Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include <linux/iversion.h>
7
8#include "xfs.h"
9#include "xfs_fs.h"
10#include "xfs_shared.h"
11#include "xfs_format.h"
12#include "xfs_log_format.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_defer.h"
16#include "xfs_inode.h"
17#include "xfs_dir2.h"
18#include "xfs_attr.h"
19#include "xfs_trans_space.h"
20#include "xfs_trans.h"
21#include "xfs_buf_item.h"
22#include "xfs_inode_item.h"
23#include "xfs_iunlink_item.h"
24#include "xfs_ialloc.h"
25#include "xfs_bmap.h"
26#include "xfs_bmap_util.h"
27#include "xfs_errortag.h"
28#include "xfs_error.h"
29#include "xfs_quota.h"
30#include "xfs_filestream.h"
31#include "xfs_trace.h"
32#include "xfs_icache.h"
33#include "xfs_symlink.h"
34#include "xfs_trans_priv.h"
35#include "xfs_log.h"
36#include "xfs_bmap_btree.h"
37#include "xfs_reflink.h"
38#include "xfs_ag.h"
39#include "xfs_log_priv.h"
40#include "xfs_health.h"
41
42struct kmem_cache *xfs_inode_cache;
43
44STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
45STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
46 struct xfs_inode *);
47
48/*
49 * helper function to extract extent size hint from inode
50 */
51xfs_extlen_t
52xfs_get_extsz_hint(
53 struct xfs_inode *ip)
54{
55 /*
56 * No point in aligning allocations if we need to COW to actually
57 * write to them.
58 */
59 if (xfs_is_always_cow_inode(ip))
60 return 0;
61 if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
62 return ip->i_extsize;
63 if (XFS_IS_REALTIME_INODE(ip))
64 return ip->i_mount->m_sb.sb_rextsize;
65 return 0;
66}
67
68/*
69 * Helper function to extract CoW extent size hint from inode.
70 * Between the extent size hint and the CoW extent size hint, we
71 * return the greater of the two. If the value is zero (automatic),
72 * use the default size.
73 */
74xfs_extlen_t
75xfs_get_cowextsz_hint(
76 struct xfs_inode *ip)
77{
78 xfs_extlen_t a, b;
79
80 a = 0;
81 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
82 a = ip->i_cowextsize;
83 b = xfs_get_extsz_hint(ip);
84
85 a = max(a, b);
86 if (a == 0)
87 return XFS_DEFAULT_COWEXTSZ_HINT;
88 return a;
89}
90
91/*
92 * These two are wrapper routines around the xfs_ilock() routine used to
93 * centralize some grungy code. They are used in places that wish to lock the
94 * inode solely for reading the extents. The reason these places can't just
95 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
96 * bringing in of the extents from disk for a file in b-tree format. If the
97 * inode is in b-tree format, then we need to lock the inode exclusively until
98 * the extents are read in. Locking it exclusively all the time would limit
99 * our parallelism unnecessarily, though. What we do instead is check to see
100 * if the extents have been read in yet, and only lock the inode exclusively
101 * if they have not.
102 *
103 * The functions return a value which should be given to the corresponding
104 * xfs_iunlock() call.
105 */
106uint
107xfs_ilock_data_map_shared(
108 struct xfs_inode *ip)
109{
110 uint lock_mode = XFS_ILOCK_SHARED;
111
112 if (xfs_need_iread_extents(&ip->i_df))
113 lock_mode = XFS_ILOCK_EXCL;
114 xfs_ilock(ip, lock_mode);
115 return lock_mode;
116}
117
118uint
119xfs_ilock_attr_map_shared(
120 struct xfs_inode *ip)
121{
122 uint lock_mode = XFS_ILOCK_SHARED;
123
124 if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
125 lock_mode = XFS_ILOCK_EXCL;
126 xfs_ilock(ip, lock_mode);
127 return lock_mode;
128}
129
130/*
131 * You can't set both SHARED and EXCL for the same lock,
132 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
133 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
134 * to set in lock_flags.
135 */
136static inline void
137xfs_lock_flags_assert(
138 uint lock_flags)
139{
140 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
141 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
142 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
143 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
144 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
145 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
146 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
147 ASSERT(lock_flags != 0);
148}
149
150/*
151 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
152 * multi-reader locks: invalidate_lock and the i_lock. This routine allows
153 * various combinations of the locks to be obtained.
154 *
155 * The 3 locks should always be ordered so that the IO lock is obtained first,
156 * the mmap lock second and the ilock last in order to prevent deadlock.
157 *
158 * Basic locking order:
159 *
160 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
161 *
162 * mmap_lock locking order:
163 *
164 * i_rwsem -> page lock -> mmap_lock
165 * mmap_lock -> invalidate_lock -> page_lock
166 *
167 * The difference in mmap_lock locking order mean that we cannot hold the
168 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
169 * can fault in pages during copy in/out (for buffered IO) or require the
170 * mmap_lock in get_user_pages() to map the user pages into the kernel address
171 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
172 * fault because page faults already hold the mmap_lock.
173 *
174 * Hence to serialise fully against both syscall and mmap based IO, we need to
175 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
176 * both taken in places where we need to invalidate the page cache in a race
177 * free manner (e.g. truncate, hole punch and other extent manipulation
178 * functions).
179 */
180void
181xfs_ilock(
182 xfs_inode_t *ip,
183 uint lock_flags)
184{
185 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
186
187 xfs_lock_flags_assert(lock_flags);
188
189 if (lock_flags & XFS_IOLOCK_EXCL) {
190 down_write_nested(&VFS_I(ip)->i_rwsem,
191 XFS_IOLOCK_DEP(lock_flags));
192 } else if (lock_flags & XFS_IOLOCK_SHARED) {
193 down_read_nested(&VFS_I(ip)->i_rwsem,
194 XFS_IOLOCK_DEP(lock_flags));
195 }
196
197 if (lock_flags & XFS_MMAPLOCK_EXCL) {
198 down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
199 XFS_MMAPLOCK_DEP(lock_flags));
200 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
201 down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
202 XFS_MMAPLOCK_DEP(lock_flags));
203 }
204
205 if (lock_flags & XFS_ILOCK_EXCL)
206 down_write_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
207 else if (lock_flags & XFS_ILOCK_SHARED)
208 down_read_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
209}
210
211/*
212 * This is just like xfs_ilock(), except that the caller
213 * is guaranteed not to sleep. It returns 1 if it gets
214 * the requested locks and 0 otherwise. If the IO lock is
215 * obtained but the inode lock cannot be, then the IO lock
216 * is dropped before returning.
217 *
218 * ip -- the inode being locked
219 * lock_flags -- this parameter indicates the inode's locks to be
220 * to be locked. See the comment for xfs_ilock() for a list
221 * of valid values.
222 */
223int
224xfs_ilock_nowait(
225 xfs_inode_t *ip,
226 uint lock_flags)
227{
228 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
229
230 xfs_lock_flags_assert(lock_flags);
231
232 if (lock_flags & XFS_IOLOCK_EXCL) {
233 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
234 goto out;
235 } else if (lock_flags & XFS_IOLOCK_SHARED) {
236 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
237 goto out;
238 }
239
240 if (lock_flags & XFS_MMAPLOCK_EXCL) {
241 if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
242 goto out_undo_iolock;
243 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
244 if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
245 goto out_undo_iolock;
246 }
247
248 if (lock_flags & XFS_ILOCK_EXCL) {
249 if (!down_write_trylock(&ip->i_lock))
250 goto out_undo_mmaplock;
251 } else if (lock_flags & XFS_ILOCK_SHARED) {
252 if (!down_read_trylock(&ip->i_lock))
253 goto out_undo_mmaplock;
254 }
255 return 1;
256
257out_undo_mmaplock:
258 if (lock_flags & XFS_MMAPLOCK_EXCL)
259 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
260 else if (lock_flags & XFS_MMAPLOCK_SHARED)
261 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
262out_undo_iolock:
263 if (lock_flags & XFS_IOLOCK_EXCL)
264 up_write(&VFS_I(ip)->i_rwsem);
265 else if (lock_flags & XFS_IOLOCK_SHARED)
266 up_read(&VFS_I(ip)->i_rwsem);
267out:
268 return 0;
269}
270
271/*
272 * xfs_iunlock() is used to drop the inode locks acquired with
273 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
274 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
275 * that we know which locks to drop.
276 *
277 * ip -- the inode being unlocked
278 * lock_flags -- this parameter indicates the inode's locks to be
279 * to be unlocked. See the comment for xfs_ilock() for a list
280 * of valid values for this parameter.
281 *
282 */
283void
284xfs_iunlock(
285 xfs_inode_t *ip,
286 uint lock_flags)
287{
288 xfs_lock_flags_assert(lock_flags);
289
290 if (lock_flags & XFS_IOLOCK_EXCL)
291 up_write(&VFS_I(ip)->i_rwsem);
292 else if (lock_flags & XFS_IOLOCK_SHARED)
293 up_read(&VFS_I(ip)->i_rwsem);
294
295 if (lock_flags & XFS_MMAPLOCK_EXCL)
296 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
297 else if (lock_flags & XFS_MMAPLOCK_SHARED)
298 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
299
300 if (lock_flags & XFS_ILOCK_EXCL)
301 up_write(&ip->i_lock);
302 else if (lock_flags & XFS_ILOCK_SHARED)
303 up_read(&ip->i_lock);
304
305 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
306}
307
308/*
309 * give up write locks. the i/o lock cannot be held nested
310 * if it is being demoted.
311 */
312void
313xfs_ilock_demote(
314 xfs_inode_t *ip,
315 uint lock_flags)
316{
317 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
318 ASSERT((lock_flags &
319 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
320
321 if (lock_flags & XFS_ILOCK_EXCL)
322 downgrade_write(&ip->i_lock);
323 if (lock_flags & XFS_MMAPLOCK_EXCL)
324 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
325 if (lock_flags & XFS_IOLOCK_EXCL)
326 downgrade_write(&VFS_I(ip)->i_rwsem);
327
328 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
329}
330
331void
332xfs_assert_ilocked(
333 struct xfs_inode *ip,
334 uint lock_flags)
335{
336 /*
337 * Sometimes we assert the ILOCK is held exclusively, but we're in
338 * a workqueue, so lockdep doesn't know we're the owner.
339 */
340 if (lock_flags & XFS_ILOCK_SHARED)
341 rwsem_assert_held(&ip->i_lock);
342 else if (lock_flags & XFS_ILOCK_EXCL)
343 rwsem_assert_held_write_nolockdep(&ip->i_lock);
344
345 if (lock_flags & XFS_MMAPLOCK_SHARED)
346 rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
347 else if (lock_flags & XFS_MMAPLOCK_EXCL)
348 rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
349
350 if (lock_flags & XFS_IOLOCK_SHARED)
351 rwsem_assert_held(&VFS_I(ip)->i_rwsem);
352 else if (lock_flags & XFS_IOLOCK_EXCL)
353 rwsem_assert_held_write(&VFS_I(ip)->i_rwsem);
354}
355
356/*
357 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
358 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
359 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
360 * errors and warnings.
361 */
362#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
363static bool
364xfs_lockdep_subclass_ok(
365 int subclass)
366{
367 return subclass < MAX_LOCKDEP_SUBCLASSES;
368}
369#else
370#define xfs_lockdep_subclass_ok(subclass) (true)
371#endif
372
373/*
374 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
375 * value. This can be called for any type of inode lock combination, including
376 * parent locking. Care must be taken to ensure we don't overrun the subclass
377 * storage fields in the class mask we build.
378 */
379static inline uint
380xfs_lock_inumorder(
381 uint lock_mode,
382 uint subclass)
383{
384 uint class = 0;
385
386 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
387 XFS_ILOCK_RTSUM)));
388 ASSERT(xfs_lockdep_subclass_ok(subclass));
389
390 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
391 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
392 class += subclass << XFS_IOLOCK_SHIFT;
393 }
394
395 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
396 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
397 class += subclass << XFS_MMAPLOCK_SHIFT;
398 }
399
400 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
401 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
402 class += subclass << XFS_ILOCK_SHIFT;
403 }
404
405 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
406}
407
408/*
409 * The following routine will lock n inodes in exclusive mode. We assume the
410 * caller calls us with the inodes in i_ino order.
411 *
412 * We need to detect deadlock where an inode that we lock is in the AIL and we
413 * start waiting for another inode that is locked by a thread in a long running
414 * transaction (such as truncate). This can result in deadlock since the long
415 * running trans might need to wait for the inode we just locked in order to
416 * push the tail and free space in the log.
417 *
418 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
419 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
420 * lock more than one at a time, lockdep will report false positives saying we
421 * have violated locking orders.
422 */
423static void
424xfs_lock_inodes(
425 struct xfs_inode **ips,
426 int inodes,
427 uint lock_mode)
428{
429 int attempts = 0;
430 uint i;
431 int j;
432 bool try_lock;
433 struct xfs_log_item *lp;
434
435 /*
436 * Currently supports between 2 and 5 inodes with exclusive locking. We
437 * support an arbitrary depth of locking here, but absolute limits on
438 * inodes depend on the type of locking and the limits placed by
439 * lockdep annotations in xfs_lock_inumorder. These are all checked by
440 * the asserts.
441 */
442 ASSERT(ips && inodes >= 2 && inodes <= 5);
443 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
444 XFS_ILOCK_EXCL));
445 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
446 XFS_ILOCK_SHARED)));
447 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
448 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
449 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
450 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
451
452 if (lock_mode & XFS_IOLOCK_EXCL) {
453 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
454 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
455 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
456
457again:
458 try_lock = false;
459 i = 0;
460 for (; i < inodes; i++) {
461 ASSERT(ips[i]);
462
463 if (i && (ips[i] == ips[i - 1])) /* Already locked */
464 continue;
465
466 /*
467 * If try_lock is not set yet, make sure all locked inodes are
468 * not in the AIL. If any are, set try_lock to be used later.
469 */
470 if (!try_lock) {
471 for (j = (i - 1); j >= 0 && !try_lock; j--) {
472 lp = &ips[j]->i_itemp->ili_item;
473 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
474 try_lock = true;
475 }
476 }
477
478 /*
479 * If any of the previous locks we have locked is in the AIL,
480 * we must TRY to get the second and subsequent locks. If
481 * we can't get any, we must release all we have
482 * and try again.
483 */
484 if (!try_lock) {
485 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
486 continue;
487 }
488
489 /* try_lock means we have an inode locked that is in the AIL. */
490 ASSERT(i != 0);
491 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
492 continue;
493
494 /*
495 * Unlock all previous guys and try again. xfs_iunlock will try
496 * to push the tail if the inode is in the AIL.
497 */
498 attempts++;
499 for (j = i - 1; j >= 0; j--) {
500 /*
501 * Check to see if we've already unlocked this one. Not
502 * the first one going back, and the inode ptr is the
503 * same.
504 */
505 if (j != (i - 1) && ips[j] == ips[j + 1])
506 continue;
507
508 xfs_iunlock(ips[j], lock_mode);
509 }
510
511 if ((attempts % 5) == 0) {
512 delay(1); /* Don't just spin the CPU */
513 }
514 goto again;
515 }
516}
517
518/*
519 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
520 * mmaplock must be double-locked separately since we use i_rwsem and
521 * invalidate_lock for that. We now support taking one lock EXCL and the
522 * other SHARED.
523 */
524void
525xfs_lock_two_inodes(
526 struct xfs_inode *ip0,
527 uint ip0_mode,
528 struct xfs_inode *ip1,
529 uint ip1_mode)
530{
531 int attempts = 0;
532 struct xfs_log_item *lp;
533
534 ASSERT(hweight32(ip0_mode) == 1);
535 ASSERT(hweight32(ip1_mode) == 1);
536 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
537 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
538 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
539 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
540 ASSERT(ip0->i_ino != ip1->i_ino);
541
542 if (ip0->i_ino > ip1->i_ino) {
543 swap(ip0, ip1);
544 swap(ip0_mode, ip1_mode);
545 }
546
547 again:
548 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
549
550 /*
551 * If the first lock we have locked is in the AIL, we must TRY to get
552 * the second lock. If we can't get it, we must release the first one
553 * and try again.
554 */
555 lp = &ip0->i_itemp->ili_item;
556 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
557 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
558 xfs_iunlock(ip0, ip0_mode);
559 if ((++attempts % 5) == 0)
560 delay(1); /* Don't just spin the CPU */
561 goto again;
562 }
563 } else {
564 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
565 }
566}
567
568uint
569xfs_ip2xflags(
570 struct xfs_inode *ip)
571{
572 uint flags = 0;
573
574 if (ip->i_diflags & XFS_DIFLAG_ANY) {
575 if (ip->i_diflags & XFS_DIFLAG_REALTIME)
576 flags |= FS_XFLAG_REALTIME;
577 if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
578 flags |= FS_XFLAG_PREALLOC;
579 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
580 flags |= FS_XFLAG_IMMUTABLE;
581 if (ip->i_diflags & XFS_DIFLAG_APPEND)
582 flags |= FS_XFLAG_APPEND;
583 if (ip->i_diflags & XFS_DIFLAG_SYNC)
584 flags |= FS_XFLAG_SYNC;
585 if (ip->i_diflags & XFS_DIFLAG_NOATIME)
586 flags |= FS_XFLAG_NOATIME;
587 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
588 flags |= FS_XFLAG_NODUMP;
589 if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
590 flags |= FS_XFLAG_RTINHERIT;
591 if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
592 flags |= FS_XFLAG_PROJINHERIT;
593 if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
594 flags |= FS_XFLAG_NOSYMLINKS;
595 if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
596 flags |= FS_XFLAG_EXTSIZE;
597 if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
598 flags |= FS_XFLAG_EXTSZINHERIT;
599 if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
600 flags |= FS_XFLAG_NODEFRAG;
601 if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
602 flags |= FS_XFLAG_FILESTREAM;
603 }
604
605 if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
606 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
607 flags |= FS_XFLAG_DAX;
608 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
609 flags |= FS_XFLAG_COWEXTSIZE;
610 }
611
612 if (xfs_inode_has_attr_fork(ip))
613 flags |= FS_XFLAG_HASATTR;
614 return flags;
615}
616
617/*
618 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
619 * is allowed, otherwise it has to be an exact match. If a CI match is found,
620 * ci_name->name will point to a the actual name (caller must free) or
621 * will be set to NULL if an exact match is found.
622 */
623int
624xfs_lookup(
625 struct xfs_inode *dp,
626 const struct xfs_name *name,
627 struct xfs_inode **ipp,
628 struct xfs_name *ci_name)
629{
630 xfs_ino_t inum;
631 int error;
632
633 trace_xfs_lookup(dp, name);
634
635 if (xfs_is_shutdown(dp->i_mount))
636 return -EIO;
637 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
638 return -EIO;
639
640 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
641 if (error)
642 goto out_unlock;
643
644 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
645 if (error)
646 goto out_free_name;
647
648 return 0;
649
650out_free_name:
651 if (ci_name)
652 kfree(ci_name->name);
653out_unlock:
654 *ipp = NULL;
655 return error;
656}
657
658/* Propagate di_flags from a parent inode to a child inode. */
659static void
660xfs_inode_inherit_flags(
661 struct xfs_inode *ip,
662 const struct xfs_inode *pip)
663{
664 unsigned int di_flags = 0;
665 xfs_failaddr_t failaddr;
666 umode_t mode = VFS_I(ip)->i_mode;
667
668 if (S_ISDIR(mode)) {
669 if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
670 di_flags |= XFS_DIFLAG_RTINHERIT;
671 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
672 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
673 ip->i_extsize = pip->i_extsize;
674 }
675 if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
676 di_flags |= XFS_DIFLAG_PROJINHERIT;
677 } else if (S_ISREG(mode)) {
678 if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
679 xfs_has_realtime(ip->i_mount))
680 di_flags |= XFS_DIFLAG_REALTIME;
681 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
682 di_flags |= XFS_DIFLAG_EXTSIZE;
683 ip->i_extsize = pip->i_extsize;
684 }
685 }
686 if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
687 xfs_inherit_noatime)
688 di_flags |= XFS_DIFLAG_NOATIME;
689 if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
690 xfs_inherit_nodump)
691 di_flags |= XFS_DIFLAG_NODUMP;
692 if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
693 xfs_inherit_sync)
694 di_flags |= XFS_DIFLAG_SYNC;
695 if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
696 xfs_inherit_nosymlinks)
697 di_flags |= XFS_DIFLAG_NOSYMLINKS;
698 if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
699 xfs_inherit_nodefrag)
700 di_flags |= XFS_DIFLAG_NODEFRAG;
701 if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
702 di_flags |= XFS_DIFLAG_FILESTREAM;
703
704 ip->i_diflags |= di_flags;
705
706 /*
707 * Inode verifiers on older kernels only check that the extent size
708 * hint is an integer multiple of the rt extent size on realtime files.
709 * They did not check the hint alignment on a directory with both
710 * rtinherit and extszinherit flags set. If the misaligned hint is
711 * propagated from a directory into a new realtime file, new file
712 * allocations will fail due to math errors in the rt allocator and/or
713 * trip the verifiers. Validate the hint settings in the new file so
714 * that we don't let broken hints propagate.
715 */
716 failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
717 VFS_I(ip)->i_mode, ip->i_diflags);
718 if (failaddr) {
719 ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
720 XFS_DIFLAG_EXTSZINHERIT);
721 ip->i_extsize = 0;
722 }
723}
724
725/* Propagate di_flags2 from a parent inode to a child inode. */
726static void
727xfs_inode_inherit_flags2(
728 struct xfs_inode *ip,
729 const struct xfs_inode *pip)
730{
731 xfs_failaddr_t failaddr;
732
733 if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
734 ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
735 ip->i_cowextsize = pip->i_cowextsize;
736 }
737 if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
738 ip->i_diflags2 |= XFS_DIFLAG2_DAX;
739
740 /* Don't let invalid cowextsize hints propagate. */
741 failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
742 VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
743 if (failaddr) {
744 ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
745 ip->i_cowextsize = 0;
746 }
747}
748
749/*
750 * Initialise a newly allocated inode and return the in-core inode to the
751 * caller locked exclusively.
752 */
753int
754xfs_init_new_inode(
755 struct mnt_idmap *idmap,
756 struct xfs_trans *tp,
757 struct xfs_inode *pip,
758 xfs_ino_t ino,
759 umode_t mode,
760 xfs_nlink_t nlink,
761 dev_t rdev,
762 prid_t prid,
763 bool init_xattrs,
764 struct xfs_inode **ipp)
765{
766 struct inode *dir = pip ? VFS_I(pip) : NULL;
767 struct xfs_mount *mp = tp->t_mountp;
768 struct xfs_inode *ip;
769 unsigned int flags;
770 int error;
771 struct timespec64 tv;
772 struct inode *inode;
773
774 /*
775 * Protect against obviously corrupt allocation btree records. Later
776 * xfs_iget checks will catch re-allocation of other active in-memory
777 * and on-disk inodes. If we don't catch reallocating the parent inode
778 * here we will deadlock in xfs_iget() so we have to do these checks
779 * first.
780 */
781 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
782 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
783 xfs_agno_mark_sick(mp, XFS_INO_TO_AGNO(mp, ino),
784 XFS_SICK_AG_INOBT);
785 return -EFSCORRUPTED;
786 }
787
788 /*
789 * Get the in-core inode with the lock held exclusively to prevent
790 * others from looking at until we're done.
791 */
792 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
793 if (error)
794 return error;
795
796 ASSERT(ip != NULL);
797 inode = VFS_I(ip);
798 set_nlink(inode, nlink);
799 inode->i_rdev = rdev;
800 ip->i_projid = prid;
801
802 if (dir && !(dir->i_mode & S_ISGID) && xfs_has_grpid(mp)) {
803 inode_fsuid_set(inode, idmap);
804 inode->i_gid = dir->i_gid;
805 inode->i_mode = mode;
806 } else {
807 inode_init_owner(idmap, inode, dir, mode);
808 }
809
810 /*
811 * If the group ID of the new file does not match the effective group
812 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
813 * (and only if the irix_sgid_inherit compatibility variable is set).
814 */
815 if (irix_sgid_inherit && (inode->i_mode & S_ISGID) &&
816 !vfsgid_in_group_p(i_gid_into_vfsgid(idmap, inode)))
817 inode->i_mode &= ~S_ISGID;
818
819 ip->i_disk_size = 0;
820 ip->i_df.if_nextents = 0;
821 ASSERT(ip->i_nblocks == 0);
822
823 tv = inode_set_ctime_current(inode);
824 inode_set_mtime_to_ts(inode, tv);
825 inode_set_atime_to_ts(inode, tv);
826
827 ip->i_extsize = 0;
828 ip->i_diflags = 0;
829
830 if (xfs_has_v3inodes(mp)) {
831 inode_set_iversion(inode, 1);
832 ip->i_cowextsize = 0;
833 ip->i_crtime = tv;
834 }
835
836 flags = XFS_ILOG_CORE;
837 switch (mode & S_IFMT) {
838 case S_IFIFO:
839 case S_IFCHR:
840 case S_IFBLK:
841 case S_IFSOCK:
842 ip->i_df.if_format = XFS_DINODE_FMT_DEV;
843 flags |= XFS_ILOG_DEV;
844 break;
845 case S_IFREG:
846 case S_IFDIR:
847 if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
848 xfs_inode_inherit_flags(ip, pip);
849 if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
850 xfs_inode_inherit_flags2(ip, pip);
851 fallthrough;
852 case S_IFLNK:
853 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
854 ip->i_df.if_bytes = 0;
855 ip->i_df.if_data = NULL;
856 break;
857 default:
858 ASSERT(0);
859 }
860
861 /*
862 * If we need to create attributes immediately after allocating the
863 * inode, initialise an empty attribute fork right now. We use the
864 * default fork offset for attributes here as we don't know exactly what
865 * size or how many attributes we might be adding. We can do this
866 * safely here because we know the data fork is completely empty and
867 * this saves us from needing to run a separate transaction to set the
868 * fork offset in the immediate future.
869 */
870 if (init_xattrs && xfs_has_attr(mp)) {
871 ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
872 xfs_ifork_init_attr(ip, XFS_DINODE_FMT_EXTENTS, 0);
873 }
874
875 /*
876 * Log the new values stuffed into the inode.
877 */
878 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
879 xfs_trans_log_inode(tp, ip, flags);
880
881 /* now that we have an i_mode we can setup the inode structure */
882 xfs_setup_inode(ip);
883
884 *ipp = ip;
885 return 0;
886}
887
888/*
889 * Decrement the link count on an inode & log the change. If this causes the
890 * link count to go to zero, move the inode to AGI unlinked list so that it can
891 * be freed when the last active reference goes away via xfs_inactive().
892 */
893static int /* error */
894xfs_droplink(
895 xfs_trans_t *tp,
896 xfs_inode_t *ip)
897{
898 if (VFS_I(ip)->i_nlink == 0) {
899 xfs_alert(ip->i_mount,
900 "%s: Attempt to drop inode (%llu) with nlink zero.",
901 __func__, ip->i_ino);
902 return -EFSCORRUPTED;
903 }
904
905 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
906
907 drop_nlink(VFS_I(ip));
908 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
909
910 if (VFS_I(ip)->i_nlink)
911 return 0;
912
913 return xfs_iunlink(tp, ip);
914}
915
916/*
917 * Increment the link count on an inode & log the change.
918 */
919static void
920xfs_bumplink(
921 xfs_trans_t *tp,
922 xfs_inode_t *ip)
923{
924 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
925
926 inc_nlink(VFS_I(ip));
927 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
928}
929
930#ifdef CONFIG_XFS_LIVE_HOOKS
931/*
932 * Use a static key here to reduce the overhead of directory live update hooks.
933 * If the compiler supports jump labels, the static branch will be replaced by
934 * a nop sled when there are no hook users. Online fsck is currently the only
935 * caller, so this is a reasonable tradeoff.
936 *
937 * Note: Patching the kernel code requires taking the cpu hotplug lock. Other
938 * parts of the kernel allocate memory with that lock held, which means that
939 * XFS callers cannot hold any locks that might be used by memory reclaim or
940 * writeback when calling the static_branch_{inc,dec} functions.
941 */
942DEFINE_STATIC_XFS_HOOK_SWITCH(xfs_dir_hooks_switch);
943
944void
945xfs_dir_hook_disable(void)
946{
947 xfs_hooks_switch_off(&xfs_dir_hooks_switch);
948}
949
950void
951xfs_dir_hook_enable(void)
952{
953 xfs_hooks_switch_on(&xfs_dir_hooks_switch);
954}
955
956/* Call hooks for a directory update relating to a child dirent update. */
957inline void
958xfs_dir_update_hook(
959 struct xfs_inode *dp,
960 struct xfs_inode *ip,
961 int delta,
962 const struct xfs_name *name)
963{
964 if (xfs_hooks_switched_on(&xfs_dir_hooks_switch)) {
965 struct xfs_dir_update_params p = {
966 .dp = dp,
967 .ip = ip,
968 .delta = delta,
969 .name = name,
970 };
971 struct xfs_mount *mp = ip->i_mount;
972
973 xfs_hooks_call(&mp->m_dir_update_hooks, 0, &p);
974 }
975}
976
977/* Call the specified function during a directory update. */
978int
979xfs_dir_hook_add(
980 struct xfs_mount *mp,
981 struct xfs_dir_hook *hook)
982{
983 return xfs_hooks_add(&mp->m_dir_update_hooks, &hook->dirent_hook);
984}
985
986/* Stop calling the specified function during a directory update. */
987void
988xfs_dir_hook_del(
989 struct xfs_mount *mp,
990 struct xfs_dir_hook *hook)
991{
992 xfs_hooks_del(&mp->m_dir_update_hooks, &hook->dirent_hook);
993}
994
995/* Configure directory update hook functions. */
996void
997xfs_dir_hook_setup(
998 struct xfs_dir_hook *hook,
999 notifier_fn_t mod_fn)
1000{
1001 xfs_hook_setup(&hook->dirent_hook, mod_fn);
1002}
1003#endif /* CONFIG_XFS_LIVE_HOOKS */
1004
1005int
1006xfs_create(
1007 struct mnt_idmap *idmap,
1008 xfs_inode_t *dp,
1009 struct xfs_name *name,
1010 umode_t mode,
1011 dev_t rdev,
1012 bool init_xattrs,
1013 xfs_inode_t **ipp)
1014{
1015 int is_dir = S_ISDIR(mode);
1016 struct xfs_mount *mp = dp->i_mount;
1017 struct xfs_inode *ip = NULL;
1018 struct xfs_trans *tp = NULL;
1019 int error;
1020 bool unlock_dp_on_error = false;
1021 prid_t prid;
1022 struct xfs_dquot *udqp = NULL;
1023 struct xfs_dquot *gdqp = NULL;
1024 struct xfs_dquot *pdqp = NULL;
1025 struct xfs_trans_res *tres;
1026 uint resblks;
1027 xfs_ino_t ino;
1028
1029 trace_xfs_create(dp, name);
1030
1031 if (xfs_is_shutdown(mp))
1032 return -EIO;
1033 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1034 return -EIO;
1035
1036 prid = xfs_get_initial_prid(dp);
1037
1038 /*
1039 * Make sure that we have allocated dquot(s) on disk.
1040 */
1041 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
1042 mapped_fsgid(idmap, &init_user_ns), prid,
1043 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1044 &udqp, &gdqp, &pdqp);
1045 if (error)
1046 return error;
1047
1048 if (is_dir) {
1049 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1050 tres = &M_RES(mp)->tr_mkdir;
1051 } else {
1052 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1053 tres = &M_RES(mp)->tr_create;
1054 }
1055
1056 /*
1057 * Initially assume that the file does not exist and
1058 * reserve the resources for that case. If that is not
1059 * the case we'll drop the one we have and get a more
1060 * appropriate transaction later.
1061 */
1062 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1063 &tp);
1064 if (error == -ENOSPC) {
1065 /* flush outstanding delalloc blocks and retry */
1066 xfs_flush_inodes(mp);
1067 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1068 resblks, &tp);
1069 }
1070 if (error)
1071 goto out_release_dquots;
1072
1073 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1074 unlock_dp_on_error = true;
1075
1076 /*
1077 * A newly created regular or special file just has one directory
1078 * entry pointing to them, but a directory also the "." entry
1079 * pointing to itself.
1080 */
1081 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1082 if (!error)
1083 error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1084 is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1085 if (error)
1086 goto out_trans_cancel;
1087
1088 /*
1089 * Now we join the directory inode to the transaction. We do not do it
1090 * earlier because xfs_dialloc might commit the previous transaction
1091 * (and release all the locks). An error from here on will result in
1092 * the transaction cancel unlocking dp so don't do it explicitly in the
1093 * error path.
1094 */
1095 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1096 unlock_dp_on_error = false;
1097
1098 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1099 resblks - XFS_IALLOC_SPACE_RES(mp));
1100 if (error) {
1101 ASSERT(error != -ENOSPC);
1102 goto out_trans_cancel;
1103 }
1104 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1105 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1106
1107 if (is_dir) {
1108 error = xfs_dir_init(tp, ip, dp);
1109 if (error)
1110 goto out_trans_cancel;
1111
1112 xfs_bumplink(tp, dp);
1113 }
1114
1115 /*
1116 * Create ip with a reference from dp, and add '.' and '..' references
1117 * if it's a directory.
1118 */
1119 xfs_dir_update_hook(dp, ip, 1, name);
1120
1121 /*
1122 * If this is a synchronous mount, make sure that the
1123 * create transaction goes to disk before returning to
1124 * the user.
1125 */
1126 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1127 xfs_trans_set_sync(tp);
1128
1129 /*
1130 * Attach the dquot(s) to the inodes and modify them incore.
1131 * These ids of the inode couldn't have changed since the new
1132 * inode has been locked ever since it was created.
1133 */
1134 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1135
1136 error = xfs_trans_commit(tp);
1137 if (error)
1138 goto out_release_inode;
1139
1140 xfs_qm_dqrele(udqp);
1141 xfs_qm_dqrele(gdqp);
1142 xfs_qm_dqrele(pdqp);
1143
1144 *ipp = ip;
1145 return 0;
1146
1147 out_trans_cancel:
1148 xfs_trans_cancel(tp);
1149 out_release_inode:
1150 /*
1151 * Wait until after the current transaction is aborted to finish the
1152 * setup of the inode and release the inode. This prevents recursive
1153 * transactions and deadlocks from xfs_inactive.
1154 */
1155 if (ip) {
1156 xfs_finish_inode_setup(ip);
1157 xfs_irele(ip);
1158 }
1159 out_release_dquots:
1160 xfs_qm_dqrele(udqp);
1161 xfs_qm_dqrele(gdqp);
1162 xfs_qm_dqrele(pdqp);
1163
1164 if (unlock_dp_on_error)
1165 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1166 return error;
1167}
1168
1169int
1170xfs_create_tmpfile(
1171 struct mnt_idmap *idmap,
1172 struct xfs_inode *dp,
1173 umode_t mode,
1174 struct xfs_inode **ipp)
1175{
1176 struct xfs_mount *mp = dp->i_mount;
1177 struct xfs_inode *ip = NULL;
1178 struct xfs_trans *tp = NULL;
1179 int error;
1180 prid_t prid;
1181 struct xfs_dquot *udqp = NULL;
1182 struct xfs_dquot *gdqp = NULL;
1183 struct xfs_dquot *pdqp = NULL;
1184 struct xfs_trans_res *tres;
1185 uint resblks;
1186 xfs_ino_t ino;
1187
1188 if (xfs_is_shutdown(mp))
1189 return -EIO;
1190
1191 prid = xfs_get_initial_prid(dp);
1192
1193 /*
1194 * Make sure that we have allocated dquot(s) on disk.
1195 */
1196 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
1197 mapped_fsgid(idmap, &init_user_ns), prid,
1198 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1199 &udqp, &gdqp, &pdqp);
1200 if (error)
1201 return error;
1202
1203 resblks = XFS_IALLOC_SPACE_RES(mp);
1204 tres = &M_RES(mp)->tr_create_tmpfile;
1205
1206 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1207 &tp);
1208 if (error)
1209 goto out_release_dquots;
1210
1211 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1212 if (!error)
1213 error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1214 0, 0, prid, false, &ip);
1215 if (error)
1216 goto out_trans_cancel;
1217
1218 if (xfs_has_wsync(mp))
1219 xfs_trans_set_sync(tp);
1220
1221 /*
1222 * Attach the dquot(s) to the inodes and modify them incore.
1223 * These ids of the inode couldn't have changed since the new
1224 * inode has been locked ever since it was created.
1225 */
1226 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1227
1228 error = xfs_iunlink(tp, ip);
1229 if (error)
1230 goto out_trans_cancel;
1231
1232 error = xfs_trans_commit(tp);
1233 if (error)
1234 goto out_release_inode;
1235
1236 xfs_qm_dqrele(udqp);
1237 xfs_qm_dqrele(gdqp);
1238 xfs_qm_dqrele(pdqp);
1239
1240 *ipp = ip;
1241 return 0;
1242
1243 out_trans_cancel:
1244 xfs_trans_cancel(tp);
1245 out_release_inode:
1246 /*
1247 * Wait until after the current transaction is aborted to finish the
1248 * setup of the inode and release the inode. This prevents recursive
1249 * transactions and deadlocks from xfs_inactive.
1250 */
1251 if (ip) {
1252 xfs_finish_inode_setup(ip);
1253 xfs_irele(ip);
1254 }
1255 out_release_dquots:
1256 xfs_qm_dqrele(udqp);
1257 xfs_qm_dqrele(gdqp);
1258 xfs_qm_dqrele(pdqp);
1259
1260 return error;
1261}
1262
1263int
1264xfs_link(
1265 xfs_inode_t *tdp,
1266 xfs_inode_t *sip,
1267 struct xfs_name *target_name)
1268{
1269 xfs_mount_t *mp = tdp->i_mount;
1270 xfs_trans_t *tp;
1271 int error, nospace_error = 0;
1272 int resblks;
1273
1274 trace_xfs_link(tdp, target_name);
1275
1276 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1277
1278 if (xfs_is_shutdown(mp))
1279 return -EIO;
1280 if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
1281 return -EIO;
1282
1283 error = xfs_qm_dqattach(sip);
1284 if (error)
1285 goto std_return;
1286
1287 error = xfs_qm_dqattach(tdp);
1288 if (error)
1289 goto std_return;
1290
1291 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1292 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
1293 &tp, &nospace_error);
1294 if (error)
1295 goto std_return;
1296
1297 /*
1298 * If we are using project inheritance, we only allow hard link
1299 * creation in our tree when the project IDs are the same; else
1300 * the tree quota mechanism could be circumvented.
1301 */
1302 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1303 tdp->i_projid != sip->i_projid)) {
1304 /*
1305 * Project quota setup skips special files which can
1306 * leave inodes in a PROJINHERIT directory without a
1307 * project ID set. We need to allow links to be made
1308 * to these "project-less" inodes because userspace
1309 * expects them to succeed after project ID setup,
1310 * but everything else should be rejected.
1311 */
1312 if (!special_file(VFS_I(sip)->i_mode) ||
1313 sip->i_projid != 0) {
1314 error = -EXDEV;
1315 goto error_return;
1316 }
1317 }
1318
1319 if (!resblks) {
1320 error = xfs_dir_canenter(tp, tdp, target_name);
1321 if (error)
1322 goto error_return;
1323 }
1324
1325 /*
1326 * Handle initial link state of O_TMPFILE inode
1327 */
1328 if (VFS_I(sip)->i_nlink == 0) {
1329 struct xfs_perag *pag;
1330
1331 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1332 error = xfs_iunlink_remove(tp, pag, sip);
1333 xfs_perag_put(pag);
1334 if (error)
1335 goto error_return;
1336 }
1337
1338 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1339 resblks);
1340 if (error)
1341 goto error_return;
1342 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1343 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1344
1345 xfs_bumplink(tp, sip);
1346 xfs_dir_update_hook(tdp, sip, 1, target_name);
1347
1348 /*
1349 * If this is a synchronous mount, make sure that the
1350 * link transaction goes to disk before returning to
1351 * the user.
1352 */
1353 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1354 xfs_trans_set_sync(tp);
1355
1356 return xfs_trans_commit(tp);
1357
1358 error_return:
1359 xfs_trans_cancel(tp);
1360 std_return:
1361 if (error == -ENOSPC && nospace_error)
1362 error = nospace_error;
1363 return error;
1364}
1365
1366/* Clear the reflink flag and the cowblocks tag if possible. */
1367static void
1368xfs_itruncate_clear_reflink_flags(
1369 struct xfs_inode *ip)
1370{
1371 struct xfs_ifork *dfork;
1372 struct xfs_ifork *cfork;
1373
1374 if (!xfs_is_reflink_inode(ip))
1375 return;
1376 dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1377 cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
1378 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1379 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1380 if (cfork->if_bytes == 0)
1381 xfs_inode_clear_cowblocks_tag(ip);
1382}
1383
1384/*
1385 * Free up the underlying blocks past new_size. The new size must be smaller
1386 * than the current size. This routine can be used both for the attribute and
1387 * data fork, and does not modify the inode size, which is left to the caller.
1388 *
1389 * The transaction passed to this routine must have made a permanent log
1390 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1391 * given transaction and start new ones, so make sure everything involved in
1392 * the transaction is tidy before calling here. Some transaction will be
1393 * returned to the caller to be committed. The incoming transaction must
1394 * already include the inode, and both inode locks must be held exclusively.
1395 * The inode must also be "held" within the transaction. On return the inode
1396 * will be "held" within the returned transaction. This routine does NOT
1397 * require any disk space to be reserved for it within the transaction.
1398 *
1399 * If we get an error, we must return with the inode locked and linked into the
1400 * current transaction. This keeps things simple for the higher level code,
1401 * because it always knows that the inode is locked and held in the transaction
1402 * that returns to it whether errors occur or not. We don't mark the inode
1403 * dirty on error so that transactions can be easily aborted if possible.
1404 */
1405int
1406xfs_itruncate_extents_flags(
1407 struct xfs_trans **tpp,
1408 struct xfs_inode *ip,
1409 int whichfork,
1410 xfs_fsize_t new_size,
1411 int flags)
1412{
1413 struct xfs_mount *mp = ip->i_mount;
1414 struct xfs_trans *tp = *tpp;
1415 xfs_fileoff_t first_unmap_block;
1416 int error = 0;
1417
1418 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1419 if (atomic_read(&VFS_I(ip)->i_count))
1420 xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1421 ASSERT(new_size <= XFS_ISIZE(ip));
1422 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1423 ASSERT(ip->i_itemp != NULL);
1424 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1425 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1426
1427 trace_xfs_itruncate_extents_start(ip, new_size);
1428
1429 flags |= xfs_bmapi_aflag(whichfork);
1430
1431 /*
1432 * Since it is possible for space to become allocated beyond
1433 * the end of the file (in a crash where the space is allocated
1434 * but the inode size is not yet updated), simply remove any
1435 * blocks which show up between the new EOF and the maximum
1436 * possible file size.
1437 *
1438 * We have to free all the blocks to the bmbt maximum offset, even if
1439 * the page cache can't scale that far.
1440 */
1441 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1442 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1443 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1444 return 0;
1445 }
1446
1447 error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1448 XFS_MAX_FILEOFF);
1449 if (error)
1450 goto out;
1451
1452 if (whichfork == XFS_DATA_FORK) {
1453 /* Remove all pending CoW reservations. */
1454 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1455 first_unmap_block, XFS_MAX_FILEOFF, true);
1456 if (error)
1457 goto out;
1458
1459 xfs_itruncate_clear_reflink_flags(ip);
1460 }
1461
1462 /*
1463 * Always re-log the inode so that our permanent transaction can keep
1464 * on rolling it forward in the log.
1465 */
1466 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1467
1468 trace_xfs_itruncate_extents_end(ip, new_size);
1469
1470out:
1471 *tpp = tp;
1472 return error;
1473}
1474
1475int
1476xfs_release(
1477 xfs_inode_t *ip)
1478{
1479 xfs_mount_t *mp = ip->i_mount;
1480 int error = 0;
1481
1482 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1483 return 0;
1484
1485 /* If this is a read-only mount, don't do this (would generate I/O) */
1486 if (xfs_is_readonly(mp))
1487 return 0;
1488
1489 if (!xfs_is_shutdown(mp)) {
1490 int truncated;
1491
1492 /*
1493 * If we previously truncated this file and removed old data
1494 * in the process, we want to initiate "early" writeout on
1495 * the last close. This is an attempt to combat the notorious
1496 * NULL files problem which is particularly noticeable from a
1497 * truncate down, buffered (re-)write (delalloc), followed by
1498 * a crash. What we are effectively doing here is
1499 * significantly reducing the time window where we'd otherwise
1500 * be exposed to that problem.
1501 */
1502 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1503 if (truncated) {
1504 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1505 if (ip->i_delayed_blks > 0) {
1506 error = filemap_flush(VFS_I(ip)->i_mapping);
1507 if (error)
1508 return error;
1509 }
1510 }
1511 }
1512
1513 if (VFS_I(ip)->i_nlink == 0)
1514 return 0;
1515
1516 /*
1517 * If we can't get the iolock just skip truncating the blocks past EOF
1518 * because we could deadlock with the mmap_lock otherwise. We'll get
1519 * another chance to drop them once the last reference to the inode is
1520 * dropped, so we'll never leak blocks permanently.
1521 */
1522 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1523 return 0;
1524
1525 if (xfs_can_free_eofblocks(ip, false)) {
1526 /*
1527 * Check if the inode is being opened, written and closed
1528 * frequently and we have delayed allocation blocks outstanding
1529 * (e.g. streaming writes from the NFS server), truncating the
1530 * blocks past EOF will cause fragmentation to occur.
1531 *
1532 * In this case don't do the truncation, but we have to be
1533 * careful how we detect this case. Blocks beyond EOF show up as
1534 * i_delayed_blks even when the inode is clean, so we need to
1535 * truncate them away first before checking for a dirty release.
1536 * Hence on the first dirty close we will still remove the
1537 * speculative allocation, but after that we will leave it in
1538 * place.
1539 */
1540 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1541 goto out_unlock;
1542
1543 error = xfs_free_eofblocks(ip);
1544 if (error)
1545 goto out_unlock;
1546
1547 /* delalloc blocks after truncation means it really is dirty */
1548 if (ip->i_delayed_blks)
1549 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1550 }
1551
1552out_unlock:
1553 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1554 return error;
1555}
1556
1557/*
1558 * xfs_inactive_truncate
1559 *
1560 * Called to perform a truncate when an inode becomes unlinked.
1561 */
1562STATIC int
1563xfs_inactive_truncate(
1564 struct xfs_inode *ip)
1565{
1566 struct xfs_mount *mp = ip->i_mount;
1567 struct xfs_trans *tp;
1568 int error;
1569
1570 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1571 if (error) {
1572 ASSERT(xfs_is_shutdown(mp));
1573 return error;
1574 }
1575 xfs_ilock(ip, XFS_ILOCK_EXCL);
1576 xfs_trans_ijoin(tp, ip, 0);
1577
1578 /*
1579 * Log the inode size first to prevent stale data exposure in the event
1580 * of a system crash before the truncate completes. See the related
1581 * comment in xfs_vn_setattr_size() for details.
1582 */
1583 ip->i_disk_size = 0;
1584 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1585
1586 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1587 if (error)
1588 goto error_trans_cancel;
1589
1590 ASSERT(ip->i_df.if_nextents == 0);
1591
1592 error = xfs_trans_commit(tp);
1593 if (error)
1594 goto error_unlock;
1595
1596 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1597 return 0;
1598
1599error_trans_cancel:
1600 xfs_trans_cancel(tp);
1601error_unlock:
1602 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1603 return error;
1604}
1605
1606/*
1607 * xfs_inactive_ifree()
1608 *
1609 * Perform the inode free when an inode is unlinked.
1610 */
1611STATIC int
1612xfs_inactive_ifree(
1613 struct xfs_inode *ip)
1614{
1615 struct xfs_mount *mp = ip->i_mount;
1616 struct xfs_trans *tp;
1617 int error;
1618
1619 /*
1620 * We try to use a per-AG reservation for any block needed by the finobt
1621 * tree, but as the finobt feature predates the per-AG reservation
1622 * support a degraded file system might not have enough space for the
1623 * reservation at mount time. In that case try to dip into the reserved
1624 * pool and pray.
1625 *
1626 * Send a warning if the reservation does happen to fail, as the inode
1627 * now remains allocated and sits on the unlinked list until the fs is
1628 * repaired.
1629 */
1630 if (unlikely(mp->m_finobt_nores)) {
1631 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1632 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1633 &tp);
1634 } else {
1635 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1636 }
1637 if (error) {
1638 if (error == -ENOSPC) {
1639 xfs_warn_ratelimited(mp,
1640 "Failed to remove inode(s) from unlinked list. "
1641 "Please free space, unmount and run xfs_repair.");
1642 } else {
1643 ASSERT(xfs_is_shutdown(mp));
1644 }
1645 return error;
1646 }
1647
1648 /*
1649 * We do not hold the inode locked across the entire rolling transaction
1650 * here. We only need to hold it for the first transaction that
1651 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1652 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1653 * here breaks the relationship between cluster buffer invalidation and
1654 * stale inode invalidation on cluster buffer item journal commit
1655 * completion, and can result in leaving dirty stale inodes hanging
1656 * around in memory.
1657 *
1658 * We have no need for serialising this inode operation against other
1659 * operations - we freed the inode and hence reallocation is required
1660 * and that will serialise on reallocating the space the deferops need
1661 * to free. Hence we can unlock the inode on the first commit of
1662 * the transaction rather than roll it right through the deferops. This
1663 * avoids relogging the XFS_ISTALE inode.
1664 *
1665 * We check that xfs_ifree() hasn't grown an internal transaction roll
1666 * by asserting that the inode is still locked when it returns.
1667 */
1668 xfs_ilock(ip, XFS_ILOCK_EXCL);
1669 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1670
1671 error = xfs_ifree(tp, ip);
1672 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1673 if (error) {
1674 /*
1675 * If we fail to free the inode, shut down. The cancel
1676 * might do that, we need to make sure. Otherwise the
1677 * inode might be lost for a long time or forever.
1678 */
1679 if (!xfs_is_shutdown(mp)) {
1680 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1681 __func__, error);
1682 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1683 }
1684 xfs_trans_cancel(tp);
1685 return error;
1686 }
1687
1688 /*
1689 * Credit the quota account(s). The inode is gone.
1690 */
1691 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1692
1693 return xfs_trans_commit(tp);
1694}
1695
1696/*
1697 * Returns true if we need to update the on-disk metadata before we can free
1698 * the memory used by this inode. Updates include freeing post-eof
1699 * preallocations; freeing COW staging extents; and marking the inode free in
1700 * the inobt if it is on the unlinked list.
1701 */
1702bool
1703xfs_inode_needs_inactive(
1704 struct xfs_inode *ip)
1705{
1706 struct xfs_mount *mp = ip->i_mount;
1707 struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1708
1709 /*
1710 * If the inode is already free, then there can be nothing
1711 * to clean up here.
1712 */
1713 if (VFS_I(ip)->i_mode == 0)
1714 return false;
1715
1716 /*
1717 * If this is a read-only mount, don't do this (would generate I/O)
1718 * unless we're in log recovery and cleaning the iunlinked list.
1719 */
1720 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1721 return false;
1722
1723 /* If the log isn't running, push inodes straight to reclaim. */
1724 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1725 return false;
1726
1727 /* Metadata inodes require explicit resource cleanup. */
1728 if (xfs_is_metadata_inode(ip))
1729 return false;
1730
1731 /* Want to clean out the cow blocks if there are any. */
1732 if (cow_ifp && cow_ifp->if_bytes > 0)
1733 return true;
1734
1735 /* Unlinked files must be freed. */
1736 if (VFS_I(ip)->i_nlink == 0)
1737 return true;
1738
1739 /*
1740 * This file isn't being freed, so check if there are post-eof blocks
1741 * to free. @force is true because we are evicting an inode from the
1742 * cache. Post-eof blocks must be freed, lest we end up with broken
1743 * free space accounting.
1744 *
1745 * Note: don't bother with iolock here since lockdep complains about
1746 * acquiring it in reclaim context. We have the only reference to the
1747 * inode at this point anyways.
1748 */
1749 return xfs_can_free_eofblocks(ip, true);
1750}
1751
1752/*
1753 * Save health status somewhere, if we're dumping an inode with uncorrected
1754 * errors and online repair isn't running.
1755 */
1756static inline void
1757xfs_inactive_health(
1758 struct xfs_inode *ip)
1759{
1760 struct xfs_mount *mp = ip->i_mount;
1761 struct xfs_perag *pag;
1762 unsigned int sick;
1763 unsigned int checked;
1764
1765 xfs_inode_measure_sickness(ip, &sick, &checked);
1766 if (!sick)
1767 return;
1768
1769 trace_xfs_inode_unfixed_corruption(ip, sick);
1770
1771 if (sick & XFS_SICK_INO_FORGET)
1772 return;
1773
1774 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1775 if (!pag) {
1776 /* There had better still be a perag structure! */
1777 ASSERT(0);
1778 return;
1779 }
1780
1781 xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1782 xfs_perag_put(pag);
1783}
1784
1785/*
1786 * xfs_inactive
1787 *
1788 * This is called when the vnode reference count for the vnode
1789 * goes to zero. If the file has been unlinked, then it must
1790 * now be truncated. Also, we clear all of the read-ahead state
1791 * kept for the inode here since the file is now closed.
1792 */
1793int
1794xfs_inactive(
1795 xfs_inode_t *ip)
1796{
1797 struct xfs_mount *mp;
1798 int error = 0;
1799 int truncate = 0;
1800
1801 /*
1802 * If the inode is already free, then there can be nothing
1803 * to clean up here.
1804 */
1805 if (VFS_I(ip)->i_mode == 0) {
1806 ASSERT(ip->i_df.if_broot_bytes == 0);
1807 goto out;
1808 }
1809
1810 mp = ip->i_mount;
1811 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1812
1813 xfs_inactive_health(ip);
1814
1815 /*
1816 * If this is a read-only mount, don't do this (would generate I/O)
1817 * unless we're in log recovery and cleaning the iunlinked list.
1818 */
1819 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1820 goto out;
1821
1822 /* Metadata inodes require explicit resource cleanup. */
1823 if (xfs_is_metadata_inode(ip))
1824 goto out;
1825
1826 /* Try to clean out the cow blocks if there are any. */
1827 if (xfs_inode_has_cow_data(ip))
1828 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1829
1830 if (VFS_I(ip)->i_nlink != 0) {
1831 /*
1832 * force is true because we are evicting an inode from the
1833 * cache. Post-eof blocks must be freed, lest we end up with
1834 * broken free space accounting.
1835 *
1836 * Note: don't bother with iolock here since lockdep complains
1837 * about acquiring it in reclaim context. We have the only
1838 * reference to the inode at this point anyways.
1839 */
1840 if (xfs_can_free_eofblocks(ip, true))
1841 error = xfs_free_eofblocks(ip);
1842
1843 goto out;
1844 }
1845
1846 if (S_ISREG(VFS_I(ip)->i_mode) &&
1847 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1848 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1849 truncate = 1;
1850
1851 if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1852 /*
1853 * If this inode is being inactivated during a quotacheck and
1854 * has not yet been scanned by quotacheck, we /must/ remove
1855 * the dquots from the inode before inactivation changes the
1856 * block and inode counts. Most probably this is a result of
1857 * reloading the incore iunlinked list to purge unrecovered
1858 * unlinked inodes.
1859 */
1860 xfs_qm_dqdetach(ip);
1861 } else {
1862 error = xfs_qm_dqattach(ip);
1863 if (error)
1864 goto out;
1865 }
1866
1867 if (S_ISLNK(VFS_I(ip)->i_mode))
1868 error = xfs_inactive_symlink(ip);
1869 else if (truncate)
1870 error = xfs_inactive_truncate(ip);
1871 if (error)
1872 goto out;
1873
1874 /*
1875 * If there are attributes associated with the file then blow them away
1876 * now. The code calls a routine that recursively deconstructs the
1877 * attribute fork. If also blows away the in-core attribute fork.
1878 */
1879 if (xfs_inode_has_attr_fork(ip)) {
1880 error = xfs_attr_inactive(ip);
1881 if (error)
1882 goto out;
1883 }
1884
1885 ASSERT(ip->i_forkoff == 0);
1886
1887 /*
1888 * Free the inode.
1889 */
1890 error = xfs_inactive_ifree(ip);
1891
1892out:
1893 /*
1894 * We're done making metadata updates for this inode, so we can release
1895 * the attached dquots.
1896 */
1897 xfs_qm_dqdetach(ip);
1898 return error;
1899}
1900
1901/*
1902 * In-Core Unlinked List Lookups
1903 * =============================
1904 *
1905 * Every inode is supposed to be reachable from some other piece of metadata
1906 * with the exception of the root directory. Inodes with a connection to a
1907 * file descriptor but not linked from anywhere in the on-disk directory tree
1908 * are collectively known as unlinked inodes, though the filesystem itself
1909 * maintains links to these inodes so that on-disk metadata are consistent.
1910 *
1911 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1912 * header contains a number of buckets that point to an inode, and each inode
1913 * record has a pointer to the next inode in the hash chain. This
1914 * singly-linked list causes scaling problems in the iunlink remove function
1915 * because we must walk that list to find the inode that points to the inode
1916 * being removed from the unlinked hash bucket list.
1917 *
1918 * Hence we keep an in-memory double linked list to link each inode on an
1919 * unlinked list. Because there are 64 unlinked lists per AGI, keeping pointer
1920 * based lists would require having 64 list heads in the perag, one for each
1921 * list. This is expensive in terms of memory (think millions of AGs) and cache
1922 * misses on lookups. Instead, use the fact that inodes on the unlinked list
1923 * must be referenced at the VFS level to keep them on the list and hence we
1924 * have an existence guarantee for inodes on the unlinked list.
1925 *
1926 * Given we have an existence guarantee, we can use lockless inode cache lookups
1927 * to resolve aginos to xfs inodes. This means we only need 8 bytes per inode
1928 * for the double linked unlinked list, and we don't need any extra locking to
1929 * keep the list safe as all manipulations are done under the AGI buffer lock.
1930 * Keeping the list up to date does not require memory allocation, just finding
1931 * the XFS inode and updating the next/prev unlinked list aginos.
1932 */
1933
1934/*
1935 * Find an inode on the unlinked list. This does not take references to the
1936 * inode as we have existence guarantees by holding the AGI buffer lock and that
1937 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1938 * don't find the inode in cache, then let the caller handle the situation.
1939 */
1940static struct xfs_inode *
1941xfs_iunlink_lookup(
1942 struct xfs_perag *pag,
1943 xfs_agino_t agino)
1944{
1945 struct xfs_inode *ip;
1946
1947 rcu_read_lock();
1948 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1949 if (!ip) {
1950 /* Caller can handle inode not being in memory. */
1951 rcu_read_unlock();
1952 return NULL;
1953 }
1954
1955 /*
1956 * Inode in RCU freeing limbo should not happen. Warn about this and
1957 * let the caller handle the failure.
1958 */
1959 if (WARN_ON_ONCE(!ip->i_ino)) {
1960 rcu_read_unlock();
1961 return NULL;
1962 }
1963 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1964 rcu_read_unlock();
1965 return ip;
1966}
1967
1968/*
1969 * Update the prev pointer of the next agino. Returns -ENOLINK if the inode
1970 * is not in cache.
1971 */
1972static int
1973xfs_iunlink_update_backref(
1974 struct xfs_perag *pag,
1975 xfs_agino_t prev_agino,
1976 xfs_agino_t next_agino)
1977{
1978 struct xfs_inode *ip;
1979
1980 /* No update necessary if we are at the end of the list. */
1981 if (next_agino == NULLAGINO)
1982 return 0;
1983
1984 ip = xfs_iunlink_lookup(pag, next_agino);
1985 if (!ip)
1986 return -ENOLINK;
1987
1988 ip->i_prev_unlinked = prev_agino;
1989 return 0;
1990}
1991
1992/*
1993 * Point the AGI unlinked bucket at an inode and log the results. The caller
1994 * is responsible for validating the old value.
1995 */
1996STATIC int
1997xfs_iunlink_update_bucket(
1998 struct xfs_trans *tp,
1999 struct xfs_perag *pag,
2000 struct xfs_buf *agibp,
2001 unsigned int bucket_index,
2002 xfs_agino_t new_agino)
2003{
2004 struct xfs_agi *agi = agibp->b_addr;
2005 xfs_agino_t old_value;
2006 int offset;
2007
2008 ASSERT(xfs_verify_agino_or_null(pag, new_agino));
2009
2010 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2011 trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
2012 old_value, new_agino);
2013
2014 /*
2015 * We should never find the head of the list already set to the value
2016 * passed in because either we're adding or removing ourselves from the
2017 * head of the list.
2018 */
2019 if (old_value == new_agino) {
2020 xfs_buf_mark_corrupt(agibp);
2021 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2022 return -EFSCORRUPTED;
2023 }
2024
2025 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2026 offset = offsetof(struct xfs_agi, agi_unlinked) +
2027 (sizeof(xfs_agino_t) * bucket_index);
2028 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2029 return 0;
2030}
2031
2032/*
2033 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
2034 * to @prev_agino. Caller must hold the AGI to synchronize with other changes
2035 * to the unlinked list.
2036 */
2037STATIC int
2038xfs_iunlink_reload_next(
2039 struct xfs_trans *tp,
2040 struct xfs_buf *agibp,
2041 xfs_agino_t prev_agino,
2042 xfs_agino_t next_agino)
2043{
2044 struct xfs_perag *pag = agibp->b_pag;
2045 struct xfs_mount *mp = pag->pag_mount;
2046 struct xfs_inode *next_ip = NULL;
2047 xfs_ino_t ino;
2048 int error;
2049
2050 ASSERT(next_agino != NULLAGINO);
2051
2052#ifdef DEBUG
2053 rcu_read_lock();
2054 next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
2055 ASSERT(next_ip == NULL);
2056 rcu_read_unlock();
2057#endif
2058
2059 xfs_info_ratelimited(mp,
2060 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
2061 next_agino, pag->pag_agno);
2062
2063 /*
2064 * Use an untrusted lookup just to be cautious in case the AGI has been
2065 * corrupted and now points at a free inode. That shouldn't happen,
2066 * but we'd rather shut down now since we're already running in a weird
2067 * situation.
2068 */
2069 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, next_agino);
2070 error = xfs_iget(mp, tp, ino, XFS_IGET_UNTRUSTED, 0, &next_ip);
2071 if (error) {
2072 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2073 return error;
2074 }
2075
2076 /* If this is not an unlinked inode, something is very wrong. */
2077 if (VFS_I(next_ip)->i_nlink != 0) {
2078 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2079 error = -EFSCORRUPTED;
2080 goto rele;
2081 }
2082
2083 next_ip->i_prev_unlinked = prev_agino;
2084 trace_xfs_iunlink_reload_next(next_ip);
2085rele:
2086 ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
2087 if (xfs_is_quotacheck_running(mp) && next_ip)
2088 xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
2089 xfs_irele(next_ip);
2090 return error;
2091}
2092
2093static int
2094xfs_iunlink_insert_inode(
2095 struct xfs_trans *tp,
2096 struct xfs_perag *pag,
2097 struct xfs_buf *agibp,
2098 struct xfs_inode *ip)
2099{
2100 struct xfs_mount *mp = tp->t_mountp;
2101 struct xfs_agi *agi = agibp->b_addr;
2102 xfs_agino_t next_agino;
2103 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2104 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2105 int error;
2106
2107 /*
2108 * Get the index into the agi hash table for the list this inode will
2109 * go on. Make sure the pointer isn't garbage and that this inode
2110 * isn't already on the list.
2111 */
2112 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2113 if (next_agino == agino ||
2114 !xfs_verify_agino_or_null(pag, next_agino)) {
2115 xfs_buf_mark_corrupt(agibp);
2116 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2117 return -EFSCORRUPTED;
2118 }
2119
2120 /*
2121 * Update the prev pointer in the next inode to point back to this
2122 * inode.
2123 */
2124 error = xfs_iunlink_update_backref(pag, agino, next_agino);
2125 if (error == -ENOLINK)
2126 error = xfs_iunlink_reload_next(tp, agibp, agino, next_agino);
2127 if (error)
2128 return error;
2129
2130 if (next_agino != NULLAGINO) {
2131 /*
2132 * There is already another inode in the bucket, so point this
2133 * inode to the current head of the list.
2134 */
2135 error = xfs_iunlink_log_inode(tp, ip, pag, next_agino);
2136 if (error)
2137 return error;
2138 ip->i_next_unlinked = next_agino;
2139 }
2140
2141 /* Point the head of the list to point to this inode. */
2142 ip->i_prev_unlinked = NULLAGINO;
2143 return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2144}
2145
2146/*
2147 * This is called when the inode's link count has gone to 0 or we are creating
2148 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2149 *
2150 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2151 * list when the inode is freed.
2152 */
2153STATIC int
2154xfs_iunlink(
2155 struct xfs_trans *tp,
2156 struct xfs_inode *ip)
2157{
2158 struct xfs_mount *mp = tp->t_mountp;
2159 struct xfs_perag *pag;
2160 struct xfs_buf *agibp;
2161 int error;
2162
2163 ASSERT(VFS_I(ip)->i_nlink == 0);
2164 ASSERT(VFS_I(ip)->i_mode != 0);
2165 trace_xfs_iunlink(ip);
2166
2167 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2168
2169 /* Get the agi buffer first. It ensures lock ordering on the list. */
2170 error = xfs_read_agi(pag, tp, &agibp);
2171 if (error)
2172 goto out;
2173
2174 error = xfs_iunlink_insert_inode(tp, pag, agibp, ip);
2175out:
2176 xfs_perag_put(pag);
2177 return error;
2178}
2179
2180static int
2181xfs_iunlink_remove_inode(
2182 struct xfs_trans *tp,
2183 struct xfs_perag *pag,
2184 struct xfs_buf *agibp,
2185 struct xfs_inode *ip)
2186{
2187 struct xfs_mount *mp = tp->t_mountp;
2188 struct xfs_agi *agi = agibp->b_addr;
2189 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2190 xfs_agino_t head_agino;
2191 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2192 int error;
2193
2194 trace_xfs_iunlink_remove(ip);
2195
2196 /*
2197 * Get the index into the agi hash table for the list this inode will
2198 * go on. Make sure the head pointer isn't garbage.
2199 */
2200 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2201 if (!xfs_verify_agino(pag, head_agino)) {
2202 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2203 agi, sizeof(*agi));
2204 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2205 return -EFSCORRUPTED;
2206 }
2207
2208 /*
2209 * Set our inode's next_unlinked pointer to NULL and then return
2210 * the old pointer value so that we can update whatever was previous
2211 * to us in the list to point to whatever was next in the list.
2212 */
2213 error = xfs_iunlink_log_inode(tp, ip, pag, NULLAGINO);
2214 if (error)
2215 return error;
2216
2217 /*
2218 * Update the prev pointer in the next inode to point back to previous
2219 * inode in the chain.
2220 */
2221 error = xfs_iunlink_update_backref(pag, ip->i_prev_unlinked,
2222 ip->i_next_unlinked);
2223 if (error == -ENOLINK)
2224 error = xfs_iunlink_reload_next(tp, agibp, ip->i_prev_unlinked,
2225 ip->i_next_unlinked);
2226 if (error)
2227 return error;
2228
2229 if (head_agino != agino) {
2230 struct xfs_inode *prev_ip;
2231
2232 prev_ip = xfs_iunlink_lookup(pag, ip->i_prev_unlinked);
2233 if (!prev_ip) {
2234 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2235 return -EFSCORRUPTED;
2236 }
2237
2238 error = xfs_iunlink_log_inode(tp, prev_ip, pag,
2239 ip->i_next_unlinked);
2240 prev_ip->i_next_unlinked = ip->i_next_unlinked;
2241 } else {
2242 /* Point the head of the list to the next unlinked inode. */
2243 error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2244 ip->i_next_unlinked);
2245 }
2246
2247 ip->i_next_unlinked = NULLAGINO;
2248 ip->i_prev_unlinked = 0;
2249 return error;
2250}
2251
2252/*
2253 * Pull the on-disk inode from the AGI unlinked list.
2254 */
2255STATIC int
2256xfs_iunlink_remove(
2257 struct xfs_trans *tp,
2258 struct xfs_perag *pag,
2259 struct xfs_inode *ip)
2260{
2261 struct xfs_buf *agibp;
2262 int error;
2263
2264 trace_xfs_iunlink_remove(ip);
2265
2266 /* Get the agi buffer first. It ensures lock ordering on the list. */
2267 error = xfs_read_agi(pag, tp, &agibp);
2268 if (error)
2269 return error;
2270
2271 return xfs_iunlink_remove_inode(tp, pag, agibp, ip);
2272}
2273
2274/*
2275 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2276 * mark it stale. We should only find clean inodes in this lookup that aren't
2277 * already stale.
2278 */
2279static void
2280xfs_ifree_mark_inode_stale(
2281 struct xfs_perag *pag,
2282 struct xfs_inode *free_ip,
2283 xfs_ino_t inum)
2284{
2285 struct xfs_mount *mp = pag->pag_mount;
2286 struct xfs_inode_log_item *iip;
2287 struct xfs_inode *ip;
2288
2289retry:
2290 rcu_read_lock();
2291 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2292
2293 /* Inode not in memory, nothing to do */
2294 if (!ip) {
2295 rcu_read_unlock();
2296 return;
2297 }
2298
2299 /*
2300 * because this is an RCU protected lookup, we could find a recently
2301 * freed or even reallocated inode during the lookup. We need to check
2302 * under the i_flags_lock for a valid inode here. Skip it if it is not
2303 * valid, the wrong inode or stale.
2304 */
2305 spin_lock(&ip->i_flags_lock);
2306 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2307 goto out_iflags_unlock;
2308
2309 /*
2310 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2311 * other inodes that we did not find in the list attached to the buffer
2312 * and are not already marked stale. If we can't lock it, back off and
2313 * retry.
2314 */
2315 if (ip != free_ip) {
2316 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2317 spin_unlock(&ip->i_flags_lock);
2318 rcu_read_unlock();
2319 delay(1);
2320 goto retry;
2321 }
2322 }
2323 ip->i_flags |= XFS_ISTALE;
2324
2325 /*
2326 * If the inode is flushing, it is already attached to the buffer. All
2327 * we needed to do here is mark the inode stale so buffer IO completion
2328 * will remove it from the AIL.
2329 */
2330 iip = ip->i_itemp;
2331 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2332 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2333 ASSERT(iip->ili_last_fields);
2334 goto out_iunlock;
2335 }
2336
2337 /*
2338 * Inodes not attached to the buffer can be released immediately.
2339 * Everything else has to go through xfs_iflush_abort() on journal
2340 * commit as the flock synchronises removal of the inode from the
2341 * cluster buffer against inode reclaim.
2342 */
2343 if (!iip || list_empty(&iip->ili_item.li_bio_list))
2344 goto out_iunlock;
2345
2346 __xfs_iflags_set(ip, XFS_IFLUSHING);
2347 spin_unlock(&ip->i_flags_lock);
2348 rcu_read_unlock();
2349
2350 /* we have a dirty inode in memory that has not yet been flushed. */
2351 spin_lock(&iip->ili_lock);
2352 iip->ili_last_fields = iip->ili_fields;
2353 iip->ili_fields = 0;
2354 iip->ili_fsync_fields = 0;
2355 spin_unlock(&iip->ili_lock);
2356 ASSERT(iip->ili_last_fields);
2357
2358 if (ip != free_ip)
2359 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2360 return;
2361
2362out_iunlock:
2363 if (ip != free_ip)
2364 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2365out_iflags_unlock:
2366 spin_unlock(&ip->i_flags_lock);
2367 rcu_read_unlock();
2368}
2369
2370/*
2371 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2372 * inodes that are in memory - they all must be marked stale and attached to
2373 * the cluster buffer.
2374 */
2375static int
2376xfs_ifree_cluster(
2377 struct xfs_trans *tp,
2378 struct xfs_perag *pag,
2379 struct xfs_inode *free_ip,
2380 struct xfs_icluster *xic)
2381{
2382 struct xfs_mount *mp = free_ip->i_mount;
2383 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2384 struct xfs_buf *bp;
2385 xfs_daddr_t blkno;
2386 xfs_ino_t inum = xic->first_ino;
2387 int nbufs;
2388 int i, j;
2389 int ioffset;
2390 int error;
2391
2392 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2393
2394 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2395 /*
2396 * The allocation bitmap tells us which inodes of the chunk were
2397 * physically allocated. Skip the cluster if an inode falls into
2398 * a sparse region.
2399 */
2400 ioffset = inum - xic->first_ino;
2401 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2402 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2403 continue;
2404 }
2405
2406 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2407 XFS_INO_TO_AGBNO(mp, inum));
2408
2409 /*
2410 * We obtain and lock the backing buffer first in the process
2411 * here to ensure dirty inodes attached to the buffer remain in
2412 * the flushing state while we mark them stale.
2413 *
2414 * If we scan the in-memory inodes first, then buffer IO can
2415 * complete before we get a lock on it, and hence we may fail
2416 * to mark all the active inodes on the buffer stale.
2417 */
2418 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2419 mp->m_bsize * igeo->blocks_per_cluster,
2420 XBF_UNMAPPED, &bp);
2421 if (error)
2422 return error;
2423
2424 /*
2425 * This buffer may not have been correctly initialised as we
2426 * didn't read it from disk. That's not important because we are
2427 * only using to mark the buffer as stale in the log, and to
2428 * attach stale cached inodes on it. That means it will never be
2429 * dispatched for IO. If it is, we want to know about it, and we
2430 * want it to fail. We can acheive this by adding a write
2431 * verifier to the buffer.
2432 */
2433 bp->b_ops = &xfs_inode_buf_ops;
2434
2435 /*
2436 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2437 * too. This requires lookups, and will skip inodes that we've
2438 * already marked XFS_ISTALE.
2439 */
2440 for (i = 0; i < igeo->inodes_per_cluster; i++)
2441 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2442
2443 xfs_trans_stale_inode_buf(tp, bp);
2444 xfs_trans_binval(tp, bp);
2445 }
2446 return 0;
2447}
2448
2449/*
2450 * This is called to return an inode to the inode free list. The inode should
2451 * already be truncated to 0 length and have no pages associated with it. This
2452 * routine also assumes that the inode is already a part of the transaction.
2453 *
2454 * The on-disk copy of the inode will have been added to the list of unlinked
2455 * inodes in the AGI. We need to remove the inode from that list atomically with
2456 * respect to freeing it here.
2457 */
2458int
2459xfs_ifree(
2460 struct xfs_trans *tp,
2461 struct xfs_inode *ip)
2462{
2463 struct xfs_mount *mp = ip->i_mount;
2464 struct xfs_perag *pag;
2465 struct xfs_icluster xic = { 0 };
2466 struct xfs_inode_log_item *iip = ip->i_itemp;
2467 int error;
2468
2469 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
2470 ASSERT(VFS_I(ip)->i_nlink == 0);
2471 ASSERT(ip->i_df.if_nextents == 0);
2472 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2473 ASSERT(ip->i_nblocks == 0);
2474
2475 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2476
2477 /*
2478 * Free the inode first so that we guarantee that the AGI lock is going
2479 * to be taken before we remove the inode from the unlinked list. This
2480 * makes the AGI lock -> unlinked list modification order the same as
2481 * used in O_TMPFILE creation.
2482 */
2483 error = xfs_difree(tp, pag, ip->i_ino, &xic);
2484 if (error)
2485 goto out;
2486
2487 error = xfs_iunlink_remove(tp, pag, ip);
2488 if (error)
2489 goto out;
2490
2491 /*
2492 * Free any local-format data sitting around before we reset the
2493 * data fork to extents format. Note that the attr fork data has
2494 * already been freed by xfs_attr_inactive.
2495 */
2496 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2497 kfree(ip->i_df.if_data);
2498 ip->i_df.if_data = NULL;
2499 ip->i_df.if_bytes = 0;
2500 }
2501
2502 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2503 ip->i_diflags = 0;
2504 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2505 ip->i_forkoff = 0; /* mark the attr fork not in use */
2506 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2507 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2508 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2509
2510 /* Don't attempt to replay owner changes for a deleted inode */
2511 spin_lock(&iip->ili_lock);
2512 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2513 spin_unlock(&iip->ili_lock);
2514
2515 /*
2516 * Bump the generation count so no one will be confused
2517 * by reincarnations of this inode.
2518 */
2519 VFS_I(ip)->i_generation++;
2520 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2521
2522 if (xic.deleted)
2523 error = xfs_ifree_cluster(tp, pag, ip, &xic);
2524out:
2525 xfs_perag_put(pag);
2526 return error;
2527}
2528
2529/*
2530 * This is called to unpin an inode. The caller must have the inode locked
2531 * in at least shared mode so that the buffer cannot be subsequently pinned
2532 * once someone is waiting for it to be unpinned.
2533 */
2534static void
2535xfs_iunpin(
2536 struct xfs_inode *ip)
2537{
2538 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2539
2540 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2541
2542 /* Give the log a push to start the unpinning I/O */
2543 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2544
2545}
2546
2547static void
2548__xfs_iunpin_wait(
2549 struct xfs_inode *ip)
2550{
2551 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2552 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2553
2554 xfs_iunpin(ip);
2555
2556 do {
2557 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2558 if (xfs_ipincount(ip))
2559 io_schedule();
2560 } while (xfs_ipincount(ip));
2561 finish_wait(wq, &wait.wq_entry);
2562}
2563
2564void
2565xfs_iunpin_wait(
2566 struct xfs_inode *ip)
2567{
2568 if (xfs_ipincount(ip))
2569 __xfs_iunpin_wait(ip);
2570}
2571
2572/*
2573 * Removing an inode from the namespace involves removing the directory entry
2574 * and dropping the link count on the inode. Removing the directory entry can
2575 * result in locking an AGF (directory blocks were freed) and removing a link
2576 * count can result in placing the inode on an unlinked list which results in
2577 * locking an AGI.
2578 *
2579 * The big problem here is that we have an ordering constraint on AGF and AGI
2580 * locking - inode allocation locks the AGI, then can allocate a new extent for
2581 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2582 * removes the inode from the unlinked list, requiring that we lock the AGI
2583 * first, and then freeing the inode can result in an inode chunk being freed
2584 * and hence freeing disk space requiring that we lock an AGF.
2585 *
2586 * Hence the ordering that is imposed by other parts of the code is AGI before
2587 * AGF. This means we cannot remove the directory entry before we drop the inode
2588 * reference count and put it on the unlinked list as this results in a lock
2589 * order of AGF then AGI, and this can deadlock against inode allocation and
2590 * freeing. Therefore we must drop the link counts before we remove the
2591 * directory entry.
2592 *
2593 * This is still safe from a transactional point of view - it is not until we
2594 * get to xfs_defer_finish() that we have the possibility of multiple
2595 * transactions in this operation. Hence as long as we remove the directory
2596 * entry and drop the link count in the first transaction of the remove
2597 * operation, there are no transactional constraints on the ordering here.
2598 */
2599int
2600xfs_remove(
2601 xfs_inode_t *dp,
2602 struct xfs_name *name,
2603 xfs_inode_t *ip)
2604{
2605 xfs_mount_t *mp = dp->i_mount;
2606 xfs_trans_t *tp = NULL;
2607 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2608 int dontcare;
2609 int error = 0;
2610 uint resblks;
2611
2612 trace_xfs_remove(dp, name);
2613
2614 if (xfs_is_shutdown(mp))
2615 return -EIO;
2616 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
2617 return -EIO;
2618
2619 error = xfs_qm_dqattach(dp);
2620 if (error)
2621 goto std_return;
2622
2623 error = xfs_qm_dqattach(ip);
2624 if (error)
2625 goto std_return;
2626
2627 /*
2628 * We try to get the real space reservation first, allowing for
2629 * directory btree deletion(s) implying possible bmap insert(s). If we
2630 * can't get the space reservation then we use 0 instead, and avoid the
2631 * bmap btree insert(s) in the directory code by, if the bmap insert
2632 * tries to happen, instead trimming the LAST block from the directory.
2633 *
2634 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
2635 * the directory code can handle a reservationless update and we don't
2636 * want to prevent a user from trying to free space by deleting things.
2637 */
2638 resblks = XFS_REMOVE_SPACE_RES(mp);
2639 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2640 &tp, &dontcare);
2641 if (error) {
2642 ASSERT(error != -ENOSPC);
2643 goto std_return;
2644 }
2645
2646 /*
2647 * If we're removing a directory perform some additional validation.
2648 */
2649 if (is_dir) {
2650 ASSERT(VFS_I(ip)->i_nlink >= 2);
2651 if (VFS_I(ip)->i_nlink != 2) {
2652 error = -ENOTEMPTY;
2653 goto out_trans_cancel;
2654 }
2655 if (!xfs_dir_isempty(ip)) {
2656 error = -ENOTEMPTY;
2657 goto out_trans_cancel;
2658 }
2659
2660 /* Drop the link from ip's "..". */
2661 error = xfs_droplink(tp, dp);
2662 if (error)
2663 goto out_trans_cancel;
2664
2665 /* Drop the "." link from ip to self. */
2666 error = xfs_droplink(tp, ip);
2667 if (error)
2668 goto out_trans_cancel;
2669
2670 /*
2671 * Point the unlinked child directory's ".." entry to the root
2672 * directory to eliminate back-references to inodes that may
2673 * get freed before the child directory is closed. If the fs
2674 * gets shrunk, this can lead to dirent inode validation errors.
2675 */
2676 if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2677 error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2678 tp->t_mountp->m_sb.sb_rootino, 0);
2679 if (error)
2680 goto out_trans_cancel;
2681 }
2682 } else {
2683 /*
2684 * When removing a non-directory we need to log the parent
2685 * inode here. For a directory this is done implicitly
2686 * by the xfs_droplink call for the ".." entry.
2687 */
2688 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2689 }
2690 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2691
2692 /* Drop the link from dp to ip. */
2693 error = xfs_droplink(tp, ip);
2694 if (error)
2695 goto out_trans_cancel;
2696
2697 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2698 if (error) {
2699 ASSERT(error != -ENOENT);
2700 goto out_trans_cancel;
2701 }
2702
2703 /*
2704 * Drop the link from dp to ip, and if ip was a directory, remove the
2705 * '.' and '..' references since we freed the directory.
2706 */
2707 xfs_dir_update_hook(dp, ip, -1, name);
2708
2709 /*
2710 * If this is a synchronous mount, make sure that the
2711 * remove transaction goes to disk before returning to
2712 * the user.
2713 */
2714 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2715 xfs_trans_set_sync(tp);
2716
2717 error = xfs_trans_commit(tp);
2718 if (error)
2719 goto std_return;
2720
2721 if (is_dir && xfs_inode_is_filestream(ip))
2722 xfs_filestream_deassociate(ip);
2723
2724 return 0;
2725
2726 out_trans_cancel:
2727 xfs_trans_cancel(tp);
2728 std_return:
2729 return error;
2730}
2731
2732/*
2733 * Enter all inodes for a rename transaction into a sorted array.
2734 */
2735#define __XFS_SORT_INODES 5
2736STATIC void
2737xfs_sort_for_rename(
2738 struct xfs_inode *dp1, /* in: old (source) directory inode */
2739 struct xfs_inode *dp2, /* in: new (target) directory inode */
2740 struct xfs_inode *ip1, /* in: inode of old entry */
2741 struct xfs_inode *ip2, /* in: inode of new entry */
2742 struct xfs_inode *wip, /* in: whiteout inode */
2743 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2744 int *num_inodes) /* in/out: inodes in array */
2745{
2746 int i, j;
2747
2748 ASSERT(*num_inodes == __XFS_SORT_INODES);
2749 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2750
2751 /*
2752 * i_tab contains a list of pointers to inodes. We initialize
2753 * the table here & we'll sort it. We will then use it to
2754 * order the acquisition of the inode locks.
2755 *
2756 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2757 */
2758 i = 0;
2759 i_tab[i++] = dp1;
2760 i_tab[i++] = dp2;
2761 i_tab[i++] = ip1;
2762 if (ip2)
2763 i_tab[i++] = ip2;
2764 if (wip)
2765 i_tab[i++] = wip;
2766 *num_inodes = i;
2767
2768 /*
2769 * Sort the elements via bubble sort. (Remember, there are at
2770 * most 5 elements to sort, so this is adequate.)
2771 */
2772 for (i = 0; i < *num_inodes; i++) {
2773 for (j = 1; j < *num_inodes; j++) {
2774 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2775 struct xfs_inode *temp = i_tab[j];
2776 i_tab[j] = i_tab[j-1];
2777 i_tab[j-1] = temp;
2778 }
2779 }
2780 }
2781}
2782
2783static int
2784xfs_finish_rename(
2785 struct xfs_trans *tp)
2786{
2787 /*
2788 * If this is a synchronous mount, make sure that the rename transaction
2789 * goes to disk before returning to the user.
2790 */
2791 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2792 xfs_trans_set_sync(tp);
2793
2794 return xfs_trans_commit(tp);
2795}
2796
2797/*
2798 * xfs_cross_rename()
2799 *
2800 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2801 */
2802STATIC int
2803xfs_cross_rename(
2804 struct xfs_trans *tp,
2805 struct xfs_inode *dp1,
2806 struct xfs_name *name1,
2807 struct xfs_inode *ip1,
2808 struct xfs_inode *dp2,
2809 struct xfs_name *name2,
2810 struct xfs_inode *ip2,
2811 int spaceres)
2812{
2813 int error = 0;
2814 int ip1_flags = 0;
2815 int ip2_flags = 0;
2816 int dp2_flags = 0;
2817
2818 /* Swap inode number for dirent in first parent */
2819 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2820 if (error)
2821 goto out_trans_abort;
2822
2823 /* Swap inode number for dirent in second parent */
2824 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2825 if (error)
2826 goto out_trans_abort;
2827
2828 /*
2829 * If we're renaming one or more directories across different parents,
2830 * update the respective ".." entries (and link counts) to match the new
2831 * parents.
2832 */
2833 if (dp1 != dp2) {
2834 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2835
2836 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2837 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2838 dp1->i_ino, spaceres);
2839 if (error)
2840 goto out_trans_abort;
2841
2842 /* transfer ip2 ".." reference to dp1 */
2843 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2844 error = xfs_droplink(tp, dp2);
2845 if (error)
2846 goto out_trans_abort;
2847 xfs_bumplink(tp, dp1);
2848 }
2849
2850 /*
2851 * Although ip1 isn't changed here, userspace needs
2852 * to be warned about the change, so that applications
2853 * relying on it (like backup ones), will properly
2854 * notify the change
2855 */
2856 ip1_flags |= XFS_ICHGTIME_CHG;
2857 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2858 }
2859
2860 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2861 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2862 dp2->i_ino, spaceres);
2863 if (error)
2864 goto out_trans_abort;
2865
2866 /* transfer ip1 ".." reference to dp2 */
2867 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2868 error = xfs_droplink(tp, dp1);
2869 if (error)
2870 goto out_trans_abort;
2871 xfs_bumplink(tp, dp2);
2872 }
2873
2874 /*
2875 * Although ip2 isn't changed here, userspace needs
2876 * to be warned about the change, so that applications
2877 * relying on it (like backup ones), will properly
2878 * notify the change
2879 */
2880 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2881 ip2_flags |= XFS_ICHGTIME_CHG;
2882 }
2883 }
2884
2885 if (ip1_flags) {
2886 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2887 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2888 }
2889 if (ip2_flags) {
2890 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2891 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2892 }
2893 if (dp2_flags) {
2894 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2895 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2896 }
2897 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2898 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2899
2900 /*
2901 * Inform our hook clients that we've finished an exchange operation as
2902 * follows: removed the source and target files from their directories;
2903 * added the target to the source directory; and added the source to
2904 * the target directory. All inodes are locked, so it's ok to model a
2905 * rename this way so long as we say we deleted entries before we add
2906 * new ones.
2907 */
2908 xfs_dir_update_hook(dp1, ip1, -1, name1);
2909 xfs_dir_update_hook(dp2, ip2, -1, name2);
2910 xfs_dir_update_hook(dp1, ip2, 1, name1);
2911 xfs_dir_update_hook(dp2, ip1, 1, name2);
2912
2913 return xfs_finish_rename(tp);
2914
2915out_trans_abort:
2916 xfs_trans_cancel(tp);
2917 return error;
2918}
2919
2920/*
2921 * xfs_rename_alloc_whiteout()
2922 *
2923 * Return a referenced, unlinked, unlocked inode that can be used as a
2924 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2925 * crash between allocating the inode and linking it into the rename transaction
2926 * recovery will free the inode and we won't leak it.
2927 */
2928static int
2929xfs_rename_alloc_whiteout(
2930 struct mnt_idmap *idmap,
2931 struct xfs_name *src_name,
2932 struct xfs_inode *dp,
2933 struct xfs_inode **wip)
2934{
2935 struct xfs_inode *tmpfile;
2936 struct qstr name;
2937 int error;
2938
2939 error = xfs_create_tmpfile(idmap, dp, S_IFCHR | WHITEOUT_MODE,
2940 &tmpfile);
2941 if (error)
2942 return error;
2943
2944 name.name = src_name->name;
2945 name.len = src_name->len;
2946 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2947 if (error) {
2948 xfs_finish_inode_setup(tmpfile);
2949 xfs_irele(tmpfile);
2950 return error;
2951 }
2952
2953 /*
2954 * Prepare the tmpfile inode as if it were created through the VFS.
2955 * Complete the inode setup and flag it as linkable. nlink is already
2956 * zero, so we can skip the drop_nlink.
2957 */
2958 xfs_setup_iops(tmpfile);
2959 xfs_finish_inode_setup(tmpfile);
2960 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2961
2962 *wip = tmpfile;
2963 return 0;
2964}
2965
2966/*
2967 * xfs_rename
2968 */
2969int
2970xfs_rename(
2971 struct mnt_idmap *idmap,
2972 struct xfs_inode *src_dp,
2973 struct xfs_name *src_name,
2974 struct xfs_inode *src_ip,
2975 struct xfs_inode *target_dp,
2976 struct xfs_name *target_name,
2977 struct xfs_inode *target_ip,
2978 unsigned int flags)
2979{
2980 struct xfs_mount *mp = src_dp->i_mount;
2981 struct xfs_trans *tp;
2982 struct xfs_inode *wip = NULL; /* whiteout inode */
2983 struct xfs_inode *inodes[__XFS_SORT_INODES];
2984 int i;
2985 int num_inodes = __XFS_SORT_INODES;
2986 bool new_parent = (src_dp != target_dp);
2987 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2988 int spaceres;
2989 bool retried = false;
2990 int error, nospace_error = 0;
2991
2992 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2993
2994 if ((flags & RENAME_EXCHANGE) && !target_ip)
2995 return -EINVAL;
2996
2997 /*
2998 * If we are doing a whiteout operation, allocate the whiteout inode
2999 * we will be placing at the target and ensure the type is set
3000 * appropriately.
3001 */
3002 if (flags & RENAME_WHITEOUT) {
3003 error = xfs_rename_alloc_whiteout(idmap, src_name,
3004 target_dp, &wip);
3005 if (error)
3006 return error;
3007
3008 /* setup target dirent info as whiteout */
3009 src_name->type = XFS_DIR3_FT_CHRDEV;
3010 }
3011
3012 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3013 inodes, &num_inodes);
3014
3015retry:
3016 nospace_error = 0;
3017 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3018 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3019 if (error == -ENOSPC) {
3020 nospace_error = error;
3021 spaceres = 0;
3022 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3023 &tp);
3024 }
3025 if (error)
3026 goto out_release_wip;
3027
3028 /*
3029 * Attach the dquots to the inodes
3030 */
3031 error = xfs_qm_vop_rename_dqattach(inodes);
3032 if (error)
3033 goto out_trans_cancel;
3034
3035 /*
3036 * Lock all the participating inodes. Depending upon whether
3037 * the target_name exists in the target directory, and
3038 * whether the target directory is the same as the source
3039 * directory, we can lock from 2 to 5 inodes.
3040 */
3041 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3042
3043 /*
3044 * Join all the inodes to the transaction. From this point on,
3045 * we can rely on either trans_commit or trans_cancel to unlock
3046 * them.
3047 */
3048 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3049 if (new_parent)
3050 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3051 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3052 if (target_ip)
3053 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3054 if (wip)
3055 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3056
3057 /*
3058 * If we are using project inheritance, we only allow renames
3059 * into our tree when the project IDs are the same; else the
3060 * tree quota mechanism would be circumvented.
3061 */
3062 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
3063 target_dp->i_projid != src_ip->i_projid)) {
3064 error = -EXDEV;
3065 goto out_trans_cancel;
3066 }
3067
3068 /* RENAME_EXCHANGE is unique from here on. */
3069 if (flags & RENAME_EXCHANGE)
3070 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3071 target_dp, target_name, target_ip,
3072 spaceres);
3073
3074 /*
3075 * Try to reserve quota to handle an expansion of the target directory.
3076 * We'll allow the rename to continue in reservationless mode if we hit
3077 * a space usage constraint. If we trigger reservationless mode, save
3078 * the errno if there isn't any free space in the target directory.
3079 */
3080 if (spaceres != 0) {
3081 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
3082 0, false);
3083 if (error == -EDQUOT || error == -ENOSPC) {
3084 if (!retried) {
3085 xfs_trans_cancel(tp);
3086 xfs_blockgc_free_quota(target_dp, 0);
3087 retried = true;
3088 goto retry;
3089 }
3090
3091 nospace_error = error;
3092 spaceres = 0;
3093 error = 0;
3094 }
3095 if (error)
3096 goto out_trans_cancel;
3097 }
3098
3099 /*
3100 * Check for expected errors before we dirty the transaction
3101 * so we can return an error without a transaction abort.
3102 */
3103 if (target_ip == NULL) {
3104 /*
3105 * If there's no space reservation, check the entry will
3106 * fit before actually inserting it.
3107 */
3108 if (!spaceres) {
3109 error = xfs_dir_canenter(tp, target_dp, target_name);
3110 if (error)
3111 goto out_trans_cancel;
3112 }
3113 } else {
3114 /*
3115 * If target exists and it's a directory, check that whether
3116 * it can be destroyed.
3117 */
3118 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3119 (!xfs_dir_isempty(target_ip) ||
3120 (VFS_I(target_ip)->i_nlink > 2))) {
3121 error = -EEXIST;
3122 goto out_trans_cancel;
3123 }
3124 }
3125
3126 /*
3127 * Lock the AGI buffers we need to handle bumping the nlink of the
3128 * whiteout inode off the unlinked list and to handle dropping the
3129 * nlink of the target inode. Per locking order rules, do this in
3130 * increasing AG order and before directory block allocation tries to
3131 * grab AGFs because we grab AGIs before AGFs.
3132 *
3133 * The (vfs) caller must ensure that if src is a directory then
3134 * target_ip is either null or an empty directory.
3135 */
3136 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3137 if (inodes[i] == wip ||
3138 (inodes[i] == target_ip &&
3139 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3140 struct xfs_perag *pag;
3141 struct xfs_buf *bp;
3142
3143 pag = xfs_perag_get(mp,
3144 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
3145 error = xfs_read_agi(pag, tp, &bp);
3146 xfs_perag_put(pag);
3147 if (error)
3148 goto out_trans_cancel;
3149 }
3150 }
3151
3152 /*
3153 * Directory entry creation below may acquire the AGF. Remove
3154 * the whiteout from the unlinked list first to preserve correct
3155 * AGI/AGF locking order. This dirties the transaction so failures
3156 * after this point will abort and log recovery will clean up the
3157 * mess.
3158 *
3159 * For whiteouts, we need to bump the link count on the whiteout
3160 * inode. After this point, we have a real link, clear the tmpfile
3161 * state flag from the inode so it doesn't accidentally get misused
3162 * in future.
3163 */
3164 if (wip) {
3165 struct xfs_perag *pag;
3166
3167 ASSERT(VFS_I(wip)->i_nlink == 0);
3168
3169 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3170 error = xfs_iunlink_remove(tp, pag, wip);
3171 xfs_perag_put(pag);
3172 if (error)
3173 goto out_trans_cancel;
3174
3175 xfs_bumplink(tp, wip);
3176 VFS_I(wip)->i_state &= ~I_LINKABLE;
3177 }
3178
3179 /*
3180 * Set up the target.
3181 */
3182 if (target_ip == NULL) {
3183 /*
3184 * If target does not exist and the rename crosses
3185 * directories, adjust the target directory link count
3186 * to account for the ".." reference from the new entry.
3187 */
3188 error = xfs_dir_createname(tp, target_dp, target_name,
3189 src_ip->i_ino, spaceres);
3190 if (error)
3191 goto out_trans_cancel;
3192
3193 xfs_trans_ichgtime(tp, target_dp,
3194 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3195
3196 if (new_parent && src_is_directory) {
3197 xfs_bumplink(tp, target_dp);
3198 }
3199 } else { /* target_ip != NULL */
3200 /*
3201 * Link the source inode under the target name.
3202 * If the source inode is a directory and we are moving
3203 * it across directories, its ".." entry will be
3204 * inconsistent until we replace that down below.
3205 *
3206 * In case there is already an entry with the same
3207 * name at the destination directory, remove it first.
3208 */
3209 error = xfs_dir_replace(tp, target_dp, target_name,
3210 src_ip->i_ino, spaceres);
3211 if (error)
3212 goto out_trans_cancel;
3213
3214 xfs_trans_ichgtime(tp, target_dp,
3215 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3216
3217 /*
3218 * Decrement the link count on the target since the target
3219 * dir no longer points to it.
3220 */
3221 error = xfs_droplink(tp, target_ip);
3222 if (error)
3223 goto out_trans_cancel;
3224
3225 if (src_is_directory) {
3226 /*
3227 * Drop the link from the old "." entry.
3228 */
3229 error = xfs_droplink(tp, target_ip);
3230 if (error)
3231 goto out_trans_cancel;
3232 }
3233 } /* target_ip != NULL */
3234
3235 /*
3236 * Remove the source.
3237 */
3238 if (new_parent && src_is_directory) {
3239 /*
3240 * Rewrite the ".." entry to point to the new
3241 * directory.
3242 */
3243 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3244 target_dp->i_ino, spaceres);
3245 ASSERT(error != -EEXIST);
3246 if (error)
3247 goto out_trans_cancel;
3248 }
3249
3250 /*
3251 * We always want to hit the ctime on the source inode.
3252 *
3253 * This isn't strictly required by the standards since the source
3254 * inode isn't really being changed, but old unix file systems did
3255 * it and some incremental backup programs won't work without it.
3256 */
3257 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3258 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3259
3260 /*
3261 * Adjust the link count on src_dp. This is necessary when
3262 * renaming a directory, either within one parent when
3263 * the target existed, or across two parent directories.
3264 */
3265 if (src_is_directory && (new_parent || target_ip != NULL)) {
3266
3267 /*
3268 * Decrement link count on src_directory since the
3269 * entry that's moved no longer points to it.
3270 */
3271 error = xfs_droplink(tp, src_dp);
3272 if (error)
3273 goto out_trans_cancel;
3274 }
3275
3276 /*
3277 * For whiteouts, we only need to update the source dirent with the
3278 * inode number of the whiteout inode rather than removing it
3279 * altogether.
3280 */
3281 if (wip)
3282 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3283 spaceres);
3284 else
3285 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3286 spaceres);
3287
3288 if (error)
3289 goto out_trans_cancel;
3290
3291 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3292 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3293 if (new_parent)
3294 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3295
3296 /*
3297 * Inform our hook clients that we've finished a rename operation as
3298 * follows: removed the source and target files from their directories;
3299 * that we've added the source to the target directory; and finally
3300 * that we've added the whiteout, if there was one. All inodes are
3301 * locked, so it's ok to model a rename this way so long as we say we
3302 * deleted entries before we add new ones.
3303 */
3304 if (target_ip)
3305 xfs_dir_update_hook(target_dp, target_ip, -1, target_name);
3306 xfs_dir_update_hook(src_dp, src_ip, -1, src_name);
3307 xfs_dir_update_hook(target_dp, src_ip, 1, target_name);
3308 if (wip)
3309 xfs_dir_update_hook(src_dp, wip, 1, src_name);
3310
3311 error = xfs_finish_rename(tp);
3312 if (wip)
3313 xfs_irele(wip);
3314 return error;
3315
3316out_trans_cancel:
3317 xfs_trans_cancel(tp);
3318out_release_wip:
3319 if (wip)
3320 xfs_irele(wip);
3321 if (error == -ENOSPC && nospace_error)
3322 error = nospace_error;
3323 return error;
3324}
3325
3326static int
3327xfs_iflush(
3328 struct xfs_inode *ip,
3329 struct xfs_buf *bp)
3330{
3331 struct xfs_inode_log_item *iip = ip->i_itemp;
3332 struct xfs_dinode *dip;
3333 struct xfs_mount *mp = ip->i_mount;
3334 int error;
3335
3336 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
3337 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3338 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3339 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3340 ASSERT(iip->ili_item.li_buf == bp);
3341
3342 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3343
3344 /*
3345 * We don't flush the inode if any of the following checks fail, but we
3346 * do still update the log item and attach to the backing buffer as if
3347 * the flush happened. This is a formality to facilitate predictable
3348 * error handling as the caller will shutdown and fail the buffer.
3349 */
3350 error = -EFSCORRUPTED;
3351 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3352 mp, XFS_ERRTAG_IFLUSH_1)) {
3353 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3354 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
3355 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3356 goto flush_out;
3357 }
3358 if (S_ISREG(VFS_I(ip)->i_mode)) {
3359 if (XFS_TEST_ERROR(
3360 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3361 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3362 mp, XFS_ERRTAG_IFLUSH_3)) {
3363 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3364 "%s: Bad regular inode %llu, ptr "PTR_FMT,
3365 __func__, ip->i_ino, ip);
3366 goto flush_out;
3367 }
3368 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3369 if (XFS_TEST_ERROR(
3370 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3371 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3372 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3373 mp, XFS_ERRTAG_IFLUSH_4)) {
3374 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3375 "%s: Bad directory inode %llu, ptr "PTR_FMT,
3376 __func__, ip->i_ino, ip);
3377 goto flush_out;
3378 }
3379 }
3380 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
3381 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3382 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3383 "%s: detected corrupt incore inode %llu, "
3384 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
3385 __func__, ip->i_ino,
3386 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
3387 ip->i_nblocks, ip);
3388 goto flush_out;
3389 }
3390 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3391 mp, XFS_ERRTAG_IFLUSH_6)) {
3392 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3393 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
3394 __func__, ip->i_ino, ip->i_forkoff, ip);
3395 goto flush_out;
3396 }
3397
3398 /*
3399 * Inode item log recovery for v2 inodes are dependent on the flushiter
3400 * count for correct sequencing. We bump the flush iteration count so
3401 * we can detect flushes which postdate a log record during recovery.
3402 * This is redundant as we now log every change and hence this can't
3403 * happen but we need to still do it to ensure backwards compatibility
3404 * with old kernels that predate logging all inode changes.
3405 */
3406 if (!xfs_has_v3inodes(mp))
3407 ip->i_flushiter++;
3408
3409 /*
3410 * If there are inline format data / attr forks attached to this inode,
3411 * make sure they are not corrupt.
3412 */
3413 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3414 xfs_ifork_verify_local_data(ip))
3415 goto flush_out;
3416 if (xfs_inode_has_attr_fork(ip) &&
3417 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
3418 xfs_ifork_verify_local_attr(ip))
3419 goto flush_out;
3420
3421 /*
3422 * Copy the dirty parts of the inode into the on-disk inode. We always
3423 * copy out the core of the inode, because if the inode is dirty at all
3424 * the core must be.
3425 */
3426 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3427
3428 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3429 if (!xfs_has_v3inodes(mp)) {
3430 if (ip->i_flushiter == DI_MAX_FLUSH)
3431 ip->i_flushiter = 0;
3432 }
3433
3434 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3435 if (xfs_inode_has_attr_fork(ip))
3436 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3437
3438 /*
3439 * We've recorded everything logged in the inode, so we'd like to clear
3440 * the ili_fields bits so we don't log and flush things unnecessarily.
3441 * However, we can't stop logging all this information until the data
3442 * we've copied into the disk buffer is written to disk. If we did we
3443 * might overwrite the copy of the inode in the log with all the data
3444 * after re-logging only part of it, and in the face of a crash we
3445 * wouldn't have all the data we need to recover.
3446 *
3447 * What we do is move the bits to the ili_last_fields field. When
3448 * logging the inode, these bits are moved back to the ili_fields field.
3449 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3450 * we know that the information those bits represent is permanently on
3451 * disk. As long as the flush completes before the inode is logged
3452 * again, then both ili_fields and ili_last_fields will be cleared.
3453 */
3454 error = 0;
3455flush_out:
3456 spin_lock(&iip->ili_lock);
3457 iip->ili_last_fields = iip->ili_fields;
3458 iip->ili_fields = 0;
3459 iip->ili_fsync_fields = 0;
3460 spin_unlock(&iip->ili_lock);
3461
3462 /*
3463 * Store the current LSN of the inode so that we can tell whether the
3464 * item has moved in the AIL from xfs_buf_inode_iodone().
3465 */
3466 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3467 &iip->ili_item.li_lsn);
3468
3469 /* generate the checksum. */
3470 xfs_dinode_calc_crc(mp, dip);
3471 if (error)
3472 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
3473 return error;
3474}
3475
3476/*
3477 * Non-blocking flush of dirty inode metadata into the backing buffer.
3478 *
3479 * The caller must have a reference to the inode and hold the cluster buffer
3480 * locked. The function will walk across all the inodes on the cluster buffer it
3481 * can find and lock without blocking, and flush them to the cluster buffer.
3482 *
3483 * On successful flushing of at least one inode, the caller must write out the
3484 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3485 * the caller needs to release the buffer. On failure, the filesystem will be
3486 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3487 * will be returned.
3488 */
3489int
3490xfs_iflush_cluster(
3491 struct xfs_buf *bp)
3492{
3493 struct xfs_mount *mp = bp->b_mount;
3494 struct xfs_log_item *lip, *n;
3495 struct xfs_inode *ip;
3496 struct xfs_inode_log_item *iip;
3497 int clcount = 0;
3498 int error = 0;
3499
3500 /*
3501 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3502 * will remove itself from the list.
3503 */
3504 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3505 iip = (struct xfs_inode_log_item *)lip;
3506 ip = iip->ili_inode;
3507
3508 /*
3509 * Quick and dirty check to avoid locks if possible.
3510 */
3511 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3512 continue;
3513 if (xfs_ipincount(ip))
3514 continue;
3515
3516 /*
3517 * The inode is still attached to the buffer, which means it is
3518 * dirty but reclaim might try to grab it. Check carefully for
3519 * that, and grab the ilock while still holding the i_flags_lock
3520 * to guarantee reclaim will not be able to reclaim this inode
3521 * once we drop the i_flags_lock.
3522 */
3523 spin_lock(&ip->i_flags_lock);
3524 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3525 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3526 spin_unlock(&ip->i_flags_lock);
3527 continue;
3528 }
3529
3530 /*
3531 * ILOCK will pin the inode against reclaim and prevent
3532 * concurrent transactions modifying the inode while we are
3533 * flushing the inode. If we get the lock, set the flushing
3534 * state before we drop the i_flags_lock.
3535 */
3536 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3537 spin_unlock(&ip->i_flags_lock);
3538 continue;
3539 }
3540 __xfs_iflags_set(ip, XFS_IFLUSHING);
3541 spin_unlock(&ip->i_flags_lock);
3542
3543 /*
3544 * Abort flushing this inode if we are shut down because the
3545 * inode may not currently be in the AIL. This can occur when
3546 * log I/O failure unpins the inode without inserting into the
3547 * AIL, leaving a dirty/unpinned inode attached to the buffer
3548 * that otherwise looks like it should be flushed.
3549 */
3550 if (xlog_is_shutdown(mp->m_log)) {
3551 xfs_iunpin_wait(ip);
3552 xfs_iflush_abort(ip);
3553 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3554 error = -EIO;
3555 continue;
3556 }
3557
3558 /* don't block waiting on a log force to unpin dirty inodes */
3559 if (xfs_ipincount(ip)) {
3560 xfs_iflags_clear(ip, XFS_IFLUSHING);
3561 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3562 continue;
3563 }
3564
3565 if (!xfs_inode_clean(ip))
3566 error = xfs_iflush(ip, bp);
3567 else
3568 xfs_iflags_clear(ip, XFS_IFLUSHING);
3569 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3570 if (error)
3571 break;
3572 clcount++;
3573 }
3574
3575 if (error) {
3576 /*
3577 * Shutdown first so we kill the log before we release this
3578 * buffer. If it is an INODE_ALLOC buffer and pins the tail
3579 * of the log, failing it before the _log_ is shut down can
3580 * result in the log tail being moved forward in the journal
3581 * on disk because log writes can still be taking place. Hence
3582 * unpinning the tail will allow the ICREATE intent to be
3583 * removed from the log an recovery will fail with uninitialised
3584 * inode cluster buffers.
3585 */
3586 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3587 bp->b_flags |= XBF_ASYNC;
3588 xfs_buf_ioend_fail(bp);
3589 return error;
3590 }
3591
3592 if (!clcount)
3593 return -EAGAIN;
3594
3595 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3596 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3597 return 0;
3598
3599}
3600
3601/* Release an inode. */
3602void
3603xfs_irele(
3604 struct xfs_inode *ip)
3605{
3606 trace_xfs_irele(ip, _RET_IP_);
3607 iput(VFS_I(ip));
3608}
3609
3610/*
3611 * Ensure all commited transactions touching the inode are written to the log.
3612 */
3613int
3614xfs_log_force_inode(
3615 struct xfs_inode *ip)
3616{
3617 xfs_csn_t seq = 0;
3618
3619 xfs_ilock(ip, XFS_ILOCK_SHARED);
3620 if (xfs_ipincount(ip))
3621 seq = ip->i_itemp->ili_commit_seq;
3622 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3623
3624 if (!seq)
3625 return 0;
3626 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3627}
3628
3629/*
3630 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3631 * abide vfs locking order (lowest pointer value goes first) and breaking the
3632 * layout leases before proceeding. The loop is needed because we cannot call
3633 * the blocking break_layout() with the iolocks held, and therefore have to
3634 * back out both locks.
3635 */
3636static int
3637xfs_iolock_two_inodes_and_break_layout(
3638 struct inode *src,
3639 struct inode *dest)
3640{
3641 int error;
3642
3643 if (src > dest)
3644 swap(src, dest);
3645
3646retry:
3647 /* Wait to break both inodes' layouts before we start locking. */
3648 error = break_layout(src, true);
3649 if (error)
3650 return error;
3651 if (src != dest) {
3652 error = break_layout(dest, true);
3653 if (error)
3654 return error;
3655 }
3656
3657 /* Lock one inode and make sure nobody got in and leased it. */
3658 inode_lock(src);
3659 error = break_layout(src, false);
3660 if (error) {
3661 inode_unlock(src);
3662 if (error == -EWOULDBLOCK)
3663 goto retry;
3664 return error;
3665 }
3666
3667 if (src == dest)
3668 return 0;
3669
3670 /* Lock the other inode and make sure nobody got in and leased it. */
3671 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3672 error = break_layout(dest, false);
3673 if (error) {
3674 inode_unlock(src);
3675 inode_unlock(dest);
3676 if (error == -EWOULDBLOCK)
3677 goto retry;
3678 return error;
3679 }
3680
3681 return 0;
3682}
3683
3684static int
3685xfs_mmaplock_two_inodes_and_break_dax_layout(
3686 struct xfs_inode *ip1,
3687 struct xfs_inode *ip2)
3688{
3689 int error;
3690 bool retry;
3691 struct page *page;
3692
3693 if (ip1->i_ino > ip2->i_ino)
3694 swap(ip1, ip2);
3695
3696again:
3697 retry = false;
3698 /* Lock the first inode */
3699 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3700 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
3701 if (error || retry) {
3702 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3703 if (error == 0 && retry)
3704 goto again;
3705 return error;
3706 }
3707
3708 if (ip1 == ip2)
3709 return 0;
3710
3711 /* Nested lock the second inode */
3712 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
3713 /*
3714 * We cannot use xfs_break_dax_layouts() directly here because it may
3715 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
3716 * for this nested lock case.
3717 */
3718 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
3719 if (page && page_ref_count(page) != 1) {
3720 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3721 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3722 goto again;
3723 }
3724
3725 return 0;
3726}
3727
3728/*
3729 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3730 * mmap activity.
3731 */
3732int
3733xfs_ilock2_io_mmap(
3734 struct xfs_inode *ip1,
3735 struct xfs_inode *ip2)
3736{
3737 int ret;
3738
3739 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3740 if (ret)
3741 return ret;
3742
3743 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3744 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
3745 if (ret) {
3746 inode_unlock(VFS_I(ip2));
3747 if (ip1 != ip2)
3748 inode_unlock(VFS_I(ip1));
3749 return ret;
3750 }
3751 } else
3752 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
3753 VFS_I(ip2)->i_mapping);
3754
3755 return 0;
3756}
3757
3758/* Unlock both inodes to allow IO and mmap activity. */
3759void
3760xfs_iunlock2_io_mmap(
3761 struct xfs_inode *ip1,
3762 struct xfs_inode *ip2)
3763{
3764 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3765 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3766 if (ip1 != ip2)
3767 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3768 } else
3769 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
3770 VFS_I(ip2)->i_mapping);
3771
3772 inode_unlock(VFS_I(ip2));
3773 if (ip1 != ip2)
3774 inode_unlock(VFS_I(ip1));
3775}
3776
3777/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
3778void
3779xfs_iunlock2_remapping(
3780 struct xfs_inode *ip1,
3781 struct xfs_inode *ip2)
3782{
3783 xfs_iflags_clear(ip1, XFS_IREMAPPING);
3784
3785 if (ip1 != ip2)
3786 xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
3787 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3788
3789 if (ip1 != ip2)
3790 inode_unlock_shared(VFS_I(ip1));
3791 inode_unlock(VFS_I(ip2));
3792}
3793
3794/*
3795 * Reload the incore inode list for this inode. Caller should ensure that
3796 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
3797 * preventing other threads from executing.
3798 */
3799int
3800xfs_inode_reload_unlinked_bucket(
3801 struct xfs_trans *tp,
3802 struct xfs_inode *ip)
3803{
3804 struct xfs_mount *mp = tp->t_mountp;
3805 struct xfs_buf *agibp;
3806 struct xfs_agi *agi;
3807 struct xfs_perag *pag;
3808 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
3809 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
3810 xfs_agino_t prev_agino, next_agino;
3811 unsigned int bucket;
3812 bool foundit = false;
3813 int error;
3814
3815 /* Grab the first inode in the list */
3816 pag = xfs_perag_get(mp, agno);
3817 error = xfs_ialloc_read_agi(pag, tp, &agibp);
3818 xfs_perag_put(pag);
3819 if (error)
3820 return error;
3821
3822 /*
3823 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
3824 * incore unlinked list pointers for this inode. Check once more to
3825 * see if we raced with anyone else to reload the unlinked list.
3826 */
3827 if (!xfs_inode_unlinked_incomplete(ip)) {
3828 foundit = true;
3829 goto out_agibp;
3830 }
3831
3832 bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
3833 agi = agibp->b_addr;
3834
3835 trace_xfs_inode_reload_unlinked_bucket(ip);
3836
3837 xfs_info_ratelimited(mp,
3838 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
3839 agino, agno);
3840
3841 prev_agino = NULLAGINO;
3842 next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3843 while (next_agino != NULLAGINO) {
3844 struct xfs_inode *next_ip = NULL;
3845
3846 /* Found this caller's inode, set its backlink. */
3847 if (next_agino == agino) {
3848 next_ip = ip;
3849 next_ip->i_prev_unlinked = prev_agino;
3850 foundit = true;
3851 goto next_inode;
3852 }
3853
3854 /* Try in-memory lookup first. */
3855 next_ip = xfs_iunlink_lookup(pag, next_agino);
3856 if (next_ip)
3857 goto next_inode;
3858
3859 /* Inode not in memory, try reloading it. */
3860 error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
3861 next_agino);
3862 if (error)
3863 break;
3864
3865 /* Grab the reloaded inode. */
3866 next_ip = xfs_iunlink_lookup(pag, next_agino);
3867 if (!next_ip) {
3868 /* No incore inode at all? We reloaded it... */
3869 ASSERT(next_ip != NULL);
3870 error = -EFSCORRUPTED;
3871 break;
3872 }
3873
3874next_inode:
3875 prev_agino = next_agino;
3876 next_agino = next_ip->i_next_unlinked;
3877 }
3878
3879out_agibp:
3880 xfs_trans_brelse(tp, agibp);
3881 /* Should have found this inode somewhere in the iunlinked bucket. */
3882 if (!error && !foundit)
3883 error = -EFSCORRUPTED;
3884 return error;
3885}
3886
3887/* Decide if this inode is missing its unlinked list and reload it. */
3888int
3889xfs_inode_reload_unlinked(
3890 struct xfs_inode *ip)
3891{
3892 struct xfs_trans *tp;
3893 int error;
3894
3895 error = xfs_trans_alloc_empty(ip->i_mount, &tp);
3896 if (error)
3897 return error;
3898
3899 xfs_ilock(ip, XFS_ILOCK_SHARED);
3900 if (xfs_inode_unlinked_incomplete(ip))
3901 error = xfs_inode_reload_unlinked_bucket(tp, ip);
3902 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3903 xfs_trans_cancel(tp);
3904
3905 return error;
3906}
3907
3908/* Has this inode fork been zapped by repair? */
3909bool
3910xfs_ifork_zapped(
3911 const struct xfs_inode *ip,
3912 int whichfork)
3913{
3914 unsigned int datamask = 0;
3915
3916 switch (whichfork) {
3917 case XFS_DATA_FORK:
3918 switch (ip->i_vnode.i_mode & S_IFMT) {
3919 case S_IFDIR:
3920 datamask = XFS_SICK_INO_DIR_ZAPPED;
3921 break;
3922 case S_IFLNK:
3923 datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
3924 break;
3925 }
3926 return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
3927 case XFS_ATTR_FORK:
3928 return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
3929 default:
3930 return false;
3931 }
3932}
3933
3934/* Compute the number of data and realtime blocks used by a file. */
3935void
3936xfs_inode_count_blocks(
3937 struct xfs_trans *tp,
3938 struct xfs_inode *ip,
3939 xfs_filblks_t *dblocks,
3940 xfs_filblks_t *rblocks)
3941{
3942 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
3943
3944 *rblocks = 0;
3945 if (XFS_IS_REALTIME_INODE(ip))
3946 xfs_bmap_count_leaves(ifp, rblocks);
3947 *dblocks = ip->i_nblocks - *rblocks;
3948}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include <linux/iversion.h>
7
8#include "xfs.h"
9#include "xfs_fs.h"
10#include "xfs_shared.h"
11#include "xfs_format.h"
12#include "xfs_log_format.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_defer.h"
16#include "xfs_inode.h"
17#include "xfs_dir2.h"
18#include "xfs_attr.h"
19#include "xfs_bit.h"
20#include "xfs_trans_space.h"
21#include "xfs_trans.h"
22#include "xfs_buf_item.h"
23#include "xfs_inode_item.h"
24#include "xfs_iunlink_item.h"
25#include "xfs_ialloc.h"
26#include "xfs_bmap.h"
27#include "xfs_bmap_util.h"
28#include "xfs_errortag.h"
29#include "xfs_error.h"
30#include "xfs_quota.h"
31#include "xfs_filestream.h"
32#include "xfs_trace.h"
33#include "xfs_icache.h"
34#include "xfs_symlink.h"
35#include "xfs_trans_priv.h"
36#include "xfs_log.h"
37#include "xfs_bmap_btree.h"
38#include "xfs_reflink.h"
39#include "xfs_ag.h"
40#include "xfs_log_priv.h"
41#include "xfs_health.h"
42#include "xfs_pnfs.h"
43#include "xfs_parent.h"
44#include "xfs_xattr.h"
45#include "xfs_inode_util.h"
46#include "xfs_metafile.h"
47
48struct kmem_cache *xfs_inode_cache;
49
50/*
51 * These two are wrapper routines around the xfs_ilock() routine used to
52 * centralize some grungy code. They are used in places that wish to lock the
53 * inode solely for reading the extents. The reason these places can't just
54 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
55 * bringing in of the extents from disk for a file in b-tree format. If the
56 * inode is in b-tree format, then we need to lock the inode exclusively until
57 * the extents are read in. Locking it exclusively all the time would limit
58 * our parallelism unnecessarily, though. What we do instead is check to see
59 * if the extents have been read in yet, and only lock the inode exclusively
60 * if they have not.
61 *
62 * The functions return a value which should be given to the corresponding
63 * xfs_iunlock() call.
64 */
65uint
66xfs_ilock_data_map_shared(
67 struct xfs_inode *ip)
68{
69 uint lock_mode = XFS_ILOCK_SHARED;
70
71 if (xfs_need_iread_extents(&ip->i_df))
72 lock_mode = XFS_ILOCK_EXCL;
73 xfs_ilock(ip, lock_mode);
74 return lock_mode;
75}
76
77uint
78xfs_ilock_attr_map_shared(
79 struct xfs_inode *ip)
80{
81 uint lock_mode = XFS_ILOCK_SHARED;
82
83 if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
84 lock_mode = XFS_ILOCK_EXCL;
85 xfs_ilock(ip, lock_mode);
86 return lock_mode;
87}
88
89/*
90 * You can't set both SHARED and EXCL for the same lock,
91 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
92 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
93 * to set in lock_flags.
94 */
95static inline void
96xfs_lock_flags_assert(
97 uint lock_flags)
98{
99 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
100 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
101 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
102 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
103 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
104 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
105 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
106 ASSERT(lock_flags != 0);
107}
108
109/*
110 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
111 * multi-reader locks: invalidate_lock and the i_lock. This routine allows
112 * various combinations of the locks to be obtained.
113 *
114 * The 3 locks should always be ordered so that the IO lock is obtained first,
115 * the mmap lock second and the ilock last in order to prevent deadlock.
116 *
117 * Basic locking order:
118 *
119 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
120 *
121 * mmap_lock locking order:
122 *
123 * i_rwsem -> page lock -> mmap_lock
124 * mmap_lock -> invalidate_lock -> page_lock
125 *
126 * The difference in mmap_lock locking order mean that we cannot hold the
127 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
128 * can fault in pages during copy in/out (for buffered IO) or require the
129 * mmap_lock in get_user_pages() to map the user pages into the kernel address
130 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
131 * fault because page faults already hold the mmap_lock.
132 *
133 * Hence to serialise fully against both syscall and mmap based IO, we need to
134 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
135 * both taken in places where we need to invalidate the page cache in a race
136 * free manner (e.g. truncate, hole punch and other extent manipulation
137 * functions).
138 */
139void
140xfs_ilock(
141 xfs_inode_t *ip,
142 uint lock_flags)
143{
144 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
145
146 xfs_lock_flags_assert(lock_flags);
147
148 if (lock_flags & XFS_IOLOCK_EXCL) {
149 down_write_nested(&VFS_I(ip)->i_rwsem,
150 XFS_IOLOCK_DEP(lock_flags));
151 } else if (lock_flags & XFS_IOLOCK_SHARED) {
152 down_read_nested(&VFS_I(ip)->i_rwsem,
153 XFS_IOLOCK_DEP(lock_flags));
154 }
155
156 if (lock_flags & XFS_MMAPLOCK_EXCL) {
157 down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
158 XFS_MMAPLOCK_DEP(lock_flags));
159 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
160 down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
161 XFS_MMAPLOCK_DEP(lock_flags));
162 }
163
164 if (lock_flags & XFS_ILOCK_EXCL)
165 down_write_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
166 else if (lock_flags & XFS_ILOCK_SHARED)
167 down_read_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
168}
169
170/*
171 * This is just like xfs_ilock(), except that the caller
172 * is guaranteed not to sleep. It returns 1 if it gets
173 * the requested locks and 0 otherwise. If the IO lock is
174 * obtained but the inode lock cannot be, then the IO lock
175 * is dropped before returning.
176 *
177 * ip -- the inode being locked
178 * lock_flags -- this parameter indicates the inode's locks to be
179 * to be locked. See the comment for xfs_ilock() for a list
180 * of valid values.
181 */
182int
183xfs_ilock_nowait(
184 xfs_inode_t *ip,
185 uint lock_flags)
186{
187 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
188
189 xfs_lock_flags_assert(lock_flags);
190
191 if (lock_flags & XFS_IOLOCK_EXCL) {
192 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
193 goto out;
194 } else if (lock_flags & XFS_IOLOCK_SHARED) {
195 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
196 goto out;
197 }
198
199 if (lock_flags & XFS_MMAPLOCK_EXCL) {
200 if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
201 goto out_undo_iolock;
202 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
203 if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
204 goto out_undo_iolock;
205 }
206
207 if (lock_flags & XFS_ILOCK_EXCL) {
208 if (!down_write_trylock(&ip->i_lock))
209 goto out_undo_mmaplock;
210 } else if (lock_flags & XFS_ILOCK_SHARED) {
211 if (!down_read_trylock(&ip->i_lock))
212 goto out_undo_mmaplock;
213 }
214 return 1;
215
216out_undo_mmaplock:
217 if (lock_flags & XFS_MMAPLOCK_EXCL)
218 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
219 else if (lock_flags & XFS_MMAPLOCK_SHARED)
220 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
221out_undo_iolock:
222 if (lock_flags & XFS_IOLOCK_EXCL)
223 up_write(&VFS_I(ip)->i_rwsem);
224 else if (lock_flags & XFS_IOLOCK_SHARED)
225 up_read(&VFS_I(ip)->i_rwsem);
226out:
227 return 0;
228}
229
230/*
231 * xfs_iunlock() is used to drop the inode locks acquired with
232 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
233 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
234 * that we know which locks to drop.
235 *
236 * ip -- the inode being unlocked
237 * lock_flags -- this parameter indicates the inode's locks to be
238 * to be unlocked. See the comment for xfs_ilock() for a list
239 * of valid values for this parameter.
240 *
241 */
242void
243xfs_iunlock(
244 xfs_inode_t *ip,
245 uint lock_flags)
246{
247 xfs_lock_flags_assert(lock_flags);
248
249 if (lock_flags & XFS_IOLOCK_EXCL)
250 up_write(&VFS_I(ip)->i_rwsem);
251 else if (lock_flags & XFS_IOLOCK_SHARED)
252 up_read(&VFS_I(ip)->i_rwsem);
253
254 if (lock_flags & XFS_MMAPLOCK_EXCL)
255 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
256 else if (lock_flags & XFS_MMAPLOCK_SHARED)
257 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
258
259 if (lock_flags & XFS_ILOCK_EXCL)
260 up_write(&ip->i_lock);
261 else if (lock_flags & XFS_ILOCK_SHARED)
262 up_read(&ip->i_lock);
263
264 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
265}
266
267/*
268 * give up write locks. the i/o lock cannot be held nested
269 * if it is being demoted.
270 */
271void
272xfs_ilock_demote(
273 xfs_inode_t *ip,
274 uint lock_flags)
275{
276 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
277 ASSERT((lock_flags &
278 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
279
280 if (lock_flags & XFS_ILOCK_EXCL)
281 downgrade_write(&ip->i_lock);
282 if (lock_flags & XFS_MMAPLOCK_EXCL)
283 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
284 if (lock_flags & XFS_IOLOCK_EXCL)
285 downgrade_write(&VFS_I(ip)->i_rwsem);
286
287 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
288}
289
290void
291xfs_assert_ilocked(
292 struct xfs_inode *ip,
293 uint lock_flags)
294{
295 /*
296 * Sometimes we assert the ILOCK is held exclusively, but we're in
297 * a workqueue, so lockdep doesn't know we're the owner.
298 */
299 if (lock_flags & XFS_ILOCK_SHARED)
300 rwsem_assert_held(&ip->i_lock);
301 else if (lock_flags & XFS_ILOCK_EXCL)
302 rwsem_assert_held_write_nolockdep(&ip->i_lock);
303
304 if (lock_flags & XFS_MMAPLOCK_SHARED)
305 rwsem_assert_held(&VFS_I(ip)->i_mapping->invalidate_lock);
306 else if (lock_flags & XFS_MMAPLOCK_EXCL)
307 rwsem_assert_held_write(&VFS_I(ip)->i_mapping->invalidate_lock);
308
309 if (lock_flags & XFS_IOLOCK_SHARED)
310 rwsem_assert_held(&VFS_I(ip)->i_rwsem);
311 else if (lock_flags & XFS_IOLOCK_EXCL)
312 rwsem_assert_held_write(&VFS_I(ip)->i_rwsem);
313}
314
315/*
316 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
317 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
318 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
319 * errors and warnings.
320 */
321#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
322static bool
323xfs_lockdep_subclass_ok(
324 int subclass)
325{
326 return subclass < MAX_LOCKDEP_SUBCLASSES;
327}
328#else
329#define xfs_lockdep_subclass_ok(subclass) (true)
330#endif
331
332/*
333 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
334 * value. This can be called for any type of inode lock combination, including
335 * parent locking. Care must be taken to ensure we don't overrun the subclass
336 * storage fields in the class mask we build.
337 */
338static inline uint
339xfs_lock_inumorder(
340 uint lock_mode,
341 uint subclass)
342{
343 uint class = 0;
344
345 ASSERT(!(lock_mode & XFS_ILOCK_PARENT));
346 ASSERT(xfs_lockdep_subclass_ok(subclass));
347
348 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
349 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
350 class += subclass << XFS_IOLOCK_SHIFT;
351 }
352
353 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
354 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
355 class += subclass << XFS_MMAPLOCK_SHIFT;
356 }
357
358 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
359 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
360 class += subclass << XFS_ILOCK_SHIFT;
361 }
362
363 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
364}
365
366/*
367 * The following routine will lock n inodes in exclusive mode. We assume the
368 * caller calls us with the inodes in i_ino order.
369 *
370 * We need to detect deadlock where an inode that we lock is in the AIL and we
371 * start waiting for another inode that is locked by a thread in a long running
372 * transaction (such as truncate). This can result in deadlock since the long
373 * running trans might need to wait for the inode we just locked in order to
374 * push the tail and free space in the log.
375 *
376 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
377 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
378 * lock more than one at a time, lockdep will report false positives saying we
379 * have violated locking orders.
380 */
381void
382xfs_lock_inodes(
383 struct xfs_inode **ips,
384 int inodes,
385 uint lock_mode)
386{
387 int attempts = 0;
388 uint i;
389 int j;
390 bool try_lock;
391 struct xfs_log_item *lp;
392
393 /*
394 * Currently supports between 2 and 5 inodes with exclusive locking. We
395 * support an arbitrary depth of locking here, but absolute limits on
396 * inodes depend on the type of locking and the limits placed by
397 * lockdep annotations in xfs_lock_inumorder. These are all checked by
398 * the asserts.
399 */
400 ASSERT(ips && inodes >= 2 && inodes <= 5);
401 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
402 XFS_ILOCK_EXCL));
403 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
404 XFS_ILOCK_SHARED)));
405 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
406 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
407 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
408 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
409
410 if (lock_mode & XFS_IOLOCK_EXCL) {
411 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
412 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
413 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
414
415again:
416 try_lock = false;
417 i = 0;
418 for (; i < inodes; i++) {
419 ASSERT(ips[i]);
420
421 if (i && (ips[i] == ips[i - 1])) /* Already locked */
422 continue;
423
424 /*
425 * If try_lock is not set yet, make sure all locked inodes are
426 * not in the AIL. If any are, set try_lock to be used later.
427 */
428 if (!try_lock) {
429 for (j = (i - 1); j >= 0 && !try_lock; j--) {
430 lp = &ips[j]->i_itemp->ili_item;
431 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
432 try_lock = true;
433 }
434 }
435
436 /*
437 * If any of the previous locks we have locked is in the AIL,
438 * we must TRY to get the second and subsequent locks. If
439 * we can't get any, we must release all we have
440 * and try again.
441 */
442 if (!try_lock) {
443 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
444 continue;
445 }
446
447 /* try_lock means we have an inode locked that is in the AIL. */
448 ASSERT(i != 0);
449 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
450 continue;
451
452 /*
453 * Unlock all previous guys and try again. xfs_iunlock will try
454 * to push the tail if the inode is in the AIL.
455 */
456 attempts++;
457 for (j = i - 1; j >= 0; j--) {
458 /*
459 * Check to see if we've already unlocked this one. Not
460 * the first one going back, and the inode ptr is the
461 * same.
462 */
463 if (j != (i - 1) && ips[j] == ips[j + 1])
464 continue;
465
466 xfs_iunlock(ips[j], lock_mode);
467 }
468
469 if ((attempts % 5) == 0) {
470 delay(1); /* Don't just spin the CPU */
471 }
472 goto again;
473 }
474}
475
476/*
477 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
478 * mmaplock must be double-locked separately since we use i_rwsem and
479 * invalidate_lock for that. We now support taking one lock EXCL and the
480 * other SHARED.
481 */
482void
483xfs_lock_two_inodes(
484 struct xfs_inode *ip0,
485 uint ip0_mode,
486 struct xfs_inode *ip1,
487 uint ip1_mode)
488{
489 int attempts = 0;
490 struct xfs_log_item *lp;
491
492 ASSERT(hweight32(ip0_mode) == 1);
493 ASSERT(hweight32(ip1_mode) == 1);
494 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
495 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
496 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
497 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
498 ASSERT(ip0->i_ino != ip1->i_ino);
499
500 if (ip0->i_ino > ip1->i_ino) {
501 swap(ip0, ip1);
502 swap(ip0_mode, ip1_mode);
503 }
504
505 again:
506 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
507
508 /*
509 * If the first lock we have locked is in the AIL, we must TRY to get
510 * the second lock. If we can't get it, we must release the first one
511 * and try again.
512 */
513 lp = &ip0->i_itemp->ili_item;
514 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
515 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
516 xfs_iunlock(ip0, ip0_mode);
517 if ((++attempts % 5) == 0)
518 delay(1); /* Don't just spin the CPU */
519 goto again;
520 }
521 } else {
522 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
523 }
524}
525
526/*
527 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
528 * is allowed, otherwise it has to be an exact match. If a CI match is found,
529 * ci_name->name will point to a the actual name (caller must free) or
530 * will be set to NULL if an exact match is found.
531 */
532int
533xfs_lookup(
534 struct xfs_inode *dp,
535 const struct xfs_name *name,
536 struct xfs_inode **ipp,
537 struct xfs_name *ci_name)
538{
539 xfs_ino_t inum;
540 int error;
541
542 trace_xfs_lookup(dp, name);
543
544 if (xfs_is_shutdown(dp->i_mount))
545 return -EIO;
546 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
547 return -EIO;
548
549 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
550 if (error)
551 goto out_unlock;
552
553 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
554 if (error)
555 goto out_free_name;
556
557 /*
558 * Fail if a directory entry in the regular directory tree points to
559 * a metadata file.
560 */
561 if (XFS_IS_CORRUPT(dp->i_mount, xfs_is_metadir_inode(*ipp))) {
562 xfs_fs_mark_sick(dp->i_mount, XFS_SICK_FS_METADIR);
563 error = -EFSCORRUPTED;
564 goto out_irele;
565 }
566
567 return 0;
568
569out_irele:
570 xfs_irele(*ipp);
571out_free_name:
572 if (ci_name)
573 kfree(ci_name->name);
574out_unlock:
575 *ipp = NULL;
576 return error;
577}
578
579/*
580 * Initialise a newly allocated inode and return the in-core inode to the
581 * caller locked exclusively.
582 *
583 * Caller is responsible for unlocking the inode manually upon return
584 */
585int
586xfs_icreate(
587 struct xfs_trans *tp,
588 xfs_ino_t ino,
589 const struct xfs_icreate_args *args,
590 struct xfs_inode **ipp)
591{
592 struct xfs_mount *mp = tp->t_mountp;
593 struct xfs_inode *ip = NULL;
594 int error;
595
596 /*
597 * Get the in-core inode with the lock held exclusively to prevent
598 * others from looking at until we're done.
599 */
600 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
601 if (error)
602 return error;
603
604 ASSERT(ip != NULL);
605 xfs_trans_ijoin(tp, ip, 0);
606 xfs_inode_init(tp, args, ip);
607
608 /* now that we have an i_mode we can setup the inode structure */
609 xfs_setup_inode(ip);
610
611 *ipp = ip;
612 return 0;
613}
614
615/* Return dquots for the ids that will be assigned to a new file. */
616int
617xfs_icreate_dqalloc(
618 const struct xfs_icreate_args *args,
619 struct xfs_dquot **udqpp,
620 struct xfs_dquot **gdqpp,
621 struct xfs_dquot **pdqpp)
622{
623 struct inode *dir = VFS_I(args->pip);
624 kuid_t uid = GLOBAL_ROOT_UID;
625 kgid_t gid = GLOBAL_ROOT_GID;
626 prid_t prid = 0;
627 unsigned int flags = XFS_QMOPT_QUOTALL;
628
629 if (args->idmap) {
630 /*
631 * The uid/gid computation code must match what the VFS uses to
632 * assign i_[ug]id. INHERIT adjusts the gid computation for
633 * setgid/grpid systems.
634 */
635 uid = mapped_fsuid(args->idmap, i_user_ns(dir));
636 gid = mapped_fsgid(args->idmap, i_user_ns(dir));
637 prid = xfs_get_initial_prid(args->pip);
638 flags |= XFS_QMOPT_INHERIT;
639 }
640
641 *udqpp = *gdqpp = *pdqpp = NULL;
642
643 return xfs_qm_vop_dqalloc(args->pip, uid, gid, prid, flags, udqpp,
644 gdqpp, pdqpp);
645}
646
647int
648xfs_create(
649 const struct xfs_icreate_args *args,
650 struct xfs_name *name,
651 struct xfs_inode **ipp)
652{
653 struct xfs_inode *dp = args->pip;
654 struct xfs_dir_update du = {
655 .dp = dp,
656 .name = name,
657 };
658 struct xfs_mount *mp = dp->i_mount;
659 struct xfs_trans *tp = NULL;
660 struct xfs_dquot *udqp;
661 struct xfs_dquot *gdqp;
662 struct xfs_dquot *pdqp;
663 struct xfs_trans_res *tres;
664 xfs_ino_t ino;
665 bool unlock_dp_on_error = false;
666 bool is_dir = S_ISDIR(args->mode);
667 uint resblks;
668 int error;
669
670 trace_xfs_create(dp, name);
671
672 if (xfs_is_shutdown(mp))
673 return -EIO;
674 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
675 return -EIO;
676
677 /* Make sure that we have allocated dquot(s) on disk. */
678 error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
679 if (error)
680 return error;
681
682 if (is_dir) {
683 resblks = xfs_mkdir_space_res(mp, name->len);
684 tres = &M_RES(mp)->tr_mkdir;
685 } else {
686 resblks = xfs_create_space_res(mp, name->len);
687 tres = &M_RES(mp)->tr_create;
688 }
689
690 error = xfs_parent_start(mp, &du.ppargs);
691 if (error)
692 goto out_release_dquots;
693
694 /*
695 * Initially assume that the file does not exist and
696 * reserve the resources for that case. If that is not
697 * the case we'll drop the one we have and get a more
698 * appropriate transaction later.
699 */
700 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
701 &tp);
702 if (error == -ENOSPC) {
703 /* flush outstanding delalloc blocks and retry */
704 xfs_flush_inodes(mp);
705 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
706 resblks, &tp);
707 }
708 if (error)
709 goto out_parent;
710
711 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
712 unlock_dp_on_error = true;
713
714 /*
715 * A newly created regular or special file just has one directory
716 * entry pointing to them, but a directory also the "." entry
717 * pointing to itself.
718 */
719 error = xfs_dialloc(&tp, args, &ino);
720 if (!error)
721 error = xfs_icreate(tp, ino, args, &du.ip);
722 if (error)
723 goto out_trans_cancel;
724
725 /*
726 * Now we join the directory inode to the transaction. We do not do it
727 * earlier because xfs_dialloc might commit the previous transaction
728 * (and release all the locks). An error from here on will result in
729 * the transaction cancel unlocking dp so don't do it explicitly in the
730 * error path.
731 */
732 xfs_trans_ijoin(tp, dp, 0);
733
734 error = xfs_dir_create_child(tp, resblks, &du);
735 if (error)
736 goto out_trans_cancel;
737
738 /*
739 * If this is a synchronous mount, make sure that the
740 * create transaction goes to disk before returning to
741 * the user.
742 */
743 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
744 xfs_trans_set_sync(tp);
745
746 /*
747 * Attach the dquot(s) to the inodes and modify them incore.
748 * These ids of the inode couldn't have changed since the new
749 * inode has been locked ever since it was created.
750 */
751 xfs_qm_vop_create_dqattach(tp, du.ip, udqp, gdqp, pdqp);
752
753 error = xfs_trans_commit(tp);
754 if (error)
755 goto out_release_inode;
756
757 xfs_qm_dqrele(udqp);
758 xfs_qm_dqrele(gdqp);
759 xfs_qm_dqrele(pdqp);
760
761 *ipp = du.ip;
762 xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
763 xfs_iunlock(dp, XFS_ILOCK_EXCL);
764 xfs_parent_finish(mp, du.ppargs);
765 return 0;
766
767 out_trans_cancel:
768 xfs_trans_cancel(tp);
769 out_release_inode:
770 /*
771 * Wait until after the current transaction is aborted to finish the
772 * setup of the inode and release the inode. This prevents recursive
773 * transactions and deadlocks from xfs_inactive.
774 */
775 if (du.ip) {
776 xfs_iunlock(du.ip, XFS_ILOCK_EXCL);
777 xfs_finish_inode_setup(du.ip);
778 xfs_irele(du.ip);
779 }
780 out_parent:
781 xfs_parent_finish(mp, du.ppargs);
782 out_release_dquots:
783 xfs_qm_dqrele(udqp);
784 xfs_qm_dqrele(gdqp);
785 xfs_qm_dqrele(pdqp);
786
787 if (unlock_dp_on_error)
788 xfs_iunlock(dp, XFS_ILOCK_EXCL);
789 return error;
790}
791
792int
793xfs_create_tmpfile(
794 const struct xfs_icreate_args *args,
795 struct xfs_inode **ipp)
796{
797 struct xfs_inode *dp = args->pip;
798 struct xfs_mount *mp = dp->i_mount;
799 struct xfs_inode *ip = NULL;
800 struct xfs_trans *tp = NULL;
801 struct xfs_dquot *udqp;
802 struct xfs_dquot *gdqp;
803 struct xfs_dquot *pdqp;
804 struct xfs_trans_res *tres;
805 xfs_ino_t ino;
806 uint resblks;
807 int error;
808
809 ASSERT(args->flags & XFS_ICREATE_TMPFILE);
810
811 if (xfs_is_shutdown(mp))
812 return -EIO;
813
814 /* Make sure that we have allocated dquot(s) on disk. */
815 error = xfs_icreate_dqalloc(args, &udqp, &gdqp, &pdqp);
816 if (error)
817 return error;
818
819 resblks = XFS_IALLOC_SPACE_RES(mp);
820 tres = &M_RES(mp)->tr_create_tmpfile;
821
822 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
823 &tp);
824 if (error)
825 goto out_release_dquots;
826
827 error = xfs_dialloc(&tp, args, &ino);
828 if (!error)
829 error = xfs_icreate(tp, ino, args, &ip);
830 if (error)
831 goto out_trans_cancel;
832
833 if (xfs_has_wsync(mp))
834 xfs_trans_set_sync(tp);
835
836 /*
837 * Attach the dquot(s) to the inodes and modify them incore.
838 * These ids of the inode couldn't have changed since the new
839 * inode has been locked ever since it was created.
840 */
841 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
842
843 error = xfs_iunlink(tp, ip);
844 if (error)
845 goto out_trans_cancel;
846
847 error = xfs_trans_commit(tp);
848 if (error)
849 goto out_release_inode;
850
851 xfs_qm_dqrele(udqp);
852 xfs_qm_dqrele(gdqp);
853 xfs_qm_dqrele(pdqp);
854
855 *ipp = ip;
856 xfs_iunlock(ip, XFS_ILOCK_EXCL);
857 return 0;
858
859 out_trans_cancel:
860 xfs_trans_cancel(tp);
861 out_release_inode:
862 /*
863 * Wait until after the current transaction is aborted to finish the
864 * setup of the inode and release the inode. This prevents recursive
865 * transactions and deadlocks from xfs_inactive.
866 */
867 if (ip) {
868 xfs_iunlock(ip, XFS_ILOCK_EXCL);
869 xfs_finish_inode_setup(ip);
870 xfs_irele(ip);
871 }
872 out_release_dquots:
873 xfs_qm_dqrele(udqp);
874 xfs_qm_dqrele(gdqp);
875 xfs_qm_dqrele(pdqp);
876
877 return error;
878}
879
880int
881xfs_link(
882 struct xfs_inode *tdp,
883 struct xfs_inode *sip,
884 struct xfs_name *target_name)
885{
886 struct xfs_dir_update du = {
887 .dp = tdp,
888 .name = target_name,
889 .ip = sip,
890 };
891 struct xfs_mount *mp = tdp->i_mount;
892 struct xfs_trans *tp;
893 int error, nospace_error = 0;
894 int resblks;
895
896 trace_xfs_link(tdp, target_name);
897
898 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
899
900 if (xfs_is_shutdown(mp))
901 return -EIO;
902 if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
903 return -EIO;
904
905 error = xfs_qm_dqattach(sip);
906 if (error)
907 goto std_return;
908
909 error = xfs_qm_dqattach(tdp);
910 if (error)
911 goto std_return;
912
913 error = xfs_parent_start(mp, &du.ppargs);
914 if (error)
915 goto std_return;
916
917 resblks = xfs_link_space_res(mp, target_name->len);
918 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
919 &tp, &nospace_error);
920 if (error)
921 goto out_parent;
922
923 /*
924 * We don't allow reservationless or quotaless hardlinking when parent
925 * pointers are enabled because we can't back out if the xattrs must
926 * grow.
927 */
928 if (du.ppargs && nospace_error) {
929 error = nospace_error;
930 goto error_return;
931 }
932
933 /*
934 * If we are using project inheritance, we only allow hard link
935 * creation in our tree when the project IDs are the same; else
936 * the tree quota mechanism could be circumvented.
937 */
938 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
939 tdp->i_projid != sip->i_projid)) {
940 /*
941 * Project quota setup skips special files which can
942 * leave inodes in a PROJINHERIT directory without a
943 * project ID set. We need to allow links to be made
944 * to these "project-less" inodes because userspace
945 * expects them to succeed after project ID setup,
946 * but everything else should be rejected.
947 */
948 if (!special_file(VFS_I(sip)->i_mode) ||
949 sip->i_projid != 0) {
950 error = -EXDEV;
951 goto error_return;
952 }
953 }
954
955 error = xfs_dir_add_child(tp, resblks, &du);
956 if (error)
957 goto error_return;
958
959 /*
960 * If this is a synchronous mount, make sure that the
961 * link transaction goes to disk before returning to
962 * the user.
963 */
964 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
965 xfs_trans_set_sync(tp);
966
967 error = xfs_trans_commit(tp);
968 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
969 xfs_iunlock(sip, XFS_ILOCK_EXCL);
970 xfs_parent_finish(mp, du.ppargs);
971 return error;
972
973 error_return:
974 xfs_trans_cancel(tp);
975 xfs_iunlock(tdp, XFS_ILOCK_EXCL);
976 xfs_iunlock(sip, XFS_ILOCK_EXCL);
977 out_parent:
978 xfs_parent_finish(mp, du.ppargs);
979 std_return:
980 if (error == -ENOSPC && nospace_error)
981 error = nospace_error;
982 return error;
983}
984
985/* Clear the reflink flag and the cowblocks tag if possible. */
986static void
987xfs_itruncate_clear_reflink_flags(
988 struct xfs_inode *ip)
989{
990 struct xfs_ifork *dfork;
991 struct xfs_ifork *cfork;
992
993 if (!xfs_is_reflink_inode(ip))
994 return;
995 dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
996 cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
997 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
998 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
999 if (cfork->if_bytes == 0)
1000 xfs_inode_clear_cowblocks_tag(ip);
1001}
1002
1003/*
1004 * Free up the underlying blocks past new_size. The new size must be smaller
1005 * than the current size. This routine can be used both for the attribute and
1006 * data fork, and does not modify the inode size, which is left to the caller.
1007 *
1008 * The transaction passed to this routine must have made a permanent log
1009 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1010 * given transaction and start new ones, so make sure everything involved in
1011 * the transaction is tidy before calling here. Some transaction will be
1012 * returned to the caller to be committed. The incoming transaction must
1013 * already include the inode, and both inode locks must be held exclusively.
1014 * The inode must also be "held" within the transaction. On return the inode
1015 * will be "held" within the returned transaction. This routine does NOT
1016 * require any disk space to be reserved for it within the transaction.
1017 *
1018 * If we get an error, we must return with the inode locked and linked into the
1019 * current transaction. This keeps things simple for the higher level code,
1020 * because it always knows that the inode is locked and held in the transaction
1021 * that returns to it whether errors occur or not. We don't mark the inode
1022 * dirty on error so that transactions can be easily aborted if possible.
1023 */
1024int
1025xfs_itruncate_extents_flags(
1026 struct xfs_trans **tpp,
1027 struct xfs_inode *ip,
1028 int whichfork,
1029 xfs_fsize_t new_size,
1030 int flags)
1031{
1032 struct xfs_mount *mp = ip->i_mount;
1033 struct xfs_trans *tp = *tpp;
1034 xfs_fileoff_t first_unmap_block;
1035 int error = 0;
1036
1037 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1038 if (atomic_read(&VFS_I(ip)->i_count))
1039 xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL);
1040 ASSERT(new_size <= XFS_ISIZE(ip));
1041 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1042 ASSERT(ip->i_itemp != NULL);
1043 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1044 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1045
1046 trace_xfs_itruncate_extents_start(ip, new_size);
1047
1048 flags |= xfs_bmapi_aflag(whichfork);
1049
1050 /*
1051 * Since it is possible for space to become allocated beyond
1052 * the end of the file (in a crash where the space is allocated
1053 * but the inode size is not yet updated), simply remove any
1054 * blocks which show up between the new EOF and the maximum
1055 * possible file size.
1056 *
1057 * We have to free all the blocks to the bmbt maximum offset, even if
1058 * the page cache can't scale that far.
1059 */
1060 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1061 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1062 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1063 return 0;
1064 }
1065
1066 error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1067 XFS_MAX_FILEOFF);
1068 if (error)
1069 goto out;
1070
1071 if (whichfork == XFS_DATA_FORK) {
1072 /* Remove all pending CoW reservations. */
1073 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1074 first_unmap_block, XFS_MAX_FILEOFF, true);
1075 if (error)
1076 goto out;
1077
1078 xfs_itruncate_clear_reflink_flags(ip);
1079 }
1080
1081 /*
1082 * Always re-log the inode so that our permanent transaction can keep
1083 * on rolling it forward in the log.
1084 */
1085 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1086
1087 trace_xfs_itruncate_extents_end(ip, new_size);
1088
1089out:
1090 *tpp = tp;
1091 return error;
1092}
1093
1094/*
1095 * Mark all the buffers attached to this directory stale. In theory we should
1096 * never be freeing a directory with any blocks at all, but this covers the
1097 * case where we've recovered a directory swap with a "temporary" directory
1098 * created by online repair and now need to dump it.
1099 */
1100STATIC void
1101xfs_inactive_dir(
1102 struct xfs_inode *dp)
1103{
1104 struct xfs_iext_cursor icur;
1105 struct xfs_bmbt_irec got;
1106 struct xfs_mount *mp = dp->i_mount;
1107 struct xfs_da_geometry *geo = mp->m_dir_geo;
1108 struct xfs_ifork *ifp = xfs_ifork_ptr(dp, XFS_DATA_FORK);
1109 xfs_fileoff_t off;
1110
1111 /*
1112 * Invalidate each directory block. All directory blocks are of
1113 * fsbcount length and alignment, so we only need to walk those same
1114 * offsets. We hold the only reference to this inode, so we must wait
1115 * for the buffer locks.
1116 */
1117 for_each_xfs_iext(ifp, &icur, &got) {
1118 for (off = round_up(got.br_startoff, geo->fsbcount);
1119 off < got.br_startoff + got.br_blockcount;
1120 off += geo->fsbcount) {
1121 struct xfs_buf *bp = NULL;
1122 xfs_fsblock_t fsbno;
1123 int error;
1124
1125 fsbno = (off - got.br_startoff) + got.br_startblock;
1126 error = xfs_buf_incore(mp->m_ddev_targp,
1127 XFS_FSB_TO_DADDR(mp, fsbno),
1128 XFS_FSB_TO_BB(mp, geo->fsbcount),
1129 XBF_LIVESCAN, &bp);
1130 if (error)
1131 continue;
1132
1133 xfs_buf_stale(bp);
1134 xfs_buf_relse(bp);
1135 }
1136 }
1137}
1138
1139/*
1140 * xfs_inactive_truncate
1141 *
1142 * Called to perform a truncate when an inode becomes unlinked.
1143 */
1144STATIC int
1145xfs_inactive_truncate(
1146 struct xfs_inode *ip)
1147{
1148 struct xfs_mount *mp = ip->i_mount;
1149 struct xfs_trans *tp;
1150 int error;
1151
1152 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1153 if (error) {
1154 ASSERT(xfs_is_shutdown(mp));
1155 return error;
1156 }
1157 xfs_ilock(ip, XFS_ILOCK_EXCL);
1158 xfs_trans_ijoin(tp, ip, 0);
1159
1160 /*
1161 * Log the inode size first to prevent stale data exposure in the event
1162 * of a system crash before the truncate completes. See the related
1163 * comment in xfs_vn_setattr_size() for details.
1164 */
1165 ip->i_disk_size = 0;
1166 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1167
1168 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1169 if (error)
1170 goto error_trans_cancel;
1171
1172 ASSERT(ip->i_df.if_nextents == 0);
1173
1174 error = xfs_trans_commit(tp);
1175 if (error)
1176 goto error_unlock;
1177
1178 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1179 return 0;
1180
1181error_trans_cancel:
1182 xfs_trans_cancel(tp);
1183error_unlock:
1184 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1185 return error;
1186}
1187
1188/*
1189 * xfs_inactive_ifree()
1190 *
1191 * Perform the inode free when an inode is unlinked.
1192 */
1193STATIC int
1194xfs_inactive_ifree(
1195 struct xfs_inode *ip)
1196{
1197 struct xfs_mount *mp = ip->i_mount;
1198 struct xfs_trans *tp;
1199 int error;
1200
1201 /*
1202 * We try to use a per-AG reservation for any block needed by the finobt
1203 * tree, but as the finobt feature predates the per-AG reservation
1204 * support a degraded file system might not have enough space for the
1205 * reservation at mount time. In that case try to dip into the reserved
1206 * pool and pray.
1207 *
1208 * Send a warning if the reservation does happen to fail, as the inode
1209 * now remains allocated and sits on the unlinked list until the fs is
1210 * repaired.
1211 */
1212 if (unlikely(mp->m_finobt_nores)) {
1213 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1214 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1215 &tp);
1216 } else {
1217 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1218 }
1219 if (error) {
1220 if (error == -ENOSPC) {
1221 xfs_warn_ratelimited(mp,
1222 "Failed to remove inode(s) from unlinked list. "
1223 "Please free space, unmount and run xfs_repair.");
1224 } else {
1225 ASSERT(xfs_is_shutdown(mp));
1226 }
1227 return error;
1228 }
1229
1230 /*
1231 * We do not hold the inode locked across the entire rolling transaction
1232 * here. We only need to hold it for the first transaction that
1233 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1234 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1235 * here breaks the relationship between cluster buffer invalidation and
1236 * stale inode invalidation on cluster buffer item journal commit
1237 * completion, and can result in leaving dirty stale inodes hanging
1238 * around in memory.
1239 *
1240 * We have no need for serialising this inode operation against other
1241 * operations - we freed the inode and hence reallocation is required
1242 * and that will serialise on reallocating the space the deferops need
1243 * to free. Hence we can unlock the inode on the first commit of
1244 * the transaction rather than roll it right through the deferops. This
1245 * avoids relogging the XFS_ISTALE inode.
1246 *
1247 * We check that xfs_ifree() hasn't grown an internal transaction roll
1248 * by asserting that the inode is still locked when it returns.
1249 */
1250 xfs_ilock(ip, XFS_ILOCK_EXCL);
1251 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1252
1253 error = xfs_ifree(tp, ip);
1254 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1255 if (error) {
1256 /*
1257 * If we fail to free the inode, shut down. The cancel
1258 * might do that, we need to make sure. Otherwise the
1259 * inode might be lost for a long time or forever.
1260 */
1261 if (!xfs_is_shutdown(mp)) {
1262 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1263 __func__, error);
1264 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1265 }
1266 xfs_trans_cancel(tp);
1267 return error;
1268 }
1269
1270 /*
1271 * Credit the quota account(s). The inode is gone.
1272 */
1273 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1274
1275 return xfs_trans_commit(tp);
1276}
1277
1278/*
1279 * Returns true if we need to update the on-disk metadata before we can free
1280 * the memory used by this inode. Updates include freeing post-eof
1281 * preallocations; freeing COW staging extents; and marking the inode free in
1282 * the inobt if it is on the unlinked list.
1283 */
1284bool
1285xfs_inode_needs_inactive(
1286 struct xfs_inode *ip)
1287{
1288 struct xfs_mount *mp = ip->i_mount;
1289 struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1290
1291 /*
1292 * If the inode is already free, then there can be nothing
1293 * to clean up here.
1294 */
1295 if (VFS_I(ip)->i_mode == 0)
1296 return false;
1297
1298 /*
1299 * If this is a read-only mount, don't do this (would generate I/O)
1300 * unless we're in log recovery and cleaning the iunlinked list.
1301 */
1302 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1303 return false;
1304
1305 /* If the log isn't running, push inodes straight to reclaim. */
1306 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1307 return false;
1308
1309 /* Metadata inodes require explicit resource cleanup. */
1310 if (xfs_is_internal_inode(ip))
1311 return false;
1312
1313 /* Want to clean out the cow blocks if there are any. */
1314 if (cow_ifp && cow_ifp->if_bytes > 0)
1315 return true;
1316
1317 /* Unlinked files must be freed. */
1318 if (VFS_I(ip)->i_nlink == 0)
1319 return true;
1320
1321 /*
1322 * This file isn't being freed, so check if there are post-eof blocks
1323 * to free.
1324 *
1325 * Note: don't bother with iolock here since lockdep complains about
1326 * acquiring it in reclaim context. We have the only reference to the
1327 * inode at this point anyways.
1328 */
1329 return xfs_can_free_eofblocks(ip);
1330}
1331
1332/*
1333 * Save health status somewhere, if we're dumping an inode with uncorrected
1334 * errors and online repair isn't running.
1335 */
1336static inline void
1337xfs_inactive_health(
1338 struct xfs_inode *ip)
1339{
1340 struct xfs_mount *mp = ip->i_mount;
1341 struct xfs_perag *pag;
1342 unsigned int sick;
1343 unsigned int checked;
1344
1345 xfs_inode_measure_sickness(ip, &sick, &checked);
1346 if (!sick)
1347 return;
1348
1349 trace_xfs_inode_unfixed_corruption(ip, sick);
1350
1351 if (sick & XFS_SICK_INO_FORGET)
1352 return;
1353
1354 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1355 if (!pag) {
1356 /* There had better still be a perag structure! */
1357 ASSERT(0);
1358 return;
1359 }
1360
1361 xfs_ag_mark_sick(pag, XFS_SICK_AG_INODES);
1362 xfs_perag_put(pag);
1363}
1364
1365/*
1366 * xfs_inactive
1367 *
1368 * This is called when the vnode reference count for the vnode
1369 * goes to zero. If the file has been unlinked, then it must
1370 * now be truncated. Also, we clear all of the read-ahead state
1371 * kept for the inode here since the file is now closed.
1372 */
1373int
1374xfs_inactive(
1375 xfs_inode_t *ip)
1376{
1377 struct xfs_mount *mp;
1378 int error = 0;
1379 int truncate = 0;
1380
1381 /*
1382 * If the inode is already free, then there can be nothing
1383 * to clean up here.
1384 */
1385 if (VFS_I(ip)->i_mode == 0) {
1386 ASSERT(ip->i_df.if_broot_bytes == 0);
1387 goto out;
1388 }
1389
1390 mp = ip->i_mount;
1391 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1392
1393 xfs_inactive_health(ip);
1394
1395 /*
1396 * If this is a read-only mount, don't do this (would generate I/O)
1397 * unless we're in log recovery and cleaning the iunlinked list.
1398 */
1399 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1400 goto out;
1401
1402 /* Metadata inodes require explicit resource cleanup. */
1403 if (xfs_is_internal_inode(ip))
1404 goto out;
1405
1406 /* Try to clean out the cow blocks if there are any. */
1407 if (xfs_inode_has_cow_data(ip)) {
1408 error = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1409 if (error)
1410 goto out;
1411 }
1412
1413 if (VFS_I(ip)->i_nlink != 0) {
1414 /*
1415 * Note: don't bother with iolock here since lockdep complains
1416 * about acquiring it in reclaim context. We have the only
1417 * reference to the inode at this point anyways.
1418 */
1419 if (xfs_can_free_eofblocks(ip))
1420 error = xfs_free_eofblocks(ip);
1421
1422 goto out;
1423 }
1424
1425 if (S_ISREG(VFS_I(ip)->i_mode) &&
1426 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1427 xfs_inode_has_filedata(ip)))
1428 truncate = 1;
1429
1430 if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1431 /*
1432 * If this inode is being inactivated during a quotacheck and
1433 * has not yet been scanned by quotacheck, we /must/ remove
1434 * the dquots from the inode before inactivation changes the
1435 * block and inode counts. Most probably this is a result of
1436 * reloading the incore iunlinked list to purge unrecovered
1437 * unlinked inodes.
1438 */
1439 xfs_qm_dqdetach(ip);
1440 } else {
1441 error = xfs_qm_dqattach(ip);
1442 if (error)
1443 goto out;
1444 }
1445
1446 if (S_ISDIR(VFS_I(ip)->i_mode) && ip->i_df.if_nextents > 0) {
1447 xfs_inactive_dir(ip);
1448 truncate = 1;
1449 }
1450
1451 if (S_ISLNK(VFS_I(ip)->i_mode))
1452 error = xfs_inactive_symlink(ip);
1453 else if (truncate)
1454 error = xfs_inactive_truncate(ip);
1455 if (error)
1456 goto out;
1457
1458 /*
1459 * If there are attributes associated with the file then blow them away
1460 * now. The code calls a routine that recursively deconstructs the
1461 * attribute fork. If also blows away the in-core attribute fork.
1462 */
1463 if (xfs_inode_has_attr_fork(ip)) {
1464 error = xfs_attr_inactive(ip);
1465 if (error)
1466 goto out;
1467 }
1468
1469 ASSERT(ip->i_forkoff == 0);
1470
1471 /*
1472 * Free the inode.
1473 */
1474 error = xfs_inactive_ifree(ip);
1475
1476out:
1477 /*
1478 * We're done making metadata updates for this inode, so we can release
1479 * the attached dquots.
1480 */
1481 xfs_qm_dqdetach(ip);
1482 return error;
1483}
1484
1485/*
1486 * Find an inode on the unlinked list. This does not take references to the
1487 * inode as we have existence guarantees by holding the AGI buffer lock and that
1488 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1489 * don't find the inode in cache, then let the caller handle the situation.
1490 */
1491struct xfs_inode *
1492xfs_iunlink_lookup(
1493 struct xfs_perag *pag,
1494 xfs_agino_t agino)
1495{
1496 struct xfs_inode *ip;
1497
1498 rcu_read_lock();
1499 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1500 if (!ip) {
1501 /* Caller can handle inode not being in memory. */
1502 rcu_read_unlock();
1503 return NULL;
1504 }
1505
1506 /*
1507 * Inode in RCU freeing limbo should not happen. Warn about this and
1508 * let the caller handle the failure.
1509 */
1510 if (WARN_ON_ONCE(!ip->i_ino)) {
1511 rcu_read_unlock();
1512 return NULL;
1513 }
1514 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1515 rcu_read_unlock();
1516 return ip;
1517}
1518
1519/*
1520 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1521 * to @prev_agino. Caller must hold the AGI to synchronize with other changes
1522 * to the unlinked list.
1523 */
1524int
1525xfs_iunlink_reload_next(
1526 struct xfs_trans *tp,
1527 struct xfs_buf *agibp,
1528 xfs_agino_t prev_agino,
1529 xfs_agino_t next_agino)
1530{
1531 struct xfs_perag *pag = agibp->b_pag;
1532 struct xfs_mount *mp = pag_mount(pag);
1533 struct xfs_inode *next_ip = NULL;
1534 int error;
1535
1536 ASSERT(next_agino != NULLAGINO);
1537
1538#ifdef DEBUG
1539 rcu_read_lock();
1540 next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1541 ASSERT(next_ip == NULL);
1542 rcu_read_unlock();
1543#endif
1544
1545 xfs_info_ratelimited(mp,
1546 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
1547 next_agino, pag_agno(pag));
1548
1549 /*
1550 * Use an untrusted lookup just to be cautious in case the AGI has been
1551 * corrupted and now points at a free inode. That shouldn't happen,
1552 * but we'd rather shut down now since we're already running in a weird
1553 * situation.
1554 */
1555 error = xfs_iget(mp, tp, xfs_agino_to_ino(pag, next_agino),
1556 XFS_IGET_UNTRUSTED, 0, &next_ip);
1557 if (error) {
1558 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1559 return error;
1560 }
1561
1562 /* If this is not an unlinked inode, something is very wrong. */
1563 if (VFS_I(next_ip)->i_nlink != 0) {
1564 xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
1565 error = -EFSCORRUPTED;
1566 goto rele;
1567 }
1568
1569 next_ip->i_prev_unlinked = prev_agino;
1570 trace_xfs_iunlink_reload_next(next_ip);
1571rele:
1572 ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
1573 if (xfs_is_quotacheck_running(mp) && next_ip)
1574 xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1575 xfs_irele(next_ip);
1576 return error;
1577}
1578
1579/*
1580 * Look up the inode number specified and if it is not already marked XFS_ISTALE
1581 * mark it stale. We should only find clean inodes in this lookup that aren't
1582 * already stale.
1583 */
1584static void
1585xfs_ifree_mark_inode_stale(
1586 struct xfs_perag *pag,
1587 struct xfs_inode *free_ip,
1588 xfs_ino_t inum)
1589{
1590 struct xfs_mount *mp = pag_mount(pag);
1591 struct xfs_inode_log_item *iip;
1592 struct xfs_inode *ip;
1593
1594retry:
1595 rcu_read_lock();
1596 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
1597
1598 /* Inode not in memory, nothing to do */
1599 if (!ip) {
1600 rcu_read_unlock();
1601 return;
1602 }
1603
1604 /*
1605 * because this is an RCU protected lookup, we could find a recently
1606 * freed or even reallocated inode during the lookup. We need to check
1607 * under the i_flags_lock for a valid inode here. Skip it if it is not
1608 * valid, the wrong inode or stale.
1609 */
1610 spin_lock(&ip->i_flags_lock);
1611 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
1612 goto out_iflags_unlock;
1613
1614 /*
1615 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
1616 * other inodes that we did not find in the list attached to the buffer
1617 * and are not already marked stale. If we can't lock it, back off and
1618 * retry.
1619 */
1620 if (ip != free_ip) {
1621 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
1622 spin_unlock(&ip->i_flags_lock);
1623 rcu_read_unlock();
1624 delay(1);
1625 goto retry;
1626 }
1627 }
1628 ip->i_flags |= XFS_ISTALE;
1629
1630 /*
1631 * If the inode is flushing, it is already attached to the buffer. All
1632 * we needed to do here is mark the inode stale so buffer IO completion
1633 * will remove it from the AIL.
1634 */
1635 iip = ip->i_itemp;
1636 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
1637 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
1638 ASSERT(iip->ili_last_fields);
1639 goto out_iunlock;
1640 }
1641
1642 /*
1643 * Inodes not attached to the buffer can be released immediately.
1644 * Everything else has to go through xfs_iflush_abort() on journal
1645 * commit as the flock synchronises removal of the inode from the
1646 * cluster buffer against inode reclaim.
1647 */
1648 if (!iip || list_empty(&iip->ili_item.li_bio_list))
1649 goto out_iunlock;
1650
1651 __xfs_iflags_set(ip, XFS_IFLUSHING);
1652 spin_unlock(&ip->i_flags_lock);
1653 rcu_read_unlock();
1654
1655 /* we have a dirty inode in memory that has not yet been flushed. */
1656 spin_lock(&iip->ili_lock);
1657 iip->ili_last_fields = iip->ili_fields;
1658 iip->ili_fields = 0;
1659 iip->ili_fsync_fields = 0;
1660 spin_unlock(&iip->ili_lock);
1661 ASSERT(iip->ili_last_fields);
1662
1663 if (ip != free_ip)
1664 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1665 return;
1666
1667out_iunlock:
1668 if (ip != free_ip)
1669 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1670out_iflags_unlock:
1671 spin_unlock(&ip->i_flags_lock);
1672 rcu_read_unlock();
1673}
1674
1675/*
1676 * A big issue when freeing the inode cluster is that we _cannot_ skip any
1677 * inodes that are in memory - they all must be marked stale and attached to
1678 * the cluster buffer.
1679 */
1680static int
1681xfs_ifree_cluster(
1682 struct xfs_trans *tp,
1683 struct xfs_perag *pag,
1684 struct xfs_inode *free_ip,
1685 struct xfs_icluster *xic)
1686{
1687 struct xfs_mount *mp = free_ip->i_mount;
1688 struct xfs_ino_geometry *igeo = M_IGEO(mp);
1689 struct xfs_buf *bp;
1690 xfs_daddr_t blkno;
1691 xfs_ino_t inum = xic->first_ino;
1692 int nbufs;
1693 int i, j;
1694 int ioffset;
1695 int error;
1696
1697 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
1698
1699 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
1700 /*
1701 * The allocation bitmap tells us which inodes of the chunk were
1702 * physically allocated. Skip the cluster if an inode falls into
1703 * a sparse region.
1704 */
1705 ioffset = inum - xic->first_ino;
1706 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
1707 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
1708 continue;
1709 }
1710
1711 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
1712 XFS_INO_TO_AGBNO(mp, inum));
1713
1714 /*
1715 * We obtain and lock the backing buffer first in the process
1716 * here to ensure dirty inodes attached to the buffer remain in
1717 * the flushing state while we mark them stale.
1718 *
1719 * If we scan the in-memory inodes first, then buffer IO can
1720 * complete before we get a lock on it, and hence we may fail
1721 * to mark all the active inodes on the buffer stale.
1722 */
1723 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
1724 mp->m_bsize * igeo->blocks_per_cluster,
1725 XBF_UNMAPPED, &bp);
1726 if (error)
1727 return error;
1728
1729 /*
1730 * This buffer may not have been correctly initialised as we
1731 * didn't read it from disk. That's not important because we are
1732 * only using to mark the buffer as stale in the log, and to
1733 * attach stale cached inodes on it.
1734 *
1735 * For the inode that triggered the cluster freeing, this
1736 * attachment may occur in xfs_inode_item_precommit() after we
1737 * have marked this buffer stale. If this buffer was not in
1738 * memory before xfs_ifree_cluster() started, it will not be
1739 * marked XBF_DONE and this will cause problems later in
1740 * xfs_inode_item_precommit() when we trip over a (stale, !done)
1741 * buffer to attached to the transaction.
1742 *
1743 * Hence we have to mark the buffer as XFS_DONE here. This is
1744 * safe because we are also marking the buffer as XBF_STALE and
1745 * XFS_BLI_STALE. That means it will never be dispatched for
1746 * IO and it won't be unlocked until the cluster freeing has
1747 * been committed to the journal and the buffer unpinned. If it
1748 * is written, we want to know about it, and we want it to
1749 * fail. We can acheive this by adding a write verifier to the
1750 * buffer.
1751 */
1752 bp->b_flags |= XBF_DONE;
1753 bp->b_ops = &xfs_inode_buf_ops;
1754
1755 /*
1756 * Now we need to set all the cached clean inodes as XFS_ISTALE,
1757 * too. This requires lookups, and will skip inodes that we've
1758 * already marked XFS_ISTALE.
1759 */
1760 for (i = 0; i < igeo->inodes_per_cluster; i++)
1761 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
1762
1763 xfs_trans_stale_inode_buf(tp, bp);
1764 xfs_trans_binval(tp, bp);
1765 }
1766 return 0;
1767}
1768
1769/*
1770 * This is called to return an inode to the inode free list. The inode should
1771 * already be truncated to 0 length and have no pages associated with it. This
1772 * routine also assumes that the inode is already a part of the transaction.
1773 *
1774 * The on-disk copy of the inode will have been added to the list of unlinked
1775 * inodes in the AGI. We need to remove the inode from that list atomically with
1776 * respect to freeing it here.
1777 */
1778int
1779xfs_ifree(
1780 struct xfs_trans *tp,
1781 struct xfs_inode *ip)
1782{
1783 struct xfs_mount *mp = ip->i_mount;
1784 struct xfs_perag *pag;
1785 struct xfs_icluster xic = { 0 };
1786 struct xfs_inode_log_item *iip = ip->i_itemp;
1787 int error;
1788
1789 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1790 ASSERT(VFS_I(ip)->i_nlink == 0);
1791 ASSERT(ip->i_df.if_nextents == 0);
1792 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
1793 ASSERT(ip->i_nblocks == 0);
1794
1795 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1796
1797 error = xfs_inode_uninit(tp, pag, ip, &xic);
1798 if (error)
1799 goto out;
1800
1801 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
1802 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
1803
1804 /* Don't attempt to replay owner changes for a deleted inode */
1805 spin_lock(&iip->ili_lock);
1806 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
1807 spin_unlock(&iip->ili_lock);
1808
1809 if (xic.deleted)
1810 error = xfs_ifree_cluster(tp, pag, ip, &xic);
1811out:
1812 xfs_perag_put(pag);
1813 return error;
1814}
1815
1816/*
1817 * This is called to unpin an inode. The caller must have the inode locked
1818 * in at least shared mode so that the buffer cannot be subsequently pinned
1819 * once someone is waiting for it to be unpinned.
1820 */
1821static void
1822xfs_iunpin(
1823 struct xfs_inode *ip)
1824{
1825 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
1826
1827 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
1828
1829 /* Give the log a push to start the unpinning I/O */
1830 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
1831
1832}
1833
1834static void
1835__xfs_iunpin_wait(
1836 struct xfs_inode *ip)
1837{
1838 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
1839 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
1840
1841 xfs_iunpin(ip);
1842
1843 do {
1844 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
1845 if (xfs_ipincount(ip))
1846 io_schedule();
1847 } while (xfs_ipincount(ip));
1848 finish_wait(wq, &wait.wq_entry);
1849}
1850
1851void
1852xfs_iunpin_wait(
1853 struct xfs_inode *ip)
1854{
1855 if (xfs_ipincount(ip))
1856 __xfs_iunpin_wait(ip);
1857}
1858
1859/*
1860 * Removing an inode from the namespace involves removing the directory entry
1861 * and dropping the link count on the inode. Removing the directory entry can
1862 * result in locking an AGF (directory blocks were freed) and removing a link
1863 * count can result in placing the inode on an unlinked list which results in
1864 * locking an AGI.
1865 *
1866 * The big problem here is that we have an ordering constraint on AGF and AGI
1867 * locking - inode allocation locks the AGI, then can allocate a new extent for
1868 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
1869 * removes the inode from the unlinked list, requiring that we lock the AGI
1870 * first, and then freeing the inode can result in an inode chunk being freed
1871 * and hence freeing disk space requiring that we lock an AGF.
1872 *
1873 * Hence the ordering that is imposed by other parts of the code is AGI before
1874 * AGF. This means we cannot remove the directory entry before we drop the inode
1875 * reference count and put it on the unlinked list as this results in a lock
1876 * order of AGF then AGI, and this can deadlock against inode allocation and
1877 * freeing. Therefore we must drop the link counts before we remove the
1878 * directory entry.
1879 *
1880 * This is still safe from a transactional point of view - it is not until we
1881 * get to xfs_defer_finish() that we have the possibility of multiple
1882 * transactions in this operation. Hence as long as we remove the directory
1883 * entry and drop the link count in the first transaction of the remove
1884 * operation, there are no transactional constraints on the ordering here.
1885 */
1886int
1887xfs_remove(
1888 struct xfs_inode *dp,
1889 struct xfs_name *name,
1890 struct xfs_inode *ip)
1891{
1892 struct xfs_dir_update du = {
1893 .dp = dp,
1894 .name = name,
1895 .ip = ip,
1896 };
1897 struct xfs_mount *mp = dp->i_mount;
1898 struct xfs_trans *tp = NULL;
1899 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
1900 int dontcare;
1901 int error = 0;
1902 uint resblks;
1903
1904 trace_xfs_remove(dp, name);
1905
1906 if (xfs_is_shutdown(mp))
1907 return -EIO;
1908 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
1909 return -EIO;
1910
1911 error = xfs_qm_dqattach(dp);
1912 if (error)
1913 goto std_return;
1914
1915 error = xfs_qm_dqattach(ip);
1916 if (error)
1917 goto std_return;
1918
1919 error = xfs_parent_start(mp, &du.ppargs);
1920 if (error)
1921 goto std_return;
1922
1923 /*
1924 * We try to get the real space reservation first, allowing for
1925 * directory btree deletion(s) implying possible bmap insert(s). If we
1926 * can't get the space reservation then we use 0 instead, and avoid the
1927 * bmap btree insert(s) in the directory code by, if the bmap insert
1928 * tries to happen, instead trimming the LAST block from the directory.
1929 *
1930 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
1931 * the directory code can handle a reservationless update and we don't
1932 * want to prevent a user from trying to free space by deleting things.
1933 */
1934 resblks = xfs_remove_space_res(mp, name->len);
1935 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
1936 &tp, &dontcare);
1937 if (error) {
1938 ASSERT(error != -ENOSPC);
1939 goto out_parent;
1940 }
1941
1942 error = xfs_dir_remove_child(tp, resblks, &du);
1943 if (error)
1944 goto out_trans_cancel;
1945
1946 /*
1947 * If this is a synchronous mount, make sure that the
1948 * remove transaction goes to disk before returning to
1949 * the user.
1950 */
1951 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1952 xfs_trans_set_sync(tp);
1953
1954 error = xfs_trans_commit(tp);
1955 if (error)
1956 goto out_unlock;
1957
1958 if (is_dir && xfs_inode_is_filestream(ip))
1959 xfs_filestream_deassociate(ip);
1960
1961 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1962 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1963 xfs_parent_finish(mp, du.ppargs);
1964 return 0;
1965
1966 out_trans_cancel:
1967 xfs_trans_cancel(tp);
1968 out_unlock:
1969 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1970 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1971 out_parent:
1972 xfs_parent_finish(mp, du.ppargs);
1973 std_return:
1974 return error;
1975}
1976
1977static inline void
1978xfs_iunlock_rename(
1979 struct xfs_inode **i_tab,
1980 int num_inodes)
1981{
1982 int i;
1983
1984 for (i = num_inodes - 1; i >= 0; i--) {
1985 /* Skip duplicate inodes if src and target dps are the same */
1986 if (!i_tab[i] || (i > 0 && i_tab[i] == i_tab[i - 1]))
1987 continue;
1988 xfs_iunlock(i_tab[i], XFS_ILOCK_EXCL);
1989 }
1990}
1991
1992/*
1993 * Enter all inodes for a rename transaction into a sorted array.
1994 */
1995#define __XFS_SORT_INODES 5
1996STATIC void
1997xfs_sort_for_rename(
1998 struct xfs_inode *dp1, /* in: old (source) directory inode */
1999 struct xfs_inode *dp2, /* in: new (target) directory inode */
2000 struct xfs_inode *ip1, /* in: inode of old entry */
2001 struct xfs_inode *ip2, /* in: inode of new entry */
2002 struct xfs_inode *wip, /* in: whiteout inode */
2003 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2004 int *num_inodes) /* in/out: inodes in array */
2005{
2006 int i;
2007
2008 ASSERT(*num_inodes == __XFS_SORT_INODES);
2009 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2010
2011 /*
2012 * i_tab contains a list of pointers to inodes. We initialize
2013 * the table here & we'll sort it. We will then use it to
2014 * order the acquisition of the inode locks.
2015 *
2016 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2017 */
2018 i = 0;
2019 i_tab[i++] = dp1;
2020 i_tab[i++] = dp2;
2021 i_tab[i++] = ip1;
2022 if (ip2)
2023 i_tab[i++] = ip2;
2024 if (wip)
2025 i_tab[i++] = wip;
2026 *num_inodes = i;
2027
2028 xfs_sort_inodes(i_tab, *num_inodes);
2029}
2030
2031void
2032xfs_sort_inodes(
2033 struct xfs_inode **i_tab,
2034 unsigned int num_inodes)
2035{
2036 int i, j;
2037
2038 ASSERT(num_inodes <= __XFS_SORT_INODES);
2039
2040 /*
2041 * Sort the elements via bubble sort. (Remember, there are at
2042 * most 5 elements to sort, so this is adequate.)
2043 */
2044 for (i = 0; i < num_inodes; i++) {
2045 for (j = 1; j < num_inodes; j++) {
2046 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino)
2047 swap(i_tab[j], i_tab[j - 1]);
2048 }
2049 }
2050}
2051
2052/*
2053 * xfs_rename_alloc_whiteout()
2054 *
2055 * Return a referenced, unlinked, unlocked inode that can be used as a
2056 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2057 * crash between allocating the inode and linking it into the rename transaction
2058 * recovery will free the inode and we won't leak it.
2059 */
2060static int
2061xfs_rename_alloc_whiteout(
2062 struct mnt_idmap *idmap,
2063 struct xfs_name *src_name,
2064 struct xfs_inode *dp,
2065 struct xfs_inode **wip)
2066{
2067 struct xfs_icreate_args args = {
2068 .idmap = idmap,
2069 .pip = dp,
2070 .mode = S_IFCHR | WHITEOUT_MODE,
2071 .flags = XFS_ICREATE_TMPFILE,
2072 };
2073 struct xfs_inode *tmpfile;
2074 struct qstr name;
2075 int error;
2076
2077 error = xfs_create_tmpfile(&args, &tmpfile);
2078 if (error)
2079 return error;
2080
2081 name.name = src_name->name;
2082 name.len = src_name->len;
2083 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2084 if (error) {
2085 xfs_finish_inode_setup(tmpfile);
2086 xfs_irele(tmpfile);
2087 return error;
2088 }
2089
2090 /*
2091 * Prepare the tmpfile inode as if it were created through the VFS.
2092 * Complete the inode setup and flag it as linkable. nlink is already
2093 * zero, so we can skip the drop_nlink.
2094 */
2095 xfs_setup_iops(tmpfile);
2096 xfs_finish_inode_setup(tmpfile);
2097 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2098
2099 *wip = tmpfile;
2100 return 0;
2101}
2102
2103/*
2104 * xfs_rename
2105 */
2106int
2107xfs_rename(
2108 struct mnt_idmap *idmap,
2109 struct xfs_inode *src_dp,
2110 struct xfs_name *src_name,
2111 struct xfs_inode *src_ip,
2112 struct xfs_inode *target_dp,
2113 struct xfs_name *target_name,
2114 struct xfs_inode *target_ip,
2115 unsigned int flags)
2116{
2117 struct xfs_dir_update du_src = {
2118 .dp = src_dp,
2119 .name = src_name,
2120 .ip = src_ip,
2121 };
2122 struct xfs_dir_update du_tgt = {
2123 .dp = target_dp,
2124 .name = target_name,
2125 .ip = target_ip,
2126 };
2127 struct xfs_dir_update du_wip = { };
2128 struct xfs_mount *mp = src_dp->i_mount;
2129 struct xfs_trans *tp;
2130 struct xfs_inode *inodes[__XFS_SORT_INODES];
2131 int i;
2132 int num_inodes = __XFS_SORT_INODES;
2133 bool new_parent = (src_dp != target_dp);
2134 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2135 int spaceres;
2136 bool retried = false;
2137 int error, nospace_error = 0;
2138
2139 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2140
2141 if ((flags & RENAME_EXCHANGE) && !target_ip)
2142 return -EINVAL;
2143
2144 /*
2145 * If we are doing a whiteout operation, allocate the whiteout inode
2146 * we will be placing at the target and ensure the type is set
2147 * appropriately.
2148 */
2149 if (flags & RENAME_WHITEOUT) {
2150 error = xfs_rename_alloc_whiteout(idmap, src_name, target_dp,
2151 &du_wip.ip);
2152 if (error)
2153 return error;
2154
2155 /* setup target dirent info as whiteout */
2156 src_name->type = XFS_DIR3_FT_CHRDEV;
2157 }
2158
2159 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, du_wip.ip,
2160 inodes, &num_inodes);
2161
2162 error = xfs_parent_start(mp, &du_src.ppargs);
2163 if (error)
2164 goto out_release_wip;
2165
2166 if (du_wip.ip) {
2167 error = xfs_parent_start(mp, &du_wip.ppargs);
2168 if (error)
2169 goto out_src_ppargs;
2170 }
2171
2172 if (target_ip) {
2173 error = xfs_parent_start(mp, &du_tgt.ppargs);
2174 if (error)
2175 goto out_wip_ppargs;
2176 }
2177
2178retry:
2179 nospace_error = 0;
2180 spaceres = xfs_rename_space_res(mp, src_name->len, target_ip != NULL,
2181 target_name->len, du_wip.ip != NULL);
2182 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2183 if (error == -ENOSPC) {
2184 nospace_error = error;
2185 spaceres = 0;
2186 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2187 &tp);
2188 }
2189 if (error)
2190 goto out_tgt_ppargs;
2191
2192 /*
2193 * We don't allow reservationless renaming when parent pointers are
2194 * enabled because we can't back out if the xattrs must grow.
2195 */
2196 if (du_src.ppargs && nospace_error) {
2197 error = nospace_error;
2198 xfs_trans_cancel(tp);
2199 goto out_tgt_ppargs;
2200 }
2201
2202 /*
2203 * Attach the dquots to the inodes
2204 */
2205 error = xfs_qm_vop_rename_dqattach(inodes);
2206 if (error) {
2207 xfs_trans_cancel(tp);
2208 goto out_tgt_ppargs;
2209 }
2210
2211 /*
2212 * Lock all the participating inodes. Depending upon whether
2213 * the target_name exists in the target directory, and
2214 * whether the target directory is the same as the source
2215 * directory, we can lock from 2 to 5 inodes.
2216 */
2217 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2218
2219 /*
2220 * Join all the inodes to the transaction.
2221 */
2222 xfs_trans_ijoin(tp, src_dp, 0);
2223 if (new_parent)
2224 xfs_trans_ijoin(tp, target_dp, 0);
2225 xfs_trans_ijoin(tp, src_ip, 0);
2226 if (target_ip)
2227 xfs_trans_ijoin(tp, target_ip, 0);
2228 if (du_wip.ip)
2229 xfs_trans_ijoin(tp, du_wip.ip, 0);
2230
2231 /*
2232 * If we are using project inheritance, we only allow renames
2233 * into our tree when the project IDs are the same; else the
2234 * tree quota mechanism would be circumvented.
2235 */
2236 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2237 target_dp->i_projid != src_ip->i_projid)) {
2238 error = -EXDEV;
2239 goto out_trans_cancel;
2240 }
2241
2242 /* RENAME_EXCHANGE is unique from here on. */
2243 if (flags & RENAME_EXCHANGE) {
2244 error = xfs_dir_exchange_children(tp, &du_src, &du_tgt,
2245 spaceres);
2246 if (error)
2247 goto out_trans_cancel;
2248 goto out_commit;
2249 }
2250
2251 /*
2252 * Try to reserve quota to handle an expansion of the target directory.
2253 * We'll allow the rename to continue in reservationless mode if we hit
2254 * a space usage constraint. If we trigger reservationless mode, save
2255 * the errno if there isn't any free space in the target directory.
2256 */
2257 if (spaceres != 0) {
2258 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2259 0, false);
2260 if (error == -EDQUOT || error == -ENOSPC) {
2261 if (!retried) {
2262 xfs_trans_cancel(tp);
2263 xfs_iunlock_rename(inodes, num_inodes);
2264 xfs_blockgc_free_quota(target_dp, 0);
2265 retried = true;
2266 goto retry;
2267 }
2268
2269 nospace_error = error;
2270 spaceres = 0;
2271 error = 0;
2272 }
2273 if (error)
2274 goto out_trans_cancel;
2275 }
2276
2277 /*
2278 * We don't allow quotaless renaming when parent pointers are enabled
2279 * because we can't back out if the xattrs must grow.
2280 */
2281 if (du_src.ppargs && nospace_error) {
2282 error = nospace_error;
2283 goto out_trans_cancel;
2284 }
2285
2286 /*
2287 * Lock the AGI buffers we need to handle bumping the nlink of the
2288 * whiteout inode off the unlinked list and to handle dropping the
2289 * nlink of the target inode. Per locking order rules, do this in
2290 * increasing AG order and before directory block allocation tries to
2291 * grab AGFs because we grab AGIs before AGFs.
2292 *
2293 * The (vfs) caller must ensure that if src is a directory then
2294 * target_ip is either null or an empty directory.
2295 */
2296 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2297 if (inodes[i] == du_wip.ip ||
2298 (inodes[i] == target_ip &&
2299 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2300 struct xfs_perag *pag;
2301 struct xfs_buf *bp;
2302
2303 pag = xfs_perag_get(mp,
2304 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2305 error = xfs_read_agi(pag, tp, 0, &bp);
2306 xfs_perag_put(pag);
2307 if (error)
2308 goto out_trans_cancel;
2309 }
2310 }
2311
2312 error = xfs_dir_rename_children(tp, &du_src, &du_tgt, spaceres,
2313 &du_wip);
2314 if (error)
2315 goto out_trans_cancel;
2316
2317 if (du_wip.ip) {
2318 /*
2319 * Now we have a real link, clear the "I'm a tmpfile" state
2320 * flag from the inode so it doesn't accidentally get misused in
2321 * future.
2322 */
2323 VFS_I(du_wip.ip)->i_state &= ~I_LINKABLE;
2324 }
2325
2326out_commit:
2327 /*
2328 * If this is a synchronous mount, make sure that the rename
2329 * transaction goes to disk before returning to the user.
2330 */
2331 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2332 xfs_trans_set_sync(tp);
2333
2334 error = xfs_trans_commit(tp);
2335 nospace_error = 0;
2336 goto out_unlock;
2337
2338out_trans_cancel:
2339 xfs_trans_cancel(tp);
2340out_unlock:
2341 xfs_iunlock_rename(inodes, num_inodes);
2342out_tgt_ppargs:
2343 xfs_parent_finish(mp, du_tgt.ppargs);
2344out_wip_ppargs:
2345 xfs_parent_finish(mp, du_wip.ppargs);
2346out_src_ppargs:
2347 xfs_parent_finish(mp, du_src.ppargs);
2348out_release_wip:
2349 if (du_wip.ip)
2350 xfs_irele(du_wip.ip);
2351 if (error == -ENOSPC && nospace_error)
2352 error = nospace_error;
2353 return error;
2354}
2355
2356static int
2357xfs_iflush(
2358 struct xfs_inode *ip,
2359 struct xfs_buf *bp)
2360{
2361 struct xfs_inode_log_item *iip = ip->i_itemp;
2362 struct xfs_dinode *dip;
2363 struct xfs_mount *mp = ip->i_mount;
2364 int error;
2365
2366 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED);
2367 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
2368 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
2369 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
2370 ASSERT(iip->ili_item.li_buf == bp);
2371
2372 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
2373
2374 /*
2375 * We don't flush the inode if any of the following checks fail, but we
2376 * do still update the log item and attach to the backing buffer as if
2377 * the flush happened. This is a formality to facilitate predictable
2378 * error handling as the caller will shutdown and fail the buffer.
2379 */
2380 error = -EFSCORRUPTED;
2381 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
2382 mp, XFS_ERRTAG_IFLUSH_1)) {
2383 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2384 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
2385 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
2386 goto flush_out;
2387 }
2388 if (S_ISREG(VFS_I(ip)->i_mode)) {
2389 if (XFS_TEST_ERROR(
2390 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2391 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
2392 mp, XFS_ERRTAG_IFLUSH_3)) {
2393 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2394 "%s: Bad regular inode %llu, ptr "PTR_FMT,
2395 __func__, ip->i_ino, ip);
2396 goto flush_out;
2397 }
2398 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
2399 if (XFS_TEST_ERROR(
2400 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
2401 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
2402 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
2403 mp, XFS_ERRTAG_IFLUSH_4)) {
2404 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2405 "%s: Bad directory inode %llu, ptr "PTR_FMT,
2406 __func__, ip->i_ino, ip);
2407 goto flush_out;
2408 }
2409 }
2410 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
2411 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
2412 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2413 "%s: detected corrupt incore inode %llu, "
2414 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
2415 __func__, ip->i_ino,
2416 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
2417 ip->i_nblocks, ip);
2418 goto flush_out;
2419 }
2420 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
2421 mp, XFS_ERRTAG_IFLUSH_6)) {
2422 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
2423 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
2424 __func__, ip->i_ino, ip->i_forkoff, ip);
2425 goto flush_out;
2426 }
2427
2428 /*
2429 * Inode item log recovery for v2 inodes are dependent on the flushiter
2430 * count for correct sequencing. We bump the flush iteration count so
2431 * we can detect flushes which postdate a log record during recovery.
2432 * This is redundant as we now log every change and hence this can't
2433 * happen but we need to still do it to ensure backwards compatibility
2434 * with old kernels that predate logging all inode changes.
2435 */
2436 if (!xfs_has_v3inodes(mp))
2437 ip->i_flushiter++;
2438
2439 /*
2440 * If there are inline format data / attr forks attached to this inode,
2441 * make sure they are not corrupt.
2442 */
2443 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
2444 xfs_ifork_verify_local_data(ip))
2445 goto flush_out;
2446 if (xfs_inode_has_attr_fork(ip) &&
2447 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
2448 xfs_ifork_verify_local_attr(ip))
2449 goto flush_out;
2450
2451 /*
2452 * Copy the dirty parts of the inode into the on-disk inode. We always
2453 * copy out the core of the inode, because if the inode is dirty at all
2454 * the core must be.
2455 */
2456 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
2457
2458 /* Wrap, we never let the log put out DI_MAX_FLUSH */
2459 if (!xfs_has_v3inodes(mp)) {
2460 if (ip->i_flushiter == DI_MAX_FLUSH)
2461 ip->i_flushiter = 0;
2462 }
2463
2464 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
2465 if (xfs_inode_has_attr_fork(ip))
2466 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
2467
2468 /*
2469 * We've recorded everything logged in the inode, so we'd like to clear
2470 * the ili_fields bits so we don't log and flush things unnecessarily.
2471 * However, we can't stop logging all this information until the data
2472 * we've copied into the disk buffer is written to disk. If we did we
2473 * might overwrite the copy of the inode in the log with all the data
2474 * after re-logging only part of it, and in the face of a crash we
2475 * wouldn't have all the data we need to recover.
2476 *
2477 * What we do is move the bits to the ili_last_fields field. When
2478 * logging the inode, these bits are moved back to the ili_fields field.
2479 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
2480 * we know that the information those bits represent is permanently on
2481 * disk. As long as the flush completes before the inode is logged
2482 * again, then both ili_fields and ili_last_fields will be cleared.
2483 */
2484 error = 0;
2485flush_out:
2486 spin_lock(&iip->ili_lock);
2487 iip->ili_last_fields = iip->ili_fields;
2488 iip->ili_fields = 0;
2489 iip->ili_fsync_fields = 0;
2490 set_bit(XFS_LI_FLUSHING, &iip->ili_item.li_flags);
2491 spin_unlock(&iip->ili_lock);
2492
2493 /*
2494 * Store the current LSN of the inode so that we can tell whether the
2495 * item has moved in the AIL from xfs_buf_inode_iodone().
2496 */
2497 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2498 &iip->ili_item.li_lsn);
2499
2500 /* generate the checksum. */
2501 xfs_dinode_calc_crc(mp, dip);
2502 if (error)
2503 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
2504 return error;
2505}
2506
2507/*
2508 * Non-blocking flush of dirty inode metadata into the backing buffer.
2509 *
2510 * The caller must have a reference to the inode and hold the cluster buffer
2511 * locked. The function will walk across all the inodes on the cluster buffer it
2512 * can find and lock without blocking, and flush them to the cluster buffer.
2513 *
2514 * On successful flushing of at least one inode, the caller must write out the
2515 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
2516 * the caller needs to release the buffer. On failure, the filesystem will be
2517 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
2518 * will be returned.
2519 */
2520int
2521xfs_iflush_cluster(
2522 struct xfs_buf *bp)
2523{
2524 struct xfs_mount *mp = bp->b_mount;
2525 struct xfs_log_item *lip, *n;
2526 struct xfs_inode *ip;
2527 struct xfs_inode_log_item *iip;
2528 int clcount = 0;
2529 int error = 0;
2530
2531 /*
2532 * We must use the safe variant here as on shutdown xfs_iflush_abort()
2533 * will remove itself from the list.
2534 */
2535 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
2536 iip = (struct xfs_inode_log_item *)lip;
2537 ip = iip->ili_inode;
2538
2539 /*
2540 * Quick and dirty check to avoid locks if possible.
2541 */
2542 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
2543 continue;
2544 if (xfs_ipincount(ip))
2545 continue;
2546
2547 /*
2548 * The inode is still attached to the buffer, which means it is
2549 * dirty but reclaim might try to grab it. Check carefully for
2550 * that, and grab the ilock while still holding the i_flags_lock
2551 * to guarantee reclaim will not be able to reclaim this inode
2552 * once we drop the i_flags_lock.
2553 */
2554 spin_lock(&ip->i_flags_lock);
2555 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
2556 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
2557 spin_unlock(&ip->i_flags_lock);
2558 continue;
2559 }
2560
2561 /*
2562 * ILOCK will pin the inode against reclaim and prevent
2563 * concurrent transactions modifying the inode while we are
2564 * flushing the inode. If we get the lock, set the flushing
2565 * state before we drop the i_flags_lock.
2566 */
2567 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2568 spin_unlock(&ip->i_flags_lock);
2569 continue;
2570 }
2571 __xfs_iflags_set(ip, XFS_IFLUSHING);
2572 spin_unlock(&ip->i_flags_lock);
2573
2574 /*
2575 * Abort flushing this inode if we are shut down because the
2576 * inode may not currently be in the AIL. This can occur when
2577 * log I/O failure unpins the inode without inserting into the
2578 * AIL, leaving a dirty/unpinned inode attached to the buffer
2579 * that otherwise looks like it should be flushed.
2580 */
2581 if (xlog_is_shutdown(mp->m_log)) {
2582 xfs_iunpin_wait(ip);
2583 xfs_iflush_abort(ip);
2584 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2585 error = -EIO;
2586 continue;
2587 }
2588
2589 /* don't block waiting on a log force to unpin dirty inodes */
2590 if (xfs_ipincount(ip)) {
2591 xfs_iflags_clear(ip, XFS_IFLUSHING);
2592 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2593 continue;
2594 }
2595
2596 if (!xfs_inode_clean(ip))
2597 error = xfs_iflush(ip, bp);
2598 else
2599 xfs_iflags_clear(ip, XFS_IFLUSHING);
2600 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2601 if (error)
2602 break;
2603 clcount++;
2604 }
2605
2606 if (error) {
2607 /*
2608 * Shutdown first so we kill the log before we release this
2609 * buffer. If it is an INODE_ALLOC buffer and pins the tail
2610 * of the log, failing it before the _log_ is shut down can
2611 * result in the log tail being moved forward in the journal
2612 * on disk because log writes can still be taking place. Hence
2613 * unpinning the tail will allow the ICREATE intent to be
2614 * removed from the log an recovery will fail with uninitialised
2615 * inode cluster buffers.
2616 */
2617 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
2618 bp->b_flags |= XBF_ASYNC;
2619 xfs_buf_ioend_fail(bp);
2620 return error;
2621 }
2622
2623 if (!clcount)
2624 return -EAGAIN;
2625
2626 XFS_STATS_INC(mp, xs_icluster_flushcnt);
2627 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
2628 return 0;
2629
2630}
2631
2632/* Release an inode. */
2633void
2634xfs_irele(
2635 struct xfs_inode *ip)
2636{
2637 trace_xfs_irele(ip, _RET_IP_);
2638 iput(VFS_I(ip));
2639}
2640
2641/*
2642 * Ensure all commited transactions touching the inode are written to the log.
2643 */
2644int
2645xfs_log_force_inode(
2646 struct xfs_inode *ip)
2647{
2648 xfs_csn_t seq = 0;
2649
2650 xfs_ilock(ip, XFS_ILOCK_SHARED);
2651 if (xfs_ipincount(ip))
2652 seq = ip->i_itemp->ili_commit_seq;
2653 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2654
2655 if (!seq)
2656 return 0;
2657 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
2658}
2659
2660/*
2661 * Grab the exclusive iolock for a data copy from src to dest, making sure to
2662 * abide vfs locking order (lowest pointer value goes first) and breaking the
2663 * layout leases before proceeding. The loop is needed because we cannot call
2664 * the blocking break_layout() with the iolocks held, and therefore have to
2665 * back out both locks.
2666 */
2667static int
2668xfs_iolock_two_inodes_and_break_layout(
2669 struct inode *src,
2670 struct inode *dest)
2671{
2672 int error;
2673
2674 if (src > dest)
2675 swap(src, dest);
2676
2677retry:
2678 /* Wait to break both inodes' layouts before we start locking. */
2679 error = break_layout(src, true);
2680 if (error)
2681 return error;
2682 if (src != dest) {
2683 error = break_layout(dest, true);
2684 if (error)
2685 return error;
2686 }
2687
2688 /* Lock one inode and make sure nobody got in and leased it. */
2689 inode_lock(src);
2690 error = break_layout(src, false);
2691 if (error) {
2692 inode_unlock(src);
2693 if (error == -EWOULDBLOCK)
2694 goto retry;
2695 return error;
2696 }
2697
2698 if (src == dest)
2699 return 0;
2700
2701 /* Lock the other inode and make sure nobody got in and leased it. */
2702 inode_lock_nested(dest, I_MUTEX_NONDIR2);
2703 error = break_layout(dest, false);
2704 if (error) {
2705 inode_unlock(src);
2706 inode_unlock(dest);
2707 if (error == -EWOULDBLOCK)
2708 goto retry;
2709 return error;
2710 }
2711
2712 return 0;
2713}
2714
2715static int
2716xfs_mmaplock_two_inodes_and_break_dax_layout(
2717 struct xfs_inode *ip1,
2718 struct xfs_inode *ip2)
2719{
2720 int error;
2721 bool retry;
2722 struct page *page;
2723
2724 if (ip1->i_ino > ip2->i_ino)
2725 swap(ip1, ip2);
2726
2727again:
2728 retry = false;
2729 /* Lock the first inode */
2730 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
2731 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
2732 if (error || retry) {
2733 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2734 if (error == 0 && retry)
2735 goto again;
2736 return error;
2737 }
2738
2739 if (ip1 == ip2)
2740 return 0;
2741
2742 /* Nested lock the second inode */
2743 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
2744 /*
2745 * We cannot use xfs_break_dax_layouts() directly here because it may
2746 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
2747 * for this nested lock case.
2748 */
2749 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
2750 if (page && page_ref_count(page) != 1) {
2751 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2752 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2753 goto again;
2754 }
2755
2756 return 0;
2757}
2758
2759/*
2760 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
2761 * mmap activity.
2762 */
2763int
2764xfs_ilock2_io_mmap(
2765 struct xfs_inode *ip1,
2766 struct xfs_inode *ip2)
2767{
2768 int ret;
2769
2770 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
2771 if (ret)
2772 return ret;
2773
2774 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2775 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
2776 if (ret) {
2777 inode_unlock(VFS_I(ip2));
2778 if (ip1 != ip2)
2779 inode_unlock(VFS_I(ip1));
2780 return ret;
2781 }
2782 } else
2783 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
2784 VFS_I(ip2)->i_mapping);
2785
2786 return 0;
2787}
2788
2789/* Unlock both inodes to allow IO and mmap activity. */
2790void
2791xfs_iunlock2_io_mmap(
2792 struct xfs_inode *ip1,
2793 struct xfs_inode *ip2)
2794{
2795 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
2796 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2797 if (ip1 != ip2)
2798 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
2799 } else
2800 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
2801 VFS_I(ip2)->i_mapping);
2802
2803 inode_unlock(VFS_I(ip2));
2804 if (ip1 != ip2)
2805 inode_unlock(VFS_I(ip1));
2806}
2807
2808/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
2809void
2810xfs_iunlock2_remapping(
2811 struct xfs_inode *ip1,
2812 struct xfs_inode *ip2)
2813{
2814 xfs_iflags_clear(ip1, XFS_IREMAPPING);
2815
2816 if (ip1 != ip2)
2817 xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
2818 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
2819
2820 if (ip1 != ip2)
2821 inode_unlock_shared(VFS_I(ip1));
2822 inode_unlock(VFS_I(ip2));
2823}
2824
2825/*
2826 * Reload the incore inode list for this inode. Caller should ensure that
2827 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
2828 * preventing other threads from executing.
2829 */
2830int
2831xfs_inode_reload_unlinked_bucket(
2832 struct xfs_trans *tp,
2833 struct xfs_inode *ip)
2834{
2835 struct xfs_mount *mp = tp->t_mountp;
2836 struct xfs_buf *agibp;
2837 struct xfs_agi *agi;
2838 struct xfs_perag *pag;
2839 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2840 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2841 xfs_agino_t prev_agino, next_agino;
2842 unsigned int bucket;
2843 bool foundit = false;
2844 int error;
2845
2846 /* Grab the first inode in the list */
2847 pag = xfs_perag_get(mp, agno);
2848 error = xfs_ialloc_read_agi(pag, tp, 0, &agibp);
2849 xfs_perag_put(pag);
2850 if (error)
2851 return error;
2852
2853 /*
2854 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
2855 * incore unlinked list pointers for this inode. Check once more to
2856 * see if we raced with anyone else to reload the unlinked list.
2857 */
2858 if (!xfs_inode_unlinked_incomplete(ip)) {
2859 foundit = true;
2860 goto out_agibp;
2861 }
2862
2863 bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
2864 agi = agibp->b_addr;
2865
2866 trace_xfs_inode_reload_unlinked_bucket(ip);
2867
2868 xfs_info_ratelimited(mp,
2869 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
2870 agino, agno);
2871
2872 prev_agino = NULLAGINO;
2873 next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2874 while (next_agino != NULLAGINO) {
2875 struct xfs_inode *next_ip = NULL;
2876
2877 /* Found this caller's inode, set its backlink. */
2878 if (next_agino == agino) {
2879 next_ip = ip;
2880 next_ip->i_prev_unlinked = prev_agino;
2881 foundit = true;
2882 goto next_inode;
2883 }
2884
2885 /* Try in-memory lookup first. */
2886 next_ip = xfs_iunlink_lookup(pag, next_agino);
2887 if (next_ip)
2888 goto next_inode;
2889
2890 /* Inode not in memory, try reloading it. */
2891 error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
2892 next_agino);
2893 if (error)
2894 break;
2895
2896 /* Grab the reloaded inode. */
2897 next_ip = xfs_iunlink_lookup(pag, next_agino);
2898 if (!next_ip) {
2899 /* No incore inode at all? We reloaded it... */
2900 ASSERT(next_ip != NULL);
2901 error = -EFSCORRUPTED;
2902 break;
2903 }
2904
2905next_inode:
2906 prev_agino = next_agino;
2907 next_agino = next_ip->i_next_unlinked;
2908 }
2909
2910out_agibp:
2911 xfs_trans_brelse(tp, agibp);
2912 /* Should have found this inode somewhere in the iunlinked bucket. */
2913 if (!error && !foundit)
2914 error = -EFSCORRUPTED;
2915 return error;
2916}
2917
2918/* Decide if this inode is missing its unlinked list and reload it. */
2919int
2920xfs_inode_reload_unlinked(
2921 struct xfs_inode *ip)
2922{
2923 struct xfs_trans *tp;
2924 int error;
2925
2926 error = xfs_trans_alloc_empty(ip->i_mount, &tp);
2927 if (error)
2928 return error;
2929
2930 xfs_ilock(ip, XFS_ILOCK_SHARED);
2931 if (xfs_inode_unlinked_incomplete(ip))
2932 error = xfs_inode_reload_unlinked_bucket(tp, ip);
2933 xfs_iunlock(ip, XFS_ILOCK_SHARED);
2934 xfs_trans_cancel(tp);
2935
2936 return error;
2937}
2938
2939/* Has this inode fork been zapped by repair? */
2940bool
2941xfs_ifork_zapped(
2942 const struct xfs_inode *ip,
2943 int whichfork)
2944{
2945 unsigned int datamask = 0;
2946
2947 switch (whichfork) {
2948 case XFS_DATA_FORK:
2949 switch (ip->i_vnode.i_mode & S_IFMT) {
2950 case S_IFDIR:
2951 datamask = XFS_SICK_INO_DIR_ZAPPED;
2952 break;
2953 case S_IFLNK:
2954 datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
2955 break;
2956 }
2957 return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
2958 case XFS_ATTR_FORK:
2959 return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
2960 default:
2961 return false;
2962 }
2963}
2964
2965/* Compute the number of data and realtime blocks used by a file. */
2966void
2967xfs_inode_count_blocks(
2968 struct xfs_trans *tp,
2969 struct xfs_inode *ip,
2970 xfs_filblks_t *dblocks,
2971 xfs_filblks_t *rblocks)
2972{
2973 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
2974
2975 *rblocks = 0;
2976 if (XFS_IS_REALTIME_INODE(ip))
2977 xfs_bmap_count_leaves(ifp, rblocks);
2978 *dblocks = ip->i_nblocks - *rblocks;
2979}
2980
2981static void
2982xfs_wait_dax_page(
2983 struct inode *inode)
2984{
2985 struct xfs_inode *ip = XFS_I(inode);
2986
2987 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
2988 schedule();
2989 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
2990}
2991
2992int
2993xfs_break_dax_layouts(
2994 struct inode *inode,
2995 bool *retry)
2996{
2997 struct page *page;
2998
2999 xfs_assert_ilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL);
3000
3001 page = dax_layout_busy_page(inode->i_mapping);
3002 if (!page)
3003 return 0;
3004
3005 *retry = true;
3006 return ___wait_var_event(&page->_refcount,
3007 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
3008 0, 0, xfs_wait_dax_page(inode));
3009}
3010
3011int
3012xfs_break_layouts(
3013 struct inode *inode,
3014 uint *iolock,
3015 enum layout_break_reason reason)
3016{
3017 bool retry;
3018 int error;
3019
3020 xfs_assert_ilocked(XFS_I(inode), XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL);
3021
3022 do {
3023 retry = false;
3024 switch (reason) {
3025 case BREAK_UNMAP:
3026 error = xfs_break_dax_layouts(inode, &retry);
3027 if (error || retry)
3028 break;
3029 fallthrough;
3030 case BREAK_WRITE:
3031 error = xfs_break_leased_layouts(inode, iolock, &retry);
3032 break;
3033 default:
3034 WARN_ON_ONCE(1);
3035 error = -EINVAL;
3036 }
3037 } while (error == 0 && retry);
3038
3039 return error;
3040}
3041
3042/* Returns the size of fundamental allocation unit for a file, in bytes. */
3043unsigned int
3044xfs_inode_alloc_unitsize(
3045 struct xfs_inode *ip)
3046{
3047 unsigned int blocks = 1;
3048
3049 if (XFS_IS_REALTIME_INODE(ip))
3050 blocks = ip->i_mount->m_sb.sb_rextsize;
3051
3052 return XFS_FSB_TO_B(ip->i_mount, blocks);
3053}
3054
3055/* Should we always be using copy on write for file writes? */
3056bool
3057xfs_is_always_cow_inode(
3058 const struct xfs_inode *ip)
3059{
3060 return ip->i_mount->m_always_cow && xfs_has_reflink(ip->i_mount);
3061}