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
47 struct kmem_cache *xfs_inode_cache;
48
49 /*
50 * These two are wrapper routines around the xfs_ilock() routine used to
51 * centralize some grungy code. They are used in places that wish to lock the
52 * inode solely for reading the extents. The reason these places can't just
53 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
54 * bringing in of the extents from disk for a file in b-tree format. If the
55 * inode is in b-tree format, then we need to lock the inode exclusively until
56 * the extents are read in. Locking it exclusively all the time would limit
57 * our parallelism unnecessarily, though. What we do instead is check to see
58 * if the extents have been read in yet, and only lock the inode exclusively
59 * if they have not.
60 *
61 * The functions return a value which should be given to the corresponding
62 * xfs_iunlock() call.
63 */
64 uint
65 xfs_ilock_data_map_shared(
66 struct xfs_inode *ip)
67 {
68 uint lock_mode = XFS_ILOCK_SHARED;
69
70 if (xfs_need_iread_extents(&ip->i_df))
71 lock_mode = XFS_ILOCK_EXCL;
72 xfs_ilock(ip, lock_mode);
73 return lock_mode;
74 }
75
76 uint
77 xfs_ilock_attr_map_shared(
78 struct xfs_inode *ip)
79 {
80 uint lock_mode = XFS_ILOCK_SHARED;
81
82 if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
83 lock_mode = XFS_ILOCK_EXCL;
84 xfs_ilock(ip, lock_mode);
85 return lock_mode;
86 }
87
88 /*
89 * You can't set both SHARED and EXCL for the same lock,
90 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
91 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
92 * to set in lock_flags.
93 */
94 static inline void
95 xfs_lock_flags_assert(
96 uint lock_flags)
97 {
98 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
99 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
100 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
101 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
102 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
103 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
104 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
105 ASSERT(lock_flags != 0);
106 }
107
108 /*
109 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
110 * multi-reader locks: invalidate_lock and the i_lock. This routine allows
111 * various combinations of the locks to be obtained.
112 *
113 * The 3 locks should always be ordered so that the IO lock is obtained first,
114 * the mmap lock second and the ilock last in order to prevent deadlock.
115 *
116 * Basic locking order:
117 *
118 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
119 *
120 * mmap_lock locking order:
121 *
122 * i_rwsem -> page lock -> mmap_lock
123 * mmap_lock -> invalidate_lock -> page_lock
124 *
125 * The difference in mmap_lock locking order mean that we cannot hold the
126 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
127 * can fault in pages during copy in/out (for buffered IO) or require the
128 * mmap_lock in get_user_pages() to map the user pages into the kernel address
129 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
130 * fault because page faults already hold the mmap_lock.
131 *
132 * Hence to serialise fully against both syscall and mmap based IO, we need to
133 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
134 * both taken in places where we need to invalidate the page cache in a race
135 * free manner (e.g. truncate, hole punch and other extent manipulation
136 * functions).
137 */
138 void
139 xfs_ilock(
140 xfs_inode_t *ip,
141 uint lock_flags)
142 {
143 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
144
145 xfs_lock_flags_assert(lock_flags);
146
147 if (lock_flags & XFS_IOLOCK_EXCL) {
148 down_write_nested(&VFS_I(ip)->i_rwsem,
149 XFS_IOLOCK_DEP(lock_flags));
150 } else if (lock_flags & XFS_IOLOCK_SHARED) {
151 down_read_nested(&VFS_I(ip)->i_rwsem,
152 XFS_IOLOCK_DEP(lock_flags));
153 }
154
155 if (lock_flags & XFS_MMAPLOCK_EXCL) {
156 down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
157 XFS_MMAPLOCK_DEP(lock_flags));
158 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
159 down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
160 XFS_MMAPLOCK_DEP(lock_flags));
161 }
162
163 if (lock_flags & XFS_ILOCK_EXCL)
164 down_write_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
165 else if (lock_flags & XFS_ILOCK_SHARED)
166 down_read_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
167 }
168
169 /*
170 * This is just like xfs_ilock(), except that the caller
171 * is guaranteed not to sleep. It returns 1 if it gets
172 * the requested locks and 0 otherwise. If the IO lock is
173 * obtained but the inode lock cannot be, then the IO lock
174 * is dropped before returning.
175 *
176 * ip -- the inode being locked
177 * lock_flags -- this parameter indicates the inode's locks to be
178 * to be locked. See the comment for xfs_ilock() for a list
179 * of valid values.
180 */
181 int
182 xfs_ilock_nowait(
183 xfs_inode_t *ip,
184 uint lock_flags)
185 {
186 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
187
188 xfs_lock_flags_assert(lock_flags);
189
190 if (lock_flags & XFS_IOLOCK_EXCL) {
191 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
192 goto out;
193 } else if (lock_flags & XFS_IOLOCK_SHARED) {
194 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
195 goto out;
196 }
197
198 if (lock_flags & XFS_MMAPLOCK_EXCL) {
199 if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
200 goto out_undo_iolock;
201 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
202 if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
203 goto out_undo_iolock;
204 }
205
206 if (lock_flags & XFS_ILOCK_EXCL) {
207 if (!down_write_trylock(&ip->i_lock))
208 goto out_undo_mmaplock;
209 } else if (lock_flags & XFS_ILOCK_SHARED) {
210 if (!down_read_trylock(&ip->i_lock))
211 goto out_undo_mmaplock;
212 }
213 return 1;
214
215 out_undo_mmaplock:
216 if (lock_flags & XFS_MMAPLOCK_EXCL)
217 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
218 else if (lock_flags & XFS_MMAPLOCK_SHARED)
219 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
220 out_undo_iolock:
221 if (lock_flags & XFS_IOLOCK_EXCL)
222 up_write(&VFS_I(ip)->i_rwsem);
223 else if (lock_flags & XFS_IOLOCK_SHARED)
224 up_read(&VFS_I(ip)->i_rwsem);
225 out:
226 return 0;
227 }
228
229 /*
230 * xfs_iunlock() is used to drop the inode locks acquired with
231 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
232 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
233 * that we know which locks to drop.
234 *
235 * ip -- the inode being unlocked
236 * lock_flags -- this parameter indicates the inode's locks to be
237 * to be unlocked. See the comment for xfs_ilock() for a list
238 * of valid values for this parameter.
239 *
240 */
241 void
242 xfs_iunlock(
243 xfs_inode_t *ip,
244 uint lock_flags)
245 {
246 xfs_lock_flags_assert(lock_flags);
247
248 if (lock_flags & XFS_IOLOCK_EXCL)
249 up_write(&VFS_I(ip)->i_rwsem);
250 else if (lock_flags & XFS_IOLOCK_SHARED)
251 up_read(&VFS_I(ip)->i_rwsem);
252
253 if (lock_flags & XFS_MMAPLOCK_EXCL)
254 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
255 else if (lock_flags & XFS_MMAPLOCK_SHARED)
256 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
257
258 if (lock_flags & XFS_ILOCK_EXCL)
259 up_write(&ip->i_lock);
260 else if (lock_flags & XFS_ILOCK_SHARED)
261 up_read(&ip->i_lock);
262
263 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
264 }
265
266 /*
267 * give up write locks. the i/o lock cannot be held nested
268 * if it is being demoted.
269 */
270 void
271 xfs_ilock_demote(
272 xfs_inode_t *ip,
273 uint lock_flags)
274 {
275 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
276 ASSERT((lock_flags &
277 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
278
279 if (lock_flags & XFS_ILOCK_EXCL)
280 downgrade_write(&ip->i_lock);
281 if (lock_flags & XFS_MMAPLOCK_EXCL)
282 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
283 if (lock_flags & XFS_IOLOCK_EXCL)
284 downgrade_write(&VFS_I(ip)->i_rwsem);
285
286 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
287 }
288
289 void
290 xfs_assert_ilocked(
291 struct xfs_inode *ip,
292 uint lock_flags)
293 {
294 /*
295 * Sometimes we assert the ILOCK is held exclusively, but we're in
296 * a workqueue, so lockdep doesn't know we're the owner.
297 */
298 if (lock_flags & XFS_ILOCK_SHARED)
299 rwsem_assert_held(&ip->i_lock);
300 else if (lock_flags & XFS_ILOCK_EXCL)
301 rwsem_assert_held_write_nolockdep(&ip->i_lock);
302
303 if (lock_flags & XFS_MMAPLOCK_SHARED)
304 rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
305 else if (lock_flags & XFS_MMAPLOCK_EXCL)
306 rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
307
308 if (lock_flags & XFS_IOLOCK_SHARED)
309 rwsem_assert_held(&VFS_I(ip)->i_rwsem);
310 else if (lock_flags & XFS_IOLOCK_EXCL)
311 rwsem_assert_held_write(&VFS_I(ip)->i_rwsem);
312 }
313
314 /*
315 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
316 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
317 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
318 * errors and warnings.
319 */
320 #if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
321 static bool
322 xfs_lockdep_subclass_ok(
323 int subclass)
324 {
325 return subclass < MAX_LOCKDEP_SUBCLASSES;
326 }
327 #else
328 #define xfs_lockdep_subclass_ok(subclass) (true)
329 #endif
330
331 /*
332 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
333 * value. This can be called for any type of inode lock combination, including
334 * parent locking. Care must be taken to ensure we don't overrun the subclass
335 * storage fields in the class mask we build.
336 */
337 static inline uint
338 xfs_lock_inumorder(
339 uint lock_mode,
340 uint subclass)
341 {
342 uint class = 0;
343
344 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
345 XFS_ILOCK_RTSUM)));
346 ASSERT(xfs_lockdep_subclass_ok(subclass));
347
348 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
349 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
350 class += subclass << XFS_IOLOCK_SHIFT;
351 }
352
353 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
354 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
355 class += subclass << XFS_MMAPLOCK_SHIFT;
356 }
357
358 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
359 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
360 class += subclass << XFS_ILOCK_SHIFT;
361 }
362
363 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
364 }
365
366 /*
367 * The following routine will lock n inodes in exclusive mode. We assume the
368 * caller calls us with the inodes in i_ino order.
369 *
370 * We need to detect deadlock where an inode that we lock is in the AIL and we
371 * start waiting for another inode that is locked by a thread in a long running
372 * transaction (such as truncate). This can result in deadlock since the long
373 * running trans might need to wait for the inode we just locked in order to
374 * push the tail and free space in the log.
375 *
376 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
377 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
378 * lock more than one at a time, lockdep will report false positives saying we
379 * have violated locking orders.
380 */
381 void
382 xfs_lock_inodes(
383 struct xfs_inode **ips,
384 int inodes,
385 uint lock_mode)
386 {
387 int attempts = 0;
388 uint i;
389 int j;
390 bool try_lock;
391 struct xfs_log_item *lp;
392
393 /*
394 * Currently supports between 2 and 5 inodes with exclusive locking. We
395 * support an arbitrary depth of locking here, but absolute limits on
396 * inodes depend on the type of locking and the limits placed by
397 * lockdep annotations in xfs_lock_inumorder. These are all checked by
398 * the asserts.
399 */
400 ASSERT(ips && inodes >= 2 && inodes <= 5);
401 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
402 XFS_ILOCK_EXCL));
403 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
404 XFS_ILOCK_SHARED)));
405 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
406 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
407 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
408 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
409
410 if (lock_mode & XFS_IOLOCK_EXCL) {
411 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
412 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
413 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
414
415 again:
416 try_lock = false;
417 i = 0;
418 for (; i < inodes; i++) {
419 ASSERT(ips[i]);
420
421 if (i && (ips[i] == ips[i - 1])) /* Already locked */
422 continue;
423
424 /*
425 * If try_lock is not set yet, make sure all locked inodes are
426 * not in the AIL. If any are, set try_lock to be used later.
427 */
428 if (!try_lock) {
429 for (j = (i - 1); j >= 0 && !try_lock; j--) {
430 lp = &ips[j]->i_itemp->ili_item;
431 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
432 try_lock = true;
433 }
434 }
435
436 /*
437 * If any of the previous locks we have locked is in the AIL,
438 * we must TRY to get the second and subsequent locks. If
439 * we can't get any, we must release all we have
440 * and try again.
441 */
442 if (!try_lock) {
443 xfs_ilock(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 */
482 void
483 xfs_lock_two_inodes(
484 struct xfs_inode *ip0,
485 uint ip0_mode,
486 struct xfs_inode *ip1,
487 uint ip1_mode)
488 {
489 int attempts = 0;
490 struct xfs_log_item *lp;
491
492 ASSERT(hweight32(ip0_mode) == 1);
493 ASSERT(hweight32(ip1_mode) == 1);
494 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
495 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
496 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
497 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
498 ASSERT(ip0->i_ino != ip1->i_ino);
499
500 if (ip0->i_ino > ip1->i_ino) {
501 swap(ip0, ip1);
502 swap(ip0_mode, ip1_mode);
503 }
504
505 again:
506 xfs_ilock(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 */
532 int
533 xfs_lookup(
534 struct xfs_inode *dp,
535 const struct xfs_name *name,
536 struct xfs_inode **ipp,
537 struct xfs_name *ci_name)
538 {
539 xfs_ino_t inum;
540 int error;
541
542 trace_xfs_lookup(dp, 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 return 0;
558
559 out_free_name:
560 if (ci_name)
561 kfree(ci_name->name);
562 out_unlock:
563 *ipp = NULL;
564 return error;
565 }
566
567 /*
568 * Initialise a newly allocated inode and return the in-core inode to the
569 * caller locked exclusively.
570 *
571 * Caller is responsible for unlocking the inode manually upon return
572 */
573 int
574 xfs_icreate(
575 struct xfs_trans *tp,
576 xfs_ino_t ino,
577 const struct xfs_icreate_args *args,
578 struct xfs_inode **ipp)
579 {
580 struct xfs_mount *mp = tp->t_mountp;
581 struct xfs_inode *ip = NULL;
582 int error;
583
584 /*
585 * Get the in-core inode with the lock held exclusively to prevent
586 * others from looking at until we're done.
587 */
588 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
589 if (error)
590 return error;
591
592 ASSERT(ip != NULL);
593 xfs_trans_ijoin(tp, ip, 0);
594 xfs_inode_init(tp, args, ip);
595
596 /* now that we have an i_mode we can setup the inode structure */
597 xfs_setup_inode(ip);
598
599 *ipp = ip;
600 return 0;
601 }
602
603 /* Return dquots for the ids that will be assigned to a new file. */
604 int
605 xfs_icreate_dqalloc(
606 const struct xfs_icreate_args *args,
607 struct xfs_dquot **udqpp,
608 struct xfs_dquot **gdqpp,
609 struct xfs_dquot **pdqpp)
610 {
611 struct inode *dir = VFS_I(args->pip);
612 kuid_t uid = GLOBAL_ROOT_UID;
613 kgid_t gid = GLOBAL_ROOT_GID;
614 prid_t prid = 0;
615 unsigned int flags = XFS_QMOPT_QUOTALL;
616
617 if (args->idmap) {
618 /*
619 * The uid/gid computation code must match what the VFS uses to
620 * assign i_[ug]id. INHERIT adjusts the gid computation for
621 * setgid/grpid systems.
622 */
623 uid = mapped_fsuid(args->idmap, i_user_ns(dir));
624 gid = mapped_fsgid(args->idmap, i_user_ns(dir));
625 prid = xfs_get_initial_prid(args->pip);
626 flags |= XFS_QMOPT_INHERIT;
627 }
628
629 *udqpp = *gdqpp = *pdqpp = NULL;
630
631 return xfs_qm_vop_dqalloc(args->pip, uid, gid, prid, flags, udqpp,
632 gdqpp, pdqpp);
633 }
634
635 int
636 xfs_create(
637 const struct xfs_icreate_args *args,
638 struct xfs_name *name,
639 struct xfs_inode **ipp)
640 {
641 struct xfs_inode *dp = args->pip;
642 struct xfs_dir_update du = {
643 .dp = dp,
644 .name = name,
645 };
646 struct xfs_mount *mp = dp->i_mount;
647 struct xfs_trans *tp = NULL;
648 struct xfs_dquot *udqp;
649 struct xfs_dquot *gdqp;
650 struct xfs_dquot *pdqp;
651 struct xfs_trans_res *tres;
652 xfs_ino_t ino;
653 bool unlock_dp_on_error = false;
654 bool is_dir = S_ISDIR(args->mode);
655 uint resblks;
656 int error;
657
658 trace_xfs_create(dp, name);
659
660 if (xfs_is_shutdown(mp))
661 return -EIO;
662 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
663 return -EIO;
664
665 /* Make sure that we have allocated dquot(s) on disk. */
666 error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
667 if (error)
668 return error;
669
670 if (is_dir) {
671 resblks = xfs_mkdir_space_res(mp, name->len);
672 tres = &M_RES(mp)->tr_mkdir;
673 } else {
674 resblks = xfs_create_space_res(mp, name->len);
675 tres = &M_RES(mp)->tr_create;
676 }
677
678 error = xfs_parent_start(mp, &du.ppargs);
679 if (error)
680 goto out_release_dquots;
681
682 /*
683 * Initially assume that the file does not exist and
684 * reserve the resources for that case. If that is not
685 * the case we'll drop the one we have and get a more
686 * appropriate transaction later.
687 */
688 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
689 &tp);
690 if (error == -ENOSPC) {
691 /* flush outstanding delalloc blocks and retry */
692 xfs_flush_inodes(mp);
693 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
694 resblks, &tp);
695 }
696 if (error)
697 goto out_parent;
698
699 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
700 unlock_dp_on_error = true;
701
702 /*
703 * A newly created regular or special file just has one directory
704 * entry pointing to them, but a directory also the "." entry
705 * pointing to itself.
706 */
707 error = xfs_dialloc(&tp, dp->i_ino, args->mode, &ino);
708 if (!error)
709 error = xfs_icreate(tp, ino, args, &du.ip);
710 if (error)
711 goto out_trans_cancel;
712
713 /*
714 * Now we join the directory inode to the transaction. We do not do it
715 * earlier because xfs_dialloc might commit the previous transaction
716 * (and release all the locks). An error from here on will result in
717 * the transaction cancel unlocking dp so don't do it explicitly in the
718 * error path.
719 */
720 xfs_trans_ijoin(tp, dp, 0);
721
722 error = xfs_dir_create_child(tp, resblks, &du);
723 if (error)
724 goto out_trans_cancel;
725
726 /*
727 * If this is a synchronous mount, make sure that the
728 * create transaction goes to disk before returning to
729 * the user.
730 */
731 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
732 xfs_trans_set_sync(tp);
733
734 /*
735 * Attach the dquot(s) to the inodes and modify them incore.
736 * These ids of the inode couldn't have changed since the new
737 * inode has been locked ever since it was created.
738 */
739 xfs_qm_vop_create_dqattach(tp, du.ip, udqp, gdqp, pdqp);
740
741 error = xfs_trans_commit(tp);
742 if (error)
743 goto out_release_inode;
744
745 xfs_qm_dqrele(udqp);
746 xfs_qm_dqrele(gdqp);
747 xfs_qm_dqrele(pdqp);
748
749 *ipp = du.ip;
750 xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
751 xfs_iunlock(dp, XFS_ILOCK_EXCL);
752 xfs_parent_finish(mp, du.ppargs);
753 return 0;
754
755 out_trans_cancel:
756 xfs_trans_cancel(tp);
757 out_release_inode:
758 /*
759 * Wait until after the current transaction is aborted to finish the
760 * setup of the inode and release the inode. This prevents recursive
761 * transactions and deadlocks from xfs_inactive.
762 */
763 if (du.ip) {
764 xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
765 xfs_finish_inode_setup(du.ip);
766 xfs_irele(du.ip);
767 }
768 out_parent:
769 xfs_parent_finish(mp, du.ppargs);
770 out_release_dquots:
771 xfs_qm_dqrele(udqp);
772 xfs_qm_dqrele(gdqp);
773 xfs_qm_dqrele(pdqp);
774
775 if (unlock_dp_on_error)
776 xfs_iunlock(dp, XFS_ILOCK_EXCL);
777 return error;
778 }
779
780 int
781 xfs_create_tmpfile(
782 const struct xfs_icreate_args *args,
783 struct xfs_inode **ipp)
784 {
785 struct xfs_inode *dp = args->pip;
786 struct xfs_mount *mp = dp->i_mount;
787 struct xfs_inode *ip = NULL;
788 struct xfs_trans *tp = NULL;
789 struct xfs_dquot *udqp;
790 struct xfs_dquot *gdqp;
791 struct xfs_dquot *pdqp;
792 struct xfs_trans_res *tres;
793 xfs_ino_t ino;
794 uint resblks;
795 int error;
796
797 ASSERT(args->flags & XFS_ICREATE_TMPFILE);
798
799 if (xfs_is_shutdown(mp))
800 return -EIO;
801
802 /* Make sure that we have allocated dquot(s) on disk. */
803 error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
804 if (error)
805 return error;
806
807 resblks = XFS_IALLOC_SPACE_RES(mp);
808 tres = &M_RES(mp)->tr_create_tmpfile;
809
810 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
811 &tp);
812 if (error)
813 goto out_release_dquots;
814
815 error = xfs_dialloc(&tp, dp->i_ino, args->mode, &ino);
816 if (!error)
817 error = xfs_icreate(tp, ino, args, &ip);
818 if (error)
819 goto out_trans_cancel;
820
821 if (xfs_has_wsync(mp))
822 xfs_trans_set_sync(tp);
823
824 /*
825 * Attach the dquot(s) to the inodes and modify them incore.
826 * These ids of the inode couldn't have changed since the new
827 * inode has been locked ever since it was created.
828 */
829 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
830
831 error = xfs_iunlink(tp, ip);
832 if (error)
833 goto out_trans_cancel;
834
835 error = xfs_trans_commit(tp);
836 if (error)
837 goto out_release_inode;
838
839 xfs_qm_dqrele(udqp);
840 xfs_qm_dqrele(gdqp);
841 xfs_qm_dqrele(pdqp);
842
843 *ipp = ip;
844 xfs_iunlock(ip, XFS_ILOCK_EXCL);
845 return 0;
846
847 out_trans_cancel:
848 xfs_trans_cancel(tp);
849 out_release_inode:
850 /*
851 * Wait until after the current transaction is aborted to finish the
852 * setup of the inode and release the inode. This prevents recursive
853 * transactions and deadlocks from xfs_inactive.
854 */
855 if (ip) {
856 xfs_iunlock(ip, XFS_ILOCK_EXCL);
857 xfs_finish_inode_setup(ip);
858 xfs_irele(ip);
859 }
860 out_release_dquots:
861 xfs_qm_dqrele(udqp);
862 xfs_qm_dqrele(gdqp);
863 xfs_qm_dqrele(pdqp);
864
865 return error;
866 }
867
868 int
869 xfs_link(
870 struct xfs_inode *tdp,
871 struct xfs_inode *sip,
872 struct xfs_name *target_name)
873 {
874 struct xfs_dir_update du = {
875 .dp = tdp,
876 .name = target_name,
877 .ip = sip,
878 };
879 struct xfs_mount *mp = tdp->i_mount;
880 struct xfs_trans *tp;
881 int error, nospace_error = 0;
882 int resblks;
883
884 trace_xfs_link(tdp, target_name);
885
886 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
887
888 if (xfs_is_shutdown(mp))
889 return -EIO;
890 if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
891 return -EIO;
892
893 error = xfs_qm_dqattach(sip);
894 if (error)
895 goto std_return;
896
897 error = xfs_qm_dqattach(tdp);
898 if (error)
899 goto std_return;
900
901 error = xfs_parent_start(mp, &du.ppargs);
902 if (error)
903 goto std_return;
904
905 resblks = xfs_link_space_res(mp, target_name->len);
906 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
907 &tp, &nospace_error);
908 if (error)
909 goto out_parent;
910
911 /*
912 * We don't allow reservationless or quotaless hardlinking when parent
913 * pointers are enabled because we can't back out if the xattrs must
914 * grow.
915 */
916 if (du.ppargs && nospace_error) {
917 error = nospace_error;
918 goto error_return;
919 }
920
921 /*
922 * If we are using project inheritance, we only allow hard link
923 * creation in our tree when the project IDs are the same; else
924 * the tree quota mechanism could be circumvented.
925 */
926 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
927 tdp->i_projid != sip->i_projid)) {
928 /*
929 * Project quota setup skips special files which can
930 * leave inodes in a PROJINHERIT directory without a
931 * project ID set. We need to allow links to be made
932 * to these "project-less" inodes because userspace
933 * expects them to succeed after project ID setup,
934 * but everything else should be rejected.
935 */
936 if (!special_file(VFS_I(sip)->i_mode) ||
937 sip->i_projid != 0) {
938 error = -EXDEV;
939 goto error_return;
940 }
941 }
942
943 error = xfs_dir_add_child(tp, resblks, &du);
944 if (error)
945 goto error_return;
946
947 /*
948 * If this is a synchronous mount, make sure that the
949 * link transaction goes to disk before returning to
950 * the user.
951 */
952 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
953 xfs_trans_set_sync(tp);
954
955 error = xfs_trans_commit(tp);
956 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
957 xfs_iunlock(sip, XFS_ILOCK_EXCL);
958 xfs_parent_finish(mp, du.ppargs);
959 return error;
960
961 error_return:
962 xfs_trans_cancel(tp);
963 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
964 xfs_iunlock(sip, XFS_ILOCK_EXCL);
965 out_parent:
966 xfs_parent_finish(mp, du.ppargs);
967 std_return:
968 if (error == -ENOSPC && nospace_error)
969 error = nospace_error;
970 return error;
971 }
972
973 /* Clear the reflink flag and the cowblocks tag if possible. */
974 static void
975 xfs_itruncate_clear_reflink_flags(
976 struct xfs_inode *ip)
977 {
978 struct xfs_ifork *dfork;
979 struct xfs_ifork *cfork;
980
981 if (!xfs_is_reflink_inode(ip))
982 return;
983 dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
984 cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
985 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
986 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
987 if (cfork->if_bytes == 0)
988 xfs_inode_clear_cowblocks_tag(ip);
989 }
990
991 /*
992 * Free up the underlying blocks past new_size. The new size must be smaller
993 * than the current size. This routine can be used both for the attribute and
994 * data fork, and does not modify the inode size, which is left to the caller.
995 *
996 * The transaction passed to this routine must have made a permanent log
997 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
998 * given transaction and start new ones, so make sure everything involved in
999 * the transaction is tidy before calling here. Some transaction will be
1000 * returned to the caller to be committed. The incoming transaction must
1001 * already include the inode, and both inode locks must be held exclusively.
1002 * The inode must also be "held" within the transaction. On return the inode
1003 * will be "held" within the returned transaction. This routine does NOT
1004 * require any disk space to be reserved for it within the transaction.
1005 *
1006 * If we get an error, we must return with the inode locked and linked into the
1007 * current transaction. This keeps things simple for the higher level code,
1008 * because it always knows that the inode is locked and held in the transaction
1009 * that returns to it whether errors occur or not. We don't mark the inode
1010 * dirty on error so that transactions can be easily aborted if possible.
1011 */
1012 int
1013 xfs_itruncate_extents_flags(
1014 struct xfs_trans **tpp,
1015 struct xfs_inode *ip,
1016 int whichfork,
1017 xfs_fsize_t new_size,
1018 int flags)
1019 {
1020 struct xfs_mount *mp = ip->i_mount;
1021 struct xfs_trans *tp = *tpp;
1022 xfs_fileoff_t first_unmap_block;
1023 int error = 0;
1024
1025 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1026 if (atomic_read(&VFS_I(ip)->i_count))
1027 xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1028 ASSERT(new_size <= XFS_ISIZE(ip));
1029 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1030 ASSERT(ip->i_itemp != NULL);
1031 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1032 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1033
1034 trace_xfs_itruncate_extents_start(ip, new_size);
1035
1036 flags |= xfs_bmapi_aflag(whichfork);
1037
1038 /*
1039 * Since it is possible for space to become allocated beyond
1040 * the end of the file (in a crash where the space is allocated
1041 * but the inode size is not yet updated), simply remove any
1042 * blocks which show up between the new EOF and the maximum
1043 * possible file size.
1044 *
1045 * We have to free all the blocks to the bmbt maximum offset, even if
1046 * the page cache can't scale that far.
1047 */
1048 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1049 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1050 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1051 return 0;
1052 }
1053
1054 error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1055 XFS_MAX_FILEOFF);
1056 if (error)
1057 goto out;
1058
1059 if (whichfork == XFS_DATA_FORK) {
1060 /* Remove all pending CoW reservations. */
1061 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1062 first_unmap_block, XFS_MAX_FILEOFF, true);
1063 if (error)
1064 goto out;
1065
1066 xfs_itruncate_clear_reflink_flags(ip);
1067 }
1068
1069 /*
1070 * Always re-log the inode so that our permanent transaction can keep
1071 * on rolling it forward in the log.
1072 */
1073 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1074
1075 trace_xfs_itruncate_extents_end(ip, new_size);
1076
1077 out:
1078 *tpp = tp;
1079 return error;
1080 }
1081
1082 int
1083 xfs_release(
1084 xfs_inode_t *ip)
1085 {
1086 xfs_mount_t *mp = ip->i_mount;
1087 int error = 0;
1088
1089 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1090 return 0;
1091
1092 /* If this is a read-only mount, don't do this (would generate I/O) */
1093 if (xfs_is_readonly(mp))
1094 return 0;
1095
1096 if (!xfs_is_shutdown(mp)) {
1097 int truncated;
1098
1099 /*
1100 * If we previously truncated this file and removed old data
1101 * in the process, we want to initiate "early" writeout on
1102 * the last close. This is an attempt to combat the notorious
1103 * NULL files problem which is particularly noticeable from a
1104 * truncate down, buffered (re-)write (delalloc), followed by
1105 * a crash. What we are effectively doing here is
1106 * significantly reducing the time window where we'd otherwise
1107 * be exposed to that problem.
1108 */
1109 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1110 if (truncated) {
1111 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1112 if (ip->i_delayed_blks > 0) {
1113 error = filemap_flush(VFS_I(ip)->i_mapping);
1114 if (error)
1115 return error;
1116 }
1117 }
1118 }
1119
1120 if (VFS_I(ip)->i_nlink == 0)
1121 return 0;
1122
1123 /*
1124 * If we can't get the iolock just skip truncating the blocks past EOF
1125 * because we could deadlock with the mmap_lock otherwise. We'll get
1126 * another chance to drop them once the last reference to the inode is
1127 * dropped, so we'll never leak blocks permanently.
1128 */
1129 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1130 return 0;
1131
1132 if (xfs_can_free_eofblocks(ip)) {
1133 /*
1134 * Check if the inode is being opened, written and closed
1135 * frequently and we have delayed allocation blocks outstanding
1136 * (e.g. streaming writes from the NFS server), truncating the
1137 * blocks past EOF will cause fragmentation to occur.
1138 *
1139 * In this case don't do the truncation, but we have to be
1140 * careful how we detect this case. Blocks beyond EOF show up as
1141 * i_delayed_blks even when the inode is clean, so we need to
1142 * truncate them away first before checking for a dirty release.
1143 * Hence on the first dirty close we will still remove the
1144 * speculative allocation, but after that we will leave it in
1145 * place.
1146 */
1147 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1148 goto out_unlock;
1149
1150 error = xfs_free_eofblocks(ip);
1151 if (error)
1152 goto out_unlock;
1153
1154 /* delalloc blocks after truncation means it really is dirty */
1155 if (ip->i_delayed_blks)
1156 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1157 }
1158
1159 out_unlock:
1160 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1161 return error;
1162 }
1163
1164 /*
1165 * Mark all the buffers attached to this directory stale. In theory we should
1166 * never be freeing a directory with any blocks at all, but this covers the
1167 * case where we've recovered a directory swap with a "temporary" directory
1168 * created by online repair and now need to dump it.
1169 */
1170 STATIC void
1171 xfs_inactive_dir(
1172 struct xfs_inode *dp)
1173 {
1174 struct xfs_iext_cursor icur;
1175 struct xfs_bmbt_irec got;
1176 struct xfs_mount *mp = dp->i_mount;
1177 struct xfs_da_geometry *geo = mp->m_dir_geo;
1178 struct xfs_ifork *ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1179 xfs_fileoff_t off;
1180
1181 /*
1182 * Invalidate each directory block. All directory blocks are of
1183 * fsbcount length and alignment, so we only need to walk those same
1184 * offsets. We hold the only reference to this inode, so we must wait
1185 * for the buffer locks.
1186 */
1187 for_each_xfs_iext(ifp, &icur, &got) {
1188 for (off = round_up(got.br_startoff, geo->fsbcount);
1189 off < got.br_startoff + got.br_blockcount;
1190 off += geo->fsbcount) {
1191 struct xfs_buf *bp = NULL;
1192 xfs_fsblock_t fsbno;
1193 int error;
1194
1195 fsbno = (off - got.br_startoff) + got.br_startblock;
1196 error = xfs_buf_incore(mp->m_ddev_targp,
1197 XFS_FSB_TO_DADDR(mp, fsbno),
1198 XFS_FSB_TO_BB(mp, geo->fsbcount),
1199 XBF_LIVESCAN, &bp);
1200 if (error)
1201 continue;
1202
1203 xfs_buf_stale(bp);
1204 xfs_buf_relse(bp);
1205 }
1206 }
1207 }
1208
1209 /*
1210 * xfs_inactive_truncate
1211 *
1212 * Called to perform a truncate when an inode becomes unlinked.
1213 */
1214 STATIC int
1215 xfs_inactive_truncate(
1216 struct xfs_inode *ip)
1217 {
1218 struct xfs_mount *mp = ip->i_mount;
1219 struct xfs_trans *tp;
1220 int error;
1221
1222 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1223 if (error) {
1224 ASSERT(xfs_is_shutdown(mp));
1225 return error;
1226 }
1227 xfs_ilock(ip, XFS_ILOCK_EXCL);
1228 xfs_trans_ijoin(tp, ip, 0);
1229
1230 /*
1231 * Log the inode size first to prevent stale data exposure in the event
1232 * of a system crash before the truncate completes. See the related
1233 * comment in xfs_vn_setattr_size() for details.
1234 */
1235 ip->i_disk_size = 0;
1236 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1237
1238 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1239 if (error)
1240 goto error_trans_cancel;
1241
1242 ASSERT(ip->i_df.if_nextents == 0);
1243
1244 error = xfs_trans_commit(tp);
1245 if (error)
1246 goto error_unlock;
1247
1248 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1249 return 0;
1250
1251 error_trans_cancel:
1252 xfs_trans_cancel(tp);
1253 error_unlock:
1254 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1255 return error;
1256 }
1257
1258 /*
1259 * xfs_inactive_ifree()
1260 *
1261 * Perform the inode free when an inode is unlinked.
1262 */
1263 STATIC int
1264 xfs_inactive_ifree(
1265 struct xfs_inode *ip)
1266 {
1267 struct xfs_mount *mp = ip->i_mount;
1268 struct xfs_trans *tp;
1269 int error;
1270
1271 /*
1272 * We try to use a per-AG reservation for any block needed by the finobt
1273 * tree, but as the finobt feature predates the per-AG reservation
1274 * support a degraded file system might not have enough space for the
1275 * reservation at mount time. In that case try to dip into the reserved
1276 * pool and pray.
1277 *
1278 * Send a warning if the reservation does happen to fail, as the inode
1279 * now remains allocated and sits on the unlinked list until the fs is
1280 * repaired.
1281 */
1282 if (unlikely(mp->m_finobt_nores)) {
1283 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1284 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1285 &tp);
1286 } else {
1287 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1288 }
1289 if (error) {
1290 if (error == -ENOSPC) {
1291 xfs_warn_ratelimited(mp,
1292 "Failed to remove inode(s) from unlinked list. "
1293 "Please free space, unmount and run xfs_repair.");
1294 } else {
1295 ASSERT(xfs_is_shutdown(mp));
1296 }
1297 return error;
1298 }
1299
1300 /*
1301 * We do not hold the inode locked across the entire rolling transaction
1302 * here. We only need to hold it for the first transaction that
1303 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1304 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1305 * here breaks the relationship between cluster buffer invalidation and
1306 * stale inode invalidation on cluster buffer item journal commit
1307 * completion, and can result in leaving dirty stale inodes hanging
1308 * around in memory.
1309 *
1310 * We have no need for serialising this inode operation against other
1311 * operations - we freed the inode and hence reallocation is required
1312 * and that will serialise on reallocating the space the deferops need
1313 * to free. Hence we can unlock the inode on the first commit of
1314 * the transaction rather than roll it right through the deferops. This
1315 * avoids relogging the XFS_ISTALE inode.
1316 *
1317 * We check that xfs_ifree() hasn't grown an internal transaction roll
1318 * by asserting that the inode is still locked when it returns.
1319 */
1320 xfs_ilock(ip, XFS_ILOCK_EXCL);
1321 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1322
1323 error = xfs_ifree(tp, ip);
1324 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1325 if (error) {
1326 /*
1327 * If we fail to free the inode, shut down. The cancel
1328 * might do that, we need to make sure. Otherwise the
1329 * inode might be lost for a long time or forever.
1330 */
1331 if (!xfs_is_shutdown(mp)) {
1332 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1333 __func__, error);
1334 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1335 }
1336 xfs_trans_cancel(tp);
1337 return error;
1338 }
1339
1340 /*
1341 * Credit the quota account(s). The inode is gone.
1342 */
1343 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1344
1345 return xfs_trans_commit(tp);
1346 }
1347
1348 /*
1349 * Returns true if we need to update the on-disk metadata before we can free
1350 * the memory used by this inode. Updates include freeing post-eof
1351 * preallocations; freeing COW staging extents; and marking the inode free in
1352 * the inobt if it is on the unlinked list.
1353 */
1354 bool
1355 xfs_inode_needs_inactive(
1356 struct xfs_inode *ip)
1357 {
1358 struct xfs_mount *mp = ip->i_mount;
1359 struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1360
1361 /*
1362 * If the inode is already free, then there can be nothing
1363 * to clean up here.
1364 */
1365 if (VFS_I(ip)->i_mode == 0)
1366 return false;
1367
1368 /*
1369 * If this is a read-only mount, don't do this (would generate I/O)
1370 * unless we're in log recovery and cleaning the iunlinked list.
1371 */
1372 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1373 return false;
1374
1375 /* If the log isn't running, push inodes straight to reclaim. */
1376 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1377 return false;
1378
1379 /* Metadata inodes require explicit resource cleanup. */
1380 if (xfs_is_metadata_inode(ip))
1381 return false;
1382
1383 /* Want to clean out the cow blocks if there are any. */
1384 if (cow_ifp && cow_ifp->if_bytes > 0)
1385 return true;
1386
1387 /* Unlinked files must be freed. */
1388 if (VFS_I(ip)->i_nlink == 0)
1389 return true;
1390
1391 /*
1392 * This file isn't being freed, so check if there are post-eof blocks
1393 * to free.
1394 *
1395 * Note: don't bother with iolock here since lockdep complains about
1396 * acquiring it in reclaim context. We have the only reference to the
1397 * inode at this point anyways.
1398 */
1399 return xfs_can_free_eofblocks(ip);
1400 }
1401
1402 /*
1403 * Save health status somewhere, if we're dumping an inode with uncorrected
1404 * errors and online repair isn't running.
1405 */
1406 static inline void
1407 xfs_inactive_health(
1408 struct xfs_inode *ip)
1409 {
1410 struct xfs_mount *mp = ip->i_mount;
1411 struct xfs_perag *pag;
1412 unsigned int sick;
1413 unsigned int checked;
1414
1415 xfs_inode_measure_sickness(ip, &sick, &checked);
1416 if (!sick)
1417 return;
1418
1419 trace_xfs_inode_unfixed_corruption(ip, sick);
1420
1421 if (sick & XFS_SICK_INO_FORGET)
1422 return;
1423
1424 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1425 if (!pag) {
1426 /* There had better still be a perag structure! */
1427 ASSERT(0);
1428 return;
1429 }
1430
1431 xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1432 xfs_perag_put(pag);
1433 }
1434
1435 /*
1436 * xfs_inactive
1437 *
1438 * This is called when the vnode reference count for the vnode
1439 * goes to zero. If the file has been unlinked, then it must
1440 * now be truncated. Also, we clear all of the read-ahead state
1441 * kept for the inode here since the file is now closed.
1442 */
1443 int
1444 xfs_inactive(
1445 xfs_inode_t *ip)
1446 {
1447 struct xfs_mount *mp;
1448 int error = 0;
1449 int truncate = 0;
1450
1451 /*
1452 * If the inode is already free, then there can be nothing
1453 * to clean up here.
1454 */
1455 if (VFS_I(ip)->i_mode == 0) {
1456 ASSERT(ip->i_df.if_broot_bytes == 0);
1457 goto out;
1458 }
1459
1460 mp = ip->i_mount;
1461 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1462
1463 xfs_inactive_health(ip);
1464
1465 /*
1466 * If this is a read-only mount, don't do this (would generate I/O)
1467 * unless we're in log recovery and cleaning the iunlinked list.
1468 */
1469 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1470 goto out;
1471
1472 /* Metadata inodes require explicit resource cleanup. */
1473 if (xfs_is_metadata_inode(ip))
1474 goto out;
1475
1476 /* Try to clean out the cow blocks if there are any. */
1477 if (xfs_inode_has_cow_data(ip))
1478 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1479
1480 if (VFS_I(ip)->i_nlink != 0) {
1481 /*
1482 * Note: don't bother with iolock here since lockdep complains
1483 * about acquiring it in reclaim context. We have the only
1484 * reference to the inode at this point anyways.
1485 */
1486 if (xfs_can_free_eofblocks(ip))
1487 error = xfs_free_eofblocks(ip);
1488
1489 goto out;
1490 }
1491
1492 if (S_ISREG(VFS_I(ip)->i_mode) &&
1493 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1494 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1495 truncate = 1;
1496
1497 if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1498 /*
1499 * If this inode is being inactivated during a quotacheck and
1500 * has not yet been scanned by quotacheck, we /must/ remove
1501 * the dquots from the inode before inactivation changes the
1502 * block and inode counts. Most probably this is a result of
1503 * reloading the incore iunlinked list to purge unrecovered
1504 * unlinked inodes.
1505 */
1506 xfs_qm_dqdetach(ip);
1507 } else {
1508 error = xfs_qm_dqattach(ip);
1509 if (error)
1510 goto out;
1511 }
1512
1513 if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1514 xfs_inactive_dir(ip);
1515 truncate = 1;
1516 }
1517
1518 if (S_ISLNK(VFS_I(ip)->i_mode))
1519 error = xfs_inactive_symlink(ip);
1520 else if (truncate)
1521 error = xfs_inactive_truncate(ip);
1522 if (error)
1523 goto out;
1524
1525 /*
1526 * If there are attributes associated with the file then blow them away
1527 * now. The code calls a routine that recursively deconstructs the
1528 * attribute fork. If also blows away the in-core attribute fork.
1529 */
1530 if (xfs_inode_has_attr_fork(ip)) {
1531 error = xfs_attr_inactive(ip);
1532 if (error)
1533 goto out;
1534 }
1535
1536 ASSERT(ip->i_forkoff == 0);
1537
1538 /*
1539 * Free the inode.
1540 */
1541 error = xfs_inactive_ifree(ip);
1542
1543 out:
1544 /*
1545 * We're done making metadata updates for this inode, so we can release
1546 * the attached dquots.
1547 */
1548 xfs_qm_dqdetach(ip);
1549 return error;
1550 }
1551
1552 /*
1553 * Find an inode on the unlinked list. This does not take references to the
1554 * inode as we have existence guarantees by holding the AGI buffer lock and that
1555 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1556 * don't find the inode in cache, then let the caller handle the situation.
1557 */
1558 struct xfs_inode *
1559 xfs_iunlink_lookup(
1560 struct xfs_perag *pag,
1561 xfs_agino_t agino)
1562 {
1563 struct xfs_inode *ip;
1564
1565 rcu_read_lock();
1566 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1567 if (!ip) {
1568 /* Caller can handle inode not being in memory. */
1569 rcu_read_unlock();
1570 return NULL;
1571 }
1572
1573 /*
1574 * Inode in RCU freeing limbo should not happen. Warn about this and
1575 * let the caller handle the failure.
1576 */
1577 if (WARN_ON_ONCE(!ip->i_ino)) {
1578 rcu_read_unlock();
1579 return NULL;
1580 }
1581 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1582 rcu_read_unlock();
1583 return ip;
1584 }
1585
1586 /*
1587 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1588 * to @prev_agino. Caller must hold the AGI to synchronize with other changes
1589 * to the unlinked list.
1590 */
1591 int
1592 xfs_iunlink_reload_next(
1593 struct xfs_trans *tp,
1594 struct xfs_buf *agibp,
1595 xfs_agino_t prev_agino,
1596 xfs_agino_t next_agino)
1597 {
1598 struct xfs_perag *pag = agibp->b_pag;
1599 struct xfs_mount *mp = pag->pag_mount;
1600 struct xfs_inode *next_ip = NULL;
1601 xfs_ino_t ino;
1602 int error;
1603
1604 ASSERT(next_agino != NULLAGINO);
1605
1606 #ifdef DEBUG
1607 rcu_read_lock();
1608 next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1609 ASSERT(next_ip == NULL);
1610 rcu_read_unlock();
1611 #endif
1612
1613 xfs_info_ratelimited(mp,
1614 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
1615 next_agino, pag->pag_agno);
1616
1617 /*
1618 * Use an untrusted lookup just to be cautious in case the AGI has been
1619 * corrupted and now points at a free inode. That shouldn't happen,
1620 * but we'd rather shut down now since we're already running in a weird
1621 * situation.
1622 */
1623 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, next_agino);
1624 error = xfs_iget(mp, tp, ino, XFS_IGET_UNTRUSTED, 0, &next_ip);
1625 if (error) {
1626 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1627 return error;
1628 }
1629
1630 /* If this is not an unlinked inode, something is very wrong. */
1631 if (VFS_I(next_ip)->i_nlink != 0) {
1632 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1633 error = -EFSCORRUPTED;
1634 goto rele;
1635 }
1636
1637 next_ip->i_prev_unlinked = prev_agino;
1638 trace_xfs_iunlink_reload_next(next_ip);
1639 rele:
1640 ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
1641 if (xfs_is_quotacheck_running(mp) && next_ip)
1642 xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1643 xfs_irele(next_ip);
1644 return error;
1645 }
1646
1647 /*
1648 * Look up the inode number specified and if it is not already marked XFS_ISTALE
1649 * mark it stale. We should only find clean inodes in this lookup that aren't
1650 * already stale.
1651 */
1652 static void
1653 xfs_ifree_mark_inode_stale(
1654 struct xfs_perag *pag,
1655 struct xfs_inode *free_ip,
1656 xfs_ino_t inum)
1657 {
1658 struct xfs_mount *mp = pag->pag_mount;
1659 struct xfs_inode_log_item *iip;
1660 struct xfs_inode *ip;
1661
1662 retry:
1663 rcu_read_lock();
1664 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1665
1666 /* Inode not in memory, nothing to do */
1667 if (!ip) {
1668 rcu_read_unlock();
1669 return;
1670 }
1671
1672 /*
1673 * because this is an RCU protected lookup, we could find a recently
1674 * freed or even reallocated inode during the lookup. We need to check
1675 * under the i_flags_lock for a valid inode here. Skip it if it is not
1676 * valid, the wrong inode or stale.
1677 */
1678 spin_lock(&ip->i_flags_lock);
1679 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1680 goto out_iflags_unlock;
1681
1682 /*
1683 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
1684 * other inodes that we did not find in the list attached to the buffer
1685 * and are not already marked stale. If we can't lock it, back off and
1686 * retry.
1687 */
1688 if (ip != free_ip) {
1689 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1690 spin_unlock(&ip->i_flags_lock);
1691 rcu_read_unlock();
1692 delay(1);
1693 goto retry;
1694 }
1695 }
1696 ip->i_flags |= XFS_ISTALE;
1697
1698 /*
1699 * If the inode is flushing, it is already attached to the buffer. All
1700 * we needed to do here is mark the inode stale so buffer IO completion
1701 * will remove it from the AIL.
1702 */
1703 iip = ip->i_itemp;
1704 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1705 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1706 ASSERT(iip->ili_last_fields);
1707 goto out_iunlock;
1708 }
1709
1710 /*
1711 * Inodes not attached to the buffer can be released immediately.
1712 * Everything else has to go through xfs_iflush_abort() on journal
1713 * commit as the flock synchronises removal of the inode from the
1714 * cluster buffer against inode reclaim.
1715 */
1716 if (!iip || list_empty(&iip->ili_item.li_bio_list))
1717 goto out_iunlock;
1718
1719 __xfs_iflags_set(ip, XFS_IFLUSHING);
1720 spin_unlock(&ip->i_flags_lock);
1721 rcu_read_unlock();
1722
1723 /* we have a dirty inode in memory that has not yet been flushed. */
1724 spin_lock(&iip->ili_lock);
1725 iip->ili_last_fields = iip->ili_fields;
1726 iip->ili_fields = 0;
1727 iip->ili_fsync_fields = 0;
1728 spin_unlock(&iip->ili_lock);
1729 ASSERT(iip->ili_last_fields);
1730
1731 if (ip != free_ip)
1732 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1733 return;
1734
1735 out_iunlock:
1736 if (ip != free_ip)
1737 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1738 out_iflags_unlock:
1739 spin_unlock(&ip->i_flags_lock);
1740 rcu_read_unlock();
1741 }
1742
1743 /*
1744 * A big issue when freeing the inode cluster is that we _cannot_ skip any
1745 * inodes that are in memory - they all must be marked stale and attached to
1746 * the cluster buffer.
1747 */
1748 static int
1749 xfs_ifree_cluster(
1750 struct xfs_trans *tp,
1751 struct xfs_perag *pag,
1752 struct xfs_inode *free_ip,
1753 struct xfs_icluster *xic)
1754 {
1755 struct xfs_mount *mp = free_ip->i_mount;
1756 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1757 struct xfs_buf *bp;
1758 xfs_daddr_t blkno;
1759 xfs_ino_t inum = xic->first_ino;
1760 int nbufs;
1761 int i, j;
1762 int ioffset;
1763 int error;
1764
1765 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1766
1767 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1768 /*
1769 * The allocation bitmap tells us which inodes of the chunk were
1770 * physically allocated. Skip the cluster if an inode falls into
1771 * a sparse region.
1772 */
1773 ioffset = inum - xic->first_ino;
1774 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1775 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1776 continue;
1777 }
1778
1779 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1780 XFS_INO_TO_AGBNO(mp, inum));
1781
1782 /*
1783 * We obtain and lock the backing buffer first in the process
1784 * here to ensure dirty inodes attached to the buffer remain in
1785 * the flushing state while we mark them stale.
1786 *
1787 * If we scan the in-memory inodes first, then buffer IO can
1788 * complete before we get a lock on it, and hence we may fail
1789 * to mark all the active inodes on the buffer stale.
1790 */
1791 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
1792 mp->m_bsize * igeo->blocks_per_cluster,
1793 XBF_UNMAPPED, &bp);
1794 if (error)
1795 return error;
1796
1797 /*
1798 * This buffer may not have been correctly initialised as we
1799 * didn't read it from disk. That's not important because we are
1800 * only using to mark the buffer as stale in the log, and to
1801 * attach stale cached inodes on it.
1802 *
1803 * For the inode that triggered the cluster freeing, this
1804 * attachment may occur in xfs_inode_item_precommit() after we
1805 * have marked this buffer stale. If this buffer was not in
1806 * memory before xfs_ifree_cluster() started, it will not be
1807 * marked XBF_DONE and this will cause problems later in
1808 * xfs_inode_item_precommit() when we trip over a (stale, !done)
1809 * buffer to attached to the transaction.
1810 *
1811 * Hence we have to mark the buffer as XFS_DONE here. This is
1812 * safe because we are also marking the buffer as XBF_STALE and
1813 * XFS_BLI_STALE. That means it will never be dispatched for
1814 * IO and it won't be unlocked until the cluster freeing has
1815 * been committed to the journal and the buffer unpinned. If it
1816 * is written, we want to know about it, and we want it to
1817 * fail. We can acheive this by adding a write verifier to the
1818 * buffer.
1819 */
1820 bp->b_flags |= XBF_DONE;
1821 bp->b_ops = &xfs_inode_buf_ops;
1822
1823 /*
1824 * Now we need to set all the cached clean inodes as XFS_ISTALE,
1825 * too. This requires lookups, and will skip inodes that we've
1826 * already marked XFS_ISTALE.
1827 */
1828 for (i = 0; i < igeo->inodes_per_cluster; i++)
1829 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
1830
1831 xfs_trans_stale_inode_buf(tp, bp);
1832 xfs_trans_binval(tp, bp);
1833 }
1834 return 0;
1835 }
1836
1837 /*
1838 * This is called to return an inode to the inode free list. The inode should
1839 * already be truncated to 0 length and have no pages associated with it. This
1840 * routine also assumes that the inode is already a part of the transaction.
1841 *
1842 * The on-disk copy of the inode will have been added to the list of unlinked
1843 * inodes in the AGI. We need to remove the inode from that list atomically with
1844 * respect to freeing it here.
1845 */
1846 int
1847 xfs_ifree(
1848 struct xfs_trans *tp,
1849 struct xfs_inode *ip)
1850 {
1851 struct xfs_mount *mp = ip->i_mount;
1852 struct xfs_perag *pag;
1853 struct xfs_icluster xic = { 0 };
1854 struct xfs_inode_log_item *iip = ip->i_itemp;
1855 int error;
1856
1857 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1858 ASSERT(VFS_I(ip)->i_nlink == 0);
1859 ASSERT(ip->i_df.if_nextents == 0);
1860 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1861 ASSERT(ip->i_nblocks == 0);
1862
1863 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1864
1865 error = xfs_inode_uninit(tp, pag, ip, &xic);
1866 if (error)
1867 goto out;
1868
1869 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1870 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1871
1872 /* Don't attempt to replay owner changes for a deleted inode */
1873 spin_lock(&iip->ili_lock);
1874 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1875 spin_unlock(&iip->ili_lock);
1876
1877 if (xic.deleted)
1878 error = xfs_ifree_cluster(tp, pag, ip, &xic);
1879 out:
1880 xfs_perag_put(pag);
1881 return error;
1882 }
1883
1884 /*
1885 * This is called to unpin an inode. The caller must have the inode locked
1886 * in at least shared mode so that the buffer cannot be subsequently pinned
1887 * once someone is waiting for it to be unpinned.
1888 */
1889 static void
1890 xfs_iunpin(
1891 struct xfs_inode *ip)
1892 {
1893 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1894
1895 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1896
1897 /* Give the log a push to start the unpinning I/O */
1898 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
1899
1900 }
1901
1902 static void
1903 __xfs_iunpin_wait(
1904 struct xfs_inode *ip)
1905 {
1906 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
1907 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1908
1909 xfs_iunpin(ip);
1910
1911 do {
1912 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1913 if (xfs_ipincount(ip))
1914 io_schedule();
1915 } while (xfs_ipincount(ip));
1916 finish_wait(wq, &wait.wq_entry);
1917 }
1918
1919 void
1920 xfs_iunpin_wait(
1921 struct xfs_inode *ip)
1922 {
1923 if (xfs_ipincount(ip))
1924 __xfs_iunpin_wait(ip);
1925 }
1926
1927 /*
1928 * Removing an inode from the namespace involves removing the directory entry
1929 * and dropping the link count on the inode. Removing the directory entry can
1930 * result in locking an AGF (directory blocks were freed) and removing a link
1931 * count can result in placing the inode on an unlinked list which results in
1932 * locking an AGI.
1933 *
1934 * The big problem here is that we have an ordering constraint on AGF and AGI
1935 * locking - inode allocation locks the AGI, then can allocate a new extent for
1936 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1937 * removes the inode from the unlinked list, requiring that we lock the AGI
1938 * first, and then freeing the inode can result in an inode chunk being freed
1939 * and hence freeing disk space requiring that we lock an AGF.
1940 *
1941 * Hence the ordering that is imposed by other parts of the code is AGI before
1942 * AGF. This means we cannot remove the directory entry before we drop the inode
1943 * reference count and put it on the unlinked list as this results in a lock
1944 * order of AGF then AGI, and this can deadlock against inode allocation and
1945 * freeing. Therefore we must drop the link counts before we remove the
1946 * directory entry.
1947 *
1948 * This is still safe from a transactional point of view - it is not until we
1949 * get to xfs_defer_finish() that we have the possibility of multiple
1950 * transactions in this operation. Hence as long as we remove the directory
1951 * entry and drop the link count in the first transaction of the remove
1952 * operation, there are no transactional constraints on the ordering here.
1953 */
1954 int
1955 xfs_remove(
1956 struct xfs_inode *dp,
1957 struct xfs_name *name,
1958 struct xfs_inode *ip)
1959 {
1960 struct xfs_dir_update du = {
1961 .dp = dp,
1962 .name = name,
1963 .ip = ip,
1964 };
1965 struct xfs_mount *mp = dp->i_mount;
1966 struct xfs_trans *tp = NULL;
1967 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1968 int dontcare;
1969 int error = 0;
1970 uint resblks;
1971
1972 trace_xfs_remove(dp, name);
1973
1974 if (xfs_is_shutdown(mp))
1975 return -EIO;
1976 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1977 return -EIO;
1978
1979 error = xfs_qm_dqattach(dp);
1980 if (error)
1981 goto std_return;
1982
1983 error = xfs_qm_dqattach(ip);
1984 if (error)
1985 goto std_return;
1986
1987 error = xfs_parent_start(mp, &du.ppargs);
1988 if (error)
1989 goto std_return;
1990
1991 /*
1992 * We try to get the real space reservation first, allowing for
1993 * directory btree deletion(s) implying possible bmap insert(s). If we
1994 * can't get the space reservation then we use 0 instead, and avoid the
1995 * bmap btree insert(s) in the directory code by, if the bmap insert
1996 * tries to happen, instead trimming the LAST block from the directory.
1997 *
1998 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
1999 * the directory code can handle a reservationless update and we don't
2000 * want to prevent a user from trying to free space by deleting things.
2001 */
2002 resblks = xfs_remove_space_res(mp, name->len);
2003 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2004 &tp, &dontcare);
2005 if (error) {
2006 ASSERT(error != -ENOSPC);
2007 goto out_parent;
2008 }
2009
2010 error = xfs_dir_remove_child(tp, resblks, &du);
2011 if (error)
2012 goto out_trans_cancel;
2013
2014 /*
2015 * If this is a synchronous mount, make sure that the
2016 * remove transaction goes to disk before returning to
2017 * the user.
2018 */
2019 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2020 xfs_trans_set_sync(tp);
2021
2022 error = xfs_trans_commit(tp);
2023 if (error)
2024 goto out_unlock;
2025
2026 if (is_dir && xfs_inode_is_filestream(ip))
2027 xfs_filestream_deassociate(ip);
2028
2029 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2030 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2031 xfs_parent_finish(mp, du.ppargs);
2032 return 0;
2033
2034 out_trans_cancel:
2035 xfs_trans_cancel(tp);
2036 out_unlock:
2037 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2038 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2039 out_parent:
2040 xfs_parent_finish(mp, du.ppargs);
2041 std_return:
2042 return error;
2043 }
2044
2045 static inline void
2046 xfs_iunlock_rename(
2047 struct xfs_inode **i_tab,
2048 int num_inodes)
2049 {
2050 int i;
2051
2052 for (i = num_inodes - 1; i >= 0; i--) {
2053 /* Skip duplicate inodes if src and target dps are the same */
2054 if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
2055 continue;
2056 xfs_iunlock(i_tab[i], XFS_ILOCK_EXCL);
2057 }
2058 }
2059
2060 /*
2061 * Enter all inodes for a rename transaction into a sorted array.
2062 */
2063 #define __XFS_SORT_INODES 5
2064 STATIC void
2065 xfs_sort_for_rename(
2066 struct xfs_inode *dp1, /* in: old (source) directory inode */
2067 struct xfs_inode *dp2, /* in: new (target) directory inode */
2068 struct xfs_inode *ip1, /* in: inode of old entry */
2069 struct xfs_inode *ip2, /* in: inode of new entry */
2070 struct xfs_inode *wip, /* in: whiteout inode */
2071 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2072 int *num_inodes) /* in/out: inodes in array */
2073 {
2074 int i;
2075
2076 ASSERT(*num_inodes == __XFS_SORT_INODES);
2077 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2078
2079 /*
2080 * i_tab contains a list of pointers to inodes. We initialize
2081 * the table here & we'll sort it. We will then use it to
2082 * order the acquisition of the inode locks.
2083 *
2084 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2085 */
2086 i = 0;
2087 i_tab[i++] = dp1;
2088 i_tab[i++] = dp2;
2089 i_tab[i++] = ip1;
2090 if (ip2)
2091 i_tab[i++] = ip2;
2092 if (wip)
2093 i_tab[i++] = wip;
2094 *num_inodes = i;
2095
2096 xfs_sort_inodes(i_tab, *num_inodes);
2097 }
2098
2099 void
2100 xfs_sort_inodes(
2101 struct xfs_inode **i_tab,
2102 unsigned int num_inodes)
2103 {
2104 int i, j;
2105
2106 ASSERT(num_inodes <= __XFS_SORT_INODES);
2107
2108 /*
2109 * Sort the elements via bubble sort. (Remember, there are at
2110 * most 5 elements to sort, so this is adequate.)
2111 */
2112 for (i = 0; i < num_inodes; i++) {
2113 for (j = 1; j < num_inodes; j++) {
2114 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2115 swap(i_tab[j], i_tab[j - 1]);
2116 }
2117 }
2118 }
2119
2120 /*
2121 * xfs_rename_alloc_whiteout()
2122 *
2123 * Return a referenced, unlinked, unlocked inode that can be used as a
2124 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2125 * crash between allocating the inode and linking it into the rename transaction
2126 * recovery will free the inode and we won't leak it.
2127 */
2128 static int
2129 xfs_rename_alloc_whiteout(
2130 struct mnt_idmap *idmap,
2131 struct xfs_name *src_name,
2132 struct xfs_inode *dp,
2133 struct xfs_inode **wip)
2134 {
2135 struct xfs_icreate_args args = {
2136 .idmap = idmap,
2137 .pip = dp,
2138 .mode = S_IFCHR | WHITEOUT_MODE,
2139 .flags = XFS_ICREATE_TMPFILE,
2140 };
2141 struct xfs_inode *tmpfile;
2142 struct qstr name;
2143 int error;
2144
2145 error = xfs_create_tmpfile(&args, &tmpfile);
2146 if (error)
2147 return error;
2148
2149 name.name = src_name->name;
2150 name.len = src_name->len;
2151 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2152 if (error) {
2153 xfs_finish_inode_setup(tmpfile);
2154 xfs_irele(tmpfile);
2155 return error;
2156 }
2157
2158 /*
2159 * Prepare the tmpfile inode as if it were created through the VFS.
2160 * Complete the inode setup and flag it as linkable. nlink is already
2161 * zero, so we can skip the drop_nlink.
2162 */
2163 xfs_setup_iops(tmpfile);
2164 xfs_finish_inode_setup(tmpfile);
2165 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2166
2167 *wip = tmpfile;
2168 return 0;
2169 }
2170
2171 /*
2172 * xfs_rename
2173 */
2174 int
2175 xfs_rename(
2176 struct mnt_idmap *idmap,
2177 struct xfs_inode *src_dp,
2178 struct xfs_name *src_name,
2179 struct xfs_inode *src_ip,
2180 struct xfs_inode *target_dp,
2181 struct xfs_name *target_name,
2182 struct xfs_inode *target_ip,
2183 unsigned int flags)
2184 {
2185 struct xfs_dir_update du_src = {
2186 .dp = src_dp,
2187 .name = src_name,
2188 .ip = src_ip,
2189 };
2190 struct xfs_dir_update du_tgt = {
2191 .dp = target_dp,
2192 .name = target_name,
2193 .ip = target_ip,
2194 };
2195 struct xfs_dir_update du_wip = { };
2196 struct xfs_mount *mp = src_dp->i_mount;
2197 struct xfs_trans *tp;
2198 struct xfs_inode *inodes[__XFS_SORT_INODES];
2199 int i;
2200 int num_inodes = __XFS_SORT_INODES;
2201 bool new_parent = (src_dp != target_dp);
2202 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2203 int spaceres;
2204 bool retried = false;
2205 int error, nospace_error = 0;
2206
2207 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2208
2209 if ((flags & RENAME_EXCHANGE) && !target_ip)
2210 return -EINVAL;
2211
2212 /*
2213 * If we are doing a whiteout operation, allocate the whiteout inode
2214 * we will be placing at the target and ensure the type is set
2215 * appropriately.
2216 */
2217 if (flags & RENAME_WHITEOUT) {
2218 error = xfs_rename_alloc_whiteout(idmap, src_name, target_dp,
2219 &du_wip.ip);
2220 if (error)
2221 return error;
2222
2223 /* setup target dirent info as whiteout */
2224 src_name->type = XFS_DIR3_FT_CHRDEV;
2225 }
2226
2227 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, du_wip.ip,
2228 inodes, &num_inodes);
2229
2230 error = xfs_parent_start(mp, &du_src.ppargs);
2231 if (error)
2232 goto out_release_wip;
2233
2234 if (du_wip.ip) {
2235 error = xfs_parent_start(mp, &du_wip.ppargs);
2236 if (error)
2237 goto out_src_ppargs;
2238 }
2239
2240 if (target_ip) {
2241 error = xfs_parent_start(mp, &du_tgt.ppargs);
2242 if (error)
2243 goto out_wip_ppargs;
2244 }
2245
2246 retry:
2247 nospace_error = 0;
2248 spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2249 target_name->len, du_wip.ip != NULL);
2250 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2251 if (error == -ENOSPC) {
2252 nospace_error = error;
2253 spaceres = 0;
2254 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2255 &tp);
2256 }
2257 if (error)
2258 goto out_tgt_ppargs;
2259
2260 /*
2261 * We don't allow reservationless renaming when parent pointers are
2262 * enabled because we can't back out if the xattrs must grow.
2263 */
2264 if (du_src.ppargs && nospace_error) {
2265 error = nospace_error;
2266 xfs_trans_cancel(tp);
2267 goto out_tgt_ppargs;
2268 }
2269
2270 /*
2271 * Attach the dquots to the inodes
2272 */
2273 error = xfs_qm_vop_rename_dqattach(inodes);
2274 if (error) {
2275 xfs_trans_cancel(tp);
2276 goto out_tgt_ppargs;
2277 }
2278
2279 /*
2280 * Lock all the participating inodes. Depending upon whether
2281 * the target_name exists in the target directory, and
2282 * whether the target directory is the same as the source
2283 * directory, we can lock from 2 to 5 inodes.
2284 */
2285 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2286
2287 /*
2288 * Join all the inodes to the transaction.
2289 */
2290 xfs_trans_ijoin(tp, src_dp, 0);
2291 if (new_parent)
2292 xfs_trans_ijoin(tp, target_dp, 0);
2293 xfs_trans_ijoin(tp, src_ip, 0);
2294 if (target_ip)
2295 xfs_trans_ijoin(tp, target_ip, 0);
2296 if (du_wip.ip)
2297 xfs_trans_ijoin(tp, du_wip.ip, 0);
2298
2299 /*
2300 * If we are using project inheritance, we only allow renames
2301 * into our tree when the project IDs are the same; else the
2302 * tree quota mechanism would be circumvented.
2303 */
2304 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2305 target_dp->i_projid != src_ip->i_projid)) {
2306 error = -EXDEV;
2307 goto out_trans_cancel;
2308 }
2309
2310 /* RENAME_EXCHANGE is unique from here on. */
2311 if (flags & RENAME_EXCHANGE) {
2312 error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2313 spaceres);
2314 if (error)
2315 goto out_trans_cancel;
2316 goto out_commit;
2317 }
2318
2319 /*
2320 * Try to reserve quota to handle an expansion of the target directory.
2321 * We'll allow the rename to continue in reservationless mode if we hit
2322 * a space usage constraint. If we trigger reservationless mode, save
2323 * the errno if there isn't any free space in the target directory.
2324 */
2325 if (spaceres != 0) {
2326 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2327 0, false);
2328 if (error == -EDQUOT || error == -ENOSPC) {
2329 if (!retried) {
2330 xfs_trans_cancel(tp);
2331 xfs_iunlock_rename(inodes, num_inodes);
2332 xfs_blockgc_free_quota(target_dp, 0);
2333 retried = true;
2334 goto retry;
2335 }
2336
2337 nospace_error = error;
2338 spaceres = 0;
2339 error = 0;
2340 }
2341 if (error)
2342 goto out_trans_cancel;
2343 }
2344
2345 /*
2346 * We don't allow quotaless renaming when parent pointers are enabled
2347 * because we can't back out if the xattrs must grow.
2348 */
2349 if (du_src.ppargs && nospace_error) {
2350 error = nospace_error;
2351 goto out_trans_cancel;
2352 }
2353
2354 /*
2355 * Lock the AGI buffers we need to handle bumping the nlink of the
2356 * whiteout inode off the unlinked list and to handle dropping the
2357 * nlink of the target inode. Per locking order rules, do this in
2358 * increasing AG order and before directory block allocation tries to
2359 * grab AGFs because we grab AGIs before AGFs.
2360 *
2361 * The (vfs) caller must ensure that if src is a directory then
2362 * target_ip is either null or an empty directory.
2363 */
2364 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2365 if (inodes[i] == du_wip.ip ||
2366 (inodes[i] == target_ip &&
2367 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2368 struct xfs_perag *pag;
2369 struct xfs_buf *bp;
2370
2371 pag = xfs_perag_get(mp,
2372 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2373 error = xfs_read_agi(pag, tp, 0, &bp);
2374 xfs_perag_put(pag);
2375 if (error)
2376 goto out_trans_cancel;
2377 }
2378 }
2379
2380 error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2381 &du_wip);
2382 if (error)
2383 goto out_trans_cancel;
2384
2385 if (du_wip.ip) {
2386 /*
2387 * Now we have a real link, clear the "I'm a tmpfile" state
2388 * flag from the inode so it doesn't accidentally get misused in
2389 * future.
2390 */
2391 VFS_I(du_wip.ip)->i_state &= ~I_LINKABLE;
2392 }
2393
2394 out_commit:
2395 /*
2396 * If this is a synchronous mount, make sure that the rename
2397 * transaction goes to disk before returning to the user.
2398 */
2399 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2400 xfs_trans_set_sync(tp);
2401
2402 error = xfs_trans_commit(tp);
2403 nospace_error = 0;
2404 goto out_unlock;
2405
2406 out_trans_cancel:
2407 xfs_trans_cancel(tp);
2408 out_unlock:
2409 xfs_iunlock_rename(inodes, num_inodes);
2410 out_tgt_ppargs:
2411 xfs_parent_finish(mp, du_tgt.ppargs);
2412 out_wip_ppargs:
2413 xfs_parent_finish(mp, du_wip.ppargs);
2414 out_src_ppargs:
2415 xfs_parent_finish(mp, du_src.ppargs);
2416 out_release_wip:
2417 if (du_wip.ip)
2418 xfs_irele(du_wip.ip);
2419 if (error == -ENOSPC && nospace_error)
2420 error = nospace_error;
2421 return error;
2422 }
2423
2424 static int
2425 xfs_iflush(
2426 struct xfs_inode *ip,
2427 struct xfs_buf *bp)
2428 {
2429 struct xfs_inode_log_item *iip = ip->i_itemp;
2430 struct xfs_dinode *dip;
2431 struct xfs_mount *mp = ip->i_mount;
2432 int error;
2433
2434 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2435 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2436 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2437 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2438 ASSERT(iip->ili_item.li_buf == bp);
2439
2440 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
2441
2442 /*
2443 * We don't flush the inode if any of the following checks fail, but we
2444 * do still update the log item and attach to the backing buffer as if
2445 * the flush happened. This is a formality to facilitate predictable
2446 * error handling as the caller will shutdown and fail the buffer.
2447 */
2448 error = -EFSCORRUPTED;
2449 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
2450 mp, XFS_ERRTAG_IFLUSH_1)) {
2451 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2452 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2453 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2454 goto flush_out;
2455 }
2456 if (S_ISREG(VFS_I(ip)->i_mode)) {
2457 if (XFS_TEST_ERROR(
2458 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2459 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
2460 mp, XFS_ERRTAG_IFLUSH_3)) {
2461 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2462 "%s: Bad regular inode %llu, ptr "PTR_FMT,
2463 __func__, ip->i_ino, ip);
2464 goto flush_out;
2465 }
2466 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2467 if (XFS_TEST_ERROR(
2468 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2469 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2470 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
2471 mp, XFS_ERRTAG_IFLUSH_4)) {
2472 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2473 "%s: Bad directory inode %llu, ptr "PTR_FMT,
2474 __func__, ip->i_ino, ip);
2475 goto flush_out;
2476 }
2477 }
2478 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2479 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
2480 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2481 "%s: detected corrupt incore inode %llu, "
2482 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2483 __func__, ip->i_ino,
2484 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2485 ip->i_nblocks, ip);
2486 goto flush_out;
2487 }
2488 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
2489 mp, XFS_ERRTAG_IFLUSH_6)) {
2490 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2491 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2492 __func__, ip->i_ino, ip->i_forkoff, ip);
2493 goto flush_out;
2494 }
2495
2496 /*
2497 * Inode item log recovery for v2 inodes are dependent on the flushiter
2498 * count for correct sequencing. We bump the flush iteration count so
2499 * we can detect flushes which postdate a log record during recovery.
2500 * This is redundant as we now log every change and hence this can't
2501 * happen but we need to still do it to ensure backwards compatibility
2502 * with old kernels that predate logging all inode changes.
2503 */
2504 if (!xfs_has_v3inodes(mp))
2505 ip->i_flushiter++;
2506
2507 /*
2508 * If there are inline format data / attr forks attached to this inode,
2509 * make sure they are not corrupt.
2510 */
2511 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2512 xfs_ifork_verify_local_data(ip))
2513 goto flush_out;
2514 if (xfs_inode_has_attr_fork(ip) &&
2515 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2516 xfs_ifork_verify_local_attr(ip))
2517 goto flush_out;
2518
2519 /*
2520 * Copy the dirty parts of the inode into the on-disk inode. We always
2521 * copy out the core of the inode, because if the inode is dirty at all
2522 * the core must be.
2523 */
2524 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2525
2526 /* Wrap, we never let the log put out DI_MAX_FLUSH */
2527 if (!xfs_has_v3inodes(mp)) {
2528 if (ip->i_flushiter == DI_MAX_FLUSH)
2529 ip->i_flushiter = 0;
2530 }
2531
2532 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2533 if (xfs_inode_has_attr_fork(ip))
2534 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2535
2536 /*
2537 * We've recorded everything logged in the inode, so we'd like to clear
2538 * the ili_fields bits so we don't log and flush things unnecessarily.
2539 * However, we can't stop logging all this information until the data
2540 * we've copied into the disk buffer is written to disk. If we did we
2541 * might overwrite the copy of the inode in the log with all the data
2542 * after re-logging only part of it, and in the face of a crash we
2543 * wouldn't have all the data we need to recover.
2544 *
2545 * What we do is move the bits to the ili_last_fields field. When
2546 * logging the inode, these bits are moved back to the ili_fields field.
2547 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2548 * we know that the information those bits represent is permanently on
2549 * disk. As long as the flush completes before the inode is logged
2550 * again, then both ili_fields and ili_last_fields will be cleared.
2551 */
2552 error = 0;
2553 flush_out:
2554 spin_lock(&iip->ili_lock);
2555 iip->ili_last_fields = iip->ili_fields;
2556 iip->ili_fields = 0;
2557 iip->ili_fsync_fields = 0;
2558 set_bit(XFS_LI_FLUSHING, &iip->ili_item.li_flags);
2559 spin_unlock(&iip->ili_lock);
2560
2561 /*
2562 * Store the current LSN of the inode so that we can tell whether the
2563 * item has moved in the AIL from xfs_buf_inode_iodone().
2564 */
2565 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2566 &iip->ili_item.li_lsn);
2567
2568 /* generate the checksum. */
2569 xfs_dinode_calc_crc(mp, dip);
2570 if (error)
2571 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2572 return error;
2573 }
2574
2575 /*
2576 * Non-blocking flush of dirty inode metadata into the backing buffer.
2577 *
2578 * The caller must have a reference to the inode and hold the cluster buffer
2579 * locked. The function will walk across all the inodes on the cluster buffer it
2580 * can find and lock without blocking, and flush them to the cluster buffer.
2581 *
2582 * On successful flushing of at least one inode, the caller must write out the
2583 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2584 * the caller needs to release the buffer. On failure, the filesystem will be
2585 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2586 * will be returned.
2587 */
2588 int
2589 xfs_iflush_cluster(
2590 struct xfs_buf *bp)
2591 {
2592 struct xfs_mount *mp = bp->b_mount;
2593 struct xfs_log_item *lip, *n;
2594 struct xfs_inode *ip;
2595 struct xfs_inode_log_item *iip;
2596 int clcount = 0;
2597 int error = 0;
2598
2599 /*
2600 * We must use the safe variant here as on shutdown xfs_iflush_abort()
2601 * will remove itself from the list.
2602 */
2603 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2604 iip = (struct xfs_inode_log_item *)lip;
2605 ip = iip->ili_inode;
2606
2607 /*
2608 * Quick and dirty check to avoid locks if possible.
2609 */
2610 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2611 continue;
2612 if (xfs_ipincount(ip))
2613 continue;
2614
2615 /*
2616 * The inode is still attached to the buffer, which means it is
2617 * dirty but reclaim might try to grab it. Check carefully for
2618 * that, and grab the ilock while still holding the i_flags_lock
2619 * to guarantee reclaim will not be able to reclaim this inode
2620 * once we drop the i_flags_lock.
2621 */
2622 spin_lock(&ip->i_flags_lock);
2623 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2624 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2625 spin_unlock(&ip->i_flags_lock);
2626 continue;
2627 }
2628
2629 /*
2630 * ILOCK will pin the inode against reclaim and prevent
2631 * concurrent transactions modifying the inode while we are
2632 * flushing the inode. If we get the lock, set the flushing
2633 * state before we drop the i_flags_lock.
2634 */
2635 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2636 spin_unlock(&ip->i_flags_lock);
2637 continue;
2638 }
2639 __xfs_iflags_set(ip, XFS_IFLUSHING);
2640 spin_unlock(&ip->i_flags_lock);
2641
2642 /*
2643 * Abort flushing this inode if we are shut down because the
2644 * inode may not currently be in the AIL. This can occur when
2645 * log I/O failure unpins the inode without inserting into the
2646 * AIL, leaving a dirty/unpinned inode attached to the buffer
2647 * that otherwise looks like it should be flushed.
2648 */
2649 if (xlog_is_shutdown(mp->m_log)) {
2650 xfs_iunpin_wait(ip);
2651 xfs_iflush_abort(ip);
2652 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2653 error = -EIO;
2654 continue;
2655 }
2656
2657 /* don't block waiting on a log force to unpin dirty inodes */
2658 if (xfs_ipincount(ip)) {
2659 xfs_iflags_clear(ip, XFS_IFLUSHING);
2660 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2661 continue;
2662 }
2663
2664 if (!xfs_inode_clean(ip))
2665 error = xfs_iflush(ip, bp);
2666 else
2667 xfs_iflags_clear(ip, XFS_IFLUSHING);
2668 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2669 if (error)
2670 break;
2671 clcount++;
2672 }
2673
2674 if (error) {
2675 /*
2676 * Shutdown first so we kill the log before we release this
2677 * buffer. If it is an INODE_ALLOC buffer and pins the tail
2678 * of the log, failing it before the _log_ is shut down can
2679 * result in the log tail being moved forward in the journal
2680 * on disk because log writes can still be taking place. Hence
2681 * unpinning the tail will allow the ICREATE intent to be
2682 * removed from the log an recovery will fail with uninitialised
2683 * inode cluster buffers.
2684 */
2685 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2686 bp->b_flags |= XBF_ASYNC;
2687 xfs_buf_ioend_fail(bp);
2688 return error;
2689 }
2690
2691 if (!clcount)
2692 return -EAGAIN;
2693
2694 XFS_STATS_INC(mp, xs_icluster_flushcnt);
2695 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2696 return 0;
2697
2698 }
2699
2700 /* Release an inode. */
2701 void
2702 xfs_irele(
2703 struct xfs_inode *ip)
2704 {
2705 trace_xfs_irele(ip, _RET_IP_);
2706 iput(VFS_I(ip));
2707 }
2708
2709 /*
2710 * Ensure all commited transactions touching the inode are written to the log.
2711 */
2712 int
2713 xfs_log_force_inode(
2714 struct xfs_inode *ip)
2715 {
2716 xfs_csn_t seq = 0;
2717
2718 xfs_ilock(ip, XFS_ILOCK_SHARED);
2719 if (xfs_ipincount(ip))
2720 seq = ip->i_itemp->ili_commit_seq;
2721 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2722
2723 if (!seq)
2724 return 0;
2725 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2726 }
2727
2728 /*
2729 * Grab the exclusive iolock for a data copy from src to dest, making sure to
2730 * abide vfs locking order (lowest pointer value goes first) and breaking the
2731 * layout leases before proceeding. The loop is needed because we cannot call
2732 * the blocking break_layout() with the iolocks held, and therefore have to
2733 * back out both locks.
2734 */
2735 static int
2736 xfs_iolock_two_inodes_and_break_layout(
2737 struct inode *src,
2738 struct inode *dest)
2739 {
2740 int error;
2741
2742 if (src > dest)
2743 swap(src, dest);
2744
2745 retry:
2746 /* Wait to break both inodes' layouts before we start locking. */
2747 error = break_layout(src, true);
2748 if (error)
2749 return error;
2750 if (src != dest) {
2751 error = break_layout(dest, true);
2752 if (error)
2753 return error;
2754 }
2755
2756 /* Lock one inode and make sure nobody got in and leased it. */
2757 inode_lock(src);
2758 error = break_layout(src, false);
2759 if (error) {
2760 inode_unlock(src);
2761 if (error == -EWOULDBLOCK)
2762 goto retry;
2763 return error;
2764 }
2765
2766 if (src == dest)
2767 return 0;
2768
2769 /* Lock the other inode and make sure nobody got in and leased it. */
2770 inode_lock_nested(dest, I_MUTEX_NONDIR2);
2771 error = break_layout(dest, false);
2772 if (error) {
2773 inode_unlock(src);
2774 inode_unlock(dest);
2775 if (error == -EWOULDBLOCK)
2776 goto retry;
2777 return error;
2778 }
2779
2780 return 0;
2781 }
2782
2783 static int
2784 xfs_mmaplock_two_inodes_and_break_dax_layout(
2785 struct xfs_inode *ip1,
2786 struct xfs_inode *ip2)
2787 {
2788 int error;
2789 bool retry;
2790 struct page *page;
2791
2792 if (ip1->i_ino > ip2->i_ino)
2793 swap(ip1, ip2);
2794
2795 again:
2796 retry = false;
2797 /* Lock the first inode */
2798 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
2799 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
2800 if (error || retry) {
2801 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2802 if (error == 0 && retry)
2803 goto again;
2804 return error;
2805 }
2806
2807 if (ip1 == ip2)
2808 return 0;
2809
2810 /* Nested lock the second inode */
2811 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
2812 /*
2813 * We cannot use xfs_break_dax_layouts() directly here because it may
2814 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2815 * for this nested lock case.
2816 */
2817 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
2818 if (page && page_ref_count(page) != 1) {
2819 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2820 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2821 goto again;
2822 }
2823
2824 return 0;
2825 }
2826
2827 /*
2828 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2829 * mmap activity.
2830 */
2831 int
2832 xfs_ilock2_io_mmap(
2833 struct xfs_inode *ip1,
2834 struct xfs_inode *ip2)
2835 {
2836 int ret;
2837
2838 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
2839 if (ret)
2840 return ret;
2841
2842 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2843 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2844 if (ret) {
2845 inode_unlock(VFS_I(ip2));
2846 if (ip1 != ip2)
2847 inode_unlock(VFS_I(ip1));
2848 return ret;
2849 }
2850 } else
2851 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
2852 VFS_I(ip2)->i_mapping);
2853
2854 return 0;
2855 }
2856
2857 /* Unlock both inodes to allow IO and mmap activity. */
2858 void
2859 xfs_iunlock2_io_mmap(
2860 struct xfs_inode *ip1,
2861 struct xfs_inode *ip2)
2862 {
2863 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2864 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2865 if (ip1 != ip2)
2866 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2867 } else
2868 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
2869 VFS_I(ip2)->i_mapping);
2870
2871 inode_unlock(VFS_I(ip2));
2872 if (ip1 != ip2)
2873 inode_unlock(VFS_I(ip1));
2874 }
2875
2876 /* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2877 void
2878 xfs_iunlock2_remapping(
2879 struct xfs_inode *ip1,
2880 struct xfs_inode *ip2)
2881 {
2882 xfs_iflags_clear(ip1, XFS_IREMAPPING);
2883
2884 if (ip1 != ip2)
2885 xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
2886 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2887
2888 if (ip1 != ip2)
2889 inode_unlock_shared(VFS_I(ip1));
2890 inode_unlock(VFS_I(ip2));
2891 }
2892
2893 /*
2894 * Reload the incore inode list for this inode. Caller should ensure that
2895 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
2896 * preventing other threads from executing.
2897 */
2898 int
2899 xfs_inode_reload_unlinked_bucket(
2900 struct xfs_trans *tp,
2901 struct xfs_inode *ip)
2902 {
2903 struct xfs_mount *mp = tp->t_mountp;
2904 struct xfs_buf *agibp;
2905 struct xfs_agi *agi;
2906 struct xfs_perag *pag;
2907 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2908 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2909 xfs_agino_t prev_agino, next_agino;
2910 unsigned int bucket;
2911 bool foundit = false;
2912 int error;
2913
2914 /* Grab the first inode in the list */
2915 pag = xfs_perag_get(mp, agno);
2916 error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2917 xfs_perag_put(pag);
2918 if (error)
2919 return error;
2920
2921 /*
2922 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2923 * incore unlinked list pointers for this inode. Check once more to
2924 * see if we raced with anyone else to reload the unlinked list.
2925 */
2926 if (!xfs_inode_unlinked_incomplete(ip)) {
2927 foundit = true;
2928 goto out_agibp;
2929 }
2930
2931 bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2932 agi = agibp->b_addr;
2933
2934 trace_xfs_inode_reload_unlinked_bucket(ip);
2935
2936 xfs_info_ratelimited(mp,
2937 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
2938 agino, agno);
2939
2940 prev_agino = NULLAGINO;
2941 next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2942 while (next_agino != NULLAGINO) {
2943 struct xfs_inode *next_ip = NULL;
2944
2945 /* Found this caller's inode, set its backlink. */
2946 if (next_agino == agino) {
2947 next_ip = ip;
2948 next_ip->i_prev_unlinked = prev_agino;
2949 foundit = true;
2950 goto next_inode;
2951 }
2952
2953 /* Try in-memory lookup first. */
2954 next_ip = xfs_iunlink_lookup(pag, next_agino);
2955 if (next_ip)
2956 goto next_inode;
2957
2958 /* Inode not in memory, try reloading it. */
2959 error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2960 next_agino);
2961 if (error)
2962 break;
2963
2964 /* Grab the reloaded inode. */
2965 next_ip = xfs_iunlink_lookup(pag, next_agino);
2966 if (!next_ip) {
2967 /* No incore inode at all? We reloaded it... */
2968 ASSERT(next_ip != NULL);
2969 error = -EFSCORRUPTED;
2970 break;
2971 }
2972
2973 next_inode:
2974 prev_agino = next_agino;
2975 next_agino = next_ip->i_next_unlinked;
2976 }
2977
2978 out_agibp:
2979 xfs_trans_brelse(tp, agibp);
2980 /* Should have found this inode somewhere in the iunlinked bucket. */
2981 if (!error && !foundit)
2982 error = -EFSCORRUPTED;
2983 return error;
2984 }
2985
2986 /* Decide if this inode is missing its unlinked list and reload it. */
2987 int
2988 xfs_inode_reload_unlinked(
2989 struct xfs_inode *ip)
2990 {
2991 struct xfs_trans *tp;
2992 int error;
2993
2994 error = xfs_trans_alloc_empty(ip->i_mount, &tp);
2995 if (error)
2996 return error;
2997
2998 xfs_ilock(ip, XFS_ILOCK_SHARED);
2999 if (xfs_inode_unlinked_incomplete(ip))
3000 error = xfs_inode_reload_unlinked_bucket(tp, ip);
3001 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3002 xfs_trans_cancel(tp);
3003
3004 return error;
3005 }
3006
3007 /* Has this inode fork been zapped by repair? */
3008 bool
3009 xfs_ifork_zapped(
3010 const struct xfs_inode *ip,
3011 int whichfork)
3012 {
3013 unsigned int datamask = 0;
3014
3015 switch (whichfork) {
3016 case XFS_DATA_FORK:
3017 switch (ip->i_vnode.i_mode & S_IFMT) {
3018 case S_IFDIR:
3019 datamask = XFS_SICK_INO_DIR_ZAPPED;
3020 break;
3021 case S_IFLNK:
3022 datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
3023 break;
3024 }
3025 return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
3026 case XFS_ATTR_FORK:
3027 return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
3028 default:
3029 return false;
3030 }
3031 }
3032
3033 /* Compute the number of data and realtime blocks used by a file. */
3034 void
3035 xfs_inode_count_blocks(
3036 struct xfs_trans *tp,
3037 struct xfs_inode *ip,
3038 xfs_filblks_t *dblocks,
3039 xfs_filblks_t *rblocks)
3040 {
3041 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
3042
3043 *rblocks = 0;
3044 if (XFS_IS_REALTIME_INODE(ip))
3045 xfs_bmap_count_leaves(ifp, rblocks);
3046 *dblocks = ip->i_nblocks - *rblocks;
3047 }
3048
3049 static void
3050 xfs_wait_dax_page(
3051 struct inode *inode)
3052 {
3053 struct xfs_inode *ip = XFS_I(inode);
3054
3055 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
3056 schedule();
3057 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
3058 }
3059
3060 int
3061 xfs_break_dax_layouts(
3062 struct inode *inode,
3063 bool *retry)
3064 {
3065 struct page *page;
3066
3067 xfs_assert_ilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL);
3068
3069 page = dax_layout_busy_page(inode->i_mapping);
3070 if (!page)
3071 return 0;
3072
3073 *retry = true;
3074 return ___wait_var_event(&page->_refcount,
3075 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
3076 0, 0, xfs_wait_dax_page(inode));
3077 }
3078
3079 int
3080 xfs_break_layouts(
3081 struct inode *inode,
3082 uint *iolock,
3083 enum layout_break_reason reason)
3084 {
3085 bool retry;
3086 int error;
3087
3088 xfs_assert_ilocked(XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3089
3090 do {
3091 retry = false;
3092 switch (reason) {
3093 case BREAK_UNMAP:
3094 error = xfs_break_dax_layouts(inode, &retry);
3095 if (error || retry)
3096 break;
3097 fallthrough;
3098 case BREAK_WRITE:
3099 error = xfs_break_leased_layouts(inode, iolock, &retry);
3100 break;
3101 default:
3102 WARN_ON_ONCE(1);
3103 error = -EINVAL;
3104 }
3105 } while (error == 0 && retry);
3106
3107 return error;
3108 }
3109
3110 /* Returns the size of fundamental allocation unit for a file, in bytes. */
3111 unsigned int
3112 xfs_inode_alloc_unitsize(
3113 struct xfs_inode *ip)
3114 {
3115 unsigned int blocks = 1;
3116
3117 if (XFS_IS_REALTIME_INODE(ip))
3118 blocks = ip->i_mount->m_sb.sb_rextsize;
3119
3120 return XFS_FSB_TO_B(ip->i_mount, blocks);
3121 }
3122
3123 /* Should we always be using copy on write for file writes? */
3124 bool
3125 xfs_is_always_cow_inode(
3126 struct xfs_inode *ip)
3127 {
3128 return ip->i_mount->m_always_cow && xfs_has_reflink(ip->i_mount);
3129 }
3130
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.