if (has_attr)
664 flags |= FS_XFLAG_HASATTR;
665
666 return flags;
667}
668
669uint
670xfs_ip2xflags(
671 struct xfs_inode *ip)
672{
673 struct xfs_icdinode *dic = &ip->i_d;
674
675 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
676}
677
678/*
679 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
680 * is allowed, otherwise it has to be an exact match. If a CI match is found,
681 * ci_name->name will point to a the actual name (caller must free) or
682 * will be set to NULL if an exact match is found.
683 */
684int
685xfs_lookup(
686 xfs_inode_t *dp,
687 struct xfs_name *name,
688 xfs_inode_t **ipp,
689 struct xfs_name *ci_name)
690{
691 xfs_ino_t inum;
692 int error;
693
694 trace_xfs_lookup(dp, name);
695
696 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
697 return -EIO;
698
699 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
700 if (error)
701 goto out_unlock;
702
703 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
704 if (error)
705 goto out_free_name;
706
707 return 0;
708
709out_free_name:
710 if (ci_name)
711 kmem_free(ci_name->name);
712out_unlock:
713 *ipp = NULL;
714 return error;
715}
716
717/*
718 * Allocate an inode on disk and return a copy of its in-core version.
719 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
720 * appropriately within the inode. The uid and gid for the inode are
721 * set according to the contents of the given cred structure.
722 *
723 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
724 * has a free inode available, call xfs_iget() to obtain the in-core
725 * version of the allocated inode. Finally, fill in the inode and
726 * log its initial contents. In this case, ialloc_context would be
727 * set to NULL.
728 *
729 * If xfs_dialloc() does not have an available inode, it will replenish
730 * its supply by doing an allocation. Since we can only do one
731 * allocation within a transaction without deadlocks, we must commit
732 * the current transaction before returning the inode itself.
733 * In this case, therefore, we will set ialloc_context and return.
734 * The caller should then commit the current transaction, start a new
735 * transaction, and call xfs_ialloc() again to actually get the inode.
736 *
737 * To ensure that some other process does not grab the inode that
738 * was allocated during the first call to xfs_ialloc(), this routine
739 * also returns the [locked] bp pointing to the head of the freelist
740 * as ialloc_context. The caller should hold this buffer across
741 * the commit and pass it back into this routine on the second call.
742 *
743 * If we are allocating quota inodes, we do not have a parent inode
744 * to attach to or associate with (i.e. pip == NULL) because they
745 * are not linked into the directory structure - they are attached
746 * directly to the superblock - and so have no parent.
747 */
748static int
749xfs_ialloc(
750 xfs_trans_t *tp,
751 xfs_inode_t *pip,
752 umode_t mode,
753 xfs_nlink_t nlink,
754 dev_t rdev,
755 prid_t prid,
756 xfs_buf_t **ialloc_context,
757 xfs_inode_t **ipp)
758{
759 struct xfs_mount *mp = tp->t_mountp;
760 xfs_ino_t ino;
761 xfs_inode_t *ip;
762 uint flags;
763 int error;
764 struct timespec64 tv;
765 struct inode *inode;
766
767 /*
768 * Call the space management code to pick
769 * the on-disk inode to be allocated.
770 */
771 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode,
772 ialloc_context, &ino);
773 if (error)
774 return error;
775 if (*ialloc_context || ino == NULLFSINO) {
776 *ipp = NULL;
777 return 0;
778 }
779 ASSERT(*ialloc_context == NULL);
780
781 /*
782 * Protect against obviously corrupt allocation btree records. Later
783 * xfs_iget checks will catch re-allocation of other active in-memory
784 * and on-disk inodes. If we don't catch reallocating the parent inode
785 * here we will deadlock in xfs_iget() so we have to do these checks
786 * first.
787 */
788 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
789 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
790 return -EFSCORRUPTED;
791 }
792
793 /*
794 * Get the in-core inode with the lock held exclusively.
795 * This is because we're setting fields here we need
796 * to prevent others from looking at until we're done.
797 */
798 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
799 XFS_ILOCK_EXCL, &ip);
800 if (error)
801 return error;
802 ASSERT(ip != NULL);
803 inode = VFS_I(ip);
804 inode->i_mode = mode;
805 set_nlink(inode, nlink);
806 inode->i_uid = current_fsuid();
807 inode->i_rdev = rdev;
808 ip->i_d.di_projid = prid;
809
810 if (pip && XFS_INHERIT_GID(pip)) {
811 inode->i_gid = VFS_I(pip)->i_gid;
812 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
813 inode->i_mode |= S_ISGID;
814 } else {
815 inode->i_gid = current_fsgid();
816 }
817
818 /*
819 * If the group ID of the new file does not match the effective group
820 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
821 * (and only if the irix_sgid_inherit compatibility variable is set).
822 */
823 if (irix_sgid_inherit &&
824 (inode->i_mode & S_ISGID) && !in_group_p(inode->i_gid))
825 inode->i_mode &= ~S_ISGID;
826
827 ip->i_d.di_size = 0;
828 ip->i_df.if_nextents = 0;
829 ASSERT(ip->i_d.di_nblocks == 0);
830
831 tv = current_time(inode);
832 inode->i_mtime = tv;
833 inode->i_atime = tv;
834 inode->i_ctime = tv;
835
836 ip->i_d.di_extsize = 0;
837 ip->i_d.di_dmevmask = 0;
838 ip->i_d.di_dmstate = 0;
839 ip->i_d.di_flags = 0;
840
841 if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
842 inode_set_iversion(inode, 1);
843 ip->i_d.di_flags2 = 0;
844 ip->i_d.di_cowextsize = 0;
845 ip->i_d.di_crtime = tv;
846 }
847
848 flags = XFS_ILOG_CORE;
849 switch (mode & S_IFMT) {
850 case S_IFIFO:
851 case S_IFCHR:
852 case S_IFBLK:
853 case S_IFSOCK:
854 ip->i_df.if_format = XFS_DINODE_FMT_DEV;
855 ip->i_df.if_flags = 0;
856 flags |= XFS_ILOG_DEV;
857 break;
858 case S_IFREG:
859 case S_IFDIR:
860 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
861 uint di_flags = 0;
862
863 if (S_ISDIR(mode)) {
864 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
865 di_flags |= XFS_DIFLAG_RTINHERIT;
866 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
867 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
868 ip->i_d.di_extsize = pip->i_d.di_extsize;
869 }
870 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
871 di_flags |= XFS_DIFLAG_PROJINHERIT;
872 } else if (S_ISREG(mode)) {
873 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
874 di_flags |= XFS_DIFLAG_REALTIME;
875 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
876 di_flags |= XFS_DIFLAG_EXTSIZE;
877 ip->i_d.di_extsize = pip->i_d.di_extsize;
878 }
879 }
880 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
881 xfs_inherit_noatime)
882 di_flags |= XFS_DIFLAG_NOATIME;
883 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
884 xfs_inherit_nodump)
885 di_flags |= XFS_DIFLAG_NODUMP;
886 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
887 xfs_inherit_sync)
888 di_flags |= XFS_DIFLAG_SYNC;
889 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
890 xfs_inherit_nosymlinks)
891 di_flags |= XFS_DIFLAG_NOSYMLINKS;
892 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
893 xfs_inherit_nodefrag)
894 di_flags |= XFS_DIFLAG_NODEFRAG;
895 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
896 di_flags |= XFS_DIFLAG_FILESTREAM;
897
898 ip->i_d.di_flags |= di_flags;
899 }
900 if (pip && (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY)) {
901 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
902 ip->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
903 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
904 }
905 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
906 ip->i_d.di_flags2 |= XFS_DIFLAG2_DAX;
907 }
908 /* FALLTHROUGH */
909 case S_IFLNK:
910 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
911 ip->i_df.if_flags = XFS_IFEXTENTS;
912 ip->i_df.if_bytes = 0;
913 ip->i_df.if_u1.if_root = NULL;
914 break;
915 default:
916 ASSERT(0);
917 }
918
919 /*
920 * Log the new values stuffed into the inode.
921 */
922 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
923 xfs_trans_log_inode(tp, ip, flags);
924
925 /* now that we have an i_mode we can setup the inode structure */
926 xfs_setup_inode(ip);
927
928 *ipp = ip;
929 return 0;
930}
931
932/*
933 * Allocates a new inode from disk and return a pointer to the
934 * incore copy. This routine will internally commit the current
935 * transaction and allocate a new one if the Space Manager needed
936 * to do an allocation to replenish the inode free-list.
937 *
938 * This routine is designed to be called from xfs_create and
939 * xfs_create_dir.
940 *
941 */
942int
943xfs_dir_ialloc(
944 xfs_trans_t **tpp, /* input: current transaction;
945 output: may be a new transaction. */
946 xfs_inode_t *dp, /* directory within whose allocate
947 the inode. */
948 umode_t mode,
949 xfs_nlink_t nlink,
950 dev_t rdev,
951 prid_t prid, /* project id */
952 xfs_inode_t **ipp) /* pointer to inode; it will be
953 locked. */
954{
955 xfs_trans_t *tp;
956 xfs_inode_t *ip;
957 xfs_buf_t *ialloc_context = NULL;
958 int code;
959 void *dqinfo;
960 uint tflags;
961
962 tp = *tpp;
963 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
964
965 /*
966 * xfs_ialloc will return a pointer to an incore inode if
967 * the Space Manager has an available inode on the free
968 * list. Otherwise, it will do an allocation and replenish
969 * the freelist. Since we can only do one allocation per
970 * transaction without deadlocks, we will need to commit the
971 * current transaction and start a new one. We will then
972 * need to call xfs_ialloc again to get the inode.
973 *
974 * If xfs_ialloc did an allocation to replenish the freelist,
975 * it returns the bp containing the head of the freelist as
976 * ialloc_context. We will hold a lock on it across the
977 * transaction commit so that no other process can steal
978 * the inode(s) that we've just allocated.
979 */
980 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, &ialloc_context,
981 &ip);
982
983 /*
984 * Return an error if we were unable to allocate a new inode.
985 * This should only happen if we run out of space on disk or
986 * encounter a disk error.
987 */
988 if (code) {
989 *ipp = NULL;
990 return code;
991 }
992 if (!ialloc_context && !ip) {
993 *ipp = NULL;
994 return -ENOSPC;
995 }
996
997 /*
998 * If the AGI buffer is non-NULL, then we were unable to get an
999 * inode in one operation. We need to commit the current
1000 * transaction and call xfs_ialloc() again. It is guaranteed
1001 * to succeed the second time.
1002 */
1003 if (ialloc_context) {
1004 /*
1005 * Normally, xfs_trans_commit releases all the locks.
1006 * We call bhold to hang on to the ialloc_context across
1007 * the commit. Holding this buffer prevents any other
1008 * processes from doing any allocations in this
1009 * allocation group.
1010 */
1011 xfs_trans_bhold(tp, ialloc_context);
1012
1013 /*
1014 * We want the quota changes to be associated with the next
1015 * transaction, NOT this one. So, detach the dqinfo from this
1016 * and attach it to the next transaction.
1017 */
1018 dqinfo = NULL;
1019 tflags = 0;
1020 if (tp->t_dqinfo) {
1021 dqinfo = (void *)tp->t_dqinfo;
1022 tp->t_dqinfo = NULL;
1023 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1024 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1025 }
1026
1027 code = xfs_trans_roll(&tp);
1028
1029 /*
1030 * Re-attach the quota info that we detached from prev trx.
1031 */
1032 if (dqinfo) {
1033 tp->t_dqinfo = dqinfo;
1034 tp->t_flags |= tflags;
1035 }
1036
1037 if (code) {
1038 xfs_buf_relse(ialloc_context);
1039 *tpp = tp;
1040 *ipp = NULL;
1041 return code;
1042 }
1043 xfs_trans_bjoin(tp, ialloc_context);
1044
1045 /*
1046 * Call ialloc again. Since we've locked out all
1047 * other allocations in this allocation group,
1048 * this call should always succeed.
1049 */
1050 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1051 &ialloc_context, &ip);
1052
1053 /*
1054 * If we get an error at this point, return to the caller
1055 * so that the current transaction can be aborted.
1056 */
1057 if (code) {
1058 *tpp = tp;
1059 *ipp = NULL;
1060 return code;
1061 }
1062 ASSERT(!ialloc_context && ip);
1063
1064 }
1065
1066 *ipp = ip;
1067 *tpp = tp;
1068
1069 return 0;
1070}
1071
1072/*
1073 * Decrement the link count on an inode & log the change. If this causes the
1074 * link count to go to zero, move the inode to AGI unlinked list so that it can
1075 * be freed when the last active reference goes away via xfs_inactive().
1076 */
1077static int /* error */
1078xfs_droplink(
1079 xfs_trans_t *tp,
1080 xfs_inode_t *ip)
1081{
1082 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1083
1084 drop_nlink(VFS_I(ip));
1085 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1086
1087 if (VFS_I(ip)->i_nlink)
1088 return 0;
1089
1090 return xfs_iunlink(tp, ip);
1091}
1092
1093/*
1094 * Increment the link count on an inode & log the change.
1095 */
1096static void
1097xfs_bumplink(
1098 xfs_trans_t *tp,
1099 xfs_inode_t *ip)
1100{
1101 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1102
1103 inc_nlink(VFS_I(ip));
1104 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1105}
1106
1107int
1108xfs_create(
1109 xfs_inode_t *dp,
1110 struct xfs_name *name,
1111 umode_t mode,
1112 dev_t rdev,
1113 xfs_inode_t **ipp)
1114{
1115 int is_dir = S_ISDIR(mode);
1116 struct xfs_mount *mp = dp->i_mount;
1117 struct xfs_inode *ip = NULL;
1118 struct xfs_trans *tp = NULL;
1119 int error;
1120 bool unlock_dp_on_error = false;
1121 prid_t prid;
1122 struct xfs_dquot *udqp = NULL;
1123 struct xfs_dquot *gdqp = NULL;
1124 struct xfs_dquot *pdqp = NULL;
1125 struct xfs_trans_res *tres;
1126 uint resblks;
1127
1128 trace_xfs_create(dp, name);
1129
1130 if (XFS_FORCED_SHUTDOWN(mp))
1131 return -EIO;
1132
1133 prid = xfs_get_initial_prid(dp);
1134
1135 /*
1136 * Make sure that we have allocated dquot(s) on disk.
1137 */
1138 error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1139 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1140 &udqp, &gdqp, &pdqp);
1141 if (error)
1142 return error;
1143
1144 if (is_dir) {
1145 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1146 tres = &M_RES(mp)->tr_mkdir;
1147 } else {
1148 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1149 tres = &M_RES(mp)->tr_create;
1150 }
1151
1152 /*
1153 * Initially assume that the file does not exist and
1154 * reserve the resources for that case. If that is not
1155 * the case we'll drop the one we have and get a more
1156 * appropriate transaction later.
1157 */
1158 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1159 if (error == -ENOSPC) {
1160 /* flush outstanding delalloc blocks and retry */
1161 xfs_flush_inodes(mp);
1162 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1163 }
1164 if (error)
1165 goto out_release_inode;
1166
1167 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1168 unlock_dp_on_error = true;
1169
1170 /*
1171 * Reserve disk quota and the inode.
1172 */
1173 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1174 pdqp, resblks, 1, 0);
1175 if (error)
1176 goto out_trans_cancel;
1177
1178 /*
1179 * A newly created regular or special file just has one directory
1180 * entry pointing to them, but a directory also the "." entry
1181 * pointing to itself.
1182 */
1183 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev, prid, &ip);
1184 if (error)
1185 goto out_trans_cancel;
1186
1187 /*
1188 * Now we join the directory inode to the transaction. We do not do it
1189 * earlier because xfs_dir_ialloc might commit the previous transaction
1190 * (and release all the locks). An error from here on will result in
1191 * the transaction cancel unlocking dp so don't do it explicitly in the
1192 * error path.
1193 */
1194 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1195 unlock_dp_on_error = false;
1196
1197 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1198 resblks - XFS_IALLOC_SPACE_RES(mp));
1199 if (error) {
1200 ASSERT(error != -ENOSPC);
1201 goto out_trans_cancel;
1202 }
1203 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1204 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1205
1206 if (is_dir) {
1207 error = xfs_dir_init(tp, ip, dp);
1208 if (error)
1209 goto out_trans_cancel;
1210
1211 xfs_bumplink(tp, dp);
1212 }
1213
1214 /*
1215 * If this is a synchronous mount, make sure that the
1216 * create transaction goes to disk before returning to
1217 * the user.
1218 */
1219 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1220 xfs_trans_set_sync(tp);
1221
1222 /*
1223 * Attach the dquot(s) to the inodes and modify them incore.
1224 * These ids of the inode couldn't have changed since the new
1225 * inode has been locked ever since it was created.
1226 */
1227 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1228
1229 error = xfs_trans_commit(tp);
1230 if (error)
1231 goto out_release_inode;
1232
1233 xfs_qm_dqrele(udqp);
1234 xfs_qm_dqrele(gdqp);
1235 xfs_qm_dqrele(pdqp);
1236
1237 *ipp = ip;
1238 return 0;
1239
1240 out_trans_cancel:
1241 xfs_trans_cancel(tp);
1242 out_release_inode:
1243 /*
1244 * Wait until after the current transaction is aborted to finish the
1245 * setup of the inode and release the inode. This prevents recursive
1246 * transactions and deadlocks from xfs_inactive.
1247 */
1248 if (ip) {
1249 xfs_finish_inode_setup(ip);
1250 xfs_irele(ip);
1251 }
1252
1253 xfs_qm_dqrele(udqp);
1254 xfs_qm_dqrele(gdqp);
1255 xfs_qm_dqrele(pdqp);
1256
1257 if (unlock_dp_on_error)
1258 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1259 return error;
1260}
1261
1262int
1263xfs_create_tmpfile(
1264 struct xfs_inode *dp,
1265 umode_t mode,
1266 struct xfs_inode **ipp)
1267{
1268 struct xfs_mount *mp = dp->i_mount;
1269 struct xfs_inode *ip = NULL;
1270 struct xfs_trans *tp = NULL;
1271 int error;
1272 prid_t prid;
1273 struct xfs_dquot *udqp = NULL;
1274 struct xfs_dquot *gdqp = NULL;
1275 struct xfs_dquot *pdqp = NULL;
1276 struct xfs_trans_res *tres;
1277 uint resblks;
1278
1279 if (XFS_FORCED_SHUTDOWN(mp))
1280 return -EIO;
1281
1282 prid = xfs_get_initial_prid(dp);
1283
1284 /*
1285 * Make sure that we have allocated dquot(s) on disk.
1286 */
1287 error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1288 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1289 &udqp, &gdqp, &pdqp);
1290 if (error)
1291 return error;
1292
1293 resblks = XFS_IALLOC_SPACE_RES(mp);
1294 tres = &M_RES(mp)->tr_create_tmpfile;
1295
1296 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1297 if (error)
1298 goto out_release_inode;
1299
1300 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1301 pdqp, resblks, 1, 0);
1302 if (error)
1303 goto out_trans_cancel;
1304
1305 error = xfs_dir_ialloc(&tp, dp, mode, 0, 0, prid, &ip);
1306 if (error)
1307 goto out_trans_cancel;
1308
1309 if (mp->m_flags & XFS_MOUNT_WSYNC)
1310 xfs_trans_set_sync(tp);
1311
1312 /*
1313 * Attach the dquot(s) to the inodes and modify them incore.
1314 * These ids of the inode couldn't have changed since the new
1315 * inode has been locked ever since it was created.
1316 */
1317 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1318
1319 error = xfs_iunlink(tp, ip);
1320 if (error)
1321 goto out_trans_cancel;
1322
1323 error = xfs_trans_commit(tp);
1324 if (error)
1325 goto out_release_inode;
1326
1327 xfs_qm_dqrele(udqp);
1328 xfs_qm_dqrele(gdqp);
1329 xfs_qm_dqrele(pdqp);
1330
1331 *ipp = ip;
1332 return 0;
1333
1334 out_trans_cancel:
1335 xfs_trans_cancel(tp);
1336 out_release_inode:
1337 /*
1338 * Wait until after the current transaction is aborted to finish the
1339 * setup of the inode and release the inode. This prevents recursive
1340 * transactions and deadlocks from xfs_inactive.
1341 */
1342 if (ip) {
1343 xfs_finish_inode_setup(ip);
1344 xfs_irele(ip);
1345 }
1346
1347 xfs_qm_dqrele(udqp);
1348 xfs_qm_dqrele(gdqp);
1349 xfs_qm_dqrele(pdqp);
1350
1351 return error;
1352}
1353
1354int
1355xfs_link(
1356 xfs_inode_t *tdp,
1357 xfs_inode_t *sip,
1358 struct xfs_name *target_name)
1359{
1360 xfs_mount_t *mp = tdp->i_mount;
1361 xfs_trans_t *tp;
1362 int error;
1363 int resblks;
1364
1365 trace_xfs_link(tdp, target_name);
1366
1367 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1368
1369 if (XFS_FORCED_SHUTDOWN(mp))
1370 return -EIO;
1371
1372 error = xfs_qm_dqattach(sip);
1373 if (error)
1374 goto std_return;
1375
1376 error = xfs_qm_dqattach(tdp);
1377 if (error)
1378 goto std_return;
1379
1380 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1381 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1382 if (error == -ENOSPC) {
1383 resblks = 0;
1384 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1385 }
1386 if (error)
1387 goto std_return;
1388
1389 xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1390
1391 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1392 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1393
1394 /*
1395 * If we are using project inheritance, we only allow hard link
1396 * creation in our tree when the project IDs are the same; else
1397 * the tree quota mechanism could be circumvented.
1398 */
1399 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1400 tdp->i_d.di_projid != sip->i_d.di_projid)) {
1401 error = -EXDEV;
1402 goto error_return;
1403 }
1404
1405 if (!resblks) {
1406 error = xfs_dir_canenter(tp, tdp, target_name);
1407 if (error)
1408 goto error_return;
1409 }
1410
1411 /*
1412 * Handle initial link state of O_TMPFILE inode
1413 */
1414 if (VFS_I(sip)->i_nlink == 0) {
1415 error = xfs_iunlink_remove(tp, sip);
1416 if (error)
1417 goto error_return;
1418 }
1419
1420 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1421 resblks);
1422 if (error)
1423 goto error_return;
1424 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1425 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1426
1427 xfs_bumplink(tp, sip);
1428
1429 /*
1430 * If this is a synchronous mount, make sure that the
1431 * link transaction goes to disk before returning to
1432 * the user.
1433 */
1434 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1435 xfs_trans_set_sync(tp);
1436
1437 return xfs_trans_commit(tp);
1438
1439 error_return:
1440 xfs_trans_cancel(tp);
1441 std_return:
1442 return error;
1443}
1444
1445/* Clear the reflink flag and the cowblocks tag if possible. */
1446static void
1447xfs_itruncate_clear_reflink_flags(
1448 struct xfs_inode *ip)
1449{
1450 struct xfs_ifork *dfork;
1451 struct xfs_ifork *cfork;
1452
1453 if (!xfs_is_reflink_inode(ip))
1454 return;
1455 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1456 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1457 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1458 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1459 if (cfork->if_bytes == 0)
1460 xfs_inode_clear_cowblocks_tag(ip);
1461}
1462
1463/*
1464 * Free up the underlying blocks past new_size. The new size must be smaller
1465 * than the current size. This routine can be used both for the attribute and
1466 * data fork, and does not modify the inode size, which is left to the caller.
1467 *
1468 * The transaction passed to this routine must have made a permanent log
1469 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1470 * given transaction and start new ones, so make sure everything involved in
1471 * the transaction is tidy before calling here. Some transaction will be
1472 * returned to the caller to be committed. The incoming transaction must
1473 * already include the inode, and both inode locks must be held exclusively.
1474 * The inode must also be "held" within the transaction. On return the inode
1475 * will be "held" within the returned transaction. This routine does NOT
1476 * require any disk space to be reserved for it within the transaction.
1477 *
1478 * If we get an error, we must return with the inode locked and linked into the
1479 * current transaction. This keeps things simple for the higher level code,
1480 * because it always knows that the inode is locked and held in the transaction
1481 * that returns to it whether errors occur or not. We don't mark the inode
1482 * dirty on error so that transactions can be easily aborted if possible.
1483 */
1484int
1485xfs_itruncate_extents_flags(
1486 struct xfs_trans **tpp,
1487 struct xfs_inode *ip,
1488 int whichfork,
1489 xfs_fsize_t new_size,
1490 int flags)
1491{
1492 struct xfs_mount *mp = ip->i_mount;
1493 struct xfs_trans *tp = *tpp;
1494 xfs_fileoff_t first_unmap_block;
1495 xfs_filblks_t unmap_len;
1496 int error = 0;
1497
1498 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1499 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1500 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1501 ASSERT(new_size <= XFS_ISIZE(ip));
1502 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1503 ASSERT(ip->i_itemp != NULL);
1504 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1505 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1506
1507 trace_xfs_itruncate_extents_start(ip, new_size);
1508
1509 flags |= xfs_bmapi_aflag(whichfork);
1510
1511 /*
1512 * Since it is possible for space to become allocated beyond
1513 * the end of the file (in a crash where the space is allocated
1514 * but the inode size is not yet updated), simply remove any
1515 * blocks which show up between the new EOF and the maximum
1516 * possible file size.
1517 *
1518 * We have to free all the blocks to the bmbt maximum offset, even if
1519 * the page cache can't scale that far.
1520 */
1521 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1522 if (first_unmap_block >= XFS_MAX_FILEOFF) {
1523 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1524 return 0;
1525 }
1526
1527 unmap_len = XFS_MAX_FILEOFF - first_unmap_block + 1;
1528 while (unmap_len > 0) {
1529 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1530 error = __xfs_bunmapi(tp, ip, first_unmap_block, &unmap_len,
1531 flags, XFS_ITRUNC_MAX_EXTENTS);
1532 if (error)
1533 goto out;
1534
1535 /*
1536 * Duplicate the transaction that has the permanent
1537 * reservation and commit the old transaction.
1538 */
1539 error = xfs_defer_finish(&tp);
1540 if (error)
1541 goto out;
1542
1543 error = xfs_trans_roll_inode(&tp, ip);
1544 if (error)
1545 goto out;
1546 }
1547
1548 if (whichfork == XFS_DATA_FORK) {
1549 /* Remove all pending CoW reservations. */
1550 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1551 first_unmap_block, XFS_MAX_FILEOFF, true);
1552 if (error)
1553 goto out;
1554
1555 xfs_itruncate_clear_reflink_flags(ip);
1556 }
1557
1558 /*
1559 * Always re-log the inode so that our permanent transaction can keep
1560 * on rolling it forward in the log.
1561 */
1562 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1563
1564 trace_xfs_itruncate_extents_end(ip, new_size);
1565
1566out:
1567 *tpp = tp;
1568 return error;
1569}
1570
1571int
1572xfs_release(
1573 xfs_inode_t *ip)
1574{
1575 xfs_mount_t *mp = ip->i_mount;
1576 int error;
1577
1578 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1579 return 0;
1580
1581 /* If this is a read-only mount, don't do this (would generate I/O) */
1582 if (mp->m_flags & XFS_MOUNT_RDONLY)
1583 return 0;
1584
1585 if (!XFS_FORCED_SHUTDOWN(mp)) {
1586 int truncated;
1587
1588 /*
1589 * If we previously truncated this file and removed old data
1590 * in the process, we want to initiate "early" writeout on
1591 * the last close. This is an attempt to combat the notorious
1592 * NULL files problem which is particularly noticeable from a
1593 * truncate down, buffered (re-)write (delalloc), followed by
1594 * a crash. What we are effectively doing here is
1595 * significantly reducing the time window where we'd otherwise
1596 * be exposed to that problem.
1597 */
1598 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1599 if (truncated) {
1600 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1601 if (ip->i_delayed_blks > 0) {
1602 error = filemap_flush(VFS_I(ip)->i_mapping);
1603 if (error)
1604 return error;
1605 }
1606 }
1607 }
1608
1609 if (VFS_I(ip)->i_nlink == 0)
1610 return 0;
1611
1612 if (xfs_can_free_eofblocks(ip, false)) {
1613
1614 /*
1615 * Check if the inode is being opened, written and closed
1616 * frequently and we have delayed allocation blocks outstanding
1617 * (e.g. streaming writes from the NFS server), truncating the
1618 * blocks past EOF will cause fragmentation to occur.
1619 *
1620 * In this case don't do the truncation, but we have to be
1621 * careful how we detect this case. Blocks beyond EOF show up as
1622 * i_delayed_blks even when the inode is clean, so we need to
1623 * truncate them away first before checking for a dirty release.
1624 * Hence on the first dirty close we will still remove the
1625 * speculative allocation, but after that we will leave it in
1626 * place.
1627 */
1628 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1629 return 0;
1630 /*
1631 * If we can't get the iolock just skip truncating the blocks
1632 * past EOF because we could deadlock with the mmap_lock
1633 * otherwise. We'll get another chance to drop them once the
1634 * last reference to the inode is dropped, so we'll never leak
1635 * blocks permanently.
1636 */
1637 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1638 error = xfs_free_eofblocks(ip);
1639 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1640 if (error)
1641 return error;
1642 }
1643
1644 /* delalloc blocks after truncation means it really is dirty */
1645 if (ip->i_delayed_blks)
1646 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1647 }
1648 return 0;
1649}
1650
1651/*
1652 * xfs_inactive_truncate
1653 *
1654 * Called to perform a truncate when an inode becomes unlinked.
1655 */
1656STATIC int
1657xfs_inactive_truncate(
1658 struct xfs_inode *ip)
1659{
1660 struct xfs_mount *mp = ip->i_mount;
1661 struct xfs_trans *tp;
1662 int error;
1663
1664 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1665 if (error) {
1666 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1667 return error;
1668 }
1669 xfs_ilock(ip, XFS_ILOCK_EXCL);
1670 xfs_trans_ijoin(tp, ip, 0);
1671
1672 /*
1673 * Log the inode size first to prevent stale data exposure in the event
1674 * of a system crash before the truncate completes. See the related
1675 * comment in xfs_vn_setattr_size() for details.
1676 */
1677 ip->i_d.di_size = 0;
1678 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1679
1680 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1681 if (error)
1682 goto error_trans_cancel;
1683
1684 ASSERT(ip->i_df.if_nextents == 0);
1685
1686 error = xfs_trans_commit(tp);
1687 if (error)
1688 goto error_unlock;
1689
1690 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1691 return 0;
1692
1693error_trans_cancel:
1694 xfs_trans_cancel(tp);
1695error_unlock:
1696 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1697 return error;
1698}
1699
1700/*
1701 * xfs_inactive_ifree()
1702 *
1703 * Perform the inode free when an inode is unlinked.
1704 */
1705STATIC int
1706xfs_inactive_ifree(
1707 struct xfs_inode *ip)
1708{
1709 struct xfs_mount *mp = ip->i_mount;
1710 struct xfs_trans *tp;
1711 int error;
1712
1713 /*
1714 * We try to use a per-AG reservation for any block needed by the finobt
1715 * tree, but as the finobt feature predates the per-AG reservation
1716 * support a degraded file system might not have enough space for the
1717 * reservation at mount time. In that case try to dip into the reserved
1718 * pool and pray.
1719 *
1720 * Send a warning if the reservation does happen to fail, as the inode
1721 * now remains allocated and sits on the unlinked list until the fs is
1722 * repaired.
1723 */
1724 if (unlikely(mp->m_finobt_nores)) {
1725 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1726 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1727 &tp);
1728 } else {
1729 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1730 }
1731 if (error) {
1732 if (error == -ENOSPC) {
1733 xfs_warn_ratelimited(mp,
1734 "Failed to remove inode(s) from unlinked list. "
1735 "Please free space, unmount and run xfs_repair.");
1736 } else {
1737 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1738 }
1739 return error;
1740 }
1741
1742 /*
1743 * We do not hold the inode locked across the entire rolling transaction
1744 * here. We only need to hold it for the first transaction that
1745 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1746 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1747 * here breaks the relationship between cluster buffer invalidation and
1748 * stale inode invalidation on cluster buffer item journal commit
1749 * completion, and can result in leaving dirty stale inodes hanging
1750 * around in memory.
1751 *
1752 * We have no need for serialising this inode operation against other
1753 * operations - we freed the inode and hence reallocation is required
1754 * and that will serialise on reallocating the space the deferops need
1755 * to free. Hence we can unlock the inode on the first commit of
1756 * the transaction rather than roll it right through the deferops. This
1757 * avoids relogging the XFS_ISTALE inode.
1758 *
1759 * We check that xfs_ifree() hasn't grown an internal transaction roll
1760 * by asserting that the inode is still locked when it returns.
1761 */
1762 xfs_ilock(ip, XFS_ILOCK_EXCL);
1763 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1764
1765 error = xfs_ifree(tp, ip);
1766 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1767 if (error) {
1768 /*
1769 * If we fail to free the inode, shut down. The cancel
1770 * might do that, we need to make sure. Otherwise the
1771 * inode might be lost for a long time or forever.
1772 */
1773 if (!XFS_FORCED_SHUTDOWN(mp)) {
1774 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1775 __func__, error);
1776 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1777 }
1778 xfs_trans_cancel(tp);
1779 return error;
1780 }
1781
1782 /*
1783 * Credit the quota account(s). The inode is gone.
1784 */
1785 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1786
1787 /*
1788 * Just ignore errors at this point. There is nothing we can do except
1789 * to try to keep going. Make sure it's not a silent error.
1790 */
1791 error = xfs_trans_commit(tp);
1792 if (error)
1793 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1794 __func__, error);
1795
1796 return 0;
1797}
1798
1799/*
1800 * xfs_inactive
1801 *
1802 * This is called when the vnode reference count for the vnode
1803 * goes to zero. If the file has been unlinked, then it must
1804 * now be truncated. Also, we clear all of the read-ahead state
1805 * kept for the inode here since the file is now closed.
1806 */
1807void
1808xfs_inactive(
1809 xfs_inode_t *ip)
1810{
1811 struct xfs_mount *mp;
1812 int error;
1813 int truncate = 0;
1814
1815 /*
1816 * If the inode is already free, then there can be nothing
1817 * to clean up here.
1818 */
1819 if (VFS_I(ip)->i_mode == 0) {
1820 ASSERT(ip->i_df.if_broot_bytes == 0);
1821 return;
1822 }
1823
1824 mp = ip->i_mount;
1825 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1826
1827 /* If this is a read-only mount, don't do this (would generate I/O) */
1828 if (mp->m_flags & XFS_MOUNT_RDONLY)
1829 return;
1830
1831 /* Try to clean out the cow blocks if there are any. */
1832 if (xfs_inode_has_cow_data(ip))
1833 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1834
1835 if (VFS_I(ip)->i_nlink != 0) {
1836 /*
1837 * force is true because we are evicting an inode from the
1838 * cache. Post-eof blocks must be freed, lest we end up with
1839 * broken free space accounting.
1840 *
1841 * Note: don't bother with iolock here since lockdep complains
1842 * about acquiring it in reclaim context. We have the only
1843 * reference to the inode at this point anyways.
1844 */
1845 if (xfs_can_free_eofblocks(ip, true))
1846 xfs_free_eofblocks(ip);
1847
1848 return;
1849 }
1850
1851 if (S_ISREG(VFS_I(ip)->i_mode) &&
1852 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1853 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1854 truncate = 1;
1855
1856 error = xfs_qm_dqattach(ip);
1857 if (error)
1858 return;
1859
1860 if (S_ISLNK(VFS_I(ip)->i_mode))
1861 error = xfs_inactive_symlink(ip);
1862 else if (truncate)
1863 error = xfs_inactive_truncate(ip);
1864 if (error)
1865 return;
1866
1867 /*
1868 * If there are attributes associated with the file then blow them away
1869 * now. The code calls a routine that recursively deconstructs the
1870 * attribute fork. If also blows away the in-core attribute fork.
1871 */
1872 if (XFS_IFORK_Q(ip)) {
1873 error = xfs_attr_inactive(ip);
1874 if (error)
1875 return;
1876 }
1877
1878 ASSERT(!ip->i_afp);
1879 ASSERT(ip->i_d.di_forkoff == 0);
1880
1881 /*
1882 * Free the inode.
1883 */
1884 error = xfs_inactive_ifree(ip);
1885 if (error)
1886 return;
1887
1888 /*
1889 * Release the dquots held by inode, if any.
1890 */
1891 xfs_qm_dqdetach(ip);
1892}
1893
1894/*
1895 * In-Core Unlinked List Lookups
1896 * =============================
1897 *
1898 * Every inode is supposed to be reachable from some other piece of metadata
1899 * with the exception of the root directory. Inodes with a connection to a
1900 * file descriptor but not linked from anywhere in the on-disk directory tree
1901 * are collectively known as unlinked inodes, though the filesystem itself
1902 * maintains links to these inodes so that on-disk metadata are consistent.
1903 *
1904 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1905 * header contains a number of buckets that point to an inode, and each inode
1906 * record has a pointer to the next inode in the hash chain. This
1907 * singly-linked list causes scaling problems in the iunlink remove function
1908 * because we must walk that list to find the inode that points to the inode
1909 * being removed from the unlinked hash bucket list.
1910 *
1911 * What if we modelled the unlinked list as a collection of records capturing
1912 * "X.next_unlinked = Y" relations? If we indexed those records on Y, we'd
1913 * have a fast way to look up unlinked list predecessors, which avoids the
1914 * slow list walk. That's exactly what we do here (in-core) with a per-AG
1915 * rhashtable.
1916 *
1917 * Because this is a backref cache, we ignore operational failures since the
1918 * iunlink code can fall back to the slow bucket walk. The only errors that
1919 * should bubble out are for obviously incorrect situations.
1920 *
1921 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1922 * access or have otherwise provided for concurrency control.
1923 */
1924
1925/* Capture a "X.next_unlinked = Y" relationship. */
1926struct xfs_iunlink {
1927 struct rhash_head iu_rhash_head;
1928 xfs_agino_t iu_agino; /* X */
1929 xfs_agino_t iu_next_unlinked; /* Y */
1930};
1931
1932/* Unlinked list predecessor lookup hashtable construction */
1933static int
1934xfs_iunlink_obj_cmpfn(
1935 struct rhashtable_compare_arg *arg,
1936 const void *obj)
1937{
1938 const xfs_agino_t *key = arg->key;
1939 const struct xfs_iunlink *iu = obj;
1940
1941 if (iu->iu_next_unlinked != *key)
1942 return 1;
1943 return 0;
1944}
1945
1946static const struct rhashtable_params xfs_iunlink_hash_params = {
1947 .min_size = XFS_AGI_UNLINKED_BUCKETS,
1948 .key_len = sizeof(xfs_agino_t),
1949 .key_offset = offsetof(struct xfs_iunlink,
1950 iu_next_unlinked),
1951 .head_offset = offsetof(struct xfs_iunlink, iu_rhash_head),
1952 .automatic_shrinking = true,
1953 .obj_cmpfn = xfs_iunlink_obj_cmpfn,
1954};
1955
1956/*
1957 * Return X, where X.next_unlinked == @agino. Returns NULLAGINO if no such
1958 * relation is found.
1959 */
1960static xfs_agino_t
1961xfs_iunlink_lookup_backref(
1962 struct xfs_perag *pag,
1963 xfs_agino_t agino)
1964{
1965 struct xfs_iunlink *iu;
1966
1967 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1968 xfs_iunlink_hash_params);
1969 return iu ? iu->iu_agino : NULLAGINO;
1970}
1971
1972/*
1973 * Take ownership of an iunlink cache entry and insert it into the hash table.
1974 * If successful, the entry will be owned by the cache; if not, it is freed.
1975 * Either way, the caller does not own @iu after this call.
1976 */
1977static int
1978xfs_iunlink_insert_backref(
1979 struct xfs_perag *pag,
1980 struct xfs_iunlink *iu)
1981{
1982 int error;
1983
1984 error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1985 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1986 /*
1987 * Fail loudly if there already was an entry because that's a sign of
1988 * corruption of in-memory data. Also fail loudly if we see an error
1989 * code we didn't anticipate from the rhashtable code. Currently we
1990 * only anticipate ENOMEM.
1991 */
1992 if (error) {
1993 WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1994 kmem_free(iu);
1995 }
1996 /*
1997 * Absorb any runtime errors that aren't a result of corruption because
1998 * this is a cache and we can always fall back to bucket list scanning.
1999 */
2000 if (error != 0 && error != -EEXIST)
2001 error = 0;
2002 return error;
2003}
2004
2005/* Remember that @prev_agino.next_unlinked = @this_agino. */
2006static int
2007xfs_iunlink_add_backref(
2008 struct xfs_perag *pag,
2009 xfs_agino_t prev_agino,
2010 xfs_agino_t this_agino)
2011{
2012 struct xfs_iunlink *iu;
2013
2014 if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
2015 return 0;
2016
2017 iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
2018 iu->iu_agino = prev_agino;
2019 iu->iu_next_unlinked = this_agino;
2020
2021 return xfs_iunlink_insert_backref(pag, iu);
2022}
2023
2024/*
2025 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
2026 * If @next_unlinked is NULLAGINO, we drop the backref and exit. If there
2027 * wasn't any such entry then we don't bother.
2028 */
2029static int
2030xfs_iunlink_change_backref(
2031 struct xfs_perag *pag,
2032 xfs_agino_t agino,
2033 xfs_agino_t next_unlinked)
2034{
2035 struct xfs_iunlink *iu;
2036 int error;
2037
2038 /* Look up the old entry; if there wasn't one then exit. */
2039 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
2040 xfs_iunlink_hash_params);
2041 if (!iu)
2042 return 0;
2043
2044 /*
2045 * Remove the entry. This shouldn't ever return an error, but if we
2046 * couldn't remove the old entry we don't want to add it again to the
2047 * hash table, and if the entry disappeared on us then someone's
2048 * violated the locking rules and we need to fail loudly. Either way
2049 * we cannot remove the inode because internal state is or would have
2050 * been corrupt.
2051 */
2052 error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
2053 &iu->iu_rhash_head, xfs_iunlink_hash_params);
2054 if (error)
2055 return error;
2056
2057 /* If there is no new next entry just free our item and return. */
2058 if (next_unlinked == NULLAGINO) {
2059 kmem_free(iu);
2060 return 0;
2061 }
2062
2063 /* Update the entry and re-add it to the hash table. */
2064 iu->iu_next_unlinked = next_unlinked;
2065 return xfs_iunlink_insert_backref(pag, iu);
2066}
2067
2068/* Set up the in-core predecessor structures. */
2069int
2070xfs_iunlink_init(
2071 struct xfs_perag *pag)
2072{
2073 return rhashtable_init(&pag->pagi_unlinked_hash,
2074 &xfs_iunlink_hash_params);
2075}
2076
2077/* Free the in-core predecessor structures. */
2078static void
2079xfs_iunlink_free_item(
2080 void *ptr,
2081 void *arg)
2082{
2083 struct xfs_iunlink *iu = ptr;
2084 bool *freed_anything = arg;
2085
2086 *freed_anything = true;
2087 kmem_free(iu);
2088}
2089
2090void
2091xfs_iunlink_destroy(
2092 struct xfs_perag *pag)
2093{
2094 bool freed_anything = false;
2095
2096 rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
2097 xfs_iunlink_free_item, &freed_anything);
2098
2099 ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
2100}
2101
2102/*
2103 * Point the AGI unlinked bucket at an inode and log the results. The caller
2104 * is responsible for validating the old value.
2105 */
2106STATIC int
2107xfs_iunlink_update_bucket(
2108 struct xfs_trans *tp,
2109 xfs_agnumber_t agno,
2110 struct xfs_buf *agibp,
2111 unsigned int bucket_index,
2112 xfs_agino_t new_agino)
2113{
2114 struct xfs_agi *agi = agibp->b_addr;
2115 xfs_agino_t old_value;
2116 int offset;
2117
2118 ASSERT(xfs_verify_agino_or_null(tp->t_mountp, agno, new_agino));
2119
2120 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2121 trace_xfs_iunlink_update_bucket(tp->t_mountp, agno, bucket_index,
2122 old_value, new_agino);
2123
2124 /*
2125 * We should never find the head of the list already set to the value
2126 * passed in because either we're adding or removing ourselves from the
2127 * head of the list.
2128 */
2129 if (old_value == new_agino) {
2130 xfs_buf_mark_corrupt(agibp);
2131 return -EFSCORRUPTED;
2132 }
2133
2134 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2135 offset = offsetof(struct xfs_agi, agi_unlinked) +
2136 (sizeof(xfs_agino_t) * bucket_index);
2137 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2138 return 0;
2139}
2140
2141/* Set an on-disk inode's next_unlinked pointer. */
2142STATIC void
2143xfs_iunlink_update_dinode(
2144 struct xfs_trans *tp,
2145 xfs_agnumber_t agno,
2146 xfs_agino_t agino,
2147 struct xfs_buf *ibp,
2148 struct xfs_dinode *dip,
2149 struct xfs_imap *imap,
2150 xfs_agino_t next_agino)
2151{
2152 struct xfs_mount *mp = tp->t_mountp;
2153 int offset;
2154
2155 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2156
2157 trace_xfs_iunlink_update_dinode(mp, agno, agino,
2158 be32_to_cpu(dip->di_next_unlinked), next_agino);
2159
2160 dip->di_next_unlinked = cpu_to_be32(next_agino);
2161 offset = imap->im_boffset +
2162 offsetof(struct xfs_dinode, di_next_unlinked);
2163
2164 /* need to recalc the inode CRC if appropriate */
2165 xfs_dinode_calc_crc(mp, dip);
2166 xfs_trans_inode_buf(tp, ibp);
2167 xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2168}
2169
2170/* Set an in-core inode's unlinked pointer and return the old value. */
2171STATIC int
2172xfs_iunlink_update_inode(
2173 struct xfs_trans *tp,
2174 struct xfs_inode *ip,
2175 xfs_agnumber_t agno,
2176 xfs_agino_t next_agino,
2177 xfs_agino_t *old_next_agino)
2178{
2179 struct xfs_mount *mp = tp->t_mountp;
2180 struct xfs_dinode *dip;
2181 struct xfs_buf *ibp;
2182 xfs_agino_t old_value;
2183 int error;
2184
2185 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2186
2187 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp, 0);
2188 if (error)
2189 return error;
2190
2191 /* Make sure the old pointer isn't garbage. */
2192 old_value = be32_to_cpu(dip->di_next_unlinked);
2193 if (!xfs_verify_agino_or_null(mp, agno, old_value)) {
2194 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__, dip,
2195 sizeof(*dip), __this_address);
2196 error = -EFSCORRUPTED;
2197 goto out;
2198 }
2199
2200 /*
2201 * Since we're updating a linked list, we should never find that the
2202 * current pointer is the same as the new value, unless we're
2203 * terminating the list.
2204 */
2205 *old_next_agino = old_value;
2206 if (old_value == next_agino) {
2207 if (next_agino != NULLAGINO) {
2208 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__,
2209 dip, sizeof(*dip), __this_address);
2210 error = -EFSCORRUPTED;
2211 }
2212 goto out;
2213 }
2214
2215 /* Ok, update the new pointer. */
2216 xfs_iunlink_update_dinode(tp, agno, XFS_INO_TO_AGINO(mp, ip->i_ino),
2217 ibp, dip, &ip->i_imap, next_agino);
2218 return 0;
2219out:
2220 xfs_trans_brelse(tp, ibp);
2221 return error;
2222}
2223
2224/*
2225 * This is called when the inode's link count has gone to 0 or we are creating
2226 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2227 *
2228 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2229 * list when the inode is freed.
2230 */
2231STATIC int
2232xfs_iunlink(
2233 struct xfs_trans *tp,
2234 struct xfs_inode *ip)
2235{
2236 struct xfs_mount *mp = tp->t_mountp;
2237 struct xfs_agi *agi;
2238 struct xfs_buf *agibp;
2239 xfs_agino_t next_agino;
2240 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2241 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2242 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2243 int error;
2244
2245 ASSERT(VFS_I(ip)->i_nlink == 0);
2246 ASSERT(VFS_I(ip)->i_mode != 0);
2247 trace_xfs_iunlink(ip);
2248
2249 /* Get the agi buffer first. It ensures lock ordering on the list. */
2250 error = xfs_read_agi(mp, tp, agno, &agibp);
2251 if (error)
2252 return error;
2253 agi = agibp->b_addr;
2254
2255 /*
2256 * Get the index into the agi hash table for the list this inode will
2257 * go on. Make sure the pointer isn't garbage and that this inode
2258 * isn't already on the list.
2259 */
2260 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2261 if (next_agino == agino ||
2262 !xfs_verify_agino_or_null(mp, agno, next_agino)) {
2263 xfs_buf_mark_corrupt(agibp);
2264 return -EFSCORRUPTED;
2265 }
2266
2267 if (next_agino != NULLAGINO) {
2268 xfs_agino_t old_agino;
2269
2270 /*
2271 * There is already another inode in the bucket, so point this
2272 * inode to the current head of the list.
2273 */
2274 error = xfs_iunlink_update_inode(tp, ip, agno, next_agino,
2275 &old_agino);
2276 if (error)
2277 return error;
2278 ASSERT(old_agino == NULLAGINO);
2279
2280 /*
2281 * agino has been unlinked, add a backref from the next inode
2282 * back to agino.
2283 */
2284 error = xfs_iunlink_add_backref(agibp->b_pag, agino, next_agino);
2285 if (error)
2286 return error;
2287 }
2288
2289 /* Point the head of the list to point to this inode. */
2290 return xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index, agino);
2291}
2292
2293/* Return the imap, dinode pointer, and buffer for an inode. */
2294STATIC int
2295xfs_iunlink_map_ino(
2296 struct xfs_trans *tp,
2297 xfs_agnumber_t agno,
2298 xfs_agino_t agino,
2299 struct xfs_imap *imap,
2300 struct xfs_dinode **dipp,
2301 struct xfs_buf **bpp)
2302{
2303 struct xfs_mount *mp = tp->t_mountp;
2304 int error;
2305
2306 imap->im_blkno = 0;
2307 error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2308 if (error) {
2309 xfs_warn(mp, "%s: xfs_imap returned error %d.",
2310 __func__, error);
2311 return error;
2312 }
2313
2314 error = xfs_imap_to_bp(mp, tp, imap, dipp, bpp, 0);
2315 if (error) {
2316 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2317 __func__, error);
2318 return error;
2319 }
2320
2321 return 0;
2322}
2323
2324/*
2325 * Walk the unlinked chain from @head_agino until we find the inode that
2326 * points to @target_agino. Return the inode number, map, dinode pointer,
2327 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2328 *
2329 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2330 * @agino, @imap, @dipp, and @bpp are all output parameters.
2331 *
2332 * Do not call this function if @target_agino is the head of the list.
2333 */
2334STATIC int
2335xfs_iunlink_map_prev(
2336 struct xfs_trans *tp,
2337 xfs_agnumber_t agno,
2338 xfs_agino_t head_agino,
2339 xfs_agino_t target_agino,
2340 xfs_agino_t *agino,
2341 struct xfs_imap *imap,
2342 struct xfs_dinode **dipp,
2343 struct xfs_buf **bpp,
2344 struct xfs_perag *pag)
2345{
2346 struct xfs_mount *mp = tp->t_mountp;
2347 xfs_agino_t next_agino;
2348 int error;
2349
2350 ASSERT(head_agino != target_agino);
2351 *bpp = NULL;
2352
2353 /* See if our backref cache can find it faster. */
2354 *agino = xfs_iunlink_lookup_backref(pag, target_agino);
2355 if (*agino != NULLAGINO) {
2356 error = xfs_iunlink_map_ino(tp, agno, *agino, imap, dipp, bpp);
2357 if (error)
2358 return error;
2359
2360 if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2361 return 0;
2362
2363 /*
2364 * If we get here the cache contents were corrupt, so drop the
2365 * buffer and fall back to walking the bucket list.
2366 */
2367 xfs_trans_brelse(tp, *bpp);
2368 *bpp = NULL;
2369 WARN_ON_ONCE(1);
2370 }
2371
2372 trace_xfs_iunlink_map_prev_fallback(mp, agno);
2373
2374 /* Otherwise, walk the entire bucket until we find it. */
2375 next_agino = head_agino;
2376 while (next_agino != target_agino) {
2377 xfs_agino_t unlinked_agino;
2378
2379 if (*bpp)
2380 xfs_trans_brelse(tp, *bpp);
2381
2382 *agino = next_agino;
2383 error = xfs_iunlink_map_ino(tp, agno, next_agino, imap, dipp,
2384 bpp);
2385 if (error)
2386 return error;
2387
2388 unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2389 /*
2390 * Make sure this pointer is valid and isn't an obvious
2391 * infinite loop.
2392 */
2393 if (!xfs_verify_agino(mp, agno, unlinked_agino) ||
2394 next_agino == unlinked_agino) {
2395 XFS_CORRUPTION_ERROR(__func__,
2396 XFS_ERRLEVEL_LOW, mp,
2397 *dipp, sizeof(**dipp));
2398 error = -EFSCORRUPTED;
2399 return error;
2400 }
2401 next_agino = unlinked_agino;
2402 }
2403
2404 return 0;
2405}
2406
2407/*
2408 * Pull the on-disk inode from the AGI unlinked list.
2409 */
2410STATIC int
2411xfs_iunlink_remove(
2412 struct xfs_trans *tp,
2413 struct xfs_inode *ip)
2414{
2415 struct xfs_mount *mp = tp->t_mountp;
2416 struct xfs_agi *agi;
2417 struct xfs_buf *agibp;
2418 struct xfs_buf *last_ibp;
2419 struct xfs_dinode *last_dip = NULL;
2420 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2421 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2422 xfs_agino_t next_agino;
2423 xfs_agino_t head_agino;
2424 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2425 int error;
2426
2427 trace_xfs_iunlink_remove(ip);
2428
2429 /* Get the agi buffer first. It ensures lock ordering on the list. */
2430 error = xfs_read_agi(mp, tp, agno, &agibp);
2431 if (error)
2432 return error;
2433 agi = agibp->b_addr;
2434
2435 /*
2436 * Get the index into the agi hash table for the list this inode will
2437 * go on. Make sure the head pointer isn't garbage.
2438 */
2439 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2440 if (!xfs_verify_agino(mp, agno, head_agino)) {
2441 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2442 agi, sizeof(*agi));
2443 return -EFSCORRUPTED;
2444 }
2445
2446 /*
2447 * Set our inode's next_unlinked pointer to NULL and then return
2448 * the old pointer value so that we can update whatever was previous
2449 * to us in the list to point to whatever was next in the list.
2450 */
2451 error = xfs_iunlink_update_inode(tp, ip, agno, NULLAGINO, &next_agino);
2452 if (error)
2453 return error;
2454
2455 /*
2456 * If there was a backref pointing from the next inode back to this
2457 * one, remove it because we've removed this inode from the list.
2458 *
2459 * Later, if this inode was in the middle of the list we'll update
2460 * this inode's backref to point from the next inode.
2461 */
2462 if (next_agino != NULLAGINO) {
2463 error = xfs_iunlink_change_backref(agibp->b_pag, next_agino,
2464 NULLAGINO);
2465 if (error)
2466 return error;
2467 }
2468
2469 if (head_agino != agino) {
2470 struct xfs_imap imap;
2471 xfs_agino_t prev_agino;
2472
2473 /* We need to search the list for the inode being freed. */
2474 error = xfs_iunlink_map_prev(tp, agno, head_agino, agino,
2475 &prev_agino, &imap, &last_dip, &last_ibp,
2476 agibp->b_pag);
2477 if (error)
2478 return error;
2479
2480 /* Point the previous inode on the list to the next inode. */
2481 xfs_iunlink_update_dinode(tp, agno, prev_agino, last_ibp,
2482 last_dip, &imap, next_agino);
2483
2484 /*
2485 * Now we deal with the backref for this inode. If this inode
2486 * pointed at a real inode, change the backref that pointed to
2487 * us to point to our old next. If this inode was the end of
2488 * the list, delete the backref that pointed to us. Note that
2489 * change_backref takes care of deleting the backref if
2490 * next_agino is NULLAGINO.
2491 */
2492 return xfs_iunlink_change_backref(agibp->b_pag, agino,
2493 next_agino);
2494 }
2495
2496 /* Point the head of the list to the next unlinked inode. */
2497 return xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index,
2498 next_agino);
2499}
2500
2501/*
2502 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2503 * mark it stale. We should only find clean inodes in this lookup that aren't
2504 * already stale.
2505 */
2506static void
2507xfs_ifree_mark_inode_stale(
2508 struct xfs_buf *bp,
2509 struct xfs_inode *free_ip,
2510 xfs_ino_t inum)
2511{
2512 struct xfs_mount *mp = bp->b_mount;
2513 struct xfs_perag *pag = bp->b_pag;
2514 struct xfs_inode_log_item *iip;
2515 struct xfs_inode *ip;
2516
2517retry:
2518 rcu_read_lock();
2519 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2520
2521 /* Inode not in memory, nothing to do */
2522 if (!ip) {
2523 rcu_read_unlock();
2524 return;
2525 }
2526
2527 /*
2528 * because this is an RCU protected lookup, we could find a recently
2529 * freed or even reallocated inode during the lookup. We need to check
2530 * under the i_flags_lock for a valid inode here. Skip it if it is not
2531 * valid, the wrong inode or stale.
2532 */
2533 spin_lock(&ip->i_flags_lock);
2534 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE)) {
2535 spin_unlock(&ip->i_flags_lock);
2536 rcu_read_unlock();
2537 return;
2538 }
2539
2540 /*
2541 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2542 * other inodes that we did not find in the list attached to the buffer
2543 * and are not already marked stale. If we can't lock it, back off and
2544 * retry.
2545 */
2546 if (ip != free_ip) {
2547 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2548 spin_unlock(&ip->i_flags_lock);
2549 rcu_read_unlock();
2550 delay(1);
2551 goto retry;
2552 }
2553 }
2554 ip->i_flags |= XFS_ISTALE;
2555 spin_unlock(&ip->i_flags_lock);
2556 rcu_read_unlock();
2557
2558 /*
2559 * If we can't get the flush lock, the inode is already attached. All
2560 * we needed to do here is mark the inode stale so buffer IO completion
2561 * will remove it from the AIL.
2562 */
2563 iip = ip->i_itemp;
2564 if (!xfs_iflock_nowait(ip)) {
2565 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2566 ASSERT(iip->ili_last_fields);
2567 goto out_iunlock;
2568 }
2569
2570 /*
2571 * Inodes not attached to the buffer can be released immediately.
2572 * Everything else has to go through xfs_iflush_abort() on journal
2573 * commit as the flock synchronises removal of the inode from the
2574 * cluster buffer against inode reclaim.
2575 */
2576 if (!iip || list_empty(&iip->ili_item.li_bio_list)) {
2577 xfs_ifunlock(ip);
2578 goto out_iunlock;
2579 }
2580
2581 /* we have a dirty inode in memory that has not yet been flushed. */
2582 spin_lock(&iip->ili_lock);
2583 iip->ili_last_fields = iip->ili_fields;
2584 iip->ili_fields = 0;
2585 iip->ili_fsync_fields = 0;
2586 spin_unlock(&iip->ili_lock);
2587 ASSERT(iip->ili_last_fields);
2588
2589out_iunlock:
2590 if (ip != free_ip)
2591 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2592}
2593
2594/*
2595 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2596 * inodes that are in memory - they all must be marked stale and attached to
2597 * the cluster buffer.
2598 */
2599STATIC int
2600xfs_ifree_cluster(
2601 struct xfs_inode *free_ip,
2602 struct xfs_trans *tp,
2603 struct xfs_icluster *xic)
2604{
2605 struct xfs_mount *mp = free_ip->i_mount;
2606 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2607 struct xfs_buf *bp;
2608 xfs_daddr_t blkno;
2609 xfs_ino_t inum = xic->first_ino;
2610 int nbufs;
2611 int i, j;
2612 int ioffset;
2613 int error;
2614
2615 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2616
2617 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2618 /*
2619 * The allocation bitmap tells us which inodes of the chunk were
2620 * physically allocated. Skip the cluster if an inode falls into
2621 * a sparse region.
2622 */
2623 ioffset = inum - xic->first_ino;
2624 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2625 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2626 continue;
2627 }
2628
2629 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2630 XFS_INO_TO_AGBNO(mp, inum));
2631
2632 /*
2633 * We obtain and lock the backing buffer first in the process
2634 * here, as we have to ensure that any dirty inode that we
2635 * can't get the flush lock on is attached to the buffer.
2636 * If we scan the in-memory inodes first, then buffer IO can
2637 * complete before we get a lock on it, and hence we may fail
2638 * to mark all the active inodes on the buffer stale.
2639 */
2640 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2641 mp->m_bsize * igeo->blocks_per_cluster,
2642 XBF_UNMAPPED, &bp);
2643 if (error)
2644 return error;
2645
2646 /*
2647 * This buffer may not have been correctly initialised as we
2648 * didn't read it from disk. That's not important because we are
2649 * only using to mark the buffer as stale in the log, and to
2650 * attach stale cached inodes on it. That means it will never be
2651 * dispatched for IO. If it is, we want to know about it, and we
2652 * want it to fail. We can acheive this by adding a write
2653 * verifier to the buffer.
2654 */
2655 bp->b_ops = &xfs_inode_buf_ops;
2656
2657 /*
2658 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2659 * too. This requires lookups, and will skip inodes that we've
2660 * already marked XFS_ISTALE.
2661 */
2662 for (i = 0; i < igeo->inodes_per_cluster; i++)
2663 xfs_ifree_mark_inode_stale(bp, free_ip, inum + i);
2664
2665 xfs_trans_stale_inode_buf(tp, bp);
2666 xfs_trans_binval(tp, bp);
2667 }
2668 return 0;
2669}
2670
2671/*
2672 * This is called to return an inode to the inode free list.
2673 * The inode should already be truncated to 0 length and have
2674 * no pages associated with it. This routine also assumes that
2675 * the inode is already a part of the transaction.
2676 *
2677 * The on-disk copy of the inode will have been added to the list
2678 * of unlinked inodes in the AGI. We need to remove the inode from
2679 * that list atomically with respect to freeing it here.
2680 */
2681int
2682xfs_ifree(
2683 struct xfs_trans *tp,
2684 struct xfs_inode *ip)
2685{
2686 int error;
2687 struct xfs_icluster xic = { 0 };
2688 struct xfs_inode_log_item *iip = ip->i_itemp;
2689
2690 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2691 ASSERT(VFS_I(ip)->i_nlink == 0);
2692 ASSERT(ip->i_df.if_nextents == 0);
2693 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2694 ASSERT(ip->i_d.di_nblocks == 0);
2695
2696 /*
2697 * Pull the on-disk inode from the AGI unlinked list.
2698 */
2699 error = xfs_iunlink_remove(tp, ip);
2700 if (error)
2701 return error;
2702
2703 error = xfs_difree(tp, ip->i_ino, &xic);
2704 if (error)
2705 return error;
2706
2707 /*
2708 * Free any local-format data sitting around before we reset the
2709 * data fork to extents format. Note that the attr fork data has
2710 * already been freed by xfs_attr_inactive.
2711 */
2712 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2713 kmem_free(ip->i_df.if_u1.if_data);
2714 ip->i_df.if_u1.if_data = NULL;
2715 ip->i_df.if_bytes = 0;
2716 }
2717
2718 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2719 ip->i_d.di_flags = 0;
2720 ip->i_d.di_flags2 = 0;
2721 ip->i_d.di_dmevmask = 0;
2722 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2723 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2724
2725 /* Don't attempt to replay owner changes for a deleted inode */
2726 spin_lock(&iip->ili_lock);
2727 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2728 spin_unlock(&iip->ili_lock);
2729
2730 /*
2731 * Bump the generation count so no one will be confused
2732 * by reincarnations of this inode.
2733 */
2734 VFS_I(ip)->i_generation++;
2735 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2736
2737 if (xic.deleted)
2738 error = xfs_ifree_cluster(ip, tp, &xic);
2739
2740 return error;
2741}
2742
2743/*
2744 * This is called to unpin an inode. The caller must have the inode locked
2745 * in at least shared mode so that the buffer cannot be subsequently pinned
2746 * once someone is waiting for it to be unpinned.
2747 */
2748static void
2749xfs_iunpin(
2750 struct xfs_inode *ip)
2751{
2752 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2753
2754 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2755
2756 /* Give the log a push to start the unpinning I/O */
2757 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0, NULL);
2758
2759}
2760
2761static void
2762__xfs_iunpin_wait(
2763 struct xfs_inode *ip)
2764{
2765 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2766 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2767
2768 xfs_iunpin(ip);
2769
2770 do {
2771 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2772 if (xfs_ipincount(ip))
2773 io_schedule();
2774 } while (xfs_ipincount(ip));
2775 finish_wait(wq, &wait.wq_entry);
2776}
2777
2778void
2779xfs_iunpin_wait(
2780 struct xfs_inode *ip)
2781{
2782 if (xfs_ipincount(ip))
2783 __xfs_iunpin_wait(ip);
2784}
2785
2786/*
2787 * Removing an inode from the namespace involves removing the directory entry
2788 * and dropping the link count on the inode. Removing the directory entry can
2789 * result in locking an AGF (directory blocks were freed) and removing a link
2790 * count can result in placing the inode on an unlinked list which results in
2791 * locking an AGI.
2792 *
2793 * The big problem here is that we have an ordering constraint on AGF and AGI
2794 * locking - inode allocation locks the AGI, then can allocate a new extent for
2795 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2796 * removes the inode from the unlinked list, requiring that we lock the AGI
2797 * first, and then freeing the inode can result in an inode chunk being freed
2798 * and hence freeing disk space requiring that we lock an AGF.
2799 *
2800 * Hence the ordering that is imposed by other parts of the code is AGI before
2801 * AGF. This means we cannot remove the directory entry before we drop the inode
2802 * reference count and put it on the unlinked list as this results in a lock
2803 * order of AGF then AGI, and this can deadlock against inode allocation and
2804 * freeing. Therefore we must drop the link counts before we remove the
2805 * directory entry.
2806 *
2807 * This is still safe from a transactional point of view - it is not until we
2808 * get to xfs_defer_finish() that we have the possibility of multiple
2809 * transactions in this operation. Hence as long as we remove the directory
2810 * entry and drop the link count in the first transaction of the remove
2811 * operation, there are no transactional constraints on the ordering here.
2812 */
2813int
2814xfs_remove(
2815 xfs_inode_t *dp,
2816 struct xfs_name *name,
2817 xfs_inode_t *ip)
2818{
2819 xfs_mount_t *mp = dp->i_mount;
2820 xfs_trans_t *tp = NULL;
2821 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2822 int error = 0;
2823 uint resblks;
2824
2825 trace_xfs_remove(dp, name);
2826
2827 if (XFS_FORCED_SHUTDOWN(mp))
2828 return -EIO;
2829
2830 error = xfs_qm_dqattach(dp);
2831 if (error)
2832 goto std_return;
2833
2834 error = xfs_qm_dqattach(ip);
2835 if (error)
2836 goto std_return;
2837
2838 /*
2839 * We try to get the real space reservation first,
2840 * allowing for directory btree deletion(s) implying
2841 * possible bmap insert(s). If we can't get the space
2842 * reservation then we use 0 instead, and avoid the bmap
2843 * btree insert(s) in the directory code by, if the bmap
2844 * insert tries to happen, instead trimming the LAST
2845 * block from the directory.
2846 */
2847 resblks = XFS_REMOVE_SPACE_RES(mp);
2848 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2849 if (error == -ENOSPC) {
2850 resblks = 0;
2851 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2852 &tp);
2853 }
2854 if (error) {
2855 ASSERT(error != -ENOSPC);
2856 goto std_return;
2857 }
2858
2859 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2860
2861 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2862 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2863
2864 /*
2865 * If we're removing a directory perform some additional validation.
2866 */
2867 if (is_dir) {
2868 ASSERT(VFS_I(ip)->i_nlink >= 2);
2869 if (VFS_I(ip)->i_nlink != 2) {
2870 error = -ENOTEMPTY;
2871 goto out_trans_cancel;
2872 }
2873 if (!xfs_dir_isempty(ip)) {
2874 error = -ENOTEMPTY;
2875 goto out_trans_cancel;
2876 }
2877
2878 /* Drop the link from ip's "..". */
2879 error = xfs_droplink(tp, dp);
2880 if (error)
2881 goto out_trans_cancel;
2882
2883 /* Drop the "." link from ip to self. */
2884 error = xfs_droplink(tp, ip);
2885 if (error)
2886 goto out_trans_cancel;
2887 } else {
2888 /*
2889 * When removing a non-directory we need to log the parent
2890 * inode here. For a directory this is done implicitly
2891 * by the xfs_droplink call for the ".." entry.
2892 */
2893 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2894 }
2895 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2896
2897 /* Drop the link from dp to ip. */
2898 error = xfs_droplink(tp, ip);
2899 if (error)
2900 goto out_trans_cancel;
2901
2902 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2903 if (error) {
2904 ASSERT(error != -ENOENT);
2905 goto out_trans_cancel;
2906 }
2907
2908 /*
2909 * If this is a synchronous mount, make sure that the
2910 * remove transaction goes to disk before returning to
2911 * the user.
2912 */
2913 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2914 xfs_trans_set_sync(tp);
2915
2916 error = xfs_trans_commit(tp);
2917 if (error)
2918 goto std_return;
2919
2920 if (is_dir && xfs_inode_is_filestream(ip))
2921 xfs_filestream_deassociate(ip);
2922
2923 return 0;
2924
2925 out_trans_cancel:
2926 xfs_trans_cancel(tp);
2927 std_return:
2928 return error;
2929}
2930
2931/*
2932 * Enter all inodes for a rename transaction into a sorted array.
2933 */
2934#define __XFS_SORT_INODES 5
2935STATIC void
2936xfs_sort_for_rename(
2937 struct xfs_inode *dp1, /* in: old (source) directory inode */
2938 struct xfs_inode *dp2, /* in: new (target) directory inode */
2939 struct xfs_inode *ip1, /* in: inode of old entry */
2940 struct xfs_inode *ip2, /* in: inode of new entry */
2941 struct xfs_inode *wip, /* in: whiteout inode */
2942 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2943 int *num_inodes) /* in/out: inodes in array */
2944{
2945 int i, j;
2946
2947 ASSERT(*num_inodes == __XFS_SORT_INODES);
2948 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2949
2950 /*
2951 * i_tab contains a list of pointers to inodes. We initialize
2952 * the table here & we'll sort it. We will then use it to
2953 * order the acquisition of the inode locks.
2954 *
2955 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2956 */
2957 i = 0;
2958 i_tab[i++] = dp1;
2959 i_tab[i++] = dp2;
2960 i_tab[i++] = ip1;
2961 if (ip2)
2962 i_tab[i++] = ip2;
2963 if (wip)
2964 i_tab[i++] = wip;
2965 *num_inodes = i;
2966
2967 /*
2968 * Sort the elements via bubble sort. (Remember, there are at
2969 * most 5 elements to sort, so this is adequate.)
2970 */
2971 for (i = 0; i < *num_inodes; i++) {
2972 for (j = 1; j < *num_inodes; j++) {
2973 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2974 struct xfs_inode *temp = i_tab[j];
2975 i_tab[j] = i_tab[j-1];
2976 i_tab[j-1] = temp;
2977 }
2978 }
2979 }
2980}
2981
2982static int
2983xfs_finish_rename(
2984 struct xfs_trans *tp)
2985{
2986 /*
2987 * If this is a synchronous mount, make sure that the rename transaction
2988 * goes to disk before returning to the user.
2989 */
2990 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2991 xfs_trans_set_sync(tp);
2992
2993 return xfs_trans_commit(tp);
2994}
2995
2996/*
2997 * xfs_cross_rename()
2998 *
2999 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
3000 */
3001STATIC int
3002xfs_cross_rename(
3003 struct xfs_trans *tp,
3004 struct xfs_inode *dp1,
3005 struct xfs_name *name1,
3006 struct xfs_inode *ip1,
3007 struct xfs_inode *dp2,
3008 struct xfs_name *name2,
3009 struct xfs_inode *ip2,
3010 int spaceres)
3011{
3012 int error = 0;
3013 int ip1_flags = 0;
3014 int ip2_flags = 0;
3015 int dp2_flags = 0;
3016
3017 /* Swap inode number for dirent in first parent */
3018 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
3019 if (error)
3020 goto out_trans_abort;
3021
3022 /* Swap inode number for dirent in second parent */
3023 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
3024 if (error)
3025 goto out_trans_abort;
3026
3027 /*
3028 * If we're renaming one or more directories across different parents,
3029 * update the respective ".." entries (and link counts) to match the new
3030 * parents.
3031 */
3032 if (dp1 != dp2) {
3033 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3034
3035 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
3036 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
3037 dp1->i_ino, spaceres);
3038 if (error)
3039 goto out_trans_abort;
3040
3041 /* transfer ip2 ".." reference to dp1 */
3042 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
3043 error = xfs_droplink(tp, dp2);
3044 if (error)
3045 goto out_trans_abort;
3046 xfs_bumplink(tp, dp1);
3047 }
3048
3049 /*
3050 * Although ip1 isn't changed here, userspace needs
3051 * to be warned about the change, so that applications
3052 * relying on it (like backup ones), will properly
3053 * notify the change
3054 */
3055 ip1_flags |= XFS_ICHGTIME_CHG;
3056 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3057 }
3058
3059 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
3060 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
3061 dp2->i_ino, spaceres);
3062 if (error)
3063 goto out_trans_abort;
3064
3065 /* transfer ip1 ".." reference to dp2 */
3066 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
3067 error = xfs_droplink(tp, dp1);
3068 if (error)
3069 goto out_trans_abort;
3070 xfs_bumplink(tp, dp2);
3071 }
3072
3073 /*
3074 * Although ip2 isn't changed here, userspace needs
3075 * to be warned about the change, so that applications
3076 * relying on it (like backup ones), will properly
3077 * notify the change
3078 */
3079 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3080 ip2_flags |= XFS_ICHGTIME_CHG;
3081 }
3082 }
3083
3084 if (ip1_flags) {
3085 xfs_trans_ichgtime(tp, ip1, ip1_flags);
3086 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
3087 }
3088 if (ip2_flags) {
3089 xfs_trans_ichgtime(tp, ip2, ip2_flags);
3090 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
3091 }
3092 if (dp2_flags) {
3093 xfs_trans_ichgtime(tp, dp2, dp2_flags);
3094 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
3095 }
3096 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3097 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
3098 return xfs_finish_rename(tp);
3099
3100out_trans_abort:
3101 xfs_trans_cancel(tp);
3102 return error;
3103}
3104
3105/*
3106 * xfs_rename_alloc_whiteout()
3107 *
3108 * Return a referenced, unlinked, unlocked inode that can be used as a
3109 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3110 * crash between allocating the inode and linking it into the rename transaction
3111 * recovery will free the inode and we won't leak it.
3112 */
3113static int
3114xfs_rename_alloc_whiteout(
3115 struct xfs_inode *dp,
3116 struct xfs_inode **wip)
3117{
3118 struct xfs_inode *tmpfile;
3119 int error;
3120
3121 error = xfs_create_tmpfile(dp, S_IFCHR | WHITEOUT_MODE, &tmpfile);
3122 if (error)
3123 return error;
3124
3125 /*
3126 * Prepare the tmpfile inode as if it were created through the VFS.
3127 * Complete the inode setup and flag it as linkable. nlink is already
3128 * zero, so we can skip the drop_nlink.
3129 */
3130 xfs_setup_iops(tmpfile);
3131 xfs_finish_inode_setup(tmpfile);
3132 VFS_I(tmpfile)->i_state |= I_LINKABLE;
3133
3134 *wip = tmpfile;
3135 return 0;
3136}
3137
3138/*
3139 * xfs_rename
3140 */
3141int
3142xfs_rename(
3143 struct xfs_inode *src_dp,
3144 struct xfs_name *src_name,
3145 struct xfs_inode *src_ip,
3146 struct xfs_inode *target_dp,
3147 struct xfs_name *target_name,
3148 struct xfs_inode *target_ip,
3149 unsigned int flags)
3150{
3151 struct xfs_mount *mp = src_dp->i_mount;
3152 struct xfs_trans *tp;
3153 struct xfs_inode *wip = NULL; /* whiteout inode */
3154 struct xfs_inode *inodes[__XFS_SORT_INODES];
3155 struct xfs_buf *agibp;
3156 int num_inodes = __XFS_SORT_INODES;
3157 bool new_parent = (src_dp != target_dp);
3158 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3159 int spaceres;
3160 int error;
3161
3162 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3163
3164 if ((flags & RENAME_EXCHANGE) && !target_ip)
3165 return -EINVAL;
3166
3167 /*
3168 * If we are doing a whiteout operation, allocate the whiteout inode
3169 * we will be placing at the target and ensure the type is set
3170 * appropriately.
3171 */
3172 if (flags & RENAME_WHITEOUT) {
3173 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3174 error = xfs_rename_alloc_whiteout(target_dp, &wip);
3175 if (error)
3176 return error;
3177
3178 /* setup target dirent info as whiteout */
3179 src_name->type = XFS_DIR3_FT_CHRDEV;
3180 }
3181
3182 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3183 inodes, &num_inodes);
3184
3185 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3186 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3187 if (error == -ENOSPC) {
3188 spaceres = 0;
3189 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3190 &tp);
3191 }
3192 if (error)
3193 goto out_release_wip;
3194
3195 /*
3196 * Attach the dquots to the inodes
3197 */
3198 error = xfs_qm_vop_rename_dqattach(inodes);
3199 if (error)
3200 goto out_trans_cancel;
3201
3202 /*
3203 * Lock all the participating inodes. Depending upon whether
3204 * the target_name exists in the target directory, and
3205 * whether the target directory is the same as the source
3206 * directory, we can lock from 2 to 4 inodes.
3207 */
3208 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3209
3210 /*
3211 * Join all the inodes to the transaction. From this point on,
3212 * we can rely on either trans_commit or trans_cancel to unlock
3213 * them.
3214 */
3215 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3216 if (new_parent)
3217 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3218 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3219 if (target_ip)
3220 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3221 if (wip)
3222 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3223
3224 /*
3225 * If we are using project inheritance, we only allow renames
3226 * into our tree when the project IDs are the same; else the
3227 * tree quota mechanism would be circumvented.
3228 */
3229 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
3230 target_dp->i_d.di_projid != src_ip->i_d.di_projid)) {
3231 error = -EXDEV;
3232 goto out_trans_cancel;
3233 }
3234
3235 /* RENAME_EXCHANGE is unique from here on. */
3236 if (flags & RENAME_EXCHANGE)
3237 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3238 target_dp, target_name, target_ip,
3239 spaceres);
3240
3241 /*
3242 * Check for expected errors before we dirty the transaction
3243 * so we can return an error without a transaction abort.
3244 */
3245 if (target_ip == NULL) {
3246 /*
3247 * If there's no space reservation, check the entry will
3248 * fit before actually inserting it.
3249 */
3250 if (!spaceres) {
3251 error = xfs_dir_canenter(tp, target_dp, target_name);
3252 if (error)
3253 goto out_trans_cancel;
3254 }
3255 } else {
3256 /*
3257 * If target exists and it's a directory, check that whether
3258 * it can be destroyed.
3259 */
3260 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3261 (!xfs_dir_isempty(target_ip) ||
3262 (VFS_I(target_ip)->i_nlink > 2))) {
3263 error = -EEXIST;
3264 goto out_trans_cancel;
3265 }
3266 }
3267
3268 /*
3269 * Directory entry creation below may acquire the AGF. Remove
3270 * the whiteout from the unlinked list first to preserve correct
3271 * AGI/AGF locking order. This dirties the transaction so failures
3272 * after this point will abort and log recovery will clean up the
3273 * mess.
3274 *
3275 * For whiteouts, we need to bump the link count on the whiteout
3276 * inode. After this point, we have a real link, clear the tmpfile
3277 * state flag from the inode so it doesn't accidentally get misused
3278 * in future.
3279 */
3280 if (wip) {
3281 ASSERT(VFS_I(wip)->i_nlink == 0);
3282 error = xfs_iunlink_remove(tp, wip);
3283 if (error)
3284 goto out_trans_cancel;
3285
3286 xfs_bumplink(tp, wip);
3287 VFS_I(wip)->i_state &= ~I_LINKABLE;
3288 }
3289
3290 /*
3291 * Set up the target.
3292 */
3293 if (target_ip == NULL) {
3294 /*
3295 * If target does not exist and the rename crosses
3296 * directories, adjust the target directory link count
3297 * to account for the ".." reference from the new entry.
3298 */
3299 error = xfs_dir_createname(tp, target_dp, target_name,
3300 src_ip->i_ino, spaceres);
3301 if (error)
3302 goto out_trans_cancel;
3303
3304 xfs_trans_ichgtime(tp, target_dp,
3305 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3306
3307 if (new_parent && src_is_directory) {
3308 xfs_bumplink(tp, target_dp);
3309 }
3310 } else { /* target_ip != NULL */
3311 /*
3312 * Link the source inode under the target name.
3313 * If the source inode is a directory and we are moving
3314 * it across directories, its ".." entry will be
3315 * inconsistent until we replace that down below.
3316 *
3317 * In case there is already an entry with the same
3318 * name at the destination directory, remove it first.
3319 */
3320
3321 /*
3322 * Check whether the replace operation will need to allocate
3323 * blocks. This happens when the shortform directory lacks
3324 * space and we have to convert it to a block format directory.
3325 * When more blocks are necessary, we must lock the AGI first
3326 * to preserve locking order (AGI -> AGF).
3327 */
3328 if (xfs_dir2_sf_replace_needblock(target_dp, src_ip->i_ino)) {
3329 error = xfs_read_agi(mp, tp,
3330 XFS_INO_TO_AGNO(mp, target_ip->i_ino),
3331 &agibp);
3332 if (error)
3333 goto out_trans_cancel;
3334 }
3335
3336 error = xfs_dir_replace(tp, target_dp, target_name,
3337 src_ip->i_ino, spaceres);
3338 if (error)
3339 goto out_trans_cancel;
3340
3341 xfs_trans_ichgtime(tp, target_dp,
3342 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3343
3344 /*
3345 * Decrement the link count on the target since the target
3346 * dir no longer points to it.
3347 */
3348 error = xfs_droplink(tp, target_ip);
3349 if (error)
3350 goto out_trans_cancel;
3351
3352 if (src_is_directory) {
3353 /*
3354 * Drop the link from the old "." entry.
3355 */
3356 error = xfs_droplink(tp, target_ip);
3357 if (error)
3358 goto out_trans_cancel;
3359 }
3360 } /* target_ip != NULL */
3361
3362 /*
3363 * Remove the source.
3364 */
3365 if (new_parent && src_is_directory) {
3366 /*
3367 * Rewrite the ".." entry to point to the new
3368 * directory.
3369 */
3370 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3371 target_dp->i_ino, spaceres);
3372 ASSERT(error != -EEXIST);
3373 if (error)
3374 goto out_trans_cancel;
3375 }
3376
3377 /*
3378 * We always want to hit the ctime on the source inode.
3379 *
3380 * This isn't strictly required by the standards since the source
3381 * inode isn't really being changed, but old unix file systems did
3382 * it and some incremental backup programs won't work without it.
3383 */
3384 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3385 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3386
3387 /*
3388 * Adjust the link count on src_dp. This is necessary when
3389 * renaming a directory, either within one parent when
3390 * the target existed, or across two parent directories.
3391 */
3392 if (src_is_directory && (new_parent || target_ip != NULL)) {
3393
3394 /*
3395 * Decrement link count on src_directory since the
3396 * entry that's moved no longer points to it.
3397 */
3398 error = xfs_droplink(tp, src_dp);
3399 if (error)
3400 goto out_trans_cancel;
3401 }
3402
3403 /*
3404 * For whiteouts, we only need to update the source dirent with the
3405 * inode number of the whiteout inode rather than removing it
3406 * altogether.
3407 */
3408 if (wip) {
3409 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3410 spaceres);
3411 } else
3412 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3413 spaceres);
3414 if (error)
3415 goto out_trans_cancel;
3416
3417 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3418 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3419 if (new_parent)
3420 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3421
3422 error = xfs_finish_rename(tp);
3423 if (wip)
3424 xfs_irele(wip);
3425 return error;
3426
3427out_trans_cancel:
3428 xfs_trans_cancel(tp);
3429out_release_wip:
3430 if (wip)
3431 xfs_irele(wip);
3432 return error;
3433}
3434
3435static int
3436xfs_iflush(
3437 struct xfs_inode *ip,
3438 struct xfs_buf *bp)
3439{
3440 struct xfs_inode_log_item *iip = ip->i_itemp;
3441 struct xfs_dinode *dip;
3442 struct xfs_mount *mp = ip->i_mount;
3443 int error;
3444
3445 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3446 ASSERT(xfs_isiflocked(ip));
3447 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3448 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3449 ASSERT(iip->ili_item.li_buf == bp);
3450
3451 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3452
3453 /*
3454 * We don't flush the inode if any of the following checks fail, but we
3455 * do still update the log item and attach to the backing buffer as if
3456 * the flush happened. This is a formality to facilitate predictable
3457 * error handling as the caller will shutdown and fail the buffer.
3458 */
3459 error = -EFSCORRUPTED;
3460 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3461 mp, XFS_ERRTAG_IFLUSH_1)) {
3462 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3463 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3464 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3465 goto flush_out;
3466 }
3467 if (S_ISREG(VFS_I(ip)->i_mode)) {
3468 if (XFS_TEST_ERROR(
3469 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3470 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3471 mp, XFS_ERRTAG_IFLUSH_3)) {
3472 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3473 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3474 __func__, ip->i_ino, ip);
3475 goto flush_out;
3476 }
3477 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3478 if (XFS_TEST_ERROR(
3479 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3480 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3481 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3482 mp, XFS_ERRTAG_IFLUSH_4)) {
3483 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3484 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3485 __func__, ip->i_ino, ip);
3486 goto flush_out;
3487 }
3488 }
3489 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp) >
3490 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3491 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3492 "%s: detected corrupt incore inode %Lu, "
3493 "total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3494 __func__, ip->i_ino,
3495 ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp),
3496 ip->i_d.di_nblocks, ip);
3497 goto flush_out;
3498 }
3499 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3500 mp, XFS_ERRTAG_IFLUSH_6)) {
3501 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3502 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3503 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3504 goto flush_out;
3505 }
3506
3507 /*
3508 * Inode item log recovery for v2 inodes are dependent on the
3509 * di_flushiter count for correct sequencing. We bump the flush
3510 * iteration count so we can detect flushes which postdate a log record
3511 * during recovery. This is redundant as we now log every change and
3512 * hence this can't happen but we need to still do it to ensure
3513 * backwards compatibility with old kernels that predate logging all
3514 * inode changes.
3515 */
3516 if (!xfs_sb_version_has_v3inode(&mp->m_sb))
3517 ip->i_d.di_flushiter++;
3518
3519 /*
3520 * If there are inline format data / attr forks attached to this inode,
3521 * make sure they are not corrupt.
3522 */
3523 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3524 xfs_ifork_verify_local_data(ip))
3525 goto flush_out;
3526 if (ip->i_afp && ip->i_afp->if_format == XFS_DINODE_FMT_LOCAL &&
3527 xfs_ifork_verify_local_attr(ip))
3528 goto flush_out;
3529
3530 /*
3531 * Copy the dirty parts of the inode into the on-disk inode. We always
3532 * copy out the core of the inode, because if the inode is dirty at all
3533 * the core must be.
3534 */
3535 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3536
3537 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3538 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3539 ip->i_d.di_flushiter = 0;
3540
3541 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3542 if (XFS_IFORK_Q(ip))
3543 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3544
3545 /*
3546 * We've recorded everything logged in the inode, so we'd like to clear
3547 * the ili_fields bits so we don't log and flush things unnecessarily.
3548 * However, we can't stop logging all this information until the data
3549 * we've copied into the disk buffer is written to disk. If we did we
3550 * might overwrite the copy of the inode in the log with all the data
3551 * after re-logging only part of it, and in the face of a crash we
3552 * wouldn't have all the data we need to recover.
3553 *
3554 * What we do is move the bits to the ili_last_fields field. When
3555 * logging the inode, these bits are moved back to the ili_fields field.
3556 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3557 * know that the information those bits represent is permanently on
3558 * disk. As long as the flush completes before the inode is logged
3559 * again, then both ili_fields and ili_last_fields will be cleared.
3560 */
3561 error = 0;
3562flush_out:
3563 spin_lock(&iip->ili_lock);
3564 iip->ili_last_fields = iip->ili_fields;
3565 iip->ili_fields = 0;
3566 iip->ili_fsync_fields = 0;
3567 spin_unlock(&iip->ili_lock);
3568
3569 /*
3570 * Store the current LSN of the inode so that we can tell whether the
3571 * item has moved in the AIL from xfs_iflush_done().
3572 */
3573 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3574 &iip->ili_item.li_lsn);
3575
3576 /* generate the checksum. */
3577 xfs_dinode_calc_crc(mp, dip);
3578 return error;
3579}
3580
3581/*
3582 * Non-blocking flush of dirty inode metadata into the backing buffer.
3583 *
3584 * The caller must have a reference to the inode and hold the cluster buffer
3585 * locked. The function will walk across all the inodes on the cluster buffer it
3586 * can find and lock without blocking, and flush them to the cluster buffer.
3587 *
3588 * On successful flushing of at least one inode, the caller must write out the
3589 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3590 * the caller needs to release the buffer. On failure, the filesystem will be
3591 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3592 * will be returned.
3593 */
3594int
3595xfs_iflush_cluster(
3596 struct xfs_buf *bp)
3597{
3598 struct xfs_mount *mp = bp->b_mount;
3599 struct xfs_log_item *lip, *n;
3600 struct xfs_inode *ip;
3601 struct xfs_inode_log_item *iip;
3602 int clcount = 0;
3603 int error = 0;
3604
3605 /*
3606 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3607 * can remove itself from the list.
3608 */
3609 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3610 iip = (struct xfs_inode_log_item *)lip;
3611 ip = iip->ili_inode;
3612
3613 /*
3614 * Quick and dirty check to avoid locks if possible.
3615 */
3616 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLOCK))
3617 continue;
3618 if (xfs_ipincount(ip))
3619 continue;
3620
3621 /*
3622 * The inode is still attached to the buffer, which means it is
3623 * dirty but reclaim might try to grab it. Check carefully for
3624 * that, and grab the ilock while still holding the i_flags_lock
3625 * to guarantee reclaim will not be able to reclaim this inode
3626 * once we drop the i_flags_lock.
3627 */
3628 spin_lock(&ip->i_flags_lock);
3629 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3630 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLOCK)) {
3631 spin_unlock(&ip->i_flags_lock);
3632 continue;
3633 }
3634
3635 /*
3636 * ILOCK will pin the inode against reclaim and prevent
3637 * concurrent transactions modifying the inode while we are
3638 * flushing the inode.
3639 */
3640 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3641 spin_unlock(&ip->i_flags_lock);
3642 continue;
3643 }
3644 spin_unlock(&ip->i_flags_lock);
3645
3646 /*
3647 * Skip inodes that are already flush locked as they have
3648 * already been written to the buffer.
3649 */
3650 if (!xfs_iflock_nowait(ip)) {
3651 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3652 continue;
3653 }
3654
3655 /*
3656 * Abort flushing this inode if we are shut down because the
3657 * inode may not currently be in the AIL. This can occur when
3658 * log I/O failure unpins the inode without inserting into the
3659 * AIL, leaving a dirty/unpinned inode attached to the buffer
3660 * that otherwise looks like it should be flushed.
3661 */
3662 if (XFS_FORCED_SHUTDOWN(mp)) {
3663 xfs_iunpin_wait(ip);
3664 /* xfs_iflush_abort() drops the flush lock */
3665 xfs_iflush_abort(ip);
3666 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3667 error = -EIO;
3668 continue;
3669 }
3670
3671 /* don't block waiting on a log force to unpin dirty inodes */
3672 if (xfs_ipincount(ip)) {
3673 xfs_ifunlock(ip);
3674 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3675 continue;
3676 }
3677
3678 if (!xfs_inode_clean(ip))
3679 error = xfs_iflush(ip, bp);
3680 else
3681 xfs_ifunlock(ip);
3682 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3683 if (error)
3684 break;
3685 clcount++;
3686 }
3687
3688 if (error) {
3689 bp->b_flags |= XBF_ASYNC;
3690 xfs_buf_ioend_fail(bp);
3691 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3692 return error;
3693 }
3694
3695 if (!clcount)
3696 return -EAGAIN;
3697
3698 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3699 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3700 return 0;
3701
3702}
3703
3704/* Release an inode. */
3705void
3706xfs_irele(
3707 struct xfs_inode *ip)
3708{
3709 trace_xfs_irele(ip, _RET_IP_);
3710 iput(VFS_I(ip));
3711}
3712
3713/*
3714 * Ensure all commited transactions touching the inode are written to the log.
3715 */
3716int
3717xfs_log_force_inode(
3718 struct xfs_inode *ip)
3719{
3720 xfs_lsn_t lsn = 0;
3721
3722 xfs_ilock(ip, XFS_ILOCK_SHARED);
3723 if (xfs_ipincount(ip))
3724 lsn = ip->i_itemp->ili_last_lsn;
3725 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3726
3727 if (!lsn)
3728 return 0;
3729 return xfs_log_force_lsn(ip->i_mount, lsn, XFS_LOG_SYNC, NULL);
3730}
3731
3732/*
3733 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3734 * abide vfs locking order (lowest pointer value goes first) and breaking the
3735 * layout leases before proceeding. The loop is needed because we cannot call
3736 * the blocking break_layout() with the iolocks held, and therefore have to
3737 * back out both locks.
3738 */
3739static int
3740xfs_iolock_two_inodes_and_break_layout(
3741 struct inode *src,
3742 struct inode *dest)
3743{
3744 int error;
3745
3746 if (src > dest)
3747 swap(src, dest);
3748
3749retry:
3750 /* Wait to break both inodes' layouts before we start locking. */
3751 error = break_layout(src, true);
3752 if (error)
3753 return error;
3754 if (src != dest) {
3755 error = break_layout(dest, true);
3756 if (error)
3757 return error;
3758 }
3759
3760 /* Lock one inode and make sure nobody got in and leased it. */
3761 inode_lock(src);
3762 error = break_layout(src, false);
3763 if (error) {
3764 inode_unlock(src);
3765 if (error == -EWOULDBLOCK)
3766 goto retry;
3767 return error;
3768 }
3769
3770 if (src == dest)
3771 return 0;
3772
3773 /* Lock the other inode and make sure nobody got in and leased it. */
3774 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3775 error = break_layout(dest, false);
3776 if (error) {
3777 inode_unlock(src);
3778 inode_unlock(dest);
3779 if (error == -EWOULDBLOCK)
3780 goto retry;
3781 return error;
3782 }
3783
3784 return 0;
3785}
3786
3787/*
3788 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3789 * mmap activity.
3790 */
3791int
3792xfs_ilock2_io_mmap(
3793 struct xfs_inode *ip1,
3794 struct xfs_inode *ip2)
3795{
3796 int ret;
3797
3798 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3799 if (ret)
3800 return ret;
3801 if (ip1 == ip2)
3802 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3803 else
3804 xfs_lock_two_inodes(ip1, XFS_MMAPLOCK_EXCL,
3805 ip2, XFS_MMAPLOCK_EXCL);
3806 return 0;
3807}
3808
3809/* Unlock both inodes to allow IO and mmap activity. */
3810void
3811xfs_iunlock2_io_mmap(
3812 struct xfs_inode *ip1,
3813 struct xfs_inode *ip2)
3814{
3815 bool same_inode = (ip1 == ip2);
3816
3817 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3818 if (!same_inode)
3819 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3820 inode_unlock(VFS_I(ip2));
3821 if (!same_inode)
3822 inode_unlock(VFS_I(ip1));
3823}