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_iflush_int(struct xfs_inode *, struct xfs_buf *);
48STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
49STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
50
51/*
52 * helper function to extract extent size hint from inode
53 */
54xfs_extlen_t
55xfs_get_extsz_hint(
56 struct xfs_inode *ip)
57{
58 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
59 return ip->i_d.di_extsize;
60 if (XFS_IS_REALTIME_INODE(ip))
61 return ip->i_mount->m_sb.sb_rextsize;
62 return 0;
63}
64
65/*
66 * Helper function to extract CoW extent size hint from inode.
67 * Between the extent size hint and the CoW extent size hint, we
68 * return the greater of the two. If the value is zero (automatic),
69 * use the default size.
70 */
71xfs_extlen_t
72xfs_get_cowextsz_hint(
73 struct xfs_inode *ip)
74{
75 xfs_extlen_t a, b;
76
77 a = 0;
78 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
79 a = ip->i_d.di_cowextsize;
80 b = xfs_get_extsz_hint(ip);
81
82 a = max(a, b);
83 if (a == 0)
84 return XFS_DEFAULT_COWEXTSZ_HINT;
85 return a;
86}
87
88/*
89 * These two are wrapper routines around the xfs_ilock() routine used to
90 * centralize some grungy code. They are used in places that wish to lock the
91 * inode solely for reading the extents. The reason these places can't just
92 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
93 * bringing in of the extents from disk for a file in b-tree format. If the
94 * inode is in b-tree format, then we need to lock the inode exclusively until
95 * the extents are read in. Locking it exclusively all the time would limit
96 * our parallelism unnecessarily, though. What we do instead is check to see
97 * if the extents have been read in yet, and only lock the inode exclusively
98 * if they have not.
99 *
100 * The functions return a value which should be given to the corresponding
101 * xfs_iunlock() call.
102 */
103uint
104xfs_ilock_data_map_shared(
105 struct xfs_inode *ip)
106{
107 uint lock_mode = XFS_ILOCK_SHARED;
108
109 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
110 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
111 lock_mode = XFS_ILOCK_EXCL;
112 xfs_ilock(ip, lock_mode);
113 return lock_mode;
114}
115
116uint
117xfs_ilock_attr_map_shared(
118 struct xfs_inode *ip)
119{
120 uint lock_mode = XFS_ILOCK_SHARED;
121
122 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
123 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
124 lock_mode = XFS_ILOCK_EXCL;
125 xfs_ilock(ip, lock_mode);
126 return lock_mode;
127}
128
129/*
130 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
131 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
132 * various combinations of the locks to be obtained.
133 *
134 * The 3 locks should always be ordered so that the IO lock is obtained first,
135 * the mmap lock second and the ilock last in order to prevent deadlock.
136 *
137 * Basic locking order:
138 *
139 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
140 *
141 * mmap_sem locking order:
142 *
143 * i_rwsem -> page lock -> mmap_sem
144 * mmap_sem -> i_mmap_lock -> page_lock
145 *
146 * The difference in mmap_sem locking order mean that we cannot hold the
147 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
148 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
149 * in get_user_pages() to map the user pages into the kernel address space for
150 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
151 * page faults already hold the mmap_sem.
152 *
153 * Hence to serialise fully against both syscall and mmap based IO, we need to
154 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
155 * taken in places where we need to invalidate the page cache in a race
156 * free manner (e.g. truncate, hole punch and other extent manipulation
157 * functions).
158 */
159void
160xfs_ilock(
161 xfs_inode_t *ip,
162 uint lock_flags)
163{
164 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
165
166 /*
167 * You can't set both SHARED and EXCL for the same lock,
168 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
169 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
170 */
171 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
172 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
173 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
174 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
175 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
176 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
177 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
178
179 if (lock_flags & XFS_IOLOCK_EXCL) {
180 down_write_nested(&VFS_I(ip)->i_rwsem,
181 XFS_IOLOCK_DEP(lock_flags));
182 } else if (lock_flags & XFS_IOLOCK_SHARED) {
183 down_read_nested(&VFS_I(ip)->i_rwsem,
184 XFS_IOLOCK_DEP(lock_flags));
185 }
186
187 if (lock_flags & XFS_MMAPLOCK_EXCL)
188 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
189 else if (lock_flags & XFS_MMAPLOCK_SHARED)
190 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
191
192 if (lock_flags & XFS_ILOCK_EXCL)
193 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
194 else if (lock_flags & XFS_ILOCK_SHARED)
195 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
196}
197
198/*
199 * This is just like xfs_ilock(), except that the caller
200 * is guaranteed not to sleep. It returns 1 if it gets
201 * the requested locks and 0 otherwise. If the IO lock is
202 * obtained but the inode lock cannot be, then the IO lock
203 * is dropped before returning.
204 *
205 * ip -- the inode being locked
206 * lock_flags -- this parameter indicates the inode's locks to be
207 * to be locked. See the comment for xfs_ilock() for a list
208 * of valid values.
209 */
210int
211xfs_ilock_nowait(
212 xfs_inode_t *ip,
213 uint lock_flags)
214{
215 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
216
217 /*
218 * You can't set both SHARED and EXCL for the same lock,
219 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
220 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
221 */
222 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
223 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
224 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
225 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
226 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
227 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
228 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
229
230 if (lock_flags & XFS_IOLOCK_EXCL) {
231 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
232 goto out;
233 } else if (lock_flags & XFS_IOLOCK_SHARED) {
234 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
235 goto out;
236 }
237
238 if (lock_flags & XFS_MMAPLOCK_EXCL) {
239 if (!mrtryupdate(&ip->i_mmaplock))
240 goto out_undo_iolock;
241 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
242 if (!mrtryaccess(&ip->i_mmaplock))
243 goto out_undo_iolock;
244 }
245
246 if (lock_flags & XFS_ILOCK_EXCL) {
247 if (!mrtryupdate(&ip->i_lock))
248 goto out_undo_mmaplock;
249 } else if (lock_flags & XFS_ILOCK_SHARED) {
250 if (!mrtryaccess(&ip->i_lock))
251 goto out_undo_mmaplock;
252 }
253 return 1;
254
255out_undo_mmaplock:
256 if (lock_flags & XFS_MMAPLOCK_EXCL)
257 mrunlock_excl(&ip->i_mmaplock);
258 else if (lock_flags & XFS_MMAPLOCK_SHARED)
259 mrunlock_shared(&ip->i_mmaplock);
260out_undo_iolock:
261 if (lock_flags & XFS_IOLOCK_EXCL)
262 up_write(&VFS_I(ip)->i_rwsem);
263 else if (lock_flags & XFS_IOLOCK_SHARED)
264 up_read(&VFS_I(ip)->i_rwsem);
265out:
266 return 0;
267}
268
269/*
270 * xfs_iunlock() is used to drop the inode locks acquired with
271 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
272 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
273 * that we know which locks to drop.
274 *
275 * ip -- the inode being unlocked
276 * lock_flags -- this parameter indicates the inode's locks to be
277 * to be unlocked. See the comment for xfs_ilock() for a list
278 * of valid values for this parameter.
279 *
280 */
281void
282xfs_iunlock(
283 xfs_inode_t *ip,
284 uint lock_flags)
285{
286 /*
287 * You can't set both SHARED and EXCL for the same lock,
288 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
289 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
290 */
291 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
292 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
293 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
294 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
295 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
296 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
297 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
298 ASSERT(lock_flags != 0);
299
300 if (lock_flags & XFS_IOLOCK_EXCL)
301 up_write(&VFS_I(ip)->i_rwsem);
302 else if (lock_flags & XFS_IOLOCK_SHARED)
303 up_read(&VFS_I(ip)->i_rwsem);
304
305 if (lock_flags & XFS_MMAPLOCK_EXCL)
306 mrunlock_excl(&ip->i_mmaplock);
307 else if (lock_flags & XFS_MMAPLOCK_SHARED)
308 mrunlock_shared(&ip->i_mmaplock);
309
310 if (lock_flags & XFS_ILOCK_EXCL)
311 mrunlock_excl(&ip->i_lock);
312 else if (lock_flags & XFS_ILOCK_SHARED)
313 mrunlock_shared(&ip->i_lock);
314
315 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
316}
317
318/*
319 * give up write locks. the i/o lock cannot be held nested
320 * if it is being demoted.
321 */
322void
323xfs_ilock_demote(
324 xfs_inode_t *ip,
325 uint lock_flags)
326{
327 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
328 ASSERT((lock_flags &
329 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
330
331 if (lock_flags & XFS_ILOCK_EXCL)
332 mrdemote(&ip->i_lock);
333 if (lock_flags & XFS_MMAPLOCK_EXCL)
334 mrdemote(&ip->i_mmaplock);
335 if (lock_flags & XFS_IOLOCK_EXCL)
336 downgrade_write(&VFS_I(ip)->i_rwsem);
337
338 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
339}
340
341#if defined(DEBUG) || defined(XFS_WARN)
342int
343xfs_isilocked(
344 xfs_inode_t *ip,
345 uint lock_flags)
346{
347 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
348 if (!(lock_flags & XFS_ILOCK_SHARED))
349 return !!ip->i_lock.mr_writer;
350 return rwsem_is_locked(&ip->i_lock.mr_lock);
351 }
352
353 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
354 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
355 return !!ip->i_mmaplock.mr_writer;
356 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
357 }
358
359 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
360 if (!(lock_flags & XFS_IOLOCK_SHARED))
361 return !debug_locks ||
362 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
363 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
364 }
365
366 ASSERT(0);
367 return 0;
368}
369#endif
370
371/*
372 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
373 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
374 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
375 * errors and warnings.
376 */
377#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
378static bool
379xfs_lockdep_subclass_ok(
380 int subclass)
381{
382 return subclass < MAX_LOCKDEP_SUBCLASSES;
383}
384#else
385#define xfs_lockdep_subclass_ok(subclass) (true)
386#endif
387
388/*
389 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
390 * value. This can be called for any type of inode lock combination, including
391 * parent locking. Care must be taken to ensure we don't overrun the subclass
392 * storage fields in the class mask we build.
393 */
394static inline int
395xfs_lock_inumorder(int lock_mode, int subclass)
396{
397 int class = 0;
398
399 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
400 XFS_ILOCK_RTSUM)));
401 ASSERT(xfs_lockdep_subclass_ok(subclass));
402
403 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
404 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
405 class += subclass << XFS_IOLOCK_SHIFT;
406 }
407
408 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
409 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
410 class += subclass << XFS_MMAPLOCK_SHIFT;
411 }
412
413 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
414 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
415 class += subclass << XFS_ILOCK_SHIFT;
416 }
417
418 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
419}
420
421/*
422 * The following routine will lock n inodes in exclusive mode. We assume the
423 * caller calls us with the inodes in i_ino order.
424 *
425 * We need to detect deadlock where an inode that we lock is in the AIL and we
426 * start waiting for another inode that is locked by a thread in a long running
427 * transaction (such as truncate). This can result in deadlock since the long
428 * running trans might need to wait for the inode we just locked in order to
429 * push the tail and free space in the log.
430 *
431 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
432 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
433 * lock more than one at a time, lockdep will report false positives saying we
434 * have violated locking orders.
435 */
436static void
437xfs_lock_inodes(
438 struct xfs_inode **ips,
439 int inodes,
440 uint lock_mode)
441{
442 int attempts = 0, i, j, try_lock;
443 struct xfs_log_item *lp;
444
445 /*
446 * Currently supports between 2 and 5 inodes with exclusive locking. We
447 * support an arbitrary depth of locking here, but absolute limits on
448 * inodes depend on the the type of locking and the limits placed by
449 * lockdep annotations in xfs_lock_inumorder. These are all checked by
450 * the asserts.
451 */
452 ASSERT(ips && inodes >= 2 && inodes <= 5);
453 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
454 XFS_ILOCK_EXCL));
455 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
456 XFS_ILOCK_SHARED)));
457 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
458 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
459 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
460 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
461
462 if (lock_mode & XFS_IOLOCK_EXCL) {
463 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
464 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
465 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
466
467 try_lock = 0;
468 i = 0;
469again:
470 for (; i < inodes; i++) {
471 ASSERT(ips[i]);
472
473 if (i && (ips[i] == ips[i - 1])) /* Already locked */
474 continue;
475
476 /*
477 * If try_lock is not set yet, make sure all locked inodes are
478 * not in the AIL. If any are, set try_lock to be used later.
479 */
480 if (!try_lock) {
481 for (j = (i - 1); j >= 0 && !try_lock; j--) {
482 lp = &ips[j]->i_itemp->ili_item;
483 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
484 try_lock++;
485 }
486 }
487
488 /*
489 * If any of the previous locks we have locked is in the AIL,
490 * we must TRY to get the second and subsequent locks. If
491 * we can't get any, we must release all we have
492 * and try again.
493 */
494 if (!try_lock) {
495 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
496 continue;
497 }
498
499 /* try_lock means we have an inode locked that is in the AIL. */
500 ASSERT(i != 0);
501 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
502 continue;
503
504 /*
505 * Unlock all previous guys and try again. xfs_iunlock will try
506 * to push the tail if the inode is in the AIL.
507 */
508 attempts++;
509 for (j = i - 1; j >= 0; j--) {
510 /*
511 * Check to see if we've already unlocked this one. Not
512 * the first one going back, and the inode ptr is the
513 * same.
514 */
515 if (j != (i - 1) && ips[j] == ips[j + 1])
516 continue;
517
518 xfs_iunlock(ips[j], lock_mode);
519 }
520
521 if ((attempts % 5) == 0) {
522 delay(1); /* Don't just spin the CPU */
523 }
524 i = 0;
525 try_lock = 0;
526 goto again;
527 }
528}
529
530/*
531 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
532 * the mmaplock or the ilock, but not more than one type at a time. If we lock
533 * more than one at a time, lockdep will report false positives saying we have
534 * violated locking orders. The iolock must be double-locked separately since
535 * we use i_rwsem for that. We now support taking one lock EXCL and the other
536 * SHARED.
537 */
538void
539xfs_lock_two_inodes(
540 struct xfs_inode *ip0,
541 uint ip0_mode,
542 struct xfs_inode *ip1,
543 uint ip1_mode)
544{
545 struct xfs_inode *temp;
546 uint mode_temp;
547 int attempts = 0;
548 struct xfs_log_item *lp;
549
550 ASSERT(hweight32(ip0_mode) == 1);
551 ASSERT(hweight32(ip1_mode) == 1);
552 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
553 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
554 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
555 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
556 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
557 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
558 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
559 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
560 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
561 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
562
563 ASSERT(ip0->i_ino != ip1->i_ino);
564
565 if (ip0->i_ino > ip1->i_ino) {
566 temp = ip0;
567 ip0 = ip1;
568 ip1 = temp;
569 mode_temp = ip0_mode;
570 ip0_mode = ip1_mode;
571 ip1_mode = mode_temp;
572 }
573
574 again:
575 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
576
577 /*
578 * If the first lock we have locked is in the AIL, we must TRY to get
579 * the second lock. If we can't get it, we must release the first one
580 * and try again.
581 */
582 lp = &ip0->i_itemp->ili_item;
583 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
584 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
585 xfs_iunlock(ip0, ip0_mode);
586 if ((++attempts % 5) == 0)
587 delay(1); /* Don't just spin the CPU */
588 goto again;
589 }
590 } else {
591 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
592 }
593}
594
595void
596__xfs_iflock(
597 struct xfs_inode *ip)
598{
599 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
600 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
601
602 do {
603 prepare_to_wait_exclusive(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
604 if (xfs_isiflocked(ip))
605 io_schedule();
606 } while (!xfs_iflock_nowait(ip));
607
608 finish_wait(wq, &wait.wq_entry);
609}
610
611STATIC uint
612_xfs_dic2xflags(
613 uint16_t di_flags,
614 uint64_t di_flags2,
615 bool has_attr)
616{
617 uint flags = 0;
618
619 if (di_flags & XFS_DIFLAG_ANY) {
620 if (di_flags & XFS_DIFLAG_REALTIME)
621 flags |= FS_XFLAG_REALTIME;
622 if (di_flags & XFS_DIFLAG_PREALLOC)
623 flags |= FS_XFLAG_PREALLOC;
624 if (di_flags & XFS_DIFLAG_IMMUTABLE)
625 flags |= FS_XFLAG_IMMUTABLE;
626 if (di_flags & XFS_DIFLAG_APPEND)
627 flags |= FS_XFLAG_APPEND;
628 if (di_flags & XFS_DIFLAG_SYNC)
629 flags |= FS_XFLAG_SYNC;
630 if (di_flags & XFS_DIFLAG_NOATIME)
631 flags |= FS_XFLAG_NOATIME;
632 if (di_flags & XFS_DIFLAG_NODUMP)
633 flags |= FS_XFLAG_NODUMP;
634 if (di_flags & XFS_DIFLAG_RTINHERIT)
635 flags |= FS_XFLAG_RTINHERIT;
636 if (di_flags & XFS_DIFLAG_PROJINHERIT)
637 flags |= FS_XFLAG_PROJINHERIT;
638 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
639 flags |= FS_XFLAG_NOSYMLINKS;
640 if (di_flags & XFS_DIFLAG_EXTSIZE)
641 flags |= FS_XFLAG_EXTSIZE;
642 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
643 flags |= FS_XFLAG_EXTSZINHERIT;
644 if (di_flags & XFS_DIFLAG_NODEFRAG)
645 flags |= FS_XFLAG_NODEFRAG;
646 if (di_flags & XFS_DIFLAG_FILESTREAM)
647 flags |= FS_XFLAG_FILESTREAM;
648 }
649
650 if (di_flags2 & XFS_DIFLAG2_ANY) {
651 if (di_flags2 & XFS_DIFLAG2_DAX)
652 flags |= FS_XFLAG_DAX;
653 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
654 flags |= FS_XFLAG_COWEXTSIZE;
655 }
656
657 if (has_attr)
658 flags |= FS_XFLAG_HASATTR;
659
660 return flags;
661}
662
663uint
664xfs_ip2xflags(
665 struct xfs_inode *ip)
666{
667 struct xfs_icdinode *dic = &ip->i_d;
668
669 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
670}
671
672/*
673 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
674 * is allowed, otherwise it has to be an exact match. If a CI match is found,
675 * ci_name->name will point to a the actual name (caller must free) or
676 * will be set to NULL if an exact match is found.
677 */
678int
679xfs_lookup(
680 xfs_inode_t *dp,
681 struct xfs_name *name,
682 xfs_inode_t **ipp,
683 struct xfs_name *ci_name)
684{
685 xfs_ino_t inum;
686 int error;
687
688 trace_xfs_lookup(dp, name);
689
690 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
691 return -EIO;
692
693 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
694 if (error)
695 goto out_unlock;
696
697 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
698 if (error)
699 goto out_free_name;
700
701 return 0;
702
703out_free_name:
704 if (ci_name)
705 kmem_free(ci_name->name);
706out_unlock:
707 *ipp = NULL;
708 return error;
709}
710
711/*
712 * Allocate an inode on disk and return a copy of its in-core version.
713 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
714 * appropriately within the inode. The uid and gid for the inode are
715 * set according to the contents of the given cred structure.
716 *
717 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
718 * has a free inode available, call xfs_iget() to obtain the in-core
719 * version of the allocated inode. Finally, fill in the inode and
720 * log its initial contents. In this case, ialloc_context would be
721 * set to NULL.
722 *
723 * If xfs_dialloc() does not have an available inode, it will replenish
724 * its supply by doing an allocation. Since we can only do one
725 * allocation within a transaction without deadlocks, we must commit
726 * the current transaction before returning the inode itself.
727 * In this case, therefore, we will set ialloc_context and return.
728 * The caller should then commit the current transaction, start a new
729 * transaction, and call xfs_ialloc() again to actually get the inode.
730 *
731 * To ensure that some other process does not grab the inode that
732 * was allocated during the first call to xfs_ialloc(), this routine
733 * also returns the [locked] bp pointing to the head of the freelist
734 * as ialloc_context. The caller should hold this buffer across
735 * the commit and pass it back into this routine on the second call.
736 *
737 * If we are allocating quota inodes, we do not have a parent inode
738 * to attach to or associate with (i.e. pip == NULL) because they
739 * are not linked into the directory structure - they are attached
740 * directly to the superblock - and so have no parent.
741 */
742static int
743xfs_ialloc(
744 xfs_trans_t *tp,
745 xfs_inode_t *pip,
746 umode_t mode,
747 xfs_nlink_t nlink,
748 dev_t rdev,
749 prid_t prid,
750 xfs_buf_t **ialloc_context,
751 xfs_inode_t **ipp)
752{
753 struct xfs_mount *mp = tp->t_mountp;
754 xfs_ino_t ino;
755 xfs_inode_t *ip;
756 uint flags;
757 int error;
758 struct timespec64 tv;
759 struct inode *inode;
760
761 /*
762 * Call the space management code to pick
763 * the on-disk inode to be allocated.
764 */
765 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode,
766 ialloc_context, &ino);
767 if (error)
768 return error;
769 if (*ialloc_context || ino == NULLFSINO) {
770 *ipp = NULL;
771 return 0;
772 }
773 ASSERT(*ialloc_context == NULL);
774
775 /*
776 * Protect against obviously corrupt allocation btree records. Later
777 * xfs_iget checks will catch re-allocation of other active in-memory
778 * and on-disk inodes. If we don't catch reallocating the parent inode
779 * here we will deadlock in xfs_iget() so we have to do these checks
780 * first.
781 */
782 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
783 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
784 return -EFSCORRUPTED;
785 }
786
787 /*
788 * Get the in-core inode with the lock held exclusively.
789 * This is because we're setting fields here we need
790 * to prevent others from looking at until we're done.
791 */
792 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
793 XFS_ILOCK_EXCL, &ip);
794 if (error)
795 return error;
796 ASSERT(ip != NULL);
797 inode = VFS_I(ip);
798
799 /*
800 * We always convert v1 inodes to v2 now - we only support filesystems
801 * with >= v2 inode capability, so there is no reason for ever leaving
802 * an inode in v1 format.
803 */
804 if (ip->i_d.di_version == 1)
805 ip->i_d.di_version = 2;
806
807 inode->i_mode = mode;
808 set_nlink(inode, nlink);
809 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
810 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
811 inode->i_rdev = rdev;
812 xfs_set_projid(ip, prid);
813
814 if (pip && XFS_INHERIT_GID(pip)) {
815 ip->i_d.di_gid = pip->i_d.di_gid;
816 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
817 inode->i_mode |= S_ISGID;
818 }
819
820 /*
821 * If the group ID of the new file does not match the effective group
822 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
823 * (and only if the irix_sgid_inherit compatibility variable is set).
824 */
825 if ((irix_sgid_inherit) &&
826 (inode->i_mode & S_ISGID) &&
827 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
828 inode->i_mode &= ~S_ISGID;
829
830 ip->i_d.di_size = 0;
831 ip->i_d.di_nextents = 0;
832 ASSERT(ip->i_d.di_nblocks == 0);
833
834 tv = current_time(inode);
835 inode->i_mtime = tv;
836 inode->i_atime = tv;
837 inode->i_ctime = tv;
838
839 ip->i_d.di_extsize = 0;
840 ip->i_d.di_dmevmask = 0;
841 ip->i_d.di_dmstate = 0;
842 ip->i_d.di_flags = 0;
843
844 if (ip->i_d.di_version == 3) {
845 inode_set_iversion(inode, 1);
846 ip->i_d.di_flags2 = 0;
847 ip->i_d.di_cowextsize = 0;
848 ip->i_d.di_crtime.t_sec = (int32_t)tv.tv_sec;
849 ip->i_d.di_crtime.t_nsec = (int32_t)tv.tv_nsec;
850 }
851
852
853 flags = XFS_ILOG_CORE;
854 switch (mode & S_IFMT) {
855 case S_IFIFO:
856 case S_IFCHR:
857 case S_IFBLK:
858 case S_IFSOCK:
859 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
860 ip->i_df.if_flags = 0;
861 flags |= XFS_ILOG_DEV;
862 break;
863 case S_IFREG:
864 case S_IFDIR:
865 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
866 uint di_flags = 0;
867
868 if (S_ISDIR(mode)) {
869 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
870 di_flags |= XFS_DIFLAG_RTINHERIT;
871 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
872 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
873 ip->i_d.di_extsize = pip->i_d.di_extsize;
874 }
875 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
876 di_flags |= XFS_DIFLAG_PROJINHERIT;
877 } else if (S_ISREG(mode)) {
878 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
879 di_flags |= XFS_DIFLAG_REALTIME;
880 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
881 di_flags |= XFS_DIFLAG_EXTSIZE;
882 ip->i_d.di_extsize = pip->i_d.di_extsize;
883 }
884 }
885 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
886 xfs_inherit_noatime)
887 di_flags |= XFS_DIFLAG_NOATIME;
888 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
889 xfs_inherit_nodump)
890 di_flags |= XFS_DIFLAG_NODUMP;
891 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
892 xfs_inherit_sync)
893 di_flags |= XFS_DIFLAG_SYNC;
894 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
895 xfs_inherit_nosymlinks)
896 di_flags |= XFS_DIFLAG_NOSYMLINKS;
897 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
898 xfs_inherit_nodefrag)
899 di_flags |= XFS_DIFLAG_NODEFRAG;
900 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
901 di_flags |= XFS_DIFLAG_FILESTREAM;
902
903 ip->i_d.di_flags |= di_flags;
904 }
905 if (pip &&
906 (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
907 pip->i_d.di_version == 3 &&
908 ip->i_d.di_version == 3) {
909 uint64_t di_flags2 = 0;
910
911 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
912 di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
913 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
914 }
915 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
916 di_flags2 |= XFS_DIFLAG2_DAX;
917
918 ip->i_d.di_flags2 |= di_flags2;
919 }
920 /* FALLTHROUGH */
921 case S_IFLNK:
922 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
923 ip->i_df.if_flags = XFS_IFEXTENTS;
924 ip->i_df.if_bytes = 0;
925 ip->i_df.if_u1.if_root = NULL;
926 break;
927 default:
928 ASSERT(0);
929 }
930 /*
931 * Attribute fork settings for new inode.
932 */
933 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
934 ip->i_d.di_anextents = 0;
935
936 /*
937 * Log the new values stuffed into the inode.
938 */
939 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
940 xfs_trans_log_inode(tp, ip, flags);
941
942 /* now that we have an i_mode we can setup the inode structure */
943 xfs_setup_inode(ip);
944
945 *ipp = ip;
946 return 0;
947}
948
949/*
950 * Allocates a new inode from disk and return a pointer to the
951 * incore copy. This routine will internally commit the current
952 * transaction and allocate a new one if the Space Manager needed
953 * to do an allocation to replenish the inode free-list.
954 *
955 * This routine is designed to be called from xfs_create and
956 * xfs_create_dir.
957 *
958 */
959int
960xfs_dir_ialloc(
961 xfs_trans_t **tpp, /* input: current transaction;
962 output: may be a new transaction. */
963 xfs_inode_t *dp, /* directory within whose allocate
964 the inode. */
965 umode_t mode,
966 xfs_nlink_t nlink,
967 dev_t rdev,
968 prid_t prid, /* project id */
969 xfs_inode_t **ipp) /* pointer to inode; it will be
970 locked. */
971{
972 xfs_trans_t *tp;
973 xfs_inode_t *ip;
974 xfs_buf_t *ialloc_context = NULL;
975 int code;
976 void *dqinfo;
977 uint tflags;
978
979 tp = *tpp;
980 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
981
982 /*
983 * xfs_ialloc will return a pointer to an incore inode if
984 * the Space Manager has an available inode on the free
985 * list. Otherwise, it will do an allocation and replenish
986 * the freelist. Since we can only do one allocation per
987 * transaction without deadlocks, we will need to commit the
988 * current transaction and start a new one. We will then
989 * need to call xfs_ialloc again to get the inode.
990 *
991 * If xfs_ialloc did an allocation to replenish the freelist,
992 * it returns the bp containing the head of the freelist as
993 * ialloc_context. We will hold a lock on it across the
994 * transaction commit so that no other process can steal
995 * the inode(s) that we've just allocated.
996 */
997 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, &ialloc_context,
998 &ip);
999
1000 /*
1001 * Return an error if we were unable to allocate a new inode.
1002 * This should only happen if we run out of space on disk or
1003 * encounter a disk error.
1004 */
1005 if (code) {
1006 *ipp = NULL;
1007 return code;
1008 }
1009 if (!ialloc_context && !ip) {
1010 *ipp = NULL;
1011 return -ENOSPC;
1012 }
1013
1014 /*
1015 * If the AGI buffer is non-NULL, then we were unable to get an
1016 * inode in one operation. We need to commit the current
1017 * transaction and call xfs_ialloc() again. It is guaranteed
1018 * to succeed the second time.
1019 */
1020 if (ialloc_context) {
1021 /*
1022 * Normally, xfs_trans_commit releases all the locks.
1023 * We call bhold to hang on to the ialloc_context across
1024 * the commit. Holding this buffer prevents any other
1025 * processes from doing any allocations in this
1026 * allocation group.
1027 */
1028 xfs_trans_bhold(tp, ialloc_context);
1029
1030 /*
1031 * We want the quota changes to be associated with the next
1032 * transaction, NOT this one. So, detach the dqinfo from this
1033 * and attach it to the next transaction.
1034 */
1035 dqinfo = NULL;
1036 tflags = 0;
1037 if (tp->t_dqinfo) {
1038 dqinfo = (void *)tp->t_dqinfo;
1039 tp->t_dqinfo = NULL;
1040 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1041 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1042 }
1043
1044 code = xfs_trans_roll(&tp);
1045
1046 /*
1047 * Re-attach the quota info that we detached from prev trx.
1048 */
1049 if (dqinfo) {
1050 tp->t_dqinfo = dqinfo;
1051 tp->t_flags |= tflags;
1052 }
1053
1054 if (code) {
1055 xfs_buf_relse(ialloc_context);
1056 *tpp = tp;
1057 *ipp = NULL;
1058 return code;
1059 }
1060 xfs_trans_bjoin(tp, ialloc_context);
1061
1062 /*
1063 * Call ialloc again. Since we've locked out all
1064 * other allocations in this allocation group,
1065 * this call should always succeed.
1066 */
1067 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1068 &ialloc_context, &ip);
1069
1070 /*
1071 * If we get an error at this point, return to the caller
1072 * so that the current transaction can be aborted.
1073 */
1074 if (code) {
1075 *tpp = tp;
1076 *ipp = NULL;
1077 return code;
1078 }
1079 ASSERT(!ialloc_context && ip);
1080
1081 }
1082
1083 *ipp = ip;
1084 *tpp = tp;
1085
1086 return 0;
1087}
1088
1089/*
1090 * Decrement the link count on an inode & log the change. If this causes the
1091 * link count to go to zero, move the inode to AGI unlinked list so that it can
1092 * be freed when the last active reference goes away via xfs_inactive().
1093 */
1094static int /* error */
1095xfs_droplink(
1096 xfs_trans_t *tp,
1097 xfs_inode_t *ip)
1098{
1099 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1100
1101 drop_nlink(VFS_I(ip));
1102 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1103
1104 if (VFS_I(ip)->i_nlink)
1105 return 0;
1106
1107 return xfs_iunlink(tp, ip);
1108}
1109
1110/*
1111 * Increment the link count on an inode & log the change.
1112 */
1113static void
1114xfs_bumplink(
1115 xfs_trans_t *tp,
1116 xfs_inode_t *ip)
1117{
1118 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1119
1120 ASSERT(ip->i_d.di_version > 1);
1121 inc_nlink(VFS_I(ip));
1122 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1123}
1124
1125int
1126xfs_create(
1127 xfs_inode_t *dp,
1128 struct xfs_name *name,
1129 umode_t mode,
1130 dev_t rdev,
1131 xfs_inode_t **ipp)
1132{
1133 int is_dir = S_ISDIR(mode);
1134 struct xfs_mount *mp = dp->i_mount;
1135 struct xfs_inode *ip = NULL;
1136 struct xfs_trans *tp = NULL;
1137 int error;
1138 bool unlock_dp_on_error = false;
1139 prid_t prid;
1140 struct xfs_dquot *udqp = NULL;
1141 struct xfs_dquot *gdqp = NULL;
1142 struct xfs_dquot *pdqp = NULL;
1143 struct xfs_trans_res *tres;
1144 uint resblks;
1145
1146 trace_xfs_create(dp, name);
1147
1148 if (XFS_FORCED_SHUTDOWN(mp))
1149 return -EIO;
1150
1151 prid = xfs_get_initial_prid(dp);
1152
1153 /*
1154 * Make sure that we have allocated dquot(s) on disk.
1155 */
1156 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1157 xfs_kgid_to_gid(current_fsgid()), prid,
1158 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1159 &udqp, &gdqp, &pdqp);
1160 if (error)
1161 return error;
1162
1163 if (is_dir) {
1164 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1165 tres = &M_RES(mp)->tr_mkdir;
1166 } else {
1167 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1168 tres = &M_RES(mp)->tr_create;
1169 }
1170
1171 /*
1172 * Initially assume that the file does not exist and
1173 * reserve the resources for that case. If that is not
1174 * the case we'll drop the one we have and get a more
1175 * appropriate transaction later.
1176 */
1177 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1178 if (error == -ENOSPC) {
1179 /* flush outstanding delalloc blocks and retry */
1180 xfs_flush_inodes(mp);
1181 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1182 }
1183 if (error)
1184 goto out_release_inode;
1185
1186 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1187 unlock_dp_on_error = true;
1188
1189 /*
1190 * Reserve disk quota and the inode.
1191 */
1192 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1193 pdqp, resblks, 1, 0);
1194 if (error)
1195 goto out_trans_cancel;
1196
1197 /*
1198 * A newly created regular or special file just has one directory
1199 * entry pointing to them, but a directory also the "." entry
1200 * pointing to itself.
1201 */
1202 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev, prid, &ip);
1203 if (error)
1204 goto out_trans_cancel;
1205
1206 /*
1207 * Now we join the directory inode to the transaction. We do not do it
1208 * earlier because xfs_dir_ialloc might commit the previous transaction
1209 * (and release all the locks). An error from here on will result in
1210 * the transaction cancel unlocking dp so don't do it explicitly in the
1211 * error path.
1212 */
1213 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1214 unlock_dp_on_error = false;
1215
1216 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1217 resblks ?
1218 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1219 if (error) {
1220 ASSERT(error != -ENOSPC);
1221 goto out_trans_cancel;
1222 }
1223 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1224 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1225
1226 if (is_dir) {
1227 error = xfs_dir_init(tp, ip, dp);
1228 if (error)
1229 goto out_trans_cancel;
1230
1231 xfs_bumplink(tp, dp);
1232 }
1233
1234 /*
1235 * If this is a synchronous mount, make sure that the
1236 * create transaction goes to disk before returning to
1237 * the user.
1238 */
1239 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1240 xfs_trans_set_sync(tp);
1241
1242 /*
1243 * Attach the dquot(s) to the inodes and modify them incore.
1244 * These ids of the inode couldn't have changed since the new
1245 * inode has been locked ever since it was created.
1246 */
1247 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1248
1249 error = xfs_trans_commit(tp);
1250 if (error)
1251 goto out_release_inode;
1252
1253 xfs_qm_dqrele(udqp);
1254 xfs_qm_dqrele(gdqp);
1255 xfs_qm_dqrele(pdqp);
1256
1257 *ipp = ip;
1258 return 0;
1259
1260 out_trans_cancel:
1261 xfs_trans_cancel(tp);
1262 out_release_inode:
1263 /*
1264 * Wait until after the current transaction is aborted to finish the
1265 * setup of the inode and release the inode. This prevents recursive
1266 * transactions and deadlocks from xfs_inactive.
1267 */
1268 if (ip) {
1269 xfs_finish_inode_setup(ip);
1270 xfs_irele(ip);
1271 }
1272
1273 xfs_qm_dqrele(udqp);
1274 xfs_qm_dqrele(gdqp);
1275 xfs_qm_dqrele(pdqp);
1276
1277 if (unlock_dp_on_error)
1278 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1279 return error;
1280}
1281
1282int
1283xfs_create_tmpfile(
1284 struct xfs_inode *dp,
1285 umode_t mode,
1286 struct xfs_inode **ipp)
1287{
1288 struct xfs_mount *mp = dp->i_mount;
1289 struct xfs_inode *ip = NULL;
1290 struct xfs_trans *tp = NULL;
1291 int error;
1292 prid_t prid;
1293 struct xfs_dquot *udqp = NULL;
1294 struct xfs_dquot *gdqp = NULL;
1295 struct xfs_dquot *pdqp = NULL;
1296 struct xfs_trans_res *tres;
1297 uint resblks;
1298
1299 if (XFS_FORCED_SHUTDOWN(mp))
1300 return -EIO;
1301
1302 prid = xfs_get_initial_prid(dp);
1303
1304 /*
1305 * Make sure that we have allocated dquot(s) on disk.
1306 */
1307 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1308 xfs_kgid_to_gid(current_fsgid()), prid,
1309 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1310 &udqp, &gdqp, &pdqp);
1311 if (error)
1312 return error;
1313
1314 resblks = XFS_IALLOC_SPACE_RES(mp);
1315 tres = &M_RES(mp)->tr_create_tmpfile;
1316
1317 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1318 if (error)
1319 goto out_release_inode;
1320
1321 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1322 pdqp, resblks, 1, 0);
1323 if (error)
1324 goto out_trans_cancel;
1325
1326 error = xfs_dir_ialloc(&tp, dp, mode, 0, 0, prid, &ip);
1327 if (error)
1328 goto out_trans_cancel;
1329
1330 if (mp->m_flags & XFS_MOUNT_WSYNC)
1331 xfs_trans_set_sync(tp);
1332
1333 /*
1334 * Attach the dquot(s) to the inodes and modify them incore.
1335 * These ids of the inode couldn't have changed since the new
1336 * inode has been locked ever since it was created.
1337 */
1338 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1339
1340 error = xfs_iunlink(tp, ip);
1341 if (error)
1342 goto out_trans_cancel;
1343
1344 error = xfs_trans_commit(tp);
1345 if (error)
1346 goto out_release_inode;
1347
1348 xfs_qm_dqrele(udqp);
1349 xfs_qm_dqrele(gdqp);
1350 xfs_qm_dqrele(pdqp);
1351
1352 *ipp = ip;
1353 return 0;
1354
1355 out_trans_cancel:
1356 xfs_trans_cancel(tp);
1357 out_release_inode:
1358 /*
1359 * Wait until after the current transaction is aborted to finish the
1360 * setup of the inode and release the inode. This prevents recursive
1361 * transactions and deadlocks from xfs_inactive.
1362 */
1363 if (ip) {
1364 xfs_finish_inode_setup(ip);
1365 xfs_irele(ip);
1366 }
1367
1368 xfs_qm_dqrele(udqp);
1369 xfs_qm_dqrele(gdqp);
1370 xfs_qm_dqrele(pdqp);
1371
1372 return error;
1373}
1374
1375int
1376xfs_link(
1377 xfs_inode_t *tdp,
1378 xfs_inode_t *sip,
1379 struct xfs_name *target_name)
1380{
1381 xfs_mount_t *mp = tdp->i_mount;
1382 xfs_trans_t *tp;
1383 int error;
1384 int resblks;
1385
1386 trace_xfs_link(tdp, target_name);
1387
1388 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1389
1390 if (XFS_FORCED_SHUTDOWN(mp))
1391 return -EIO;
1392
1393 error = xfs_qm_dqattach(sip);
1394 if (error)
1395 goto std_return;
1396
1397 error = xfs_qm_dqattach(tdp);
1398 if (error)
1399 goto std_return;
1400
1401 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1402 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1403 if (error == -ENOSPC) {
1404 resblks = 0;
1405 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1406 }
1407 if (error)
1408 goto std_return;
1409
1410 xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1411
1412 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1413 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1414
1415 /*
1416 * If we are using project inheritance, we only allow hard link
1417 * creation in our tree when the project IDs are the same; else
1418 * the tree quota mechanism could be circumvented.
1419 */
1420 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1421 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1422 error = -EXDEV;
1423 goto error_return;
1424 }
1425
1426 if (!resblks) {
1427 error = xfs_dir_canenter(tp, tdp, target_name);
1428 if (error)
1429 goto error_return;
1430 }
1431
1432 /*
1433 * Handle initial link state of O_TMPFILE inode
1434 */
1435 if (VFS_I(sip)->i_nlink == 0) {
1436 error = xfs_iunlink_remove(tp, sip);
1437 if (error)
1438 goto error_return;
1439 }
1440
1441 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1442 resblks);
1443 if (error)
1444 goto error_return;
1445 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1446 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1447
1448 xfs_bumplink(tp, sip);
1449
1450 /*
1451 * If this is a synchronous mount, make sure that the
1452 * link transaction goes to disk before returning to
1453 * the user.
1454 */
1455 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1456 xfs_trans_set_sync(tp);
1457
1458 return xfs_trans_commit(tp);
1459
1460 error_return:
1461 xfs_trans_cancel(tp);
1462 std_return:
1463 return error;
1464}
1465
1466/* Clear the reflink flag and the cowblocks tag if possible. */
1467static void
1468xfs_itruncate_clear_reflink_flags(
1469 struct xfs_inode *ip)
1470{
1471 struct xfs_ifork *dfork;
1472 struct xfs_ifork *cfork;
1473
1474 if (!xfs_is_reflink_inode(ip))
1475 return;
1476 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1477 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1478 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1479 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1480 if (cfork->if_bytes == 0)
1481 xfs_inode_clear_cowblocks_tag(ip);
1482}
1483
1484/*
1485 * Free up the underlying blocks past new_size. The new size must be smaller
1486 * than the current size. This routine can be used both for the attribute and
1487 * data fork, and does not modify the inode size, which is left to the caller.
1488 *
1489 * The transaction passed to this routine must have made a permanent log
1490 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1491 * given transaction and start new ones, so make sure everything involved in
1492 * the transaction is tidy before calling here. Some transaction will be
1493 * returned to the caller to be committed. The incoming transaction must
1494 * already include the inode, and both inode locks must be held exclusively.
1495 * The inode must also be "held" within the transaction. On return the inode
1496 * will be "held" within the returned transaction. This routine does NOT
1497 * require any disk space to be reserved for it within the transaction.
1498 *
1499 * If we get an error, we must return with the inode locked and linked into the
1500 * current transaction. This keeps things simple for the higher level code,
1501 * because it always knows that the inode is locked and held in the transaction
1502 * that returns to it whether errors occur or not. We don't mark the inode
1503 * dirty on error so that transactions can be easily aborted if possible.
1504 */
1505int
1506xfs_itruncate_extents_flags(
1507 struct xfs_trans **tpp,
1508 struct xfs_inode *ip,
1509 int whichfork,
1510 xfs_fsize_t new_size,
1511 int flags)
1512{
1513 struct xfs_mount *mp = ip->i_mount;
1514 struct xfs_trans *tp = *tpp;
1515 xfs_fileoff_t first_unmap_block;
1516 xfs_fileoff_t last_block;
1517 xfs_filblks_t unmap_len;
1518 int error = 0;
1519 int done = 0;
1520
1521 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1522 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1523 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1524 ASSERT(new_size <= XFS_ISIZE(ip));
1525 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1526 ASSERT(ip->i_itemp != NULL);
1527 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1528 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1529
1530 trace_xfs_itruncate_extents_start(ip, new_size);
1531
1532 flags |= xfs_bmapi_aflag(whichfork);
1533
1534 /*
1535 * Since it is possible for space to become allocated beyond
1536 * the end of the file (in a crash where the space is allocated
1537 * but the inode size is not yet updated), simply remove any
1538 * blocks which show up between the new EOF and the maximum
1539 * possible file size. If the first block to be removed is
1540 * beyond the maximum file size (ie it is the same as last_block),
1541 * then there is nothing to do.
1542 */
1543 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1544 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1545 if (first_unmap_block == last_block)
1546 return 0;
1547
1548 ASSERT(first_unmap_block < last_block);
1549 unmap_len = last_block - first_unmap_block + 1;
1550 while (!done) {
1551 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1552 error = xfs_bunmapi(tp, ip, first_unmap_block, unmap_len, flags,
1553 XFS_ITRUNC_MAX_EXTENTS, &done);
1554 if (error)
1555 goto out;
1556
1557 /*
1558 * Duplicate the transaction that has the permanent
1559 * reservation and commit the old transaction.
1560 */
1561 error = xfs_defer_finish(&tp);
1562 if (error)
1563 goto out;
1564
1565 error = xfs_trans_roll_inode(&tp, ip);
1566 if (error)
1567 goto out;
1568 }
1569
1570 if (whichfork == XFS_DATA_FORK) {
1571 /* Remove all pending CoW reservations. */
1572 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1573 first_unmap_block, last_block, true);
1574 if (error)
1575 goto out;
1576
1577 xfs_itruncate_clear_reflink_flags(ip);
1578 }
1579
1580 /*
1581 * Always re-log the inode so that our permanent transaction can keep
1582 * on rolling it forward in the log.
1583 */
1584 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1585
1586 trace_xfs_itruncate_extents_end(ip, new_size);
1587
1588out:
1589 *tpp = tp;
1590 return error;
1591}
1592
1593int
1594xfs_release(
1595 xfs_inode_t *ip)
1596{
1597 xfs_mount_t *mp = ip->i_mount;
1598 int error;
1599
1600 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1601 return 0;
1602
1603 /* If this is a read-only mount, don't do this (would generate I/O) */
1604 if (mp->m_flags & XFS_MOUNT_RDONLY)
1605 return 0;
1606
1607 if (!XFS_FORCED_SHUTDOWN(mp)) {
1608 int truncated;
1609
1610 /*
1611 * If we previously truncated this file and removed old data
1612 * in the process, we want to initiate "early" writeout on
1613 * the last close. This is an attempt to combat the notorious
1614 * NULL files problem which is particularly noticeable from a
1615 * truncate down, buffered (re-)write (delalloc), followed by
1616 * a crash. What we are effectively doing here is
1617 * significantly reducing the time window where we'd otherwise
1618 * be exposed to that problem.
1619 */
1620 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1621 if (truncated) {
1622 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1623 if (ip->i_delayed_blks > 0) {
1624 error = filemap_flush(VFS_I(ip)->i_mapping);
1625 if (error)
1626 return error;
1627 }
1628 }
1629 }
1630
1631 if (VFS_I(ip)->i_nlink == 0)
1632 return 0;
1633
1634 if (xfs_can_free_eofblocks(ip, false)) {
1635
1636 /*
1637 * Check if the inode is being opened, written and closed
1638 * frequently and we have delayed allocation blocks outstanding
1639 * (e.g. streaming writes from the NFS server), truncating the
1640 * blocks past EOF will cause fragmentation to occur.
1641 *
1642 * In this case don't do the truncation, but we have to be
1643 * careful how we detect this case. Blocks beyond EOF show up as
1644 * i_delayed_blks even when the inode is clean, so we need to
1645 * truncate them away first before checking for a dirty release.
1646 * Hence on the first dirty close we will still remove the
1647 * speculative allocation, but after that we will leave it in
1648 * place.
1649 */
1650 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1651 return 0;
1652 /*
1653 * If we can't get the iolock just skip truncating the blocks
1654 * past EOF because we could deadlock with the mmap_sem
1655 * otherwise. We'll get another chance to drop them once the
1656 * last reference to the inode is dropped, so we'll never leak
1657 * blocks permanently.
1658 */
1659 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1660 error = xfs_free_eofblocks(ip);
1661 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1662 if (error)
1663 return error;
1664 }
1665
1666 /* delalloc blocks after truncation means it really is dirty */
1667 if (ip->i_delayed_blks)
1668 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1669 }
1670 return 0;
1671}
1672
1673/*
1674 * xfs_inactive_truncate
1675 *
1676 * Called to perform a truncate when an inode becomes unlinked.
1677 */
1678STATIC int
1679xfs_inactive_truncate(
1680 struct xfs_inode *ip)
1681{
1682 struct xfs_mount *mp = ip->i_mount;
1683 struct xfs_trans *tp;
1684 int error;
1685
1686 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1687 if (error) {
1688 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1689 return error;
1690 }
1691 xfs_ilock(ip, XFS_ILOCK_EXCL);
1692 xfs_trans_ijoin(tp, ip, 0);
1693
1694 /*
1695 * Log the inode size first to prevent stale data exposure in the event
1696 * of a system crash before the truncate completes. See the related
1697 * comment in xfs_vn_setattr_size() for details.
1698 */
1699 ip->i_d.di_size = 0;
1700 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1701
1702 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1703 if (error)
1704 goto error_trans_cancel;
1705
1706 ASSERT(ip->i_d.di_nextents == 0);
1707
1708 error = xfs_trans_commit(tp);
1709 if (error)
1710 goto error_unlock;
1711
1712 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1713 return 0;
1714
1715error_trans_cancel:
1716 xfs_trans_cancel(tp);
1717error_unlock:
1718 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1719 return error;
1720}
1721
1722/*
1723 * xfs_inactive_ifree()
1724 *
1725 * Perform the inode free when an inode is unlinked.
1726 */
1727STATIC int
1728xfs_inactive_ifree(
1729 struct xfs_inode *ip)
1730{
1731 struct xfs_mount *mp = ip->i_mount;
1732 struct xfs_trans *tp;
1733 int error;
1734
1735 /*
1736 * We try to use a per-AG reservation for any block needed by the finobt
1737 * tree, but as the finobt feature predates the per-AG reservation
1738 * support a degraded file system might not have enough space for the
1739 * reservation at mount time. In that case try to dip into the reserved
1740 * pool and pray.
1741 *
1742 * Send a warning if the reservation does happen to fail, as the inode
1743 * now remains allocated and sits on the unlinked list until the fs is
1744 * repaired.
1745 */
1746 if (unlikely(mp->m_finobt_nores)) {
1747 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1748 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1749 &tp);
1750 } else {
1751 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1752 }
1753 if (error) {
1754 if (error == -ENOSPC) {
1755 xfs_warn_ratelimited(mp,
1756 "Failed to remove inode(s) from unlinked list. "
1757 "Please free space, unmount and run xfs_repair.");
1758 } else {
1759 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1760 }
1761 return error;
1762 }
1763
1764 xfs_ilock(ip, XFS_ILOCK_EXCL);
1765 xfs_trans_ijoin(tp, ip, 0);
1766
1767 error = xfs_ifree(tp, ip);
1768 if (error) {
1769 /*
1770 * If we fail to free the inode, shut down. The cancel
1771 * might do that, we need to make sure. Otherwise the
1772 * inode might be lost for a long time or forever.
1773 */
1774 if (!XFS_FORCED_SHUTDOWN(mp)) {
1775 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1776 __func__, error);
1777 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1778 }
1779 xfs_trans_cancel(tp);
1780 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1781 return error;
1782 }
1783
1784 /*
1785 * Credit the quota account(s). The inode is gone.
1786 */
1787 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1788
1789 /*
1790 * Just ignore errors at this point. There is nothing we can do except
1791 * to try to keep going. Make sure it's not a silent error.
1792 */
1793 error = xfs_trans_commit(tp);
1794 if (error)
1795 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1796 __func__, error);
1797
1798 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1799 return 0;
1800}
1801
1802/*
1803 * xfs_inactive
1804 *
1805 * This is called when the vnode reference count for the vnode
1806 * goes to zero. If the file has been unlinked, then it must
1807 * now be truncated. Also, we clear all of the read-ahead state
1808 * kept for the inode here since the file is now closed.
1809 */
1810void
1811xfs_inactive(
1812 xfs_inode_t *ip)
1813{
1814 struct xfs_mount *mp;
1815 int error;
1816 int truncate = 0;
1817
1818 /*
1819 * If the inode is already free, then there can be nothing
1820 * to clean up here.
1821 */
1822 if (VFS_I(ip)->i_mode == 0) {
1823 ASSERT(ip->i_df.if_broot_bytes == 0);
1824 return;
1825 }
1826
1827 mp = ip->i_mount;
1828 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1829
1830 /* If this is a read-only mount, don't do this (would generate I/O) */
1831 if (mp->m_flags & XFS_MOUNT_RDONLY)
1832 return;
1833
1834 /* Try to clean out the cow blocks if there are any. */
1835 if (xfs_inode_has_cow_data(ip))
1836 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1837
1838 if (VFS_I(ip)->i_nlink != 0) {
1839 /*
1840 * force is true because we are evicting an inode from the
1841 * cache. Post-eof blocks must be freed, lest we end up with
1842 * broken free space accounting.
1843 *
1844 * Note: don't bother with iolock here since lockdep complains
1845 * about acquiring it in reclaim context. We have the only
1846 * reference to the inode at this point anyways.
1847 */
1848 if (xfs_can_free_eofblocks(ip, true))
1849 xfs_free_eofblocks(ip);
1850
1851 return;
1852 }
1853
1854 if (S_ISREG(VFS_I(ip)->i_mode) &&
1855 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1856 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1857 truncate = 1;
1858
1859 error = xfs_qm_dqattach(ip);
1860 if (error)
1861 return;
1862
1863 if (S_ISLNK(VFS_I(ip)->i_mode))
1864 error = xfs_inactive_symlink(ip);
1865 else if (truncate)
1866 error = xfs_inactive_truncate(ip);
1867 if (error)
1868 return;
1869
1870 /*
1871 * If there are attributes associated with the file then blow them away
1872 * now. The code calls a routine that recursively deconstructs the
1873 * attribute fork. If also blows away the in-core attribute fork.
1874 */
1875 if (XFS_IFORK_Q(ip)) {
1876 error = xfs_attr_inactive(ip);
1877 if (error)
1878 return;
1879 }
1880
1881 ASSERT(!ip->i_afp);
1882 ASSERT(ip->i_d.di_anextents == 0);
1883 ASSERT(ip->i_d.di_forkoff == 0);
1884
1885 /*
1886 * Free the inode.
1887 */
1888 error = xfs_inactive_ifree(ip);
1889 if (error)
1890 return;
1891
1892 /*
1893 * Release the dquots held by inode, if any.
1894 */
1895 xfs_qm_dqdetach(ip);
1896}
1897
1898/*
1899 * In-Core Unlinked List Lookups
1900 * =============================
1901 *
1902 * Every inode is supposed to be reachable from some other piece of metadata
1903 * with the exception of the root directory. Inodes with a connection to a
1904 * file descriptor but not linked from anywhere in the on-disk directory tree
1905 * are collectively known as unlinked inodes, though the filesystem itself
1906 * maintains links to these inodes so that on-disk metadata are consistent.
1907 *
1908 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1909 * header contains a number of buckets that point to an inode, and each inode
1910 * record has a pointer to the next inode in the hash chain. This
1911 * singly-linked list causes scaling problems in the iunlink remove function
1912 * because we must walk that list to find the inode that points to the inode
1913 * being removed from the unlinked hash bucket list.
1914 *
1915 * What if we modelled the unlinked list as a collection of records capturing
1916 * "X.next_unlinked = Y" relations? If we indexed those records on Y, we'd
1917 * have a fast way to look up unlinked list predecessors, which avoids the
1918 * slow list walk. That's exactly what we do here (in-core) with a per-AG
1919 * rhashtable.
1920 *
1921 * Because this is a backref cache, we ignore operational failures since the
1922 * iunlink code can fall back to the slow bucket walk. The only errors that
1923 * should bubble out are for obviously incorrect situations.
1924 *
1925 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1926 * access or have otherwise provided for concurrency control.
1927 */
1928
1929/* Capture a "X.next_unlinked = Y" relationship. */
1930struct xfs_iunlink {
1931 struct rhash_head iu_rhash_head;
1932 xfs_agino_t iu_agino; /* X */
1933 xfs_agino_t iu_next_unlinked; /* Y */
1934};
1935
1936/* Unlinked list predecessor lookup hashtable construction */
1937static int
1938xfs_iunlink_obj_cmpfn(
1939 struct rhashtable_compare_arg *arg,
1940 const void *obj)
1941{
1942 const xfs_agino_t *key = arg->key;
1943 const struct xfs_iunlink *iu = obj;
1944
1945 if (iu->iu_next_unlinked != *key)
1946 return 1;
1947 return 0;
1948}
1949
1950static const struct rhashtable_params xfs_iunlink_hash_params = {
1951 .min_size = XFS_AGI_UNLINKED_BUCKETS,
1952 .key_len = sizeof(xfs_agino_t),
1953 .key_offset = offsetof(struct xfs_iunlink,
1954 iu_next_unlinked),
1955 .head_offset = offsetof(struct xfs_iunlink, iu_rhash_head),
1956 .automatic_shrinking = true,
1957 .obj_cmpfn = xfs_iunlink_obj_cmpfn,
1958};
1959
1960/*
1961 * Return X, where X.next_unlinked == @agino. Returns NULLAGINO if no such
1962 * relation is found.
1963 */
1964static xfs_agino_t
1965xfs_iunlink_lookup_backref(
1966 struct xfs_perag *pag,
1967 xfs_agino_t agino)
1968{
1969 struct xfs_iunlink *iu;
1970
1971 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1972 xfs_iunlink_hash_params);
1973 return iu ? iu->iu_agino : NULLAGINO;
1974}
1975
1976/*
1977 * Take ownership of an iunlink cache entry and insert it into the hash table.
1978 * If successful, the entry will be owned by the cache; if not, it is freed.
1979 * Either way, the caller does not own @iu after this call.
1980 */
1981static int
1982xfs_iunlink_insert_backref(
1983 struct xfs_perag *pag,
1984 struct xfs_iunlink *iu)
1985{
1986 int error;
1987
1988 error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1989 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1990 /*
1991 * Fail loudly if there already was an entry because that's a sign of
1992 * corruption of in-memory data. Also fail loudly if we see an error
1993 * code we didn't anticipate from the rhashtable code. Currently we
1994 * only anticipate ENOMEM.
1995 */
1996 if (error) {
1997 WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1998 kmem_free(iu);
1999 }
2000 /*
2001 * Absorb any runtime errors that aren't a result of corruption because
2002 * this is a cache and we can always fall back to bucket list scanning.
2003 */
2004 if (error != 0 && error != -EEXIST)
2005 error = 0;
2006 return error;
2007}
2008
2009/* Remember that @prev_agino.next_unlinked = @this_agino. */
2010static int
2011xfs_iunlink_add_backref(
2012 struct xfs_perag *pag,
2013 xfs_agino_t prev_agino,
2014 xfs_agino_t this_agino)
2015{
2016 struct xfs_iunlink *iu;
2017
2018 if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
2019 return 0;
2020
2021 iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
2022 iu->iu_agino = prev_agino;
2023 iu->iu_next_unlinked = this_agino;
2024
2025 return xfs_iunlink_insert_backref(pag, iu);
2026}
2027
2028/*
2029 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
2030 * If @next_unlinked is NULLAGINO, we drop the backref and exit. If there
2031 * wasn't any such entry then we don't bother.
2032 */
2033static int
2034xfs_iunlink_change_backref(
2035 struct xfs_perag *pag,
2036 xfs_agino_t agino,
2037 xfs_agino_t next_unlinked)
2038{
2039 struct xfs_iunlink *iu;
2040 int error;
2041
2042 /* Look up the old entry; if there wasn't one then exit. */
2043 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
2044 xfs_iunlink_hash_params);
2045 if (!iu)
2046 return 0;
2047
2048 /*
2049 * Remove the entry. This shouldn't ever return an error, but if we
2050 * couldn't remove the old entry we don't want to add it again to the
2051 * hash table, and if the entry disappeared on us then someone's
2052 * violated the locking rules and we need to fail loudly. Either way
2053 * we cannot remove the inode because internal state is or would have
2054 * been corrupt.
2055 */
2056 error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
2057 &iu->iu_rhash_head, xfs_iunlink_hash_params);
2058 if (error)
2059 return error;
2060
2061 /* If there is no new next entry just free our item and return. */
2062 if (next_unlinked == NULLAGINO) {
2063 kmem_free(iu);
2064 return 0;
2065 }
2066
2067 /* Update the entry and re-add it to the hash table. */
2068 iu->iu_next_unlinked = next_unlinked;
2069 return xfs_iunlink_insert_backref(pag, iu);
2070}
2071
2072/* Set up the in-core predecessor structures. */
2073int
2074xfs_iunlink_init(
2075 struct xfs_perag *pag)
2076{
2077 return rhashtable_init(&pag->pagi_unlinked_hash,
2078 &xfs_iunlink_hash_params);
2079}
2080
2081/* Free the in-core predecessor structures. */
2082static void
2083xfs_iunlink_free_item(
2084 void *ptr,
2085 void *arg)
2086{
2087 struct xfs_iunlink *iu = ptr;
2088 bool *freed_anything = arg;
2089
2090 *freed_anything = true;
2091 kmem_free(iu);
2092}
2093
2094void
2095xfs_iunlink_destroy(
2096 struct xfs_perag *pag)
2097{
2098 bool freed_anything = false;
2099
2100 rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
2101 xfs_iunlink_free_item, &freed_anything);
2102
2103 ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
2104}
2105
2106/*
2107 * Point the AGI unlinked bucket at an inode and log the results. The caller
2108 * is responsible for validating the old value.
2109 */
2110STATIC int
2111xfs_iunlink_update_bucket(
2112 struct xfs_trans *tp,
2113 xfs_agnumber_t agno,
2114 struct xfs_buf *agibp,
2115 unsigned int bucket_index,
2116 xfs_agino_t new_agino)
2117{
2118 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
2119 xfs_agino_t old_value;
2120 int offset;
2121
2122 ASSERT(xfs_verify_agino_or_null(tp->t_mountp, agno, new_agino));
2123
2124 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2125 trace_xfs_iunlink_update_bucket(tp->t_mountp, agno, bucket_index,
2126 old_value, new_agino);
2127
2128 /*
2129 * We should never find the head of the list already set to the value
2130 * passed in because either we're adding or removing ourselves from the
2131 * head of the list.
2132 */
2133 if (old_value == new_agino)
2134 return -EFSCORRUPTED;
2135
2136 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2137 offset = offsetof(struct xfs_agi, agi_unlinked) +
2138 (sizeof(xfs_agino_t) * bucket_index);
2139 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2140 return 0;
2141}
2142
2143/* Set an on-disk inode's next_unlinked pointer. */
2144STATIC void
2145xfs_iunlink_update_dinode(
2146 struct xfs_trans *tp,
2147 xfs_agnumber_t agno,
2148 xfs_agino_t agino,
2149 struct xfs_buf *ibp,
2150 struct xfs_dinode *dip,
2151 struct xfs_imap *imap,
2152 xfs_agino_t next_agino)
2153{
2154 struct xfs_mount *mp = tp->t_mountp;
2155 int offset;
2156
2157 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2158
2159 trace_xfs_iunlink_update_dinode(mp, agno, agino,
2160 be32_to_cpu(dip->di_next_unlinked), next_agino);
2161
2162 dip->di_next_unlinked = cpu_to_be32(next_agino);
2163 offset = imap->im_boffset +
2164 offsetof(struct xfs_dinode, di_next_unlinked);
2165
2166 /* need to recalc the inode CRC if appropriate */
2167 xfs_dinode_calc_crc(mp, dip);
2168 xfs_trans_inode_buf(tp, ibp);
2169 xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2170 xfs_inobp_check(mp, ibp);
2171}
2172
2173/* Set an in-core inode's unlinked pointer and return the old value. */
2174STATIC int
2175xfs_iunlink_update_inode(
2176 struct xfs_trans *tp,
2177 struct xfs_inode *ip,
2178 xfs_agnumber_t agno,
2179 xfs_agino_t next_agino,
2180 xfs_agino_t *old_next_agino)
2181{
2182 struct xfs_mount *mp = tp->t_mountp;
2183 struct xfs_dinode *dip;
2184 struct xfs_buf *ibp;
2185 xfs_agino_t old_value;
2186 int error;
2187
2188 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2189
2190 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp, 0, 0);
2191 if (error)
2192 return error;
2193
2194 /* Make sure the old pointer isn't garbage. */
2195 old_value = be32_to_cpu(dip->di_next_unlinked);
2196 if (!xfs_verify_agino_or_null(mp, agno, old_value)) {
2197 error = -EFSCORRUPTED;
2198 goto out;
2199 }
2200
2201 /*
2202 * Since we're updating a linked list, we should never find that the
2203 * current pointer is the same as the new value, unless we're
2204 * terminating the list.
2205 */
2206 *old_next_agino = old_value;
2207 if (old_value == next_agino) {
2208 if (next_agino != NULLAGINO)
2209 error = -EFSCORRUPTED;
2210 goto out;
2211 }
2212
2213 /* Ok, update the new pointer. */
2214 xfs_iunlink_update_dinode(tp, agno, XFS_INO_TO_AGINO(mp, ip->i_ino),
2215 ibp, dip, &ip->i_imap, next_agino);
2216 return 0;
2217out:
2218 xfs_trans_brelse(tp, ibp);
2219 return error;
2220}
2221
2222/*
2223 * This is called when the inode's link count has gone to 0 or we are creating
2224 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2225 *
2226 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2227 * list when the inode is freed.
2228 */
2229STATIC int
2230xfs_iunlink(
2231 struct xfs_trans *tp,
2232 struct xfs_inode *ip)
2233{
2234 struct xfs_mount *mp = tp->t_mountp;
2235 struct xfs_agi *agi;
2236 struct xfs_buf *agibp;
2237 xfs_agino_t next_agino;
2238 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2239 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2240 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2241 int error;
2242
2243 ASSERT(VFS_I(ip)->i_nlink == 0);
2244 ASSERT(VFS_I(ip)->i_mode != 0);
2245 trace_xfs_iunlink(ip);
2246
2247 /* Get the agi buffer first. It ensures lock ordering on the list. */
2248 error = xfs_read_agi(mp, tp, agno, &agibp);
2249 if (error)
2250 return error;
2251 agi = XFS_BUF_TO_AGI(agibp);
2252
2253 /*
2254 * Get the index into the agi hash table for the list this inode will
2255 * go on. Make sure the pointer isn't garbage and that this inode
2256 * isn't already on the list.
2257 */
2258 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2259 if (next_agino == agino ||
2260 !xfs_verify_agino_or_null(mp, agno, next_agino))
2261 return -EFSCORRUPTED;
2262
2263 if (next_agino != NULLAGINO) {
2264 struct xfs_perag *pag;
2265 xfs_agino_t old_agino;
2266
2267 /*
2268 * There is already another inode in the bucket, so point this
2269 * inode to the current head of the list.
2270 */
2271 error = xfs_iunlink_update_inode(tp, ip, agno, next_agino,
2272 &old_agino);
2273 if (error)
2274 return error;
2275 ASSERT(old_agino == NULLAGINO);
2276
2277 /*
2278 * agino has been unlinked, add a backref from the next inode
2279 * back to agino.
2280 */
2281 pag = xfs_perag_get(mp, agno);
2282 error = xfs_iunlink_add_backref(pag, agino, next_agino);
2283 xfs_perag_put(pag);
2284 if (error)
2285 return error;
2286 }
2287
2288 /* Point the head of the list to point to this inode. */
2289 return xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index, agino);
2290}
2291
2292/* Return the imap, dinode pointer, and buffer for an inode. */
2293STATIC int
2294xfs_iunlink_map_ino(
2295 struct xfs_trans *tp,
2296 xfs_agnumber_t agno,
2297 xfs_agino_t agino,
2298 struct xfs_imap *imap,
2299 struct xfs_dinode **dipp,
2300 struct xfs_buf **bpp)
2301{
2302 struct xfs_mount *mp = tp->t_mountp;
2303 int error;
2304
2305 imap->im_blkno = 0;
2306 error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2307 if (error) {
2308 xfs_warn(mp, "%s: xfs_imap returned error %d.",
2309 __func__, error);
2310 return error;
2311 }
2312
2313 error = xfs_imap_to_bp(mp, tp, imap, dipp, bpp, 0, 0);
2314 if (error) {
2315 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2316 __func__, error);
2317 return error;
2318 }
2319
2320 return 0;
2321}
2322
2323/*
2324 * Walk the unlinked chain from @head_agino until we find the inode that
2325 * points to @target_agino. Return the inode number, map, dinode pointer,
2326 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2327 *
2328 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2329 * @agino, @imap, @dipp, and @bpp are all output parameters.
2330 *
2331 * Do not call this function if @target_agino is the head of the list.
2332 */
2333STATIC int
2334xfs_iunlink_map_prev(
2335 struct xfs_trans *tp,
2336 xfs_agnumber_t agno,
2337 xfs_agino_t head_agino,
2338 xfs_agino_t target_agino,
2339 xfs_agino_t *agino,
2340 struct xfs_imap *imap,
2341 struct xfs_dinode **dipp,
2342 struct xfs_buf **bpp,
2343 struct xfs_perag *pag)
2344{
2345 struct xfs_mount *mp = tp->t_mountp;
2346 xfs_agino_t next_agino;
2347 int error;
2348
2349 ASSERT(head_agino != target_agino);
2350 *bpp = NULL;
2351
2352 /* See if our backref cache can find it faster. */
2353 *agino = xfs_iunlink_lookup_backref(pag, target_agino);
2354 if (*agino != NULLAGINO) {
2355 error = xfs_iunlink_map_ino(tp, agno, *agino, imap, dipp, bpp);
2356 if (error)
2357 return error;
2358
2359 if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2360 return 0;
2361
2362 /*
2363 * If we get here the cache contents were corrupt, so drop the
2364 * buffer and fall back to walking the bucket list.
2365 */
2366 xfs_trans_brelse(tp, *bpp);
2367 *bpp = NULL;
2368 WARN_ON_ONCE(1);
2369 }
2370
2371 trace_xfs_iunlink_map_prev_fallback(mp, agno);
2372
2373 /* Otherwise, walk the entire bucket until we find it. */
2374 next_agino = head_agino;
2375 while (next_agino != target_agino) {
2376 xfs_agino_t unlinked_agino;
2377
2378 if (*bpp)
2379 xfs_trans_brelse(tp, *bpp);
2380
2381 *agino = next_agino;
2382 error = xfs_iunlink_map_ino(tp, agno, next_agino, imap, dipp,
2383 bpp);
2384 if (error)
2385 return error;
2386
2387 unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2388 /*
2389 * Make sure this pointer is valid and isn't an obvious
2390 * infinite loop.
2391 */
2392 if (!xfs_verify_agino(mp, agno, unlinked_agino) ||
2393 next_agino == unlinked_agino) {
2394 XFS_CORRUPTION_ERROR(__func__,
2395 XFS_ERRLEVEL_LOW, mp,
2396 *dipp, sizeof(**dipp));
2397 error = -EFSCORRUPTED;
2398 return error;
2399 }
2400 next_agino = unlinked_agino;
2401 }
2402
2403 return 0;
2404}
2405
2406/*
2407 * Pull the on-disk inode from the AGI unlinked list.
2408 */
2409STATIC int
2410xfs_iunlink_remove(
2411 struct xfs_trans *tp,
2412 struct xfs_inode *ip)
2413{
2414 struct xfs_mount *mp = tp->t_mountp;
2415 struct xfs_agi *agi;
2416 struct xfs_buf *agibp;
2417 struct xfs_buf *last_ibp;
2418 struct xfs_dinode *last_dip = NULL;
2419 struct xfs_perag *pag = 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 = XFS_BUF_TO_AGI(agibp);
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 pag = xfs_perag_get(mp, agno);
2464 error = xfs_iunlink_change_backref(pag, next_agino,
2465 NULLAGINO);
2466 if (error)
2467 goto out;
2468 }
2469
2470 if (head_agino == agino) {
2471 /* Point the head of the list to the next unlinked inode. */
2472 error = xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index,
2473 next_agino);
2474 if (error)
2475 goto out;
2476 } else {
2477 struct xfs_imap imap;
2478 xfs_agino_t prev_agino;
2479
2480 if (!pag)
2481 pag = xfs_perag_get(mp, agno);
2482
2483 /* We need to search the list for the inode being freed. */
2484 error = xfs_iunlink_map_prev(tp, agno, head_agino, agino,
2485 &prev_agino, &imap, &last_dip, &last_ibp,
2486 pag);
2487 if (error)
2488 goto out;
2489
2490 /* Point the previous inode on the list to the next inode. */
2491 xfs_iunlink_update_dinode(tp, agno, prev_agino, last_ibp,
2492 last_dip, &imap, next_agino);
2493
2494 /*
2495 * Now we deal with the backref for this inode. If this inode
2496 * pointed at a real inode, change the backref that pointed to
2497 * us to point to our old next. If this inode was the end of
2498 * the list, delete the backref that pointed to us. Note that
2499 * change_backref takes care of deleting the backref if
2500 * next_agino is NULLAGINO.
2501 */
2502 error = xfs_iunlink_change_backref(pag, agino, next_agino);
2503 if (error)
2504 goto out;
2505 }
2506
2507out:
2508 if (pag)
2509 xfs_perag_put(pag);
2510 return error;
2511}
2512
2513/*
2514 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2515 * inodes that are in memory - they all must be marked stale and attached to
2516 * the cluster buffer.
2517 */
2518STATIC int
2519xfs_ifree_cluster(
2520 xfs_inode_t *free_ip,
2521 xfs_trans_t *tp,
2522 struct xfs_icluster *xic)
2523{
2524 xfs_mount_t *mp = free_ip->i_mount;
2525 int nbufs;
2526 int i, j;
2527 int ioffset;
2528 xfs_daddr_t blkno;
2529 xfs_buf_t *bp;
2530 xfs_inode_t *ip;
2531 xfs_inode_log_item_t *iip;
2532 struct xfs_log_item *lip;
2533 struct xfs_perag *pag;
2534 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2535 xfs_ino_t inum;
2536
2537 inum = xic->first_ino;
2538 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2539 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2540
2541 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2542 /*
2543 * The allocation bitmap tells us which inodes of the chunk were
2544 * physically allocated. Skip the cluster if an inode falls into
2545 * a sparse region.
2546 */
2547 ioffset = inum - xic->first_ino;
2548 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2549 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2550 continue;
2551 }
2552
2553 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2554 XFS_INO_TO_AGBNO(mp, inum));
2555
2556 /*
2557 * We obtain and lock the backing buffer first in the process
2558 * here, as we have to ensure that any dirty inode that we
2559 * can't get the flush lock on is attached to the buffer.
2560 * If we scan the in-memory inodes first, then buffer IO can
2561 * complete before we get a lock on it, and hence we may fail
2562 * to mark all the active inodes on the buffer stale.
2563 */
2564 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2565 mp->m_bsize * igeo->blocks_per_cluster,
2566 XBF_UNMAPPED);
2567
2568 if (!bp)
2569 return -ENOMEM;
2570
2571 /*
2572 * This buffer may not have been correctly initialised as we
2573 * didn't read it from disk. That's not important because we are
2574 * only using to mark the buffer as stale in the log, and to
2575 * attach stale cached inodes on it. That means it will never be
2576 * dispatched for IO. If it is, we want to know about it, and we
2577 * want it to fail. We can acheive this by adding a write
2578 * verifier to the buffer.
2579 */
2580 bp->b_ops = &xfs_inode_buf_ops;
2581
2582 /*
2583 * Walk the inodes already attached to the buffer and mark them
2584 * stale. These will all have the flush locks held, so an
2585 * in-memory inode walk can't lock them. By marking them all
2586 * stale first, we will not attempt to lock them in the loop
2587 * below as the XFS_ISTALE flag will be set.
2588 */
2589 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
2590 if (lip->li_type == XFS_LI_INODE) {
2591 iip = (xfs_inode_log_item_t *)lip;
2592 ASSERT(iip->ili_logged == 1);
2593 lip->li_cb = xfs_istale_done;
2594 xfs_trans_ail_copy_lsn(mp->m_ail,
2595 &iip->ili_flush_lsn,
2596 &iip->ili_item.li_lsn);
2597 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2598 }
2599 }
2600
2601
2602 /*
2603 * For each inode in memory attempt to add it to the inode
2604 * buffer and set it up for being staled on buffer IO
2605 * completion. This is safe as we've locked out tail pushing
2606 * and flushing by locking the buffer.
2607 *
2608 * We have already marked every inode that was part of a
2609 * transaction stale above, which means there is no point in
2610 * even trying to lock them.
2611 */
2612 for (i = 0; i < igeo->inodes_per_cluster; i++) {
2613retry:
2614 rcu_read_lock();
2615 ip = radix_tree_lookup(&pag->pag_ici_root,
2616 XFS_INO_TO_AGINO(mp, (inum + i)));
2617
2618 /* Inode not in memory, nothing to do */
2619 if (!ip) {
2620 rcu_read_unlock();
2621 continue;
2622 }
2623
2624 /*
2625 * because this is an RCU protected lookup, we could
2626 * find a recently freed or even reallocated inode
2627 * during the lookup. We need to check under the
2628 * i_flags_lock for a valid inode here. Skip it if it
2629 * is not valid, the wrong inode or stale.
2630 */
2631 spin_lock(&ip->i_flags_lock);
2632 if (ip->i_ino != inum + i ||
2633 __xfs_iflags_test(ip, XFS_ISTALE)) {
2634 spin_unlock(&ip->i_flags_lock);
2635 rcu_read_unlock();
2636 continue;
2637 }
2638 spin_unlock(&ip->i_flags_lock);
2639
2640 /*
2641 * Don't try to lock/unlock the current inode, but we
2642 * _cannot_ skip the other inodes that we did not find
2643 * in the list attached to the buffer and are not
2644 * already marked stale. If we can't lock it, back off
2645 * and retry.
2646 */
2647 if (ip != free_ip) {
2648 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2649 rcu_read_unlock();
2650 delay(1);
2651 goto retry;
2652 }
2653
2654 /*
2655 * Check the inode number again in case we're
2656 * racing with freeing in xfs_reclaim_inode().
2657 * See the comments in that function for more
2658 * information as to why the initial check is
2659 * not sufficient.
2660 */
2661 if (ip->i_ino != inum + i) {
2662 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2663 rcu_read_unlock();
2664 continue;
2665 }
2666 }
2667 rcu_read_unlock();
2668
2669 xfs_iflock(ip);
2670 xfs_iflags_set(ip, XFS_ISTALE);
2671
2672 /*
2673 * we don't need to attach clean inodes or those only
2674 * with unlogged changes (which we throw away, anyway).
2675 */
2676 iip = ip->i_itemp;
2677 if (!iip || xfs_inode_clean(ip)) {
2678 ASSERT(ip != free_ip);
2679 xfs_ifunlock(ip);
2680 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2681 continue;
2682 }
2683
2684 iip->ili_last_fields = iip->ili_fields;
2685 iip->ili_fields = 0;
2686 iip->ili_fsync_fields = 0;
2687 iip->ili_logged = 1;
2688 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2689 &iip->ili_item.li_lsn);
2690
2691 xfs_buf_attach_iodone(bp, xfs_istale_done,
2692 &iip->ili_item);
2693
2694 if (ip != free_ip)
2695 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2696 }
2697
2698 xfs_trans_stale_inode_buf(tp, bp);
2699 xfs_trans_binval(tp, bp);
2700 }
2701
2702 xfs_perag_put(pag);
2703 return 0;
2704}
2705
2706/*
2707 * Free any local-format buffers sitting around before we reset to
2708 * extents format.
2709 */
2710static inline void
2711xfs_ifree_local_data(
2712 struct xfs_inode *ip,
2713 int whichfork)
2714{
2715 struct xfs_ifork *ifp;
2716
2717 if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
2718 return;
2719
2720 ifp = XFS_IFORK_PTR(ip, whichfork);
2721 xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
2722}
2723
2724/*
2725 * This is called to return an inode to the inode free list.
2726 * The inode should already be truncated to 0 length and have
2727 * no pages associated with it. This routine also assumes that
2728 * the inode is already a part of the transaction.
2729 *
2730 * The on-disk copy of the inode will have been added to the list
2731 * of unlinked inodes in the AGI. We need to remove the inode from
2732 * that list atomically with respect to freeing it here.
2733 */
2734int
2735xfs_ifree(
2736 struct xfs_trans *tp,
2737 struct xfs_inode *ip)
2738{
2739 int error;
2740 struct xfs_icluster xic = { 0 };
2741
2742 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2743 ASSERT(VFS_I(ip)->i_nlink == 0);
2744 ASSERT(ip->i_d.di_nextents == 0);
2745 ASSERT(ip->i_d.di_anextents == 0);
2746 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2747 ASSERT(ip->i_d.di_nblocks == 0);
2748
2749 /*
2750 * Pull the on-disk inode from the AGI unlinked list.
2751 */
2752 error = xfs_iunlink_remove(tp, ip);
2753 if (error)
2754 return error;
2755
2756 error = xfs_difree(tp, ip->i_ino, &xic);
2757 if (error)
2758 return error;
2759
2760 xfs_ifree_local_data(ip, XFS_DATA_FORK);
2761 xfs_ifree_local_data(ip, XFS_ATTR_FORK);
2762
2763 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2764 ip->i_d.di_flags = 0;
2765 ip->i_d.di_flags2 = 0;
2766 ip->i_d.di_dmevmask = 0;
2767 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2768 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2769 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2770
2771 /* Don't attempt to replay owner changes for a deleted inode */
2772 ip->i_itemp->ili_fields &= ~(XFS_ILOG_AOWNER|XFS_ILOG_DOWNER);
2773
2774 /*
2775 * Bump the generation count so no one will be confused
2776 * by reincarnations of this inode.
2777 */
2778 VFS_I(ip)->i_generation++;
2779 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2780
2781 if (xic.deleted)
2782 error = xfs_ifree_cluster(ip, tp, &xic);
2783
2784 return error;
2785}
2786
2787/*
2788 * This is called to unpin an inode. The caller must have the inode locked
2789 * in at least shared mode so that the buffer cannot be subsequently pinned
2790 * once someone is waiting for it to be unpinned.
2791 */
2792static void
2793xfs_iunpin(
2794 struct xfs_inode *ip)
2795{
2796 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2797
2798 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2799
2800 /* Give the log a push to start the unpinning I/O */
2801 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0, NULL);
2802
2803}
2804
2805static void
2806__xfs_iunpin_wait(
2807 struct xfs_inode *ip)
2808{
2809 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2810 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2811
2812 xfs_iunpin(ip);
2813
2814 do {
2815 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2816 if (xfs_ipincount(ip))
2817 io_schedule();
2818 } while (xfs_ipincount(ip));
2819 finish_wait(wq, &wait.wq_entry);
2820}
2821
2822void
2823xfs_iunpin_wait(
2824 struct xfs_inode *ip)
2825{
2826 if (xfs_ipincount(ip))
2827 __xfs_iunpin_wait(ip);
2828}
2829
2830/*
2831 * Removing an inode from the namespace involves removing the directory entry
2832 * and dropping the link count on the inode. Removing the directory entry can
2833 * result in locking an AGF (directory blocks were freed) and removing a link
2834 * count can result in placing the inode on an unlinked list which results in
2835 * locking an AGI.
2836 *
2837 * The big problem here is that we have an ordering constraint on AGF and AGI
2838 * locking - inode allocation locks the AGI, then can allocate a new extent for
2839 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2840 * removes the inode from the unlinked list, requiring that we lock the AGI
2841 * first, and then freeing the inode can result in an inode chunk being freed
2842 * and hence freeing disk space requiring that we lock an AGF.
2843 *
2844 * Hence the ordering that is imposed by other parts of the code is AGI before
2845 * AGF. This means we cannot remove the directory entry before we drop the inode
2846 * reference count and put it on the unlinked list as this results in a lock
2847 * order of AGF then AGI, and this can deadlock against inode allocation and
2848 * freeing. Therefore we must drop the link counts before we remove the
2849 * directory entry.
2850 *
2851 * This is still safe from a transactional point of view - it is not until we
2852 * get to xfs_defer_finish() that we have the possibility of multiple
2853 * transactions in this operation. Hence as long as we remove the directory
2854 * entry and drop the link count in the first transaction of the remove
2855 * operation, there are no transactional constraints on the ordering here.
2856 */
2857int
2858xfs_remove(
2859 xfs_inode_t *dp,
2860 struct xfs_name *name,
2861 xfs_inode_t *ip)
2862{
2863 xfs_mount_t *mp = dp->i_mount;
2864 xfs_trans_t *tp = NULL;
2865 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2866 int error = 0;
2867 uint resblks;
2868
2869 trace_xfs_remove(dp, name);
2870
2871 if (XFS_FORCED_SHUTDOWN(mp))
2872 return -EIO;
2873
2874 error = xfs_qm_dqattach(dp);
2875 if (error)
2876 goto std_return;
2877
2878 error = xfs_qm_dqattach(ip);
2879 if (error)
2880 goto std_return;
2881
2882 /*
2883 * We try to get the real space reservation first,
2884 * allowing for directory btree deletion(s) implying
2885 * possible bmap insert(s). If we can't get the space
2886 * reservation then we use 0 instead, and avoid the bmap
2887 * btree insert(s) in the directory code by, if the bmap
2888 * insert tries to happen, instead trimming the LAST
2889 * block from the directory.
2890 */
2891 resblks = XFS_REMOVE_SPACE_RES(mp);
2892 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2893 if (error == -ENOSPC) {
2894 resblks = 0;
2895 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2896 &tp);
2897 }
2898 if (error) {
2899 ASSERT(error != -ENOSPC);
2900 goto std_return;
2901 }
2902
2903 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2904
2905 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2906 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2907
2908 /*
2909 * If we're removing a directory perform some additional validation.
2910 */
2911 if (is_dir) {
2912 ASSERT(VFS_I(ip)->i_nlink >= 2);
2913 if (VFS_I(ip)->i_nlink != 2) {
2914 error = -ENOTEMPTY;
2915 goto out_trans_cancel;
2916 }
2917 if (!xfs_dir_isempty(ip)) {
2918 error = -ENOTEMPTY;
2919 goto out_trans_cancel;
2920 }
2921
2922 /* Drop the link from ip's "..". */
2923 error = xfs_droplink(tp, dp);
2924 if (error)
2925 goto out_trans_cancel;
2926
2927 /* Drop the "." link from ip to self. */
2928 error = xfs_droplink(tp, ip);
2929 if (error)
2930 goto out_trans_cancel;
2931 } else {
2932 /*
2933 * When removing a non-directory we need to log the parent
2934 * inode here. For a directory this is done implicitly
2935 * by the xfs_droplink call for the ".." entry.
2936 */
2937 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2938 }
2939 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2940
2941 /* Drop the link from dp to ip. */
2942 error = xfs_droplink(tp, ip);
2943 if (error)
2944 goto out_trans_cancel;
2945
2946 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2947 if (error) {
2948 ASSERT(error != -ENOENT);
2949 goto out_trans_cancel;
2950 }
2951
2952 /*
2953 * If this is a synchronous mount, make sure that the
2954 * remove transaction goes to disk before returning to
2955 * the user.
2956 */
2957 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2958 xfs_trans_set_sync(tp);
2959
2960 error = xfs_trans_commit(tp);
2961 if (error)
2962 goto std_return;
2963
2964 if (is_dir && xfs_inode_is_filestream(ip))
2965 xfs_filestream_deassociate(ip);
2966
2967 return 0;
2968
2969 out_trans_cancel:
2970 xfs_trans_cancel(tp);
2971 std_return:
2972 return error;
2973}
2974
2975/*
2976 * Enter all inodes for a rename transaction into a sorted array.
2977 */
2978#define __XFS_SORT_INODES 5
2979STATIC void
2980xfs_sort_for_rename(
2981 struct xfs_inode *dp1, /* in: old (source) directory inode */
2982 struct xfs_inode *dp2, /* in: new (target) directory inode */
2983 struct xfs_inode *ip1, /* in: inode of old entry */
2984 struct xfs_inode *ip2, /* in: inode of new entry */
2985 struct xfs_inode *wip, /* in: whiteout inode */
2986 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2987 int *num_inodes) /* in/out: inodes in array */
2988{
2989 int i, j;
2990
2991 ASSERT(*num_inodes == __XFS_SORT_INODES);
2992 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2993
2994 /*
2995 * i_tab contains a list of pointers to inodes. We initialize
2996 * the table here & we'll sort it. We will then use it to
2997 * order the acquisition of the inode locks.
2998 *
2999 * Note that the table may contain duplicates. e.g., dp1 == dp2.
3000 */
3001 i = 0;
3002 i_tab[i++] = dp1;
3003 i_tab[i++] = dp2;
3004 i_tab[i++] = ip1;
3005 if (ip2)
3006 i_tab[i++] = ip2;
3007 if (wip)
3008 i_tab[i++] = wip;
3009 *num_inodes = i;
3010
3011 /*
3012 * Sort the elements via bubble sort. (Remember, there are at
3013 * most 5 elements to sort, so this is adequate.)
3014 */
3015 for (i = 0; i < *num_inodes; i++) {
3016 for (j = 1; j < *num_inodes; j++) {
3017 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
3018 struct xfs_inode *temp = i_tab[j];
3019 i_tab[j] = i_tab[j-1];
3020 i_tab[j-1] = temp;
3021 }
3022 }
3023 }
3024}
3025
3026static int
3027xfs_finish_rename(
3028 struct xfs_trans *tp)
3029{
3030 /*
3031 * If this is a synchronous mount, make sure that the rename transaction
3032 * goes to disk before returning to the user.
3033 */
3034 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
3035 xfs_trans_set_sync(tp);
3036
3037 return xfs_trans_commit(tp);
3038}
3039
3040/*
3041 * xfs_cross_rename()
3042 *
3043 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
3044 */
3045STATIC int
3046xfs_cross_rename(
3047 struct xfs_trans *tp,
3048 struct xfs_inode *dp1,
3049 struct xfs_name *name1,
3050 struct xfs_inode *ip1,
3051 struct xfs_inode *dp2,
3052 struct xfs_name *name2,
3053 struct xfs_inode *ip2,
3054 int spaceres)
3055{
3056 int error = 0;
3057 int ip1_flags = 0;
3058 int ip2_flags = 0;
3059 int dp2_flags = 0;
3060
3061 /* Swap inode number for dirent in first parent */
3062 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
3063 if (error)
3064 goto out_trans_abort;
3065
3066 /* Swap inode number for dirent in second parent */
3067 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
3068 if (error)
3069 goto out_trans_abort;
3070
3071 /*
3072 * If we're renaming one or more directories across different parents,
3073 * update the respective ".." entries (and link counts) to match the new
3074 * parents.
3075 */
3076 if (dp1 != dp2) {
3077 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3078
3079 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
3080 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
3081 dp1->i_ino, spaceres);
3082 if (error)
3083 goto out_trans_abort;
3084
3085 /* transfer ip2 ".." reference to dp1 */
3086 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
3087 error = xfs_droplink(tp, dp2);
3088 if (error)
3089 goto out_trans_abort;
3090 xfs_bumplink(tp, dp1);
3091 }
3092
3093 /*
3094 * Although ip1 isn't changed here, userspace needs
3095 * to be warned about the change, so that applications
3096 * relying on it (like backup ones), will properly
3097 * notify the change
3098 */
3099 ip1_flags |= XFS_ICHGTIME_CHG;
3100 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3101 }
3102
3103 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
3104 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
3105 dp2->i_ino, spaceres);
3106 if (error)
3107 goto out_trans_abort;
3108
3109 /* transfer ip1 ".." reference to dp2 */
3110 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
3111 error = xfs_droplink(tp, dp1);
3112 if (error)
3113 goto out_trans_abort;
3114 xfs_bumplink(tp, dp2);
3115 }
3116
3117 /*
3118 * Although ip2 isn't changed here, userspace needs
3119 * to be warned about the change, so that applications
3120 * relying on it (like backup ones), will properly
3121 * notify the change
3122 */
3123 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3124 ip2_flags |= XFS_ICHGTIME_CHG;
3125 }
3126 }
3127
3128 if (ip1_flags) {
3129 xfs_trans_ichgtime(tp, ip1, ip1_flags);
3130 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
3131 }
3132 if (ip2_flags) {
3133 xfs_trans_ichgtime(tp, ip2, ip2_flags);
3134 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
3135 }
3136 if (dp2_flags) {
3137 xfs_trans_ichgtime(tp, dp2, dp2_flags);
3138 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
3139 }
3140 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3141 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
3142 return xfs_finish_rename(tp);
3143
3144out_trans_abort:
3145 xfs_trans_cancel(tp);
3146 return error;
3147}
3148
3149/*
3150 * xfs_rename_alloc_whiteout()
3151 *
3152 * Return a referenced, unlinked, unlocked inode that that can be used as a
3153 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3154 * crash between allocating the inode and linking it into the rename transaction
3155 * recovery will free the inode and we won't leak it.
3156 */
3157static int
3158xfs_rename_alloc_whiteout(
3159 struct xfs_inode *dp,
3160 struct xfs_inode **wip)
3161{
3162 struct xfs_inode *tmpfile;
3163 int error;
3164
3165 error = xfs_create_tmpfile(dp, S_IFCHR | WHITEOUT_MODE, &tmpfile);
3166 if (error)
3167 return error;
3168
3169 /*
3170 * Prepare the tmpfile inode as if it were created through the VFS.
3171 * Complete the inode setup and flag it as linkable. nlink is already
3172 * zero, so we can skip the drop_nlink.
3173 */
3174 xfs_setup_iops(tmpfile);
3175 xfs_finish_inode_setup(tmpfile);
3176 VFS_I(tmpfile)->i_state |= I_LINKABLE;
3177
3178 *wip = tmpfile;
3179 return 0;
3180}
3181
3182/*
3183 * xfs_rename
3184 */
3185int
3186xfs_rename(
3187 struct xfs_inode *src_dp,
3188 struct xfs_name *src_name,
3189 struct xfs_inode *src_ip,
3190 struct xfs_inode *target_dp,
3191 struct xfs_name *target_name,
3192 struct xfs_inode *target_ip,
3193 unsigned int flags)
3194{
3195 struct xfs_mount *mp = src_dp->i_mount;
3196 struct xfs_trans *tp;
3197 struct xfs_inode *wip = NULL; /* whiteout inode */
3198 struct xfs_inode *inodes[__XFS_SORT_INODES];
3199 int num_inodes = __XFS_SORT_INODES;
3200 bool new_parent = (src_dp != target_dp);
3201 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3202 int spaceres;
3203 int error;
3204
3205 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3206
3207 if ((flags & RENAME_EXCHANGE) && !target_ip)
3208 return -EINVAL;
3209
3210 /*
3211 * If we are doing a whiteout operation, allocate the whiteout inode
3212 * we will be placing at the target and ensure the type is set
3213 * appropriately.
3214 */
3215 if (flags & RENAME_WHITEOUT) {
3216 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3217 error = xfs_rename_alloc_whiteout(target_dp, &wip);
3218 if (error)
3219 return error;
3220
3221 /* setup target dirent info as whiteout */
3222 src_name->type = XFS_DIR3_FT_CHRDEV;
3223 }
3224
3225 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3226 inodes, &num_inodes);
3227
3228 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3229 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3230 if (error == -ENOSPC) {
3231 spaceres = 0;
3232 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3233 &tp);
3234 }
3235 if (error)
3236 goto out_release_wip;
3237
3238 /*
3239 * Attach the dquots to the inodes
3240 */
3241 error = xfs_qm_vop_rename_dqattach(inodes);
3242 if (error)
3243 goto out_trans_cancel;
3244
3245 /*
3246 * Lock all the participating inodes. Depending upon whether
3247 * the target_name exists in the target directory, and
3248 * whether the target directory is the same as the source
3249 * directory, we can lock from 2 to 4 inodes.
3250 */
3251 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3252
3253 /*
3254 * Join all the inodes to the transaction. From this point on,
3255 * we can rely on either trans_commit or trans_cancel to unlock
3256 * them.
3257 */
3258 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3259 if (new_parent)
3260 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3261 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3262 if (target_ip)
3263 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3264 if (wip)
3265 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3266
3267 /*
3268 * If we are using project inheritance, we only allow renames
3269 * into our tree when the project IDs are the same; else the
3270 * tree quota mechanism would be circumvented.
3271 */
3272 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
3273 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
3274 error = -EXDEV;
3275 goto out_trans_cancel;
3276 }
3277
3278 /* RENAME_EXCHANGE is unique from here on. */
3279 if (flags & RENAME_EXCHANGE)
3280 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3281 target_dp, target_name, target_ip,
3282 spaceres);
3283
3284 /*
3285 * Check for expected errors before we dirty the transaction
3286 * so we can return an error without a transaction abort.
3287 */
3288 if (target_ip == NULL) {
3289 /*
3290 * If there's no space reservation, check the entry will
3291 * fit before actually inserting it.
3292 */
3293 if (!spaceres) {
3294 error = xfs_dir_canenter(tp, target_dp, target_name);
3295 if (error)
3296 goto out_trans_cancel;
3297 }
3298 } else {
3299 /*
3300 * If target exists and it's a directory, check that whether
3301 * it can be destroyed.
3302 */
3303 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3304 (!xfs_dir_isempty(target_ip) ||
3305 (VFS_I(target_ip)->i_nlink > 2))) {
3306 error = -EEXIST;
3307 goto out_trans_cancel;
3308 }
3309 }
3310
3311 /*
3312 * Directory entry creation below may acquire the AGF. Remove
3313 * the whiteout from the unlinked list first to preserve correct
3314 * AGI/AGF locking order. This dirties the transaction so failures
3315 * after this point will abort and log recovery will clean up the
3316 * mess.
3317 *
3318 * For whiteouts, we need to bump the link count on the whiteout
3319 * inode. After this point, we have a real link, clear the tmpfile
3320 * state flag from the inode so it doesn't accidentally get misused
3321 * in future.
3322 */
3323 if (wip) {
3324 ASSERT(VFS_I(wip)->i_nlink == 0);
3325 error = xfs_iunlink_remove(tp, wip);
3326 if (error)
3327 goto out_trans_cancel;
3328
3329 xfs_bumplink(tp, wip);
3330 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3331 VFS_I(wip)->i_state &= ~I_LINKABLE;
3332 }
3333
3334 /*
3335 * Set up the target.
3336 */
3337 if (target_ip == NULL) {
3338 /*
3339 * If target does not exist and the rename crosses
3340 * directories, adjust the target directory link count
3341 * to account for the ".." reference from the new entry.
3342 */
3343 error = xfs_dir_createname(tp, target_dp, target_name,
3344 src_ip->i_ino, spaceres);
3345 if (error)
3346 goto out_trans_cancel;
3347
3348 xfs_trans_ichgtime(tp, target_dp,
3349 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3350
3351 if (new_parent && src_is_directory) {
3352 xfs_bumplink(tp, target_dp);
3353 }
3354 } else { /* target_ip != NULL */
3355 /*
3356 * Link the source inode under the target name.
3357 * If the source inode is a directory and we are moving
3358 * it across directories, its ".." entry will be
3359 * inconsistent until we replace that down below.
3360 *
3361 * In case there is already an entry with the same
3362 * name at the destination directory, remove it first.
3363 */
3364 error = xfs_dir_replace(tp, target_dp, target_name,
3365 src_ip->i_ino, spaceres);
3366 if (error)
3367 goto out_trans_cancel;
3368
3369 xfs_trans_ichgtime(tp, target_dp,
3370 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3371
3372 /*
3373 * Decrement the link count on the target since the target
3374 * dir no longer points to it.
3375 */
3376 error = xfs_droplink(tp, target_ip);
3377 if (error)
3378 goto out_trans_cancel;
3379
3380 if (src_is_directory) {
3381 /*
3382 * Drop the link from the old "." entry.
3383 */
3384 error = xfs_droplink(tp, target_ip);
3385 if (error)
3386 goto out_trans_cancel;
3387 }
3388 } /* target_ip != NULL */
3389
3390 /*
3391 * Remove the source.
3392 */
3393 if (new_parent && src_is_directory) {
3394 /*
3395 * Rewrite the ".." entry to point to the new
3396 * directory.
3397 */
3398 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3399 target_dp->i_ino, spaceres);
3400 ASSERT(error != -EEXIST);
3401 if (error)
3402 goto out_trans_cancel;
3403 }
3404
3405 /*
3406 * We always want to hit the ctime on the source inode.
3407 *
3408 * This isn't strictly required by the standards since the source
3409 * inode isn't really being changed, but old unix file systems did
3410 * it and some incremental backup programs won't work without it.
3411 */
3412 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3413 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3414
3415 /*
3416 * Adjust the link count on src_dp. This is necessary when
3417 * renaming a directory, either within one parent when
3418 * the target existed, or across two parent directories.
3419 */
3420 if (src_is_directory && (new_parent || target_ip != NULL)) {
3421
3422 /*
3423 * Decrement link count on src_directory since the
3424 * entry that's moved no longer points to it.
3425 */
3426 error = xfs_droplink(tp, src_dp);
3427 if (error)
3428 goto out_trans_cancel;
3429 }
3430
3431 /*
3432 * For whiteouts, we only need to update the source dirent with the
3433 * inode number of the whiteout inode rather than removing it
3434 * altogether.
3435 */
3436 if (wip) {
3437 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3438 spaceres);
3439 } else
3440 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3441 spaceres);
3442 if (error)
3443 goto out_trans_cancel;
3444
3445 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3446 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3447 if (new_parent)
3448 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3449
3450 error = xfs_finish_rename(tp);
3451 if (wip)
3452 xfs_irele(wip);
3453 return error;
3454
3455out_trans_cancel:
3456 xfs_trans_cancel(tp);
3457out_release_wip:
3458 if (wip)
3459 xfs_irele(wip);
3460 return error;
3461}
3462
3463STATIC int
3464xfs_iflush_cluster(
3465 struct xfs_inode *ip,
3466 struct xfs_buf *bp)
3467{
3468 struct xfs_mount *mp = ip->i_mount;
3469 struct xfs_perag *pag;
3470 unsigned long first_index, mask;
3471 int cilist_size;
3472 struct xfs_inode **cilist;
3473 struct xfs_inode *cip;
3474 struct xfs_ino_geometry *igeo = M_IGEO(mp);
3475 int nr_found;
3476 int clcount = 0;
3477 int i;
3478
3479 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3480
3481 cilist_size = igeo->inodes_per_cluster * sizeof(struct xfs_inode *);
3482 cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3483 if (!cilist)
3484 goto out_put;
3485
3486 mask = ~(igeo->inodes_per_cluster - 1);
3487 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3488 rcu_read_lock();
3489 /* really need a gang lookup range call here */
3490 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3491 first_index, igeo->inodes_per_cluster);
3492 if (nr_found == 0)
3493 goto out_free;
3494
3495 for (i = 0; i < nr_found; i++) {
3496 cip = cilist[i];
3497 if (cip == ip)
3498 continue;
3499
3500 /*
3501 * because this is an RCU protected lookup, we could find a
3502 * recently freed or even reallocated inode during the lookup.
3503 * We need to check under the i_flags_lock for a valid inode
3504 * here. Skip it if it is not valid or the wrong inode.
3505 */
3506 spin_lock(&cip->i_flags_lock);
3507 if (!cip->i_ino ||
3508 __xfs_iflags_test(cip, XFS_ISTALE)) {
3509 spin_unlock(&cip->i_flags_lock);
3510 continue;
3511 }
3512
3513 /*
3514 * Once we fall off the end of the cluster, no point checking
3515 * any more inodes in the list because they will also all be
3516 * outside the cluster.
3517 */
3518 if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3519 spin_unlock(&cip->i_flags_lock);
3520 break;
3521 }
3522 spin_unlock(&cip->i_flags_lock);
3523
3524 /*
3525 * Do an un-protected check to see if the inode is dirty and
3526 * is a candidate for flushing. These checks will be repeated
3527 * later after the appropriate locks are acquired.
3528 */
3529 if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3530 continue;
3531
3532 /*
3533 * Try to get locks. If any are unavailable or it is pinned,
3534 * then this inode cannot be flushed and is skipped.
3535 */
3536
3537 if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3538 continue;
3539 if (!xfs_iflock_nowait(cip)) {
3540 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3541 continue;
3542 }
3543 if (xfs_ipincount(cip)) {
3544 xfs_ifunlock(cip);
3545 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3546 continue;
3547 }
3548
3549
3550 /*
3551 * Check the inode number again, just to be certain we are not
3552 * racing with freeing in xfs_reclaim_inode(). See the comments
3553 * in that function for more information as to why the initial
3554 * check is not sufficient.
3555 */
3556 if (!cip->i_ino) {
3557 xfs_ifunlock(cip);
3558 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3559 continue;
3560 }
3561
3562 /*
3563 * arriving here means that this inode can be flushed. First
3564 * re-check that it's dirty before flushing.
3565 */
3566 if (!xfs_inode_clean(cip)) {
3567 int error;
3568 error = xfs_iflush_int(cip, bp);
3569 if (error) {
3570 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3571 goto cluster_corrupt_out;
3572 }
3573 clcount++;
3574 } else {
3575 xfs_ifunlock(cip);
3576 }
3577 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3578 }
3579
3580 if (clcount) {
3581 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3582 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3583 }
3584
3585out_free:
3586 rcu_read_unlock();
3587 kmem_free(cilist);
3588out_put:
3589 xfs_perag_put(pag);
3590 return 0;
3591
3592
3593cluster_corrupt_out:
3594 /*
3595 * Corruption detected in the clustering loop. Invalidate the
3596 * inode buffer and shut down the filesystem.
3597 */
3598 rcu_read_unlock();
3599
3600 /*
3601 * We'll always have an inode attached to the buffer for completion
3602 * process by the time we are called from xfs_iflush(). Hence we have
3603 * always need to do IO completion processing to abort the inodes
3604 * attached to the buffer. handle them just like the shutdown case in
3605 * xfs_buf_submit().
3606 */
3607 ASSERT(bp->b_iodone);
3608 bp->b_flags |= XBF_ASYNC;
3609 bp->b_flags &= ~XBF_DONE;
3610 xfs_buf_stale(bp);
3611 xfs_buf_ioerror(bp, -EIO);
3612 xfs_buf_ioend(bp);
3613
3614 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3615
3616 /* abort the corrupt inode, as it was not attached to the buffer */
3617 xfs_iflush_abort(cip, false);
3618 kmem_free(cilist);
3619 xfs_perag_put(pag);
3620 return -EFSCORRUPTED;
3621}
3622
3623/*
3624 * Flush dirty inode metadata into the backing buffer.
3625 *
3626 * The caller must have the inode lock and the inode flush lock held. The
3627 * inode lock will still be held upon return to the caller, and the inode
3628 * flush lock will be released after the inode has reached the disk.
3629 *
3630 * The caller must write out the buffer returned in *bpp and release it.
3631 */
3632int
3633xfs_iflush(
3634 struct xfs_inode *ip,
3635 struct xfs_buf **bpp)
3636{
3637 struct xfs_mount *mp = ip->i_mount;
3638 struct xfs_buf *bp = NULL;
3639 struct xfs_dinode *dip;
3640 int error;
3641
3642 XFS_STATS_INC(mp, xs_iflush_count);
3643
3644 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3645 ASSERT(xfs_isiflocked(ip));
3646 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3647 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3648
3649 *bpp = NULL;
3650
3651 xfs_iunpin_wait(ip);
3652
3653 /*
3654 * For stale inodes we cannot rely on the backing buffer remaining
3655 * stale in cache for the remaining life of the stale inode and so
3656 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3657 * inodes below. We have to check this after ensuring the inode is
3658 * unpinned so that it is safe to reclaim the stale inode after the
3659 * flush call.
3660 */
3661 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3662 xfs_ifunlock(ip);
3663 return 0;
3664 }
3665
3666 /*
3667 * This may have been unpinned because the filesystem is shutting
3668 * down forcibly. If that's the case we must not write this inode
3669 * to disk, because the log record didn't make it to disk.
3670 *
3671 * We also have to remove the log item from the AIL in this case,
3672 * as we wait for an empty AIL as part of the unmount process.
3673 */
3674 if (XFS_FORCED_SHUTDOWN(mp)) {
3675 error = -EIO;
3676 goto abort_out;
3677 }
3678
3679 /*
3680 * Get the buffer containing the on-disk inode. We are doing a try-lock
3681 * operation here, so we may get an EAGAIN error. In that case, we
3682 * simply want to return with the inode still dirty.
3683 *
3684 * If we get any other error, we effectively have a corruption situation
3685 * and we cannot flush the inode, so we treat it the same as failing
3686 * xfs_iflush_int().
3687 */
3688 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3689 0);
3690 if (error == -EAGAIN) {
3691 xfs_ifunlock(ip);
3692 return error;
3693 }
3694 if (error)
3695 goto corrupt_out;
3696
3697 /*
3698 * First flush out the inode that xfs_iflush was called with.
3699 */
3700 error = xfs_iflush_int(ip, bp);
3701 if (error)
3702 goto corrupt_out;
3703
3704 /*
3705 * If the buffer is pinned then push on the log now so we won't
3706 * get stuck waiting in the write for too long.
3707 */
3708 if (xfs_buf_ispinned(bp))
3709 xfs_log_force(mp, 0);
3710
3711 /*
3712 * inode clustering: try to gather other inodes into this write
3713 *
3714 * Note: Any error during clustering will result in the filesystem
3715 * being shut down and completion callbacks run on the cluster buffer.
3716 * As we have already flushed and attached this inode to the buffer,
3717 * it has already been aborted and released by xfs_iflush_cluster() and
3718 * so we have no further error handling to do here.
3719 */
3720 error = xfs_iflush_cluster(ip, bp);
3721 if (error)
3722 return error;
3723
3724 *bpp = bp;
3725 return 0;
3726
3727corrupt_out:
3728 if (bp)
3729 xfs_buf_relse(bp);
3730 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3731abort_out:
3732 /* abort the corrupt inode, as it was not attached to the buffer */
3733 xfs_iflush_abort(ip, false);
3734 return error;
3735}
3736
3737/*
3738 * If there are inline format data / attr forks attached to this inode,
3739 * make sure they're not corrupt.
3740 */
3741bool
3742xfs_inode_verify_forks(
3743 struct xfs_inode *ip)
3744{
3745 struct xfs_ifork *ifp;
3746 xfs_failaddr_t fa;
3747
3748 fa = xfs_ifork_verify_data(ip, &xfs_default_ifork_ops);
3749 if (fa) {
3750 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
3751 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "data fork",
3752 ifp->if_u1.if_data, ifp->if_bytes, fa);
3753 return false;
3754 }
3755
3756 fa = xfs_ifork_verify_attr(ip, &xfs_default_ifork_ops);
3757 if (fa) {
3758 ifp = XFS_IFORK_PTR(ip, XFS_ATTR_FORK);
3759 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "attr fork",
3760 ifp ? ifp->if_u1.if_data : NULL,
3761 ifp ? ifp->if_bytes : 0, fa);
3762 return false;
3763 }
3764 return true;
3765}
3766
3767STATIC int
3768xfs_iflush_int(
3769 struct xfs_inode *ip,
3770 struct xfs_buf *bp)
3771{
3772 struct xfs_inode_log_item *iip = ip->i_itemp;
3773 struct xfs_dinode *dip;
3774 struct xfs_mount *mp = ip->i_mount;
3775
3776 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3777 ASSERT(xfs_isiflocked(ip));
3778 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3779 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3780 ASSERT(iip != NULL && iip->ili_fields != 0);
3781 ASSERT(ip->i_d.di_version > 1);
3782
3783 /* set *dip = inode's place in the buffer */
3784 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3785
3786 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3787 mp, XFS_ERRTAG_IFLUSH_1)) {
3788 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3789 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3790 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3791 goto corrupt_out;
3792 }
3793 if (S_ISREG(VFS_I(ip)->i_mode)) {
3794 if (XFS_TEST_ERROR(
3795 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3796 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3797 mp, XFS_ERRTAG_IFLUSH_3)) {
3798 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3799 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3800 __func__, ip->i_ino, ip);
3801 goto corrupt_out;
3802 }
3803 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3804 if (XFS_TEST_ERROR(
3805 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3806 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3807 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3808 mp, XFS_ERRTAG_IFLUSH_4)) {
3809 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3810 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3811 __func__, ip->i_ino, ip);
3812 goto corrupt_out;
3813 }
3814 }
3815 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3816 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3817 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3818 "%s: detected corrupt incore inode %Lu, "
3819 "total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3820 __func__, ip->i_ino,
3821 ip->i_d.di_nextents + ip->i_d.di_anextents,
3822 ip->i_d.di_nblocks, ip);
3823 goto corrupt_out;
3824 }
3825 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3826 mp, XFS_ERRTAG_IFLUSH_6)) {
3827 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3828 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3829 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3830 goto corrupt_out;
3831 }
3832
3833 /*
3834 * Inode item log recovery for v2 inodes are dependent on the
3835 * di_flushiter count for correct sequencing. We bump the flush
3836 * iteration count so we can detect flushes which postdate a log record
3837 * during recovery. This is redundant as we now log every change and
3838 * hence this can't happen but we need to still do it to ensure
3839 * backwards compatibility with old kernels that predate logging all
3840 * inode changes.
3841 */
3842 if (ip->i_d.di_version < 3)
3843 ip->i_d.di_flushiter++;
3844
3845 /* Check the inline fork data before we write out. */
3846 if (!xfs_inode_verify_forks(ip))
3847 goto corrupt_out;
3848
3849 /*
3850 * Copy the dirty parts of the inode into the on-disk inode. We always
3851 * copy out the core of the inode, because if the inode is dirty at all
3852 * the core must be.
3853 */
3854 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3855
3856 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3857 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3858 ip->i_d.di_flushiter = 0;
3859
3860 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3861 if (XFS_IFORK_Q(ip))
3862 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3863 xfs_inobp_check(mp, bp);
3864
3865 /*
3866 * We've recorded everything logged in the inode, so we'd like to clear
3867 * the ili_fields bits so we don't log and flush things unnecessarily.
3868 * However, we can't stop logging all this information until the data
3869 * we've copied into the disk buffer is written to disk. If we did we
3870 * might overwrite the copy of the inode in the log with all the data
3871 * after re-logging only part of it, and in the face of a crash we
3872 * wouldn't have all the data we need to recover.
3873 *
3874 * What we do is move the bits to the ili_last_fields field. When
3875 * logging the inode, these bits are moved back to the ili_fields field.
3876 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3877 * know that the information those bits represent is permanently on
3878 * disk. As long as the flush completes before the inode is logged
3879 * again, then both ili_fields and ili_last_fields will be cleared.
3880 *
3881 * We can play with the ili_fields bits here, because the inode lock
3882 * must be held exclusively in order to set bits there and the flush
3883 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3884 * done routine can tell whether or not to look in the AIL. Also, store
3885 * the current LSN of the inode so that we can tell whether the item has
3886 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3887 * need the AIL lock, because it is a 64 bit value that cannot be read
3888 * atomically.
3889 */
3890 iip->ili_last_fields = iip->ili_fields;
3891 iip->ili_fields = 0;
3892 iip->ili_fsync_fields = 0;
3893 iip->ili_logged = 1;
3894
3895 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3896 &iip->ili_item.li_lsn);
3897
3898 /*
3899 * Attach the function xfs_iflush_done to the inode's
3900 * buffer. This will remove the inode from the AIL
3901 * and unlock the inode's flush lock when the inode is
3902 * completely written to disk.
3903 */
3904 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3905
3906 /* generate the checksum. */
3907 xfs_dinode_calc_crc(mp, dip);
3908
3909 ASSERT(!list_empty(&bp->b_li_list));
3910 ASSERT(bp->b_iodone != NULL);
3911 return 0;
3912
3913corrupt_out:
3914 return -EFSCORRUPTED;
3915}
3916
3917/* Release an inode. */
3918void
3919xfs_irele(
3920 struct xfs_inode *ip)
3921{
3922 trace_xfs_irele(ip, _RET_IP_);
3923 iput(VFS_I(ip));
3924}