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 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
207 else if (lock_flags & XFS_ILOCK_SHARED)
208 mraccess_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 (!mrtryupdate(&ip->i_lock))
250 goto out_undo_mmaplock;
251 } else if (lock_flags & XFS_ILOCK_SHARED) {
252 if (!mrtryaccess(&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 mrunlock_excl(&ip->i_lock);
302 else if (lock_flags & XFS_ILOCK_SHARED)
303 mrunlock_shared(&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 mrdemote(&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
331#if defined(DEBUG) || defined(XFS_WARN)
332static inline bool
333__xfs_rwsem_islocked(
334 struct rw_semaphore *rwsem,
335 bool shared)
336{
337 if (!debug_locks)
338 return rwsem_is_locked(rwsem);
339
340 if (!shared)
341 return lockdep_is_held_type(rwsem, 0);
342
343 /*
344 * We are checking that the lock is held at least in shared
345 * mode but don't care that it might be held exclusively
346 * (i.e. shared | excl). Hence we check if the lock is held
347 * in any mode rather than an explicit shared mode.
348 */
349 return lockdep_is_held_type(rwsem, -1);
350}
351
352bool
353xfs_isilocked(
354 struct xfs_inode *ip,
355 uint lock_flags)
356{
357 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
358 if (!(lock_flags & XFS_ILOCK_SHARED))
359 return !!ip->i_lock.mr_writer;
360 return rwsem_is_locked(&ip->i_lock.mr_lock);
361 }
362
363 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
364 return __xfs_rwsem_islocked(&VFS_I(ip)->i_mapping->invalidate_lock,
365 (lock_flags & XFS_MMAPLOCK_SHARED));
366 }
367
368 if (lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) {
369 return __xfs_rwsem_islocked(&VFS_I(ip)->i_rwsem,
370 (lock_flags & XFS_IOLOCK_SHARED));
371 }
372
373 ASSERT(0);
374 return false;
375}
376#endif
377
378/*
379 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
380 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
381 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
382 * errors and warnings.
383 */
384#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
385static bool
386xfs_lockdep_subclass_ok(
387 int subclass)
388{
389 return subclass < MAX_LOCKDEP_SUBCLASSES;
390}
391#else
392#define xfs_lockdep_subclass_ok(subclass) (true)
393#endif
394
395/*
396 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
397 * value. This can be called for any type of inode lock combination, including
398 * parent locking. Care must be taken to ensure we don't overrun the subclass
399 * storage fields in the class mask we build.
400 */
401static inline uint
402xfs_lock_inumorder(
403 uint lock_mode,
404 uint subclass)
405{
406 uint class = 0;
407
408 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
409 XFS_ILOCK_RTSUM)));
410 ASSERT(xfs_lockdep_subclass_ok(subclass));
411
412 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
413 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
414 class += subclass << XFS_IOLOCK_SHIFT;
415 }
416
417 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
418 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
419 class += subclass << XFS_MMAPLOCK_SHIFT;
420 }
421
422 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
423 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
424 class += subclass << XFS_ILOCK_SHIFT;
425 }
426
427 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
428}
429
430/*
431 * The following routine will lock n inodes in exclusive mode. We assume the
432 * caller calls us with the inodes in i_ino order.
433 *
434 * We need to detect deadlock where an inode that we lock is in the AIL and we
435 * start waiting for another inode that is locked by a thread in a long running
436 * transaction (such as truncate). This can result in deadlock since the long
437 * running trans might need to wait for the inode we just locked in order to
438 * push the tail and free space in the log.
439 *
440 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
441 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
442 * lock more than one at a time, lockdep will report false positives saying we
443 * have violated locking orders.
444 */
445static void
446xfs_lock_inodes(
447 struct xfs_inode **ips,
448 int inodes,
449 uint lock_mode)
450{
451 int attempts = 0;
452 uint i;
453 int j;
454 bool try_lock;
455 struct xfs_log_item *lp;
456
457 /*
458 * Currently supports between 2 and 5 inodes with exclusive locking. We
459 * support an arbitrary depth of locking here, but absolute limits on
460 * inodes depend on the type of locking and the limits placed by
461 * lockdep annotations in xfs_lock_inumorder. These are all checked by
462 * the asserts.
463 */
464 ASSERT(ips && inodes >= 2 && inodes <= 5);
465 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
466 XFS_ILOCK_EXCL));
467 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
468 XFS_ILOCK_SHARED)));
469 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
470 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
471 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
472 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
473
474 if (lock_mode & XFS_IOLOCK_EXCL) {
475 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
476 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
477 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
478
479again:
480 try_lock = false;
481 i = 0;
482 for (; i < inodes; i++) {
483 ASSERT(ips[i]);
484
485 if (i && (ips[i] == ips[i - 1])) /* Already locked */
486 continue;
487
488 /*
489 * If try_lock is not set yet, make sure all locked inodes are
490 * not in the AIL. If any are, set try_lock to be used later.
491 */
492 if (!try_lock) {
493 for (j = (i - 1); j >= 0 && !try_lock; j--) {
494 lp = &ips[j]->i_itemp->ili_item;
495 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
496 try_lock = true;
497 }
498 }
499
500 /*
501 * If any of the previous locks we have locked is in the AIL,
502 * we must TRY to get the second and subsequent locks. If
503 * we can't get any, we must release all we have
504 * and try again.
505 */
506 if (!try_lock) {
507 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
508 continue;
509 }
510
511 /* try_lock means we have an inode locked that is in the AIL. */
512 ASSERT(i != 0);
513 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
514 continue;
515
516 /*
517 * Unlock all previous guys and try again. xfs_iunlock will try
518 * to push the tail if the inode is in the AIL.
519 */
520 attempts++;
521 for (j = i - 1; j >= 0; j--) {
522 /*
523 * Check to see if we've already unlocked this one. Not
524 * the first one going back, and the inode ptr is the
525 * same.
526 */
527 if (j != (i - 1) && ips[j] == ips[j + 1])
528 continue;
529
530 xfs_iunlock(ips[j], lock_mode);
531 }
532
533 if ((attempts % 5) == 0) {
534 delay(1); /* Don't just spin the CPU */
535 }
536 goto again;
537 }
538}
539
540/*
541 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
542 * mmaplock must be double-locked separately since we use i_rwsem and
543 * invalidate_lock for that. We now support taking one lock EXCL and the
544 * other SHARED.
545 */
546void
547xfs_lock_two_inodes(
548 struct xfs_inode *ip0,
549 uint ip0_mode,
550 struct xfs_inode *ip1,
551 uint ip1_mode)
552{
553 int attempts = 0;
554 struct xfs_log_item *lp;
555
556 ASSERT(hweight32(ip0_mode) == 1);
557 ASSERT(hweight32(ip1_mode) == 1);
558 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
559 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
560 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
561 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
562 ASSERT(ip0->i_ino != ip1->i_ino);
563
564 if (ip0->i_ino > ip1->i_ino) {
565 swap(ip0, ip1);
566 swap(ip0_mode, ip1_mode);
567 }
568
569 again:
570 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
571
572 /*
573 * If the first lock we have locked is in the AIL, we must TRY to get
574 * the second lock. If we can't get it, we must release the first one
575 * and try again.
576 */
577 lp = &ip0->i_itemp->ili_item;
578 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
579 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
580 xfs_iunlock(ip0, ip0_mode);
581 if ((++attempts % 5) == 0)
582 delay(1); /* Don't just spin the CPU */
583 goto again;
584 }
585 } else {
586 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
587 }
588}
589
590uint
591xfs_ip2xflags(
592 struct xfs_inode *ip)
593{
594 uint flags = 0;
595
596 if (ip->i_diflags & XFS_DIFLAG_ANY) {
597 if (ip->i_diflags & XFS_DIFLAG_REALTIME)
598 flags |= FS_XFLAG_REALTIME;
599 if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
600 flags |= FS_XFLAG_PREALLOC;
601 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
602 flags |= FS_XFLAG_IMMUTABLE;
603 if (ip->i_diflags & XFS_DIFLAG_APPEND)
604 flags |= FS_XFLAG_APPEND;
605 if (ip->i_diflags & XFS_DIFLAG_SYNC)
606 flags |= FS_XFLAG_SYNC;
607 if (ip->i_diflags & XFS_DIFLAG_NOATIME)
608 flags |= FS_XFLAG_NOATIME;
609 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
610 flags |= FS_XFLAG_NODUMP;
611 if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
612 flags |= FS_XFLAG_RTINHERIT;
613 if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
614 flags |= FS_XFLAG_PROJINHERIT;
615 if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
616 flags |= FS_XFLAG_NOSYMLINKS;
617 if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
618 flags |= FS_XFLAG_EXTSIZE;
619 if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
620 flags |= FS_XFLAG_EXTSZINHERIT;
621 if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
622 flags |= FS_XFLAG_NODEFRAG;
623 if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
624 flags |= FS_XFLAG_FILESTREAM;
625 }
626
627 if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
628 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
629 flags |= FS_XFLAG_DAX;
630 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
631 flags |= FS_XFLAG_COWEXTSIZE;
632 }
633
634 if (xfs_inode_has_attr_fork(ip))
635 flags |= FS_XFLAG_HASATTR;
636 return flags;
637}
638
639/*
640 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
641 * is allowed, otherwise it has to be an exact match. If a CI match is found,
642 * ci_name->name will point to a the actual name (caller must free) or
643 * will be set to NULL if an exact match is found.
644 */
645int
646xfs_lookup(
647 struct xfs_inode *dp,
648 const struct xfs_name *name,
649 struct xfs_inode **ipp,
650 struct xfs_name *ci_name)
651{
652 xfs_ino_t inum;
653 int error;
654
655 trace_xfs_lookup(dp, name);
656
657 if (xfs_is_shutdown(dp->i_mount))
658 return -EIO;
659 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
660 return -EIO;
661
662 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
663 if (error)
664 goto out_unlock;
665
666 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
667 if (error)
668 goto out_free_name;
669
670 return 0;
671
672out_free_name:
673 if (ci_name)
674 kmem_free(ci_name->name);
675out_unlock:
676 *ipp = NULL;
677 return error;
678}
679
680/* Propagate di_flags from a parent inode to a child inode. */
681static void
682xfs_inode_inherit_flags(
683 struct xfs_inode *ip,
684 const struct xfs_inode *pip)
685{
686 unsigned int di_flags = 0;
687 xfs_failaddr_t failaddr;
688 umode_t mode = VFS_I(ip)->i_mode;
689
690 if (S_ISDIR(mode)) {
691 if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
692 di_flags |= XFS_DIFLAG_RTINHERIT;
693 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
694 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
695 ip->i_extsize = pip->i_extsize;
696 }
697 if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
698 di_flags |= XFS_DIFLAG_PROJINHERIT;
699 } else if (S_ISREG(mode)) {
700 if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
701 xfs_has_realtime(ip->i_mount))
702 di_flags |= XFS_DIFLAG_REALTIME;
703 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
704 di_flags |= XFS_DIFLAG_EXTSIZE;
705 ip->i_extsize = pip->i_extsize;
706 }
707 }
708 if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
709 xfs_inherit_noatime)
710 di_flags |= XFS_DIFLAG_NOATIME;
711 if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
712 xfs_inherit_nodump)
713 di_flags |= XFS_DIFLAG_NODUMP;
714 if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
715 xfs_inherit_sync)
716 di_flags |= XFS_DIFLAG_SYNC;
717 if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
718 xfs_inherit_nosymlinks)
719 di_flags |= XFS_DIFLAG_NOSYMLINKS;
720 if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
721 xfs_inherit_nodefrag)
722 di_flags |= XFS_DIFLAG_NODEFRAG;
723 if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
724 di_flags |= XFS_DIFLAG_FILESTREAM;
725
726 ip->i_diflags |= di_flags;
727
728 /*
729 * Inode verifiers on older kernels only check that the extent size
730 * hint is an integer multiple of the rt extent size on realtime files.
731 * They did not check the hint alignment on a directory with both
732 * rtinherit and extszinherit flags set. If the misaligned hint is
733 * propagated from a directory into a new realtime file, new file
734 * allocations will fail due to math errors in the rt allocator and/or
735 * trip the verifiers. Validate the hint settings in the new file so
736 * that we don't let broken hints propagate.
737 */
738 failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
739 VFS_I(ip)->i_mode, ip->i_diflags);
740 if (failaddr) {
741 ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
742 XFS_DIFLAG_EXTSZINHERIT);
743 ip->i_extsize = 0;
744 }
745}
746
747/* Propagate di_flags2 from a parent inode to a child inode. */
748static void
749xfs_inode_inherit_flags2(
750 struct xfs_inode *ip,
751 const struct xfs_inode *pip)
752{
753 xfs_failaddr_t failaddr;
754
755 if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
756 ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
757 ip->i_cowextsize = pip->i_cowextsize;
758 }
759 if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
760 ip->i_diflags2 |= XFS_DIFLAG2_DAX;
761
762 /* Don't let invalid cowextsize hints propagate. */
763 failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
764 VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
765 if (failaddr) {
766 ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
767 ip->i_cowextsize = 0;
768 }
769}
770
771/*
772 * Initialise a newly allocated inode and return the in-core inode to the
773 * caller locked exclusively.
774 */
775int
776xfs_init_new_inode(
777 struct mnt_idmap *idmap,
778 struct xfs_trans *tp,
779 struct xfs_inode *pip,
780 xfs_ino_t ino,
781 umode_t mode,
782 xfs_nlink_t nlink,
783 dev_t rdev,
784 prid_t prid,
785 bool init_xattrs,
786 struct xfs_inode **ipp)
787{
788 struct inode *dir = pip ? VFS_I(pip) : NULL;
789 struct xfs_mount *mp = tp->t_mountp;
790 struct xfs_inode *ip;
791 unsigned int flags;
792 int error;
793 struct timespec64 tv;
794 struct inode *inode;
795
796 /*
797 * Protect against obviously corrupt allocation btree records. Later
798 * xfs_iget checks will catch re-allocation of other active in-memory
799 * and on-disk inodes. If we don't catch reallocating the parent inode
800 * here we will deadlock in xfs_iget() so we have to do these checks
801 * first.
802 */
803 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
804 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
805 return -EFSCORRUPTED;
806 }
807
808 /*
809 * Get the in-core inode with the lock held exclusively to prevent
810 * others from looking at until we're done.
811 */
812 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
813 if (error)
814 return error;
815
816 ASSERT(ip != NULL);
817 inode = VFS_I(ip);
818 set_nlink(inode, nlink);
819 inode->i_rdev = rdev;
820 ip->i_projid = prid;
821
822 if (dir && !(dir->i_mode & S_ISGID) && xfs_has_grpid(mp)) {
823 inode_fsuid_set(inode, idmap);
824 inode->i_gid = dir->i_gid;
825 inode->i_mode = mode;
826 } else {
827 inode_init_owner(idmap, inode, dir, mode);
828 }
829
830 /*
831 * If the group ID of the new file does not match the effective group
832 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
833 * (and only if the irix_sgid_inherit compatibility variable is set).
834 */
835 if (irix_sgid_inherit && (inode->i_mode & S_ISGID) &&
836 !vfsgid_in_group_p(i_gid_into_vfsgid(idmap, inode)))
837 inode->i_mode &= ~S_ISGID;
838
839 ip->i_disk_size = 0;
840 ip->i_df.if_nextents = 0;
841 ASSERT(ip->i_nblocks == 0);
842
843 tv = inode_set_ctime_current(inode);
844 inode_set_mtime_to_ts(inode, tv);
845 inode_set_atime_to_ts(inode, tv);
846
847 ip->i_extsize = 0;
848 ip->i_diflags = 0;
849
850 if (xfs_has_v3inodes(mp)) {
851 inode_set_iversion(inode, 1);
852 ip->i_cowextsize = 0;
853 ip->i_crtime = tv;
854 }
855
856 flags = XFS_ILOG_CORE;
857 switch (mode & S_IFMT) {
858 case S_IFIFO:
859 case S_IFCHR:
860 case S_IFBLK:
861 case S_IFSOCK:
862 ip->i_df.if_format = XFS_DINODE_FMT_DEV;
863 flags |= XFS_ILOG_DEV;
864 break;
865 case S_IFREG:
866 case S_IFDIR:
867 if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
868 xfs_inode_inherit_flags(ip, pip);
869 if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
870 xfs_inode_inherit_flags2(ip, pip);
871 fallthrough;
872 case S_IFLNK:
873 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
874 ip->i_df.if_bytes = 0;
875 ip->i_df.if_data = NULL;
876 break;
877 default:
878 ASSERT(0);
879 }
880
881 /*
882 * If we need to create attributes immediately after allocating the
883 * inode, initialise an empty attribute fork right now. We use the
884 * default fork offset for attributes here as we don't know exactly what
885 * size or how many attributes we might be adding. We can do this
886 * safely here because we know the data fork is completely empty and
887 * this saves us from needing to run a separate transaction to set the
888 * fork offset in the immediate future.
889 */
890 if (init_xattrs && xfs_has_attr(mp)) {
891 ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
892 xfs_ifork_init_attr(ip, XFS_DINODE_FMT_EXTENTS, 0);
893 }
894
895 /*
896 * Log the new values stuffed into the inode.
897 */
898 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
899 xfs_trans_log_inode(tp, ip, flags);
900
901 /* now that we have an i_mode we can setup the inode structure */
902 xfs_setup_inode(ip);
903
904 *ipp = ip;
905 return 0;
906}
907
908/*
909 * Decrement the link count on an inode & log the change. If this causes the
910 * link count to go to zero, move the inode to AGI unlinked list so that it can
911 * be freed when the last active reference goes away via xfs_inactive().
912 */
913static int /* error */
914xfs_droplink(
915 xfs_trans_t *tp,
916 xfs_inode_t *ip)
917{
918 if (VFS_I(ip)->i_nlink == 0) {
919 xfs_alert(ip->i_mount,
920 "%s: Attempt to drop inode (%llu) with nlink zero.",
921 __func__, ip->i_ino);
922 return -EFSCORRUPTED;
923 }
924
925 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
926
927 drop_nlink(VFS_I(ip));
928 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
929
930 if (VFS_I(ip)->i_nlink)
931 return 0;
932
933 return xfs_iunlink(tp, ip);
934}
935
936/*
937 * Increment the link count on an inode & log the change.
938 */
939static void
940xfs_bumplink(
941 xfs_trans_t *tp,
942 xfs_inode_t *ip)
943{
944 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
945
946 inc_nlink(VFS_I(ip));
947 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
948}
949
950int
951xfs_create(
952 struct mnt_idmap *idmap,
953 xfs_inode_t *dp,
954 struct xfs_name *name,
955 umode_t mode,
956 dev_t rdev,
957 bool init_xattrs,
958 xfs_inode_t **ipp)
959{
960 int is_dir = S_ISDIR(mode);
961 struct xfs_mount *mp = dp->i_mount;
962 struct xfs_inode *ip = NULL;
963 struct xfs_trans *tp = NULL;
964 int error;
965 bool unlock_dp_on_error = false;
966 prid_t prid;
967 struct xfs_dquot *udqp = NULL;
968 struct xfs_dquot *gdqp = NULL;
969 struct xfs_dquot *pdqp = NULL;
970 struct xfs_trans_res *tres;
971 uint resblks;
972 xfs_ino_t ino;
973
974 trace_xfs_create(dp, name);
975
976 if (xfs_is_shutdown(mp))
977 return -EIO;
978 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
979 return -EIO;
980
981 prid = xfs_get_initial_prid(dp);
982
983 /*
984 * Make sure that we have allocated dquot(s) on disk.
985 */
986 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
987 mapped_fsgid(idmap, &init_user_ns), prid,
988 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
989 &udqp, &gdqp, &pdqp);
990 if (error)
991 return error;
992
993 if (is_dir) {
994 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
995 tres = &M_RES(mp)->tr_mkdir;
996 } else {
997 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
998 tres = &M_RES(mp)->tr_create;
999 }
1000
1001 /*
1002 * Initially assume that the file does not exist and
1003 * reserve the resources for that case. If that is not
1004 * the case we'll drop the one we have and get a more
1005 * appropriate transaction later.
1006 */
1007 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1008 &tp);
1009 if (error == -ENOSPC) {
1010 /* flush outstanding delalloc blocks and retry */
1011 xfs_flush_inodes(mp);
1012 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1013 resblks, &tp);
1014 }
1015 if (error)
1016 goto out_release_dquots;
1017
1018 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1019 unlock_dp_on_error = true;
1020
1021 /*
1022 * A newly created regular or special file just has one directory
1023 * entry pointing to them, but a directory also the "." entry
1024 * pointing to itself.
1025 */
1026 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1027 if (!error)
1028 error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1029 is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1030 if (error)
1031 goto out_trans_cancel;
1032
1033 /*
1034 * Now we join the directory inode to the transaction. We do not do it
1035 * earlier because xfs_dialloc might commit the previous transaction
1036 * (and release all the locks). An error from here on will result in
1037 * the transaction cancel unlocking dp so don't do it explicitly in the
1038 * error path.
1039 */
1040 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1041 unlock_dp_on_error = false;
1042
1043 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1044 resblks - XFS_IALLOC_SPACE_RES(mp));
1045 if (error) {
1046 ASSERT(error != -ENOSPC);
1047 goto out_trans_cancel;
1048 }
1049 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1050 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1051
1052 if (is_dir) {
1053 error = xfs_dir_init(tp, ip, dp);
1054 if (error)
1055 goto out_trans_cancel;
1056
1057 xfs_bumplink(tp, dp);
1058 }
1059
1060 /*
1061 * If this is a synchronous mount, make sure that the
1062 * create transaction goes to disk before returning to
1063 * the user.
1064 */
1065 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1066 xfs_trans_set_sync(tp);
1067
1068 /*
1069 * Attach the dquot(s) to the inodes and modify them incore.
1070 * These ids of the inode couldn't have changed since the new
1071 * inode has been locked ever since it was created.
1072 */
1073 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1074
1075 error = xfs_trans_commit(tp);
1076 if (error)
1077 goto out_release_inode;
1078
1079 xfs_qm_dqrele(udqp);
1080 xfs_qm_dqrele(gdqp);
1081 xfs_qm_dqrele(pdqp);
1082
1083 *ipp = ip;
1084 return 0;
1085
1086 out_trans_cancel:
1087 xfs_trans_cancel(tp);
1088 out_release_inode:
1089 /*
1090 * Wait until after the current transaction is aborted to finish the
1091 * setup of the inode and release the inode. This prevents recursive
1092 * transactions and deadlocks from xfs_inactive.
1093 */
1094 if (ip) {
1095 xfs_finish_inode_setup(ip);
1096 xfs_irele(ip);
1097 }
1098 out_release_dquots:
1099 xfs_qm_dqrele(udqp);
1100 xfs_qm_dqrele(gdqp);
1101 xfs_qm_dqrele(pdqp);
1102
1103 if (unlock_dp_on_error)
1104 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1105 return error;
1106}
1107
1108int
1109xfs_create_tmpfile(
1110 struct mnt_idmap *idmap,
1111 struct xfs_inode *dp,
1112 umode_t mode,
1113 struct xfs_inode **ipp)
1114{
1115 struct xfs_mount *mp = dp->i_mount;
1116 struct xfs_inode *ip = NULL;
1117 struct xfs_trans *tp = NULL;
1118 int error;
1119 prid_t prid;
1120 struct xfs_dquot *udqp = NULL;
1121 struct xfs_dquot *gdqp = NULL;
1122 struct xfs_dquot *pdqp = NULL;
1123 struct xfs_trans_res *tres;
1124 uint resblks;
1125 xfs_ino_t ino;
1126
1127 if (xfs_is_shutdown(mp))
1128 return -EIO;
1129
1130 prid = xfs_get_initial_prid(dp);
1131
1132 /*
1133 * Make sure that we have allocated dquot(s) on disk.
1134 */
1135 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(idmap, &init_user_ns),
1136 mapped_fsgid(idmap, &init_user_ns), prid,
1137 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1138 &udqp, &gdqp, &pdqp);
1139 if (error)
1140 return error;
1141
1142 resblks = XFS_IALLOC_SPACE_RES(mp);
1143 tres = &M_RES(mp)->tr_create_tmpfile;
1144
1145 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1146 &tp);
1147 if (error)
1148 goto out_release_dquots;
1149
1150 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1151 if (!error)
1152 error = xfs_init_new_inode(idmap, tp, dp, ino, mode,
1153 0, 0, prid, false, &ip);
1154 if (error)
1155 goto out_trans_cancel;
1156
1157 if (xfs_has_wsync(mp))
1158 xfs_trans_set_sync(tp);
1159
1160 /*
1161 * Attach the dquot(s) to the inodes and modify them incore.
1162 * These ids of the inode couldn't have changed since the new
1163 * inode has been locked ever since it was created.
1164 */
1165 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1166
1167 error = xfs_iunlink(tp, ip);
1168 if (error)
1169 goto out_trans_cancel;
1170
1171 error = xfs_trans_commit(tp);
1172 if (error)
1173 goto out_release_inode;
1174
1175 xfs_qm_dqrele(udqp);
1176 xfs_qm_dqrele(gdqp);
1177 xfs_qm_dqrele(pdqp);
1178
1179 *ipp = ip;
1180 return 0;
1181
1182 out_trans_cancel:
1183 xfs_trans_cancel(tp);
1184 out_release_inode:
1185 /*
1186 * Wait until after the current transaction is aborted to finish the
1187 * setup of the inode and release the inode. This prevents recursive
1188 * transactions and deadlocks from xfs_inactive.
1189 */
1190 if (ip) {
1191 xfs_finish_inode_setup(ip);
1192 xfs_irele(ip);
1193 }
1194 out_release_dquots:
1195 xfs_qm_dqrele(udqp);
1196 xfs_qm_dqrele(gdqp);
1197 xfs_qm_dqrele(pdqp);
1198
1199 return error;
1200}
1201
1202int
1203xfs_link(
1204 xfs_inode_t *tdp,
1205 xfs_inode_t *sip,
1206 struct xfs_name *target_name)
1207{
1208 xfs_mount_t *mp = tdp->i_mount;
1209 xfs_trans_t *tp;
1210 int error, nospace_error = 0;
1211 int resblks;
1212
1213 trace_xfs_link(tdp, target_name);
1214
1215 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1216
1217 if (xfs_is_shutdown(mp))
1218 return -EIO;
1219 if (xfs_ifork_zapped(tdp, XFS_DATA_FORK))
1220 return -EIO;
1221
1222 error = xfs_qm_dqattach(sip);
1223 if (error)
1224 goto std_return;
1225
1226 error = xfs_qm_dqattach(tdp);
1227 if (error)
1228 goto std_return;
1229
1230 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1231 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
1232 &tp, &nospace_error);
1233 if (error)
1234 goto std_return;
1235
1236 /*
1237 * If we are using project inheritance, we only allow hard link
1238 * creation in our tree when the project IDs are the same; else
1239 * the tree quota mechanism could be circumvented.
1240 */
1241 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1242 tdp->i_projid != sip->i_projid)) {
1243 error = -EXDEV;
1244 goto error_return;
1245 }
1246
1247 if (!resblks) {
1248 error = xfs_dir_canenter(tp, tdp, target_name);
1249 if (error)
1250 goto error_return;
1251 }
1252
1253 /*
1254 * Handle initial link state of O_TMPFILE inode
1255 */
1256 if (VFS_I(sip)->i_nlink == 0) {
1257 struct xfs_perag *pag;
1258
1259 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1260 error = xfs_iunlink_remove(tp, pag, sip);
1261 xfs_perag_put(pag);
1262 if (error)
1263 goto error_return;
1264 }
1265
1266 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1267 resblks);
1268 if (error)
1269 goto error_return;
1270 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1271 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1272
1273 xfs_bumplink(tp, sip);
1274
1275 /*
1276 * If this is a synchronous mount, make sure that the
1277 * link transaction goes to disk before returning to
1278 * the user.
1279 */
1280 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1281 xfs_trans_set_sync(tp);
1282
1283 return xfs_trans_commit(tp);
1284
1285 error_return:
1286 xfs_trans_cancel(tp);
1287 std_return:
1288 if (error == -ENOSPC && nospace_error)
1289 error = nospace_error;
1290 return error;
1291}
1292
1293/* Clear the reflink flag and the cowblocks tag if possible. */
1294static void
1295xfs_itruncate_clear_reflink_flags(
1296 struct xfs_inode *ip)
1297{
1298 struct xfs_ifork *dfork;
1299 struct xfs_ifork *cfork;
1300
1301 if (!xfs_is_reflink_inode(ip))
1302 return;
1303 dfork = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1304 cfork = xfs_ifork_ptr(ip, XFS_COW_FORK);
1305 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1306 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1307 if (cfork->if_bytes == 0)
1308 xfs_inode_clear_cowblocks_tag(ip);
1309}
1310
1311/*
1312 * Free up the underlying blocks past new_size. The new size must be smaller
1313 * than the current size. This routine can be used both for the attribute and
1314 * data fork, and does not modify the inode size, which is left to the caller.
1315 *
1316 * The transaction passed to this routine must have made a permanent log
1317 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1318 * given transaction and start new ones, so make sure everything involved in
1319 * the transaction is tidy before calling here. Some transaction will be
1320 * returned to the caller to be committed. The incoming transaction must
1321 * already include the inode, and both inode locks must be held exclusively.
1322 * The inode must also be "held" within the transaction. On return the inode
1323 * will be "held" within the returned transaction. This routine does NOT
1324 * require any disk space to be reserved for it within the transaction.
1325 *
1326 * If we get an error, we must return with the inode locked and linked into the
1327 * current transaction. This keeps things simple for the higher level code,
1328 * because it always knows that the inode is locked and held in the transaction
1329 * that returns to it whether errors occur or not. We don't mark the inode
1330 * dirty on error so that transactions can be easily aborted if possible.
1331 */
1332int
1333xfs_itruncate_extents_flags(
1334 struct xfs_trans **tpp,
1335 struct xfs_inode *ip,
1336 int whichfork,
1337 xfs_fsize_t new_size,
1338 int flags)
1339{
1340 struct xfs_mount *mp = ip->i_mount;
1341 struct xfs_trans *tp = *tpp;
1342 xfs_fileoff_t first_unmap_block;
1343 int error = 0;
1344
1345 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1346 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1347 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1348 ASSERT(new_size <= XFS_ISIZE(ip));
1349 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1350 ASSERT(ip->i_itemp != NULL);
1351 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1352 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1353
1354 trace_xfs_itruncate_extents_start(ip, new_size);
1355
1356 flags |= xfs_bmapi_aflag(whichfork);
1357
1358 /*
1359 * Since it is possible for space to become allocated beyond
1360 * the end of the file (in a crash where the space is allocated
1361 * but the inode size is not yet updated), simply remove any
1362 * blocks which show up between the new EOF and the maximum
1363 * possible file size.
1364 *
1365 * We have to free all the blocks to the bmbt maximum offset, even if
1366 * the page cache can't scale that far.
1367 */
1368 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1369 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1370 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1371 return 0;
1372 }
1373
1374 error = xfs_bunmapi_range(&tp, ip, flags, first_unmap_block,
1375 XFS_MAX_FILEOFF);
1376 if (error)
1377 goto out;
1378
1379 if (whichfork == XFS_DATA_FORK) {
1380 /* Remove all pending CoW reservations. */
1381 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1382 first_unmap_block, XFS_MAX_FILEOFF, true);
1383 if (error)
1384 goto out;
1385
1386 xfs_itruncate_clear_reflink_flags(ip);
1387 }
1388
1389 /*
1390 * Always re-log the inode so that our permanent transaction can keep
1391 * on rolling it forward in the log.
1392 */
1393 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1394
1395 trace_xfs_itruncate_extents_end(ip, new_size);
1396
1397out:
1398 *tpp = tp;
1399 return error;
1400}
1401
1402int
1403xfs_release(
1404 xfs_inode_t *ip)
1405{
1406 xfs_mount_t *mp = ip->i_mount;
1407 int error = 0;
1408
1409 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1410 return 0;
1411
1412 /* If this is a read-only mount, don't do this (would generate I/O) */
1413 if (xfs_is_readonly(mp))
1414 return 0;
1415
1416 if (!xfs_is_shutdown(mp)) {
1417 int truncated;
1418
1419 /*
1420 * If we previously truncated this file and removed old data
1421 * in the process, we want to initiate "early" writeout on
1422 * the last close. This is an attempt to combat the notorious
1423 * NULL files problem which is particularly noticeable from a
1424 * truncate down, buffered (re-)write (delalloc), followed by
1425 * a crash. What we are effectively doing here is
1426 * significantly reducing the time window where we'd otherwise
1427 * be exposed to that problem.
1428 */
1429 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1430 if (truncated) {
1431 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1432 if (ip->i_delayed_blks > 0) {
1433 error = filemap_flush(VFS_I(ip)->i_mapping);
1434 if (error)
1435 return error;
1436 }
1437 }
1438 }
1439
1440 if (VFS_I(ip)->i_nlink == 0)
1441 return 0;
1442
1443 /*
1444 * If we can't get the iolock just skip truncating the blocks past EOF
1445 * because we could deadlock with the mmap_lock otherwise. We'll get
1446 * another chance to drop them once the last reference to the inode is
1447 * dropped, so we'll never leak blocks permanently.
1448 */
1449 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1450 return 0;
1451
1452 if (xfs_can_free_eofblocks(ip, false)) {
1453 /*
1454 * Check if the inode is being opened, written and closed
1455 * frequently and we have delayed allocation blocks outstanding
1456 * (e.g. streaming writes from the NFS server), truncating the
1457 * blocks past EOF will cause fragmentation to occur.
1458 *
1459 * In this case don't do the truncation, but we have to be
1460 * careful how we detect this case. Blocks beyond EOF show up as
1461 * i_delayed_blks even when the inode is clean, so we need to
1462 * truncate them away first before checking for a dirty release.
1463 * Hence on the first dirty close we will still remove the
1464 * speculative allocation, but after that we will leave it in
1465 * place.
1466 */
1467 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1468 goto out_unlock;
1469
1470 error = xfs_free_eofblocks(ip);
1471 if (error)
1472 goto out_unlock;
1473
1474 /* delalloc blocks after truncation means it really is dirty */
1475 if (ip->i_delayed_blks)
1476 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1477 }
1478
1479out_unlock:
1480 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1481 return error;
1482}
1483
1484/*
1485 * xfs_inactive_truncate
1486 *
1487 * Called to perform a truncate when an inode becomes unlinked.
1488 */
1489STATIC int
1490xfs_inactive_truncate(
1491 struct xfs_inode *ip)
1492{
1493 struct xfs_mount *mp = ip->i_mount;
1494 struct xfs_trans *tp;
1495 int error;
1496
1497 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1498 if (error) {
1499 ASSERT(xfs_is_shutdown(mp));
1500 return error;
1501 }
1502 xfs_ilock(ip, XFS_ILOCK_EXCL);
1503 xfs_trans_ijoin(tp, ip, 0);
1504
1505 /*
1506 * Log the inode size first to prevent stale data exposure in the event
1507 * of a system crash before the truncate completes. See the related
1508 * comment in xfs_vn_setattr_size() for details.
1509 */
1510 ip->i_disk_size = 0;
1511 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1512
1513 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1514 if (error)
1515 goto error_trans_cancel;
1516
1517 ASSERT(ip->i_df.if_nextents == 0);
1518
1519 error = xfs_trans_commit(tp);
1520 if (error)
1521 goto error_unlock;
1522
1523 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1524 return 0;
1525
1526error_trans_cancel:
1527 xfs_trans_cancel(tp);
1528error_unlock:
1529 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1530 return error;
1531}
1532
1533/*
1534 * xfs_inactive_ifree()
1535 *
1536 * Perform the inode free when an inode is unlinked.
1537 */
1538STATIC int
1539xfs_inactive_ifree(
1540 struct xfs_inode *ip)
1541{
1542 struct xfs_mount *mp = ip->i_mount;
1543 struct xfs_trans *tp;
1544 int error;
1545
1546 /*
1547 * We try to use a per-AG reservation for any block needed by the finobt
1548 * tree, but as the finobt feature predates the per-AG reservation
1549 * support a degraded file system might not have enough space for the
1550 * reservation at mount time. In that case try to dip into the reserved
1551 * pool and pray.
1552 *
1553 * Send a warning if the reservation does happen to fail, as the inode
1554 * now remains allocated and sits on the unlinked list until the fs is
1555 * repaired.
1556 */
1557 if (unlikely(mp->m_finobt_nores)) {
1558 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1559 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1560 &tp);
1561 } else {
1562 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1563 }
1564 if (error) {
1565 if (error == -ENOSPC) {
1566 xfs_warn_ratelimited(mp,
1567 "Failed to remove inode(s) from unlinked list. "
1568 "Please free space, unmount and run xfs_repair.");
1569 } else {
1570 ASSERT(xfs_is_shutdown(mp));
1571 }
1572 return error;
1573 }
1574
1575 /*
1576 * We do not hold the inode locked across the entire rolling transaction
1577 * here. We only need to hold it for the first transaction that
1578 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1579 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1580 * here breaks the relationship between cluster buffer invalidation and
1581 * stale inode invalidation on cluster buffer item journal commit
1582 * completion, and can result in leaving dirty stale inodes hanging
1583 * around in memory.
1584 *
1585 * We have no need for serialising this inode operation against other
1586 * operations - we freed the inode and hence reallocation is required
1587 * and that will serialise on reallocating the space the deferops need
1588 * to free. Hence we can unlock the inode on the first commit of
1589 * the transaction rather than roll it right through the deferops. This
1590 * avoids relogging the XFS_ISTALE inode.
1591 *
1592 * We check that xfs_ifree() hasn't grown an internal transaction roll
1593 * by asserting that the inode is still locked when it returns.
1594 */
1595 xfs_ilock(ip, XFS_ILOCK_EXCL);
1596 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1597
1598 error = xfs_ifree(tp, ip);
1599 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1600 if (error) {
1601 /*
1602 * If we fail to free the inode, shut down. The cancel
1603 * might do that, we need to make sure. Otherwise the
1604 * inode might be lost for a long time or forever.
1605 */
1606 if (!xfs_is_shutdown(mp)) {
1607 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1608 __func__, error);
1609 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1610 }
1611 xfs_trans_cancel(tp);
1612 return error;
1613 }
1614
1615 /*
1616 * Credit the quota account(s). The inode is gone.
1617 */
1618 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1619
1620 return xfs_trans_commit(tp);
1621}
1622
1623/*
1624 * Returns true if we need to update the on-disk metadata before we can free
1625 * the memory used by this inode. Updates include freeing post-eof
1626 * preallocations; freeing COW staging extents; and marking the inode free in
1627 * the inobt if it is on the unlinked list.
1628 */
1629bool
1630xfs_inode_needs_inactive(
1631 struct xfs_inode *ip)
1632{
1633 struct xfs_mount *mp = ip->i_mount;
1634 struct xfs_ifork *cow_ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
1635
1636 /*
1637 * If the inode is already free, then there can be nothing
1638 * to clean up here.
1639 */
1640 if (VFS_I(ip)->i_mode == 0)
1641 return false;
1642
1643 /*
1644 * If this is a read-only mount, don't do this (would generate I/O)
1645 * unless we're in log recovery and cleaning the iunlinked list.
1646 */
1647 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1648 return false;
1649
1650 /* If the log isn't running, push inodes straight to reclaim. */
1651 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1652 return false;
1653
1654 /* Metadata inodes require explicit resource cleanup. */
1655 if (xfs_is_metadata_inode(ip))
1656 return false;
1657
1658 /* Want to clean out the cow blocks if there are any. */
1659 if (cow_ifp && cow_ifp->if_bytes > 0)
1660 return true;
1661
1662 /* Unlinked files must be freed. */
1663 if (VFS_I(ip)->i_nlink == 0)
1664 return true;
1665
1666 /*
1667 * This file isn't being freed, so check if there are post-eof blocks
1668 * to free. @force is true because we are evicting an inode from the
1669 * cache. Post-eof blocks must be freed, lest we end up with broken
1670 * free space accounting.
1671 *
1672 * Note: don't bother with iolock here since lockdep complains about
1673 * acquiring it in reclaim context. We have the only reference to the
1674 * inode at this point anyways.
1675 */
1676 return xfs_can_free_eofblocks(ip, true);
1677}
1678
1679/*
1680 * xfs_inactive
1681 *
1682 * This is called when the vnode reference count for the vnode
1683 * goes to zero. If the file has been unlinked, then it must
1684 * now be truncated. Also, we clear all of the read-ahead state
1685 * kept for the inode here since the file is now closed.
1686 */
1687int
1688xfs_inactive(
1689 xfs_inode_t *ip)
1690{
1691 struct xfs_mount *mp;
1692 int error = 0;
1693 int truncate = 0;
1694
1695 /*
1696 * If the inode is already free, then there can be nothing
1697 * to clean up here.
1698 */
1699 if (VFS_I(ip)->i_mode == 0) {
1700 ASSERT(ip->i_df.if_broot_bytes == 0);
1701 goto out;
1702 }
1703
1704 mp = ip->i_mount;
1705 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1706
1707 /*
1708 * If this is a read-only mount, don't do this (would generate I/O)
1709 * unless we're in log recovery and cleaning the iunlinked list.
1710 */
1711 if (xfs_is_readonly(mp) && !xlog_recovery_needed(mp->m_log))
1712 goto out;
1713
1714 /* Metadata inodes require explicit resource cleanup. */
1715 if (xfs_is_metadata_inode(ip))
1716 goto out;
1717
1718 /* Try to clean out the cow blocks if there are any. */
1719 if (xfs_inode_has_cow_data(ip))
1720 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1721
1722 if (VFS_I(ip)->i_nlink != 0) {
1723 /*
1724 * force is true because we are evicting an inode from the
1725 * cache. Post-eof blocks must be freed, lest we end up with
1726 * broken free space accounting.
1727 *
1728 * Note: don't bother with iolock here since lockdep complains
1729 * about acquiring it in reclaim context. We have the only
1730 * reference to the inode at this point anyways.
1731 */
1732 if (xfs_can_free_eofblocks(ip, true))
1733 error = xfs_free_eofblocks(ip);
1734
1735 goto out;
1736 }
1737
1738 if (S_ISREG(VFS_I(ip)->i_mode) &&
1739 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1740 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1741 truncate = 1;
1742
1743 if (xfs_iflags_test(ip, XFS_IQUOTAUNCHECKED)) {
1744 /*
1745 * If this inode is being inactivated during a quotacheck and
1746 * has not yet been scanned by quotacheck, we /must/ remove
1747 * the dquots from the inode before inactivation changes the
1748 * block and inode counts. Most probably this is a result of
1749 * reloading the incore iunlinked list to purge unrecovered
1750 * unlinked inodes.
1751 */
1752 xfs_qm_dqdetach(ip);
1753 } else {
1754 error = xfs_qm_dqattach(ip);
1755 if (error)
1756 goto out;
1757 }
1758
1759 if (S_ISLNK(VFS_I(ip)->i_mode))
1760 error = xfs_inactive_symlink(ip);
1761 else if (truncate)
1762 error = xfs_inactive_truncate(ip);
1763 if (error)
1764 goto out;
1765
1766 /*
1767 * If there are attributes associated with the file then blow them away
1768 * now. The code calls a routine that recursively deconstructs the
1769 * attribute fork. If also blows away the in-core attribute fork.
1770 */
1771 if (xfs_inode_has_attr_fork(ip)) {
1772 error = xfs_attr_inactive(ip);
1773 if (error)
1774 goto out;
1775 }
1776
1777 ASSERT(ip->i_forkoff == 0);
1778
1779 /*
1780 * Free the inode.
1781 */
1782 error = xfs_inactive_ifree(ip);
1783
1784out:
1785 /*
1786 * We're done making metadata updates for this inode, so we can release
1787 * the attached dquots.
1788 */
1789 xfs_qm_dqdetach(ip);
1790 return error;
1791}
1792
1793/*
1794 * In-Core Unlinked List Lookups
1795 * =============================
1796 *
1797 * Every inode is supposed to be reachable from some other piece of metadata
1798 * with the exception of the root directory. Inodes with a connection to a
1799 * file descriptor but not linked from anywhere in the on-disk directory tree
1800 * are collectively known as unlinked inodes, though the filesystem itself
1801 * maintains links to these inodes so that on-disk metadata are consistent.
1802 *
1803 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1804 * header contains a number of buckets that point to an inode, and each inode
1805 * record has a pointer to the next inode in the hash chain. This
1806 * singly-linked list causes scaling problems in the iunlink remove function
1807 * because we must walk that list to find the inode that points to the inode
1808 * being removed from the unlinked hash bucket list.
1809 *
1810 * Hence we keep an in-memory double linked list to link each inode on an
1811 * unlinked list. Because there are 64 unlinked lists per AGI, keeping pointer
1812 * based lists would require having 64 list heads in the perag, one for each
1813 * list. This is expensive in terms of memory (think millions of AGs) and cache
1814 * misses on lookups. Instead, use the fact that inodes on the unlinked list
1815 * must be referenced at the VFS level to keep them on the list and hence we
1816 * have an existence guarantee for inodes on the unlinked list.
1817 *
1818 * Given we have an existence guarantee, we can use lockless inode cache lookups
1819 * to resolve aginos to xfs inodes. This means we only need 8 bytes per inode
1820 * for the double linked unlinked list, and we don't need any extra locking to
1821 * keep the list safe as all manipulations are done under the AGI buffer lock.
1822 * Keeping the list up to date does not require memory allocation, just finding
1823 * the XFS inode and updating the next/prev unlinked list aginos.
1824 */
1825
1826/*
1827 * Find an inode on the unlinked list. This does not take references to the
1828 * inode as we have existence guarantees by holding the AGI buffer lock and that
1829 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1830 * don't find the inode in cache, then let the caller handle the situation.
1831 */
1832static struct xfs_inode *
1833xfs_iunlink_lookup(
1834 struct xfs_perag *pag,
1835 xfs_agino_t agino)
1836{
1837 struct xfs_inode *ip;
1838
1839 rcu_read_lock();
1840 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1841 if (!ip) {
1842 /* Caller can handle inode not being in memory. */
1843 rcu_read_unlock();
1844 return NULL;
1845 }
1846
1847 /*
1848 * Inode in RCU freeing limbo should not happen. Warn about this and
1849 * let the caller handle the failure.
1850 */
1851 if (WARN_ON_ONCE(!ip->i_ino)) {
1852 rcu_read_unlock();
1853 return NULL;
1854 }
1855 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1856 rcu_read_unlock();
1857 return ip;
1858}
1859
1860/*
1861 * Update the prev pointer of the next agino. Returns -ENOLINK if the inode
1862 * is not in cache.
1863 */
1864static int
1865xfs_iunlink_update_backref(
1866 struct xfs_perag *pag,
1867 xfs_agino_t prev_agino,
1868 xfs_agino_t next_agino)
1869{
1870 struct xfs_inode *ip;
1871
1872 /* No update necessary if we are at the end of the list. */
1873 if (next_agino == NULLAGINO)
1874 return 0;
1875
1876 ip = xfs_iunlink_lookup(pag, next_agino);
1877 if (!ip)
1878 return -ENOLINK;
1879
1880 ip->i_prev_unlinked = prev_agino;
1881 return 0;
1882}
1883
1884/*
1885 * Point the AGI unlinked bucket at an inode and log the results. The caller
1886 * is responsible for validating the old value.
1887 */
1888STATIC int
1889xfs_iunlink_update_bucket(
1890 struct xfs_trans *tp,
1891 struct xfs_perag *pag,
1892 struct xfs_buf *agibp,
1893 unsigned int bucket_index,
1894 xfs_agino_t new_agino)
1895{
1896 struct xfs_agi *agi = agibp->b_addr;
1897 xfs_agino_t old_value;
1898 int offset;
1899
1900 ASSERT(xfs_verify_agino_or_null(pag, new_agino));
1901
1902 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
1903 trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
1904 old_value, new_agino);
1905
1906 /*
1907 * We should never find the head of the list already set to the value
1908 * passed in because either we're adding or removing ourselves from the
1909 * head of the list.
1910 */
1911 if (old_value == new_agino) {
1912 xfs_buf_mark_corrupt(agibp);
1913 return -EFSCORRUPTED;
1914 }
1915
1916 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
1917 offset = offsetof(struct xfs_agi, agi_unlinked) +
1918 (sizeof(xfs_agino_t) * bucket_index);
1919 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
1920 return 0;
1921}
1922
1923/*
1924 * Load the inode @next_agino into the cache and set its prev_unlinked pointer
1925 * to @prev_agino. Caller must hold the AGI to synchronize with other changes
1926 * to the unlinked list.
1927 */
1928STATIC int
1929xfs_iunlink_reload_next(
1930 struct xfs_trans *tp,
1931 struct xfs_buf *agibp,
1932 xfs_agino_t prev_agino,
1933 xfs_agino_t next_agino)
1934{
1935 struct xfs_perag *pag = agibp->b_pag;
1936 struct xfs_mount *mp = pag->pag_mount;
1937 struct xfs_inode *next_ip = NULL;
1938 xfs_ino_t ino;
1939 int error;
1940
1941 ASSERT(next_agino != NULLAGINO);
1942
1943#ifdef DEBUG
1944 rcu_read_lock();
1945 next_ip = radix_tree_lookup(&pag->pag_ici_root, next_agino);
1946 ASSERT(next_ip == NULL);
1947 rcu_read_unlock();
1948#endif
1949
1950 xfs_info_ratelimited(mp,
1951 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating recovery.",
1952 next_agino, pag->pag_agno);
1953
1954 /*
1955 * Use an untrusted lookup just to be cautious in case the AGI has been
1956 * corrupted and now points at a free inode. That shouldn't happen,
1957 * but we'd rather shut down now since we're already running in a weird
1958 * situation.
1959 */
1960 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, next_agino);
1961 error = xfs_iget(mp, tp, ino, XFS_IGET_UNTRUSTED, 0, &next_ip);
1962 if (error)
1963 return error;
1964
1965 /* If this is not an unlinked inode, something is very wrong. */
1966 if (VFS_I(next_ip)->i_nlink != 0) {
1967 error = -EFSCORRUPTED;
1968 goto rele;
1969 }
1970
1971 next_ip->i_prev_unlinked = prev_agino;
1972 trace_xfs_iunlink_reload_next(next_ip);
1973rele:
1974 ASSERT(!(VFS_I(next_ip)->i_state & I_DONTCACHE));
1975 if (xfs_is_quotacheck_running(mp) && next_ip)
1976 xfs_iflags_set(next_ip, XFS_IQUOTAUNCHECKED);
1977 xfs_irele(next_ip);
1978 return error;
1979}
1980
1981static int
1982xfs_iunlink_insert_inode(
1983 struct xfs_trans *tp,
1984 struct xfs_perag *pag,
1985 struct xfs_buf *agibp,
1986 struct xfs_inode *ip)
1987{
1988 struct xfs_mount *mp = tp->t_mountp;
1989 struct xfs_agi *agi = agibp->b_addr;
1990 xfs_agino_t next_agino;
1991 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
1992 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
1993 int error;
1994
1995 /*
1996 * Get the index into the agi hash table for the list this inode will
1997 * go on. Make sure the pointer isn't garbage and that this inode
1998 * isn't already on the list.
1999 */
2000 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2001 if (next_agino == agino ||
2002 !xfs_verify_agino_or_null(pag, next_agino)) {
2003 xfs_buf_mark_corrupt(agibp);
2004 return -EFSCORRUPTED;
2005 }
2006
2007 /*
2008 * Update the prev pointer in the next inode to point back to this
2009 * inode.
2010 */
2011 error = xfs_iunlink_update_backref(pag, agino, next_agino);
2012 if (error == -ENOLINK)
2013 error = xfs_iunlink_reload_next(tp, agibp, agino, next_agino);
2014 if (error)
2015 return error;
2016
2017 if (next_agino != NULLAGINO) {
2018 /*
2019 * There is already another inode in the bucket, so point this
2020 * inode to the current head of the list.
2021 */
2022 error = xfs_iunlink_log_inode(tp, ip, pag, next_agino);
2023 if (error)
2024 return error;
2025 ip->i_next_unlinked = next_agino;
2026 }
2027
2028 /* Point the head of the list to point to this inode. */
2029 ip->i_prev_unlinked = NULLAGINO;
2030 return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2031}
2032
2033/*
2034 * This is called when the inode's link count has gone to 0 or we are creating
2035 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2036 *
2037 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2038 * list when the inode is freed.
2039 */
2040STATIC int
2041xfs_iunlink(
2042 struct xfs_trans *tp,
2043 struct xfs_inode *ip)
2044{
2045 struct xfs_mount *mp = tp->t_mountp;
2046 struct xfs_perag *pag;
2047 struct xfs_buf *agibp;
2048 int error;
2049
2050 ASSERT(VFS_I(ip)->i_nlink == 0);
2051 ASSERT(VFS_I(ip)->i_mode != 0);
2052 trace_xfs_iunlink(ip);
2053
2054 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2055
2056 /* Get the agi buffer first. It ensures lock ordering on the list. */
2057 error = xfs_read_agi(pag, tp, &agibp);
2058 if (error)
2059 goto out;
2060
2061 error = xfs_iunlink_insert_inode(tp, pag, agibp, ip);
2062out:
2063 xfs_perag_put(pag);
2064 return error;
2065}
2066
2067static int
2068xfs_iunlink_remove_inode(
2069 struct xfs_trans *tp,
2070 struct xfs_perag *pag,
2071 struct xfs_buf *agibp,
2072 struct xfs_inode *ip)
2073{
2074 struct xfs_mount *mp = tp->t_mountp;
2075 struct xfs_agi *agi = agibp->b_addr;
2076 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2077 xfs_agino_t head_agino;
2078 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2079 int error;
2080
2081 trace_xfs_iunlink_remove(ip);
2082
2083 /*
2084 * Get the index into the agi hash table for the list this inode will
2085 * go on. Make sure the head pointer isn't garbage.
2086 */
2087 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2088 if (!xfs_verify_agino(pag, head_agino)) {
2089 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2090 agi, sizeof(*agi));
2091 return -EFSCORRUPTED;
2092 }
2093
2094 /*
2095 * Set our inode's next_unlinked pointer to NULL and then return
2096 * the old pointer value so that we can update whatever was previous
2097 * to us in the list to point to whatever was next in the list.
2098 */
2099 error = xfs_iunlink_log_inode(tp, ip, pag, NULLAGINO);
2100 if (error)
2101 return error;
2102
2103 /*
2104 * Update the prev pointer in the next inode to point back to previous
2105 * inode in the chain.
2106 */
2107 error = xfs_iunlink_update_backref(pag, ip->i_prev_unlinked,
2108 ip->i_next_unlinked);
2109 if (error == -ENOLINK)
2110 error = xfs_iunlink_reload_next(tp, agibp, ip->i_prev_unlinked,
2111 ip->i_next_unlinked);
2112 if (error)
2113 return error;
2114
2115 if (head_agino != agino) {
2116 struct xfs_inode *prev_ip;
2117
2118 prev_ip = xfs_iunlink_lookup(pag, ip->i_prev_unlinked);
2119 if (!prev_ip)
2120 return -EFSCORRUPTED;
2121
2122 error = xfs_iunlink_log_inode(tp, prev_ip, pag,
2123 ip->i_next_unlinked);
2124 prev_ip->i_next_unlinked = ip->i_next_unlinked;
2125 } else {
2126 /* Point the head of the list to the next unlinked inode. */
2127 error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2128 ip->i_next_unlinked);
2129 }
2130
2131 ip->i_next_unlinked = NULLAGINO;
2132 ip->i_prev_unlinked = 0;
2133 return error;
2134}
2135
2136/*
2137 * Pull the on-disk inode from the AGI unlinked list.
2138 */
2139STATIC int
2140xfs_iunlink_remove(
2141 struct xfs_trans *tp,
2142 struct xfs_perag *pag,
2143 struct xfs_inode *ip)
2144{
2145 struct xfs_buf *agibp;
2146 int error;
2147
2148 trace_xfs_iunlink_remove(ip);
2149
2150 /* Get the agi buffer first. It ensures lock ordering on the list. */
2151 error = xfs_read_agi(pag, tp, &agibp);
2152 if (error)
2153 return error;
2154
2155 return xfs_iunlink_remove_inode(tp, pag, agibp, ip);
2156}
2157
2158/*
2159 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2160 * mark it stale. We should only find clean inodes in this lookup that aren't
2161 * already stale.
2162 */
2163static void
2164xfs_ifree_mark_inode_stale(
2165 struct xfs_perag *pag,
2166 struct xfs_inode *free_ip,
2167 xfs_ino_t inum)
2168{
2169 struct xfs_mount *mp = pag->pag_mount;
2170 struct xfs_inode_log_item *iip;
2171 struct xfs_inode *ip;
2172
2173retry:
2174 rcu_read_lock();
2175 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2176
2177 /* Inode not in memory, nothing to do */
2178 if (!ip) {
2179 rcu_read_unlock();
2180 return;
2181 }
2182
2183 /*
2184 * because this is an RCU protected lookup, we could find a recently
2185 * freed or even reallocated inode during the lookup. We need to check
2186 * under the i_flags_lock for a valid inode here. Skip it if it is not
2187 * valid, the wrong inode or stale.
2188 */
2189 spin_lock(&ip->i_flags_lock);
2190 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2191 goto out_iflags_unlock;
2192
2193 /*
2194 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2195 * other inodes that we did not find in the list attached to the buffer
2196 * and are not already marked stale. If we can't lock it, back off and
2197 * retry.
2198 */
2199 if (ip != free_ip) {
2200 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2201 spin_unlock(&ip->i_flags_lock);
2202 rcu_read_unlock();
2203 delay(1);
2204 goto retry;
2205 }
2206 }
2207 ip->i_flags |= XFS_ISTALE;
2208
2209 /*
2210 * If the inode is flushing, it is already attached to the buffer. All
2211 * we needed to do here is mark the inode stale so buffer IO completion
2212 * will remove it from the AIL.
2213 */
2214 iip = ip->i_itemp;
2215 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2216 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2217 ASSERT(iip->ili_last_fields);
2218 goto out_iunlock;
2219 }
2220
2221 /*
2222 * Inodes not attached to the buffer can be released immediately.
2223 * Everything else has to go through xfs_iflush_abort() on journal
2224 * commit as the flock synchronises removal of the inode from the
2225 * cluster buffer against inode reclaim.
2226 */
2227 if (!iip || list_empty(&iip->ili_item.li_bio_list))
2228 goto out_iunlock;
2229
2230 __xfs_iflags_set(ip, XFS_IFLUSHING);
2231 spin_unlock(&ip->i_flags_lock);
2232 rcu_read_unlock();
2233
2234 /* we have a dirty inode in memory that has not yet been flushed. */
2235 spin_lock(&iip->ili_lock);
2236 iip->ili_last_fields = iip->ili_fields;
2237 iip->ili_fields = 0;
2238 iip->ili_fsync_fields = 0;
2239 spin_unlock(&iip->ili_lock);
2240 ASSERT(iip->ili_last_fields);
2241
2242 if (ip != free_ip)
2243 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2244 return;
2245
2246out_iunlock:
2247 if (ip != free_ip)
2248 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2249out_iflags_unlock:
2250 spin_unlock(&ip->i_flags_lock);
2251 rcu_read_unlock();
2252}
2253
2254/*
2255 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2256 * inodes that are in memory - they all must be marked stale and attached to
2257 * the cluster buffer.
2258 */
2259static int
2260xfs_ifree_cluster(
2261 struct xfs_trans *tp,
2262 struct xfs_perag *pag,
2263 struct xfs_inode *free_ip,
2264 struct xfs_icluster *xic)
2265{
2266 struct xfs_mount *mp = free_ip->i_mount;
2267 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2268 struct xfs_buf *bp;
2269 xfs_daddr_t blkno;
2270 xfs_ino_t inum = xic->first_ino;
2271 int nbufs;
2272 int i, j;
2273 int ioffset;
2274 int error;
2275
2276 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2277
2278 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2279 /*
2280 * The allocation bitmap tells us which inodes of the chunk were
2281 * physically allocated. Skip the cluster if an inode falls into
2282 * a sparse region.
2283 */
2284 ioffset = inum - xic->first_ino;
2285 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2286 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2287 continue;
2288 }
2289
2290 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2291 XFS_INO_TO_AGBNO(mp, inum));
2292
2293 /*
2294 * We obtain and lock the backing buffer first in the process
2295 * here to ensure dirty inodes attached to the buffer remain in
2296 * the flushing state while we mark them stale.
2297 *
2298 * If we scan the in-memory inodes first, then buffer IO can
2299 * complete before we get a lock on it, and hence we may fail
2300 * to mark all the active inodes on the buffer stale.
2301 */
2302 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2303 mp->m_bsize * igeo->blocks_per_cluster,
2304 XBF_UNMAPPED, &bp);
2305 if (error)
2306 return error;
2307
2308 /*
2309 * This buffer may not have been correctly initialised as we
2310 * didn't read it from disk. That's not important because we are
2311 * only using to mark the buffer as stale in the log, and to
2312 * attach stale cached inodes on it. That means it will never be
2313 * dispatched for IO. If it is, we want to know about it, and we
2314 * want it to fail. We can acheive this by adding a write
2315 * verifier to the buffer.
2316 */
2317 bp->b_ops = &xfs_inode_buf_ops;
2318
2319 /*
2320 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2321 * too. This requires lookups, and will skip inodes that we've
2322 * already marked XFS_ISTALE.
2323 */
2324 for (i = 0; i < igeo->inodes_per_cluster; i++)
2325 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2326
2327 xfs_trans_stale_inode_buf(tp, bp);
2328 xfs_trans_binval(tp, bp);
2329 }
2330 return 0;
2331}
2332
2333/*
2334 * This is called to return an inode to the inode free list. The inode should
2335 * already be truncated to 0 length and have no pages associated with it. This
2336 * routine also assumes that the inode is already a part of the transaction.
2337 *
2338 * The on-disk copy of the inode will have been added to the list of unlinked
2339 * inodes in the AGI. We need to remove the inode from that list atomically with
2340 * respect to freeing it here.
2341 */
2342int
2343xfs_ifree(
2344 struct xfs_trans *tp,
2345 struct xfs_inode *ip)
2346{
2347 struct xfs_mount *mp = ip->i_mount;
2348 struct xfs_perag *pag;
2349 struct xfs_icluster xic = { 0 };
2350 struct xfs_inode_log_item *iip = ip->i_itemp;
2351 int error;
2352
2353 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2354 ASSERT(VFS_I(ip)->i_nlink == 0);
2355 ASSERT(ip->i_df.if_nextents == 0);
2356 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2357 ASSERT(ip->i_nblocks == 0);
2358
2359 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2360
2361 /*
2362 * Free the inode first so that we guarantee that the AGI lock is going
2363 * to be taken before we remove the inode from the unlinked list. This
2364 * makes the AGI lock -> unlinked list modification order the same as
2365 * used in O_TMPFILE creation.
2366 */
2367 error = xfs_difree(tp, pag, ip->i_ino, &xic);
2368 if (error)
2369 goto out;
2370
2371 error = xfs_iunlink_remove(tp, pag, ip);
2372 if (error)
2373 goto out;
2374
2375 /*
2376 * Free any local-format data sitting around before we reset the
2377 * data fork to extents format. Note that the attr fork data has
2378 * already been freed by xfs_attr_inactive.
2379 */
2380 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2381 kmem_free(ip->i_df.if_data);
2382 ip->i_df.if_data = NULL;
2383 ip->i_df.if_bytes = 0;
2384 }
2385
2386 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2387 ip->i_diflags = 0;
2388 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2389 ip->i_forkoff = 0; /* mark the attr fork not in use */
2390 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2391 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2392 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2393
2394 /* Don't attempt to replay owner changes for a deleted inode */
2395 spin_lock(&iip->ili_lock);
2396 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2397 spin_unlock(&iip->ili_lock);
2398
2399 /*
2400 * Bump the generation count so no one will be confused
2401 * by reincarnations of this inode.
2402 */
2403 VFS_I(ip)->i_generation++;
2404 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2405
2406 if (xic.deleted)
2407 error = xfs_ifree_cluster(tp, pag, ip, &xic);
2408out:
2409 xfs_perag_put(pag);
2410 return error;
2411}
2412
2413/*
2414 * This is called to unpin an inode. The caller must have the inode locked
2415 * in at least shared mode so that the buffer cannot be subsequently pinned
2416 * once someone is waiting for it to be unpinned.
2417 */
2418static void
2419xfs_iunpin(
2420 struct xfs_inode *ip)
2421{
2422 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2423
2424 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2425
2426 /* Give the log a push to start the unpinning I/O */
2427 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2428
2429}
2430
2431static void
2432__xfs_iunpin_wait(
2433 struct xfs_inode *ip)
2434{
2435 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2436 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2437
2438 xfs_iunpin(ip);
2439
2440 do {
2441 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2442 if (xfs_ipincount(ip))
2443 io_schedule();
2444 } while (xfs_ipincount(ip));
2445 finish_wait(wq, &wait.wq_entry);
2446}
2447
2448void
2449xfs_iunpin_wait(
2450 struct xfs_inode *ip)
2451{
2452 if (xfs_ipincount(ip))
2453 __xfs_iunpin_wait(ip);
2454}
2455
2456/*
2457 * Removing an inode from the namespace involves removing the directory entry
2458 * and dropping the link count on the inode. Removing the directory entry can
2459 * result in locking an AGF (directory blocks were freed) and removing a link
2460 * count can result in placing the inode on an unlinked list which results in
2461 * locking an AGI.
2462 *
2463 * The big problem here is that we have an ordering constraint on AGF and AGI
2464 * locking - inode allocation locks the AGI, then can allocate a new extent for
2465 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2466 * removes the inode from the unlinked list, requiring that we lock the AGI
2467 * first, and then freeing the inode can result in an inode chunk being freed
2468 * and hence freeing disk space requiring that we lock an AGF.
2469 *
2470 * Hence the ordering that is imposed by other parts of the code is AGI before
2471 * AGF. This means we cannot remove the directory entry before we drop the inode
2472 * reference count and put it on the unlinked list as this results in a lock
2473 * order of AGF then AGI, and this can deadlock against inode allocation and
2474 * freeing. Therefore we must drop the link counts before we remove the
2475 * directory entry.
2476 *
2477 * This is still safe from a transactional point of view - it is not until we
2478 * get to xfs_defer_finish() that we have the possibility of multiple
2479 * transactions in this operation. Hence as long as we remove the directory
2480 * entry and drop the link count in the first transaction of the remove
2481 * operation, there are no transactional constraints on the ordering here.
2482 */
2483int
2484xfs_remove(
2485 xfs_inode_t *dp,
2486 struct xfs_name *name,
2487 xfs_inode_t *ip)
2488{
2489 xfs_mount_t *mp = dp->i_mount;
2490 xfs_trans_t *tp = NULL;
2491 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2492 int dontcare;
2493 int error = 0;
2494 uint resblks;
2495
2496 trace_xfs_remove(dp, name);
2497
2498 if (xfs_is_shutdown(mp))
2499 return -EIO;
2500 if (xfs_ifork_zapped(dp, XFS_DATA_FORK))
2501 return -EIO;
2502
2503 error = xfs_qm_dqattach(dp);
2504 if (error)
2505 goto std_return;
2506
2507 error = xfs_qm_dqattach(ip);
2508 if (error)
2509 goto std_return;
2510
2511 /*
2512 * We try to get the real space reservation first, allowing for
2513 * directory btree deletion(s) implying possible bmap insert(s). If we
2514 * can't get the space reservation then we use 0 instead, and avoid the
2515 * bmap btree insert(s) in the directory code by, if the bmap insert
2516 * tries to happen, instead trimming the LAST block from the directory.
2517 *
2518 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
2519 * the directory code can handle a reservationless update and we don't
2520 * want to prevent a user from trying to free space by deleting things.
2521 */
2522 resblks = XFS_REMOVE_SPACE_RES(mp);
2523 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2524 &tp, &dontcare);
2525 if (error) {
2526 ASSERT(error != -ENOSPC);
2527 goto std_return;
2528 }
2529
2530 /*
2531 * If we're removing a directory perform some additional validation.
2532 */
2533 if (is_dir) {
2534 ASSERT(VFS_I(ip)->i_nlink >= 2);
2535 if (VFS_I(ip)->i_nlink != 2) {
2536 error = -ENOTEMPTY;
2537 goto out_trans_cancel;
2538 }
2539 if (!xfs_dir_isempty(ip)) {
2540 error = -ENOTEMPTY;
2541 goto out_trans_cancel;
2542 }
2543
2544 /* Drop the link from ip's "..". */
2545 error = xfs_droplink(tp, dp);
2546 if (error)
2547 goto out_trans_cancel;
2548
2549 /* Drop the "." link from ip to self. */
2550 error = xfs_droplink(tp, ip);
2551 if (error)
2552 goto out_trans_cancel;
2553
2554 /*
2555 * Point the unlinked child directory's ".." entry to the root
2556 * directory to eliminate back-references to inodes that may
2557 * get freed before the child directory is closed. If the fs
2558 * gets shrunk, this can lead to dirent inode validation errors.
2559 */
2560 if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2561 error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2562 tp->t_mountp->m_sb.sb_rootino, 0);
2563 if (error)
2564 goto out_trans_cancel;
2565 }
2566 } else {
2567 /*
2568 * When removing a non-directory we need to log the parent
2569 * inode here. For a directory this is done implicitly
2570 * by the xfs_droplink call for the ".." entry.
2571 */
2572 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2573 }
2574 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2575
2576 /* Drop the link from dp to ip. */
2577 error = xfs_droplink(tp, ip);
2578 if (error)
2579 goto out_trans_cancel;
2580
2581 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2582 if (error) {
2583 ASSERT(error != -ENOENT);
2584 goto out_trans_cancel;
2585 }
2586
2587 /*
2588 * If this is a synchronous mount, make sure that the
2589 * remove transaction goes to disk before returning to
2590 * the user.
2591 */
2592 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2593 xfs_trans_set_sync(tp);
2594
2595 error = xfs_trans_commit(tp);
2596 if (error)
2597 goto std_return;
2598
2599 if (is_dir && xfs_inode_is_filestream(ip))
2600 xfs_filestream_deassociate(ip);
2601
2602 return 0;
2603
2604 out_trans_cancel:
2605 xfs_trans_cancel(tp);
2606 std_return:
2607 return error;
2608}
2609
2610/*
2611 * Enter all inodes for a rename transaction into a sorted array.
2612 */
2613#define __XFS_SORT_INODES 5
2614STATIC void
2615xfs_sort_for_rename(
2616 struct xfs_inode *dp1, /* in: old (source) directory inode */
2617 struct xfs_inode *dp2, /* in: new (target) directory inode */
2618 struct xfs_inode *ip1, /* in: inode of old entry */
2619 struct xfs_inode *ip2, /* in: inode of new entry */
2620 struct xfs_inode *wip, /* in: whiteout inode */
2621 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2622 int *num_inodes) /* in/out: inodes in array */
2623{
2624 int i, j;
2625
2626 ASSERT(*num_inodes == __XFS_SORT_INODES);
2627 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2628
2629 /*
2630 * i_tab contains a list of pointers to inodes. We initialize
2631 * the table here & we'll sort it. We will then use it to
2632 * order the acquisition of the inode locks.
2633 *
2634 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2635 */
2636 i = 0;
2637 i_tab[i++] = dp1;
2638 i_tab[i++] = dp2;
2639 i_tab[i++] = ip1;
2640 if (ip2)
2641 i_tab[i++] = ip2;
2642 if (wip)
2643 i_tab[i++] = wip;
2644 *num_inodes = i;
2645
2646 /*
2647 * Sort the elements via bubble sort. (Remember, there are at
2648 * most 5 elements to sort, so this is adequate.)
2649 */
2650 for (i = 0; i < *num_inodes; i++) {
2651 for (j = 1; j < *num_inodes; j++) {
2652 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2653 struct xfs_inode *temp = i_tab[j];
2654 i_tab[j] = i_tab[j-1];
2655 i_tab[j-1] = temp;
2656 }
2657 }
2658 }
2659}
2660
2661static int
2662xfs_finish_rename(
2663 struct xfs_trans *tp)
2664{
2665 /*
2666 * If this is a synchronous mount, make sure that the rename transaction
2667 * goes to disk before returning to the user.
2668 */
2669 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2670 xfs_trans_set_sync(tp);
2671
2672 return xfs_trans_commit(tp);
2673}
2674
2675/*
2676 * xfs_cross_rename()
2677 *
2678 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2679 */
2680STATIC int
2681xfs_cross_rename(
2682 struct xfs_trans *tp,
2683 struct xfs_inode *dp1,
2684 struct xfs_name *name1,
2685 struct xfs_inode *ip1,
2686 struct xfs_inode *dp2,
2687 struct xfs_name *name2,
2688 struct xfs_inode *ip2,
2689 int spaceres)
2690{
2691 int error = 0;
2692 int ip1_flags = 0;
2693 int ip2_flags = 0;
2694 int dp2_flags = 0;
2695
2696 /* Swap inode number for dirent in first parent */
2697 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2698 if (error)
2699 goto out_trans_abort;
2700
2701 /* Swap inode number for dirent in second parent */
2702 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2703 if (error)
2704 goto out_trans_abort;
2705
2706 /*
2707 * If we're renaming one or more directories across different parents,
2708 * update the respective ".." entries (and link counts) to match the new
2709 * parents.
2710 */
2711 if (dp1 != dp2) {
2712 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2713
2714 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2715 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2716 dp1->i_ino, spaceres);
2717 if (error)
2718 goto out_trans_abort;
2719
2720 /* transfer ip2 ".." reference to dp1 */
2721 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2722 error = xfs_droplink(tp, dp2);
2723 if (error)
2724 goto out_trans_abort;
2725 xfs_bumplink(tp, dp1);
2726 }
2727
2728 /*
2729 * Although ip1 isn't changed here, userspace needs
2730 * to be warned about the change, so that applications
2731 * relying on it (like backup ones), will properly
2732 * notify the change
2733 */
2734 ip1_flags |= XFS_ICHGTIME_CHG;
2735 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2736 }
2737
2738 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2739 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2740 dp2->i_ino, spaceres);
2741 if (error)
2742 goto out_trans_abort;
2743
2744 /* transfer ip1 ".." reference to dp2 */
2745 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2746 error = xfs_droplink(tp, dp1);
2747 if (error)
2748 goto out_trans_abort;
2749 xfs_bumplink(tp, dp2);
2750 }
2751
2752 /*
2753 * Although ip2 isn't changed here, userspace needs
2754 * to be warned about the change, so that applications
2755 * relying on it (like backup ones), will properly
2756 * notify the change
2757 */
2758 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2759 ip2_flags |= XFS_ICHGTIME_CHG;
2760 }
2761 }
2762
2763 if (ip1_flags) {
2764 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2765 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2766 }
2767 if (ip2_flags) {
2768 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2769 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2770 }
2771 if (dp2_flags) {
2772 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2773 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2774 }
2775 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2776 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2777 return xfs_finish_rename(tp);
2778
2779out_trans_abort:
2780 xfs_trans_cancel(tp);
2781 return error;
2782}
2783
2784/*
2785 * xfs_rename_alloc_whiteout()
2786 *
2787 * Return a referenced, unlinked, unlocked inode that can be used as a
2788 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2789 * crash between allocating the inode and linking it into the rename transaction
2790 * recovery will free the inode and we won't leak it.
2791 */
2792static int
2793xfs_rename_alloc_whiteout(
2794 struct mnt_idmap *idmap,
2795 struct xfs_name *src_name,
2796 struct xfs_inode *dp,
2797 struct xfs_inode **wip)
2798{
2799 struct xfs_inode *tmpfile;
2800 struct qstr name;
2801 int error;
2802
2803 error = xfs_create_tmpfile(idmap, dp, S_IFCHR | WHITEOUT_MODE,
2804 &tmpfile);
2805 if (error)
2806 return error;
2807
2808 name.name = src_name->name;
2809 name.len = src_name->len;
2810 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2811 if (error) {
2812 xfs_finish_inode_setup(tmpfile);
2813 xfs_irele(tmpfile);
2814 return error;
2815 }
2816
2817 /*
2818 * Prepare the tmpfile inode as if it were created through the VFS.
2819 * Complete the inode setup and flag it as linkable. nlink is already
2820 * zero, so we can skip the drop_nlink.
2821 */
2822 xfs_setup_iops(tmpfile);
2823 xfs_finish_inode_setup(tmpfile);
2824 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2825
2826 *wip = tmpfile;
2827 return 0;
2828}
2829
2830/*
2831 * xfs_rename
2832 */
2833int
2834xfs_rename(
2835 struct mnt_idmap *idmap,
2836 struct xfs_inode *src_dp,
2837 struct xfs_name *src_name,
2838 struct xfs_inode *src_ip,
2839 struct xfs_inode *target_dp,
2840 struct xfs_name *target_name,
2841 struct xfs_inode *target_ip,
2842 unsigned int flags)
2843{
2844 struct xfs_mount *mp = src_dp->i_mount;
2845 struct xfs_trans *tp;
2846 struct xfs_inode *wip = NULL; /* whiteout inode */
2847 struct xfs_inode *inodes[__XFS_SORT_INODES];
2848 int i;
2849 int num_inodes = __XFS_SORT_INODES;
2850 bool new_parent = (src_dp != target_dp);
2851 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2852 int spaceres;
2853 bool retried = false;
2854 int error, nospace_error = 0;
2855
2856 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2857
2858 if ((flags & RENAME_EXCHANGE) && !target_ip)
2859 return -EINVAL;
2860
2861 /*
2862 * If we are doing a whiteout operation, allocate the whiteout inode
2863 * we will be placing at the target and ensure the type is set
2864 * appropriately.
2865 */
2866 if (flags & RENAME_WHITEOUT) {
2867 error = xfs_rename_alloc_whiteout(idmap, src_name,
2868 target_dp, &wip);
2869 if (error)
2870 return error;
2871
2872 /* setup target dirent info as whiteout */
2873 src_name->type = XFS_DIR3_FT_CHRDEV;
2874 }
2875
2876 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2877 inodes, &num_inodes);
2878
2879retry:
2880 nospace_error = 0;
2881 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2882 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2883 if (error == -ENOSPC) {
2884 nospace_error = error;
2885 spaceres = 0;
2886 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2887 &tp);
2888 }
2889 if (error)
2890 goto out_release_wip;
2891
2892 /*
2893 * Attach the dquots to the inodes
2894 */
2895 error = xfs_qm_vop_rename_dqattach(inodes);
2896 if (error)
2897 goto out_trans_cancel;
2898
2899 /*
2900 * Lock all the participating inodes. Depending upon whether
2901 * the target_name exists in the target directory, and
2902 * whether the target directory is the same as the source
2903 * directory, we can lock from 2 to 5 inodes.
2904 */
2905 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2906
2907 /*
2908 * Join all the inodes to the transaction. From this point on,
2909 * we can rely on either trans_commit or trans_cancel to unlock
2910 * them.
2911 */
2912 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2913 if (new_parent)
2914 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2915 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2916 if (target_ip)
2917 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2918 if (wip)
2919 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
2920
2921 /*
2922 * If we are using project inheritance, we only allow renames
2923 * into our tree when the project IDs are the same; else the
2924 * tree quota mechanism would be circumvented.
2925 */
2926 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2927 target_dp->i_projid != src_ip->i_projid)) {
2928 error = -EXDEV;
2929 goto out_trans_cancel;
2930 }
2931
2932 /* RENAME_EXCHANGE is unique from here on. */
2933 if (flags & RENAME_EXCHANGE)
2934 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
2935 target_dp, target_name, target_ip,
2936 spaceres);
2937
2938 /*
2939 * Try to reserve quota to handle an expansion of the target directory.
2940 * We'll allow the rename to continue in reservationless mode if we hit
2941 * a space usage constraint. If we trigger reservationless mode, save
2942 * the errno if there isn't any free space in the target directory.
2943 */
2944 if (spaceres != 0) {
2945 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2946 0, false);
2947 if (error == -EDQUOT || error == -ENOSPC) {
2948 if (!retried) {
2949 xfs_trans_cancel(tp);
2950 xfs_blockgc_free_quota(target_dp, 0);
2951 retried = true;
2952 goto retry;
2953 }
2954
2955 nospace_error = error;
2956 spaceres = 0;
2957 error = 0;
2958 }
2959 if (error)
2960 goto out_trans_cancel;
2961 }
2962
2963 /*
2964 * Check for expected errors before we dirty the transaction
2965 * so we can return an error without a transaction abort.
2966 */
2967 if (target_ip == NULL) {
2968 /*
2969 * If there's no space reservation, check the entry will
2970 * fit before actually inserting it.
2971 */
2972 if (!spaceres) {
2973 error = xfs_dir_canenter(tp, target_dp, target_name);
2974 if (error)
2975 goto out_trans_cancel;
2976 }
2977 } else {
2978 /*
2979 * If target exists and it's a directory, check that whether
2980 * it can be destroyed.
2981 */
2982 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
2983 (!xfs_dir_isempty(target_ip) ||
2984 (VFS_I(target_ip)->i_nlink > 2))) {
2985 error = -EEXIST;
2986 goto out_trans_cancel;
2987 }
2988 }
2989
2990 /*
2991 * Lock the AGI buffers we need to handle bumping the nlink of the
2992 * whiteout inode off the unlinked list and to handle dropping the
2993 * nlink of the target inode. Per locking order rules, do this in
2994 * increasing AG order and before directory block allocation tries to
2995 * grab AGFs because we grab AGIs before AGFs.
2996 *
2997 * The (vfs) caller must ensure that if src is a directory then
2998 * target_ip is either null or an empty directory.
2999 */
3000 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3001 if (inodes[i] == wip ||
3002 (inodes[i] == target_ip &&
3003 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3004 struct xfs_perag *pag;
3005 struct xfs_buf *bp;
3006
3007 pag = xfs_perag_get(mp,
3008 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
3009 error = xfs_read_agi(pag, tp, &bp);
3010 xfs_perag_put(pag);
3011 if (error)
3012 goto out_trans_cancel;
3013 }
3014 }
3015
3016 /*
3017 * Directory entry creation below may acquire the AGF. Remove
3018 * the whiteout from the unlinked list first to preserve correct
3019 * AGI/AGF locking order. This dirties the transaction so failures
3020 * after this point will abort and log recovery will clean up the
3021 * mess.
3022 *
3023 * For whiteouts, we need to bump the link count on the whiteout
3024 * inode. After this point, we have a real link, clear the tmpfile
3025 * state flag from the inode so it doesn't accidentally get misused
3026 * in future.
3027 */
3028 if (wip) {
3029 struct xfs_perag *pag;
3030
3031 ASSERT(VFS_I(wip)->i_nlink == 0);
3032
3033 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3034 error = xfs_iunlink_remove(tp, pag, wip);
3035 xfs_perag_put(pag);
3036 if (error)
3037 goto out_trans_cancel;
3038
3039 xfs_bumplink(tp, wip);
3040 VFS_I(wip)->i_state &= ~I_LINKABLE;
3041 }
3042
3043 /*
3044 * Set up the target.
3045 */
3046 if (target_ip == NULL) {
3047 /*
3048 * If target does not exist and the rename crosses
3049 * directories, adjust the target directory link count
3050 * to account for the ".." reference from the new entry.
3051 */
3052 error = xfs_dir_createname(tp, target_dp, target_name,
3053 src_ip->i_ino, spaceres);
3054 if (error)
3055 goto out_trans_cancel;
3056
3057 xfs_trans_ichgtime(tp, target_dp,
3058 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3059
3060 if (new_parent && src_is_directory) {
3061 xfs_bumplink(tp, target_dp);
3062 }
3063 } else { /* target_ip != NULL */
3064 /*
3065 * Link the source inode under the target name.
3066 * If the source inode is a directory and we are moving
3067 * it across directories, its ".." entry will be
3068 * inconsistent until we replace that down below.
3069 *
3070 * In case there is already an entry with the same
3071 * name at the destination directory, remove it first.
3072 */
3073 error = xfs_dir_replace(tp, target_dp, target_name,
3074 src_ip->i_ino, spaceres);
3075 if (error)
3076 goto out_trans_cancel;
3077
3078 xfs_trans_ichgtime(tp, target_dp,
3079 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3080
3081 /*
3082 * Decrement the link count on the target since the target
3083 * dir no longer points to it.
3084 */
3085 error = xfs_droplink(tp, target_ip);
3086 if (error)
3087 goto out_trans_cancel;
3088
3089 if (src_is_directory) {
3090 /*
3091 * Drop the link from the old "." entry.
3092 */
3093 error = xfs_droplink(tp, target_ip);
3094 if (error)
3095 goto out_trans_cancel;
3096 }
3097 } /* target_ip != NULL */
3098
3099 /*
3100 * Remove the source.
3101 */
3102 if (new_parent && src_is_directory) {
3103 /*
3104 * Rewrite the ".." entry to point to the new
3105 * directory.
3106 */
3107 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3108 target_dp->i_ino, spaceres);
3109 ASSERT(error != -EEXIST);
3110 if (error)
3111 goto out_trans_cancel;
3112 }
3113
3114 /*
3115 * We always want to hit the ctime on the source inode.
3116 *
3117 * This isn't strictly required by the standards since the source
3118 * inode isn't really being changed, but old unix file systems did
3119 * it and some incremental backup programs won't work without it.
3120 */
3121 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3122 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3123
3124 /*
3125 * Adjust the link count on src_dp. This is necessary when
3126 * renaming a directory, either within one parent when
3127 * the target existed, or across two parent directories.
3128 */
3129 if (src_is_directory && (new_parent || target_ip != NULL)) {
3130
3131 /*
3132 * Decrement link count on src_directory since the
3133 * entry that's moved no longer points to it.
3134 */
3135 error = xfs_droplink(tp, src_dp);
3136 if (error)
3137 goto out_trans_cancel;
3138 }
3139
3140 /*
3141 * For whiteouts, we only need to update the source dirent with the
3142 * inode number of the whiteout inode rather than removing it
3143 * altogether.
3144 */
3145 if (wip)
3146 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3147 spaceres);
3148 else
3149 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3150 spaceres);
3151
3152 if (error)
3153 goto out_trans_cancel;
3154
3155 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3156 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3157 if (new_parent)
3158 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3159
3160 error = xfs_finish_rename(tp);
3161 if (wip)
3162 xfs_irele(wip);
3163 return error;
3164
3165out_trans_cancel:
3166 xfs_trans_cancel(tp);
3167out_release_wip:
3168 if (wip)
3169 xfs_irele(wip);
3170 if (error == -ENOSPC && nospace_error)
3171 error = nospace_error;
3172 return error;
3173}
3174
3175static int
3176xfs_iflush(
3177 struct xfs_inode *ip,
3178 struct xfs_buf *bp)
3179{
3180 struct xfs_inode_log_item *iip = ip->i_itemp;
3181 struct xfs_dinode *dip;
3182 struct xfs_mount *mp = ip->i_mount;
3183 int error;
3184
3185 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3186 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3187 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3188 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3189 ASSERT(iip->ili_item.li_buf == bp);
3190
3191 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3192
3193 /*
3194 * We don't flush the inode if any of the following checks fail, but we
3195 * do still update the log item and attach to the backing buffer as if
3196 * the flush happened. This is a formality to facilitate predictable
3197 * error handling as the caller will shutdown and fail the buffer.
3198 */
3199 error = -EFSCORRUPTED;
3200 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3201 mp, XFS_ERRTAG_IFLUSH_1)) {
3202 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3203 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
3204 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3205 goto flush_out;
3206 }
3207 if (S_ISREG(VFS_I(ip)->i_mode)) {
3208 if (XFS_TEST_ERROR(
3209 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3210 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3211 mp, XFS_ERRTAG_IFLUSH_3)) {
3212 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3213 "%s: Bad regular inode %llu, ptr "PTR_FMT,
3214 __func__, ip->i_ino, ip);
3215 goto flush_out;
3216 }
3217 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3218 if (XFS_TEST_ERROR(
3219 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3220 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3221 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3222 mp, XFS_ERRTAG_IFLUSH_4)) {
3223 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3224 "%s: Bad directory inode %llu, ptr "PTR_FMT,
3225 __func__, ip->i_ino, ip);
3226 goto flush_out;
3227 }
3228 }
3229 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
3230 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3231 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3232 "%s: detected corrupt incore inode %llu, "
3233 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
3234 __func__, ip->i_ino,
3235 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
3236 ip->i_nblocks, ip);
3237 goto flush_out;
3238 }
3239 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3240 mp, XFS_ERRTAG_IFLUSH_6)) {
3241 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3242 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
3243 __func__, ip->i_ino, ip->i_forkoff, ip);
3244 goto flush_out;
3245 }
3246
3247 /*
3248 * Inode item log recovery for v2 inodes are dependent on the flushiter
3249 * count for correct sequencing. We bump the flush iteration count so
3250 * we can detect flushes which postdate a log record during recovery.
3251 * This is redundant as we now log every change and hence this can't
3252 * happen but we need to still do it to ensure backwards compatibility
3253 * with old kernels that predate logging all inode changes.
3254 */
3255 if (!xfs_has_v3inodes(mp))
3256 ip->i_flushiter++;
3257
3258 /*
3259 * If there are inline format data / attr forks attached to this inode,
3260 * make sure they are not corrupt.
3261 */
3262 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3263 xfs_ifork_verify_local_data(ip))
3264 goto flush_out;
3265 if (xfs_inode_has_attr_fork(ip) &&
3266 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
3267 xfs_ifork_verify_local_attr(ip))
3268 goto flush_out;
3269
3270 /*
3271 * Copy the dirty parts of the inode into the on-disk inode. We always
3272 * copy out the core of the inode, because if the inode is dirty at all
3273 * the core must be.
3274 */
3275 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3276
3277 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3278 if (!xfs_has_v3inodes(mp)) {
3279 if (ip->i_flushiter == DI_MAX_FLUSH)
3280 ip->i_flushiter = 0;
3281 }
3282
3283 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3284 if (xfs_inode_has_attr_fork(ip))
3285 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3286
3287 /*
3288 * We've recorded everything logged in the inode, so we'd like to clear
3289 * the ili_fields bits so we don't log and flush things unnecessarily.
3290 * However, we can't stop logging all this information until the data
3291 * we've copied into the disk buffer is written to disk. If we did we
3292 * might overwrite the copy of the inode in the log with all the data
3293 * after re-logging only part of it, and in the face of a crash we
3294 * wouldn't have all the data we need to recover.
3295 *
3296 * What we do is move the bits to the ili_last_fields field. When
3297 * logging the inode, these bits are moved back to the ili_fields field.
3298 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3299 * we know that the information those bits represent is permanently on
3300 * disk. As long as the flush completes before the inode is logged
3301 * again, then both ili_fields and ili_last_fields will be cleared.
3302 */
3303 error = 0;
3304flush_out:
3305 spin_lock(&iip->ili_lock);
3306 iip->ili_last_fields = iip->ili_fields;
3307 iip->ili_fields = 0;
3308 iip->ili_fsync_fields = 0;
3309 spin_unlock(&iip->ili_lock);
3310
3311 /*
3312 * Store the current LSN of the inode so that we can tell whether the
3313 * item has moved in the AIL from xfs_buf_inode_iodone().
3314 */
3315 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3316 &iip->ili_item.li_lsn);
3317
3318 /* generate the checksum. */
3319 xfs_dinode_calc_crc(mp, dip);
3320 return error;
3321}
3322
3323/*
3324 * Non-blocking flush of dirty inode metadata into the backing buffer.
3325 *
3326 * The caller must have a reference to the inode and hold the cluster buffer
3327 * locked. The function will walk across all the inodes on the cluster buffer it
3328 * can find and lock without blocking, and flush them to the cluster buffer.
3329 *
3330 * On successful flushing of at least one inode, the caller must write out the
3331 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3332 * the caller needs to release the buffer. On failure, the filesystem will be
3333 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3334 * will be returned.
3335 */
3336int
3337xfs_iflush_cluster(
3338 struct xfs_buf *bp)
3339{
3340 struct xfs_mount *mp = bp->b_mount;
3341 struct xfs_log_item *lip, *n;
3342 struct xfs_inode *ip;
3343 struct xfs_inode_log_item *iip;
3344 int clcount = 0;
3345 int error = 0;
3346
3347 /*
3348 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3349 * will remove itself from the list.
3350 */
3351 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3352 iip = (struct xfs_inode_log_item *)lip;
3353 ip = iip->ili_inode;
3354
3355 /*
3356 * Quick and dirty check to avoid locks if possible.
3357 */
3358 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3359 continue;
3360 if (xfs_ipincount(ip))
3361 continue;
3362
3363 /*
3364 * The inode is still attached to the buffer, which means it is
3365 * dirty but reclaim might try to grab it. Check carefully for
3366 * that, and grab the ilock while still holding the i_flags_lock
3367 * to guarantee reclaim will not be able to reclaim this inode
3368 * once we drop the i_flags_lock.
3369 */
3370 spin_lock(&ip->i_flags_lock);
3371 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3372 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3373 spin_unlock(&ip->i_flags_lock);
3374 continue;
3375 }
3376
3377 /*
3378 * ILOCK will pin the inode against reclaim and prevent
3379 * concurrent transactions modifying the inode while we are
3380 * flushing the inode. If we get the lock, set the flushing
3381 * state before we drop the i_flags_lock.
3382 */
3383 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3384 spin_unlock(&ip->i_flags_lock);
3385 continue;
3386 }
3387 __xfs_iflags_set(ip, XFS_IFLUSHING);
3388 spin_unlock(&ip->i_flags_lock);
3389
3390 /*
3391 * Abort flushing this inode if we are shut down because the
3392 * inode may not currently be in the AIL. This can occur when
3393 * log I/O failure unpins the inode without inserting into the
3394 * AIL, leaving a dirty/unpinned inode attached to the buffer
3395 * that otherwise looks like it should be flushed.
3396 */
3397 if (xlog_is_shutdown(mp->m_log)) {
3398 xfs_iunpin_wait(ip);
3399 xfs_iflush_abort(ip);
3400 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3401 error = -EIO;
3402 continue;
3403 }
3404
3405 /* don't block waiting on a log force to unpin dirty inodes */
3406 if (xfs_ipincount(ip)) {
3407 xfs_iflags_clear(ip, XFS_IFLUSHING);
3408 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3409 continue;
3410 }
3411
3412 if (!xfs_inode_clean(ip))
3413 error = xfs_iflush(ip, bp);
3414 else
3415 xfs_iflags_clear(ip, XFS_IFLUSHING);
3416 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3417 if (error)
3418 break;
3419 clcount++;
3420 }
3421
3422 if (error) {
3423 /*
3424 * Shutdown first so we kill the log before we release this
3425 * buffer. If it is an INODE_ALLOC buffer and pins the tail
3426 * of the log, failing it before the _log_ is shut down can
3427 * result in the log tail being moved forward in the journal
3428 * on disk because log writes can still be taking place. Hence
3429 * unpinning the tail will allow the ICREATE intent to be
3430 * removed from the log an recovery will fail with uninitialised
3431 * inode cluster buffers.
3432 */
3433 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3434 bp->b_flags |= XBF_ASYNC;
3435 xfs_buf_ioend_fail(bp);
3436 return error;
3437 }
3438
3439 if (!clcount)
3440 return -EAGAIN;
3441
3442 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3443 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3444 return 0;
3445
3446}
3447
3448/* Release an inode. */
3449void
3450xfs_irele(
3451 struct xfs_inode *ip)
3452{
3453 trace_xfs_irele(ip, _RET_IP_);
3454 iput(VFS_I(ip));
3455}
3456
3457/*
3458 * Ensure all commited transactions touching the inode are written to the log.
3459 */
3460int
3461xfs_log_force_inode(
3462 struct xfs_inode *ip)
3463{
3464 xfs_csn_t seq = 0;
3465
3466 xfs_ilock(ip, XFS_ILOCK_SHARED);
3467 if (xfs_ipincount(ip))
3468 seq = ip->i_itemp->ili_commit_seq;
3469 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3470
3471 if (!seq)
3472 return 0;
3473 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3474}
3475
3476/*
3477 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3478 * abide vfs locking order (lowest pointer value goes first) and breaking the
3479 * layout leases before proceeding. The loop is needed because we cannot call
3480 * the blocking break_layout() with the iolocks held, and therefore have to
3481 * back out both locks.
3482 */
3483static int
3484xfs_iolock_two_inodes_and_break_layout(
3485 struct inode *src,
3486 struct inode *dest)
3487{
3488 int error;
3489
3490 if (src > dest)
3491 swap(src, dest);
3492
3493retry:
3494 /* Wait to break both inodes' layouts before we start locking. */
3495 error = break_layout(src, true);
3496 if (error)
3497 return error;
3498 if (src != dest) {
3499 error = break_layout(dest, true);
3500 if (error)
3501 return error;
3502 }
3503
3504 /* Lock one inode and make sure nobody got in and leased it. */
3505 inode_lock(src);
3506 error = break_layout(src, false);
3507 if (error) {
3508 inode_unlock(src);
3509 if (error == -EWOULDBLOCK)
3510 goto retry;
3511 return error;
3512 }
3513
3514 if (src == dest)
3515 return 0;
3516
3517 /* Lock the other inode and make sure nobody got in and leased it. */
3518 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3519 error = break_layout(dest, false);
3520 if (error) {
3521 inode_unlock(src);
3522 inode_unlock(dest);
3523 if (error == -EWOULDBLOCK)
3524 goto retry;
3525 return error;
3526 }
3527
3528 return 0;
3529}
3530
3531static int
3532xfs_mmaplock_two_inodes_and_break_dax_layout(
3533 struct xfs_inode *ip1,
3534 struct xfs_inode *ip2)
3535{
3536 int error;
3537 bool retry;
3538 struct page *page;
3539
3540 if (ip1->i_ino > ip2->i_ino)
3541 swap(ip1, ip2);
3542
3543again:
3544 retry = false;
3545 /* Lock the first inode */
3546 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3547 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
3548 if (error || retry) {
3549 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3550 if (error == 0 && retry)
3551 goto again;
3552 return error;
3553 }
3554
3555 if (ip1 == ip2)
3556 return 0;
3557
3558 /* Nested lock the second inode */
3559 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
3560 /*
3561 * We cannot use xfs_break_dax_layouts() directly here because it may
3562 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
3563 * for this nested lock case.
3564 */
3565 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
3566 if (page && page_ref_count(page) != 1) {
3567 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3568 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3569 goto again;
3570 }
3571
3572 return 0;
3573}
3574
3575/*
3576 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3577 * mmap activity.
3578 */
3579int
3580xfs_ilock2_io_mmap(
3581 struct xfs_inode *ip1,
3582 struct xfs_inode *ip2)
3583{
3584 int ret;
3585
3586 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3587 if (ret)
3588 return ret;
3589
3590 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3591 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
3592 if (ret) {
3593 inode_unlock(VFS_I(ip2));
3594 if (ip1 != ip2)
3595 inode_unlock(VFS_I(ip1));
3596 return ret;
3597 }
3598 } else
3599 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
3600 VFS_I(ip2)->i_mapping);
3601
3602 return 0;
3603}
3604
3605/* Unlock both inodes to allow IO and mmap activity. */
3606void
3607xfs_iunlock2_io_mmap(
3608 struct xfs_inode *ip1,
3609 struct xfs_inode *ip2)
3610{
3611 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3612 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3613 if (ip1 != ip2)
3614 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3615 } else
3616 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
3617 VFS_I(ip2)->i_mapping);
3618
3619 inode_unlock(VFS_I(ip2));
3620 if (ip1 != ip2)
3621 inode_unlock(VFS_I(ip1));
3622}
3623
3624/* Drop the MMAPLOCK and the IOLOCK after a remap completes. */
3625void
3626xfs_iunlock2_remapping(
3627 struct xfs_inode *ip1,
3628 struct xfs_inode *ip2)
3629{
3630 xfs_iflags_clear(ip1, XFS_IREMAPPING);
3631
3632 if (ip1 != ip2)
3633 xfs_iunlock(ip1, XFS_MMAPLOCK_SHARED);
3634 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3635
3636 if (ip1 != ip2)
3637 inode_unlock_shared(VFS_I(ip1));
3638 inode_unlock(VFS_I(ip2));
3639}
3640
3641/*
3642 * Reload the incore inode list for this inode. Caller should ensure that
3643 * the link count cannot change, either by taking ILOCK_SHARED or otherwise
3644 * preventing other threads from executing.
3645 */
3646int
3647xfs_inode_reload_unlinked_bucket(
3648 struct xfs_trans *tp,
3649 struct xfs_inode *ip)
3650{
3651 struct xfs_mount *mp = tp->t_mountp;
3652 struct xfs_buf *agibp;
3653 struct xfs_agi *agi;
3654 struct xfs_perag *pag;
3655 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
3656 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
3657 xfs_agino_t prev_agino, next_agino;
3658 unsigned int bucket;
3659 bool foundit = false;
3660 int error;
3661
3662 /* Grab the first inode in the list */
3663 pag = xfs_perag_get(mp, agno);
3664 error = xfs_ialloc_read_agi(pag, tp, &agibp);
3665 xfs_perag_put(pag);
3666 if (error)
3667 return error;
3668
3669 /*
3670 * We've taken ILOCK_SHARED and the AGI buffer lock to stabilize the
3671 * incore unlinked list pointers for this inode. Check once more to
3672 * see if we raced with anyone else to reload the unlinked list.
3673 */
3674 if (!xfs_inode_unlinked_incomplete(ip)) {
3675 foundit = true;
3676 goto out_agibp;
3677 }
3678
3679 bucket = agino % XFS_AGI_UNLINKED_BUCKETS;
3680 agi = agibp->b_addr;
3681
3682 trace_xfs_inode_reload_unlinked_bucket(ip);
3683
3684 xfs_info_ratelimited(mp,
3685 "Found unrecovered unlinked inode 0x%x in AG 0x%x. Initiating list recovery.",
3686 agino, agno);
3687
3688 prev_agino = NULLAGINO;
3689 next_agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3690 while (next_agino != NULLAGINO) {
3691 struct xfs_inode *next_ip = NULL;
3692
3693 /* Found this caller's inode, set its backlink. */
3694 if (next_agino == agino) {
3695 next_ip = ip;
3696 next_ip->i_prev_unlinked = prev_agino;
3697 foundit = true;
3698 goto next_inode;
3699 }
3700
3701 /* Try in-memory lookup first. */
3702 next_ip = xfs_iunlink_lookup(pag, next_agino);
3703 if (next_ip)
3704 goto next_inode;
3705
3706 /* Inode not in memory, try reloading it. */
3707 error = xfs_iunlink_reload_next(tp, agibp, prev_agino,
3708 next_agino);
3709 if (error)
3710 break;
3711
3712 /* Grab the reloaded inode. */
3713 next_ip = xfs_iunlink_lookup(pag, next_agino);
3714 if (!next_ip) {
3715 /* No incore inode at all? We reloaded it... */
3716 ASSERT(next_ip != NULL);
3717 error = -EFSCORRUPTED;
3718 break;
3719 }
3720
3721next_inode:
3722 prev_agino = next_agino;
3723 next_agino = next_ip->i_next_unlinked;
3724 }
3725
3726out_agibp:
3727 xfs_trans_brelse(tp, agibp);
3728 /* Should have found this inode somewhere in the iunlinked bucket. */
3729 if (!error && !foundit)
3730 error = -EFSCORRUPTED;
3731 return error;
3732}
3733
3734/* Decide if this inode is missing its unlinked list and reload it. */
3735int
3736xfs_inode_reload_unlinked(
3737 struct xfs_inode *ip)
3738{
3739 struct xfs_trans *tp;
3740 int error;
3741
3742 error = xfs_trans_alloc_empty(ip->i_mount, &tp);
3743 if (error)
3744 return error;
3745
3746 xfs_ilock(ip, XFS_ILOCK_SHARED);
3747 if (xfs_inode_unlinked_incomplete(ip))
3748 error = xfs_inode_reload_unlinked_bucket(tp, ip);
3749 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3750 xfs_trans_cancel(tp);
3751
3752 return error;
3753}
3754
3755/* Has this inode fork been zapped by repair? */
3756bool
3757xfs_ifork_zapped(
3758 const struct xfs_inode *ip,
3759 int whichfork)
3760{
3761 unsigned int datamask = 0;
3762
3763 switch (whichfork) {
3764 case XFS_DATA_FORK:
3765 switch (ip->i_vnode.i_mode & S_IFMT) {
3766 case S_IFDIR:
3767 datamask = XFS_SICK_INO_DIR_ZAPPED;
3768 break;
3769 case S_IFLNK:
3770 datamask = XFS_SICK_INO_SYMLINK_ZAPPED;
3771 break;
3772 }
3773 return ip->i_sick & (XFS_SICK_INO_BMBTD_ZAPPED | datamask);
3774 case XFS_ATTR_FORK:
3775 return ip->i_sick & XFS_SICK_INO_BMBTA_ZAPPED;
3776 default:
3777 return false;
3778 }
3779}
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include <linux/log2.h>
19#include <linux/iversion.h>
20
21#include "xfs.h"
22#include "xfs_fs.h"
23#include "xfs_shared.h"
24#include "xfs_format.h"
25#include "xfs_log_format.h"
26#include "xfs_trans_resv.h"
27#include "xfs_sb.h"
28#include "xfs_mount.h"
29#include "xfs_defer.h"
30#include "xfs_inode.h"
31#include "xfs_da_format.h"
32#include "xfs_da_btree.h"
33#include "xfs_dir2.h"
34#include "xfs_attr_sf.h"
35#include "xfs_attr.h"
36#include "xfs_trans_space.h"
37#include "xfs_trans.h"
38#include "xfs_buf_item.h"
39#include "xfs_inode_item.h"
40#include "xfs_ialloc.h"
41#include "xfs_bmap.h"
42#include "xfs_bmap_util.h"
43#include "xfs_errortag.h"
44#include "xfs_error.h"
45#include "xfs_quota.h"
46#include "xfs_filestream.h"
47#include "xfs_cksum.h"
48#include "xfs_trace.h"
49#include "xfs_icache.h"
50#include "xfs_symlink.h"
51#include "xfs_trans_priv.h"
52#include "xfs_log.h"
53#include "xfs_bmap_btree.h"
54#include "xfs_reflink.h"
55#include "xfs_dir2_priv.h"
56
57kmem_zone_t *xfs_inode_zone;
58
59/*
60 * Used in xfs_itruncate_extents(). This is the maximum number of extents
61 * freed from a file in a single transaction.
62 */
63#define XFS_ITRUNC_MAX_EXTENTS 2
64
65STATIC int xfs_iflush_int(struct xfs_inode *, struct xfs_buf *);
66STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
67STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
68
69/*
70 * helper function to extract extent size hint from inode
71 */
72xfs_extlen_t
73xfs_get_extsz_hint(
74 struct xfs_inode *ip)
75{
76 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
77 return ip->i_d.di_extsize;
78 if (XFS_IS_REALTIME_INODE(ip))
79 return ip->i_mount->m_sb.sb_rextsize;
80 return 0;
81}
82
83/*
84 * Helper function to extract CoW extent size hint from inode.
85 * Between the extent size hint and the CoW extent size hint, we
86 * return the greater of the two. If the value is zero (automatic),
87 * use the default size.
88 */
89xfs_extlen_t
90xfs_get_cowextsz_hint(
91 struct xfs_inode *ip)
92{
93 xfs_extlen_t a, b;
94
95 a = 0;
96 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
97 a = ip->i_d.di_cowextsize;
98 b = xfs_get_extsz_hint(ip);
99
100 a = max(a, b);
101 if (a == 0)
102 return XFS_DEFAULT_COWEXTSZ_HINT;
103 return a;
104}
105
106/*
107 * These two are wrapper routines around the xfs_ilock() routine used to
108 * centralize some grungy code. They are used in places that wish to lock the
109 * inode solely for reading the extents. The reason these places can't just
110 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
111 * bringing in of the extents from disk for a file in b-tree format. If the
112 * inode is in b-tree format, then we need to lock the inode exclusively until
113 * the extents are read in. Locking it exclusively all the time would limit
114 * our parallelism unnecessarily, though. What we do instead is check to see
115 * if the extents have been read in yet, and only lock the inode exclusively
116 * if they have not.
117 *
118 * The functions return a value which should be given to the corresponding
119 * xfs_iunlock() call.
120 */
121uint
122xfs_ilock_data_map_shared(
123 struct xfs_inode *ip)
124{
125 uint lock_mode = XFS_ILOCK_SHARED;
126
127 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
128 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
129 lock_mode = XFS_ILOCK_EXCL;
130 xfs_ilock(ip, lock_mode);
131 return lock_mode;
132}
133
134uint
135xfs_ilock_attr_map_shared(
136 struct xfs_inode *ip)
137{
138 uint lock_mode = XFS_ILOCK_SHARED;
139
140 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
141 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
142 lock_mode = XFS_ILOCK_EXCL;
143 xfs_ilock(ip, lock_mode);
144 return lock_mode;
145}
146
147/*
148 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
149 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
150 * various combinations of the locks to be obtained.
151 *
152 * The 3 locks should always be ordered so that the IO lock is obtained first,
153 * the mmap lock second and the ilock last in order to prevent deadlock.
154 *
155 * Basic locking order:
156 *
157 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
158 *
159 * mmap_sem locking order:
160 *
161 * i_rwsem -> page lock -> mmap_sem
162 * mmap_sem -> i_mmap_lock -> page_lock
163 *
164 * The difference in mmap_sem locking order mean that we cannot hold the
165 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
166 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
167 * in get_user_pages() to map the user pages into the kernel address space for
168 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
169 * page faults already hold the mmap_sem.
170 *
171 * Hence to serialise fully against both syscall and mmap based IO, we need to
172 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
173 * taken in places where we need to invalidate the page cache in a race
174 * free manner (e.g. truncate, hole punch and other extent manipulation
175 * functions).
176 */
177void
178xfs_ilock(
179 xfs_inode_t *ip,
180 uint lock_flags)
181{
182 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
183
184 /*
185 * You can't set both SHARED and EXCL for the same lock,
186 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
187 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
188 */
189 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
190 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
191 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
192 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
193 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
194 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
195 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
196
197 if (lock_flags & XFS_IOLOCK_EXCL) {
198 down_write_nested(&VFS_I(ip)->i_rwsem,
199 XFS_IOLOCK_DEP(lock_flags));
200 } else if (lock_flags & XFS_IOLOCK_SHARED) {
201 down_read_nested(&VFS_I(ip)->i_rwsem,
202 XFS_IOLOCK_DEP(lock_flags));
203 }
204
205 if (lock_flags & XFS_MMAPLOCK_EXCL)
206 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
207 else if (lock_flags & XFS_MMAPLOCK_SHARED)
208 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
209
210 if (lock_flags & XFS_ILOCK_EXCL)
211 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
212 else if (lock_flags & XFS_ILOCK_SHARED)
213 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
214}
215
216/*
217 * This is just like xfs_ilock(), except that the caller
218 * is guaranteed not to sleep. It returns 1 if it gets
219 * the requested locks and 0 otherwise. If the IO lock is
220 * obtained but the inode lock cannot be, then the IO lock
221 * is dropped before returning.
222 *
223 * ip -- the inode being locked
224 * lock_flags -- this parameter indicates the inode's locks to be
225 * to be locked. See the comment for xfs_ilock() for a list
226 * of valid values.
227 */
228int
229xfs_ilock_nowait(
230 xfs_inode_t *ip,
231 uint lock_flags)
232{
233 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
234
235 /*
236 * You can't set both SHARED and EXCL for the same lock,
237 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
238 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
239 */
240 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
241 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
242 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
243 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
244 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
245 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
246 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
247
248 if (lock_flags & XFS_IOLOCK_EXCL) {
249 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
250 goto out;
251 } else if (lock_flags & XFS_IOLOCK_SHARED) {
252 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
253 goto out;
254 }
255
256 if (lock_flags & XFS_MMAPLOCK_EXCL) {
257 if (!mrtryupdate(&ip->i_mmaplock))
258 goto out_undo_iolock;
259 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
260 if (!mrtryaccess(&ip->i_mmaplock))
261 goto out_undo_iolock;
262 }
263
264 if (lock_flags & XFS_ILOCK_EXCL) {
265 if (!mrtryupdate(&ip->i_lock))
266 goto out_undo_mmaplock;
267 } else if (lock_flags & XFS_ILOCK_SHARED) {
268 if (!mrtryaccess(&ip->i_lock))
269 goto out_undo_mmaplock;
270 }
271 return 1;
272
273out_undo_mmaplock:
274 if (lock_flags & XFS_MMAPLOCK_EXCL)
275 mrunlock_excl(&ip->i_mmaplock);
276 else if (lock_flags & XFS_MMAPLOCK_SHARED)
277 mrunlock_shared(&ip->i_mmaplock);
278out_undo_iolock:
279 if (lock_flags & XFS_IOLOCK_EXCL)
280 up_write(&VFS_I(ip)->i_rwsem);
281 else if (lock_flags & XFS_IOLOCK_SHARED)
282 up_read(&VFS_I(ip)->i_rwsem);
283out:
284 return 0;
285}
286
287/*
288 * xfs_iunlock() is used to drop the inode locks acquired with
289 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
290 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
291 * that we know which locks to drop.
292 *
293 * ip -- the inode being unlocked
294 * lock_flags -- this parameter indicates the inode's locks to be
295 * to be unlocked. See the comment for xfs_ilock() for a list
296 * of valid values for this parameter.
297 *
298 */
299void
300xfs_iunlock(
301 xfs_inode_t *ip,
302 uint lock_flags)
303{
304 /*
305 * You can't set both SHARED and EXCL for the same lock,
306 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
307 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
308 */
309 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
310 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
311 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
312 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
313 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
314 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
315 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
316 ASSERT(lock_flags != 0);
317
318 if (lock_flags & XFS_IOLOCK_EXCL)
319 up_write(&VFS_I(ip)->i_rwsem);
320 else if (lock_flags & XFS_IOLOCK_SHARED)
321 up_read(&VFS_I(ip)->i_rwsem);
322
323 if (lock_flags & XFS_MMAPLOCK_EXCL)
324 mrunlock_excl(&ip->i_mmaplock);
325 else if (lock_flags & XFS_MMAPLOCK_SHARED)
326 mrunlock_shared(&ip->i_mmaplock);
327
328 if (lock_flags & XFS_ILOCK_EXCL)
329 mrunlock_excl(&ip->i_lock);
330 else if (lock_flags & XFS_ILOCK_SHARED)
331 mrunlock_shared(&ip->i_lock);
332
333 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
334}
335
336/*
337 * give up write locks. the i/o lock cannot be held nested
338 * if it is being demoted.
339 */
340void
341xfs_ilock_demote(
342 xfs_inode_t *ip,
343 uint lock_flags)
344{
345 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
346 ASSERT((lock_flags &
347 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
348
349 if (lock_flags & XFS_ILOCK_EXCL)
350 mrdemote(&ip->i_lock);
351 if (lock_flags & XFS_MMAPLOCK_EXCL)
352 mrdemote(&ip->i_mmaplock);
353 if (lock_flags & XFS_IOLOCK_EXCL)
354 downgrade_write(&VFS_I(ip)->i_rwsem);
355
356 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
357}
358
359#if defined(DEBUG) || defined(XFS_WARN)
360int
361xfs_isilocked(
362 xfs_inode_t *ip,
363 uint lock_flags)
364{
365 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
366 if (!(lock_flags & XFS_ILOCK_SHARED))
367 return !!ip->i_lock.mr_writer;
368 return rwsem_is_locked(&ip->i_lock.mr_lock);
369 }
370
371 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
372 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
373 return !!ip->i_mmaplock.mr_writer;
374 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
375 }
376
377 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
378 if (!(lock_flags & XFS_IOLOCK_SHARED))
379 return !debug_locks ||
380 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
381 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
382 }
383
384 ASSERT(0);
385 return 0;
386}
387#endif
388
389/*
390 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
391 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
392 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
393 * errors and warnings.
394 */
395#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
396static bool
397xfs_lockdep_subclass_ok(
398 int subclass)
399{
400 return subclass < MAX_LOCKDEP_SUBCLASSES;
401}
402#else
403#define xfs_lockdep_subclass_ok(subclass) (true)
404#endif
405
406/*
407 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
408 * value. This can be called for any type of inode lock combination, including
409 * parent locking. Care must be taken to ensure we don't overrun the subclass
410 * storage fields in the class mask we build.
411 */
412static inline int
413xfs_lock_inumorder(int lock_mode, int subclass)
414{
415 int class = 0;
416
417 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
418 XFS_ILOCK_RTSUM)));
419 ASSERT(xfs_lockdep_subclass_ok(subclass));
420
421 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
422 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
423 class += subclass << XFS_IOLOCK_SHIFT;
424 }
425
426 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
427 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
428 class += subclass << XFS_MMAPLOCK_SHIFT;
429 }
430
431 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
432 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
433 class += subclass << XFS_ILOCK_SHIFT;
434 }
435
436 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
437}
438
439/*
440 * The following routine will lock n inodes in exclusive mode. We assume the
441 * caller calls us with the inodes in i_ino order.
442 *
443 * We need to detect deadlock where an inode that we lock is in the AIL and we
444 * start waiting for another inode that is locked by a thread in a long running
445 * transaction (such as truncate). This can result in deadlock since the long
446 * running trans might need to wait for the inode we just locked in order to
447 * push the tail and free space in the log.
448 *
449 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
450 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
451 * lock more than one at a time, lockdep will report false positives saying we
452 * have violated locking orders.
453 */
454static void
455xfs_lock_inodes(
456 xfs_inode_t **ips,
457 int inodes,
458 uint lock_mode)
459{
460 int attempts = 0, i, j, try_lock;
461 xfs_log_item_t *lp;
462
463 /*
464 * Currently supports between 2 and 5 inodes with exclusive locking. We
465 * support an arbitrary depth of locking here, but absolute limits on
466 * inodes depend on the the type of locking and the limits placed by
467 * lockdep annotations in xfs_lock_inumorder. These are all checked by
468 * the asserts.
469 */
470 ASSERT(ips && inodes >= 2 && inodes <= 5);
471 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
472 XFS_ILOCK_EXCL));
473 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
474 XFS_ILOCK_SHARED)));
475 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
476 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
477 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
478 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
479
480 if (lock_mode & XFS_IOLOCK_EXCL) {
481 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
482 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
483 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
484
485 try_lock = 0;
486 i = 0;
487again:
488 for (; i < inodes; i++) {
489 ASSERT(ips[i]);
490
491 if (i && (ips[i] == ips[i - 1])) /* Already locked */
492 continue;
493
494 /*
495 * If try_lock is not set yet, make sure all locked inodes are
496 * not in the AIL. If any are, set try_lock to be used later.
497 */
498 if (!try_lock) {
499 for (j = (i - 1); j >= 0 && !try_lock; j--) {
500 lp = (xfs_log_item_t *)ips[j]->i_itemp;
501 if (lp && (lp->li_flags & XFS_LI_IN_AIL))
502 try_lock++;
503 }
504 }
505
506 /*
507 * If any of the previous locks we have locked is in the AIL,
508 * we must TRY to get the second and subsequent locks. If
509 * we can't get any, we must release all we have
510 * and try again.
511 */
512 if (!try_lock) {
513 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
514 continue;
515 }
516
517 /* try_lock means we have an inode locked that is in the AIL. */
518 ASSERT(i != 0);
519 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
520 continue;
521
522 /*
523 * Unlock all previous guys and try again. xfs_iunlock will try
524 * to push the tail if the inode is in the AIL.
525 */
526 attempts++;
527 for (j = i - 1; j >= 0; j--) {
528 /*
529 * Check to see if we've already unlocked this one. Not
530 * the first one going back, and the inode ptr is the
531 * same.
532 */
533 if (j != (i - 1) && ips[j] == ips[j + 1])
534 continue;
535
536 xfs_iunlock(ips[j], lock_mode);
537 }
538
539 if ((attempts % 5) == 0) {
540 delay(1); /* Don't just spin the CPU */
541 }
542 i = 0;
543 try_lock = 0;
544 goto again;
545 }
546}
547
548/*
549 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
550 * the mmaplock or the ilock, but not more than one type at a time. If we lock
551 * more than one at a time, lockdep will report false positives saying we have
552 * violated locking orders. The iolock must be double-locked separately since
553 * we use i_rwsem for that. We now support taking one lock EXCL and the other
554 * SHARED.
555 */
556void
557xfs_lock_two_inodes(
558 struct xfs_inode *ip0,
559 uint ip0_mode,
560 struct xfs_inode *ip1,
561 uint ip1_mode)
562{
563 struct xfs_inode *temp;
564 uint mode_temp;
565 int attempts = 0;
566 xfs_log_item_t *lp;
567
568 ASSERT(hweight32(ip0_mode) == 1);
569 ASSERT(hweight32(ip1_mode) == 1);
570 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
571 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
572 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
573 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
574 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
575 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
576 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
577 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
578 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
579 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
580
581 ASSERT(ip0->i_ino != ip1->i_ino);
582
583 if (ip0->i_ino > ip1->i_ino) {
584 temp = ip0;
585 ip0 = ip1;
586 ip1 = temp;
587 mode_temp = ip0_mode;
588 ip0_mode = ip1_mode;
589 ip1_mode = mode_temp;
590 }
591
592 again:
593 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
594
595 /*
596 * If the first lock we have locked is in the AIL, we must TRY to get
597 * the second lock. If we can't get it, we must release the first one
598 * and try again.
599 */
600 lp = (xfs_log_item_t *)ip0->i_itemp;
601 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
602 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
603 xfs_iunlock(ip0, ip0_mode);
604 if ((++attempts % 5) == 0)
605 delay(1); /* Don't just spin the CPU */
606 goto again;
607 }
608 } else {
609 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
610 }
611}
612
613void
614__xfs_iflock(
615 struct xfs_inode *ip)
616{
617 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
618 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
619
620 do {
621 prepare_to_wait_exclusive(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
622 if (xfs_isiflocked(ip))
623 io_schedule();
624 } while (!xfs_iflock_nowait(ip));
625
626 finish_wait(wq, &wait.wq_entry);
627}
628
629STATIC uint
630_xfs_dic2xflags(
631 uint16_t di_flags,
632 uint64_t di_flags2,
633 bool has_attr)
634{
635 uint flags = 0;
636
637 if (di_flags & XFS_DIFLAG_ANY) {
638 if (di_flags & XFS_DIFLAG_REALTIME)
639 flags |= FS_XFLAG_REALTIME;
640 if (di_flags & XFS_DIFLAG_PREALLOC)
641 flags |= FS_XFLAG_PREALLOC;
642 if (di_flags & XFS_DIFLAG_IMMUTABLE)
643 flags |= FS_XFLAG_IMMUTABLE;
644 if (di_flags & XFS_DIFLAG_APPEND)
645 flags |= FS_XFLAG_APPEND;
646 if (di_flags & XFS_DIFLAG_SYNC)
647 flags |= FS_XFLAG_SYNC;
648 if (di_flags & XFS_DIFLAG_NOATIME)
649 flags |= FS_XFLAG_NOATIME;
650 if (di_flags & XFS_DIFLAG_NODUMP)
651 flags |= FS_XFLAG_NODUMP;
652 if (di_flags & XFS_DIFLAG_RTINHERIT)
653 flags |= FS_XFLAG_RTINHERIT;
654 if (di_flags & XFS_DIFLAG_PROJINHERIT)
655 flags |= FS_XFLAG_PROJINHERIT;
656 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
657 flags |= FS_XFLAG_NOSYMLINKS;
658 if (di_flags & XFS_DIFLAG_EXTSIZE)
659 flags |= FS_XFLAG_EXTSIZE;
660 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
661 flags |= FS_XFLAG_EXTSZINHERIT;
662 if (di_flags & XFS_DIFLAG_NODEFRAG)
663 flags |= FS_XFLAG_NODEFRAG;
664 if (di_flags & XFS_DIFLAG_FILESTREAM)
665 flags |= FS_XFLAG_FILESTREAM;
666 }
667
668 if (di_flags2 & XFS_DIFLAG2_ANY) {
669 if (di_flags2 & XFS_DIFLAG2_DAX)
670 flags |= FS_XFLAG_DAX;
671 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
672 flags |= FS_XFLAG_COWEXTSIZE;
673 }
674
675 if (has_attr)
676 flags |= FS_XFLAG_HASATTR;
677
678 return flags;
679}
680
681uint
682xfs_ip2xflags(
683 struct xfs_inode *ip)
684{
685 struct xfs_icdinode *dic = &ip->i_d;
686
687 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
688}
689
690/*
691 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
692 * is allowed, otherwise it has to be an exact match. If a CI match is found,
693 * ci_name->name will point to a the actual name (caller must free) or
694 * will be set to NULL if an exact match is found.
695 */
696int
697xfs_lookup(
698 xfs_inode_t *dp,
699 struct xfs_name *name,
700 xfs_inode_t **ipp,
701 struct xfs_name *ci_name)
702{
703 xfs_ino_t inum;
704 int error;
705
706 trace_xfs_lookup(dp, name);
707
708 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
709 return -EIO;
710
711 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
712 if (error)
713 goto out_unlock;
714
715 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
716 if (error)
717 goto out_free_name;
718
719 return 0;
720
721out_free_name:
722 if (ci_name)
723 kmem_free(ci_name->name);
724out_unlock:
725 *ipp = NULL;
726 return error;
727}
728
729/*
730 * Allocate an inode on disk and return a copy of its in-core version.
731 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
732 * appropriately within the inode. The uid and gid for the inode are
733 * set according to the contents of the given cred structure.
734 *
735 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
736 * has a free inode available, call xfs_iget() to obtain the in-core
737 * version of the allocated inode. Finally, fill in the inode and
738 * log its initial contents. In this case, ialloc_context would be
739 * set to NULL.
740 *
741 * If xfs_dialloc() does not have an available inode, it will replenish
742 * its supply by doing an allocation. Since we can only do one
743 * allocation within a transaction without deadlocks, we must commit
744 * the current transaction before returning the inode itself.
745 * In this case, therefore, we will set ialloc_context and return.
746 * The caller should then commit the current transaction, start a new
747 * transaction, and call xfs_ialloc() again to actually get the inode.
748 *
749 * To ensure that some other process does not grab the inode that
750 * was allocated during the first call to xfs_ialloc(), this routine
751 * also returns the [locked] bp pointing to the head of the freelist
752 * as ialloc_context. The caller should hold this buffer across
753 * the commit and pass it back into this routine on the second call.
754 *
755 * If we are allocating quota inodes, we do not have a parent inode
756 * to attach to or associate with (i.e. pip == NULL) because they
757 * are not linked into the directory structure - they are attached
758 * directly to the superblock - and so have no parent.
759 */
760static int
761xfs_ialloc(
762 xfs_trans_t *tp,
763 xfs_inode_t *pip,
764 umode_t mode,
765 xfs_nlink_t nlink,
766 dev_t rdev,
767 prid_t prid,
768 xfs_buf_t **ialloc_context,
769 xfs_inode_t **ipp)
770{
771 struct xfs_mount *mp = tp->t_mountp;
772 xfs_ino_t ino;
773 xfs_inode_t *ip;
774 uint flags;
775 int error;
776 struct timespec tv;
777 struct inode *inode;
778
779 /*
780 * Call the space management code to pick
781 * the on-disk inode to be allocated.
782 */
783 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode,
784 ialloc_context, &ino);
785 if (error)
786 return error;
787 if (*ialloc_context || ino == NULLFSINO) {
788 *ipp = NULL;
789 return 0;
790 }
791 ASSERT(*ialloc_context == NULL);
792
793 /*
794 * Get the in-core inode with the lock held exclusively.
795 * This is because we're setting fields here we need
796 * to prevent others from looking at until we're done.
797 */
798 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
799 XFS_ILOCK_EXCL, &ip);
800 if (error)
801 return error;
802 ASSERT(ip != NULL);
803 inode = VFS_I(ip);
804
805 /*
806 * We always convert v1 inodes to v2 now - we only support filesystems
807 * with >= v2 inode capability, so there is no reason for ever leaving
808 * an inode in v1 format.
809 */
810 if (ip->i_d.di_version == 1)
811 ip->i_d.di_version = 2;
812
813 inode->i_mode = mode;
814 set_nlink(inode, nlink);
815 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
816 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
817 inode->i_rdev = rdev;
818 xfs_set_projid(ip, prid);
819
820 if (pip && XFS_INHERIT_GID(pip)) {
821 ip->i_d.di_gid = pip->i_d.di_gid;
822 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
823 inode->i_mode |= S_ISGID;
824 }
825
826 /*
827 * If the group ID of the new file does not match the effective group
828 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
829 * (and only if the irix_sgid_inherit compatibility variable is set).
830 */
831 if ((irix_sgid_inherit) &&
832 (inode->i_mode & S_ISGID) &&
833 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
834 inode->i_mode &= ~S_ISGID;
835
836 ip->i_d.di_size = 0;
837 ip->i_d.di_nextents = 0;
838 ASSERT(ip->i_d.di_nblocks == 0);
839
840 tv = current_time(inode);
841 inode->i_mtime = tv;
842 inode->i_atime = tv;
843 inode->i_ctime = tv;
844
845 ip->i_d.di_extsize = 0;
846 ip->i_d.di_dmevmask = 0;
847 ip->i_d.di_dmstate = 0;
848 ip->i_d.di_flags = 0;
849
850 if (ip->i_d.di_version == 3) {
851 inode_set_iversion(inode, 1);
852 ip->i_d.di_flags2 = 0;
853 ip->i_d.di_cowextsize = 0;
854 ip->i_d.di_crtime.t_sec = (int32_t)tv.tv_sec;
855 ip->i_d.di_crtime.t_nsec = (int32_t)tv.tv_nsec;
856 }
857
858
859 flags = XFS_ILOG_CORE;
860 switch (mode & S_IFMT) {
861 case S_IFIFO:
862 case S_IFCHR:
863 case S_IFBLK:
864 case S_IFSOCK:
865 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
866 ip->i_df.if_flags = 0;
867 flags |= XFS_ILOG_DEV;
868 break;
869 case S_IFREG:
870 case S_IFDIR:
871 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
872 uint di_flags = 0;
873
874 if (S_ISDIR(mode)) {
875 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
876 di_flags |= XFS_DIFLAG_RTINHERIT;
877 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
878 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
879 ip->i_d.di_extsize = pip->i_d.di_extsize;
880 }
881 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
882 di_flags |= XFS_DIFLAG_PROJINHERIT;
883 } else if (S_ISREG(mode)) {
884 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
885 di_flags |= XFS_DIFLAG_REALTIME;
886 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
887 di_flags |= XFS_DIFLAG_EXTSIZE;
888 ip->i_d.di_extsize = pip->i_d.di_extsize;
889 }
890 }
891 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
892 xfs_inherit_noatime)
893 di_flags |= XFS_DIFLAG_NOATIME;
894 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
895 xfs_inherit_nodump)
896 di_flags |= XFS_DIFLAG_NODUMP;
897 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
898 xfs_inherit_sync)
899 di_flags |= XFS_DIFLAG_SYNC;
900 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
901 xfs_inherit_nosymlinks)
902 di_flags |= XFS_DIFLAG_NOSYMLINKS;
903 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
904 xfs_inherit_nodefrag)
905 di_flags |= XFS_DIFLAG_NODEFRAG;
906 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
907 di_flags |= XFS_DIFLAG_FILESTREAM;
908
909 ip->i_d.di_flags |= di_flags;
910 }
911 if (pip &&
912 (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
913 pip->i_d.di_version == 3 &&
914 ip->i_d.di_version == 3) {
915 uint64_t di_flags2 = 0;
916
917 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
918 di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
919 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
920 }
921 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
922 di_flags2 |= XFS_DIFLAG2_DAX;
923
924 ip->i_d.di_flags2 |= di_flags2;
925 }
926 /* FALLTHROUGH */
927 case S_IFLNK:
928 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
929 ip->i_df.if_flags = XFS_IFEXTENTS;
930 ip->i_df.if_bytes = ip->i_df.if_real_bytes = 0;
931 ip->i_df.if_u1.if_root = NULL;
932 break;
933 default:
934 ASSERT(0);
935 }
936 /*
937 * Attribute fork settings for new inode.
938 */
939 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
940 ip->i_d.di_anextents = 0;
941
942 /*
943 * Log the new values stuffed into the inode.
944 */
945 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
946 xfs_trans_log_inode(tp, ip, flags);
947
948 /* now that we have an i_mode we can setup the inode structure */
949 xfs_setup_inode(ip);
950
951 *ipp = ip;
952 return 0;
953}
954
955/*
956 * Allocates a new inode from disk and return a pointer to the
957 * incore copy. This routine will internally commit the current
958 * transaction and allocate a new one if the Space Manager needed
959 * to do an allocation to replenish the inode free-list.
960 *
961 * This routine is designed to be called from xfs_create and
962 * xfs_create_dir.
963 *
964 */
965int
966xfs_dir_ialloc(
967 xfs_trans_t **tpp, /* input: current transaction;
968 output: may be a new transaction. */
969 xfs_inode_t *dp, /* directory within whose allocate
970 the inode. */
971 umode_t mode,
972 xfs_nlink_t nlink,
973 dev_t rdev,
974 prid_t prid, /* project id */
975 xfs_inode_t **ipp) /* pointer to inode; it will be
976 locked. */
977{
978 xfs_trans_t *tp;
979 xfs_inode_t *ip;
980 xfs_buf_t *ialloc_context = NULL;
981 int code;
982 void *dqinfo;
983 uint tflags;
984
985 tp = *tpp;
986 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
987
988 /*
989 * xfs_ialloc will return a pointer to an incore inode if
990 * the Space Manager has an available inode on the free
991 * list. Otherwise, it will do an allocation and replenish
992 * the freelist. Since we can only do one allocation per
993 * transaction without deadlocks, we will need to commit the
994 * current transaction and start a new one. We will then
995 * need to call xfs_ialloc again to get the inode.
996 *
997 * If xfs_ialloc did an allocation to replenish the freelist,
998 * it returns the bp containing the head of the freelist as
999 * ialloc_context. We will hold a lock on it across the
1000 * transaction commit so that no other process can steal
1001 * the inode(s) that we've just allocated.
1002 */
1003 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, &ialloc_context,
1004 &ip);
1005
1006 /*
1007 * Return an error if we were unable to allocate a new inode.
1008 * This should only happen if we run out of space on disk or
1009 * encounter a disk error.
1010 */
1011 if (code) {
1012 *ipp = NULL;
1013 return code;
1014 }
1015 if (!ialloc_context && !ip) {
1016 *ipp = NULL;
1017 return -ENOSPC;
1018 }
1019
1020 /*
1021 * If the AGI buffer is non-NULL, then we were unable to get an
1022 * inode in one operation. We need to commit the current
1023 * transaction and call xfs_ialloc() again. It is guaranteed
1024 * to succeed the second time.
1025 */
1026 if (ialloc_context) {
1027 /*
1028 * Normally, xfs_trans_commit releases all the locks.
1029 * We call bhold to hang on to the ialloc_context across
1030 * the commit. Holding this buffer prevents any other
1031 * processes from doing any allocations in this
1032 * allocation group.
1033 */
1034 xfs_trans_bhold(tp, ialloc_context);
1035
1036 /*
1037 * We want the quota changes to be associated with the next
1038 * transaction, NOT this one. So, detach the dqinfo from this
1039 * and attach it to the next transaction.
1040 */
1041 dqinfo = NULL;
1042 tflags = 0;
1043 if (tp->t_dqinfo) {
1044 dqinfo = (void *)tp->t_dqinfo;
1045 tp->t_dqinfo = NULL;
1046 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1047 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1048 }
1049
1050 code = xfs_trans_roll(&tp);
1051
1052 /*
1053 * Re-attach the quota info that we detached from prev trx.
1054 */
1055 if (dqinfo) {
1056 tp->t_dqinfo = dqinfo;
1057 tp->t_flags |= tflags;
1058 }
1059
1060 if (code) {
1061 xfs_buf_relse(ialloc_context);
1062 *tpp = tp;
1063 *ipp = NULL;
1064 return code;
1065 }
1066 xfs_trans_bjoin(tp, ialloc_context);
1067
1068 /*
1069 * Call ialloc again. Since we've locked out all
1070 * other allocations in this allocation group,
1071 * this call should always succeed.
1072 */
1073 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1074 &ialloc_context, &ip);
1075
1076 /*
1077 * If we get an error at this point, return to the caller
1078 * so that the current transaction can be aborted.
1079 */
1080 if (code) {
1081 *tpp = tp;
1082 *ipp = NULL;
1083 return code;
1084 }
1085 ASSERT(!ialloc_context && ip);
1086
1087 }
1088
1089 *ipp = ip;
1090 *tpp = tp;
1091
1092 return 0;
1093}
1094
1095/*
1096 * Decrement the link count on an inode & log the change. If this causes the
1097 * link count to go to zero, move the inode to AGI unlinked list so that it can
1098 * be freed when the last active reference goes away via xfs_inactive().
1099 */
1100static int /* error */
1101xfs_droplink(
1102 xfs_trans_t *tp,
1103 xfs_inode_t *ip)
1104{
1105 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1106
1107 drop_nlink(VFS_I(ip));
1108 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1109
1110 if (VFS_I(ip)->i_nlink)
1111 return 0;
1112
1113 return xfs_iunlink(tp, ip);
1114}
1115
1116/*
1117 * Increment the link count on an inode & log the change.
1118 */
1119static int
1120xfs_bumplink(
1121 xfs_trans_t *tp,
1122 xfs_inode_t *ip)
1123{
1124 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1125
1126 ASSERT(ip->i_d.di_version > 1);
1127 inc_nlink(VFS_I(ip));
1128 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1129 return 0;
1130}
1131
1132int
1133xfs_create(
1134 xfs_inode_t *dp,
1135 struct xfs_name *name,
1136 umode_t mode,
1137 dev_t rdev,
1138 xfs_inode_t **ipp)
1139{
1140 int is_dir = S_ISDIR(mode);
1141 struct xfs_mount *mp = dp->i_mount;
1142 struct xfs_inode *ip = NULL;
1143 struct xfs_trans *tp = NULL;
1144 int error;
1145 struct xfs_defer_ops dfops;
1146 xfs_fsblock_t first_block;
1147 bool unlock_dp_on_error = false;
1148 prid_t prid;
1149 struct xfs_dquot *udqp = NULL;
1150 struct xfs_dquot *gdqp = NULL;
1151 struct xfs_dquot *pdqp = NULL;
1152 struct xfs_trans_res *tres;
1153 uint resblks;
1154
1155 trace_xfs_create(dp, name);
1156
1157 if (XFS_FORCED_SHUTDOWN(mp))
1158 return -EIO;
1159
1160 prid = xfs_get_initial_prid(dp);
1161
1162 /*
1163 * Make sure that we have allocated dquot(s) on disk.
1164 */
1165 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1166 xfs_kgid_to_gid(current_fsgid()), prid,
1167 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1168 &udqp, &gdqp, &pdqp);
1169 if (error)
1170 return error;
1171
1172 if (is_dir) {
1173 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1174 tres = &M_RES(mp)->tr_mkdir;
1175 } else {
1176 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1177 tres = &M_RES(mp)->tr_create;
1178 }
1179
1180 /*
1181 * Initially assume that the file does not exist and
1182 * reserve the resources for that case. If that is not
1183 * the case we'll drop the one we have and get a more
1184 * appropriate transaction later.
1185 */
1186 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1187 if (error == -ENOSPC) {
1188 /* flush outstanding delalloc blocks and retry */
1189 xfs_flush_inodes(mp);
1190 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1191 }
1192 if (error)
1193 goto out_release_inode;
1194
1195 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1196 unlock_dp_on_error = true;
1197
1198 xfs_defer_init(&dfops, &first_block);
1199
1200 /*
1201 * Reserve disk quota and the inode.
1202 */
1203 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1204 pdqp, resblks, 1, 0);
1205 if (error)
1206 goto out_trans_cancel;
1207
1208 /*
1209 * A newly created regular or special file just has one directory
1210 * entry pointing to them, but a directory also the "." entry
1211 * pointing to itself.
1212 */
1213 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev, prid, &ip);
1214 if (error)
1215 goto out_trans_cancel;
1216
1217 /*
1218 * Now we join the directory inode to the transaction. We do not do it
1219 * earlier because xfs_dir_ialloc might commit the previous transaction
1220 * (and release all the locks). An error from here on will result in
1221 * the transaction cancel unlocking dp so don't do it explicitly in the
1222 * error path.
1223 */
1224 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1225 unlock_dp_on_error = false;
1226
1227 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1228 &first_block, &dfops, resblks ?
1229 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1230 if (error) {
1231 ASSERT(error != -ENOSPC);
1232 goto out_trans_cancel;
1233 }
1234 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1235 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1236
1237 if (is_dir) {
1238 error = xfs_dir_init(tp, ip, dp);
1239 if (error)
1240 goto out_bmap_cancel;
1241
1242 error = xfs_bumplink(tp, dp);
1243 if (error)
1244 goto out_bmap_cancel;
1245 }
1246
1247 /*
1248 * If this is a synchronous mount, make sure that the
1249 * create transaction goes to disk before returning to
1250 * the user.
1251 */
1252 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1253 xfs_trans_set_sync(tp);
1254
1255 /*
1256 * Attach the dquot(s) to the inodes and modify them incore.
1257 * These ids of the inode couldn't have changed since the new
1258 * inode has been locked ever since it was created.
1259 */
1260 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1261
1262 error = xfs_defer_finish(&tp, &dfops);
1263 if (error)
1264 goto out_bmap_cancel;
1265
1266 error = xfs_trans_commit(tp);
1267 if (error)
1268 goto out_release_inode;
1269
1270 xfs_qm_dqrele(udqp);
1271 xfs_qm_dqrele(gdqp);
1272 xfs_qm_dqrele(pdqp);
1273
1274 *ipp = ip;
1275 return 0;
1276
1277 out_bmap_cancel:
1278 xfs_defer_cancel(&dfops);
1279 out_trans_cancel:
1280 xfs_trans_cancel(tp);
1281 out_release_inode:
1282 /*
1283 * Wait until after the current transaction is aborted to finish the
1284 * setup of the inode and release the inode. This prevents recursive
1285 * transactions and deadlocks from xfs_inactive.
1286 */
1287 if (ip) {
1288 xfs_finish_inode_setup(ip);
1289 IRELE(ip);
1290 }
1291
1292 xfs_qm_dqrele(udqp);
1293 xfs_qm_dqrele(gdqp);
1294 xfs_qm_dqrele(pdqp);
1295
1296 if (unlock_dp_on_error)
1297 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1298 return error;
1299}
1300
1301int
1302xfs_create_tmpfile(
1303 struct xfs_inode *dp,
1304 umode_t mode,
1305 struct xfs_inode **ipp)
1306{
1307 struct xfs_mount *mp = dp->i_mount;
1308 struct xfs_inode *ip = NULL;
1309 struct xfs_trans *tp = NULL;
1310 int error;
1311 prid_t prid;
1312 struct xfs_dquot *udqp = NULL;
1313 struct xfs_dquot *gdqp = NULL;
1314 struct xfs_dquot *pdqp = NULL;
1315 struct xfs_trans_res *tres;
1316 uint resblks;
1317
1318 if (XFS_FORCED_SHUTDOWN(mp))
1319 return -EIO;
1320
1321 prid = xfs_get_initial_prid(dp);
1322
1323 /*
1324 * Make sure that we have allocated dquot(s) on disk.
1325 */
1326 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1327 xfs_kgid_to_gid(current_fsgid()), prid,
1328 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1329 &udqp, &gdqp, &pdqp);
1330 if (error)
1331 return error;
1332
1333 resblks = XFS_IALLOC_SPACE_RES(mp);
1334 tres = &M_RES(mp)->tr_create_tmpfile;
1335
1336 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1337 if (error)
1338 goto out_release_inode;
1339
1340 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1341 pdqp, resblks, 1, 0);
1342 if (error)
1343 goto out_trans_cancel;
1344
1345 error = xfs_dir_ialloc(&tp, dp, mode, 1, 0, prid, &ip);
1346 if (error)
1347 goto out_trans_cancel;
1348
1349 if (mp->m_flags & XFS_MOUNT_WSYNC)
1350 xfs_trans_set_sync(tp);
1351
1352 /*
1353 * Attach the dquot(s) to the inodes and modify them incore.
1354 * These ids of the inode couldn't have changed since the new
1355 * inode has been locked ever since it was created.
1356 */
1357 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1358
1359 error = xfs_iunlink(tp, ip);
1360 if (error)
1361 goto out_trans_cancel;
1362
1363 error = xfs_trans_commit(tp);
1364 if (error)
1365 goto out_release_inode;
1366
1367 xfs_qm_dqrele(udqp);
1368 xfs_qm_dqrele(gdqp);
1369 xfs_qm_dqrele(pdqp);
1370
1371 *ipp = ip;
1372 return 0;
1373
1374 out_trans_cancel:
1375 xfs_trans_cancel(tp);
1376 out_release_inode:
1377 /*
1378 * Wait until after the current transaction is aborted to finish the
1379 * setup of the inode and release the inode. This prevents recursive
1380 * transactions and deadlocks from xfs_inactive.
1381 */
1382 if (ip) {
1383 xfs_finish_inode_setup(ip);
1384 IRELE(ip);
1385 }
1386
1387 xfs_qm_dqrele(udqp);
1388 xfs_qm_dqrele(gdqp);
1389 xfs_qm_dqrele(pdqp);
1390
1391 return error;
1392}
1393
1394int
1395xfs_link(
1396 xfs_inode_t *tdp,
1397 xfs_inode_t *sip,
1398 struct xfs_name *target_name)
1399{
1400 xfs_mount_t *mp = tdp->i_mount;
1401 xfs_trans_t *tp;
1402 int error;
1403 struct xfs_defer_ops dfops;
1404 xfs_fsblock_t first_block;
1405 int resblks;
1406
1407 trace_xfs_link(tdp, target_name);
1408
1409 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1410
1411 if (XFS_FORCED_SHUTDOWN(mp))
1412 return -EIO;
1413
1414 error = xfs_qm_dqattach(sip, 0);
1415 if (error)
1416 goto std_return;
1417
1418 error = xfs_qm_dqattach(tdp, 0);
1419 if (error)
1420 goto std_return;
1421
1422 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1423 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1424 if (error == -ENOSPC) {
1425 resblks = 0;
1426 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1427 }
1428 if (error)
1429 goto std_return;
1430
1431 xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1432
1433 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1434 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1435
1436 /*
1437 * If we are using project inheritance, we only allow hard link
1438 * creation in our tree when the project IDs are the same; else
1439 * the tree quota mechanism could be circumvented.
1440 */
1441 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1442 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1443 error = -EXDEV;
1444 goto error_return;
1445 }
1446
1447 if (!resblks) {
1448 error = xfs_dir_canenter(tp, tdp, target_name);
1449 if (error)
1450 goto error_return;
1451 }
1452
1453 xfs_defer_init(&dfops, &first_block);
1454
1455 /*
1456 * Handle initial link state of O_TMPFILE inode
1457 */
1458 if (VFS_I(sip)->i_nlink == 0) {
1459 error = xfs_iunlink_remove(tp, sip);
1460 if (error)
1461 goto error_return;
1462 }
1463
1464 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1465 &first_block, &dfops, resblks);
1466 if (error)
1467 goto error_return;
1468 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1469 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1470
1471 error = xfs_bumplink(tp, sip);
1472 if (error)
1473 goto error_return;
1474
1475 /*
1476 * If this is a synchronous mount, make sure that the
1477 * link transaction goes to disk before returning to
1478 * the user.
1479 */
1480 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1481 xfs_trans_set_sync(tp);
1482
1483 error = xfs_defer_finish(&tp, &dfops);
1484 if (error) {
1485 xfs_defer_cancel(&dfops);
1486 goto error_return;
1487 }
1488
1489 return xfs_trans_commit(tp);
1490
1491 error_return:
1492 xfs_trans_cancel(tp);
1493 std_return:
1494 return error;
1495}
1496
1497/* Clear the reflink flag and the cowblocks tag if possible. */
1498static void
1499xfs_itruncate_clear_reflink_flags(
1500 struct xfs_inode *ip)
1501{
1502 struct xfs_ifork *dfork;
1503 struct xfs_ifork *cfork;
1504
1505 if (!xfs_is_reflink_inode(ip))
1506 return;
1507 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1508 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1509 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1510 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1511 if (cfork->if_bytes == 0)
1512 xfs_inode_clear_cowblocks_tag(ip);
1513}
1514
1515/*
1516 * Free up the underlying blocks past new_size. The new size must be smaller
1517 * than the current size. This routine can be used both for the attribute and
1518 * data fork, and does not modify the inode size, which is left to the caller.
1519 *
1520 * The transaction passed to this routine must have made a permanent log
1521 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1522 * given transaction and start new ones, so make sure everything involved in
1523 * the transaction is tidy before calling here. Some transaction will be
1524 * returned to the caller to be committed. The incoming transaction must
1525 * already include the inode, and both inode locks must be held exclusively.
1526 * The inode must also be "held" within the transaction. On return the inode
1527 * will be "held" within the returned transaction. This routine does NOT
1528 * require any disk space to be reserved for it within the transaction.
1529 *
1530 * If we get an error, we must return with the inode locked and linked into the
1531 * current transaction. This keeps things simple for the higher level code,
1532 * because it always knows that the inode is locked and held in the transaction
1533 * that returns to it whether errors occur or not. We don't mark the inode
1534 * dirty on error so that transactions can be easily aborted if possible.
1535 */
1536int
1537xfs_itruncate_extents(
1538 struct xfs_trans **tpp,
1539 struct xfs_inode *ip,
1540 int whichfork,
1541 xfs_fsize_t new_size)
1542{
1543 struct xfs_mount *mp = ip->i_mount;
1544 struct xfs_trans *tp = *tpp;
1545 struct xfs_defer_ops dfops;
1546 xfs_fsblock_t first_block;
1547 xfs_fileoff_t first_unmap_block;
1548 xfs_fileoff_t last_block;
1549 xfs_filblks_t unmap_len;
1550 int error = 0;
1551 int done = 0;
1552
1553 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1554 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1555 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1556 ASSERT(new_size <= XFS_ISIZE(ip));
1557 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1558 ASSERT(ip->i_itemp != NULL);
1559 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1560 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1561
1562 trace_xfs_itruncate_extents_start(ip, new_size);
1563
1564 /*
1565 * Since it is possible for space to become allocated beyond
1566 * the end of the file (in a crash where the space is allocated
1567 * but the inode size is not yet updated), simply remove any
1568 * blocks which show up between the new EOF and the maximum
1569 * possible file size. If the first block to be removed is
1570 * beyond the maximum file size (ie it is the same as last_block),
1571 * then there is nothing to do.
1572 */
1573 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1574 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1575 if (first_unmap_block == last_block)
1576 return 0;
1577
1578 ASSERT(first_unmap_block < last_block);
1579 unmap_len = last_block - first_unmap_block + 1;
1580 while (!done) {
1581 xfs_defer_init(&dfops, &first_block);
1582 error = xfs_bunmapi(tp, ip,
1583 first_unmap_block, unmap_len,
1584 xfs_bmapi_aflag(whichfork),
1585 XFS_ITRUNC_MAX_EXTENTS,
1586 &first_block, &dfops,
1587 &done);
1588 if (error)
1589 goto out_bmap_cancel;
1590
1591 /*
1592 * Duplicate the transaction that has the permanent
1593 * reservation and commit the old transaction.
1594 */
1595 xfs_defer_ijoin(&dfops, ip);
1596 error = xfs_defer_finish(&tp, &dfops);
1597 if (error)
1598 goto out_bmap_cancel;
1599
1600 error = xfs_trans_roll_inode(&tp, ip);
1601 if (error)
1602 goto out;
1603 }
1604
1605 if (whichfork == XFS_DATA_FORK) {
1606 /* Remove all pending CoW reservations. */
1607 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1608 first_unmap_block, last_block, true);
1609 if (error)
1610 goto out;
1611
1612 xfs_itruncate_clear_reflink_flags(ip);
1613 }
1614
1615 /*
1616 * Always re-log the inode so that our permanent transaction can keep
1617 * on rolling it forward in the log.
1618 */
1619 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1620
1621 trace_xfs_itruncate_extents_end(ip, new_size);
1622
1623out:
1624 *tpp = tp;
1625 return error;
1626out_bmap_cancel:
1627 /*
1628 * If the bunmapi call encounters an error, return to the caller where
1629 * the transaction can be properly aborted. We just need to make sure
1630 * we're not holding any resources that we were not when we came in.
1631 */
1632 xfs_defer_cancel(&dfops);
1633 goto out;
1634}
1635
1636int
1637xfs_release(
1638 xfs_inode_t *ip)
1639{
1640 xfs_mount_t *mp = ip->i_mount;
1641 int error;
1642
1643 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1644 return 0;
1645
1646 /* If this is a read-only mount, don't do this (would generate I/O) */
1647 if (mp->m_flags & XFS_MOUNT_RDONLY)
1648 return 0;
1649
1650 if (!XFS_FORCED_SHUTDOWN(mp)) {
1651 int truncated;
1652
1653 /*
1654 * If we previously truncated this file and removed old data
1655 * in the process, we want to initiate "early" writeout on
1656 * the last close. This is an attempt to combat the notorious
1657 * NULL files problem which is particularly noticeable from a
1658 * truncate down, buffered (re-)write (delalloc), followed by
1659 * a crash. What we are effectively doing here is
1660 * significantly reducing the time window where we'd otherwise
1661 * be exposed to that problem.
1662 */
1663 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1664 if (truncated) {
1665 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1666 if (ip->i_delayed_blks > 0) {
1667 error = filemap_flush(VFS_I(ip)->i_mapping);
1668 if (error)
1669 return error;
1670 }
1671 }
1672 }
1673
1674 if (VFS_I(ip)->i_nlink == 0)
1675 return 0;
1676
1677 if (xfs_can_free_eofblocks(ip, false)) {
1678
1679 /*
1680 * Check if the inode is being opened, written and closed
1681 * frequently and we have delayed allocation blocks outstanding
1682 * (e.g. streaming writes from the NFS server), truncating the
1683 * blocks past EOF will cause fragmentation to occur.
1684 *
1685 * In this case don't do the truncation, but we have to be
1686 * careful how we detect this case. Blocks beyond EOF show up as
1687 * i_delayed_blks even when the inode is clean, so we need to
1688 * truncate them away first before checking for a dirty release.
1689 * Hence on the first dirty close we will still remove the
1690 * speculative allocation, but after that we will leave it in
1691 * place.
1692 */
1693 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1694 return 0;
1695 /*
1696 * If we can't get the iolock just skip truncating the blocks
1697 * past EOF because we could deadlock with the mmap_sem
1698 * otherwise. We'll get another chance to drop them once the
1699 * last reference to the inode is dropped, so we'll never leak
1700 * blocks permanently.
1701 */
1702 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1703 error = xfs_free_eofblocks(ip);
1704 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1705 if (error)
1706 return error;
1707 }
1708
1709 /* delalloc blocks after truncation means it really is dirty */
1710 if (ip->i_delayed_blks)
1711 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1712 }
1713 return 0;
1714}
1715
1716/*
1717 * xfs_inactive_truncate
1718 *
1719 * Called to perform a truncate when an inode becomes unlinked.
1720 */
1721STATIC int
1722xfs_inactive_truncate(
1723 struct xfs_inode *ip)
1724{
1725 struct xfs_mount *mp = ip->i_mount;
1726 struct xfs_trans *tp;
1727 int error;
1728
1729 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1730 if (error) {
1731 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1732 return error;
1733 }
1734
1735 xfs_ilock(ip, XFS_ILOCK_EXCL);
1736 xfs_trans_ijoin(tp, ip, 0);
1737
1738 /*
1739 * Log the inode size first to prevent stale data exposure in the event
1740 * of a system crash before the truncate completes. See the related
1741 * comment in xfs_vn_setattr_size() for details.
1742 */
1743 ip->i_d.di_size = 0;
1744 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1745
1746 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1747 if (error)
1748 goto error_trans_cancel;
1749
1750 ASSERT(ip->i_d.di_nextents == 0);
1751
1752 error = xfs_trans_commit(tp);
1753 if (error)
1754 goto error_unlock;
1755
1756 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1757 return 0;
1758
1759error_trans_cancel:
1760 xfs_trans_cancel(tp);
1761error_unlock:
1762 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1763 return error;
1764}
1765
1766/*
1767 * xfs_inactive_ifree()
1768 *
1769 * Perform the inode free when an inode is unlinked.
1770 */
1771STATIC int
1772xfs_inactive_ifree(
1773 struct xfs_inode *ip)
1774{
1775 struct xfs_defer_ops dfops;
1776 xfs_fsblock_t first_block;
1777 struct xfs_mount *mp = ip->i_mount;
1778 struct xfs_trans *tp;
1779 int error;
1780
1781 /*
1782 * We try to use a per-AG reservation for any block needed by the finobt
1783 * tree, but as the finobt feature predates the per-AG reservation
1784 * support a degraded file system might not have enough space for the
1785 * reservation at mount time. In that case try to dip into the reserved
1786 * pool and pray.
1787 *
1788 * Send a warning if the reservation does happen to fail, as the inode
1789 * now remains allocated and sits on the unlinked list until the fs is
1790 * repaired.
1791 */
1792 if (unlikely(mp->m_inotbt_nores)) {
1793 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1794 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1795 &tp);
1796 } else {
1797 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1798 }
1799 if (error) {
1800 if (error == -ENOSPC) {
1801 xfs_warn_ratelimited(mp,
1802 "Failed to remove inode(s) from unlinked list. "
1803 "Please free space, unmount and run xfs_repair.");
1804 } else {
1805 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1806 }
1807 return error;
1808 }
1809
1810 xfs_ilock(ip, XFS_ILOCK_EXCL);
1811 xfs_trans_ijoin(tp, ip, 0);
1812
1813 xfs_defer_init(&dfops, &first_block);
1814 error = xfs_ifree(tp, ip, &dfops);
1815 if (error) {
1816 /*
1817 * If we fail to free the inode, shut down. The cancel
1818 * might do that, we need to make sure. Otherwise the
1819 * inode might be lost for a long time or forever.
1820 */
1821 if (!XFS_FORCED_SHUTDOWN(mp)) {
1822 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1823 __func__, error);
1824 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1825 }
1826 xfs_trans_cancel(tp);
1827 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1828 return error;
1829 }
1830
1831 /*
1832 * Credit the quota account(s). The inode is gone.
1833 */
1834 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1835
1836 /*
1837 * Just ignore errors at this point. There is nothing we can do except
1838 * to try to keep going. Make sure it's not a silent error.
1839 */
1840 error = xfs_defer_finish(&tp, &dfops);
1841 if (error) {
1842 xfs_notice(mp, "%s: xfs_defer_finish returned error %d",
1843 __func__, error);
1844 xfs_defer_cancel(&dfops);
1845 }
1846 error = xfs_trans_commit(tp);
1847 if (error)
1848 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1849 __func__, error);
1850
1851 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1852 return 0;
1853}
1854
1855/*
1856 * xfs_inactive
1857 *
1858 * This is called when the vnode reference count for the vnode
1859 * goes to zero. If the file has been unlinked, then it must
1860 * now be truncated. Also, we clear all of the read-ahead state
1861 * kept for the inode here since the file is now closed.
1862 */
1863void
1864xfs_inactive(
1865 xfs_inode_t *ip)
1866{
1867 struct xfs_mount *mp;
1868 struct xfs_ifork *cow_ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1869 int error;
1870 int truncate = 0;
1871
1872 /*
1873 * If the inode is already free, then there can be nothing
1874 * to clean up here.
1875 */
1876 if (VFS_I(ip)->i_mode == 0) {
1877 ASSERT(ip->i_df.if_real_bytes == 0);
1878 ASSERT(ip->i_df.if_broot_bytes == 0);
1879 return;
1880 }
1881
1882 mp = ip->i_mount;
1883 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1884
1885 /* If this is a read-only mount, don't do this (would generate I/O) */
1886 if (mp->m_flags & XFS_MOUNT_RDONLY)
1887 return;
1888
1889 /* Try to clean out the cow blocks if there are any. */
1890 if (xfs_is_reflink_inode(ip) && cow_ifp->if_bytes > 0)
1891 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1892
1893 if (VFS_I(ip)->i_nlink != 0) {
1894 /*
1895 * force is true because we are evicting an inode from the
1896 * cache. Post-eof blocks must be freed, lest we end up with
1897 * broken free space accounting.
1898 *
1899 * Note: don't bother with iolock here since lockdep complains
1900 * about acquiring it in reclaim context. We have the only
1901 * reference to the inode at this point anyways.
1902 */
1903 if (xfs_can_free_eofblocks(ip, true))
1904 xfs_free_eofblocks(ip);
1905
1906 return;
1907 }
1908
1909 if (S_ISREG(VFS_I(ip)->i_mode) &&
1910 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1911 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1912 truncate = 1;
1913
1914 error = xfs_qm_dqattach(ip, 0);
1915 if (error)
1916 return;
1917
1918 if (S_ISLNK(VFS_I(ip)->i_mode))
1919 error = xfs_inactive_symlink(ip);
1920 else if (truncate)
1921 error = xfs_inactive_truncate(ip);
1922 if (error)
1923 return;
1924
1925 /*
1926 * If there are attributes associated with the file then blow them away
1927 * now. The code calls a routine that recursively deconstructs the
1928 * attribute fork. If also blows away the in-core attribute fork.
1929 */
1930 if (XFS_IFORK_Q(ip)) {
1931 error = xfs_attr_inactive(ip);
1932 if (error)
1933 return;
1934 }
1935
1936 ASSERT(!ip->i_afp);
1937 ASSERT(ip->i_d.di_anextents == 0);
1938 ASSERT(ip->i_d.di_forkoff == 0);
1939
1940 /*
1941 * Free the inode.
1942 */
1943 error = xfs_inactive_ifree(ip);
1944 if (error)
1945 return;
1946
1947 /*
1948 * Release the dquots held by inode, if any.
1949 */
1950 xfs_qm_dqdetach(ip);
1951}
1952
1953/*
1954 * This is called when the inode's link count goes to 0 or we are creating a
1955 * tmpfile via O_TMPFILE. In the case of a tmpfile, @ignore_linkcount will be
1956 * set to true as the link count is dropped to zero by the VFS after we've
1957 * created the file successfully, so we have to add it to the unlinked list
1958 * while the link count is non-zero.
1959 *
1960 * We place the on-disk inode on a list in the AGI. It will be pulled from this
1961 * list when the inode is freed.
1962 */
1963STATIC int
1964xfs_iunlink(
1965 struct xfs_trans *tp,
1966 struct xfs_inode *ip)
1967{
1968 xfs_mount_t *mp = tp->t_mountp;
1969 xfs_agi_t *agi;
1970 xfs_dinode_t *dip;
1971 xfs_buf_t *agibp;
1972 xfs_buf_t *ibp;
1973 xfs_agino_t agino;
1974 short bucket_index;
1975 int offset;
1976 int error;
1977
1978 ASSERT(VFS_I(ip)->i_mode != 0);
1979
1980 /*
1981 * Get the agi buffer first. It ensures lock ordering
1982 * on the list.
1983 */
1984 error = xfs_read_agi(mp, tp, XFS_INO_TO_AGNO(mp, ip->i_ino), &agibp);
1985 if (error)
1986 return error;
1987 agi = XFS_BUF_TO_AGI(agibp);
1988
1989 /*
1990 * Get the index into the agi hash table for the
1991 * list this inode will go on.
1992 */
1993 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
1994 ASSERT(agino != 0);
1995 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
1996 ASSERT(agi->agi_unlinked[bucket_index]);
1997 ASSERT(be32_to_cpu(agi->agi_unlinked[bucket_index]) != agino);
1998
1999 if (agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO)) {
2000 /*
2001 * There is already another inode in the bucket we need
2002 * to add ourselves to. Add us at the front of the list.
2003 * Here we put the head pointer into our next pointer,
2004 * and then we fall through to point the head at us.
2005 */
2006 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2007 0, 0);
2008 if (error)
2009 return error;
2010
2011 ASSERT(dip->di_next_unlinked == cpu_to_be32(NULLAGINO));
2012 dip->di_next_unlinked = agi->agi_unlinked[bucket_index];
2013 offset = ip->i_imap.im_boffset +
2014 offsetof(xfs_dinode_t, di_next_unlinked);
2015
2016 /* need to recalc the inode CRC if appropriate */
2017 xfs_dinode_calc_crc(mp, dip);
2018
2019 xfs_trans_inode_buf(tp, ibp);
2020 xfs_trans_log_buf(tp, ibp, offset,
2021 (offset + sizeof(xfs_agino_t) - 1));
2022 xfs_inobp_check(mp, ibp);
2023 }
2024
2025 /*
2026 * Point the bucket head pointer at the inode being inserted.
2027 */
2028 ASSERT(agino != 0);
2029 agi->agi_unlinked[bucket_index] = cpu_to_be32(agino);
2030 offset = offsetof(xfs_agi_t, agi_unlinked) +
2031 (sizeof(xfs_agino_t) * bucket_index);
2032 xfs_trans_log_buf(tp, agibp, offset,
2033 (offset + sizeof(xfs_agino_t) - 1));
2034 return 0;
2035}
2036
2037/*
2038 * Pull the on-disk inode from the AGI unlinked list.
2039 */
2040STATIC int
2041xfs_iunlink_remove(
2042 xfs_trans_t *tp,
2043 xfs_inode_t *ip)
2044{
2045 xfs_ino_t next_ino;
2046 xfs_mount_t *mp;
2047 xfs_agi_t *agi;
2048 xfs_dinode_t *dip;
2049 xfs_buf_t *agibp;
2050 xfs_buf_t *ibp;
2051 xfs_agnumber_t agno;
2052 xfs_agino_t agino;
2053 xfs_agino_t next_agino;
2054 xfs_buf_t *last_ibp;
2055 xfs_dinode_t *last_dip = NULL;
2056 short bucket_index;
2057 int offset, last_offset = 0;
2058 int error;
2059
2060 mp = tp->t_mountp;
2061 agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2062
2063 /*
2064 * Get the agi buffer first. It ensures lock ordering
2065 * on the list.
2066 */
2067 error = xfs_read_agi(mp, tp, agno, &agibp);
2068 if (error)
2069 return error;
2070
2071 agi = XFS_BUF_TO_AGI(agibp);
2072
2073 /*
2074 * Get the index into the agi hash table for the
2075 * list this inode will go on.
2076 */
2077 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2078 ASSERT(agino != 0);
2079 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2080 ASSERT(agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO));
2081 ASSERT(agi->agi_unlinked[bucket_index]);
2082
2083 if (be32_to_cpu(agi->agi_unlinked[bucket_index]) == agino) {
2084 /*
2085 * We're at the head of the list. Get the inode's on-disk
2086 * buffer to see if there is anyone after us on the list.
2087 * Only modify our next pointer if it is not already NULLAGINO.
2088 * This saves us the overhead of dealing with the buffer when
2089 * there is no need to change it.
2090 */
2091 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2092 0, 0);
2093 if (error) {
2094 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2095 __func__, error);
2096 return error;
2097 }
2098 next_agino = be32_to_cpu(dip->di_next_unlinked);
2099 ASSERT(next_agino != 0);
2100 if (next_agino != NULLAGINO) {
2101 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2102 offset = ip->i_imap.im_boffset +
2103 offsetof(xfs_dinode_t, di_next_unlinked);
2104
2105 /* need to recalc the inode CRC if appropriate */
2106 xfs_dinode_calc_crc(mp, dip);
2107
2108 xfs_trans_inode_buf(tp, ibp);
2109 xfs_trans_log_buf(tp, ibp, offset,
2110 (offset + sizeof(xfs_agino_t) - 1));
2111 xfs_inobp_check(mp, ibp);
2112 } else {
2113 xfs_trans_brelse(tp, ibp);
2114 }
2115 /*
2116 * Point the bucket head pointer at the next inode.
2117 */
2118 ASSERT(next_agino != 0);
2119 ASSERT(next_agino != agino);
2120 agi->agi_unlinked[bucket_index] = cpu_to_be32(next_agino);
2121 offset = offsetof(xfs_agi_t, agi_unlinked) +
2122 (sizeof(xfs_agino_t) * bucket_index);
2123 xfs_trans_log_buf(tp, agibp, offset,
2124 (offset + sizeof(xfs_agino_t) - 1));
2125 } else {
2126 /*
2127 * We need to search the list for the inode being freed.
2128 */
2129 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2130 last_ibp = NULL;
2131 while (next_agino != agino) {
2132 struct xfs_imap imap;
2133
2134 if (last_ibp)
2135 xfs_trans_brelse(tp, last_ibp);
2136
2137 imap.im_blkno = 0;
2138 next_ino = XFS_AGINO_TO_INO(mp, agno, next_agino);
2139
2140 error = xfs_imap(mp, tp, next_ino, &imap, 0);
2141 if (error) {
2142 xfs_warn(mp,
2143 "%s: xfs_imap returned error %d.",
2144 __func__, error);
2145 return error;
2146 }
2147
2148 error = xfs_imap_to_bp(mp, tp, &imap, &last_dip,
2149 &last_ibp, 0, 0);
2150 if (error) {
2151 xfs_warn(mp,
2152 "%s: xfs_imap_to_bp returned error %d.",
2153 __func__, error);
2154 return error;
2155 }
2156
2157 last_offset = imap.im_boffset;
2158 next_agino = be32_to_cpu(last_dip->di_next_unlinked);
2159 ASSERT(next_agino != NULLAGINO);
2160 ASSERT(next_agino != 0);
2161 }
2162
2163 /*
2164 * Now last_ibp points to the buffer previous to us on the
2165 * unlinked list. Pull us from the list.
2166 */
2167 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2168 0, 0);
2169 if (error) {
2170 xfs_warn(mp, "%s: xfs_imap_to_bp(2) returned error %d.",
2171 __func__, error);
2172 return error;
2173 }
2174 next_agino = be32_to_cpu(dip->di_next_unlinked);
2175 ASSERT(next_agino != 0);
2176 ASSERT(next_agino != agino);
2177 if (next_agino != NULLAGINO) {
2178 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2179 offset = ip->i_imap.im_boffset +
2180 offsetof(xfs_dinode_t, di_next_unlinked);
2181
2182 /* need to recalc the inode CRC if appropriate */
2183 xfs_dinode_calc_crc(mp, dip);
2184
2185 xfs_trans_inode_buf(tp, ibp);
2186 xfs_trans_log_buf(tp, ibp, offset,
2187 (offset + sizeof(xfs_agino_t) - 1));
2188 xfs_inobp_check(mp, ibp);
2189 } else {
2190 xfs_trans_brelse(tp, ibp);
2191 }
2192 /*
2193 * Point the previous inode on the list to the next inode.
2194 */
2195 last_dip->di_next_unlinked = cpu_to_be32(next_agino);
2196 ASSERT(next_agino != 0);
2197 offset = last_offset + offsetof(xfs_dinode_t, di_next_unlinked);
2198
2199 /* need to recalc the inode CRC if appropriate */
2200 xfs_dinode_calc_crc(mp, last_dip);
2201
2202 xfs_trans_inode_buf(tp, last_ibp);
2203 xfs_trans_log_buf(tp, last_ibp, offset,
2204 (offset + sizeof(xfs_agino_t) - 1));
2205 xfs_inobp_check(mp, last_ibp);
2206 }
2207 return 0;
2208}
2209
2210/*
2211 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2212 * inodes that are in memory - they all must be marked stale and attached to
2213 * the cluster buffer.
2214 */
2215STATIC int
2216xfs_ifree_cluster(
2217 xfs_inode_t *free_ip,
2218 xfs_trans_t *tp,
2219 struct xfs_icluster *xic)
2220{
2221 xfs_mount_t *mp = free_ip->i_mount;
2222 int blks_per_cluster;
2223 int inodes_per_cluster;
2224 int nbufs;
2225 int i, j;
2226 int ioffset;
2227 xfs_daddr_t blkno;
2228 xfs_buf_t *bp;
2229 xfs_inode_t *ip;
2230 xfs_inode_log_item_t *iip;
2231 struct xfs_log_item *lip;
2232 struct xfs_perag *pag;
2233 xfs_ino_t inum;
2234
2235 inum = xic->first_ino;
2236 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2237 blks_per_cluster = xfs_icluster_size_fsb(mp);
2238 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
2239 nbufs = mp->m_ialloc_blks / blks_per_cluster;
2240
2241 for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
2242 /*
2243 * The allocation bitmap tells us which inodes of the chunk were
2244 * physically allocated. Skip the cluster if an inode falls into
2245 * a sparse region.
2246 */
2247 ioffset = inum - xic->first_ino;
2248 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2249 ASSERT(do_mod(ioffset, inodes_per_cluster) == 0);
2250 continue;
2251 }
2252
2253 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2254 XFS_INO_TO_AGBNO(mp, inum));
2255
2256 /*
2257 * We obtain and lock the backing buffer first in the process
2258 * here, as we have to ensure that any dirty inode that we
2259 * can't get the flush lock on is attached to the buffer.
2260 * If we scan the in-memory inodes first, then buffer IO can
2261 * complete before we get a lock on it, and hence we may fail
2262 * to mark all the active inodes on the buffer stale.
2263 */
2264 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2265 mp->m_bsize * blks_per_cluster,
2266 XBF_UNMAPPED);
2267
2268 if (!bp)
2269 return -ENOMEM;
2270
2271 /*
2272 * This buffer may not have been correctly initialised as we
2273 * didn't read it from disk. That's not important because we are
2274 * only using to mark the buffer as stale in the log, and to
2275 * attach stale cached inodes on it. That means it will never be
2276 * dispatched for IO. If it is, we want to know about it, and we
2277 * want it to fail. We can acheive this by adding a write
2278 * verifier to the buffer.
2279 */
2280 bp->b_ops = &xfs_inode_buf_ops;
2281
2282 /*
2283 * Walk the inodes already attached to the buffer and mark them
2284 * stale. These will all have the flush locks held, so an
2285 * in-memory inode walk can't lock them. By marking them all
2286 * stale first, we will not attempt to lock them in the loop
2287 * below as the XFS_ISTALE flag will be set.
2288 */
2289 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
2290 if (lip->li_type == XFS_LI_INODE) {
2291 iip = (xfs_inode_log_item_t *)lip;
2292 ASSERT(iip->ili_logged == 1);
2293 lip->li_cb = xfs_istale_done;
2294 xfs_trans_ail_copy_lsn(mp->m_ail,
2295 &iip->ili_flush_lsn,
2296 &iip->ili_item.li_lsn);
2297 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2298 }
2299 }
2300
2301
2302 /*
2303 * For each inode in memory attempt to add it to the inode
2304 * buffer and set it up for being staled on buffer IO
2305 * completion. This is safe as we've locked out tail pushing
2306 * and flushing by locking the buffer.
2307 *
2308 * We have already marked every inode that was part of a
2309 * transaction stale above, which means there is no point in
2310 * even trying to lock them.
2311 */
2312 for (i = 0; i < inodes_per_cluster; i++) {
2313retry:
2314 rcu_read_lock();
2315 ip = radix_tree_lookup(&pag->pag_ici_root,
2316 XFS_INO_TO_AGINO(mp, (inum + i)));
2317
2318 /* Inode not in memory, nothing to do */
2319 if (!ip) {
2320 rcu_read_unlock();
2321 continue;
2322 }
2323
2324 /*
2325 * because this is an RCU protected lookup, we could
2326 * find a recently freed or even reallocated inode
2327 * during the lookup. We need to check under the
2328 * i_flags_lock for a valid inode here. Skip it if it
2329 * is not valid, the wrong inode or stale.
2330 */
2331 spin_lock(&ip->i_flags_lock);
2332 if (ip->i_ino != inum + i ||
2333 __xfs_iflags_test(ip, XFS_ISTALE)) {
2334 spin_unlock(&ip->i_flags_lock);
2335 rcu_read_unlock();
2336 continue;
2337 }
2338 spin_unlock(&ip->i_flags_lock);
2339
2340 /*
2341 * Don't try to lock/unlock the current inode, but we
2342 * _cannot_ skip the other inodes that we did not find
2343 * in the list attached to the buffer and are not
2344 * already marked stale. If we can't lock it, back off
2345 * and retry.
2346 */
2347 if (ip != free_ip) {
2348 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2349 rcu_read_unlock();
2350 delay(1);
2351 goto retry;
2352 }
2353
2354 /*
2355 * Check the inode number again in case we're
2356 * racing with freeing in xfs_reclaim_inode().
2357 * See the comments in that function for more
2358 * information as to why the initial check is
2359 * not sufficient.
2360 */
2361 if (ip->i_ino != inum + i) {
2362 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2363 rcu_read_unlock();
2364 continue;
2365 }
2366 }
2367 rcu_read_unlock();
2368
2369 xfs_iflock(ip);
2370 xfs_iflags_set(ip, XFS_ISTALE);
2371
2372 /*
2373 * we don't need to attach clean inodes or those only
2374 * with unlogged changes (which we throw away, anyway).
2375 */
2376 iip = ip->i_itemp;
2377 if (!iip || xfs_inode_clean(ip)) {
2378 ASSERT(ip != free_ip);
2379 xfs_ifunlock(ip);
2380 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2381 continue;
2382 }
2383
2384 iip->ili_last_fields = iip->ili_fields;
2385 iip->ili_fields = 0;
2386 iip->ili_fsync_fields = 0;
2387 iip->ili_logged = 1;
2388 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2389 &iip->ili_item.li_lsn);
2390
2391 xfs_buf_attach_iodone(bp, xfs_istale_done,
2392 &iip->ili_item);
2393
2394 if (ip != free_ip)
2395 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2396 }
2397
2398 xfs_trans_stale_inode_buf(tp, bp);
2399 xfs_trans_binval(tp, bp);
2400 }
2401
2402 xfs_perag_put(pag);
2403 return 0;
2404}
2405
2406/*
2407 * Free any local-format buffers sitting around before we reset to
2408 * extents format.
2409 */
2410static inline void
2411xfs_ifree_local_data(
2412 struct xfs_inode *ip,
2413 int whichfork)
2414{
2415 struct xfs_ifork *ifp;
2416
2417 if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
2418 return;
2419
2420 ifp = XFS_IFORK_PTR(ip, whichfork);
2421 xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
2422}
2423
2424/*
2425 * This is called to return an inode to the inode free list.
2426 * The inode should already be truncated to 0 length and have
2427 * no pages associated with it. This routine also assumes that
2428 * the inode is already a part of the transaction.
2429 *
2430 * The on-disk copy of the inode will have been added to the list
2431 * of unlinked inodes in the AGI. We need to remove the inode from
2432 * that list atomically with respect to freeing it here.
2433 */
2434int
2435xfs_ifree(
2436 xfs_trans_t *tp,
2437 xfs_inode_t *ip,
2438 struct xfs_defer_ops *dfops)
2439{
2440 int error;
2441 struct xfs_icluster xic = { 0 };
2442
2443 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2444 ASSERT(VFS_I(ip)->i_nlink == 0);
2445 ASSERT(ip->i_d.di_nextents == 0);
2446 ASSERT(ip->i_d.di_anextents == 0);
2447 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2448 ASSERT(ip->i_d.di_nblocks == 0);
2449
2450 /*
2451 * Pull the on-disk inode from the AGI unlinked list.
2452 */
2453 error = xfs_iunlink_remove(tp, ip);
2454 if (error)
2455 return error;
2456
2457 error = xfs_difree(tp, ip->i_ino, dfops, &xic);
2458 if (error)
2459 return error;
2460
2461 xfs_ifree_local_data(ip, XFS_DATA_FORK);
2462 xfs_ifree_local_data(ip, XFS_ATTR_FORK);
2463
2464 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2465 ip->i_d.di_flags = 0;
2466 ip->i_d.di_flags2 = 0;
2467 ip->i_d.di_dmevmask = 0;
2468 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2469 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2470 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2471
2472 /* Don't attempt to replay owner changes for a deleted inode */
2473 ip->i_itemp->ili_fields &= ~(XFS_ILOG_AOWNER|XFS_ILOG_DOWNER);
2474
2475 /*
2476 * Bump the generation count so no one will be confused
2477 * by reincarnations of this inode.
2478 */
2479 VFS_I(ip)->i_generation++;
2480 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2481
2482 if (xic.deleted)
2483 error = xfs_ifree_cluster(ip, tp, &xic);
2484
2485 return error;
2486}
2487
2488/*
2489 * This is called to unpin an inode. The caller must have the inode locked
2490 * in at least shared mode so that the buffer cannot be subsequently pinned
2491 * once someone is waiting for it to be unpinned.
2492 */
2493static void
2494xfs_iunpin(
2495 struct xfs_inode *ip)
2496{
2497 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2498
2499 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2500
2501 /* Give the log a push to start the unpinning I/O */
2502 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0, NULL);
2503
2504}
2505
2506static void
2507__xfs_iunpin_wait(
2508 struct xfs_inode *ip)
2509{
2510 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2511 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2512
2513 xfs_iunpin(ip);
2514
2515 do {
2516 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2517 if (xfs_ipincount(ip))
2518 io_schedule();
2519 } while (xfs_ipincount(ip));
2520 finish_wait(wq, &wait.wq_entry);
2521}
2522
2523void
2524xfs_iunpin_wait(
2525 struct xfs_inode *ip)
2526{
2527 if (xfs_ipincount(ip))
2528 __xfs_iunpin_wait(ip);
2529}
2530
2531/*
2532 * Removing an inode from the namespace involves removing the directory entry
2533 * and dropping the link count on the inode. Removing the directory entry can
2534 * result in locking an AGF (directory blocks were freed) and removing a link
2535 * count can result in placing the inode on an unlinked list which results in
2536 * locking an AGI.
2537 *
2538 * The big problem here is that we have an ordering constraint on AGF and AGI
2539 * locking - inode allocation locks the AGI, then can allocate a new extent for
2540 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2541 * removes the inode from the unlinked list, requiring that we lock the AGI
2542 * first, and then freeing the inode can result in an inode chunk being freed
2543 * and hence freeing disk space requiring that we lock an AGF.
2544 *
2545 * Hence the ordering that is imposed by other parts of the code is AGI before
2546 * AGF. This means we cannot remove the directory entry before we drop the inode
2547 * reference count and put it on the unlinked list as this results in a lock
2548 * order of AGF then AGI, and this can deadlock against inode allocation and
2549 * freeing. Therefore we must drop the link counts before we remove the
2550 * directory entry.
2551 *
2552 * This is still safe from a transactional point of view - it is not until we
2553 * get to xfs_defer_finish() that we have the possibility of multiple
2554 * transactions in this operation. Hence as long as we remove the directory
2555 * entry and drop the link count in the first transaction of the remove
2556 * operation, there are no transactional constraints on the ordering here.
2557 */
2558int
2559xfs_remove(
2560 xfs_inode_t *dp,
2561 struct xfs_name *name,
2562 xfs_inode_t *ip)
2563{
2564 xfs_mount_t *mp = dp->i_mount;
2565 xfs_trans_t *tp = NULL;
2566 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2567 int error = 0;
2568 struct xfs_defer_ops dfops;
2569 xfs_fsblock_t first_block;
2570 uint resblks;
2571
2572 trace_xfs_remove(dp, name);
2573
2574 if (XFS_FORCED_SHUTDOWN(mp))
2575 return -EIO;
2576
2577 error = xfs_qm_dqattach(dp, 0);
2578 if (error)
2579 goto std_return;
2580
2581 error = xfs_qm_dqattach(ip, 0);
2582 if (error)
2583 goto std_return;
2584
2585 /*
2586 * We try to get the real space reservation first,
2587 * allowing for directory btree deletion(s) implying
2588 * possible bmap insert(s). If we can't get the space
2589 * reservation then we use 0 instead, and avoid the bmap
2590 * btree insert(s) in the directory code by, if the bmap
2591 * insert tries to happen, instead trimming the LAST
2592 * block from the directory.
2593 */
2594 resblks = XFS_REMOVE_SPACE_RES(mp);
2595 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2596 if (error == -ENOSPC) {
2597 resblks = 0;
2598 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2599 &tp);
2600 }
2601 if (error) {
2602 ASSERT(error != -ENOSPC);
2603 goto std_return;
2604 }
2605
2606 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2607
2608 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2609 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2610
2611 /*
2612 * If we're removing a directory perform some additional validation.
2613 */
2614 if (is_dir) {
2615 ASSERT(VFS_I(ip)->i_nlink >= 2);
2616 if (VFS_I(ip)->i_nlink != 2) {
2617 error = -ENOTEMPTY;
2618 goto out_trans_cancel;
2619 }
2620 if (!xfs_dir_isempty(ip)) {
2621 error = -ENOTEMPTY;
2622 goto out_trans_cancel;
2623 }
2624
2625 /* Drop the link from ip's "..". */
2626 error = xfs_droplink(tp, dp);
2627 if (error)
2628 goto out_trans_cancel;
2629
2630 /* Drop the "." link from ip to self. */
2631 error = xfs_droplink(tp, ip);
2632 if (error)
2633 goto out_trans_cancel;
2634 } else {
2635 /*
2636 * When removing a non-directory we need to log the parent
2637 * inode here. For a directory this is done implicitly
2638 * by the xfs_droplink call for the ".." entry.
2639 */
2640 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2641 }
2642 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2643
2644 /* Drop the link from dp to ip. */
2645 error = xfs_droplink(tp, ip);
2646 if (error)
2647 goto out_trans_cancel;
2648
2649 xfs_defer_init(&dfops, &first_block);
2650 error = xfs_dir_removename(tp, dp, name, ip->i_ino,
2651 &first_block, &dfops, resblks);
2652 if (error) {
2653 ASSERT(error != -ENOENT);
2654 goto out_bmap_cancel;
2655 }
2656
2657 /*
2658 * If this is a synchronous mount, make sure that the
2659 * remove transaction goes to disk before returning to
2660 * the user.
2661 */
2662 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2663 xfs_trans_set_sync(tp);
2664
2665 error = xfs_defer_finish(&tp, &dfops);
2666 if (error)
2667 goto out_bmap_cancel;
2668
2669 error = xfs_trans_commit(tp);
2670 if (error)
2671 goto std_return;
2672
2673 if (is_dir && xfs_inode_is_filestream(ip))
2674 xfs_filestream_deassociate(ip);
2675
2676 return 0;
2677
2678 out_bmap_cancel:
2679 xfs_defer_cancel(&dfops);
2680 out_trans_cancel:
2681 xfs_trans_cancel(tp);
2682 std_return:
2683 return error;
2684}
2685
2686/*
2687 * Enter all inodes for a rename transaction into a sorted array.
2688 */
2689#define __XFS_SORT_INODES 5
2690STATIC void
2691xfs_sort_for_rename(
2692 struct xfs_inode *dp1, /* in: old (source) directory inode */
2693 struct xfs_inode *dp2, /* in: new (target) directory inode */
2694 struct xfs_inode *ip1, /* in: inode of old entry */
2695 struct xfs_inode *ip2, /* in: inode of new entry */
2696 struct xfs_inode *wip, /* in: whiteout inode */
2697 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2698 int *num_inodes) /* in/out: inodes in array */
2699{
2700 int i, j;
2701
2702 ASSERT(*num_inodes == __XFS_SORT_INODES);
2703 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2704
2705 /*
2706 * i_tab contains a list of pointers to inodes. We initialize
2707 * the table here & we'll sort it. We will then use it to
2708 * order the acquisition of the inode locks.
2709 *
2710 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2711 */
2712 i = 0;
2713 i_tab[i++] = dp1;
2714 i_tab[i++] = dp2;
2715 i_tab[i++] = ip1;
2716 if (ip2)
2717 i_tab[i++] = ip2;
2718 if (wip)
2719 i_tab[i++] = wip;
2720 *num_inodes = i;
2721
2722 /*
2723 * Sort the elements via bubble sort. (Remember, there are at
2724 * most 5 elements to sort, so this is adequate.)
2725 */
2726 for (i = 0; i < *num_inodes; i++) {
2727 for (j = 1; j < *num_inodes; j++) {
2728 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2729 struct xfs_inode *temp = i_tab[j];
2730 i_tab[j] = i_tab[j-1];
2731 i_tab[j-1] = temp;
2732 }
2733 }
2734 }
2735}
2736
2737static int
2738xfs_finish_rename(
2739 struct xfs_trans *tp,
2740 struct xfs_defer_ops *dfops)
2741{
2742 int error;
2743
2744 /*
2745 * If this is a synchronous mount, make sure that the rename transaction
2746 * goes to disk before returning to the user.
2747 */
2748 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2749 xfs_trans_set_sync(tp);
2750
2751 error = xfs_defer_finish(&tp, dfops);
2752 if (error) {
2753 xfs_defer_cancel(dfops);
2754 xfs_trans_cancel(tp);
2755 return error;
2756 }
2757
2758 return xfs_trans_commit(tp);
2759}
2760
2761/*
2762 * xfs_cross_rename()
2763 *
2764 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
2765 */
2766STATIC int
2767xfs_cross_rename(
2768 struct xfs_trans *tp,
2769 struct xfs_inode *dp1,
2770 struct xfs_name *name1,
2771 struct xfs_inode *ip1,
2772 struct xfs_inode *dp2,
2773 struct xfs_name *name2,
2774 struct xfs_inode *ip2,
2775 struct xfs_defer_ops *dfops,
2776 xfs_fsblock_t *first_block,
2777 int spaceres)
2778{
2779 int error = 0;
2780 int ip1_flags = 0;
2781 int ip2_flags = 0;
2782 int dp2_flags = 0;
2783
2784 /* Swap inode number for dirent in first parent */
2785 error = xfs_dir_replace(tp, dp1, name1,
2786 ip2->i_ino,
2787 first_block, dfops, spaceres);
2788 if (error)
2789 goto out_trans_abort;
2790
2791 /* Swap inode number for dirent in second parent */
2792 error = xfs_dir_replace(tp, dp2, name2,
2793 ip1->i_ino,
2794 first_block, dfops, spaceres);
2795 if (error)
2796 goto out_trans_abort;
2797
2798 /*
2799 * If we're renaming one or more directories across different parents,
2800 * update the respective ".." entries (and link counts) to match the new
2801 * parents.
2802 */
2803 if (dp1 != dp2) {
2804 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2805
2806 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2807 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2808 dp1->i_ino, first_block,
2809 dfops, spaceres);
2810 if (error)
2811 goto out_trans_abort;
2812
2813 /* transfer ip2 ".." reference to dp1 */
2814 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2815 error = xfs_droplink(tp, dp2);
2816 if (error)
2817 goto out_trans_abort;
2818 error = xfs_bumplink(tp, dp1);
2819 if (error)
2820 goto out_trans_abort;
2821 }
2822
2823 /*
2824 * Although ip1 isn't changed here, userspace needs
2825 * to be warned about the change, so that applications
2826 * relying on it (like backup ones), will properly
2827 * notify the change
2828 */
2829 ip1_flags |= XFS_ICHGTIME_CHG;
2830 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2831 }
2832
2833 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2834 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2835 dp2->i_ino, first_block,
2836 dfops, spaceres);
2837 if (error)
2838 goto out_trans_abort;
2839
2840 /* transfer ip1 ".." reference to dp2 */
2841 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2842 error = xfs_droplink(tp, dp1);
2843 if (error)
2844 goto out_trans_abort;
2845 error = xfs_bumplink(tp, dp2);
2846 if (error)
2847 goto out_trans_abort;
2848 }
2849
2850 /*
2851 * Although ip2 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_MOD | XFS_ICHGTIME_CHG;
2857 ip2_flags |= XFS_ICHGTIME_CHG;
2858 }
2859 }
2860
2861 if (ip1_flags) {
2862 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2863 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2864 }
2865 if (ip2_flags) {
2866 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2867 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2868 }
2869 if (dp2_flags) {
2870 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2871 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2872 }
2873 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2874 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2875 return xfs_finish_rename(tp, dfops);
2876
2877out_trans_abort:
2878 xfs_defer_cancel(dfops);
2879 xfs_trans_cancel(tp);
2880 return error;
2881}
2882
2883/*
2884 * xfs_rename_alloc_whiteout()
2885 *
2886 * Return a referenced, unlinked, unlocked inode that that can be used as a
2887 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2888 * crash between allocating the inode and linking it into the rename transaction
2889 * recovery will free the inode and we won't leak it.
2890 */
2891static int
2892xfs_rename_alloc_whiteout(
2893 struct xfs_inode *dp,
2894 struct xfs_inode **wip)
2895{
2896 struct xfs_inode *tmpfile;
2897 int error;
2898
2899 error = xfs_create_tmpfile(dp, S_IFCHR | WHITEOUT_MODE, &tmpfile);
2900 if (error)
2901 return error;
2902
2903 /*
2904 * Prepare the tmpfile inode as if it were created through the VFS.
2905 * Otherwise, the link increment paths will complain about nlink 0->1.
2906 * Drop the link count as done by d_tmpfile(), complete the inode setup
2907 * and flag it as linkable.
2908 */
2909 drop_nlink(VFS_I(tmpfile));
2910 xfs_setup_iops(tmpfile);
2911 xfs_finish_inode_setup(tmpfile);
2912 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2913
2914 *wip = tmpfile;
2915 return 0;
2916}
2917
2918/*
2919 * xfs_rename
2920 */
2921int
2922xfs_rename(
2923 struct xfs_inode *src_dp,
2924 struct xfs_name *src_name,
2925 struct xfs_inode *src_ip,
2926 struct xfs_inode *target_dp,
2927 struct xfs_name *target_name,
2928 struct xfs_inode *target_ip,
2929 unsigned int flags)
2930{
2931 struct xfs_mount *mp = src_dp->i_mount;
2932 struct xfs_trans *tp;
2933 struct xfs_defer_ops dfops;
2934 xfs_fsblock_t first_block;
2935 struct xfs_inode *wip = NULL; /* whiteout inode */
2936 struct xfs_inode *inodes[__XFS_SORT_INODES];
2937 int num_inodes = __XFS_SORT_INODES;
2938 bool new_parent = (src_dp != target_dp);
2939 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2940 int spaceres;
2941 int error;
2942
2943 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2944
2945 if ((flags & RENAME_EXCHANGE) && !target_ip)
2946 return -EINVAL;
2947
2948 /*
2949 * If we are doing a whiteout operation, allocate the whiteout inode
2950 * we will be placing at the target and ensure the type is set
2951 * appropriately.
2952 */
2953 if (flags & RENAME_WHITEOUT) {
2954 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
2955 error = xfs_rename_alloc_whiteout(target_dp, &wip);
2956 if (error)
2957 return error;
2958
2959 /* setup target dirent info as whiteout */
2960 src_name->type = XFS_DIR3_FT_CHRDEV;
2961 }
2962
2963 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2964 inodes, &num_inodes);
2965
2966 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2967 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2968 if (error == -ENOSPC) {
2969 spaceres = 0;
2970 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2971 &tp);
2972 }
2973 if (error)
2974 goto out_release_wip;
2975
2976 /*
2977 * Attach the dquots to the inodes
2978 */
2979 error = xfs_qm_vop_rename_dqattach(inodes);
2980 if (error)
2981 goto out_trans_cancel;
2982
2983 /*
2984 * Lock all the participating inodes. Depending upon whether
2985 * the target_name exists in the target directory, and
2986 * whether the target directory is the same as the source
2987 * directory, we can lock from 2 to 4 inodes.
2988 */
2989 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2990
2991 /*
2992 * Join all the inodes to the transaction. From this point on,
2993 * we can rely on either trans_commit or trans_cancel to unlock
2994 * them.
2995 */
2996 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2997 if (new_parent)
2998 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2999 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3000 if (target_ip)
3001 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3002 if (wip)
3003 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3004
3005 /*
3006 * If we are using project inheritance, we only allow renames
3007 * into our tree when the project IDs are the same; else the
3008 * tree quota mechanism would be circumvented.
3009 */
3010 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
3011 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
3012 error = -EXDEV;
3013 goto out_trans_cancel;
3014 }
3015
3016 xfs_defer_init(&dfops, &first_block);
3017
3018 /* RENAME_EXCHANGE is unique from here on. */
3019 if (flags & RENAME_EXCHANGE)
3020 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3021 target_dp, target_name, target_ip,
3022 &dfops, &first_block, spaceres);
3023
3024 /*
3025 * Set up the target.
3026 */
3027 if (target_ip == NULL) {
3028 /*
3029 * If there's no space reservation, check the entry will
3030 * fit before actually inserting it.
3031 */
3032 if (!spaceres) {
3033 error = xfs_dir_canenter(tp, target_dp, target_name);
3034 if (error)
3035 goto out_trans_cancel;
3036 }
3037 /*
3038 * If target does not exist and the rename crosses
3039 * directories, adjust the target directory link count
3040 * to account for the ".." reference from the new entry.
3041 */
3042 error = xfs_dir_createname(tp, target_dp, target_name,
3043 src_ip->i_ino, &first_block,
3044 &dfops, spaceres);
3045 if (error)
3046 goto out_bmap_cancel;
3047
3048 xfs_trans_ichgtime(tp, target_dp,
3049 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3050
3051 if (new_parent && src_is_directory) {
3052 error = xfs_bumplink(tp, target_dp);
3053 if (error)
3054 goto out_bmap_cancel;
3055 }
3056 } else { /* target_ip != NULL */
3057 /*
3058 * If target exists and it's a directory, check that both
3059 * target and source are directories and that target can be
3060 * destroyed, or that neither is a directory.
3061 */
3062 if (S_ISDIR(VFS_I(target_ip)->i_mode)) {
3063 /*
3064 * Make sure target dir is empty.
3065 */
3066 if (!(xfs_dir_isempty(target_ip)) ||
3067 (VFS_I(target_ip)->i_nlink > 2)) {
3068 error = -EEXIST;
3069 goto out_trans_cancel;
3070 }
3071 }
3072
3073 /*
3074 * Link the source inode under the target name.
3075 * If the source inode is a directory and we are moving
3076 * it across directories, its ".." entry will be
3077 * inconsistent until we replace that down below.
3078 *
3079 * In case there is already an entry with the same
3080 * name at the destination directory, remove it first.
3081 */
3082 error = xfs_dir_replace(tp, target_dp, target_name,
3083 src_ip->i_ino,
3084 &first_block, &dfops, spaceres);
3085 if (error)
3086 goto out_bmap_cancel;
3087
3088 xfs_trans_ichgtime(tp, target_dp,
3089 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3090
3091 /*
3092 * Decrement the link count on the target since the target
3093 * dir no longer points to it.
3094 */
3095 error = xfs_droplink(tp, target_ip);
3096 if (error)
3097 goto out_bmap_cancel;
3098
3099 if (src_is_directory) {
3100 /*
3101 * Drop the link from the old "." entry.
3102 */
3103 error = xfs_droplink(tp, target_ip);
3104 if (error)
3105 goto out_bmap_cancel;
3106 }
3107 } /* target_ip != NULL */
3108
3109 /*
3110 * Remove the source.
3111 */
3112 if (new_parent && src_is_directory) {
3113 /*
3114 * Rewrite the ".." entry to point to the new
3115 * directory.
3116 */
3117 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3118 target_dp->i_ino,
3119 &first_block, &dfops, spaceres);
3120 ASSERT(error != -EEXIST);
3121 if (error)
3122 goto out_bmap_cancel;
3123 }
3124
3125 /*
3126 * We always want to hit the ctime on the source inode.
3127 *
3128 * This isn't strictly required by the standards since the source
3129 * inode isn't really being changed, but old unix file systems did
3130 * it and some incremental backup programs won't work without it.
3131 */
3132 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3133 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3134
3135 /*
3136 * Adjust the link count on src_dp. This is necessary when
3137 * renaming a directory, either within one parent when
3138 * the target existed, or across two parent directories.
3139 */
3140 if (src_is_directory && (new_parent || target_ip != NULL)) {
3141
3142 /*
3143 * Decrement link count on src_directory since the
3144 * entry that's moved no longer points to it.
3145 */
3146 error = xfs_droplink(tp, src_dp);
3147 if (error)
3148 goto out_bmap_cancel;
3149 }
3150
3151 /*
3152 * For whiteouts, we only need to update the source dirent with the
3153 * inode number of the whiteout inode rather than removing it
3154 * altogether.
3155 */
3156 if (wip) {
3157 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3158 &first_block, &dfops, spaceres);
3159 } else
3160 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3161 &first_block, &dfops, spaceres);
3162 if (error)
3163 goto out_bmap_cancel;
3164
3165 /*
3166 * For whiteouts, we need to bump the link count on the whiteout inode.
3167 * This means that failures all the way up to this point leave the inode
3168 * on the unlinked list and so cleanup is a simple matter of dropping
3169 * the remaining reference to it. If we fail here after bumping the link
3170 * count, we're shutting down the filesystem so we'll never see the
3171 * intermediate state on disk.
3172 */
3173 if (wip) {
3174 ASSERT(VFS_I(wip)->i_nlink == 0);
3175 error = xfs_bumplink(tp, wip);
3176 if (error)
3177 goto out_bmap_cancel;
3178 error = xfs_iunlink_remove(tp, wip);
3179 if (error)
3180 goto out_bmap_cancel;
3181 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3182
3183 /*
3184 * Now we have a real link, clear the "I'm a tmpfile" state
3185 * flag from the inode so it doesn't accidentally get misused in
3186 * future.
3187 */
3188 VFS_I(wip)->i_state &= ~I_LINKABLE;
3189 }
3190
3191 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3192 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3193 if (new_parent)
3194 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3195
3196 error = xfs_finish_rename(tp, &dfops);
3197 if (wip)
3198 IRELE(wip);
3199 return error;
3200
3201out_bmap_cancel:
3202 xfs_defer_cancel(&dfops);
3203out_trans_cancel:
3204 xfs_trans_cancel(tp);
3205out_release_wip:
3206 if (wip)
3207 IRELE(wip);
3208 return error;
3209}
3210
3211STATIC int
3212xfs_iflush_cluster(
3213 struct xfs_inode *ip,
3214 struct xfs_buf *bp)
3215{
3216 struct xfs_mount *mp = ip->i_mount;
3217 struct xfs_perag *pag;
3218 unsigned long first_index, mask;
3219 unsigned long inodes_per_cluster;
3220 int cilist_size;
3221 struct xfs_inode **cilist;
3222 struct xfs_inode *cip;
3223 int nr_found;
3224 int clcount = 0;
3225 int bufwasdelwri;
3226 int i;
3227
3228 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3229
3230 inodes_per_cluster = mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog;
3231 cilist_size = inodes_per_cluster * sizeof(xfs_inode_t *);
3232 cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3233 if (!cilist)
3234 goto out_put;
3235
3236 mask = ~(((mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog)) - 1);
3237 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3238 rcu_read_lock();
3239 /* really need a gang lookup range call here */
3240 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3241 first_index, inodes_per_cluster);
3242 if (nr_found == 0)
3243 goto out_free;
3244
3245 for (i = 0; i < nr_found; i++) {
3246 cip = cilist[i];
3247 if (cip == ip)
3248 continue;
3249
3250 /*
3251 * because this is an RCU protected lookup, we could find a
3252 * recently freed or even reallocated inode during the lookup.
3253 * We need to check under the i_flags_lock for a valid inode
3254 * here. Skip it if it is not valid or the wrong inode.
3255 */
3256 spin_lock(&cip->i_flags_lock);
3257 if (!cip->i_ino ||
3258 __xfs_iflags_test(cip, XFS_ISTALE)) {
3259 spin_unlock(&cip->i_flags_lock);
3260 continue;
3261 }
3262
3263 /*
3264 * Once we fall off the end of the cluster, no point checking
3265 * any more inodes in the list because they will also all be
3266 * outside the cluster.
3267 */
3268 if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3269 spin_unlock(&cip->i_flags_lock);
3270 break;
3271 }
3272 spin_unlock(&cip->i_flags_lock);
3273
3274 /*
3275 * Do an un-protected check to see if the inode is dirty and
3276 * is a candidate for flushing. These checks will be repeated
3277 * later after the appropriate locks are acquired.
3278 */
3279 if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3280 continue;
3281
3282 /*
3283 * Try to get locks. If any are unavailable or it is pinned,
3284 * then this inode cannot be flushed and is skipped.
3285 */
3286
3287 if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3288 continue;
3289 if (!xfs_iflock_nowait(cip)) {
3290 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3291 continue;
3292 }
3293 if (xfs_ipincount(cip)) {
3294 xfs_ifunlock(cip);
3295 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3296 continue;
3297 }
3298
3299
3300 /*
3301 * Check the inode number again, just to be certain we are not
3302 * racing with freeing in xfs_reclaim_inode(). See the comments
3303 * in that function for more information as to why the initial
3304 * check is not sufficient.
3305 */
3306 if (!cip->i_ino) {
3307 xfs_ifunlock(cip);
3308 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3309 continue;
3310 }
3311
3312 /*
3313 * arriving here means that this inode can be flushed. First
3314 * re-check that it's dirty before flushing.
3315 */
3316 if (!xfs_inode_clean(cip)) {
3317 int error;
3318 error = xfs_iflush_int(cip, bp);
3319 if (error) {
3320 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3321 goto cluster_corrupt_out;
3322 }
3323 clcount++;
3324 } else {
3325 xfs_ifunlock(cip);
3326 }
3327 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3328 }
3329
3330 if (clcount) {
3331 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3332 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3333 }
3334
3335out_free:
3336 rcu_read_unlock();
3337 kmem_free(cilist);
3338out_put:
3339 xfs_perag_put(pag);
3340 return 0;
3341
3342
3343cluster_corrupt_out:
3344 /*
3345 * Corruption detected in the clustering loop. Invalidate the
3346 * inode buffer and shut down the filesystem.
3347 */
3348 rcu_read_unlock();
3349 /*
3350 * Clean up the buffer. If it was delwri, just release it --
3351 * brelse can handle it with no problems. If not, shut down the
3352 * filesystem before releasing the buffer.
3353 */
3354 bufwasdelwri = (bp->b_flags & _XBF_DELWRI_Q);
3355 if (bufwasdelwri)
3356 xfs_buf_relse(bp);
3357
3358 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3359
3360 if (!bufwasdelwri) {
3361 /*
3362 * Just like incore_relse: if we have b_iodone functions,
3363 * mark the buffer as an error and call them. Otherwise
3364 * mark it as stale and brelse.
3365 */
3366 if (bp->b_iodone) {
3367 bp->b_flags &= ~XBF_DONE;
3368 xfs_buf_stale(bp);
3369 xfs_buf_ioerror(bp, -EIO);
3370 xfs_buf_ioend(bp);
3371 } else {
3372 xfs_buf_stale(bp);
3373 xfs_buf_relse(bp);
3374 }
3375 }
3376
3377 /*
3378 * Unlocks the flush lock
3379 */
3380 xfs_iflush_abort(cip, false);
3381 kmem_free(cilist);
3382 xfs_perag_put(pag);
3383 return -EFSCORRUPTED;
3384}
3385
3386/*
3387 * Flush dirty inode metadata into the backing buffer.
3388 *
3389 * The caller must have the inode lock and the inode flush lock held. The
3390 * inode lock will still be held upon return to the caller, and the inode
3391 * flush lock will be released after the inode has reached the disk.
3392 *
3393 * The caller must write out the buffer returned in *bpp and release it.
3394 */
3395int
3396xfs_iflush(
3397 struct xfs_inode *ip,
3398 struct xfs_buf **bpp)
3399{
3400 struct xfs_mount *mp = ip->i_mount;
3401 struct xfs_buf *bp = NULL;
3402 struct xfs_dinode *dip;
3403 int error;
3404
3405 XFS_STATS_INC(mp, xs_iflush_count);
3406
3407 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3408 ASSERT(xfs_isiflocked(ip));
3409 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3410 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3411
3412 *bpp = NULL;
3413
3414 xfs_iunpin_wait(ip);
3415
3416 /*
3417 * For stale inodes we cannot rely on the backing buffer remaining
3418 * stale in cache for the remaining life of the stale inode and so
3419 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3420 * inodes below. We have to check this after ensuring the inode is
3421 * unpinned so that it is safe to reclaim the stale inode after the
3422 * flush call.
3423 */
3424 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3425 xfs_ifunlock(ip);
3426 return 0;
3427 }
3428
3429 /*
3430 * This may have been unpinned because the filesystem is shutting
3431 * down forcibly. If that's the case we must not write this inode
3432 * to disk, because the log record didn't make it to disk.
3433 *
3434 * We also have to remove the log item from the AIL in this case,
3435 * as we wait for an empty AIL as part of the unmount process.
3436 */
3437 if (XFS_FORCED_SHUTDOWN(mp)) {
3438 error = -EIO;
3439 goto abort_out;
3440 }
3441
3442 /*
3443 * Get the buffer containing the on-disk inode. We are doing a try-lock
3444 * operation here, so we may get an EAGAIN error. In that case, we
3445 * simply want to return with the inode still dirty.
3446 *
3447 * If we get any other error, we effectively have a corruption situation
3448 * and we cannot flush the inode, so we treat it the same as failing
3449 * xfs_iflush_int().
3450 */
3451 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3452 0);
3453 if (error == -EAGAIN) {
3454 xfs_ifunlock(ip);
3455 return error;
3456 }
3457 if (error)
3458 goto corrupt_out;
3459
3460 /*
3461 * First flush out the inode that xfs_iflush was called with.
3462 */
3463 error = xfs_iflush_int(ip, bp);
3464 if (error)
3465 goto corrupt_out;
3466
3467 /*
3468 * If the buffer is pinned then push on the log now so we won't
3469 * get stuck waiting in the write for too long.
3470 */
3471 if (xfs_buf_ispinned(bp))
3472 xfs_log_force(mp, 0);
3473
3474 /*
3475 * inode clustering:
3476 * see if other inodes can be gathered into this write
3477 */
3478 error = xfs_iflush_cluster(ip, bp);
3479 if (error)
3480 goto cluster_corrupt_out;
3481
3482 *bpp = bp;
3483 return 0;
3484
3485corrupt_out:
3486 if (bp)
3487 xfs_buf_relse(bp);
3488 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3489cluster_corrupt_out:
3490 error = -EFSCORRUPTED;
3491abort_out:
3492 /*
3493 * Unlocks the flush lock
3494 */
3495 xfs_iflush_abort(ip, false);
3496 return error;
3497}
3498
3499/*
3500 * If there are inline format data / attr forks attached to this inode,
3501 * make sure they're not corrupt.
3502 */
3503bool
3504xfs_inode_verify_forks(
3505 struct xfs_inode *ip)
3506{
3507 struct xfs_ifork *ifp;
3508 xfs_failaddr_t fa;
3509
3510 fa = xfs_ifork_verify_data(ip, &xfs_default_ifork_ops);
3511 if (fa) {
3512 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
3513 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "data fork",
3514 ifp->if_u1.if_data, ifp->if_bytes, fa);
3515 return false;
3516 }
3517
3518 fa = xfs_ifork_verify_attr(ip, &xfs_default_ifork_ops);
3519 if (fa) {
3520 ifp = XFS_IFORK_PTR(ip, XFS_ATTR_FORK);
3521 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "attr fork",
3522 ifp ? ifp->if_u1.if_data : NULL,
3523 ifp ? ifp->if_bytes : 0, fa);
3524 return false;
3525 }
3526 return true;
3527}
3528
3529STATIC int
3530xfs_iflush_int(
3531 struct xfs_inode *ip,
3532 struct xfs_buf *bp)
3533{
3534 struct xfs_inode_log_item *iip = ip->i_itemp;
3535 struct xfs_dinode *dip;
3536 struct xfs_mount *mp = ip->i_mount;
3537
3538 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3539 ASSERT(xfs_isiflocked(ip));
3540 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3541 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3542 ASSERT(iip != NULL && iip->ili_fields != 0);
3543 ASSERT(ip->i_d.di_version > 1);
3544
3545 /* set *dip = inode's place in the buffer */
3546 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3547
3548 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3549 mp, XFS_ERRTAG_IFLUSH_1)) {
3550 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3551 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3552 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3553 goto corrupt_out;
3554 }
3555 if (S_ISREG(VFS_I(ip)->i_mode)) {
3556 if (XFS_TEST_ERROR(
3557 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3558 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3559 mp, XFS_ERRTAG_IFLUSH_3)) {
3560 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3561 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3562 __func__, ip->i_ino, ip);
3563 goto corrupt_out;
3564 }
3565 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3566 if (XFS_TEST_ERROR(
3567 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3568 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3569 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3570 mp, XFS_ERRTAG_IFLUSH_4)) {
3571 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3572 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3573 __func__, ip->i_ino, ip);
3574 goto corrupt_out;
3575 }
3576 }
3577 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3578 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3579 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3580 "%s: detected corrupt incore inode %Lu, "
3581 "total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3582 __func__, ip->i_ino,
3583 ip->i_d.di_nextents + ip->i_d.di_anextents,
3584 ip->i_d.di_nblocks, ip);
3585 goto corrupt_out;
3586 }
3587 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3588 mp, XFS_ERRTAG_IFLUSH_6)) {
3589 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3590 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3591 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3592 goto corrupt_out;
3593 }
3594
3595 /*
3596 * Inode item log recovery for v2 inodes are dependent on the
3597 * di_flushiter count for correct sequencing. We bump the flush
3598 * iteration count so we can detect flushes which postdate a log record
3599 * during recovery. This is redundant as we now log every change and
3600 * hence this can't happen but we need to still do it to ensure
3601 * backwards compatibility with old kernels that predate logging all
3602 * inode changes.
3603 */
3604 if (ip->i_d.di_version < 3)
3605 ip->i_d.di_flushiter++;
3606
3607 /* Check the inline fork data before we write out. */
3608 if (!xfs_inode_verify_forks(ip))
3609 goto corrupt_out;
3610
3611 /*
3612 * Copy the dirty parts of the inode into the on-disk inode. We always
3613 * copy out the core of the inode, because if the inode is dirty at all
3614 * the core must be.
3615 */
3616 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3617
3618 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3619 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3620 ip->i_d.di_flushiter = 0;
3621
3622 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3623 if (XFS_IFORK_Q(ip))
3624 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3625 xfs_inobp_check(mp, bp);
3626
3627 /*
3628 * We've recorded everything logged in the inode, so we'd like to clear
3629 * the ili_fields bits so we don't log and flush things unnecessarily.
3630 * However, we can't stop logging all this information until the data
3631 * we've copied into the disk buffer is written to disk. If we did we
3632 * might overwrite the copy of the inode in the log with all the data
3633 * after re-logging only part of it, and in the face of a crash we
3634 * wouldn't have all the data we need to recover.
3635 *
3636 * What we do is move the bits to the ili_last_fields field. When
3637 * logging the inode, these bits are moved back to the ili_fields field.
3638 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3639 * know that the information those bits represent is permanently on
3640 * disk. As long as the flush completes before the inode is logged
3641 * again, then both ili_fields and ili_last_fields will be cleared.
3642 *
3643 * We can play with the ili_fields bits here, because the inode lock
3644 * must be held exclusively in order to set bits there and the flush
3645 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3646 * done routine can tell whether or not to look in the AIL. Also, store
3647 * the current LSN of the inode so that we can tell whether the item has
3648 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3649 * need the AIL lock, because it is a 64 bit value that cannot be read
3650 * atomically.
3651 */
3652 iip->ili_last_fields = iip->ili_fields;
3653 iip->ili_fields = 0;
3654 iip->ili_fsync_fields = 0;
3655 iip->ili_logged = 1;
3656
3657 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3658 &iip->ili_item.li_lsn);
3659
3660 /*
3661 * Attach the function xfs_iflush_done to the inode's
3662 * buffer. This will remove the inode from the AIL
3663 * and unlock the inode's flush lock when the inode is
3664 * completely written to disk.
3665 */
3666 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3667
3668 /* generate the checksum. */
3669 xfs_dinode_calc_crc(mp, dip);
3670
3671 ASSERT(!list_empty(&bp->b_li_list));
3672 ASSERT(bp->b_iodone != NULL);
3673 return 0;
3674
3675corrupt_out:
3676 return -EFSCORRUPTED;
3677}