Loading...
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
20#include "xfs.h"
21#include "xfs_fs.h"
22#include "xfs_shared.h"
23#include "xfs_format.h"
24#include "xfs_log_format.h"
25#include "xfs_trans_resv.h"
26#include "xfs_sb.h"
27#include "xfs_mount.h"
28#include "xfs_defer.h"
29#include "xfs_inode.h"
30#include "xfs_da_format.h"
31#include "xfs_da_btree.h"
32#include "xfs_dir2.h"
33#include "xfs_attr_sf.h"
34#include "xfs_attr.h"
35#include "xfs_trans_space.h"
36#include "xfs_trans.h"
37#include "xfs_buf_item.h"
38#include "xfs_inode_item.h"
39#include "xfs_ialloc.h"
40#include "xfs_bmap.h"
41#include "xfs_bmap_util.h"
42#include "xfs_error.h"
43#include "xfs_quota.h"
44#include "xfs_filestream.h"
45#include "xfs_cksum.h"
46#include "xfs_trace.h"
47#include "xfs_icache.h"
48#include "xfs_symlink.h"
49#include "xfs_trans_priv.h"
50#include "xfs_log.h"
51#include "xfs_bmap_btree.h"
52#include "xfs_reflink.h"
53
54kmem_zone_t *xfs_inode_zone;
55
56/*
57 * Used in xfs_itruncate_extents(). This is the maximum number of extents
58 * freed from a file in a single transaction.
59 */
60#define XFS_ITRUNC_MAX_EXTENTS 2
61
62STATIC int xfs_iflush_int(struct xfs_inode *, struct xfs_buf *);
63STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
64STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
65
66/*
67 * helper function to extract extent size hint from inode
68 */
69xfs_extlen_t
70xfs_get_extsz_hint(
71 struct xfs_inode *ip)
72{
73 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
74 return ip->i_d.di_extsize;
75 if (XFS_IS_REALTIME_INODE(ip))
76 return ip->i_mount->m_sb.sb_rextsize;
77 return 0;
78}
79
80/*
81 * Helper function to extract CoW extent size hint from inode.
82 * Between the extent size hint and the CoW extent size hint, we
83 * return the greater of the two. If the value is zero (automatic),
84 * use the default size.
85 */
86xfs_extlen_t
87xfs_get_cowextsz_hint(
88 struct xfs_inode *ip)
89{
90 xfs_extlen_t a, b;
91
92 a = 0;
93 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
94 a = ip->i_d.di_cowextsize;
95 b = xfs_get_extsz_hint(ip);
96
97 a = max(a, b);
98 if (a == 0)
99 return XFS_DEFAULT_COWEXTSZ_HINT;
100 return a;
101}
102
103/*
104 * These two are wrapper routines around the xfs_ilock() routine used to
105 * centralize some grungy code. They are used in places that wish to lock the
106 * inode solely for reading the extents. The reason these places can't just
107 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
108 * bringing in of the extents from disk for a file in b-tree format. If the
109 * inode is in b-tree format, then we need to lock the inode exclusively until
110 * the extents are read in. Locking it exclusively all the time would limit
111 * our parallelism unnecessarily, though. What we do instead is check to see
112 * if the extents have been read in yet, and only lock the inode exclusively
113 * if they have not.
114 *
115 * The functions return a value which should be given to the corresponding
116 * xfs_iunlock() call.
117 */
118uint
119xfs_ilock_data_map_shared(
120 struct xfs_inode *ip)
121{
122 uint lock_mode = XFS_ILOCK_SHARED;
123
124 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
125 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
126 lock_mode = XFS_ILOCK_EXCL;
127 xfs_ilock(ip, lock_mode);
128 return lock_mode;
129}
130
131uint
132xfs_ilock_attr_map_shared(
133 struct xfs_inode *ip)
134{
135 uint lock_mode = XFS_ILOCK_SHARED;
136
137 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
138 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
139 lock_mode = XFS_ILOCK_EXCL;
140 xfs_ilock(ip, lock_mode);
141 return lock_mode;
142}
143
144/*
145 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
146 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
147 * various combinations of the locks to be obtained.
148 *
149 * The 3 locks should always be ordered so that the IO lock is obtained first,
150 * the mmap lock second and the ilock last in order to prevent deadlock.
151 *
152 * Basic locking order:
153 *
154 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
155 *
156 * mmap_sem locking order:
157 *
158 * i_rwsem -> page lock -> mmap_sem
159 * mmap_sem -> i_mmap_lock -> page_lock
160 *
161 * The difference in mmap_sem locking order mean that we cannot hold the
162 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
163 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
164 * in get_user_pages() to map the user pages into the kernel address space for
165 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
166 * page faults already hold the mmap_sem.
167 *
168 * Hence to serialise fully against both syscall and mmap based IO, we need to
169 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
170 * taken in places where we need to invalidate the page cache in a race
171 * free manner (e.g. truncate, hole punch and other extent manipulation
172 * functions).
173 */
174void
175xfs_ilock(
176 xfs_inode_t *ip,
177 uint lock_flags)
178{
179 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
180
181 /*
182 * You can't set both SHARED and EXCL for the same lock,
183 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
184 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
185 */
186 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
187 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
188 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
189 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
190 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
191 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
192 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
193
194 if (lock_flags & XFS_IOLOCK_EXCL) {
195 down_write_nested(&VFS_I(ip)->i_rwsem,
196 XFS_IOLOCK_DEP(lock_flags));
197 } else if (lock_flags & XFS_IOLOCK_SHARED) {
198 down_read_nested(&VFS_I(ip)->i_rwsem,
199 XFS_IOLOCK_DEP(lock_flags));
200 }
201
202 if (lock_flags & XFS_MMAPLOCK_EXCL)
203 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
204 else if (lock_flags & XFS_MMAPLOCK_SHARED)
205 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
206
207 if (lock_flags & XFS_ILOCK_EXCL)
208 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
209 else if (lock_flags & XFS_ILOCK_SHARED)
210 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
211}
212
213/*
214 * This is just like xfs_ilock(), except that the caller
215 * is guaranteed not to sleep. It returns 1 if it gets
216 * the requested locks and 0 otherwise. If the IO lock is
217 * obtained but the inode lock cannot be, then the IO lock
218 * is dropped before returning.
219 *
220 * ip -- the inode being locked
221 * lock_flags -- this parameter indicates the inode's locks to be
222 * to be locked. See the comment for xfs_ilock() for a list
223 * of valid values.
224 */
225int
226xfs_ilock_nowait(
227 xfs_inode_t *ip,
228 uint lock_flags)
229{
230 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
231
232 /*
233 * You can't set both SHARED and EXCL for the same lock,
234 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
235 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
236 */
237 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
238 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
239 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
240 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
241 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
242 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
243 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
244
245 if (lock_flags & XFS_IOLOCK_EXCL) {
246 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
247 goto out;
248 } else if (lock_flags & XFS_IOLOCK_SHARED) {
249 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
250 goto out;
251 }
252
253 if (lock_flags & XFS_MMAPLOCK_EXCL) {
254 if (!mrtryupdate(&ip->i_mmaplock))
255 goto out_undo_iolock;
256 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
257 if (!mrtryaccess(&ip->i_mmaplock))
258 goto out_undo_iolock;
259 }
260
261 if (lock_flags & XFS_ILOCK_EXCL) {
262 if (!mrtryupdate(&ip->i_lock))
263 goto out_undo_mmaplock;
264 } else if (lock_flags & XFS_ILOCK_SHARED) {
265 if (!mrtryaccess(&ip->i_lock))
266 goto out_undo_mmaplock;
267 }
268 return 1;
269
270out_undo_mmaplock:
271 if (lock_flags & XFS_MMAPLOCK_EXCL)
272 mrunlock_excl(&ip->i_mmaplock);
273 else if (lock_flags & XFS_MMAPLOCK_SHARED)
274 mrunlock_shared(&ip->i_mmaplock);
275out_undo_iolock:
276 if (lock_flags & XFS_IOLOCK_EXCL)
277 up_write(&VFS_I(ip)->i_rwsem);
278 else if (lock_flags & XFS_IOLOCK_SHARED)
279 up_read(&VFS_I(ip)->i_rwsem);
280out:
281 return 0;
282}
283
284/*
285 * xfs_iunlock() is used to drop the inode locks acquired with
286 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
287 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
288 * that we know which locks to drop.
289 *
290 * ip -- the inode being unlocked
291 * lock_flags -- this parameter indicates the inode's locks to be
292 * to be unlocked. See the comment for xfs_ilock() for a list
293 * of valid values for this parameter.
294 *
295 */
296void
297xfs_iunlock(
298 xfs_inode_t *ip,
299 uint lock_flags)
300{
301 /*
302 * You can't set both SHARED and EXCL for the same lock,
303 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
304 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
305 */
306 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
307 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
308 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
309 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
310 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
311 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
312 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
313 ASSERT(lock_flags != 0);
314
315 if (lock_flags & XFS_IOLOCK_EXCL)
316 up_write(&VFS_I(ip)->i_rwsem);
317 else if (lock_flags & XFS_IOLOCK_SHARED)
318 up_read(&VFS_I(ip)->i_rwsem);
319
320 if (lock_flags & XFS_MMAPLOCK_EXCL)
321 mrunlock_excl(&ip->i_mmaplock);
322 else if (lock_flags & XFS_MMAPLOCK_SHARED)
323 mrunlock_shared(&ip->i_mmaplock);
324
325 if (lock_flags & XFS_ILOCK_EXCL)
326 mrunlock_excl(&ip->i_lock);
327 else if (lock_flags & XFS_ILOCK_SHARED)
328 mrunlock_shared(&ip->i_lock);
329
330 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
331}
332
333/*
334 * give up write locks. the i/o lock cannot be held nested
335 * if it is being demoted.
336 */
337void
338xfs_ilock_demote(
339 xfs_inode_t *ip,
340 uint lock_flags)
341{
342 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
343 ASSERT((lock_flags &
344 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
345
346 if (lock_flags & XFS_ILOCK_EXCL)
347 mrdemote(&ip->i_lock);
348 if (lock_flags & XFS_MMAPLOCK_EXCL)
349 mrdemote(&ip->i_mmaplock);
350 if (lock_flags & XFS_IOLOCK_EXCL)
351 downgrade_write(&VFS_I(ip)->i_rwsem);
352
353 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
354}
355
356#if defined(DEBUG) || defined(XFS_WARN)
357int
358xfs_isilocked(
359 xfs_inode_t *ip,
360 uint lock_flags)
361{
362 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
363 if (!(lock_flags & XFS_ILOCK_SHARED))
364 return !!ip->i_lock.mr_writer;
365 return rwsem_is_locked(&ip->i_lock.mr_lock);
366 }
367
368 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
369 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
370 return !!ip->i_mmaplock.mr_writer;
371 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
372 }
373
374 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
375 if (!(lock_flags & XFS_IOLOCK_SHARED))
376 return !debug_locks ||
377 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
378 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
379 }
380
381 ASSERT(0);
382 return 0;
383}
384#endif
385
386#ifdef DEBUG
387int xfs_locked_n;
388int xfs_small_retries;
389int xfs_middle_retries;
390int xfs_lots_retries;
391int xfs_lock_delays;
392#endif
393
394/*
395 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
396 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
397 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
398 * errors and warnings.
399 */
400#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
401static bool
402xfs_lockdep_subclass_ok(
403 int subclass)
404{
405 return subclass < MAX_LOCKDEP_SUBCLASSES;
406}
407#else
408#define xfs_lockdep_subclass_ok(subclass) (true)
409#endif
410
411/*
412 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
413 * value. This can be called for any type of inode lock combination, including
414 * parent locking. Care must be taken to ensure we don't overrun the subclass
415 * storage fields in the class mask we build.
416 */
417static inline int
418xfs_lock_inumorder(int lock_mode, int subclass)
419{
420 int class = 0;
421
422 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
423 XFS_ILOCK_RTSUM)));
424 ASSERT(xfs_lockdep_subclass_ok(subclass));
425
426 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
427 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
428 class += subclass << XFS_IOLOCK_SHIFT;
429 }
430
431 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
432 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
433 class += subclass << XFS_MMAPLOCK_SHIFT;
434 }
435
436 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
437 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
438 class += subclass << XFS_ILOCK_SHIFT;
439 }
440
441 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
442}
443
444/*
445 * The following routine will lock n inodes in exclusive mode. We assume the
446 * caller calls us with the inodes in i_ino order.
447 *
448 * We need to detect deadlock where an inode that we lock is in the AIL and we
449 * start waiting for another inode that is locked by a thread in a long running
450 * transaction (such as truncate). This can result in deadlock since the long
451 * running trans might need to wait for the inode we just locked in order to
452 * push the tail and free space in the log.
453 *
454 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
455 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
456 * lock more than one at a time, lockdep will report false positives saying we
457 * have violated locking orders.
458 */
459static void
460xfs_lock_inodes(
461 xfs_inode_t **ips,
462 int inodes,
463 uint lock_mode)
464{
465 int attempts = 0, i, j, try_lock;
466 xfs_log_item_t *lp;
467
468 /*
469 * Currently supports between 2 and 5 inodes with exclusive locking. We
470 * support an arbitrary depth of locking here, but absolute limits on
471 * inodes depend on the the type of locking and the limits placed by
472 * lockdep annotations in xfs_lock_inumorder. These are all checked by
473 * the asserts.
474 */
475 ASSERT(ips && inodes >= 2 && inodes <= 5);
476 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
477 XFS_ILOCK_EXCL));
478 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
479 XFS_ILOCK_SHARED)));
480 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
481 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
482 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
483 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
484
485 if (lock_mode & XFS_IOLOCK_EXCL) {
486 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
487 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
488 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
489
490 try_lock = 0;
491 i = 0;
492again:
493 for (; i < inodes; i++) {
494 ASSERT(ips[i]);
495
496 if (i && (ips[i] == ips[i - 1])) /* Already locked */
497 continue;
498
499 /*
500 * If try_lock is not set yet, make sure all locked inodes are
501 * not in the AIL. If any are, set try_lock to be used later.
502 */
503 if (!try_lock) {
504 for (j = (i - 1); j >= 0 && !try_lock; j--) {
505 lp = (xfs_log_item_t *)ips[j]->i_itemp;
506 if (lp && (lp->li_flags & XFS_LI_IN_AIL))
507 try_lock++;
508 }
509 }
510
511 /*
512 * If any of the previous locks we have locked is in the AIL,
513 * we must TRY to get the second and subsequent locks. If
514 * we can't get any, we must release all we have
515 * and try again.
516 */
517 if (!try_lock) {
518 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
519 continue;
520 }
521
522 /* try_lock means we have an inode locked that is in the AIL. */
523 ASSERT(i != 0);
524 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
525 continue;
526
527 /*
528 * Unlock all previous guys and try again. xfs_iunlock will try
529 * to push the tail if the inode is in the AIL.
530 */
531 attempts++;
532 for (j = i - 1; j >= 0; j--) {
533 /*
534 * Check to see if we've already unlocked this one. Not
535 * the first one going back, and the inode ptr is the
536 * same.
537 */
538 if (j != (i - 1) && ips[j] == ips[j + 1])
539 continue;
540
541 xfs_iunlock(ips[j], lock_mode);
542 }
543
544 if ((attempts % 5) == 0) {
545 delay(1); /* Don't just spin the CPU */
546#ifdef DEBUG
547 xfs_lock_delays++;
548#endif
549 }
550 i = 0;
551 try_lock = 0;
552 goto again;
553 }
554
555#ifdef DEBUG
556 if (attempts) {
557 if (attempts < 5) xfs_small_retries++;
558 else if (attempts < 100) xfs_middle_retries++;
559 else xfs_lots_retries++;
560 } else {
561 xfs_locked_n++;
562 }
563#endif
564}
565
566/*
567 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
568 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
569 * lock more than one at a time, lockdep will report false positives saying we
570 * have violated locking orders.
571 */
572void
573xfs_lock_two_inodes(
574 xfs_inode_t *ip0,
575 xfs_inode_t *ip1,
576 uint lock_mode)
577{
578 xfs_inode_t *temp;
579 int attempts = 0;
580 xfs_log_item_t *lp;
581
582 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
583 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL))
584 ASSERT(!(lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
585
586 ASSERT(ip0->i_ino != ip1->i_ino);
587
588 if (ip0->i_ino > ip1->i_ino) {
589 temp = ip0;
590 ip0 = ip1;
591 ip1 = temp;
592 }
593
594 again:
595 xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
596
597 /*
598 * If the first lock we have locked is in the AIL, we must TRY to get
599 * the second lock. If we can't get it, we must release the first one
600 * and try again.
601 */
602 lp = (xfs_log_item_t *)ip0->i_itemp;
603 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
604 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
605 xfs_iunlock(ip0, lock_mode);
606 if ((++attempts % 5) == 0)
607 delay(1); /* Don't just spin the CPU */
608 goto again;
609 }
610 } else {
611 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
612 }
613}
614
615
616void
617__xfs_iflock(
618 struct xfs_inode *ip)
619{
620 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
621 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
622
623 do {
624 prepare_to_wait_exclusive(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
625 if (xfs_isiflocked(ip))
626 io_schedule();
627 } while (!xfs_iflock_nowait(ip));
628
629 finish_wait(wq, &wait.wait);
630}
631
632STATIC uint
633_xfs_dic2xflags(
634 __uint16_t di_flags,
635 uint64_t di_flags2,
636 bool has_attr)
637{
638 uint flags = 0;
639
640 if (di_flags & XFS_DIFLAG_ANY) {
641 if (di_flags & XFS_DIFLAG_REALTIME)
642 flags |= FS_XFLAG_REALTIME;
643 if (di_flags & XFS_DIFLAG_PREALLOC)
644 flags |= FS_XFLAG_PREALLOC;
645 if (di_flags & XFS_DIFLAG_IMMUTABLE)
646 flags |= FS_XFLAG_IMMUTABLE;
647 if (di_flags & XFS_DIFLAG_APPEND)
648 flags |= FS_XFLAG_APPEND;
649 if (di_flags & XFS_DIFLAG_SYNC)
650 flags |= FS_XFLAG_SYNC;
651 if (di_flags & XFS_DIFLAG_NOATIME)
652 flags |= FS_XFLAG_NOATIME;
653 if (di_flags & XFS_DIFLAG_NODUMP)
654 flags |= FS_XFLAG_NODUMP;
655 if (di_flags & XFS_DIFLAG_RTINHERIT)
656 flags |= FS_XFLAG_RTINHERIT;
657 if (di_flags & XFS_DIFLAG_PROJINHERIT)
658 flags |= FS_XFLAG_PROJINHERIT;
659 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
660 flags |= FS_XFLAG_NOSYMLINKS;
661 if (di_flags & XFS_DIFLAG_EXTSIZE)
662 flags |= FS_XFLAG_EXTSIZE;
663 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
664 flags |= FS_XFLAG_EXTSZINHERIT;
665 if (di_flags & XFS_DIFLAG_NODEFRAG)
666 flags |= FS_XFLAG_NODEFRAG;
667 if (di_flags & XFS_DIFLAG_FILESTREAM)
668 flags |= FS_XFLAG_FILESTREAM;
669 }
670
671 if (di_flags2 & XFS_DIFLAG2_ANY) {
672 if (di_flags2 & XFS_DIFLAG2_DAX)
673 flags |= FS_XFLAG_DAX;
674 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
675 flags |= FS_XFLAG_COWEXTSIZE;
676 }
677
678 if (has_attr)
679 flags |= FS_XFLAG_HASATTR;
680
681 return flags;
682}
683
684uint
685xfs_ip2xflags(
686 struct xfs_inode *ip)
687{
688 struct xfs_icdinode *dic = &ip->i_d;
689
690 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
691}
692
693/*
694 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
695 * is allowed, otherwise it has to be an exact match. If a CI match is found,
696 * ci_name->name will point to a the actual name (caller must free) or
697 * will be set to NULL if an exact match is found.
698 */
699int
700xfs_lookup(
701 xfs_inode_t *dp,
702 struct xfs_name *name,
703 xfs_inode_t **ipp,
704 struct xfs_name *ci_name)
705{
706 xfs_ino_t inum;
707 int error;
708
709 trace_xfs_lookup(dp, name);
710
711 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
712 return -EIO;
713
714 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
715 if (error)
716 goto out_unlock;
717
718 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
719 if (error)
720 goto out_free_name;
721
722 return 0;
723
724out_free_name:
725 if (ci_name)
726 kmem_free(ci_name->name);
727out_unlock:
728 *ipp = NULL;
729 return error;
730}
731
732/*
733 * Allocate an inode on disk and return a copy of its in-core version.
734 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
735 * appropriately within the inode. The uid and gid for the inode are
736 * set according to the contents of the given cred structure.
737 *
738 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
739 * has a free inode available, call xfs_iget() to obtain the in-core
740 * version of the allocated inode. Finally, fill in the inode and
741 * log its initial contents. In this case, ialloc_context would be
742 * set to NULL.
743 *
744 * If xfs_dialloc() does not have an available inode, it will replenish
745 * its supply by doing an allocation. Since we can only do one
746 * allocation within a transaction without deadlocks, we must commit
747 * the current transaction before returning the inode itself.
748 * In this case, therefore, we will set ialloc_context and return.
749 * The caller should then commit the current transaction, start a new
750 * transaction, and call xfs_ialloc() again to actually get the inode.
751 *
752 * To ensure that some other process does not grab the inode that
753 * was allocated during the first call to xfs_ialloc(), this routine
754 * also returns the [locked] bp pointing to the head of the freelist
755 * as ialloc_context. The caller should hold this buffer across
756 * the commit and pass it back into this routine on the second call.
757 *
758 * If we are allocating quota inodes, we do not have a parent inode
759 * to attach to or associate with (i.e. pip == NULL) because they
760 * are not linked into the directory structure - they are attached
761 * directly to the superblock - and so have no parent.
762 */
763static int
764xfs_ialloc(
765 xfs_trans_t *tp,
766 xfs_inode_t *pip,
767 umode_t mode,
768 xfs_nlink_t nlink,
769 xfs_dev_t rdev,
770 prid_t prid,
771 int okalloc,
772 xfs_buf_t **ialloc_context,
773 xfs_inode_t **ipp)
774{
775 struct xfs_mount *mp = tp->t_mountp;
776 xfs_ino_t ino;
777 xfs_inode_t *ip;
778 uint flags;
779 int error;
780 struct timespec tv;
781 struct inode *inode;
782
783 /*
784 * Call the space management code to pick
785 * the on-disk inode to be allocated.
786 */
787 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode, okalloc,
788 ialloc_context, &ino);
789 if (error)
790 return error;
791 if (*ialloc_context || ino == NULLFSINO) {
792 *ipp = NULL;
793 return 0;
794 }
795 ASSERT(*ialloc_context == NULL);
796
797 /*
798 * Get the in-core inode with the lock held exclusively.
799 * This is because we're setting fields here we need
800 * to prevent others from looking at until we're done.
801 */
802 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
803 XFS_ILOCK_EXCL, &ip);
804 if (error)
805 return error;
806 ASSERT(ip != NULL);
807 inode = VFS_I(ip);
808
809 /*
810 * We always convert v1 inodes to v2 now - we only support filesystems
811 * with >= v2 inode capability, so there is no reason for ever leaving
812 * an inode in v1 format.
813 */
814 if (ip->i_d.di_version == 1)
815 ip->i_d.di_version = 2;
816
817 inode->i_mode = mode;
818 set_nlink(inode, nlink);
819 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
820 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
821 xfs_set_projid(ip, prid);
822
823 if (pip && XFS_INHERIT_GID(pip)) {
824 ip->i_d.di_gid = pip->i_d.di_gid;
825 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
826 inode->i_mode |= S_ISGID;
827 }
828
829 /*
830 * If the group ID of the new file does not match the effective group
831 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
832 * (and only if the irix_sgid_inherit compatibility variable is set).
833 */
834 if ((irix_sgid_inherit) &&
835 (inode->i_mode & S_ISGID) &&
836 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
837 inode->i_mode &= ~S_ISGID;
838
839 ip->i_d.di_size = 0;
840 ip->i_d.di_nextents = 0;
841 ASSERT(ip->i_d.di_nblocks == 0);
842
843 tv = current_time(inode);
844 inode->i_mtime = tv;
845 inode->i_atime = tv;
846 inode->i_ctime = tv;
847
848 ip->i_d.di_extsize = 0;
849 ip->i_d.di_dmevmask = 0;
850 ip->i_d.di_dmstate = 0;
851 ip->i_d.di_flags = 0;
852
853 if (ip->i_d.di_version == 3) {
854 inode->i_version = 1;
855 ip->i_d.di_flags2 = 0;
856 ip->i_d.di_cowextsize = 0;
857 ip->i_d.di_crtime.t_sec = (__int32_t)tv.tv_sec;
858 ip->i_d.di_crtime.t_nsec = (__int32_t)tv.tv_nsec;
859 }
860
861
862 flags = XFS_ILOG_CORE;
863 switch (mode & S_IFMT) {
864 case S_IFIFO:
865 case S_IFCHR:
866 case S_IFBLK:
867 case S_IFSOCK:
868 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
869 ip->i_df.if_u2.if_rdev = rdev;
870 ip->i_df.if_flags = 0;
871 flags |= XFS_ILOG_DEV;
872 break;
873 case S_IFREG:
874 case S_IFDIR:
875 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
876 uint64_t di_flags2 = 0;
877 uint di_flags = 0;
878
879 if (S_ISDIR(mode)) {
880 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
881 di_flags |= XFS_DIFLAG_RTINHERIT;
882 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
883 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
884 ip->i_d.di_extsize = pip->i_d.di_extsize;
885 }
886 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
887 di_flags |= XFS_DIFLAG_PROJINHERIT;
888 } else if (S_ISREG(mode)) {
889 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
890 di_flags |= XFS_DIFLAG_REALTIME;
891 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
892 di_flags |= XFS_DIFLAG_EXTSIZE;
893 ip->i_d.di_extsize = pip->i_d.di_extsize;
894 }
895 }
896 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
897 xfs_inherit_noatime)
898 di_flags |= XFS_DIFLAG_NOATIME;
899 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
900 xfs_inherit_nodump)
901 di_flags |= XFS_DIFLAG_NODUMP;
902 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
903 xfs_inherit_sync)
904 di_flags |= XFS_DIFLAG_SYNC;
905 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
906 xfs_inherit_nosymlinks)
907 di_flags |= XFS_DIFLAG_NOSYMLINKS;
908 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
909 xfs_inherit_nodefrag)
910 di_flags |= XFS_DIFLAG_NODEFRAG;
911 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
912 di_flags |= XFS_DIFLAG_FILESTREAM;
913 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
914 di_flags2 |= XFS_DIFLAG2_DAX;
915
916 ip->i_d.di_flags |= di_flags;
917 ip->i_d.di_flags2 |= di_flags2;
918 }
919 if (pip &&
920 (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
921 pip->i_d.di_version == 3 &&
922 ip->i_d.di_version == 3) {
923 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
924 ip->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
925 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
926 }
927 }
928 /* FALLTHROUGH */
929 case S_IFLNK:
930 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
931 ip->i_df.if_flags = XFS_IFEXTENTS;
932 ip->i_df.if_bytes = ip->i_df.if_real_bytes = 0;
933 ip->i_df.if_u1.if_extents = NULL;
934 break;
935 default:
936 ASSERT(0);
937 }
938 /*
939 * Attribute fork settings for new inode.
940 */
941 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
942 ip->i_d.di_anextents = 0;
943
944 /*
945 * Log the new values stuffed into the inode.
946 */
947 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
948 xfs_trans_log_inode(tp, ip, flags);
949
950 /* now that we have an i_mode we can setup the inode structure */
951 xfs_setup_inode(ip);
952
953 *ipp = ip;
954 return 0;
955}
956
957/*
958 * Allocates a new inode from disk and return a pointer to the
959 * incore copy. This routine will internally commit the current
960 * transaction and allocate a new one if the Space Manager needed
961 * to do an allocation to replenish the inode free-list.
962 *
963 * This routine is designed to be called from xfs_create and
964 * xfs_create_dir.
965 *
966 */
967int
968xfs_dir_ialloc(
969 xfs_trans_t **tpp, /* input: current transaction;
970 output: may be a new transaction. */
971 xfs_inode_t *dp, /* directory within whose allocate
972 the inode. */
973 umode_t mode,
974 xfs_nlink_t nlink,
975 xfs_dev_t rdev,
976 prid_t prid, /* project id */
977 int okalloc, /* ok to allocate new space */
978 xfs_inode_t **ipp, /* pointer to inode; it will be
979 locked. */
980 int *committed)
981
982{
983 xfs_trans_t *tp;
984 xfs_inode_t *ip;
985 xfs_buf_t *ialloc_context = NULL;
986 int code;
987 void *dqinfo;
988 uint tflags;
989
990 tp = *tpp;
991 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
992
993 /*
994 * xfs_ialloc will return a pointer to an incore inode if
995 * the Space Manager has an available inode on the free
996 * list. Otherwise, it will do an allocation and replenish
997 * the freelist. Since we can only do one allocation per
998 * transaction without deadlocks, we will need to commit the
999 * current transaction and start a new one. We will then
1000 * need to call xfs_ialloc again to get the inode.
1001 *
1002 * If xfs_ialloc did an allocation to replenish the freelist,
1003 * it returns the bp containing the head of the freelist as
1004 * ialloc_context. We will hold a lock on it across the
1005 * transaction commit so that no other process can steal
1006 * the inode(s) that we've just allocated.
1007 */
1008 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, okalloc,
1009 &ialloc_context, &ip);
1010
1011 /*
1012 * Return an error if we were unable to allocate a new inode.
1013 * This should only happen if we run out of space on disk or
1014 * encounter a disk error.
1015 */
1016 if (code) {
1017 *ipp = NULL;
1018 return code;
1019 }
1020 if (!ialloc_context && !ip) {
1021 *ipp = NULL;
1022 return -ENOSPC;
1023 }
1024
1025 /*
1026 * If the AGI buffer is non-NULL, then we were unable to get an
1027 * inode in one operation. We need to commit the current
1028 * transaction and call xfs_ialloc() again. It is guaranteed
1029 * to succeed the second time.
1030 */
1031 if (ialloc_context) {
1032 /*
1033 * Normally, xfs_trans_commit releases all the locks.
1034 * We call bhold to hang on to the ialloc_context across
1035 * the commit. Holding this buffer prevents any other
1036 * processes from doing any allocations in this
1037 * allocation group.
1038 */
1039 xfs_trans_bhold(tp, ialloc_context);
1040
1041 /*
1042 * We want the quota changes to be associated with the next
1043 * transaction, NOT this one. So, detach the dqinfo from this
1044 * and attach it to the next transaction.
1045 */
1046 dqinfo = NULL;
1047 tflags = 0;
1048 if (tp->t_dqinfo) {
1049 dqinfo = (void *)tp->t_dqinfo;
1050 tp->t_dqinfo = NULL;
1051 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1052 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1053 }
1054
1055 code = xfs_trans_roll(&tp, NULL);
1056 if (committed != NULL)
1057 *committed = 1;
1058
1059 /*
1060 * Re-attach the quota info that we detached from prev trx.
1061 */
1062 if (dqinfo) {
1063 tp->t_dqinfo = dqinfo;
1064 tp->t_flags |= tflags;
1065 }
1066
1067 if (code) {
1068 xfs_buf_relse(ialloc_context);
1069 *tpp = tp;
1070 *ipp = NULL;
1071 return code;
1072 }
1073 xfs_trans_bjoin(tp, ialloc_context);
1074
1075 /*
1076 * Call ialloc again. Since we've locked out all
1077 * other allocations in this allocation group,
1078 * this call should always succeed.
1079 */
1080 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1081 okalloc, &ialloc_context, &ip);
1082
1083 /*
1084 * If we get an error at this point, return to the caller
1085 * so that the current transaction can be aborted.
1086 */
1087 if (code) {
1088 *tpp = tp;
1089 *ipp = NULL;
1090 return code;
1091 }
1092 ASSERT(!ialloc_context && ip);
1093
1094 } else {
1095 if (committed != NULL)
1096 *committed = 0;
1097 }
1098
1099 *ipp = ip;
1100 *tpp = tp;
1101
1102 return 0;
1103}
1104
1105/*
1106 * Decrement the link count on an inode & log the change. If this causes the
1107 * link count to go to zero, move the inode to AGI unlinked list so that it can
1108 * be freed when the last active reference goes away via xfs_inactive().
1109 */
1110static int /* error */
1111xfs_droplink(
1112 xfs_trans_t *tp,
1113 xfs_inode_t *ip)
1114{
1115 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1116
1117 drop_nlink(VFS_I(ip));
1118 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1119
1120 if (VFS_I(ip)->i_nlink)
1121 return 0;
1122
1123 return xfs_iunlink(tp, ip);
1124}
1125
1126/*
1127 * Increment the link count on an inode & log the change.
1128 */
1129static int
1130xfs_bumplink(
1131 xfs_trans_t *tp,
1132 xfs_inode_t *ip)
1133{
1134 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1135
1136 ASSERT(ip->i_d.di_version > 1);
1137 inc_nlink(VFS_I(ip));
1138 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1139 return 0;
1140}
1141
1142int
1143xfs_create(
1144 xfs_inode_t *dp,
1145 struct xfs_name *name,
1146 umode_t mode,
1147 xfs_dev_t rdev,
1148 xfs_inode_t **ipp)
1149{
1150 int is_dir = S_ISDIR(mode);
1151 struct xfs_mount *mp = dp->i_mount;
1152 struct xfs_inode *ip = NULL;
1153 struct xfs_trans *tp = NULL;
1154 int error;
1155 struct xfs_defer_ops dfops;
1156 xfs_fsblock_t first_block;
1157 bool unlock_dp_on_error = false;
1158 prid_t prid;
1159 struct xfs_dquot *udqp = NULL;
1160 struct xfs_dquot *gdqp = NULL;
1161 struct xfs_dquot *pdqp = NULL;
1162 struct xfs_trans_res *tres;
1163 uint resblks;
1164
1165 trace_xfs_create(dp, name);
1166
1167 if (XFS_FORCED_SHUTDOWN(mp))
1168 return -EIO;
1169
1170 prid = xfs_get_initial_prid(dp);
1171
1172 /*
1173 * Make sure that we have allocated dquot(s) on disk.
1174 */
1175 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1176 xfs_kgid_to_gid(current_fsgid()), prid,
1177 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1178 &udqp, &gdqp, &pdqp);
1179 if (error)
1180 return error;
1181
1182 if (is_dir) {
1183 rdev = 0;
1184 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1185 tres = &M_RES(mp)->tr_mkdir;
1186 } else {
1187 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1188 tres = &M_RES(mp)->tr_create;
1189 }
1190
1191 /*
1192 * Initially assume that the file does not exist and
1193 * reserve the resources for that case. If that is not
1194 * the case we'll drop the one we have and get a more
1195 * appropriate transaction later.
1196 */
1197 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1198 if (error == -ENOSPC) {
1199 /* flush outstanding delalloc blocks and retry */
1200 xfs_flush_inodes(mp);
1201 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1202 }
1203 if (error == -ENOSPC) {
1204 /* No space at all so try a "no-allocation" reservation */
1205 resblks = 0;
1206 error = xfs_trans_alloc(mp, tres, 0, 0, 0, &tp);
1207 }
1208 if (error)
1209 goto out_release_inode;
1210
1211 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1212 unlock_dp_on_error = true;
1213
1214 xfs_defer_init(&dfops, &first_block);
1215
1216 /*
1217 * Reserve disk quota and the inode.
1218 */
1219 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1220 pdqp, resblks, 1, 0);
1221 if (error)
1222 goto out_trans_cancel;
1223
1224 if (!resblks) {
1225 error = xfs_dir_canenter(tp, dp, name);
1226 if (error)
1227 goto out_trans_cancel;
1228 }
1229
1230 /*
1231 * A newly created regular or special file just has one directory
1232 * entry pointing to them, but a directory also the "." entry
1233 * pointing to itself.
1234 */
1235 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1236 prid, resblks > 0, &ip, NULL);
1237 if (error)
1238 goto out_trans_cancel;
1239
1240 /*
1241 * Now we join the directory inode to the transaction. We do not do it
1242 * earlier because xfs_dir_ialloc might commit the previous transaction
1243 * (and release all the locks). An error from here on will result in
1244 * the transaction cancel unlocking dp so don't do it explicitly in the
1245 * error path.
1246 */
1247 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1248 unlock_dp_on_error = false;
1249
1250 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1251 &first_block, &dfops, resblks ?
1252 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1253 if (error) {
1254 ASSERT(error != -ENOSPC);
1255 goto out_trans_cancel;
1256 }
1257 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1258 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1259
1260 if (is_dir) {
1261 error = xfs_dir_init(tp, ip, dp);
1262 if (error)
1263 goto out_bmap_cancel;
1264
1265 error = xfs_bumplink(tp, dp);
1266 if (error)
1267 goto out_bmap_cancel;
1268 }
1269
1270 /*
1271 * If this is a synchronous mount, make sure that the
1272 * create transaction goes to disk before returning to
1273 * the user.
1274 */
1275 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1276 xfs_trans_set_sync(tp);
1277
1278 /*
1279 * Attach the dquot(s) to the inodes and modify them incore.
1280 * These ids of the inode couldn't have changed since the new
1281 * inode has been locked ever since it was created.
1282 */
1283 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1284
1285 error = xfs_defer_finish(&tp, &dfops, NULL);
1286 if (error)
1287 goto out_bmap_cancel;
1288
1289 error = xfs_trans_commit(tp);
1290 if (error)
1291 goto out_release_inode;
1292
1293 xfs_qm_dqrele(udqp);
1294 xfs_qm_dqrele(gdqp);
1295 xfs_qm_dqrele(pdqp);
1296
1297 *ipp = ip;
1298 return 0;
1299
1300 out_bmap_cancel:
1301 xfs_defer_cancel(&dfops);
1302 out_trans_cancel:
1303 xfs_trans_cancel(tp);
1304 out_release_inode:
1305 /*
1306 * Wait until after the current transaction is aborted to finish the
1307 * setup of the inode and release the inode. This prevents recursive
1308 * transactions and deadlocks from xfs_inactive.
1309 */
1310 if (ip) {
1311 xfs_finish_inode_setup(ip);
1312 IRELE(ip);
1313 }
1314
1315 xfs_qm_dqrele(udqp);
1316 xfs_qm_dqrele(gdqp);
1317 xfs_qm_dqrele(pdqp);
1318
1319 if (unlock_dp_on_error)
1320 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1321 return error;
1322}
1323
1324int
1325xfs_create_tmpfile(
1326 struct xfs_inode *dp,
1327 struct dentry *dentry,
1328 umode_t mode,
1329 struct xfs_inode **ipp)
1330{
1331 struct xfs_mount *mp = dp->i_mount;
1332 struct xfs_inode *ip = NULL;
1333 struct xfs_trans *tp = NULL;
1334 int error;
1335 prid_t prid;
1336 struct xfs_dquot *udqp = NULL;
1337 struct xfs_dquot *gdqp = NULL;
1338 struct xfs_dquot *pdqp = NULL;
1339 struct xfs_trans_res *tres;
1340 uint resblks;
1341
1342 if (XFS_FORCED_SHUTDOWN(mp))
1343 return -EIO;
1344
1345 prid = xfs_get_initial_prid(dp);
1346
1347 /*
1348 * Make sure that we have allocated dquot(s) on disk.
1349 */
1350 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1351 xfs_kgid_to_gid(current_fsgid()), prid,
1352 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1353 &udqp, &gdqp, &pdqp);
1354 if (error)
1355 return error;
1356
1357 resblks = XFS_IALLOC_SPACE_RES(mp);
1358 tres = &M_RES(mp)->tr_create_tmpfile;
1359
1360 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1361 if (error == -ENOSPC) {
1362 /* No space at all so try a "no-allocation" reservation */
1363 resblks = 0;
1364 error = xfs_trans_alloc(mp, tres, 0, 0, 0, &tp);
1365 }
1366 if (error)
1367 goto out_release_inode;
1368
1369 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1370 pdqp, resblks, 1, 0);
1371 if (error)
1372 goto out_trans_cancel;
1373
1374 error = xfs_dir_ialloc(&tp, dp, mode, 1, 0,
1375 prid, resblks > 0, &ip, NULL);
1376 if (error)
1377 goto out_trans_cancel;
1378
1379 if (mp->m_flags & XFS_MOUNT_WSYNC)
1380 xfs_trans_set_sync(tp);
1381
1382 /*
1383 * Attach the dquot(s) to the inodes and modify them incore.
1384 * These ids of the inode couldn't have changed since the new
1385 * inode has been locked ever since it was created.
1386 */
1387 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1388
1389 error = xfs_iunlink(tp, ip);
1390 if (error)
1391 goto out_trans_cancel;
1392
1393 error = xfs_trans_commit(tp);
1394 if (error)
1395 goto out_release_inode;
1396
1397 xfs_qm_dqrele(udqp);
1398 xfs_qm_dqrele(gdqp);
1399 xfs_qm_dqrele(pdqp);
1400
1401 *ipp = ip;
1402 return 0;
1403
1404 out_trans_cancel:
1405 xfs_trans_cancel(tp);
1406 out_release_inode:
1407 /*
1408 * Wait until after the current transaction is aborted to finish the
1409 * setup of the inode and release the inode. This prevents recursive
1410 * transactions and deadlocks from xfs_inactive.
1411 */
1412 if (ip) {
1413 xfs_finish_inode_setup(ip);
1414 IRELE(ip);
1415 }
1416
1417 xfs_qm_dqrele(udqp);
1418 xfs_qm_dqrele(gdqp);
1419 xfs_qm_dqrele(pdqp);
1420
1421 return error;
1422}
1423
1424int
1425xfs_link(
1426 xfs_inode_t *tdp,
1427 xfs_inode_t *sip,
1428 struct xfs_name *target_name)
1429{
1430 xfs_mount_t *mp = tdp->i_mount;
1431 xfs_trans_t *tp;
1432 int error;
1433 struct xfs_defer_ops dfops;
1434 xfs_fsblock_t first_block;
1435 int resblks;
1436
1437 trace_xfs_link(tdp, target_name);
1438
1439 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1440
1441 if (XFS_FORCED_SHUTDOWN(mp))
1442 return -EIO;
1443
1444 error = xfs_qm_dqattach(sip, 0);
1445 if (error)
1446 goto std_return;
1447
1448 error = xfs_qm_dqattach(tdp, 0);
1449 if (error)
1450 goto std_return;
1451
1452 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1453 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1454 if (error == -ENOSPC) {
1455 resblks = 0;
1456 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1457 }
1458 if (error)
1459 goto std_return;
1460
1461 xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1462
1463 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1464 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1465
1466 /*
1467 * If we are using project inheritance, we only allow hard link
1468 * creation in our tree when the project IDs are the same; else
1469 * the tree quota mechanism could be circumvented.
1470 */
1471 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1472 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1473 error = -EXDEV;
1474 goto error_return;
1475 }
1476
1477 if (!resblks) {
1478 error = xfs_dir_canenter(tp, tdp, target_name);
1479 if (error)
1480 goto error_return;
1481 }
1482
1483 xfs_defer_init(&dfops, &first_block);
1484
1485 /*
1486 * Handle initial link state of O_TMPFILE inode
1487 */
1488 if (VFS_I(sip)->i_nlink == 0) {
1489 error = xfs_iunlink_remove(tp, sip);
1490 if (error)
1491 goto error_return;
1492 }
1493
1494 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1495 &first_block, &dfops, resblks);
1496 if (error)
1497 goto error_return;
1498 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1499 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1500
1501 error = xfs_bumplink(tp, sip);
1502 if (error)
1503 goto error_return;
1504
1505 /*
1506 * If this is a synchronous mount, make sure that the
1507 * link transaction goes to disk before returning to
1508 * the user.
1509 */
1510 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1511 xfs_trans_set_sync(tp);
1512
1513 error = xfs_defer_finish(&tp, &dfops, NULL);
1514 if (error) {
1515 xfs_defer_cancel(&dfops);
1516 goto error_return;
1517 }
1518
1519 return xfs_trans_commit(tp);
1520
1521 error_return:
1522 xfs_trans_cancel(tp);
1523 std_return:
1524 return error;
1525}
1526
1527/*
1528 * Free up the underlying blocks past new_size. The new size must be smaller
1529 * than the current size. This routine can be used both for the attribute and
1530 * data fork, and does not modify the inode size, which is left to the caller.
1531 *
1532 * The transaction passed to this routine must have made a permanent log
1533 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1534 * given transaction and start new ones, so make sure everything involved in
1535 * the transaction is tidy before calling here. Some transaction will be
1536 * returned to the caller to be committed. The incoming transaction must
1537 * already include the inode, and both inode locks must be held exclusively.
1538 * The inode must also be "held" within the transaction. On return the inode
1539 * will be "held" within the returned transaction. This routine does NOT
1540 * require any disk space to be reserved for it within the transaction.
1541 *
1542 * If we get an error, we must return with the inode locked and linked into the
1543 * current transaction. This keeps things simple for the higher level code,
1544 * because it always knows that the inode is locked and held in the transaction
1545 * that returns to it whether errors occur or not. We don't mark the inode
1546 * dirty on error so that transactions can be easily aborted if possible.
1547 */
1548int
1549xfs_itruncate_extents(
1550 struct xfs_trans **tpp,
1551 struct xfs_inode *ip,
1552 int whichfork,
1553 xfs_fsize_t new_size)
1554{
1555 struct xfs_mount *mp = ip->i_mount;
1556 struct xfs_trans *tp = *tpp;
1557 struct xfs_defer_ops dfops;
1558 xfs_fsblock_t first_block;
1559 xfs_fileoff_t first_unmap_block;
1560 xfs_fileoff_t last_block;
1561 xfs_filblks_t unmap_len;
1562 int error = 0;
1563 int done = 0;
1564
1565 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1566 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1567 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1568 ASSERT(new_size <= XFS_ISIZE(ip));
1569 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1570 ASSERT(ip->i_itemp != NULL);
1571 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1572 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1573
1574 trace_xfs_itruncate_extents_start(ip, new_size);
1575
1576 /*
1577 * Since it is possible for space to become allocated beyond
1578 * the end of the file (in a crash where the space is allocated
1579 * but the inode size is not yet updated), simply remove any
1580 * blocks which show up between the new EOF and the maximum
1581 * possible file size. If the first block to be removed is
1582 * beyond the maximum file size (ie it is the same as last_block),
1583 * then there is nothing to do.
1584 */
1585 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1586 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1587 if (first_unmap_block == last_block)
1588 return 0;
1589
1590 ASSERT(first_unmap_block < last_block);
1591 unmap_len = last_block - first_unmap_block + 1;
1592 while (!done) {
1593 xfs_defer_init(&dfops, &first_block);
1594 error = xfs_bunmapi(tp, ip,
1595 first_unmap_block, unmap_len,
1596 xfs_bmapi_aflag(whichfork),
1597 XFS_ITRUNC_MAX_EXTENTS,
1598 &first_block, &dfops,
1599 &done);
1600 if (error)
1601 goto out_bmap_cancel;
1602
1603 /*
1604 * Duplicate the transaction that has the permanent
1605 * reservation and commit the old transaction.
1606 */
1607 error = xfs_defer_finish(&tp, &dfops, ip);
1608 if (error)
1609 goto out_bmap_cancel;
1610
1611 error = xfs_trans_roll(&tp, ip);
1612 if (error)
1613 goto out;
1614 }
1615
1616 /* Remove all pending CoW reservations. */
1617 error = xfs_reflink_cancel_cow_blocks(ip, &tp, first_unmap_block,
1618 last_block, true);
1619 if (error)
1620 goto out;
1621
1622 /*
1623 * Clear the reflink flag if we truncated everything.
1624 */
1625 if (ip->i_d.di_nblocks == 0 && xfs_is_reflink_inode(ip)) {
1626 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1627 xfs_inode_clear_cowblocks_tag(ip);
1628 }
1629
1630 /*
1631 * Always re-log the inode so that our permanent transaction can keep
1632 * on rolling it forward in the log.
1633 */
1634 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1635
1636 trace_xfs_itruncate_extents_end(ip, new_size);
1637
1638out:
1639 *tpp = tp;
1640 return error;
1641out_bmap_cancel:
1642 /*
1643 * If the bunmapi call encounters an error, return to the caller where
1644 * the transaction can be properly aborted. We just need to make sure
1645 * we're not holding any resources that we were not when we came in.
1646 */
1647 xfs_defer_cancel(&dfops);
1648 goto out;
1649}
1650
1651int
1652xfs_release(
1653 xfs_inode_t *ip)
1654{
1655 xfs_mount_t *mp = ip->i_mount;
1656 int error;
1657
1658 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1659 return 0;
1660
1661 /* If this is a read-only mount, don't do this (would generate I/O) */
1662 if (mp->m_flags & XFS_MOUNT_RDONLY)
1663 return 0;
1664
1665 if (!XFS_FORCED_SHUTDOWN(mp)) {
1666 int truncated;
1667
1668 /*
1669 * If we previously truncated this file and removed old data
1670 * in the process, we want to initiate "early" writeout on
1671 * the last close. This is an attempt to combat the notorious
1672 * NULL files problem which is particularly noticeable from a
1673 * truncate down, buffered (re-)write (delalloc), followed by
1674 * a crash. What we are effectively doing here is
1675 * significantly reducing the time window where we'd otherwise
1676 * be exposed to that problem.
1677 */
1678 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1679 if (truncated) {
1680 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1681 if (ip->i_delayed_blks > 0) {
1682 error = filemap_flush(VFS_I(ip)->i_mapping);
1683 if (error)
1684 return error;
1685 }
1686 }
1687 }
1688
1689 if (VFS_I(ip)->i_nlink == 0)
1690 return 0;
1691
1692 if (xfs_can_free_eofblocks(ip, false)) {
1693
1694 /*
1695 * Check if the inode is being opened, written and closed
1696 * frequently and we have delayed allocation blocks outstanding
1697 * (e.g. streaming writes from the NFS server), truncating the
1698 * blocks past EOF will cause fragmentation to occur.
1699 *
1700 * In this case don't do the truncation, but we have to be
1701 * careful how we detect this case. Blocks beyond EOF show up as
1702 * i_delayed_blks even when the inode is clean, so we need to
1703 * truncate them away first before checking for a dirty release.
1704 * Hence on the first dirty close we will still remove the
1705 * speculative allocation, but after that we will leave it in
1706 * place.
1707 */
1708 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1709 return 0;
1710 /*
1711 * If we can't get the iolock just skip truncating the blocks
1712 * past EOF because we could deadlock with the mmap_sem
1713 * otherwise. We'll get another chance to drop them once the
1714 * last reference to the inode is dropped, so we'll never leak
1715 * blocks permanently.
1716 */
1717 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1718 error = xfs_free_eofblocks(ip);
1719 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1720 if (error)
1721 return error;
1722 }
1723
1724 /* delalloc blocks after truncation means it really is dirty */
1725 if (ip->i_delayed_blks)
1726 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1727 }
1728 return 0;
1729}
1730
1731/*
1732 * xfs_inactive_truncate
1733 *
1734 * Called to perform a truncate when an inode becomes unlinked.
1735 */
1736STATIC int
1737xfs_inactive_truncate(
1738 struct xfs_inode *ip)
1739{
1740 struct xfs_mount *mp = ip->i_mount;
1741 struct xfs_trans *tp;
1742 int error;
1743
1744 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1745 if (error) {
1746 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1747 return error;
1748 }
1749
1750 xfs_ilock(ip, XFS_ILOCK_EXCL);
1751 xfs_trans_ijoin(tp, ip, 0);
1752
1753 /*
1754 * Log the inode size first to prevent stale data exposure in the event
1755 * of a system crash before the truncate completes. See the related
1756 * comment in xfs_vn_setattr_size() for details.
1757 */
1758 ip->i_d.di_size = 0;
1759 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1760
1761 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1762 if (error)
1763 goto error_trans_cancel;
1764
1765 ASSERT(ip->i_d.di_nextents == 0);
1766
1767 error = xfs_trans_commit(tp);
1768 if (error)
1769 goto error_unlock;
1770
1771 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1772 return 0;
1773
1774error_trans_cancel:
1775 xfs_trans_cancel(tp);
1776error_unlock:
1777 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1778 return error;
1779}
1780
1781/*
1782 * xfs_inactive_ifree()
1783 *
1784 * Perform the inode free when an inode is unlinked.
1785 */
1786STATIC int
1787xfs_inactive_ifree(
1788 struct xfs_inode *ip)
1789{
1790 struct xfs_defer_ops dfops;
1791 xfs_fsblock_t first_block;
1792 struct xfs_mount *mp = ip->i_mount;
1793 struct xfs_trans *tp;
1794 int error;
1795
1796 /*
1797 * We try to use a per-AG reservation for any block needed by the finobt
1798 * tree, but as the finobt feature predates the per-AG reservation
1799 * support a degraded file system might not have enough space for the
1800 * reservation at mount time. In that case try to dip into the reserved
1801 * pool and pray.
1802 *
1803 * Send a warning if the reservation does happen to fail, as the inode
1804 * now remains allocated and sits on the unlinked list until the fs is
1805 * repaired.
1806 */
1807 if (unlikely(mp->m_inotbt_nores)) {
1808 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1809 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1810 &tp);
1811 } else {
1812 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1813 }
1814 if (error) {
1815 if (error == -ENOSPC) {
1816 xfs_warn_ratelimited(mp,
1817 "Failed to remove inode(s) from unlinked list. "
1818 "Please free space, unmount and run xfs_repair.");
1819 } else {
1820 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1821 }
1822 return error;
1823 }
1824
1825 xfs_ilock(ip, XFS_ILOCK_EXCL);
1826 xfs_trans_ijoin(tp, ip, 0);
1827
1828 xfs_defer_init(&dfops, &first_block);
1829 error = xfs_ifree(tp, ip, &dfops);
1830 if (error) {
1831 /*
1832 * If we fail to free the inode, shut down. The cancel
1833 * might do that, we need to make sure. Otherwise the
1834 * inode might be lost for a long time or forever.
1835 */
1836 if (!XFS_FORCED_SHUTDOWN(mp)) {
1837 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1838 __func__, error);
1839 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1840 }
1841 xfs_trans_cancel(tp);
1842 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1843 return error;
1844 }
1845
1846 /*
1847 * Credit the quota account(s). The inode is gone.
1848 */
1849 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1850
1851 /*
1852 * Just ignore errors at this point. There is nothing we can do except
1853 * to try to keep going. Make sure it's not a silent error.
1854 */
1855 error = xfs_defer_finish(&tp, &dfops, NULL);
1856 if (error) {
1857 xfs_notice(mp, "%s: xfs_defer_finish returned error %d",
1858 __func__, error);
1859 xfs_defer_cancel(&dfops);
1860 }
1861 error = xfs_trans_commit(tp);
1862 if (error)
1863 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1864 __func__, error);
1865
1866 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1867 return 0;
1868}
1869
1870/*
1871 * xfs_inactive
1872 *
1873 * This is called when the vnode reference count for the vnode
1874 * goes to zero. If the file has been unlinked, then it must
1875 * now be truncated. Also, we clear all of the read-ahead state
1876 * kept for the inode here since the file is now closed.
1877 */
1878void
1879xfs_inactive(
1880 xfs_inode_t *ip)
1881{
1882 struct xfs_mount *mp;
1883 int error;
1884 int truncate = 0;
1885
1886 /*
1887 * If the inode is already free, then there can be nothing
1888 * to clean up here.
1889 */
1890 if (VFS_I(ip)->i_mode == 0) {
1891 ASSERT(ip->i_df.if_real_bytes == 0);
1892 ASSERT(ip->i_df.if_broot_bytes == 0);
1893 return;
1894 }
1895
1896 mp = ip->i_mount;
1897 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1898
1899 /* If this is a read-only mount, don't do this (would generate I/O) */
1900 if (mp->m_flags & XFS_MOUNT_RDONLY)
1901 return;
1902
1903 if (VFS_I(ip)->i_nlink != 0) {
1904 /*
1905 * force is true because we are evicting an inode from the
1906 * cache. Post-eof blocks must be freed, lest we end up with
1907 * broken free space accounting.
1908 */
1909 if (xfs_can_free_eofblocks(ip, true)) {
1910 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1911 xfs_free_eofblocks(ip);
1912 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1913 }
1914
1915 return;
1916 }
1917
1918 if (S_ISREG(VFS_I(ip)->i_mode) &&
1919 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1920 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1921 truncate = 1;
1922
1923 error = xfs_qm_dqattach(ip, 0);
1924 if (error)
1925 return;
1926
1927 if (S_ISLNK(VFS_I(ip)->i_mode))
1928 error = xfs_inactive_symlink(ip);
1929 else if (truncate)
1930 error = xfs_inactive_truncate(ip);
1931 if (error)
1932 return;
1933
1934 /*
1935 * If there are attributes associated with the file then blow them away
1936 * now. The code calls a routine that recursively deconstructs the
1937 * attribute fork. If also blows away the in-core attribute fork.
1938 */
1939 if (XFS_IFORK_Q(ip)) {
1940 error = xfs_attr_inactive(ip);
1941 if (error)
1942 return;
1943 }
1944
1945 ASSERT(!ip->i_afp);
1946 ASSERT(ip->i_d.di_anextents == 0);
1947 ASSERT(ip->i_d.di_forkoff == 0);
1948
1949 /*
1950 * Free the inode.
1951 */
1952 error = xfs_inactive_ifree(ip);
1953 if (error)
1954 return;
1955
1956 /*
1957 * Release the dquots held by inode, if any.
1958 */
1959 xfs_qm_dqdetach(ip);
1960}
1961
1962/*
1963 * This is called when the inode's link count goes to 0 or we are creating a
1964 * tmpfile via O_TMPFILE. In the case of a tmpfile, @ignore_linkcount will be
1965 * set to true as the link count is dropped to zero by the VFS after we've
1966 * created the file successfully, so we have to add it to the unlinked list
1967 * while the link count is non-zero.
1968 *
1969 * We place the on-disk inode on a list in the AGI. It will be pulled from this
1970 * list when the inode is freed.
1971 */
1972STATIC int
1973xfs_iunlink(
1974 struct xfs_trans *tp,
1975 struct xfs_inode *ip)
1976{
1977 xfs_mount_t *mp = tp->t_mountp;
1978 xfs_agi_t *agi;
1979 xfs_dinode_t *dip;
1980 xfs_buf_t *agibp;
1981 xfs_buf_t *ibp;
1982 xfs_agino_t agino;
1983 short bucket_index;
1984 int offset;
1985 int error;
1986
1987 ASSERT(VFS_I(ip)->i_mode != 0);
1988
1989 /*
1990 * Get the agi buffer first. It ensures lock ordering
1991 * on the list.
1992 */
1993 error = xfs_read_agi(mp, tp, XFS_INO_TO_AGNO(mp, ip->i_ino), &agibp);
1994 if (error)
1995 return error;
1996 agi = XFS_BUF_TO_AGI(agibp);
1997
1998 /*
1999 * Get the index into the agi hash table for the
2000 * list this inode will go on.
2001 */
2002 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2003 ASSERT(agino != 0);
2004 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2005 ASSERT(agi->agi_unlinked[bucket_index]);
2006 ASSERT(be32_to_cpu(agi->agi_unlinked[bucket_index]) != agino);
2007
2008 if (agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO)) {
2009 /*
2010 * There is already another inode in the bucket we need
2011 * to add ourselves to. Add us at the front of the list.
2012 * Here we put the head pointer into our next pointer,
2013 * and then we fall through to point the head at us.
2014 */
2015 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2016 0, 0);
2017 if (error)
2018 return error;
2019
2020 ASSERT(dip->di_next_unlinked == cpu_to_be32(NULLAGINO));
2021 dip->di_next_unlinked = agi->agi_unlinked[bucket_index];
2022 offset = ip->i_imap.im_boffset +
2023 offsetof(xfs_dinode_t, di_next_unlinked);
2024
2025 /* need to recalc the inode CRC if appropriate */
2026 xfs_dinode_calc_crc(mp, dip);
2027
2028 xfs_trans_inode_buf(tp, ibp);
2029 xfs_trans_log_buf(tp, ibp, offset,
2030 (offset + sizeof(xfs_agino_t) - 1));
2031 xfs_inobp_check(mp, ibp);
2032 }
2033
2034 /*
2035 * Point the bucket head pointer at the inode being inserted.
2036 */
2037 ASSERT(agino != 0);
2038 agi->agi_unlinked[bucket_index] = cpu_to_be32(agino);
2039 offset = offsetof(xfs_agi_t, agi_unlinked) +
2040 (sizeof(xfs_agino_t) * bucket_index);
2041 xfs_trans_log_buf(tp, agibp, offset,
2042 (offset + sizeof(xfs_agino_t) - 1));
2043 return 0;
2044}
2045
2046/*
2047 * Pull the on-disk inode from the AGI unlinked list.
2048 */
2049STATIC int
2050xfs_iunlink_remove(
2051 xfs_trans_t *tp,
2052 xfs_inode_t *ip)
2053{
2054 xfs_ino_t next_ino;
2055 xfs_mount_t *mp;
2056 xfs_agi_t *agi;
2057 xfs_dinode_t *dip;
2058 xfs_buf_t *agibp;
2059 xfs_buf_t *ibp;
2060 xfs_agnumber_t agno;
2061 xfs_agino_t agino;
2062 xfs_agino_t next_agino;
2063 xfs_buf_t *last_ibp;
2064 xfs_dinode_t *last_dip = NULL;
2065 short bucket_index;
2066 int offset, last_offset = 0;
2067 int error;
2068
2069 mp = tp->t_mountp;
2070 agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2071
2072 /*
2073 * Get the agi buffer first. It ensures lock ordering
2074 * on the list.
2075 */
2076 error = xfs_read_agi(mp, tp, agno, &agibp);
2077 if (error)
2078 return error;
2079
2080 agi = XFS_BUF_TO_AGI(agibp);
2081
2082 /*
2083 * Get the index into the agi hash table for the
2084 * list this inode will go on.
2085 */
2086 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2087 ASSERT(agino != 0);
2088 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2089 ASSERT(agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO));
2090 ASSERT(agi->agi_unlinked[bucket_index]);
2091
2092 if (be32_to_cpu(agi->agi_unlinked[bucket_index]) == agino) {
2093 /*
2094 * We're at the head of the list. Get the inode's on-disk
2095 * buffer to see if there is anyone after us on the list.
2096 * Only modify our next pointer if it is not already NULLAGINO.
2097 * This saves us the overhead of dealing with the buffer when
2098 * there is no need to change it.
2099 */
2100 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2101 0, 0);
2102 if (error) {
2103 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2104 __func__, error);
2105 return error;
2106 }
2107 next_agino = be32_to_cpu(dip->di_next_unlinked);
2108 ASSERT(next_agino != 0);
2109 if (next_agino != NULLAGINO) {
2110 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2111 offset = ip->i_imap.im_boffset +
2112 offsetof(xfs_dinode_t, di_next_unlinked);
2113
2114 /* need to recalc the inode CRC if appropriate */
2115 xfs_dinode_calc_crc(mp, dip);
2116
2117 xfs_trans_inode_buf(tp, ibp);
2118 xfs_trans_log_buf(tp, ibp, offset,
2119 (offset + sizeof(xfs_agino_t) - 1));
2120 xfs_inobp_check(mp, ibp);
2121 } else {
2122 xfs_trans_brelse(tp, ibp);
2123 }
2124 /*
2125 * Point the bucket head pointer at the next inode.
2126 */
2127 ASSERT(next_agino != 0);
2128 ASSERT(next_agino != agino);
2129 agi->agi_unlinked[bucket_index] = cpu_to_be32(next_agino);
2130 offset = offsetof(xfs_agi_t, agi_unlinked) +
2131 (sizeof(xfs_agino_t) * bucket_index);
2132 xfs_trans_log_buf(tp, agibp, offset,
2133 (offset + sizeof(xfs_agino_t) - 1));
2134 } else {
2135 /*
2136 * We need to search the list for the inode being freed.
2137 */
2138 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2139 last_ibp = NULL;
2140 while (next_agino != agino) {
2141 struct xfs_imap imap;
2142
2143 if (last_ibp)
2144 xfs_trans_brelse(tp, last_ibp);
2145
2146 imap.im_blkno = 0;
2147 next_ino = XFS_AGINO_TO_INO(mp, agno, next_agino);
2148
2149 error = xfs_imap(mp, tp, next_ino, &imap, 0);
2150 if (error) {
2151 xfs_warn(mp,
2152 "%s: xfs_imap returned error %d.",
2153 __func__, error);
2154 return error;
2155 }
2156
2157 error = xfs_imap_to_bp(mp, tp, &imap, &last_dip,
2158 &last_ibp, 0, 0);
2159 if (error) {
2160 xfs_warn(mp,
2161 "%s: xfs_imap_to_bp returned error %d.",
2162 __func__, error);
2163 return error;
2164 }
2165
2166 last_offset = imap.im_boffset;
2167 next_agino = be32_to_cpu(last_dip->di_next_unlinked);
2168 ASSERT(next_agino != NULLAGINO);
2169 ASSERT(next_agino != 0);
2170 }
2171
2172 /*
2173 * Now last_ibp points to the buffer previous to us on the
2174 * unlinked list. Pull us from the list.
2175 */
2176 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2177 0, 0);
2178 if (error) {
2179 xfs_warn(mp, "%s: xfs_imap_to_bp(2) returned error %d.",
2180 __func__, error);
2181 return error;
2182 }
2183 next_agino = be32_to_cpu(dip->di_next_unlinked);
2184 ASSERT(next_agino != 0);
2185 ASSERT(next_agino != agino);
2186 if (next_agino != NULLAGINO) {
2187 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2188 offset = ip->i_imap.im_boffset +
2189 offsetof(xfs_dinode_t, di_next_unlinked);
2190
2191 /* need to recalc the inode CRC if appropriate */
2192 xfs_dinode_calc_crc(mp, dip);
2193
2194 xfs_trans_inode_buf(tp, ibp);
2195 xfs_trans_log_buf(tp, ibp, offset,
2196 (offset + sizeof(xfs_agino_t) - 1));
2197 xfs_inobp_check(mp, ibp);
2198 } else {
2199 xfs_trans_brelse(tp, ibp);
2200 }
2201 /*
2202 * Point the previous inode on the list to the next inode.
2203 */
2204 last_dip->di_next_unlinked = cpu_to_be32(next_agino);
2205 ASSERT(next_agino != 0);
2206 offset = last_offset + offsetof(xfs_dinode_t, di_next_unlinked);
2207
2208 /* need to recalc the inode CRC if appropriate */
2209 xfs_dinode_calc_crc(mp, last_dip);
2210
2211 xfs_trans_inode_buf(tp, last_ibp);
2212 xfs_trans_log_buf(tp, last_ibp, offset,
2213 (offset + sizeof(xfs_agino_t) - 1));
2214 xfs_inobp_check(mp, last_ibp);
2215 }
2216 return 0;
2217}
2218
2219/*
2220 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2221 * inodes that are in memory - they all must be marked stale and attached to
2222 * the cluster buffer.
2223 */
2224STATIC int
2225xfs_ifree_cluster(
2226 xfs_inode_t *free_ip,
2227 xfs_trans_t *tp,
2228 struct xfs_icluster *xic)
2229{
2230 xfs_mount_t *mp = free_ip->i_mount;
2231 int blks_per_cluster;
2232 int inodes_per_cluster;
2233 int nbufs;
2234 int i, j;
2235 int ioffset;
2236 xfs_daddr_t blkno;
2237 xfs_buf_t *bp;
2238 xfs_inode_t *ip;
2239 xfs_inode_log_item_t *iip;
2240 xfs_log_item_t *lip;
2241 struct xfs_perag *pag;
2242 xfs_ino_t inum;
2243
2244 inum = xic->first_ino;
2245 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2246 blks_per_cluster = xfs_icluster_size_fsb(mp);
2247 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
2248 nbufs = mp->m_ialloc_blks / blks_per_cluster;
2249
2250 for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
2251 /*
2252 * The allocation bitmap tells us which inodes of the chunk were
2253 * physically allocated. Skip the cluster if an inode falls into
2254 * a sparse region.
2255 */
2256 ioffset = inum - xic->first_ino;
2257 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2258 ASSERT(do_mod(ioffset, inodes_per_cluster) == 0);
2259 continue;
2260 }
2261
2262 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2263 XFS_INO_TO_AGBNO(mp, inum));
2264
2265 /*
2266 * We obtain and lock the backing buffer first in the process
2267 * here, as we have to ensure that any dirty inode that we
2268 * can't get the flush lock on is attached to the buffer.
2269 * If we scan the in-memory inodes first, then buffer IO can
2270 * complete before we get a lock on it, and hence we may fail
2271 * to mark all the active inodes on the buffer stale.
2272 */
2273 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2274 mp->m_bsize * blks_per_cluster,
2275 XBF_UNMAPPED);
2276
2277 if (!bp)
2278 return -ENOMEM;
2279
2280 /*
2281 * This buffer may not have been correctly initialised as we
2282 * didn't read it from disk. That's not important because we are
2283 * only using to mark the buffer as stale in the log, and to
2284 * attach stale cached inodes on it. That means it will never be
2285 * dispatched for IO. If it is, we want to know about it, and we
2286 * want it to fail. We can acheive this by adding a write
2287 * verifier to the buffer.
2288 */
2289 bp->b_ops = &xfs_inode_buf_ops;
2290
2291 /*
2292 * Walk the inodes already attached to the buffer and mark them
2293 * stale. These will all have the flush locks held, so an
2294 * in-memory inode walk can't lock them. By marking them all
2295 * stale first, we will not attempt to lock them in the loop
2296 * below as the XFS_ISTALE flag will be set.
2297 */
2298 lip = bp->b_fspriv;
2299 while (lip) {
2300 if (lip->li_type == XFS_LI_INODE) {
2301 iip = (xfs_inode_log_item_t *)lip;
2302 ASSERT(iip->ili_logged == 1);
2303 lip->li_cb = xfs_istale_done;
2304 xfs_trans_ail_copy_lsn(mp->m_ail,
2305 &iip->ili_flush_lsn,
2306 &iip->ili_item.li_lsn);
2307 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2308 }
2309 lip = lip->li_bio_list;
2310 }
2311
2312
2313 /*
2314 * For each inode in memory attempt to add it to the inode
2315 * buffer and set it up for being staled on buffer IO
2316 * completion. This is safe as we've locked out tail pushing
2317 * and flushing by locking the buffer.
2318 *
2319 * We have already marked every inode that was part of a
2320 * transaction stale above, which means there is no point in
2321 * even trying to lock them.
2322 */
2323 for (i = 0; i < inodes_per_cluster; i++) {
2324retry:
2325 rcu_read_lock();
2326 ip = radix_tree_lookup(&pag->pag_ici_root,
2327 XFS_INO_TO_AGINO(mp, (inum + i)));
2328
2329 /* Inode not in memory, nothing to do */
2330 if (!ip) {
2331 rcu_read_unlock();
2332 continue;
2333 }
2334
2335 /*
2336 * because this is an RCU protected lookup, we could
2337 * find a recently freed or even reallocated inode
2338 * during the lookup. We need to check under the
2339 * i_flags_lock for a valid inode here. Skip it if it
2340 * is not valid, the wrong inode or stale.
2341 */
2342 spin_lock(&ip->i_flags_lock);
2343 if (ip->i_ino != inum + i ||
2344 __xfs_iflags_test(ip, XFS_ISTALE)) {
2345 spin_unlock(&ip->i_flags_lock);
2346 rcu_read_unlock();
2347 continue;
2348 }
2349 spin_unlock(&ip->i_flags_lock);
2350
2351 /*
2352 * Don't try to lock/unlock the current inode, but we
2353 * _cannot_ skip the other inodes that we did not find
2354 * in the list attached to the buffer and are not
2355 * already marked stale. If we can't lock it, back off
2356 * and retry.
2357 */
2358 if (ip != free_ip &&
2359 !xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2360 rcu_read_unlock();
2361 delay(1);
2362 goto retry;
2363 }
2364 rcu_read_unlock();
2365
2366 xfs_iflock(ip);
2367 xfs_iflags_set(ip, XFS_ISTALE);
2368
2369 /*
2370 * we don't need to attach clean inodes or those only
2371 * with unlogged changes (which we throw away, anyway).
2372 */
2373 iip = ip->i_itemp;
2374 if (!iip || xfs_inode_clean(ip)) {
2375 ASSERT(ip != free_ip);
2376 xfs_ifunlock(ip);
2377 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2378 continue;
2379 }
2380
2381 iip->ili_last_fields = iip->ili_fields;
2382 iip->ili_fields = 0;
2383 iip->ili_fsync_fields = 0;
2384 iip->ili_logged = 1;
2385 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2386 &iip->ili_item.li_lsn);
2387
2388 xfs_buf_attach_iodone(bp, xfs_istale_done,
2389 &iip->ili_item);
2390
2391 if (ip != free_ip)
2392 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2393 }
2394
2395 xfs_trans_stale_inode_buf(tp, bp);
2396 xfs_trans_binval(tp, bp);
2397 }
2398
2399 xfs_perag_put(pag);
2400 return 0;
2401}
2402
2403/*
2404 * This is called to return an inode to the inode free list.
2405 * The inode should already be truncated to 0 length and have
2406 * no pages associated with it. This routine also assumes that
2407 * the inode is already a part of the transaction.
2408 *
2409 * The on-disk copy of the inode will have been added to the list
2410 * of unlinked inodes in the AGI. We need to remove the inode from
2411 * that list atomically with respect to freeing it here.
2412 */
2413int
2414xfs_ifree(
2415 xfs_trans_t *tp,
2416 xfs_inode_t *ip,
2417 struct xfs_defer_ops *dfops)
2418{
2419 int error;
2420 struct xfs_icluster xic = { 0 };
2421
2422 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2423 ASSERT(VFS_I(ip)->i_nlink == 0);
2424 ASSERT(ip->i_d.di_nextents == 0);
2425 ASSERT(ip->i_d.di_anextents == 0);
2426 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2427 ASSERT(ip->i_d.di_nblocks == 0);
2428
2429 /*
2430 * Pull the on-disk inode from the AGI unlinked list.
2431 */
2432 error = xfs_iunlink_remove(tp, ip);
2433 if (error)
2434 return error;
2435
2436 error = xfs_difree(tp, ip->i_ino, dfops, &xic);
2437 if (error)
2438 return error;
2439
2440 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2441 ip->i_d.di_flags = 0;
2442 ip->i_d.di_dmevmask = 0;
2443 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2444 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2445 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2446 /*
2447 * Bump the generation count so no one will be confused
2448 * by reincarnations of this inode.
2449 */
2450 VFS_I(ip)->i_generation++;
2451 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2452
2453 if (xic.deleted)
2454 error = xfs_ifree_cluster(ip, tp, &xic);
2455
2456 return error;
2457}
2458
2459/*
2460 * This is called to unpin an inode. The caller must have the inode locked
2461 * in at least shared mode so that the buffer cannot be subsequently pinned
2462 * once someone is waiting for it to be unpinned.
2463 */
2464static void
2465xfs_iunpin(
2466 struct xfs_inode *ip)
2467{
2468 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2469
2470 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2471
2472 /* Give the log a push to start the unpinning I/O */
2473 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0);
2474
2475}
2476
2477static void
2478__xfs_iunpin_wait(
2479 struct xfs_inode *ip)
2480{
2481 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2482 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2483
2484 xfs_iunpin(ip);
2485
2486 do {
2487 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2488 if (xfs_ipincount(ip))
2489 io_schedule();
2490 } while (xfs_ipincount(ip));
2491 finish_wait(wq, &wait.wait);
2492}
2493
2494void
2495xfs_iunpin_wait(
2496 struct xfs_inode *ip)
2497{
2498 if (xfs_ipincount(ip))
2499 __xfs_iunpin_wait(ip);
2500}
2501
2502/*
2503 * Removing an inode from the namespace involves removing the directory entry
2504 * and dropping the link count on the inode. Removing the directory entry can
2505 * result in locking an AGF (directory blocks were freed) and removing a link
2506 * count can result in placing the inode on an unlinked list which results in
2507 * locking an AGI.
2508 *
2509 * The big problem here is that we have an ordering constraint on AGF and AGI
2510 * locking - inode allocation locks the AGI, then can allocate a new extent for
2511 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2512 * removes the inode from the unlinked list, requiring that we lock the AGI
2513 * first, and then freeing the inode can result in an inode chunk being freed
2514 * and hence freeing disk space requiring that we lock an AGF.
2515 *
2516 * Hence the ordering that is imposed by other parts of the code is AGI before
2517 * AGF. This means we cannot remove the directory entry before we drop the inode
2518 * reference count and put it on the unlinked list as this results in a lock
2519 * order of AGF then AGI, and this can deadlock against inode allocation and
2520 * freeing. Therefore we must drop the link counts before we remove the
2521 * directory entry.
2522 *
2523 * This is still safe from a transactional point of view - it is not until we
2524 * get to xfs_defer_finish() that we have the possibility of multiple
2525 * transactions in this operation. Hence as long as we remove the directory
2526 * entry and drop the link count in the first transaction of the remove
2527 * operation, there are no transactional constraints on the ordering here.
2528 */
2529int
2530xfs_remove(
2531 xfs_inode_t *dp,
2532 struct xfs_name *name,
2533 xfs_inode_t *ip)
2534{
2535 xfs_mount_t *mp = dp->i_mount;
2536 xfs_trans_t *tp = NULL;
2537 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2538 int error = 0;
2539 struct xfs_defer_ops dfops;
2540 xfs_fsblock_t first_block;
2541 uint resblks;
2542
2543 trace_xfs_remove(dp, name);
2544
2545 if (XFS_FORCED_SHUTDOWN(mp))
2546 return -EIO;
2547
2548 error = xfs_qm_dqattach(dp, 0);
2549 if (error)
2550 goto std_return;
2551
2552 error = xfs_qm_dqattach(ip, 0);
2553 if (error)
2554 goto std_return;
2555
2556 /*
2557 * We try to get the real space reservation first,
2558 * allowing for directory btree deletion(s) implying
2559 * possible bmap insert(s). If we can't get the space
2560 * reservation then we use 0 instead, and avoid the bmap
2561 * btree insert(s) in the directory code by, if the bmap
2562 * insert tries to happen, instead trimming the LAST
2563 * block from the directory.
2564 */
2565 resblks = XFS_REMOVE_SPACE_RES(mp);
2566 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2567 if (error == -ENOSPC) {
2568 resblks = 0;
2569 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2570 &tp);
2571 }
2572 if (error) {
2573 ASSERT(error != -ENOSPC);
2574 goto std_return;
2575 }
2576
2577 xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
2578
2579 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2580 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2581
2582 /*
2583 * If we're removing a directory perform some additional validation.
2584 */
2585 if (is_dir) {
2586 ASSERT(VFS_I(ip)->i_nlink >= 2);
2587 if (VFS_I(ip)->i_nlink != 2) {
2588 error = -ENOTEMPTY;
2589 goto out_trans_cancel;
2590 }
2591 if (!xfs_dir_isempty(ip)) {
2592 error = -ENOTEMPTY;
2593 goto out_trans_cancel;
2594 }
2595
2596 /* Drop the link from ip's "..". */
2597 error = xfs_droplink(tp, dp);
2598 if (error)
2599 goto out_trans_cancel;
2600
2601 /* Drop the "." link from ip to self. */
2602 error = xfs_droplink(tp, ip);
2603 if (error)
2604 goto out_trans_cancel;
2605 } else {
2606 /*
2607 * When removing a non-directory we need to log the parent
2608 * inode here. For a directory this is done implicitly
2609 * by the xfs_droplink call for the ".." entry.
2610 */
2611 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2612 }
2613 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2614
2615 /* Drop the link from dp to ip. */
2616 error = xfs_droplink(tp, ip);
2617 if (error)
2618 goto out_trans_cancel;
2619
2620 xfs_defer_init(&dfops, &first_block);
2621 error = xfs_dir_removename(tp, dp, name, ip->i_ino,
2622 &first_block, &dfops, resblks);
2623 if (error) {
2624 ASSERT(error != -ENOENT);
2625 goto out_bmap_cancel;
2626 }
2627
2628 /*
2629 * If this is a synchronous mount, make sure that the
2630 * remove transaction goes to disk before returning to
2631 * the user.
2632 */
2633 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2634 xfs_trans_set_sync(tp);
2635
2636 error = xfs_defer_finish(&tp, &dfops, NULL);
2637 if (error)
2638 goto out_bmap_cancel;
2639
2640 error = xfs_trans_commit(tp);
2641 if (error)
2642 goto std_return;
2643
2644 if (is_dir && xfs_inode_is_filestream(ip))
2645 xfs_filestream_deassociate(ip);
2646
2647 return 0;
2648
2649 out_bmap_cancel:
2650 xfs_defer_cancel(&dfops);
2651 out_trans_cancel:
2652 xfs_trans_cancel(tp);
2653 std_return:
2654 return error;
2655}
2656
2657/*
2658 * Enter all inodes for a rename transaction into a sorted array.
2659 */
2660#define __XFS_SORT_INODES 5
2661STATIC void
2662xfs_sort_for_rename(
2663 struct xfs_inode *dp1, /* in: old (source) directory inode */
2664 struct xfs_inode *dp2, /* in: new (target) directory inode */
2665 struct xfs_inode *ip1, /* in: inode of old entry */
2666 struct xfs_inode *ip2, /* in: inode of new entry */
2667 struct xfs_inode *wip, /* in: whiteout inode */
2668 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2669 int *num_inodes) /* in/out: inodes in array */
2670{
2671 int i, j;
2672
2673 ASSERT(*num_inodes == __XFS_SORT_INODES);
2674 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2675
2676 /*
2677 * i_tab contains a list of pointers to inodes. We initialize
2678 * the table here & we'll sort it. We will then use it to
2679 * order the acquisition of the inode locks.
2680 *
2681 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2682 */
2683 i = 0;
2684 i_tab[i++] = dp1;
2685 i_tab[i++] = dp2;
2686 i_tab[i++] = ip1;
2687 if (ip2)
2688 i_tab[i++] = ip2;
2689 if (wip)
2690 i_tab[i++] = wip;
2691 *num_inodes = i;
2692
2693 /*
2694 * Sort the elements via bubble sort. (Remember, there are at
2695 * most 5 elements to sort, so this is adequate.)
2696 */
2697 for (i = 0; i < *num_inodes; i++) {
2698 for (j = 1; j < *num_inodes; j++) {
2699 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2700 struct xfs_inode *temp = i_tab[j];
2701 i_tab[j] = i_tab[j-1];
2702 i_tab[j-1] = temp;
2703 }
2704 }
2705 }
2706}
2707
2708static int
2709xfs_finish_rename(
2710 struct xfs_trans *tp,
2711 struct xfs_defer_ops *dfops)
2712{
2713 int error;
2714
2715 /*
2716 * If this is a synchronous mount, make sure that the rename transaction
2717 * goes to disk before returning to the user.
2718 */
2719 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2720 xfs_trans_set_sync(tp);
2721
2722 error = xfs_defer_finish(&tp, dfops, NULL);
2723 if (error) {
2724 xfs_defer_cancel(dfops);
2725 xfs_trans_cancel(tp);
2726 return error;
2727 }
2728
2729 return xfs_trans_commit(tp);
2730}
2731
2732/*
2733 * xfs_cross_rename()
2734 *
2735 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
2736 */
2737STATIC int
2738xfs_cross_rename(
2739 struct xfs_trans *tp,
2740 struct xfs_inode *dp1,
2741 struct xfs_name *name1,
2742 struct xfs_inode *ip1,
2743 struct xfs_inode *dp2,
2744 struct xfs_name *name2,
2745 struct xfs_inode *ip2,
2746 struct xfs_defer_ops *dfops,
2747 xfs_fsblock_t *first_block,
2748 int spaceres)
2749{
2750 int error = 0;
2751 int ip1_flags = 0;
2752 int ip2_flags = 0;
2753 int dp2_flags = 0;
2754
2755 /* Swap inode number for dirent in first parent */
2756 error = xfs_dir_replace(tp, dp1, name1,
2757 ip2->i_ino,
2758 first_block, dfops, spaceres);
2759 if (error)
2760 goto out_trans_abort;
2761
2762 /* Swap inode number for dirent in second parent */
2763 error = xfs_dir_replace(tp, dp2, name2,
2764 ip1->i_ino,
2765 first_block, dfops, spaceres);
2766 if (error)
2767 goto out_trans_abort;
2768
2769 /*
2770 * If we're renaming one or more directories across different parents,
2771 * update the respective ".." entries (and link counts) to match the new
2772 * parents.
2773 */
2774 if (dp1 != dp2) {
2775 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2776
2777 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2778 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2779 dp1->i_ino, first_block,
2780 dfops, spaceres);
2781 if (error)
2782 goto out_trans_abort;
2783
2784 /* transfer ip2 ".." reference to dp1 */
2785 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2786 error = xfs_droplink(tp, dp2);
2787 if (error)
2788 goto out_trans_abort;
2789 error = xfs_bumplink(tp, dp1);
2790 if (error)
2791 goto out_trans_abort;
2792 }
2793
2794 /*
2795 * Although ip1 isn't changed here, userspace needs
2796 * to be warned about the change, so that applications
2797 * relying on it (like backup ones), will properly
2798 * notify the change
2799 */
2800 ip1_flags |= XFS_ICHGTIME_CHG;
2801 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2802 }
2803
2804 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2805 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2806 dp2->i_ino, first_block,
2807 dfops, spaceres);
2808 if (error)
2809 goto out_trans_abort;
2810
2811 /* transfer ip1 ".." reference to dp2 */
2812 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2813 error = xfs_droplink(tp, dp1);
2814 if (error)
2815 goto out_trans_abort;
2816 error = xfs_bumplink(tp, dp2);
2817 if (error)
2818 goto out_trans_abort;
2819 }
2820
2821 /*
2822 * Although ip2 isn't changed here, userspace needs
2823 * to be warned about the change, so that applications
2824 * relying on it (like backup ones), will properly
2825 * notify the change
2826 */
2827 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2828 ip2_flags |= XFS_ICHGTIME_CHG;
2829 }
2830 }
2831
2832 if (ip1_flags) {
2833 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2834 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2835 }
2836 if (ip2_flags) {
2837 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2838 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2839 }
2840 if (dp2_flags) {
2841 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2842 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2843 }
2844 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2845 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2846 return xfs_finish_rename(tp, dfops);
2847
2848out_trans_abort:
2849 xfs_defer_cancel(dfops);
2850 xfs_trans_cancel(tp);
2851 return error;
2852}
2853
2854/*
2855 * xfs_rename_alloc_whiteout()
2856 *
2857 * Return a referenced, unlinked, unlocked inode that that can be used as a
2858 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2859 * crash between allocating the inode and linking it into the rename transaction
2860 * recovery will free the inode and we won't leak it.
2861 */
2862static int
2863xfs_rename_alloc_whiteout(
2864 struct xfs_inode *dp,
2865 struct xfs_inode **wip)
2866{
2867 struct xfs_inode *tmpfile;
2868 int error;
2869
2870 error = xfs_create_tmpfile(dp, NULL, S_IFCHR | WHITEOUT_MODE, &tmpfile);
2871 if (error)
2872 return error;
2873
2874 /*
2875 * Prepare the tmpfile inode as if it were created through the VFS.
2876 * Otherwise, the link increment paths will complain about nlink 0->1.
2877 * Drop the link count as done by d_tmpfile(), complete the inode setup
2878 * and flag it as linkable.
2879 */
2880 drop_nlink(VFS_I(tmpfile));
2881 xfs_setup_iops(tmpfile);
2882 xfs_finish_inode_setup(tmpfile);
2883 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2884
2885 *wip = tmpfile;
2886 return 0;
2887}
2888
2889/*
2890 * xfs_rename
2891 */
2892int
2893xfs_rename(
2894 struct xfs_inode *src_dp,
2895 struct xfs_name *src_name,
2896 struct xfs_inode *src_ip,
2897 struct xfs_inode *target_dp,
2898 struct xfs_name *target_name,
2899 struct xfs_inode *target_ip,
2900 unsigned int flags)
2901{
2902 struct xfs_mount *mp = src_dp->i_mount;
2903 struct xfs_trans *tp;
2904 struct xfs_defer_ops dfops;
2905 xfs_fsblock_t first_block;
2906 struct xfs_inode *wip = NULL; /* whiteout inode */
2907 struct xfs_inode *inodes[__XFS_SORT_INODES];
2908 int num_inodes = __XFS_SORT_INODES;
2909 bool new_parent = (src_dp != target_dp);
2910 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2911 int spaceres;
2912 int error;
2913
2914 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2915
2916 if ((flags & RENAME_EXCHANGE) && !target_ip)
2917 return -EINVAL;
2918
2919 /*
2920 * If we are doing a whiteout operation, allocate the whiteout inode
2921 * we will be placing at the target and ensure the type is set
2922 * appropriately.
2923 */
2924 if (flags & RENAME_WHITEOUT) {
2925 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
2926 error = xfs_rename_alloc_whiteout(target_dp, &wip);
2927 if (error)
2928 return error;
2929
2930 /* setup target dirent info as whiteout */
2931 src_name->type = XFS_DIR3_FT_CHRDEV;
2932 }
2933
2934 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2935 inodes, &num_inodes);
2936
2937 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2938 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2939 if (error == -ENOSPC) {
2940 spaceres = 0;
2941 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2942 &tp);
2943 }
2944 if (error)
2945 goto out_release_wip;
2946
2947 /*
2948 * Attach the dquots to the inodes
2949 */
2950 error = xfs_qm_vop_rename_dqattach(inodes);
2951 if (error)
2952 goto out_trans_cancel;
2953
2954 /*
2955 * Lock all the participating inodes. Depending upon whether
2956 * the target_name exists in the target directory, and
2957 * whether the target directory is the same as the source
2958 * directory, we can lock from 2 to 4 inodes.
2959 */
2960 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2961
2962 /*
2963 * Join all the inodes to the transaction. From this point on,
2964 * we can rely on either trans_commit or trans_cancel to unlock
2965 * them.
2966 */
2967 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2968 if (new_parent)
2969 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2970 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2971 if (target_ip)
2972 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2973 if (wip)
2974 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
2975
2976 /*
2977 * If we are using project inheritance, we only allow renames
2978 * into our tree when the project IDs are the same; else the
2979 * tree quota mechanism would be circumvented.
2980 */
2981 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2982 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
2983 error = -EXDEV;
2984 goto out_trans_cancel;
2985 }
2986
2987 xfs_defer_init(&dfops, &first_block);
2988
2989 /* RENAME_EXCHANGE is unique from here on. */
2990 if (flags & RENAME_EXCHANGE)
2991 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
2992 target_dp, target_name, target_ip,
2993 &dfops, &first_block, spaceres);
2994
2995 /*
2996 * Set up the target.
2997 */
2998 if (target_ip == NULL) {
2999 /*
3000 * If there's no space reservation, check the entry will
3001 * fit before actually inserting it.
3002 */
3003 if (!spaceres) {
3004 error = xfs_dir_canenter(tp, target_dp, target_name);
3005 if (error)
3006 goto out_trans_cancel;
3007 }
3008 /*
3009 * If target does not exist and the rename crosses
3010 * directories, adjust the target directory link count
3011 * to account for the ".." reference from the new entry.
3012 */
3013 error = xfs_dir_createname(tp, target_dp, target_name,
3014 src_ip->i_ino, &first_block,
3015 &dfops, spaceres);
3016 if (error)
3017 goto out_bmap_cancel;
3018
3019 xfs_trans_ichgtime(tp, target_dp,
3020 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3021
3022 if (new_parent && src_is_directory) {
3023 error = xfs_bumplink(tp, target_dp);
3024 if (error)
3025 goto out_bmap_cancel;
3026 }
3027 } else { /* target_ip != NULL */
3028 /*
3029 * If target exists and it's a directory, check that both
3030 * target and source are directories and that target can be
3031 * destroyed, or that neither is a directory.
3032 */
3033 if (S_ISDIR(VFS_I(target_ip)->i_mode)) {
3034 /*
3035 * Make sure target dir is empty.
3036 */
3037 if (!(xfs_dir_isempty(target_ip)) ||
3038 (VFS_I(target_ip)->i_nlink > 2)) {
3039 error = -EEXIST;
3040 goto out_trans_cancel;
3041 }
3042 }
3043
3044 /*
3045 * Link the source inode under the target name.
3046 * If the source inode is a directory and we are moving
3047 * it across directories, its ".." entry will be
3048 * inconsistent until we replace that down below.
3049 *
3050 * In case there is already an entry with the same
3051 * name at the destination directory, remove it first.
3052 */
3053 error = xfs_dir_replace(tp, target_dp, target_name,
3054 src_ip->i_ino,
3055 &first_block, &dfops, spaceres);
3056 if (error)
3057 goto out_bmap_cancel;
3058
3059 xfs_trans_ichgtime(tp, target_dp,
3060 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3061
3062 /*
3063 * Decrement the link count on the target since the target
3064 * dir no longer points to it.
3065 */
3066 error = xfs_droplink(tp, target_ip);
3067 if (error)
3068 goto out_bmap_cancel;
3069
3070 if (src_is_directory) {
3071 /*
3072 * Drop the link from the old "." entry.
3073 */
3074 error = xfs_droplink(tp, target_ip);
3075 if (error)
3076 goto out_bmap_cancel;
3077 }
3078 } /* target_ip != NULL */
3079
3080 /*
3081 * Remove the source.
3082 */
3083 if (new_parent && src_is_directory) {
3084 /*
3085 * Rewrite the ".." entry to point to the new
3086 * directory.
3087 */
3088 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3089 target_dp->i_ino,
3090 &first_block, &dfops, spaceres);
3091 ASSERT(error != -EEXIST);
3092 if (error)
3093 goto out_bmap_cancel;
3094 }
3095
3096 /*
3097 * We always want to hit the ctime on the source inode.
3098 *
3099 * This isn't strictly required by the standards since the source
3100 * inode isn't really being changed, but old unix file systems did
3101 * it and some incremental backup programs won't work without it.
3102 */
3103 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3104 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3105
3106 /*
3107 * Adjust the link count on src_dp. This is necessary when
3108 * renaming a directory, either within one parent when
3109 * the target existed, or across two parent directories.
3110 */
3111 if (src_is_directory && (new_parent || target_ip != NULL)) {
3112
3113 /*
3114 * Decrement link count on src_directory since the
3115 * entry that's moved no longer points to it.
3116 */
3117 error = xfs_droplink(tp, src_dp);
3118 if (error)
3119 goto out_bmap_cancel;
3120 }
3121
3122 /*
3123 * For whiteouts, we only need to update the source dirent with the
3124 * inode number of the whiteout inode rather than removing it
3125 * altogether.
3126 */
3127 if (wip) {
3128 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3129 &first_block, &dfops, spaceres);
3130 } else
3131 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3132 &first_block, &dfops, spaceres);
3133 if (error)
3134 goto out_bmap_cancel;
3135
3136 /*
3137 * For whiteouts, we need to bump the link count on the whiteout inode.
3138 * This means that failures all the way up to this point leave the inode
3139 * on the unlinked list and so cleanup is a simple matter of dropping
3140 * the remaining reference to it. If we fail here after bumping the link
3141 * count, we're shutting down the filesystem so we'll never see the
3142 * intermediate state on disk.
3143 */
3144 if (wip) {
3145 ASSERT(VFS_I(wip)->i_nlink == 0);
3146 error = xfs_bumplink(tp, wip);
3147 if (error)
3148 goto out_bmap_cancel;
3149 error = xfs_iunlink_remove(tp, wip);
3150 if (error)
3151 goto out_bmap_cancel;
3152 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3153
3154 /*
3155 * Now we have a real link, clear the "I'm a tmpfile" state
3156 * flag from the inode so it doesn't accidentally get misused in
3157 * future.
3158 */
3159 VFS_I(wip)->i_state &= ~I_LINKABLE;
3160 }
3161
3162 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3163 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3164 if (new_parent)
3165 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3166
3167 error = xfs_finish_rename(tp, &dfops);
3168 if (wip)
3169 IRELE(wip);
3170 return error;
3171
3172out_bmap_cancel:
3173 xfs_defer_cancel(&dfops);
3174out_trans_cancel:
3175 xfs_trans_cancel(tp);
3176out_release_wip:
3177 if (wip)
3178 IRELE(wip);
3179 return error;
3180}
3181
3182STATIC int
3183xfs_iflush_cluster(
3184 struct xfs_inode *ip,
3185 struct xfs_buf *bp)
3186{
3187 struct xfs_mount *mp = ip->i_mount;
3188 struct xfs_perag *pag;
3189 unsigned long first_index, mask;
3190 unsigned long inodes_per_cluster;
3191 int cilist_size;
3192 struct xfs_inode **cilist;
3193 struct xfs_inode *cip;
3194 int nr_found;
3195 int clcount = 0;
3196 int bufwasdelwri;
3197 int i;
3198
3199 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3200
3201 inodes_per_cluster = mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog;
3202 cilist_size = inodes_per_cluster * sizeof(xfs_inode_t *);
3203 cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3204 if (!cilist)
3205 goto out_put;
3206
3207 mask = ~(((mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog)) - 1);
3208 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3209 rcu_read_lock();
3210 /* really need a gang lookup range call here */
3211 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3212 first_index, inodes_per_cluster);
3213 if (nr_found == 0)
3214 goto out_free;
3215
3216 for (i = 0; i < nr_found; i++) {
3217 cip = cilist[i];
3218 if (cip == ip)
3219 continue;
3220
3221 /*
3222 * because this is an RCU protected lookup, we could find a
3223 * recently freed or even reallocated inode during the lookup.
3224 * We need to check under the i_flags_lock for a valid inode
3225 * here. Skip it if it is not valid or the wrong inode.
3226 */
3227 spin_lock(&cip->i_flags_lock);
3228 if (!cip->i_ino ||
3229 __xfs_iflags_test(cip, XFS_ISTALE)) {
3230 spin_unlock(&cip->i_flags_lock);
3231 continue;
3232 }
3233
3234 /*
3235 * Once we fall off the end of the cluster, no point checking
3236 * any more inodes in the list because they will also all be
3237 * outside the cluster.
3238 */
3239 if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3240 spin_unlock(&cip->i_flags_lock);
3241 break;
3242 }
3243 spin_unlock(&cip->i_flags_lock);
3244
3245 /*
3246 * Do an un-protected check to see if the inode is dirty and
3247 * is a candidate for flushing. These checks will be repeated
3248 * later after the appropriate locks are acquired.
3249 */
3250 if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3251 continue;
3252
3253 /*
3254 * Try to get locks. If any are unavailable or it is pinned,
3255 * then this inode cannot be flushed and is skipped.
3256 */
3257
3258 if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3259 continue;
3260 if (!xfs_iflock_nowait(cip)) {
3261 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3262 continue;
3263 }
3264 if (xfs_ipincount(cip)) {
3265 xfs_ifunlock(cip);
3266 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3267 continue;
3268 }
3269
3270
3271 /*
3272 * Check the inode number again, just to be certain we are not
3273 * racing with freeing in xfs_reclaim_inode(). See the comments
3274 * in that function for more information as to why the initial
3275 * check is not sufficient.
3276 */
3277 if (!cip->i_ino) {
3278 xfs_ifunlock(cip);
3279 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3280 continue;
3281 }
3282
3283 /*
3284 * arriving here means that this inode can be flushed. First
3285 * re-check that it's dirty before flushing.
3286 */
3287 if (!xfs_inode_clean(cip)) {
3288 int error;
3289 error = xfs_iflush_int(cip, bp);
3290 if (error) {
3291 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3292 goto cluster_corrupt_out;
3293 }
3294 clcount++;
3295 } else {
3296 xfs_ifunlock(cip);
3297 }
3298 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3299 }
3300
3301 if (clcount) {
3302 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3303 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3304 }
3305
3306out_free:
3307 rcu_read_unlock();
3308 kmem_free(cilist);
3309out_put:
3310 xfs_perag_put(pag);
3311 return 0;
3312
3313
3314cluster_corrupt_out:
3315 /*
3316 * Corruption detected in the clustering loop. Invalidate the
3317 * inode buffer and shut down the filesystem.
3318 */
3319 rcu_read_unlock();
3320 /*
3321 * Clean up the buffer. If it was delwri, just release it --
3322 * brelse can handle it with no problems. If not, shut down the
3323 * filesystem before releasing the buffer.
3324 */
3325 bufwasdelwri = (bp->b_flags & _XBF_DELWRI_Q);
3326 if (bufwasdelwri)
3327 xfs_buf_relse(bp);
3328
3329 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3330
3331 if (!bufwasdelwri) {
3332 /*
3333 * Just like incore_relse: if we have b_iodone functions,
3334 * mark the buffer as an error and call them. Otherwise
3335 * mark it as stale and brelse.
3336 */
3337 if (bp->b_iodone) {
3338 bp->b_flags &= ~XBF_DONE;
3339 xfs_buf_stale(bp);
3340 xfs_buf_ioerror(bp, -EIO);
3341 xfs_buf_ioend(bp);
3342 } else {
3343 xfs_buf_stale(bp);
3344 xfs_buf_relse(bp);
3345 }
3346 }
3347
3348 /*
3349 * Unlocks the flush lock
3350 */
3351 xfs_iflush_abort(cip, false);
3352 kmem_free(cilist);
3353 xfs_perag_put(pag);
3354 return -EFSCORRUPTED;
3355}
3356
3357/*
3358 * Flush dirty inode metadata into the backing buffer.
3359 *
3360 * The caller must have the inode lock and the inode flush lock held. The
3361 * inode lock will still be held upon return to the caller, and the inode
3362 * flush lock will be released after the inode has reached the disk.
3363 *
3364 * The caller must write out the buffer returned in *bpp and release it.
3365 */
3366int
3367xfs_iflush(
3368 struct xfs_inode *ip,
3369 struct xfs_buf **bpp)
3370{
3371 struct xfs_mount *mp = ip->i_mount;
3372 struct xfs_buf *bp = NULL;
3373 struct xfs_dinode *dip;
3374 int error;
3375
3376 XFS_STATS_INC(mp, xs_iflush_count);
3377
3378 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3379 ASSERT(xfs_isiflocked(ip));
3380 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3381 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3382
3383 *bpp = NULL;
3384
3385 xfs_iunpin_wait(ip);
3386
3387 /*
3388 * For stale inodes we cannot rely on the backing buffer remaining
3389 * stale in cache for the remaining life of the stale inode and so
3390 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3391 * inodes below. We have to check this after ensuring the inode is
3392 * unpinned so that it is safe to reclaim the stale inode after the
3393 * flush call.
3394 */
3395 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3396 xfs_ifunlock(ip);
3397 return 0;
3398 }
3399
3400 /*
3401 * This may have been unpinned because the filesystem is shutting
3402 * down forcibly. If that's the case we must not write this inode
3403 * to disk, because the log record didn't make it to disk.
3404 *
3405 * We also have to remove the log item from the AIL in this case,
3406 * as we wait for an empty AIL as part of the unmount process.
3407 */
3408 if (XFS_FORCED_SHUTDOWN(mp)) {
3409 error = -EIO;
3410 goto abort_out;
3411 }
3412
3413 /*
3414 * Get the buffer containing the on-disk inode. We are doing a try-lock
3415 * operation here, so we may get an EAGAIN error. In that case, we
3416 * simply want to return with the inode still dirty.
3417 *
3418 * If we get any other error, we effectively have a corruption situation
3419 * and we cannot flush the inode, so we treat it the same as failing
3420 * xfs_iflush_int().
3421 */
3422 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3423 0);
3424 if (error == -EAGAIN) {
3425 xfs_ifunlock(ip);
3426 return error;
3427 }
3428 if (error)
3429 goto corrupt_out;
3430
3431 /*
3432 * First flush out the inode that xfs_iflush was called with.
3433 */
3434 error = xfs_iflush_int(ip, bp);
3435 if (error)
3436 goto corrupt_out;
3437
3438 /*
3439 * If the buffer is pinned then push on the log now so we won't
3440 * get stuck waiting in the write for too long.
3441 */
3442 if (xfs_buf_ispinned(bp))
3443 xfs_log_force(mp, 0);
3444
3445 /*
3446 * inode clustering:
3447 * see if other inodes can be gathered into this write
3448 */
3449 error = xfs_iflush_cluster(ip, bp);
3450 if (error)
3451 goto cluster_corrupt_out;
3452
3453 *bpp = bp;
3454 return 0;
3455
3456corrupt_out:
3457 if (bp)
3458 xfs_buf_relse(bp);
3459 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3460cluster_corrupt_out:
3461 error = -EFSCORRUPTED;
3462abort_out:
3463 /*
3464 * Unlocks the flush lock
3465 */
3466 xfs_iflush_abort(ip, false);
3467 return error;
3468}
3469
3470STATIC int
3471xfs_iflush_int(
3472 struct xfs_inode *ip,
3473 struct xfs_buf *bp)
3474{
3475 struct xfs_inode_log_item *iip = ip->i_itemp;
3476 struct xfs_dinode *dip;
3477 struct xfs_mount *mp = ip->i_mount;
3478
3479 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3480 ASSERT(xfs_isiflocked(ip));
3481 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3482 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3483 ASSERT(iip != NULL && iip->ili_fields != 0);
3484 ASSERT(ip->i_d.di_version > 1);
3485
3486 /* set *dip = inode's place in the buffer */
3487 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3488
3489 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3490 mp, XFS_ERRTAG_IFLUSH_1, XFS_RANDOM_IFLUSH_1)) {
3491 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3492 "%s: Bad inode %Lu magic number 0x%x, ptr 0x%p",
3493 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3494 goto corrupt_out;
3495 }
3496 if (S_ISREG(VFS_I(ip)->i_mode)) {
3497 if (XFS_TEST_ERROR(
3498 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3499 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3500 mp, XFS_ERRTAG_IFLUSH_3, XFS_RANDOM_IFLUSH_3)) {
3501 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3502 "%s: Bad regular inode %Lu, ptr 0x%p",
3503 __func__, ip->i_ino, ip);
3504 goto corrupt_out;
3505 }
3506 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3507 if (XFS_TEST_ERROR(
3508 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3509 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3510 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3511 mp, XFS_ERRTAG_IFLUSH_4, XFS_RANDOM_IFLUSH_4)) {
3512 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3513 "%s: Bad directory inode %Lu, ptr 0x%p",
3514 __func__, ip->i_ino, ip);
3515 goto corrupt_out;
3516 }
3517 }
3518 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3519 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5,
3520 XFS_RANDOM_IFLUSH_5)) {
3521 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3522 "%s: detected corrupt incore inode %Lu, "
3523 "total extents = %d, nblocks = %Ld, ptr 0x%p",
3524 __func__, ip->i_ino,
3525 ip->i_d.di_nextents + ip->i_d.di_anextents,
3526 ip->i_d.di_nblocks, ip);
3527 goto corrupt_out;
3528 }
3529 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3530 mp, XFS_ERRTAG_IFLUSH_6, XFS_RANDOM_IFLUSH_6)) {
3531 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3532 "%s: bad inode %Lu, forkoff 0x%x, ptr 0x%p",
3533 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3534 goto corrupt_out;
3535 }
3536
3537 /*
3538 * Inode item log recovery for v2 inodes are dependent on the
3539 * di_flushiter count for correct sequencing. We bump the flush
3540 * iteration count so we can detect flushes which postdate a log record
3541 * during recovery. This is redundant as we now log every change and
3542 * hence this can't happen but we need to still do it to ensure
3543 * backwards compatibility with old kernels that predate logging all
3544 * inode changes.
3545 */
3546 if (ip->i_d.di_version < 3)
3547 ip->i_d.di_flushiter++;
3548
3549 /*
3550 * Copy the dirty parts of the inode into the on-disk inode. We always
3551 * copy out the core of the inode, because if the inode is dirty at all
3552 * the core must be.
3553 */
3554 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3555
3556 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3557 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3558 ip->i_d.di_flushiter = 0;
3559
3560 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3561 if (XFS_IFORK_Q(ip))
3562 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3563 xfs_inobp_check(mp, bp);
3564
3565 /*
3566 * We've recorded everything logged in the inode, so we'd like to clear
3567 * the ili_fields bits so we don't log and flush things unnecessarily.
3568 * However, we can't stop logging all this information until the data
3569 * we've copied into the disk buffer is written to disk. If we did we
3570 * might overwrite the copy of the inode in the log with all the data
3571 * after re-logging only part of it, and in the face of a crash we
3572 * wouldn't have all the data we need to recover.
3573 *
3574 * What we do is move the bits to the ili_last_fields field. When
3575 * logging the inode, these bits are moved back to the ili_fields field.
3576 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3577 * know that the information those bits represent is permanently on
3578 * disk. As long as the flush completes before the inode is logged
3579 * again, then both ili_fields and ili_last_fields will be cleared.
3580 *
3581 * We can play with the ili_fields bits here, because the inode lock
3582 * must be held exclusively in order to set bits there and the flush
3583 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3584 * done routine can tell whether or not to look in the AIL. Also, store
3585 * the current LSN of the inode so that we can tell whether the item has
3586 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3587 * need the AIL lock, because it is a 64 bit value that cannot be read
3588 * atomically.
3589 */
3590 iip->ili_last_fields = iip->ili_fields;
3591 iip->ili_fields = 0;
3592 iip->ili_fsync_fields = 0;
3593 iip->ili_logged = 1;
3594
3595 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3596 &iip->ili_item.li_lsn);
3597
3598 /*
3599 * Attach the function xfs_iflush_done to the inode's
3600 * buffer. This will remove the inode from the AIL
3601 * and unlock the inode's flush lock when the inode is
3602 * completely written to disk.
3603 */
3604 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3605
3606 /* generate the checksum. */
3607 xfs_dinode_calc_crc(mp, dip);
3608
3609 ASSERT(bp->b_fspriv != NULL);
3610 ASSERT(bp->b_iodone != NULL);
3611 return 0;
3612
3613corrupt_out:
3614 return -EFSCORRUPTED;
3615}
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}